Datasets:

Daniel-P-Gonzalez

/

OCD

like 0

# Introduction Service-oriented transformation and deep integration of manufacturing and service have become an inevitable trend. Because of its characteristics of agglomeration and heterogeneity, service-oriented transformation puts forward higher requirements for the supply of rules and institutions. In practice, there are institutional barriers to value co-creation, global network construction, digital technologies, big data applications, free flow as well as optimal allocation of human capital and productive service factors between the secondary and tertiary sectors. Discrimination in ownership, administrative monopolies and the "paradoxes in servitization" sometimes occur. The relative lags in institutional construction affect the enthusiasm and initiative of manufacturing enterprises to transform into a service-oriented industry. In addition, under the influence of de-globalization and partial trade protectionism, the uncertainty of international demand increases, while the growth of domestic demand slows down. The layout of global production networks and integrated supply chain management based on digital technology has led to the transformation of Industry 4.0 to Society 5.0. It creates an inverted pressure for manufacturing enterprises to transform into product service systems. It also innovates with the current and upcoming manufacturing and social paradigms. Efficient institutional supply and adaptation of existing institutions become major challenges in deepening structural reform on the supply side and promoting the transformation and upgrading of the manufacturing industry. Regarding the relationship between institutional innovation, institutional change and service transformation, scholars have done a lot of research and came up with a series of important conclusions. Firstly, the institution were sticky and path-dependent. There was a clear correlation between institutional change and industrial development. It was conducive to the adjustment and servitization of manufacturing industries to release the effect of institutional supply. Secondly, the government, as an important external institutional body for manufacturing enterprises, played an irreplaceable role in the transformation of the service-oriented manufacturing industry. The innovation of marketing institution can guide the coordinated development of manufacturing and productive services. Thirdly, the synergy between institutional innovation and technological innovation drove the transformation of manufacturing enterprises. This indicated that technological innovation changed the traditional product- centered manufacturing model and became a direct driver of service innovation. Moreover, technological innovation was never a stand-alone effect, but was influenced by the institutional environment in which the enterprises operated. The synergy of technological and institutional innovation drives the implementation of enterprise innovation strategies. The above research established a major theoretical basis for this paper. However, in most of these studies, the institutional environment is discussed as a whole, or the role of only one aspect is discussed. This does not allow for an in-depth interpretation of the complexity of the institution in the context of economic transformation from different perspectives. It is also difficult to guide companies to respond well to institutional pressures from different quarters. For economies in transition, the rules of the game are changing in a fundamental and comprehensive way. Digital technologies and big data are being widely applied and global value networks are gradually established. The paradigm shift from mass production to on-demand, personalized customer-driven and knowledge-based production has reshaped manufacturing, increasing institutional complexity and environmental uncertainty. It is therefore important to explore the impact of different isomorphic aspects of the institutional environment and how these different logics together influence the practice of servitization transformation within manufacturing in a complex context, in order to clarify the relationship between institutional innovation and service-oriented transformation. It is also conducive to institutional design and innovation in today’s complex institutional context. It can help the production paradigm shift rapidly from product to demand and service, and provide theoretical support for the construction of the institutional environment that matches the service-oriented transformation. An effective external institutional environment is the basis to the survival and development of a firm, providing boundaries of legitimacy to pursue and a relatively stable operating framework for innovation and strategic change. But institutional change is usually characterized by clear path dependence, which makes it difficult for efficient and superior institutions to spontaneously replace the existing institutional situation. This requires companies to comply with the rules and conditions established by the external institutional environment. while obtaining legitimacy and support, they should constantly engage in technological innovation to break through the excessive constraints of the institutional environment and avoid falling into a "legitimacy" crisis. In addition, to find a reasonable way to change the system, overcome the path dependence of the system and build an institutional environment that matches the transformation and upgrading of the manufacturing industry. Therefore, technological innovation is an indispensable support factor for the servitization transformation of manufacturing enterprises. However, there is a lack of evidence in theory and practice regarding the influence of technological innovation capabilities on the relationship between the different dimensions of the institutional environment and the servitization transformation of manufacturing. To address the above issues, this paper builds a driving model of service- oriented transformation from the perspective of institutional environment based on institutional theory. On this basis, a sample of 4313 manufacturing enterprises implementing service-oriented transformation strategies in China from 2016 to 2019 is selected as the study population. The differential impact of institutional environment on service-oriented transformation is empirically tested, and the moderating role of technological innovation capability is further examined. It is expected to provide an important empirical basis and reference for manufacturing enterprises in transition economies to realize service-oriented transformation, break down institutional barriers and optimize the institutional environment for the integrated development of the secondary and tertiary industries. The contributions of this paper are: firstly, it clarifies the mechanism of the influence of different institutional environments on service-oriented transformation and provides evidence to support manufacturing enterprises’ choice of a service-oriented transformation strategy that matches the institutional environment. Secondly, it confirms the heterogeneous role of technological innovation capabilities in the service- oriented transformation of manufacturing enterprises and enriches the boundary conditions of the institutional environment affecting it. Thirdly, the specifics of the interaction between the institutional environment, technological innovation capabilities and servitization transformation of manufacturing enterprises are elucidated from the perspective of property rights and regional differences. # Literature review and research hypothesis ## Institutional environment and service transformation The new institutionalism theory emphasized that the institutional environment is an important reference point for organizational behavior decision-making and strategy formulation. In order to obtain legitimacy and support from the environment, organizations must comply with the institutional rules and constraints in the external environment. As a result, when the institutional environment changes, enterprise behavior will adjust accordingly. In general, the better the institutional environment, the more it reduces the moral hazard of purchasing services and encourages cooperation in the process of service- oriented transformation of manufacturing enterprises, as well as improving predictability of behavior and reducing transaction costs. This facilitates the development of contract-intensive productive service activities and promotes the service-oriented transformation and upgrading of manufacturing enterprises. Conversely, if the institutional environment is imperfect, the policy continuity is weak and the business environment is poor, then it will be difficult for the market to function. The process of services-oriented transformation of manufacturing enterprises then faces the double dilemma of "legality" and "efficiency". Transition risks and market uncertainties will hinder its further transformation and upgrading. Institutions, as a set of rules of the game in human society, are formal, such as political systems, economic rules, the spirit of contract, etc. There are also informal ones, such as customs, practices, norms, etc. Scott (1995) constructed a theoretical model of a three-dimensional institution, including mandatory, normative and cognitive elements. It has been widely used by domestic and international scholars because it covers all factors of institutions. Among the domestic studies, the most influential is a set of marketization and business environment indices developed by Fan & Wang (2011). Institutional theory suggests that market mechanisms, such as market order, market rules, market competition, market supervision, market allocation of production factors, and market-based technological innovation systems, play an important role in transition economies. However, the development and strategic transformation of enterprises still rely to a large extent on non-market systems, such as government and social networks sources for support in terms of land, funding, policies, resources and others. At the same time, a legal business environment is also an important guarantee for the high-quality development of enterprises as well as the optimization and upgrading of industrial structures. So the service-oriented transformation of the manufacturing industry demands the government deregulate, improve the level of governance, enhance the legalization level, strengthen contract supervision and property rights protection, improve the market mechanism and avoid opportunistic behavior. Therefore, in the context of China’s current economic transition, this paper followed Fan & Wang’s ideas of an institutional environment, focusing on three dimensions, the level of government governance, market standardization and legalization. Governments play an important role in all stages of a country’s economic development, including the formulation of macroeconomic policies and the improvement of specific economic institutions. The government is the important external institutional subjects of the manufacturing enterprises. In the early stages of service transformation, the government plays an essential role in top- level framework design, policy guidance, tax support, information security, platform construction and priority allocation of resources. To a certain extent, the government’s administrative intervention has compensated for market failure and guided the development direction of the market economy. The supportive behavior of the government breaks down many barriers to the development of the manufacturing services industry and gradually changes the tendency of government departments to "work in their own way" and "emphasize manufacturing rather than services". The degree of integrity of a government determines its administrative capacity and efficiency. A clean government can curb distortions in public policy implementation and unfair resource allocation caused by corruption and achieve openness, fairness and impartiality. A higher level of government governance can enhance the precision of policy guidance and the effectiveness of resource allocation. It helps manufacturing enterprises expand into higher value-added services. Thus, the following assumption is made: 1. Hypothesis 1: There is a positive correlation between the level of government governance and the service-oriented transformation of manufacturing enterprises. The dual power of a strong government and a strong market has been present during the transition phase of the economic system. A well-developed market institutional environment is a hallmark of the enterprise’s pursuit of efficiency. Firstly, a comprehensive market mechanism promotes technological progress, scientific management, freedom of market competition, lower transaction costs and increases inter-firm exchanges, cooperation or contracting. As a result, there will be more mergers and acquisitions, and more opportunities for manufacturing to transform into service-oriented enterprises. Secondly, in the environment of a sophisticated market system, information is freely available. The cost of finding and matching external resources is reduced, the propensity to negotiate and cooperate is stimulated, and external communication and cooperation are increased. Most importantly, it facilitates the possibility of transformation. At the same time, with sufficient competition, effective use of information and high incentive compatibility, companies become more sensitive to the market. The closer the product research & development and service innovation efforts of enterprises are to customer needs, the higher the success rate of service transformation around customer demands. Thirdly, the improvement of market regulation mechanisms is often accompanied by the continuous development of product markets, factor markets and intermediary organizations. This leads to the optimization and enhancement of resource allocation efficiency. It makes the redundant resources of manufacturing enterprises flow to service link with high added-value, which in turn promotes the service transformation of manufacturing enterprises. In conclusion, the higher the degree of market regulation, the stronger the willingness of manufacturing enterprises to transform into service-oriented, and the greater the opportunities for manufacturing enterprises to transform into service- oriented. Thus, the following assumption is made: 1. Hypothesis 2: There is a positive correlation between the degree of market regulation and the service transformation of manufacturing enterprises. The in-depth integration of business environment optimization and legal construction is an important part of the strategy of law-based governance. The 20^th CPC Central Committee Report mentions that a good business environment implies a high level of legalization. Enterprises are more willing to use intermediary services to access external information and heterogeneous resources, thus reducing the infringement of their rights. This will facilitate the transformation of manufacturing enterprises in a rule of law, transparent and predictable external environment. Studies found that the development of service industry is highly sensitive and dependent on institutions. As for service products, they are typically institution-intensive due to their trust, posteriori and heterogeneous attributes. There is a significant positive correlation between the quality of contract maintenance systems and the share of services measured by a country’s level of rule of law. North (2008) argued that a regulated country should have an efficient judicial system to resolve disputes over the signing and enforcement of contracts between the private and public sectors. This facilitated the division of labor and transactions in society and promoted economic growth. A higher level of legalization can provide more robust public services, strengthen property rights protection, improve contract enforcement and reduce financing constraints. A legal environment that is open and transparent is conducive to lowering the risk of service-oriented transformation of manufacturing enterprises and increasing the willingness and success rate of transformation. The high level of legal regulations provides a strong guarantee for the service-oriented transformation of manufacturing enterprises. Thus, the following assumption is made: 1. Hypothesis 3: There is a positive correlation between the level of legalization and the service transformation of manufacturing enterprises. ## Technological innovation ability Technological innovation ability is gradually formed based on the theory of technological innovation, which covers both static and dynamic dimensions. Static technological innovation ability essentially refers to the knowledge stock resources possessed by enterprises that support the realization of technological innovation activities. The ability to develop new ideas and innovative products, obtain market value, and achieve dynamic development of the organization through the coordination and operation of the above knowledge stock resources, forming the enterprise’s dynamic innovation, is referred to as dynamic technological innovation ability. It not only reflects the strength of product innovation and process improvement, but is also closely integrated with the overall strategy of the enterprise’s internal system. It is a comprehensive embodiment of innovation decision-making capabilities, research & development capabilities, manufacturing capabilities, organizational management capabilities and marketing capabilities. In the current context of open innovation, technological innovation capability is a comprehensive and synergistic capability based on technological innovation. It includes the ability to integrate technological innovation elements, the ability to commercialize technology and the synergistic ability of technological innovation systems, as well as the ability to respond to changes in the external environment such as customer demands and competitors. It has been discovered that the ability of enterprises to integrate technology after collaboration with other partners has a different impact on product innovation and innovation performance. The ability to innovate technologically becomes the foundation for manufacturing businesses to gain a competitive advantage. It achieves service transformation by improving service quality and service innovation capabilities, and influencing the effectiveness of service transformation. Manufacturing companies integrate technological innovation into their product production and customer service processes. They provide feedback on new customer needs during the use of their products and interact with their customers continuously. On the one hand, technological innovation can bring technological advantages to enterprises and is the technical basis for manufacturing enterprises to provide solutions and value-added services to their customers. On the other hand, technological innovation ability can enhance service innovation ability. Research has found that the research and development intensity of manufacturing enterprises can significantly affect the development of new services. Manufacturing enterprises can gain the trust of customers through their superior technological innovation ability, overcome the weakness of inexperience in service and bring about the success of service innovation strategy. First, technological innovation breaks through the excessive restrictions imposed by the institutional environment on enterprises’ pursuit of legitimacy to a certain extent, enhancing their flexibility, adaptability and policy sensitivity. After the government’s favorable policies are introduced, enterprises with strong technological innovation capabilities are able to quickly optimize resource allocation, utilize their technological advantages, innovate service products, meet market demands and lead the development of the industry. As a result, the impact of government-initiated institutional innovation is amplified. Technological innovation reduces the dependence of enterprises on government governance. Even if government policies are unfair and administrative intervention is strong, enterprises can enhance the competitiveness of their own products, innovate service products and expand into higher value-added service segments. Second, technological innovation capability serves as a lever between improving market mechanisms and achieving technological innovation. Enterprises with strong technological innovation capabilities can quickly respond to market demand, collect all types of innovation elements, and design and develop high-end service content such as services and comprehensive project solutions. This creates a higher than average profit-added value in the market, meets the diversified needs of the market, amplifies the positive impact of market regulation and development, and generates a market spillover effect. Thirdly, studies show that the business environment has a gateway effect on technological innovation capabilities. Enterprises with strong technological innovation capabilities can first surpass the threshold value and demonstrate the comparative advantage of the business environment. Then expand the legal business environment to facilitate the service-oriented transformation of manufacturing enterprises. Further, a benefit distribution mechanism centered on intellectual property rights will build a sustainable competitive advantage. Thus, the following hypotheses are proposed: 1. Hypothesis 4: Technological innovation capabilities positively moderate the relationship between the institutional environment and the service- oriented transformation of manufacturing enterprises. 2. Hypothesis 4a: Technological innovation capabilities positively moderate the relationship between the governance of government and the service- oriented transformation of manufacturing enterprises. 3. Hypothesis 4b: Technological innovation capabilities positively moderate the relationship between market regulation and the service-oriented transformation of manufacturing enterprises. 4. Hypothesis 4c: Technological innovation capabilities positively moderate the relationship between the legalization and the service-oriented transformation of manufacturing enterprises. Based on the above analysis, the framework model of this paper is built, as shown in. # Research method ## Sample and data collection According to the 2012 edition of the *Listed Companies Industry Classification Index* for manufacturing enterprises of the China Securities Regulatory Commission, this paper conducted the sample selection. The data sources for this study were the China Stock Market Accounting Research database, the WIND database and the *Chinese Business Environment Index by Province 2020 Report (hereinafter referred to as the Report)*. To ensure the normality of the sample data, the following treatments were made, 1. refer to Sun et al. (2008) and Li et al. (2015) for the measurement of service transformation in manufacturing enterprises, and after manual screening, the data of non-service transformation enterprises were excluded. 2. the data of ST and \*ST listed companies were excluded. 3. outliers and missing samples were excluded. 4. continuous variables were winsorized by 1% up and down to avoid the effect of low end values. Finally, 4,313 observations were obtained. Stata 14.0 was used for data processing and analysis. ## Variable measurement The service-oriented transformation of manufacturing enterprises is the interpreted variable. According to the motivation for manufacturing enterprise transformation and the endogenous and exogenous connotation attributes of transformation, it can be divided into intra-industry service-oriented transformation and cross-industry service-oriented transformation (Li & Chen, 2011). The intra-industry service transformation refers to the extension of the industrial chain, that is, from the low-end value chain to the high value-added value chain at both ends. It is measured using the adjusted value-added method. Cross-industry boundary service transformation refers to enterprises that are unable to share existing resources within their established industries and therefore break down industry boundaries to serve a wider range of business areas and explore higher value-added opportunities. The Herfindahl-Hirschman Index was used to measure this (as the study sample was a manually screened sample of manufacturing companies that were confirmed to have undergone service transformation, their higher degree of diversification can reflect to some a greater number of inter-industry service transformations). The institutional environment, as an explanatory variable, includes three dimensions: the level of government governance, the level of market standardization and the level of legalization. Based on the characteristics of the institutional environment during the economic transition period, the Report by Wang et al. (2020) was used. Government governance was measured by the Index of "Openness, Fairness and Justice Policies" and the Index of "Administrative Intervention and Government Integrity and Efficiency". The market regularization was measured by the "Market Environment and Intermediary Services" Index. The level of legalization was measured by the "Legal Environment for Enterprise" Index. It includes the effective law enforcement of the public security law, the performance of the enterprise contract, the protection of intellectual products, the property and personal safety of the enterprise and the operator. The moderating variable refers to the ability to innovate technologically. The indicators with a high degree of consensus in domestic studies were all measured in terms of innovation management capability, research & development input intensity and input-output capability. Based on the characteristics of the information disclosed in the database, research & development input intensity was more objective and reliable, therefore, based on the existing studies, it was adopted in this study. In order to ensure the authenticity and reliability of the research conclusions, control variables such as enterprise size, enterprise age, asset specificity, organizational redundancy, enterprise growth, property right nature are added in this paper. The type of industry was also controlled to avoid the effects of heterogeneity arising from differences in manufacturing industry segments. The measurement of each variable is shown in. ## Model To examine the impact of the dimensions of the institutional environment on the service transformation of manufacturing enterprises, the following econometric model was constructed in this paper. $$ST_{i,t} = \beta_{0} + \beta_{1}Institution_{i,t} + \beta_{2}Size_{i,t} + \beta_{3}AS_{i,t} + \beta_{4}Slack_{i,t} + \beta_{5}Age_{i,t} + \beta_{6}Growth_{i,t} + \beta_{7}State_{i,t} + Year + IND + \mu_{i} + \varepsilon_{i,t}$$ where *ST*_*i*,*t* represents the degree of service-oriented transformation of manufacturing enterprises, including the service-oriented transformation intra-industry (*ST*1_*i*,*t*) and the across industry boundaries (*ST*2_*i*,*t*). *Institution*_3*i*,*t* represents the institutional environment for the manufacturing enterprises, including three dimensions, the government governance (*Gov*_*i*,*t*), the market regularization (*Mar*_*i*,*t*) and the legalization (*Leg*_*i*,*t*). *μ*_*i* denotes individual effects that do not vary over time. *ε*_*i*,*t* indicates random error terms. For controlled variables, there are enterprises size (*Size*_*i*,*t*), enterprises age(*Age*_*i*,*t*), asset specificity (*AS*_*i*,*t*), organizational redundancy (*Slack*_*i*,*t*), enterprises growth (*Growth*_*i*,*t*), nature of property (*State*_*i*,*t*), industry factors (*IND*) and sampling year (*Year*). To test the impact of technological innovation capability on the relationship between the institutional environment and the service-oriented transformation of manufacturing enterprises, the following econometric model was constructed. $$ST_{i,t} = \beta_{0} + \beta_{1}Institution_{i,t} + \beta_{2}TIC_{i,t} + \beta_{3}Institution_{i,t} \times TIC_{i,t} + \beta_{4}Size_{i,t} + \beta_{5}AS_{i,t} + \beta_{6}Slack_{i,t} + \beta_{7}Age_{i,t} + \beta_{8}Growth_{i,t} + \beta_{9}State_{i,t} + Year + IND + \mu_{i} + \varepsilon_{i,t}$$ where *TIC*_*i*,*t* denotes technological innovation capability. The names and meanings of the other variables are the same as in model. # Results and discussion The results of the descriptive statistical analysis of the main variables are shown in, including the mean, standard deviation, minimum and maximum values of each variable. The mean value of ST1 was 0.885 and 0.192 for ST2, respectively. It indicated that intra-industry service transformation was currently dominating service transformation in Chinese manufacturing enterprises. Only some enterprises carried out service transformation across industry boundaries as an assistance. The average value of the technological innovation ability of enterprises was 0.429. This demonstrated that the majority of companies implementing service transformation attached more importance to technological innovation and had some capacity for technological innovation. The correlation coefficients between the variables are shown in. The results revealed correlation coefficients of 0.130 (*P*\<0.01) between Gov and Mar, 0.575 (*P*\<0.01) between Gov and Leg, and 0.413 (*P*\<0.01) between Mar and Leg, respectively, indicating a significant positive correlation between them. To avoid bias in the results due to the problem of multicollinearity and heteroskedasticity among the variables, the paper also conducted the variance inflation factor (VIF) test for covariance and the WHITE test. The VIF test showed that the maximum value for each variable was 7.46, which was less than 10, indicating that the problem of multicollinearity was not serious. Models M1, M2 and M3 tested the impact of the three dimensions of the institutional environment on service transformation intra-industry, respectively. Models M4, M5 and M6 tested the impact of the three dimensions of the institutional environment on service transformation across industry boundaries, respectively. As shown in. The results showed that the level of government governance significantly positively influenced intra-industry service transformation at the 1% level (β = 0.022, *P*\<0.01) and cross-industry boundary service transformation at the 10% level (β = 0.021, *P*\<0.1), respectively. It indicated that improved government governance significantly contributed to the service transformation of manufacturing enterprises, and was more pronounced with intra-industry transformation. Hypothesis 1 was confirmed. In addition, the market regulation significantly and positively affected intra- industry service transformation at the 1% level (β = 0.040, *P*\<0.01), while there was no significant effect on service transformation across industry boundaries. The likely reason for this was that enterprises prioritized resources for in-industry transformation based on transformation risk and market efficiency considerations. Hypothesis 2 was partially confirmed. Besides, legalization significantly and positively influenced intra-industry service transformation at the 1% level (β = 0.020, *P*\<0.01) and cross-industry boundary service transformation at the 10% level (β = 0.013, *P*\<0.1). It demonstrated that the improvement of the legal business environment is conducive to the successful service-oriented transformation of enterprises. Moreover, it was significantly more conducive to transformations within industries than across industry boundaries. Hypothesis 3 was confirmed. The results of the moderating effect of technological innovation capability are shown in. The regression results showed that for intra-industry transformation, the interaction term between technological innovation capability and the level of government governance was significantly positive (β = 0.435, *P*\<0.01). This indicated that technological innovation capability strengthened the degree of influence of the level of government governance on the intra-industry service transformation of manufacturing enterprises. For cross-industry boundary transition, the interaction term between technological innovation capability and the level of government governance level was significantly negative (β = -0.479, *P*\<0.01). This showed that the technological innovation capability negatively moderated the level of government governance and the level of service transformation of manufacturing enterprises across industry boundaries. The possible explanation was that the pursuit and consideration of legitimacy and efficiency differed when there were differences in the technological innovation capabilities of different firms. This could thus have a heterogeneous impact on the areas in which enterprises are involved in service transformation. Technological innovation capability had a positive effect on the relationship between the degree of market regulation and the service transformation of manufacturing firms within their industries (β = 0.423, *P*\<0.1). However, there was no significant influence on the relationship between the degree of market regulation and the service transformation of manufacturing enterprises across industry boundaries. In addition, technological innovation ability had a positive effect on the relationship between the level of legalization and the service transformation of manufacturing enterprises (β = 0.206, *P*\<0.1). In contrast, there was no significant influence on the relationship between legalization and the service transformation of manufacturing enterprises across industry boundaries. Therefore, the relationship between technological innovation capability and the institutional environment and the service transformation of manufacturing enterprises across industry boundaries all had a significant positive effect, in line with the hypothesis. However, none of the moderating effects on the positive influence relationship between the institutional environment and the service transformation of manufacturing enterprises across industry boundaries were proven, as shown in. This led to an in-depth consideration of the mechanisms underlying the role of technological innovation capability, institutional environment and service-oriented transformation of manufacturing enterprises across industry boundaries. Firstly, the service transformation of manufacturing companies across industry boundaries required more cross-border service innovation. While this was affected by multiple factors, such as knowledge search, integration ability, managers’ focus, innovation paradigm, product complexity, etc., rather than a single factor of technological innovation acting independently. Enterprises were expected to search for information, knowledge and access to a wider range of technological resources across organizational boundaries. Further, these resources needed to be internalized, absorbed and integrated. Therefore, it was difficult to identify and measure the part of the institutional environment that played a separate role in the cross-border transformation of manufacturing enterprises. Secondly, there was heterogeneity in the choice of innovation approach by manufacturing enterprises. It has been shown that in the present state of established innovation inputs, exploitative innovation was chosen much more than exploratory innovation based on risk and difficulty considerations. Utilization-based innovation was an ongoing innovation of an existing product or service based on available knowledge. It was more based within the industry. However, exploration-based innovation was a departure from the original technological trajectory to expand into new business areas through a broader search for knowledge and access to external heterogeneous resources. There was a very high degree of uncertainty and risk. In other words, utilization-based innovation preferred to work within the industry, while exploration-based innovation was more conducive to cross-border transformation. As a result, technological innovation capabilities, mainly utilization-based innovation, played a more significant moderating role in the intra-industry servitization transformation of manufacturing enterprises. The following robustness tests were conducted to ensure the reliability of the study findings. 1. Changed the method of variable measurement. As the institutional environment changes, it first needs to act on managers’ perceptions of the external environment and make strategic adjustments to adopt service transformation strategies. Therefore, as the dependent variable, the degree of service transformation at manufacturing firms, may have a lag effect. Therefore, the degree of service transformation was re-estimated with a one- period lag. The result showed that the significance of the influence of the institutional environment and technological innovation capability on the degree of service transformation of manufacturing enterprises did not change obviously. 2. Added control variables. The overall level of economic development of a region and the business efficiency of the enterprises themselves will indirectly affect the degree of service-oriented transformation of enterprises. Therefore, this paper also controlled for the level of inter- provincial GDP per capita and the operating efficiency of enterprises for correlation tests. The main hypothesis is still valid. 3. Changed the parameter estimation method. A 50% quantile regression was used to further understand the potential relationship between the independent variable institutional environment and the dependent variable degree of servitization transition. The approach of Zhang et al. (2013) and the advantage of quantile regression being insensitive to outliers were adopted. 4. Changed the sample size. As there was policy and market efficiency variability in the institutional environment in different regions. Therefore, referring to Wei et al. (2010), a re-estimation test was conducted after excluding some of the western and northeastern provinces from the sample, with reference to the practice of Wei et al. (2010), as shown in Table 8. The new results were found to be insignificantly unchanged from the previous findings. This further indicated the robustness of the paper’s conclusions. # Further discussion ## Heterogeneity analysis of the nature of enterprise property In the sample, there were 1,257 state-owned enterprises and 3,056 non-state- owned enterprises (private, foreign or other). It showed that non-state manufacturing enterprises were the mainstay of industrial structural transformation and upgrading, and were more willing to transform into services than state-owned manufacturing enterprises. State-owned enterprises had less incentive to pursue legitimacy due to their controlling interest in the central or local government and more government intervention. Policy sensitivity and market efficiency were lower than for non-state-owned enterprises, corresponding to their weaker environmental dependence for service-oriented transformation. Non-state enterprises, on the other hand, had a higher degree of flexibility and market resilience, a high degree of sensitivity to policy and a strong desire for technological innovation out of a quest for legitimacy. Their service transformation was more dependent on the environment. Therefore, it is necessary to further explore the differential impact of the institutional environment and technological innovation capabilities on manufacturing enterprises with various property rights systems, and thus develop divergent strategies. While regulating the service-oriented transformation of state-owned enterprises, the service transformation and upgrading of non-state-owned manufacturing enterprises should be well guided and supported. The institutional environment had a far greater positive impact on non-state enterprises than on state enterprises, as shown in. The three dimensions of institutional environment, government governance, market regulation and legalization, all significantly and positively influenced the degree of intra- industry transformation of non-state manufacturing enterprises at the 1% level. However, only market regulation significantly and positively influenced the intra-industry service transformation of state-owned enterprises. This was related to the current system of state-owned enterprises in China. Whereas the development of non-state-owned enterprises was more influenced by factors such as financing constraints and fiscal and taxation policies, and had significant external institutional sensitivity. The ability to innovate technologically enhanced the positive impact of the institutional environment on the intra- industry transformation of state-owned manufacturing enterprises. In contrast, there was little influence on the relationship between the institutional environment and the intra-industry transformation of non-state-owned manufacturing enterprises. On the one hand, this was due to the significant differences between state-owned enterprises and non-state-owned enterprises in terms of innovation incentives, resource base, etc., and their sensitivity to external innovation resources. On the other hand, the intra-industry service- oriented transformation of manufacturing enterprises was mainly a utilization- based innovation with accumulated effectiveness, which was more influenced by the innovation resources that could be gathered in the relevant fields. Therefore, when state-owned enterprises had advantages in financial subsidies and financing channels, and were less constrained by the lack of innovation resources, they could effectively take benefits of innovation resources as long as their innovation capacity was improved. This promotes the service-oriented transformation and upgrading of manufacturing enterprises. Non-state enterprises, on the other hand, had more limited external financing constraints, policy support and relatively insufficient innovation resources. Even if technological innovation capabilities improved, there was no discernible moderating effect on service transformation in the short term. There was also variability in the impact of the institutional environment and technological innovation capabilities on the service transformation of manufacturing enterprises across industry boundaries. However, this variability was less than the impact on transformation within industries, as shown in. The level of government governance and legalization in the institutional environment facilitated the degree of service transformation of non-state manufacturing enterprises across industry boundaries. This positive impact was weakened by the ability to innovate with technology. This was primarily because for cross-border servitization transformation, it relied more on exploration-based innovation. This was primarily because for cross-border service transformation, it relied more on exploration-based innovation. It required access to more external heterogeneous resources, and both investment and risk were relatively high. Non- state-owned manufacturing enterprises with a certain level of technological innovation capability and innovation resources will prioritize intra-industry transformation over cross-border transformation. However, the degree of market regulation had no obvious influence on the cross-border service transformation of non-state-owned manufacturing enterprises. This was mainly due to the fact that the cross-border transformation itself spanned a relatively large distance from the existing business of the enterprise, and the initial investment was also large. The spontaneous adjustment of resources across industries guided by market laws was also a slow process, and the effects were presented with a strong lag. The dimensions of technological innovation capability and institutional environment had no obvious influence on the cross-border transformation of state-owned enterprises, which was consistent with the expectations of this paper. State-owned enterprises were influenced by the nature of their controlling stakes and undertook important national strategic planning and social responsibilities. Achieving cross-border transformation was a major strategic change and was mainly influenced by the controlling decision maker. The influence of environmental attributes was more indirect. ## Analysis of regional differences The development of China’s regions is uneven. There are significant regional differences in institutional factors such as the level of government governance, the market-oriented reform process and the legalized business environment. The institutional environment in the eastern region is significantly better than that in the western region. Therefore, the impact of regional differences on the transformation of servitization needs to be further removed to enhance the generalizability of the study’s findings. Based on the grouping of provinces in Wang et al.’s (2020) Report, the country was divided into four regions, namely the eastern, central, western and northeastern regions, as detailed in the notes. The results of the regression analysis showed a high degree of consistency between the central and western regions, so these two regions were combined for analysis. The Northeast region had a smaller sample size and was less representative and was discarded. Therefore, the results are presented in this paper in two groups for the eastern and midwest regions, as shown in. The institutional environment in the eastern region had the most significant impact on the service-oriented transformation of manufacturing enterprises. The government governance, market regulation and legalized business environment all positively influenced the service transformation of manufacturing enterprises within their industries at the 1% level. However, only legalized business environment had a significant positive impact on the service transformation across industry boundaries at the 1% level. In contrast, government governance and market regulation did not show a direct impact on service transformation across borders. The impact of the level of government governance on the service-oriented transformation of manufacturing enterprises was significantly mitigated by technological innovation capability. But neither the institutional environment nor the technological innovation capability in the midwest regions showed a significant impact on the service- oriented transformation of manufacturing enterprises. This was related to the large differences in the level of economic development, regional innovation capacity and policy and institutional environment in the midwest regions. In addition, there were fewer samples eligible for the study in these two regions. Further research showed that there was a large gap between the level of transformation of manufacturing services in the eastern region and the midwest regions. Meanwhile, there were significant differences in the roles played by the institutional environment and technological innovation capabilities. In the eastern region, the government policy environment was well established, the market was more standardized and the level of technological innovation capacity was generally higher. As a result, the influence of the institutional environment and technological innovation capacity on service transformation exhibited regular development. To some extent, the current institutional environment met the need for service-oriented transformation and upgrading of the manufacturing sector. This was in accordance with the findings of Zhu et al. (2005), Wei et al. (2010), Du et al. (2021). Affected by innovation infrastructure, quality of industry-university-research linkages, technology spillover effects and industry cluster environments, the eastern and midwest regions had a pronounced gap in their economic development levels. The gap in innovation ability was even more obvious and tended to widen. # Conclusions and recommendations A suitable institutional environment was an important driver for the implementation of a service-oriented transformation strategy in enterprises. Enterprises matched their internal strategies with the external environment through continuous strategic change and institutional innovation. This in turn led to access to the heterogeneous resources needed for organizational legitimacy and development. While technological innovation can, to a certain extent, reduce enterprises’ perception of stress in the external institutional environment, allowing them to gain more development opportunities and stronger environmental adaptability. This study showed that, 1. The improvement of government governance contributed significantly to the service transformation of manufacturing enterprises, and the transformation was most evident within industries. 2. The development of market regulation facilitated the intra-industry service transformation of manufacturing enterprises in the short term. It had no significant impact on service transformation across industry boundaries. 3. The promotion of a legislated business environment was conducive to the successful service-oriented transformation of enterprises. Moreover, the boosting effect was significantly greater within the industry than across the industry boundary. 4. Technological innovation capabilities were the basis for the strategic transformation of enterprises. In the industry, mostly utilization-based innovation played a role, influenced by the current choice of enterprise innovation approaches. This intensified the influence of the institutional environment on manufacturing firms undertaking intra-industry servitization transformations. However, the impact of the institutional environment on service transformation across industry boundaries was not significantly moderated. 5. The impact of the institutional environment on non-state enterprises was much stronger than that of state enterprises. Also, the impact on the extent of intra-industry transformation was greater than that across industry boundaries. 6. The institutional environment varied considerably across regions, and the impact on the service-oriented transformation of manufacturing enterprises also differed significantly. It was easy to observe through the service transformation practices of Chinese manufacturing enterprises that in order to avoid the risks of transformation and overcome the "service paradox", most of them preferred the relatively robust transformation within the industry. As a result, a favorable institutional environment was firstly revealed to be effective in the transformation within the industry. The risks and upfront investment of transitioning across industry boundaries were relatively large. Therefore, only a few large manufacturing enterprises with strong technological innovation capabilities and strengths chose to enter. At the same time, cross-border transformation required cross- border information and knowledge searches and the integration of more external resources. The factors that influenced it were more diverse. Specifically, the institutional environment integrated with other situational factors worked together, while the role of a single factor was not obvious. In-depth analysis revealed that government policy guidance and improvement of the business environment can more directly promote the cross-border service transformation of manufacturing enterprises than the market’s regulated development. This was because favorable policies can quickly influence the allocation of resources by the enterprises. They will prioritise resources in areas with good growth prospects and high added value. Improvements in the business environment will directly reduce the cost of doing business and the barriers to entry. It will also change the manager’s perception of the environment, and lead to more optimistic strategic decisions. Following the efficiency mechanism, the long- term effects of a regulated market are self-evident. However, it will require a prolonged transmission period and therefore the short-term effects are difficult to see. Based on the above findings, this paper proposes two main levels of government and enterprises. 1. Government departments should continue to promote institutional innovation and build an institutional environment that matches the service- oriented transformation of the manufacturing industry. This will increase manufacturing enterprises’ willingness and success rate in service transformation. At the primary stage of service-oriented transformation, the government needs to improve its governance, establish long-term market regulation mechanism and optimize the legalized business environment, all of which are crucial. This is because, firstly, the government still controls key resources to a large extent during the economic transition period. The government’s policy direction is the wind vane for business transformation. More manufacturing enterprises can benefit from the service- oriented transformation by enhancing the degree of openness, fairness and justice of government policies. In particular, for non-state enterprises, which are more policy-sensitive and numerous in number, increasing policy inclination, reducing administrative intervention, enhancing government integrity and efficiency, and lowering industry entry barriers & resource acquisition costs can effectively carry out cross-industry boundary transformation on the basis of strengthening their transformation within the industry. Secondly, the pursuit of efficiency is inevitable under the premise of conforming to government policy guidance. The establishment of a long-term market regulation mechanism will further optimize the allocation of resources. It will guide the integration of resources across borders on the basis of promoting service-oriented transformation within the industry, which will in turn lead to cross-border service-oriented transformation and enhance the degree of transformation and upgrading. Thirdly, enhancing the impartiality and efficiency of the judiciary, creating a good legalized business environment, promoting the normal performance of contracts, improving the level of social credit and reducing transaction costs can all improve the success rate of transformation. Finally, when the state carries out the top-level system design, it can appropriately tilt its policies towards the midwest regions. It should be done to take advantage of the situation and adapt to local conditions. Thus, the gap between the midwest and the eastern regions can be narrowed and the synergistic development of the eastern part leading the central and western parts can be achieved. 2. Enterprises should strengthen their sensitivity to the external environment. Identify and respond timely to the pressures brought by the different components of the institutional environment on the transformation of enterprises, while enhancing technological innovation capabilities. Ensure the positive effect of the institutional environment on service- oriented transformation and further avoid or mitigate institutional risks. Specifically, firstly, enterprises should correctly interpret the guidance of government policy documents and relevant laws and regulations, and make corresponding adjustments to their transformation strategies. When the state introduces favorable policies for service-oriented transformation, enterprises should quickly integrate internal and external resources, seize external opportunities and explore transformation opportunities. In particular, non-state enterprises with strong policy sensitivity should pay more attention to policy trends. Enterprises need to actively participate in fair competition and promote the development of product and factor markets. Based on the pursuit of efficiency, align themselves with the benchmark companies in the industry that are transforming their services. To promote a higher degree of market regulation, which in turn will optimize resource allocation to generate more transformation opportunities. Increase the protection of property rights and contract enforcement. Access to external information and heterogeneous resources through intermediary services in accordance with the law. Collaborate with the government to create a beneficial legal environment, which in turn will promote successful transformation. Secondly, technological innovation is a constant topic for enterprises. Traditional manufacturing industry should also nurture its technological innovation capabilities and choose the type of innovation that matches the type of transformation. Manufacturing enterprises undertaking service-oriented transformation in the industry should continue to enhance their utilization-based innovation capabilities. For enterprises conducting cross-industry boundary service transformation, they should focus on building breakthrough innovation capabilities alongside utilization-based innovation to better fit the external institutional environment and bring about maximum transformation effectiveness. # Limitations and future direction This study has certain limitations that must be mentioned. The data subject to institutional environment limitations are from the *Report*, which is only updated to the end of 2019. Although the institutional environment changes slowly and with some lags. However, the last three years have been in the post- epidemic era and the rapid application of digital technology could have an impact on the institutional environment and service-oriented transformation. Future research can be conducted in two ways. On the one hand, digital transformation technology and relevant data in the post-epidemic era are introduced into the model as moderating variables to comprehensively examine the impact mechanism of the institutional environment on service-oriented transformation under the influence of these combined factors. On the other hand, data from Chinese enterprises will be used as a sample to directly study the new influence of digital transformation, big data, artificial intelligence and other technologies on the servitization transformation of manufacturing enterprises, so as to provide the empirical evidence for the servitization transformation of global manufacturing enterprises. # Supporting information First I would like to thank the support of the National Social Science Foundation of China, and I also want to show my deepest gratitude to Professor Su, my supervisor, for her constant encouragement and guidance. My research partner, Shengshi Zhou, was instrumental in defining the path of my research. For this, I am extremely grateful. I would also like to thank my reviewers and those who made valuable suggestions for the article. [^1]: The authors have declared that no competing interests exist.

# Introduction Colorectal cancer (CRC) is the third most frequently diagnosed cancer and the second leading cause of cancer deaths in the United States and Europe, as well. Although CRC incidence, in economically developed countries, is stabilized or rather declined, there is still the question for a greater understanding of the biology of angiogenesis and invasion with a view to the discovery of antiangiogenic/anti-invasive molecules. Accumulating evidence demonstrates the crucial role of proteolytic enzymes such as matrix metalloproteinases (MMPs) and closely related ADAMTSs (a disintegrin and metalloproteinase with thrombospondin motif) in cancer development and progression. ADAMTS are a distinct group of zinc-dependent metalloproteinases and display sequence similarities with the reprolysin family of snake venomases. The complete human ADAMTS family comprises 19 genes (ADAMTS-5 and ADAMTS-11 share the same gene) and although they are soluble proteins, many of them appear to bind the extracellular matrix through their thrombospondin motifs or their spacer region. ADAMTS have varying functions, including specific cleavage of the matrix proteoglycans aggrecan, versican and brevican (ADAMTS-1, -4, -5, -8 and -15), inhibition of angiogenesis (ADAMTS-1, -8) and collagen processing (ADAMTS-2, -3 and -14). ADAMTS-1 and -8 are considered to be anti-angiogenic factors, because of the interaction between their thrombospondin motifs and CD36, a membrane glycoprotein receptor of endothelial cells. These proteases have been shown to inhibit VEGF-induced angiogenesis in the chorioallantoic membrane (CAM) assay and suppress FGF-2-induced vascularisation in the cornea pocket assay. However, ADAMTS-1 should be considered as a molecule with anti- or pro-metastatic effects depending on the cleavage site, during its auto-proteolytic cleavage. ADAMTS-8, seems also to inhibit epidermal growth factor receptor signaling along with decreased levels of phosphorylated MEK and ERK. Moreover, the 12q12 gene locus of ADAMTS-20 has been found to be subject to translocations and other alterations in human malignancies and also several pathologic conditions, including Parkinson’s disease. There is a wealth of evidence demonstrating that ADAMTS are deregulated in human cancer. Indeed, previous studies have shown that ADAMTS-20 is overexpressed in brain and breast carcinomas, suggesting that this protease could play a role in tumor progression. For ADAMTS-1, it has been shown that overexpression of its full-length isoform enhances tumor growth and promotes pulmonary metastasis of TA3 mammary cells or Lewis lung cells, while it has been found decreased in non- small cell lung carcinoma, pancreatic tumors and prostate cancer cell lines. It has also been shown that ADAMTS-1, -5, -9, -12, -15 and -18 gene promoters are hypermethylated in colorectal cancer, suggesting decreased expression of these enzymes in this state. In addition, both ADAMTS-15 and -18 genes undergo frequent mutations in colorectal cancer cells, but no evidence has yet presented of their effect in expression or function of the enzymes. In contrast, ADAMTS-4 and -5 are upregulated in glioblastomas (GBMs), with a possible role in increased degradation of brevican thereby increasing invasive potential. Increased versican, which may be related to modulated ADAMTS expression, is also observed in canine colonic adenomas and carcinomas. Versican but not decorin accumulation is related to malignancy in mammographically detected high density and malignant-appearing microcalcifications in non-palpable breast carcinomas. To gain further insight into whether it is likely for the overexpressed versican to be degraded by ADAMTS as a cancer invasive mechanism, we examined ADAMTS-1, -4 and -5 expression in healthy and cancerous colon tissues and also in colon cancer cell lines of different metastatic potential. Since the aim of this study was to reveal potential targets for the development of novel not only anti- invasive but also anti-angiogenic therapies, the ADAMTS anti-angiogenic subgroup expression was also estimated. Additionally, since ADAMTS-1 anti-angiogenic effect is mediated by its thrombospondin motifs and ADAMTS-20 contains 14 repeats of it, we also estimated its cellular levels in neoplastic and healthy colon. # Materials and Methods ## Materials Rabbit polyclonal antibodies, ab28284 (against N-terminal end of ADAMTS-1), ab84792 (against C-terminal end of ADAMTS-4), ab41037 (against 600–700 residues of ADAMTS-5) and ab60148 (against catalytic domain of ADAMTS-20) were purchased from Abcam (Cambridge, UK) and Goat anti-rabbit IgG conjucated with peroxidase was from EMD Millipore Corporation (Billerica, Massachusetts, USA). Mouse anti- tubulin (T9026) and peroxidase conjugated anti-mouse IgG (A4416) were obtained from Sigma-Aldrich Inc (St Luis, USA). Expose Mouse and Rabbit Specific HRP/DAB Detection IHC Kit (ab94710) was from Abcam. Total RNA was extracted from frozen tissue and cultured cells using the NucleoSpin Macherey-Nagel (Düren, Germany) extraction kit, following the manufacturer’s protocol. RT-PCR was obtained from Takara (Otsu, Shiga, Japan) One Step RT-PCR kit, used for fresh colon tissue and from Takara PrimeScript 1^st strand cDNA Synthesis kit and Finnzymes (Espoo, Finland) DyNAzyme II DNA Polymerase kit, used for cultured cells. Human Primers for all molecules were designed in the lab, using the PerlPrimer program. All other chemicals used throughout the study were of the best available analytical grade. ## Studies on cells ### Cell cultures Caco-2, DLD-1 and HT-29 colon cancer cell lines were purchased from the American Type Culture Collection (Manassas, VA). The aggressiveness of the cell lines is lower in Caco-2 cells and higher in HT-29 cells. DLD-1 and HT-29 cells were cultured in RPMI 1640 medium with 2 mM L-glutamine and supplemented with 10 mM HEPES, 1 mM sodium pyruvate, 4.5 g/L glucose, 1.5 g/L sodium bicarbonate and 10% fetal bovine serum (FBS), when required, as recommended by ATCC. Caco-2 cells were cultured in Eagle’s Minimum Essential medium with Earle’s BSS and 2 mM L-glutamine (EMEM) and supplemented with 1.0 mM sodium pyruvate, 0.1 mM nonessential amino acids, 1.5 g/L sodium bicarbonate and 20% fetal bovine serum when required, as also recommended by ATCC. Cells were cultured at 37°C, 5% CO₂ and 100% humidity. ### Cell lysis The pellets from cultured cells were treated with appropriate lysis buffer, containing 25 mM Hepes, 150 mM NaCl and 5 mM EDTA in 10% glycerol and 1% Triton X-100. Extractions were performed for 30 min at 0°C, under vortexing for 4 sec every 10 min, with the presence of phosphotyrosyl phosphatase inhibitor Na₃VO₄ (50 mM) and of the following protease inhibitors: Aprotinin (1250 μg/mL), Pefablock (1000mg/mL), Leupeptin (10mM), following centrifugation in 10.000 rpm for 5 min. ## Studies on human samples ### Patients The cancerous colon tissues were obtained from patients (38 patients, mean age; 73, age range; 50–81), who underwent surgical operation due to colorectal carcinoma. Nine patients of stage A (Duke’s), thirteen patients of stage B, twelve patients of stage C and four patients of stage D were included in this study. The patients included in the study were free of other disease and never before suffered of any disease. Healthy colon tissues (N = 6) were also included in our study. This study design had been approved by the Ethical Committee of the University Hospital of Patras, Greece, and written informed consent was obtained from all patients entering the study. ### Sequential extraction of extracellular and cell-associated components from tissues The collected normal and cancerous colon tissues were diced and sequentially extracted with 10 vols of 10 mM disodium phosphate, 0.14 M NaCl, pH 7.4 (PBS), 4 M Guanidine Hydrochloride (GdnHCl)– 0.05 M sodium acetate pH 5.8 and 4 M GdnHCl-0.05 M sodium acetate-1% Triton X–100, in order to obtain the soluble and the membrane-bound forms of ADAMTSs. Extractions were performed for 24 h at 4°C under gentle shaking, with the presence of the following proteinase inhibitors: ε-amino-n-caproic acid (0.1 M), PMSF (0.4mM), N-ethylmaleimide (10mM), disodium EDTA (10mM) and benzamidine-HCl (5mM). The extracts were collected and stored at -20°C until use. The extraction protocol was designed in such a way to separate ADAMTS existed in free forms in the tissues and should be extracted in PBS (associative conditions), from ADAMTS interacting with other extracellular macromolecules that should be extracted in 4M GdnHCl (dissociative conditions) and from ADAMTS existing in cell membranes that should be extracted especially in 4M GdnHCl-1% Triton X–100. ### Immunohistochemistry Immunohistochemistry (IHC) was performed on 5 μm sections cut from formalin fixed, paraffin-embedded blocks as described previously, using primary polyclonal antibodies against ADAMTS-1, -4, -5 and -20 diluted 1:6250, 1:2000, 1:650 and 1:6250 respectively, in TBS containing 1% (w/v) BSA. The obtained antigen-antibody complexes were visualized by incubation with Goat anti-rabbit HRP conjugate and the staining was developed with DAB/hydrogen peroxide of IHC Kit (Abcam) according to manufacturer’s instructions. Finally, the sections were counterstained with hematoxylin. Positive staining was scored according to the whole intensity of each one of the sections. ## Western blot analyses About 20 μg of protein, quantified using Bradford assay (BioRad), from the tissues and cell cultures extracts, and the cell cultures media were precipitated with 5 vols of ethanol, dissolved in 20 μl of Laemmli sample buffer and applied to Polyacrylamide gel electrophoresis (PAGE) on gels of 10% concentration in acrylamide as previously described. The macromolecules were then transferred to polyvinylidene difluoride (PVDF) membranes (Immobilon-P, Millipore) at a constant current of 80 mA at 4°C for 20 h in 0.05 M Tris/HCl, pH 8.3 and the membranes were washed with 0.14 M NaCl in 0.01 M phosphate buffer (PBS), pH 7.2, containing 0.1% (v/v) Tween-20 (PBS-T). After blocking with 5% bovine serum albumin (BSA) in PBS-T, the membranes were immersed in antibody solution against either ADAMTS-1, -4, -5 or -20 diluted 1:5,000, 1:500, 1:250 and 1:5,000 respectively, in PBS containing 1% bovine serum albumin (BSA) in PBS-T and incubated for 1 h at room temperature. After repeated washings with PBS-T, the membranes were incubated with proper secondary antibody diluted 1:10,000 in PBS-1% BSA, incubated for 1 h at room temperature and washed with PBS-T. The immunoreacting bands were visualized by enhanced chemiluminescence method (ECL) (Amersham, UK), according to the manufacturer's instructions and by exposure to Agfa Curix X-ray film, in time varied from 1 to 30 min, depending on the experiment. ## RNA extraction and RT-PCR analyses Fresh colon tissues were pulverized in liquid nitrogen and as cultured cells were subjected to total RNA extraction, using the Nucleospin extraction kit, as described by the manufacturer’s instructions and treated with RNase-free DNase to remove contaminating genomic DNA. For colon tissues, strand cDNA was synthesized from 80 ng of total RNA in 50μl reaction components for one-step RT- PCR kit, according to the manufacturer’s instructions. This reaction mixture contained in addition 1μM of the sense and antisense primers shown in. The amplification was performed in a GeneAmp 2400 thermal cycler (Perkin-Elmer Co.) and the reaction profile used for all primers sets was: 95°C for 10 min for the activation of DNA polymerase and inactivation of reverse transcriptase and then 25–35 cycles, depending on the analysis, at 94°C for 30 sec, 52–58°C depending on the primer set for 1 min and 72°C for 1 min to finalize extension. For cell cultures, two step RT-PCR was applied and the first strand cDNA was synthesized from 1 μg of total RNA in 20 μl reaction components for PrimeScript 1^st strand cDNA Synthesis kit, using random 6mers, according to manufacturer’s instructions. Then 100 ng of cDNA were amplified in 50 μl reaction components for DNA Polymerase kit. This reaction mixture contained in addition 0.5 μM of the sense and antisense primers shown in. The amplification was performed in a GeneAmp 2400 thermal cycler (Perkin-Elmer Co.) and the reaction profile used for all primers sets was: 95°C for 2 min, 94°C for 30 sec, 52–58°C depending on the primer set for 1 min and 72°C for 1 min to finalize extension and was repeated for 25–35 cycles, depending on the analysis. The reaction products were separated by electrophoresis in 2% (w/v) agarose gels containing Gelstar Stain to visualize the amplified cDNA fragments under UV. The gels were then scanned and the bands were analyzed densitometrically. Quantitative differences between cDNA samples were normalized by including GAPDH in all experiments. ## Statistical analysis Normality of distribution of values was tested with Kolmogorov-Smirnov test. Results were statistically analyzed using the unpaired t-test to detect differences between groups. *P*≤0.05 was regarded as statistically significant. # Results ## ADAMTSs expression in colon cancer cell lines ### RT-PCR analyses In order to investigate possible implication of ADAMTSs in CRC, we examined their expression at RNA and protein level in three colon cancer cell lines. The results indicated that in RNA level, ADAMTS-1 expression was slightly dependent on cancer cell aggressiveness; it was not found to be expressed in Caco-2, while it showed low expression in DLD-1, which was slightly increased in HT-29 cells. Moreover, in HT-29 cells cultured in the presence of serum, ADAMTS-1 expression was found elevated. On the contrary, high levels of ADAMTS-4 RNA were found in all three cell lines. However, ADAMTS-4 expression was found to be increased in an aggressiveness-related manner, especially when cells cultured in the absence of serum. Moreover, with the exception of DLD-1 cells, increased expression of ADAMTS-4 was observed when cells were cultured in the presence of serum. As shown in, no ADAMTS-5 expression was found in Caco-2 cells, while it was mainly observed in HT-29 cells. Unlike, ADAMTS-1 and -4, ADAMTS-5 expression was found to be down-regulated in cells cultured in the presence of serum. A different expression pattern was observed for ADAMTS-20. As shown in, in DLD-1 and HT-29 cells, it was found to be expressed only when cultured in the absence of serum. However, serum presence did not seem to influence ADAMTS-20 expression in Caco-2 cells. ### Western Blot analyses ADAMTS-1 active form was detected in DLD-1 and HT-29 cell lines, as a band of 80 kDa in the cell extracts, while only fragments of ADAMTS-1 of 35 kDa were detected in the media, regardless of the presence of serum. These fragments could be products of either autocatalytically processed ADAMTS-1 or catalytically processed ADAMTS-1 by MMPs. Given the fact that such fragments were detected only extra- and not intra-cellularly, they are more possible resulting from MMPs activity. In Caco-2 cell lines this metalloprotease was only detected in the presence of serum at same molecular form as in both other cell lines. An additional band of 65 kDa was observed in the extracts of DLD-1 cells cultured in the absence of serum. ADAMTS-4 active form was detected in DLD-1 and HT-29 cell lines and in the presence or not of serum, as a band of 90 kDa, intra- and extracellularly. In the case of Caco-2 cells, the 90 kDa band was only detected in cell extracts, whereas in cell media an extra immunoreactive band of 50 kDa was observed, probably as a result of its own autocatalytic activity. ADAMTS-5 active form was detected as a band of 74 kDa, in all three cell lines cultured under any conditions. Additional bands of various sizes, smaller than that of the active form, were observed in the cell extracts of HT-29 and mainly of Caco-2 cells cultured in the absence of serum. Interestingly, no immunoreactive band of any size was detected in the media of any of these cell cultures. However, ADAMTS-5 active form was also detected extracellularly, in HT-29 cells cultured in the presence of serum and in DLD-1 cell lines. Moreover, in HT-29 cells cultured in the presence of serum, additional bands of smaller sizes were also detected extracellularly, suggesting a posttranslational truncation of the enzyme resulting from either its own autocatalytic activity or MMPs activity. In contrast to all other ADAMTSs studied, ADAMTS-20 protein was not detected in these three colon cancer cell lines. ## In situ expression of ADAMTS-1, -4, -5 and -20 by IHC ADAMTS-1, -4, -5 and -20 in situ expression and localization was next analyzed in colon tissue sections from paraffin-embedded blocks, using the appropriate antibodies. Strong staining for ADAMTS-1 was observed in healthy colon, mainly in the muscular layers and in the connective tissue between the two layers (Healthy). In CRC, ADAMTS-1 showed strong expression in muscle tissue (arrow in, St. C) and lower mainly cytoplasmic expression in cancer cells of stage C specimens. Reduced or no staining for ADAMTS-1 was observed in muscle tissue of stage B specimens (St. B) and in specimens of other stages (C and D), respectively. Staining for ADAMTS-4 in healthy colon tissue was observed in the two layers of Muscularis Externa and in the Submucosae (, Healthy). In early cancer stages, stages A and B, ADAMTS-4 exhibited a similar expression pattern to ADAMTS-1 since it showed staining in stroma (arrows in, St. A, St B) and quite fainter or no staining in cancer cells (arrowheads in St. B and St. A, respectively). However, in specimens of stages C and D, cancer cells showed considerable cytoplasmic expression of ADAMTS-4 (arrowheads in St. C and St. D and their insets, respectively). Moreover, less or no expression of ADAMTS-4 was observed in stroma cells in specimens of stages C and D, respectively (arrows in St.C and St. D, respectively). As for ADAMTS-5, quite low staining was observed in the longitudinal layer of Muscularis Externa in healthy colon specimens (arrow in, Healthy). In CRC, no staining for ADAMTS-5 was observed either in stroma or in neoplastic epithelial cells, of stage A specimens (arrow and arrowhead in, St.A, respectively), while expression of ADAMTS-5 was detected in stroma cells of stage B specimens (arrow in, St. B). As ADAMTS-4, ADAMTS-5 also exhibited strong cytoplasmic staining in cancer cells of stage C and D specimens (arrowheads in St.C and St. D, respectively) and no stromal expression in stage D. Finally, ADAMTS-20 also exhibited low staining in the longitudinal layer of Muscularis Externa in healthy colon specimens (Healthy), while in CRC it was detected only in cancer cells of stage B and mainly of stage C specimens (arrowheads in, St.B and St. C, respectively). ## ADAMTSs expression in colon cancer tissues ### RT-PCR analyses RNA levels of ADAMTSs were examined in healthy cecum, sigmoid and rectum. Of all proteases investigated, ADAMTS-1 showed the highest expression (more than double to the internal control-GAPDH), while ADAMTS-5 was the less expressed protease (almost at zero level) (data not shown). RT-PCR analyses indicated that *ADAMTS-1* was down-regulated in cancer specimens of any stage compared to the healthy colon, with the greatest decrease to about 20% in stage A specimens. These data are in agreement with previous studies, where *ADAMTS-1* had been found to be down-regulated in many types of cancer. A different expression pattern from *ADAMTS-1* was observed for *ADAMTS-4* and-*5*, which were found to be overexpressed in stage C specimens; almost 14-fold and 20-fold compared to healthy colon tissue, respectively. These data were in agreement with previous studies in other types of cancer, where ADAMTS-4 and -5 had been found to be over-expressed in order to degrade ECM proteoglycans, such as agreecan and versican to facilitate cancer cell invasion. Following the same experimental procedure for *ADAMTS-20*, it was found elevated in stage A and B specimens compared to the healthy tissues by 50% and 100%, respectively, but in stage C specimens it exhibited a decrease to 20%. ### Western Blot analyses ADAMTS-4, -5 and -20 were detected in cecum, sigmoid and rectum, whereas ADAMTS-1 was detected only in cecum and sigmoid. However, they differed in their extractability revealing different types of interactions and/or localization of these enzymes within the colon tissue. ADAMTS-1 was detected only in PBS and GdnHCl/Triton extracts of healthy cecum and sigmoid, as a band of 120 kDa and 11 kDa respectively, representing the latent form of the enzyme and products of uncompleted synthesis, while none of the extracts contained the active form (80 kDa) of the enzyme.. ADAMTS-1 latent form was also detected in cancerous cecum of stage B, using dissociative conditions showing that it interacted with ECM components, such as TIMP-3 or a2- macroglobulin, which maintained the enzyme in its inactivated form. In cancerous rectum of stage C it was detected in GdnHCl extracts as a fragment of 35 kDa, possibly resulting from MMPs activity. A different figure was obtained from sigmoid, where ADAMTS-1 was not detected in any cancer stage. Thus, it could be claimed that ADAMTS-1 role in CRC depends at a large extent on the anatomic site. To continue, the latent form of ADAMTS-4 was detected in Triton extracts of all three anatomic sites of healthy colon as a band of 110 kDa. However, in PBS and GdnHCl extracts of both cecum and rectum, the active form of the enzyme (90 kDa) and multiple fragments of various molecular sizes and two distinct fragments of 60 kDa and 70 kDa were detected, respectively. On the contrary, ADAMTS-4 was detected in stage B and C specimens mainly in its active form. In stage B cecum specimens, the active form of the enzyme was detected in PBS and GdnHCl extracts, while in stage C rectum specimens, except for the active form which was found in PBS extracts, an additional band of 110 kDa, representing its latent form, was also detected in all tree extracts. Following the same experimental procedure, ADAMTS-5 active form was observed in PBS and GdnHCl extracts of healthy colon but also of all cancer stages, however, with a stage-related increased immunoreactivity. Finally, western blot analysis revealed the presence of latent form of ADAMTS-20 (161 kDa) in the PBS and GdnHCl extracts of the three anatomic sites of healthy tissues, sigmoid being containing the lowest amounts. Interestingly, it was also detected in all cancer stages, however with differences in its extractability and in the molecular forms. In stage A sigmoid specimens, the latent form of the enzyme was detected in PBS extracts, indicating that it existed rather freely in that tissue. ADAMTS-20 fragments were detected in GdnHCl and GdnHCl/Triton extracts. In GdnHCl extracts of stage B cecum specimens, a single fragment of 60 kDa was obtained, while in PBS and Gdn-HCl extracts of stage C rectum specimens, an additional smaller fragment of 50 kDa was obtained. # Discussion Accumulating evidence of ADAMTSs implication in human malignancies has demonstrated their significant role in tumor progression. Over-expression of these proteases is consistent with the requirements of carcinoma cells to remove proteoglycans of ECM, such as versican and aggrecan, as a complementary mechanism to collagen degradation by collagenases and gelatinases. Accordingly, increased expression of versican has been observed in neoplasias of colon and rectum. Additional significant information would give a study in ADAMTSs expression in CRC, since a potential degradation of versican would result in active versican fragments with established roles in tumor progression. On the other hand, some members of ADAMTSs family, those reported to have potential anti-angiogenic role, have been found epigenetically silenced in various types of cancer, including the CRC. In this study, a possible modulation of ADAMTS-1, -4, -5 and -20 expression and distribution in colon cancer compared to healthy colon was investigated. The experimental findings revealed a similar expression pattern for ADAMTS-4 and -5 and a completely different expression pattern for ADAMTS-1 and -20. ## ADAMTS-4 and -5 are over-expressed in CRC The results of this study support the notion that ADAMTS-4 and -5 are over- expressed in CRC possibly for tissue disruption that would facilitate cancer cell invasion and metastasis. However, slight differences between ADAMTS-4 and -5 expression in CRC were observed. RT-PCR and western blotting analyses in cultured cells of three different cancer cell lines revealed that expression levels of ADAMTS-5 were stronger related to cancer aggressiveness, as compared to ADAMTS-4. More specifically, extracellular ADAMTS-5 active form was not detected in Caco-2 cells, in contrast to ADAMTS-4 whose extracellular active form was detected in all three cell lines, regardless of cell aggressiveness. In addition to this, extracellular fragments of ADAMTS-4 and -5, possibly resulting from their own autocatalytic activity, were detected in both the Caco-2 and HT-29 cell lines. Hence, it could be suggested that among these two enzymes, ADAMTS-5 was over-expressed mainly in mediate/highly-aggressive cancer cells, while ADAMTS-4 performed a wider range of expression, regardless of cancer aggressiveness. Moreover, *ADAMTS-4* and-*5* transcription activity was found to be differentially regulated by the presence of serum. This finding most probably suggests that ADAMTS-4 up-regulation was possibly mediated via inflammatory cytokines and growth factors, and data from studies in osteoarthritis support this. On the contrary, ADAMTS-5 down-regulation was possibly mediated by growth factors, such as FGF-2. However, serum seemed to be crucial for secretion of the active form of ADAMTS-4 and -5 in both the Caco-2 and the HT-29 cells. ADAMTS-4 and -5 also exhibited the same localization pattern. In healthy colon, both enzymes, but mainly ADAMTS-4, were expressed in muscle tissue. In CRC, the expression levels of ADAMTS-4 and -5 were relatively decreased in early cancer stages (A, B) and the localization of the metalloproteinases was primarly at stroma cells. Interestingly enough, in late cancer stages (C, D) their expression levels were augmented and the examined enzymes were located mainly in malignant cells. Finally, ADAMTS-4 and -5 displayed similar expression pattern during cancer progression. They were both over-expressed at stage C, that is characterized by lymph node metastasis. Hence, it is possible that ADAMTS-4 and mainly ADAMTS-5 play a key role in tumor progression to higher stages of CRC by degrading ECM, so as to facilitate cancer cell invasion, in a similar manner as it has been previously demonstrated for hyaluronidase. Western blot analysis also confirmed the presence of active forms of both ADAMTS-4 and -5. In contrast to healthy colon, where ADAMTS-4 was fragmented, in CRC it was present mainly in its active form. This observation was in accordance with previous studies suggesting that fragments of ADAMTS-4 had an anti-metastatic role, in contrast to the active form of the enzyme which usually exhibits a pro-metastatic function. However, immunobloting also revealed differences between ADAMTS-4 and -5. Apart from its active form, ADAMTS-4 was also present in its latent form, while ADAMTS-5 was constantly present in its active form during cancer progression. Taking these data into account, it could be suggested that ADAMTS-4 only partially contributes to CRC progression. On the other hand, ADAMTS-5 serves as the cardinal component activation of which, fires CRC progression to higher stages. Notably, this finding is in harmony with previous studies having demonstrated a similar role of ADAMTS-5 in laryngeal cancer. Hence, future studies should focus on further investigating the mechanisms of ADAMTS-4 and -5 activation and to what extent these two enzymes are specialized in cleaving the modified substrates/proteoglycans which are present in CRC. ## ADAMTS-1 and ADAMTS-20 are down-regulated in CRC The different expression levels of *ADAMTS-1* and-*20* in CRC, despite their individual differences, further differentiate their implication in cancer, as compared to ADAMTS-4 and -5. ADAMTS-20 down-regulation, with the exception of Caco-2 cells where it is still expressed, was achieved in transcription level by serum components but mainly in post-transcription level, since no protein is being produced. On the contrary, despite the low transcriptional activity of *ADAMTS-1*, the catalytically active enzyme was being produced but it was not secreted by the cells. ADAMTS-1 fragments found in the media of cell cultures were possibly products of catalytically processed ADAMTS-1 by MMPs. As a result, the removal of ADAMTS-1 from cell membrane limits its anti-angiogenic properties. In healthy colon ADAMTS-1 and -20 performed the same localization, since they were both expressed in muscle tissue, though ADAMTS-1 showed much higher expression than ADAMTS-20. Interestingly, in CRC they were expressed by different cell types. ADAMTS-1 was mostly expressed by stroma cells, in contrast to ADAMTS-20 which was expressed by cancer cells. ADAMTS-1 expression by stroma cells could be an attempt to arrest the progress of cancer by inhibition of angiogenesis; an attempt finally failed as western blot analysis revealed by the presence of enzyme fragments in stage C. ADAMTS-20 expression by cancer cells could be a result of cross-talking between cancer cells and stroma cells. Given the fact that no ADAMTS-20 is being produced by cultured colon cancer cells, in contrast to colon cancer tissue, where the enzyme is being produced, as it was confirmed by both IHC and Western blot, it could be said that tumor micro- enviroment plays an important role in ADAMTS-20 expression. However, as Western blot analysis revealed the latent form of the enzyme, which was present in healthy colon, is being produced only in early cancer stage while in late cancer stages the enzyme is being fragmented. These fragments are freely circulated but also in association with ECM components; the exact role of these fragments though remains unknown. Taking together these data, it could be said that although ADAMTS-1 and -20 are both generally down-regulated in CRC, their low expression came from stroma or stroma-induced cancer cells, respectively, where they were present as fragments. # Conclusions During cancer progression, a reorganization of the extracellular matrix takes place to influence cellular proliferation and invasion. ADAMTSs are key molecules in this event, since they are extracellular proteases, possessing also anti-angiogenic activity. From the results of our study in CRC tissue and cells, we conclude that ADAMTS-4 and -5 expression positively correlates with cancer progression, whereas the anti-angiogenic ADAMTS-1 and -20 were found to be down- regulated and degraded. Thus, our results provide a mechanism of CRC progression and invasion mediated by specific ADAMTS members. We are grateful to Prof. Dr. Nikos Karamanos for critically reading the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: DHV ADT MS DJP. Performed the experiments: SF AK DK DB PR. Analyzed the data: DHV DJP PR SF. Contributed reagents/materials/analysis tools: DHV ADT DJP. Wrote the paper: SF DHV DJP ADT MS DK.

# Introduction Current understanding of mammalian herbivore foraging ecology is mainly based on studies focusing on ungulates; see for example, and. Other herbivores with a central role in many ecosystems, such as small rodents, have been less studied. Small rodents are non-ruminant herbivores with fast digestion, invest greatly in reproduction and little in growth, generally have a high risk of predation and are often territorial. They can therefore be expected to have different nutritional needs and face different trade-offs both physiologically and behaviorally than ungulates. Due to such differences in trade-offs, small rodent functional responses, i.e. the relationship between food intake and food availability, are also likely to differ from those developed using ungulates as empirical models. Functional response models can improve the understanding of how herbivores select their food and thus aid in predicting how they may cope with current vegetation changes, as well as how they may themselves affect vegetation. A range of parameters have been suggested to be incorporated into functional response models for herbivores. However, to target those parameters that are important determinants of small rodent functional responses, more exploratory empirical work is required to assess which processes shape their food selection in the wild. Within a food item category such as plant species or genus, small rodent functional responses to food availability have been studied experimentally. In these studies, food availability has, unavoidably, been found to increase food intake. However, various processes such as handling time, bite size or plant spacing have been shown to have the potential to regulate this relationship. Even though feeding trials using single food items may identify mechanisms that operate in the wild, the value of a food item to an animal is relative to what else is available. Studies investigating how the availability of alternative food items impacts on consumption of other food items by small rodents are, however, scarce. These studies experimentally demonstrate that availability of a high-quality food item can reduce the consumption of a low-quality food item. In natural environments, a range of food items of different quality are available and the composition of vegetation may vary greatly. However, small rodent functional responses to the spatially and temporally variable food availability in the setting of complex natural plant communities remain unexplored. Grey-sided voles (*Myodes rufocanus*) and tundra voles (*Microtus oeconomus*) are among the key herbivores of subarctic tundra ecosystems where they greatly modify tundra vegetation during their cyclic population density peaks. During recent decades, their cyclic population dynamics have dampened in many areas. While changes in winter climate have mainly been suggested to cause these changes in population dynamics, the role of concurrent vegetation changes is unclear,. However, any evaluation of such bottom-up effects in tundra food webs is severely hampered by the current gaps in knowledge of vole diets and how diet is affected by food availability. Grey-sided voles are considered to prefer *Vaccinium myrtillus* but to also feed on forbs during summer, while tundra voles are considered to feed primarily on monocotyledons, with an increased proportion of *Equisetum* and forbs during the summer. These generalizations are mostly based on microhistological analysis of ingested material – and observations of feeding signs on vegetation. However, taxonomic resolution of microhistological studies of small rodent diets is limited , whereas feeding signs of vegetation give limited information on proportional abundance of different food items in diets. We analyzed stomach contents of grey-sided voles and tundra voles using DNA metabarcoding. This novel methodology has lately opened new avenues of herbivore diet studies, as it enables analysis of large numbers of samples and identification of the ingested plants at a detailed taxonomic level,. We used a spatially extensive study design, spanning across two river catchment areas and two habitats and sampled vole diets and vegetation composition in common locations over two seasons. Thus, we were able to study the impact of food availability on diets and selectivity both at a taxonomically more detailed level than previous studies and across the range of food availability variation present in natural habitats of voles. We first compared vole diets to vegetation in order to determine which food plants were selected for. We then investigated how vole diets and selectivity were related to availability of food plants. We analyzed these relationships using both plant families and plant functional groups, and use hereafter the term “food item” to describe any plant group. We predicted that the proportion of any preferred food item in diets would increase with its availability but that it also would be affected by availability of alternative food items. We further predicted that selectivity for a food item would also be affected by availabilities of both the food item in question and alternative food items. Plant families allowed the most precise taxonomic units for diet and vegetation comparison. Plant functional groups coarsely reflect plant nutrient content and digestibility and allowed grouping of plants according to their presumed nutritional value for herbivores. By analyzing vole feeding habits using both taxonomic and ecological groupings we aimed to both perform a taxonomically detailed analysis of vole diet and to evaluate how the different food item units reflect vole feeding ecology. # Materials and Methods ## Study Area This study took place at Varanger peninsula, (70°N, 31°E), Finnmark, North- Eastern Norway. Two prominent habitat types of the area, dwarf-shrub heaths and meadows with scattered willow (*Salix* spp.) thickets harbor different vegetation and small rodent communities. Vegetation in the heath is mostly dominated by *Empetrum nigrum* s. lat. but also *Betula nana* and *Vaccinium myrtillus* are frequent. Field layer of the meadow vegetation is more diverse and dominated by grasses (e.g. *Avenella flexuosa*, *Deschampsia cespitosa*), forbs (e.g. *Rumex acetosa*, *Trollius europaeus*, *Viola* spp.), vascular cryptogams (mainly *Equisetum* spp.), deciduous shrubs (mainly *Salix* spp.), sedges and rushes (e.g. *Carex bigelowii*, *Carex aquatilis* coll., *Juncus filiformis*) and mosses. Average (and range) total plant biomass during this study was 525 g/m (280–1056 g/m) in the heath habitat and 206 g/m (82–439 g/m) in the meadow habitat (see details on biomass measurements below). Biomass ranges for plant functional groups are shown in. In heath habitats, grey-sided voles (*Myodes rufocanus*) are the most common small rodent species, whereas in the meadow habitats tundra voles (*Microtus oeconomus*) dominate the small rodent community. In addition to voles, Norwegian lemmings (*Lemmus lemmus*) are found in the area during their outbreak years. Small rodent populations in the region have cyclic population dynamics with high-amplitude peaks every 4–5 years, and this study included a summer season peak in 2007. In addition to small rodents, semi-domesticated reindeer are abundant in the study area, whereas other mammalian herbivores are scarce. More detailed descriptions of the study area can be found in, and. ## Study Design In order to cover the range of variation in vegetation composition present at Varanger peninsula, we used a large-scale study design encompassing two river catchment areas; Komagelva (KO) and Vestre Jakobselva (VJ). In both river catchments, we established sampling grids (15×15 m) in equal numbers in both meadows and heaths. The sampling grids were selected to represent the range of variable field layer species compositions of both habitats. In total, KO had 12 sampling grids per habitat and VJ had 13. The distance between neighboring grids had a range of 160–2200 m, while the two most distant grids were 40 km apart. In order to measure food availability in the immediate habitat of each vole individual, we used the same study design for both vole trapping and plant biomass analysis. Population dynamics of voles differ between the focal river catchments, and both vole species had lower densities in VJ compared to KO. ## Vole Trapping; Samples for Diet Analysis In order to obtain samples for diet analysis, we conducted snap-trapping of voles in each sampling grid according to. To estimate changes in diet during the growing season, the trapping was done twice, with the first period occurring between 22^nd and 24^th July and the second period occurring between 3^rd and 5^th September. Each trapping event consisted of 600 trap nights per habitat, with 12 traps in each grid, 25 grids and trapping over 2 nights. The traps were baited with raisins (*Vitis vinifera*) and oat flakes (*Avena sativa*). Voles were dissected and their stomachs stored in 70% ethanol until diet analysis. Snap-traps were required, as the rodent trapping was part of a project where also the Norwegian lemmings were studied. Norwegian lemmings are hard to trap with live-traps, a phenomenon which has been repeatedly observed by different research groups. In another study using live-traps, only one lemming was caught despite a large trapping effort (ca 6,000 trap-nights every year) and the occurrence of two small rodents peaks (2007 and 2011, resulting \>10, 000 trapped voles) (Ims and Yoccoz unpublished data). ## Vegetation Composition; Food Availability Data Vegetation of each grid was sampled during the peak of the growing season, i.e. between 22^nd July and 8^th August. We established 13 vegetation sampling plots (0.5×0.5 m) in each grid and estimated the biomass of all vascular plant species present in the plots using a non-destructive point intercept method, with 20 pins in each plot. We then converted the point intercept counts to biomass estimates (g/m) for each grid, by first converting the hits to biomass per plot using calibration described in. In another study across northern Norwegian landscapes, encompassing similar habitats as the current study, plant growth form was found to be the most important predictor for both vegetative and flowering phenology of plants. Hence, the phenology of biomass in our study area could be expected to fairly similar in both river catchements. To account for the temporal changes in biomass we included the effect of season to our analyses. ## Ethics Statements The study area is part of Varangerhalvøya National Park. No permit was required for the non-destructive vegetation sampling, as only destructive use of vegetation is prohibited in the national park (FOR-2006-12-08-1384, Regulation of Varangerhalvøya nationalpark protection plan). Vole trapping was conducted as part of the “Arctic fox in Finnmark” project (<http://www.fjellrev- finnmark.uit.no/>), which was initiated, financed and approved by The Norwegian Directorate of Nature Management (DN). The DN is the legal Norwegian authority that licenses sampling of all vertebrate wild life species for scientific purposes (LOV 1981-05-29 nr 38: Lov om jakt og fangst av vilt (viltloven) <http://www.lovdata.no/cgi-wift/ldles?doc=/all/nl-19810529-038.html&26>) and regulation about sampling wildlife for scientific or other specific purposes (FOR-2003-03-14-349 Forskrift om innfanging og innsamling av vilt for vitenskapelige eller andre særlige formal <http://www.lovdata.no/for/sf/md/md-20030314-0349.html>). No specific permit was issued for this project, but sampling protocol was approved by the DN. No protected species were sampled. ## Diet Analysis Stomach contents of grey-sided voles trapped from heath habitat (n = 82) and tundra voles trapped from meadow habitat (n = 67) were analyzed for spermatophyte (i.e. seed plant) content. Part of the dataset is published by, who described in detail the DNA metabarcoding methods used (see for additional details on the datasets). In summary, spermatophyte plant DNA was amplified from a sample of each voles stomach content using primer pair *g-h*, which targets the P6-loop of chloroplast *trn*L (UAA) intron. Samples from different individuals were thereafter individually tagged, pooled to one sample and pyrosequenced. The resulting sequences were sorted to individual voles based on the tags and compared to two taxonomic reference libraries to identify which taxon they belonged to. We first used a library containing sequences of 842 arctic vascular plants (accession numbers GQ244527 - GQ245667 in GenBank). Thereafter, we compared sequences which could not be satisfactorily identified to sequences retrieved from GenBank (available at <http://www.ncbi.nlm.nih.gov/genbank/>). For each vole individual, we thus achieved a count of sequences belonging to different taxa. To make data from different vole individuals comparable, we transformed these counts to proportions of different taxa in an individuals’ stomach content, hereafter termed as “diet proportions”. Quantitative use of DNA metabarcoding data is potentially hampered by several technical issues. However, based on a comparison with traditionally used microhistological method, DNA metabarcoding reflects well the actual proportions of spermatophytes in vole diets. We also verified that diet at the vole population level, measured as food item proportions, did not differ greatly from diets determined by frequency of occurrence (i.e. percentage of vole individuals which had ingested the taxa in question), as recommended by. Moreover, a taxon may be over-represented in a DNA metabarcoding dataset if it has a short target DNA-region in comparison to other simultaneously analyzed taxa. We therefore also confirmed that the most abundant taxa did not have clearly shorter target- DNA regions than other taxa. Both frequencies of occurrence and lengths of the targeted DNA region are given in and. We removed two vole individuals from the dataset prior to the analyses. One of these was a grey-sided vole that had seemingly eaten only one plant species, an unlikely result which could easily be due to low DNA quality of the sample. The other was a tundra vole whose diet was composed 99% of *Pinus sylvestris*, a species not present in the study area. Rather than representing a new species in the region’s flora, such a result is probably caused by errors during the analyses. The reference libraries we used included a different range of species than those present in the study area. We therefore checked for potential mis- identifications and adjusted the sequence assignments based on taxa present in Northern Fennoscandia. Taxa which are not found in the region were assigned to their next higher taxonomic level (e.g. *Cerastium maximum* was assigned to *Cerastium* sp. and *Gaylussacia* sp. to Ericaceae). Adjustments were also made to more specific taxa, i.e. when a genus (or family) was represented by only one species (or genus), it was assigned to this representative (e.g. *Bistorta* sp. was assigned to *Bistorta vivipara* and Betulaceae were assigned to *Betula* spp.). Sequences originally assigned to *Vaccinium alaskense*, which is not found in the region were grouped together with those assigned to *Vaccinium myrtillus*. These species are almost identical at the DNA region we used for identification but differ from other *Vaccinium* species of Northern Fennoscandia, namely *Vaccinium uliginosum* and *Vaccinium vitis-ideae* (accession numbers GQ245635-GQ245641 in GenBank). ## Definitions of Food Item Groups For analysis at plant functional group level we classified plants as forbs, grasses, sedges and rushes, deciduous shrubs, ericoid shrubs, or hemiparasites. The grouping was primarily based on nutritional characteristics, as well as responses to herbivory in the focal ecosystem. However, we grouped all ericoid shrubs together as less than half of the sequences within Ericaceae were identified at a detailed enough level to allow distinguishing between deciduous and evergreen shrubs. The deciduous shrubs -group was thus composed of Betulaceae and Salicaceae. Only a few non-ericoid evergreens (Pyrolaceae, Pinaceae, Cupressaceae) were recorded in the diets and each of them occurred only in one vole individual. These taxa compose a very small fraction of the biomass (on average 0.003% and 0.4% in heaths and meadows respectively) and we therefore excluded them from all analyses. We also excluded data on vascular spore plants (i.e. *Equisetum* and ferns) from the analyses, as the primer pair *g-h* is designed particularly for spermatophytes and does not reflect well the abundance of other plant groups. For analyses based on taxonomic units, we grouped the plants at family level in order to be able to include majority of the data. For example, 36% of sequences identified to Ericaceae in grey-sided voles diets could not be identified to genera. However, for two families we had sufficient data to refine the analyses to species level. One of these, Cornaceae is represented in Northern Fennoscandia by only one species (*Chamaepericlymenum suecica*). The other family for which we achieved species level resolution was Ranunculaceae for tundra voles. *Ranunculus acris* coll. was the only representative in the meadow grids and *Ranunculus* sp. constituted 99% of Ranunculaceae in tundra vole diet. We therefore used data on *Ranunculus acris* coll. in all analyses of Ranunculaceae in tundra vole diets and selectivity. ## Statistical Analysis ### Food selection: compositional analysis To determine selectivity we used compositional analysis of centered log-ratio transformed proportions of food items in individual diets and available vegetation, at both plant family and functional group level. The centered log- ratio transformation was implemented by function named “clr” in in the R library compositions. As food availability data we used for each vole individual, the biomass proportions from the grid in which it was trapped. The selectivity index was calculated as clr(diet proportions) -clr(available proportions). To test whether selectivity for different food items was significantly different, we used compana -function in adehabitat-library of R, which computes pairwise significances in preference among food items using Wilks lambda. Results of these significance tests are presented in, while diets, food availability and selectivity index values are shown in and. Both the data for diet and available food contained zeros, which have to be replaced to enable compositional analysis. We followed recommendations given by, replacing zeros with a value three orders of magnitude smaller than any observed used proportion (0.000077) to the diet proportions. This very small replacement value ensured that minute amounts of detectable DNA were not included in the analysis. We excluded plant families which never occurred in diets from the compositional analysis at family level, as their combined biomass was on average \<1% in heaths and 2% in meadows. We replaced zero availability in a given grid with average biomass of the food item in question in the same habitat and river catchment. When a food item was not recorded within a river catchment, we used average biomass of the habitat across river catchments. Campanulaceae and Apiaceae were never recorded in the heath habitat, even if they were recorded in the diets of two grey-sided voles. We replaced zero availability in these families by an order of magnitude smaller value than the smallest observed proportion of any family in the heath. We included sequences of trap bite (*Avena* and *Vitis*) in calculations of food item proportions in stomach contents, but excluded them from further analyses. ### Variability in diets and selectivity: linear mixed effect models We evaluated whether I) the proportion of a food item in diet and II), selectivity for it were related to vegetation composition using linear mixed effect models implemented by lmer-function from lme-4 library of R. In order to target important food items and retain sufficient sample size for the models, we modeled separately the response of each food item which was both selected for and eaten by at least ca. 50% of the individuals of a given vole species. For grey-sided voles these were the functional group forbs and families Ericaceae, Polygonaceae, Poaceae, Salicaceae and Cornaceae, while for tundra voles they were the functional group forbs and families Polygonaceae, Poaceae, Ranunculaceae and Salicaceae. No other functional group than forbs had several families which were eaten so commonly that they could be tested separately. For all of these food items, we modeled separately proportions in diets and the selectivity index score. For diets, we used logit transformed proportions as the response variable, avoiding zeros by adding a value which was an order of magnitude smaller than the smallest value of the respective predictor variable, while selectivity index scores were already at logit-scale. For each response variable, we created two alternative models. The first model included as predictors a) biomass of the food item itself and b) biomasses of other substantially eaten food items at family level (i.e. those listed above). However, to better fit data with the voles feeding ecology, we only used biomass of palatable deciduous *Vaccinium* species as predictor instead of Ericaceae. For grey-sided voles, Salicaceae was omitted from models which included Polygonaceae, as their biomasses in vegetation were highly correlated. In the alternative model we replaced the predictor(s) forb families by the forb functional group, leaving the response variable at family level. In all of these models, we evaluated the spatial variability of diets and selectivity in two ways, using river catchment (KO and VJ) as a fixed effect and grid identity as random effect. In addition, we included season (autumn and summer) as a fixed effect. When random effect variance was estimated as zero, we removed the term and present a model with fixed effects only (using lm-function of R). We then selected the better model using likelihood ratio test (model parameters estimated using ML). When neither model was significantly better, we show the model with less parameters. We present the final mixed effect models with model parameters estimated using REML. In addition to statistically significant effects (defined as 95% confidence intervals not encompassing zero), we included in our interpretation close-to-significant trends which seemed biologically interesting. These are statistically defined as having 95% confidence intervals crossing zero by \<0.05 and an effect size of \>0.15. After fitting fixed terms of the models, we calculated the proportion of remaining variance explained by random variable “grid identity” (i.e. grid variance/(grid variance+residual variance)). In each model, we removed those individuals which had a combination of zero availability and zero use of the response food item. We checked models for outliers and removed one heath grid where Polygonaceae biomass was approximately 8 times that of any other grid (thus removing two grey-sided vole individuals). We also verified that models showed constant variance of the residuals and approximate linearity between the fitted and observed values. For models with random effects we estimated significance of the fixed parameters with 95% confidence intervals (Markov Chain Monte Carlo estimation with 100 000 replicates using mcmcsamp -function), while for models without random effects we used the confint-function of R. We used the software R for all analyses. # Results ## Diets At the level of plant functional groups, diet of the grey-sided voles was dominated by ericoid shrubs, followed by forbs. Deciduous shrubs and grasses were also eaten but less commonly. Within ericoid shrubs, i.e. within the family Ericaceae, deciduous species *Vaccinium uliginosum* (9%) and *Vaccinium myrtillus* (8%) were the most commonly identified but also everegreen shrubs, mainly *Empetrum nigrum* s. lat. (6%) were found. Within the functional group of forbs, most abundantly eaten families were Cornaceae (10%, represented by *Chamaepericlymenum suecica*) and Polygonaceae (9%, represented mainly by *Rumex* sp.). Grey-sided voles had also consumed a range of other forb families at a lower proportion, many of which occurred in only a few individuals. Species richness of grey-sided voles diet (n = 82 vole individuals) was 28 at species level, 37 at genera level and 23 at family level. At the level of plant functional groups, the diet of tundra voles was markedly dominated by forbs, followed by deciduous shrubs and grasses. The functional group of forbs was dominated by family Polygonaceae (45%, represented mainly by *Rumex* sp.). While the family Ranunculaceae was also commonly eaten (12%), the mean proportion of other forb families was low and many of them occurred in only a few individuals. Species richness of tundra vole diet (n = 67 vole individuals) was 26 at species level, 35 at genera level and 23 at family level. ## Selectivity Both vole species had the strongest selection for forbs. After forbs, grey-sided voles selected for ericoid shrubs and thereafter grasses, whereas tundra voles selected for grasses and thereafter deciduous shrubs. The patterns of selectivity at functional group level differed somewhat from those at plant family level. For example, only one forb family, namely Polygonaceae, was more often selected than Poaceae (i.e. grasses), a pattern found for both vole species. Also within the functional group of deciduous shrubs both vole species showed a similar pattern; Salicaceae was relatively preferred whereas Betulaceae was the least preferred. ## Spatio-temporal Variation of Diets and Selectivity Both vole species showed temporal and spatial variation in their feeding habits. During summer, grey-sided voles had higher proportions of Poaceae and forbs, especially Polygonaceae in their diets than during autumn. During autumn, they selected more for Ericaceae than during summer. Tundra vole diets and selectivity were less modified by season than that of grey-sided voles but tundra voles also selected for Polygonaceae more during summer than autumn. Spatial variability in diets and selectivity was measured at two scales; river catchment and sampling grid. Of these, river catchment had little effect on the vole diets and selectivity. Only grey-sided voles use of Poaceae varied at the scale of river catchment, with diet proportions and selectivity being higher at VJ than at KO. However, both vole species showed spatial variability in diet proportions and selectivity at the scale of sampling grids, based on percentage of residual variance explained by grid identity. ## Impact of Availability on Diets and Selectivity We found few clear effects of biomass of a food item (i.e. availability) on its use. The sole statistically significant effect was that tundra voles were more selective for Ranunculaceae when its biomass was higher. In addition we found one non-significant trend whereby grey-sided voles’ selectivity for Salicaceae decreased with its biomass. However, use of several food items decreased with the availability of alternative food items. For grey-sided voles, selectivity for Polygonaceae decreased with biomass of other forbs and Polygonaceae proportion in diets had a similar trend. Moreover, increasing biomass of Salicaceae tended to decrease the proportion of Cornaceae in the diets of grey- sided voles. For tundra voles, selectivity for Poaceae decreased when biomass of Salicaceae increased and had a similar trend with biomass of Polygonaceae. We also found opposite patterns in grey-sided voles, i.e. use of a food item increasing with the availability of alternative food items. Of these, both diet proportions and selectivity for Polygonaceae increased when biomass of Poaceae increased, whereas diet proportions of Cornaceae increased with biomass of Salicaceae. Some of these indirect effects, both negative and positive ones, were caused by changes in the biomass of food items which were on average less selected for than the response food item. The use of different forb families responded differently to their respective biomass, biomass of alternative food items and season. However, combined biomass of forbs better predicted the consumption of other food items than those of separate forb families. Only the selectivity of tundra voles for Poaceae was slightly better predicted by a model which included biomass of Polygonaceae and Ranunculaceae as predictor variables than with a model using combined forb biomass ( = 3.62, d.f. = 1, p = 0.06). # Discussion Both grey-sided voles and tundra voles consumed a diverse range of food items. Although diets and selectivity varied seasonally and spatially, the biomass of a food item had little effect on its use but sometimes influenced the use of other food items. Together, these results show that both vole species exhibit flexible feeding ecology. ## New Insights into Vole Diets Most studies on the interactions between grey-sided voles and vegetation have focused on *Vaccinium myrtillus*, which has been considered as the most important food item of this species. However, our results show that during the snow-free period the species has a diverse diet which includes, in addition to *V. myrtillus*, a range of different herbaceous food items. We also found surprisingly much *V. uliginosum* (on average 9% of diets, eaten by 50% of individuals), even if it is relatively rare in the heath vegetation (2% of biomass in average), indicating that it is selected much more than previously observed (similar numbers for *V. myrtillus* are 8% in diets, eaten by 68% of individuals, 20% of biomass in average). Interestingly, *V. uliginosum* and *E. nigrum* have been suggested to be unpreferred species and eaten only when population densities are high. Further, they have been suggested to produce toxins and therefore to have a negative impact on vole population growth rate even if they constitute only a small proportion of diets. We found additional support for this hypothesis for *V. uliginosum*, which inclusion in the diets of grey-sided voles increased with population density. *V. uliginosum* was found in 25% of the individuals in VJ during summer while corresponding values were 42%, 45%, 62% for VJ autumn, KO summer, KO autumn, respectively (in comparison to population density index). For *E. nigrum*, on the other hand, we found no such pattern (proportion of individuals that had ingested it, in same order as above, were 70%, 34%, 50% and 75%). Still, as we observed both of these plant species to be eaten by more than half of the studied grey-sided voles during a peak year, it is possible that these food items play a role in the population dynamics of this vole species at the Varanger Peninsula. For tundra voles, we found forbs to dominate diets and be highly selected for, unlike previous microhistological stomach content studies which have emphasized the use of especially *Eriophorum* and *Equisetum* –. Such discrepancies between studies are probably partly due to differences in availability resulting in different diets. For example, forbs were abundant and *Eriophorum* absent in our study grids, whereas forbs were rare and *Eriophorum* abundant in habitats where tundra vole diets have been studied before. Moreover, in similar habitats the closely related field voles (*Microtus agrestis*) have also been found to have diets dominated by dicotyledons, and in a cafeteria-test tundra voles showed a preference for forbs. As *Eriophorum* was not included in that test, it remains unclear whether plant availability modifies only vole diets or also preferences. In addition, different methodology may contribute to differences in results. While results based on microhistological methods have a tendency to overestimate monocots, the DNA metabarcoding method used in this study possibly underestimates *Equisetum*. In spite of such methodological discrepancies, habitat-specific food availability is likely to be an important determinant of tundra vole diets. Both vole species selected for highly palatable functional groups, i.e. forbs and grasses, indicating that vole food preferences are related to plant nutritional quality. However, within plant functional groups different families and species were eaten and selected to a very different degree. For example, some forb species were rarely eaten even if their availability did not greatly differ from that of other, more commonly consumed species. Moreover, the use of forb families responded differently to biomass and season. Different nutritional value may explain such differences but because only few measurements of energy, nutrients or secondary metabolites in subarctic forb species exist, we cannot judge the importance of different nutritional characteristics for voles. Within the functional group deciduous shrubs, both vole species preferred willows (*Salix* spp.) but avoided birches (*Betula* spp.), a pattern consistent with palatability of these taxa as well as previous food selection studies of voles. Thus, more detailed patterns of food quality than those reflected by plant functional groups, as defined in this study, seem to direct food preferences of voles. However, field measurements of detailed food-selection units have limitations especially when the food items are scarce. For example, grey-sided voles preferred forbs as a functional group even though at family level most forbs were seemingly not preferred, a pattern which could simply be due to different forbs being available to different individuals. Plant functional groups have mainly been studied from a plant ecological perspective, and only few attempts have been made to evaluate them based on herbivores ecology. However, small rodent food-selection units may be best reflected by plant functional groups defined from a herbivores perspective. Previous analyses of diets of small rodents have used methods that are constrained to a taxonomically coarse resolution. Using DNA metabarcoding we were able to reveal that both vole species had remarkably diverse diets in terms of consuming a large number of plant taxa. In fact, diet diversity as such may be an important attribute of vole diets, as it is in general acknowledged to be an important determinant of herbivore performance. Accordingly, found that species richness of vascular plants in the sub-arctic habitats of grey-sided voles was the most important predictor of female reproductive success. It therefore seems likely, that increased understanding of the role of food item diversity for small rodents should reveal previously unknown aspects of vegetation-small rodent interactions. ## Small Rodent Functional Responses Several earlier studies on mammalian herbivore food selection have indicated that availability, both absolute and relative of a preferred food item may increase its use. Seasonal effects were common even though we found that the biomass of a food item had little effect on its consumption. Nutrient content of herbaceous plants decreases towards the end of the growing season. Moreover, while berries produced by ericoid shrubs are more palatable than leaves of these plants they are available only in the autumn. The seasonal changes in grey-sided voles feeding habits, i.e. the decrease of forbs and grasses in diets and increased selectivity for Ericaceae from summer to autumn, thus seem to be related to availability of good-quality food. However, at the resolution of our data, season was best seen as qualitative “index” of changing availability of good-quality food. In addition to availability of a food item, availability of alternative good-quality food items may modify consumption by voles. Our results indicate that a food item which has such indirect effects does not have to be more preferred at the population level. For example, tundra voles selected less for grasses (Poaceae) when the biomass of willows (Salicaceae) increased, even though at population level they preferred grasses to willows. We therefore suggest that the effect of alternative food item availability for small rodent functional responses should be further evaluated. Moreover, based on the seasonal changes of voles’ diets and selectivity, relative differences of nutritional quality between different food items are probably important determinants of such effects. The spatial effects we found, i.e. voles from the same study grids having more similar diets and preferences than those from different grids, suggest that voles from same local environment are more likely to make similar food choices than voles from different environments. In addition to food characteristics, small rodent feeding habits can be affected by competition and predation risk which could therefore contribute to spatial variation in feeding habits. Vole population densities, especially those of tundra voles, differed drastically between the river catchments. However, diets and selectivity differed little between river catchments and therefore intraspecific competition seems unlikely to have caused the spatial patterns we observed at grid level. On the other hand, population density of Norwegian lemmings was higher in VJ where vole densities were lower. Interspecific competition with lemmings may therefore have masked some of the effects of intraspecific competition on vole diets. That food biomass had little effect on diets and selectivity does not support the idea that the spatial effects would be caused by food availability either. Still, food item biomass may not necessarily reflect all vegetation characteristics which are important for determining vole feeding habits. For example, a plant species’ nutritional quality may vary, both temporally, spatially and also between plant parts. Moreover, positive responses of vole selectivity on availability, evaluated via responses to biomass and season, suggest that voles do not compensate low availability with increased selectivity. This in turn indicates that voles invested little effort in searching and selecting the most preferred food. It is well established that perceived predation risk reduces time herbivores, including small rodents, spend foraging in dangerous habitats. Nevertheless, the interplay between food availability and perceived predation risk, ‘the landscapes of food and fear’, remains poorly understood. In tundra habitats vegetation cover is generally low and predation risk high, especially during small rodent population-peak years. Flexible feeding habits of voles could thus at least partly be an adaptation to minimize time spent searching for food, as emphasized by and. The spatial variation of diets and selectivity which we observed are therefore probably caused by a combination of local vegetation characteristics and search time limitations due to predation risk. Both plant quality and search time limitations have been included in some functional response models for herbivores, and we suggest that examining the roles of these parameters for small rodent functional response models should be attempted. While we here show that the population level patterns in feeding habits of voles are flexible, it is possible that vole individuals are more conservative. At least some of the changes in vole diets are related to changes in gut morphology, indicating that individual voles may have physiological limitations related to switching quickly between highly different diets. However, little is known about the flexibility of vole diets at individual level, and it is unclear how fast and drastically individual voles may change their diets. ## Methodological Considerations While few of the observed effects of food item biomass on stomach content were statistically strong, we are confident that these patterns indeed reflect the relationships between voles and their food. The methods we used to estimate food item use and availability, i.e. stomach contents and biomass of plants, have certain shortcomings. Most importantly, food passes quickly through the digestive system of voles and stomach contents therefore give a snapshot of the vole diet during the last hour. In addition, food item availability to voles may be poorly represented by average g/m biomass of plant species. For example, some food items might reach a height which makes them unavailable for small-statured herbivores like voles and hence average biomass may differ from what is available for voles. Finally, we measured plant biomass during the peak of growing season but sampled vole diets during early and late growing season. Due to seasonal increase of biomass, this may have led to underestimation of selectivity in the summer in comparison to the autumn. However, the only seasonal increase of selectivity was that of grey-sided voles for Ericaceae, which can be well explained by an increase in the availability of berries. That we were able to relate patterns of diets and selectivity to patterns of food availability, such as the biomass of alternative food items or the seasonal changes in availability, in spite of biological and technical noise in the data indicates that those patterns are probably stronger in reality than suggested by our analyses. This explanation is supported by the difference between grey-sided voles and tundra voles, as the sample size was higher and the observed patterns both more abundant and statistically stronger for grey-sided voles. We therefore recommend a larger sample size and a more adapted way of measuring food availability for future studies on small rodent functional responses. Such larger sample size could be achieved by analyses of fecal samples, to avoid lethal methods. For example, DNA metabarcoding coupled with radiotelemetry could provide repeated individual-level data on diets, together with targeted locations for food availability estimates. Such estimates could be achieved by adjusting the point intercept method to better represent the vegetation actually available for small rodents, by for example counting only hits up to 10 cm from ground level and separating between leaves and woody plant parts. ## Conclusions We conclude that voles have diverse diets and flexible food preferences. Thus, viewing food preferences as a fixed ranking of a few species is likely to be insufficient for understanding small rodent feeding ecology. Diet diversity as such may be a functional trait of small rodent diets that previously has been underrated in the literature because of methodological constrains. Moreover, our results suggest that in order to understand small rodent functional responses, the roles of alternative food items and search time limitations should be further investigated. # Supporting Information We thank J.-A. Henden, S. T. Killengreen, M. Nilsen and numerous field assistants for data collection, S. Kaino, D. Rioux, C. Miquel and A. Valentini for help with lab work, T. Alm for taxonomical advice, J. H. Sønstebø for help with interpreting *Vaccinium*-sequences, A. Tarroux, O. Huitu and two anonymous referees for comments on the manuscript and J. Stien for checking the English. [^1]: Ludovic Gielly is one of the co-inventors of a patent concerning g-h primers and the subsequent use of the P6 loop of the chloroplast trnL (UAA) intron for plant identification using degraded template DNA. The patent has the following numbers: CA 2581347 (Canada Patent), 2006/040448 (PCT Patent), EP1797201 (EPO Patent), 20090081646 and 20110143354 (both United States Patent Application). These patents, titled “Universal primers and their use for detecting and identifying plant materials in complex mixtures”, only restrict commercial applications and have no impact on the use of this locus by academic researchers. Hence, the patents do not alter the authors’ adherence to the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: EMS RAI NGY VTR KAB LG. Performed the experiments: EMS LG VTR KAB. Analyzed the data: EMS NGY. Wrote the paper: EMS VTR RAI KAB NGY.

# Introduction The highly diverse monocot family Araceae Juss. comprises 3,319 currently recognized species classified into 118 genera. The Neotropical genus *Philodendron* Schott is of special interest as it represents the second most diverse genera of the family, with 489 accepted species, and is one of the most conspicuous and well-known groups of epiphytic and hemiepiphytic plants. The current taxonomy of *Philodendron* follows Mayo, Grayum, Croat and Sakuragui *et al.* and comprises three subgenera: *Philodendron*, *Pteromischum* and *Meconostigma*, the latter with 21 species. *Philodendron* is composed of very showy and horticulturally durable plants. These features are especially notable in subgenus *Meconostigma* and as a consequence, species of this subgenus are widely used in horticulture as valuable ornamental plants. The roots of some species, such as *P. corcovadense* and *P. williamsii* are used by traditional Brazilian populations for making rustic craft products that are widely sold in urban centers. The three subgenera of *Philodendron* were recovered as monophyletic by the molecular study of Gauthier *et al.*. Subgenus *Meconostigma* has been recognized since the first taxonomic proposal of *Philodendron* by Schott in 1829 and was most recently revised by Mayo with later updates by Gonçalves and Salviani. Members of this subgenus are well distinguished by diagnostic morphological characters as well as by a conspicuous geographical distribution, ranging from the Amazonian and Atlantic forests to savanna-like landscapes (Cerrado biome), as displayed in. Although the monophyly of subgenus *Meconostigma* is indicated by morphological, anatomical, and molecular analyses, the detailed phylogenetic relationships of its 21 extant species remained to be established. Gauthier *et al.*, who analyzed 6 species, made the largest molecular survey of subgenus *Meconostigma* hitherto. Due to their high level of variability, flowers have been one of the main sources of characters used to investigate the phylogeny of *Philodendron* species. Each female flower in *Philodendron* consists of a single gynoecium lacking staminodes or perianth parts. Together, they form a well-defined female zone at the base of the spadix. An important characteristic of the gynoecium is the presence of a separate stylar canal for each carpel, a feature used by Mayo to discuss the taxonomy and evolution of *P.* subgenus *Meconostigma*. However, in this group, morphological characters present significant plasticity, which increases the frequency of homoplasies in phylogenetic analysis. Moreover, Mayo, suggested that parallelism and convergence may be widespread in the evolution of subgenus *Meconostigma*. Such modes of evolution would hamper the phylogenetic reconstruction of this subgenus using morphology and anatomy alone. In this study, we have inferred the phylogeny of extant *P*. subgenus *Meconostigma* species using both nuclear and plastid molecular markers. With the aim of understanding the evolutionary history of the inflorescence of subgenus *Meconostigma*, this new phylogeny was used to investigate the evolution of floral characters in the group, with a focus on the gynoecium. Our findings suggest that the morphological diversity observed in the gynoecium of *Meconostigma* species is the result of an ongoing process of fusion of its flower structures leading to a reduction of energy wastage and increase in stigmatic surface and, as a consequence, the chances of fertilization. # Materials and Methods ## Species and Gene Sampling We generated molecular sequences for the nuclear 18S and external transcribed spacer (ETS) and the *trn*K intron and maturase K (*matK)* gene from the chloroplast of all extant species of subgenus *Meconostigma* and of *Philodendron* subgenus *Philodendron* (*P. pedatum*) and *Philodendron* subgenus *Pteromischum* (*P. oblongum*). We also included one species from *Homalomena* (*H. cochinchinensis*) as outgroup, the genus most closely related to *Philodendron*. Information on GenBank accession numbers and voucher of the species studied are listed in Table S1 and Table S2 in, respectively. ## Ethics Statement All living tissues were collected following the guidelines and jurisprudence of the Brazilian Ministry of Environment (MMA): SISBIO authorization no. 25755-1. Sítio Roberto Burle-Marx, Marcus Nadruz, Eduardo Gonçalves, Harri Lorenzi, Lourdes Soares and Leland Mayano provided live material from cultivated plants, which are free from MMA legislation. ## DNA Isolation, Amplification and Sequencing Genomic DNA was isolated with QIAGEN DNeasy Blood & Tissue kit from silica-dried or fresh leaves. The amplification and sequencing were conducted using the primers listed in Table S3 in. Sequencing reactions were performed in the Applied Biosystems 3730xl automatic sequencer. The consensus sequences were generated with Genious version 5.5.3. ## Alignment and Phylogenetic Analysis Sequences of the molecular markers were individually aligned in MUSCLE. Alignments were manually adjusted in SeaView and posteriorly concatenated into a single supermatrix of 3,323 base pairs. Phylogenetic inferences were performed using the maximum likelihood (ML) method as implemented in PhyML 3.0 using the GTR+G model of sequence evolution with both concatenated and isolated sequences. Bayesian analysis relied on Mr. Bayes 3.2.2, using a random starting tree and four rate categories. The Markov chain Monte Carlo (MCMC) chains ran for 1.000.000 generations, with trees sampled every 100th generation and a burn-in of 1000 trees. Model choice was conducted in Modeltest, using the likelihood ratio test. Statistical confidence of clades was assessed by the approximate likelihood ratio test statistics - aLRT. In order to evaluate the genetic similarities between species shown as very closely related in our ML and Bayesian phylogenies, we calculated their pairwise genetic distances with MEGA 5.2.2 based on our concatenated matrix. ## Ancestral State Reconstruction of Gynoecium To infer evolutionary changes in the gynoecium, we composed a reduced data set containing the 22 species that were characterized morphologically. It consisted of 19 species of subgenus *Meconostigma* (*P. xanadu* and *P. leal-costae* are absent in this analysis) with two other *Philodendron* species {subgenus *Philodendron* (1), subgenus *Pteromischum* (1)} and one *Homalomena* species as outgroup. The morphological character states of the gynoecium are listed in Table S4 in, as well as the state codes used to construct our matrix. illustrates these characters. The terminology used to name the characters follows the definitions of Mayo. Ancestral character state reconstruction was performed in Mesquite version 2.73 using the Markov k-state one-parameter model, which assumes a homogeneous probability of change between the number of states of a character. The transition parameters of the model were estimated from the phylogram obtained by the ML tree. # Results ## Phylogenetic Relationships among *Meconostigma* Species As displayed in and, subgenus *Meconostigma* was recovered as monophyletic with 100% aLRT support and 100% posterior probability, respectively. The ML likelihood and Bayesian analysis provided very similar results. In general, the species relationships in the phylogeny reflect morphological similarities among them. Within the subgenus, two major lineages could be identified in the ML and Bayesian phylogeny. One, hereafter referred to as Clade 1, consisted of two Amazonian species – *Philodendron solimoensense* and *Philodendron goeldii* (96% aLRT support, 100% posterior probability) and the other (Clade 2) consisting of a diverse assemblage of species from Amazonian and non-Amazonian biomes. Clade 2 presented *P. lundii* as the basal-most species in the ML phylogeny. However, the phylogeny based on the Bayesian approach did not recover a single species as the basal-most among the others, instead, it presented two minor clades within Clade 2, consisting of the only difference between the two trees. *Philodendron venezuelense*, the remaining Amazonian species of the subgenus, was recovered as sister species of *P*. *williamsii*, an Atlantic forest species. Interestingly, the morphologically similar species *P. brasiliense* and *P. uliginosum* were recovered as sister groups in our phylogeny; both occur in damp to periodically flooded soils in open, *campo rupestre* vegetation. Unlike the other Clade 2 species, *P. leal-costae* and *P. xanadu* presented very long branch lengths in both ML and Bayesian analysis phylogenies, possibly due to their high evolutionary rate. We calculated the genetic distance between the sister species that were most closely related on Clade 2. The results indicate low pairwise distances: *P. paludicola* and *P. dardanianum* (0.011); *P. brasiliense* and *P. uliginosum* (0.008); *P. williamsii* and *P. venezuelense* (0.008); *P. petraeum* and *P. mello-barretoanum* (0.011). In comparing the phylogenies based on the isolated markers, that based on 18S – ETS provided a better resolution for the identification of closely related species. On the other hand, *matK* – *trn*K marker provided a tree with very short branch lengths and more unresolved clades. Although these phylogenies did not comprise our complete taxon sampling, overall, the relationships among the species are similar to the ones recovered with the concatenated dataset, presenting *Meconostigma* as a monophyletic clade and Amazonian species as basal-most of the subgenus. ## Ancestral State Reconstruction of *Meconostigma* Gynoecium Ancestral character state reconstruction indicated that the common ancestor of *Meconostigma* species probably possessed short stylar lobes, long stylar canals, a stylar body, a vascular plexus in the gynoecium and druses in the stylar parenchyma. It is uncertain if raphide inclusions were present in the stylar parenchyma. Also, the ancestral lineage probably possessed up to 10 locules in the ovary. The topologies shown in suggest that the floral characters studied here are evolving independently along some branches of the tree. Specifically, two changes probably occured during the evolution of subgen. *Meconostigma* in *P. speciosum*, *P. dardanianum*, *P. williamsii* and *P. adamantinum*: an increase in height of the stylar lobes and loss of the stylar body in the gynoecium. *Philodendron speciosum* and *P. williamsii* also share two independent similarities, namely, the loss of a vascular plexus (also absent in *P. venezuelense*) and the shortening of the stylar canals, which is also observed in *P. adamantinum*. The pattern of evolution of raphides in the stylar parenchyma remains unclear. From our analyses, it is not possible to affirm whether the Amazonian species (*P. venezuelense*, *P. solimoesense* and *P. goeldii*) and *P. stenolobum* have maintained the ancestral state, because the estimated likelihood of this character in the *Meconostigma* ancestral node was 50%. However, the presence of druse inclusions was inherited from the ancestral *Meconostigma* and is maintained in all species. Finally, the Amazonian species (*P. solimoensense*, *P. venezuelense* and *P. goeldii*) and *P. speciosum*, *P. stenolobum* and *P. williamsii* show an increase in locule number in the ovary in comparison with ancestral *Meconostigma*, that probably had an ovary with less than 10 locules. # Discussion ## Evolutionary Relationships of Subgen. *Meconostigma* This is the first study to estimate the detailed phylogenetic relationships of *Meconostigma* with a complete sampling of extant species, and the first to use a well-established phylogeny to suggest the possible evolutionary scenario under which floral structures evolved in this group. In general, the evolutionary affinities among *Meconostigma* species were recovered with high statistical support and posterior probability. The Amazonian species *P. solimoesense* and *P. goeldii* were recovered as the sister group of the remaining *Meconostigma* species. *Philodendron solimoesense* was also recovered as basal by Gauthier *et al.*. This result is, however, in sharp contrast to the phylogeny of Mayo, who found *P. saxicola* as the stem species of the subgenus based on data from the ovary, stigmatic region and style. Despite the low genetic distances between the Clade 2 species pairs *P. paludicola* and *P. dardanianum*, *P. brasiliense* and *P. uliginosum*, *P. williamsii* and *P. venezuelense*, *P. petraeum* and *P. mello-barretoanum*, they are taxonomically well delimited. Thanks to the taxonomic revisions of Mayo and Gonçalves and Salviani, subgenus *Meconostigma* is a well delimited taxon, which is favored by the small size of the group and its highly distinctive morphology. These factors lead us to believe that although these species are separated by low genetic distances, they are in fact different species. ## Gynoecial Evolution of the Species of Subgenus *Meconostigma* We consider that estimating the divergence time of *Meconostigma* species is pivotal to elucidate the lineages diversification pattern and to understand the gynoecial evolution of the morphological characters based on their ancestry. However, the dating of *Meconostigma* evolutionary history still presents some obstacles. The only fossil assigned to the *Meconostigma* subgenus so far was described by Dilcher and Daghlian, based on fossilized leaves. According to the authors, the fossil would date from the Eocene of Tennessee (56,0–33,9 millions of years). However, Mayo suggested that this fossil would correspond to a Peltandreae fossil, being a member of another monocot family. As there is no convergence about the taxonomy of the referred fossil and until present there is no evidence of the occurrence of *Meconostigma*, not even *Philodendron* in North America, we have decided not to use that fossil as a calibration point. Nevertheless, as the genetic divergence is the product of the substitution rate and the time elapsed, an alternative way to estimate the divergence time would be to separate those components and to use the substitution rate. In *Meconostigma*, however, this approach presents two problems: (1) the substitution rates for ETS and *matK* markers have not been inferred for the Araceae family. Therefore, we would have to rely on rates estimates for other plant families; as those rates might range from orders of magnitude according to the taxon being considered, it would be very speculative to assume the substitution rates of these markers inferred for other families; (2) based on the available data, we are not able to calculate the specific substitution rate for *Meconostigma* or *Philodendron*, because this estimative depends on some calibration information, which is absent in the studied group. Therefore, our discussion about the gynoecial evolution of *Meconostigma* species will not consider the timing of the characters evolution. Considering the morphological characters analysed in this study, except for the length of the stylar lobes of the ancestral *Meconostigma*, other floral characters differ from those observed in extant *Meconostigma* species, indicating a trend towards the maintenance of the length of the stylar lobes, the shortening of the stylar canals and the loss of vascular plexus during the evolutionary history of this group. Our findings point to an interesting scenario underlying the evolution of these structures. Currently, it is widely accepted that the gynoecium in angiosperms corresponds to a modified form of the carpel leaf of gymnosperms and that flower structures such as petals and stamens similarly correspond to modified leaves. In the Araceae, it is notable that flower structures tend towards naked and fused states, so that we observe a rather peculiar flower morphology. Our data indicate that species that present a less developed style, in other words, shorter style lobes, do not have a vascular plexus. Considering the plexus as a structure that favours the distribution of water and nutrients to the gynoecium, it would be more advantageous in flowers with more elaborate styles. On the contrary, for flowers with shorter style lobes it would be energy saving not to maintain the vascular plexus, which might be the case of *P. speciosum*, *P. williamsii*, *P. venezuelense* and *P. adamantinum*. Likewise, the longer stylar lobes and the shortening of the stylar canals over the evolutionary history of the group can be also considered advantageous to the extent that it increases the stigmatic surface, hence the likelihood that the pollen reaches the stylar canals and, consequently, the chances of fertilization. Interestingly, these features occur concomitantly in *P. speciosum*, *P. williamsii* and *P. adamantinum*. The presence of idioblasts with different calcium oxalate inclusions in the stylar body, such as druses and raphides, is commonly associated with tissue protection. It is unclear if raphides were present in ancestral subgenus *Meconostigma*. Among the extant species, they are absent in *P. venezuelense*, *P. goeldii*, *P. solimoesense*, *P. mello-barretoanum* and *P. stenolobum*. It is interesting to notice that these species also have a large number of loci in the ovary. Although it is not possible to establish a direct relationship between the number of locules and ovules, overall, the increase of locule number can be associated with the possibility of increase in ovule number. Thus, although these species potentially lack protection in the gynoecium, they are investing more energy in the chances of fertilization. An alternative hypothesis driving the increase in locule number is that this has occurred as a defense against parasitic wasp predation. It is interesting that all *Meconostigma* species have druses, probably a character inherited from the common ancestor. In view of the absence of raphide idioblasts in the gynoecium of *P. venezuelense*, *P. goeldii*, *P. solimoesense*, *P. mello-barretoanum* and *P. stenolobum*, as previously discussed, the persistence of druses might be considered advantageous in providing protection against herbivory but at a lower cost since they are less elaborate structures. Under this evolutionary scenario, we propose that the morphological diversity observed in the gynoecium subgenus *Meconostigma* species is the result of an ongoing process of fusion of its floral structures. The resulting reduction of energy wastage and increase in stigmatic surface are likely to be evolving under positive selection. However, the role of natural selection and other evolutionary forces in this process still needs to be directly evaluated. Future studies addressing these issues should prove fruitful in confirming the hypotheses put forward here and, ultimately, contribute towards the understanding of inflorescence evolution in *Meconostigma* and other flowering plants. # Supporting Information We thank Petrobrás and INPA for allowing field expeditions in their biological reserves; Esdras Sakuragui for travel facilities; Nerivaldo Antas and Felipe Bastos for fieldwork support; Marco Octávio Pellegrini for assisting with images compositions and for fieldwork support; Aníbal Alves de Carvalho Jr. for supporting with images compositions. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LLO CGS CMS. Performed the experiments: LLO LSBC EBM. Analyzed the data: LLO LSBC CGS CMS. Contributed reagents/materials/analysis tools: CGS CMS. Wrote the paper: LLO. Performed the taxon collection: LLO LSBC EBM. Generated and error-checked the DNA sequences: LLO EBM. Analyzed the morphological characters and generated the morphological matrix: LSBC. Submitted the sequences to GenBank: LLO. Performed the phylogenetic and ancestral character reconstruction analyzes: LLO. Analyzed phylogenetic and ancestral character reconstruction data: LLO LSBC CGS CMS. Commented on the manuscript: LLO LSBC EBM SJM CGS CMS.

# 1. Introduction Emergency Departments (EDs) worldwide have to deal with rising patient volumes causing significant pressures on both Emergency Medicine (EM) and entire healthcare systems. Therefore, many EDs are in a situation where the number of patients occupying the ED is temporarily beyond the capacity for which the ED is designed and resourced to manage―a phenomenon called *Emergency Department (ED) crowding*. Particularly, ED crowding can lead to (i) reduced quality of care, (ii) longer patient waiting times for doctor’s consult (time to provider), (iii) increased numbers of patients who leave without being seen, and (iv) more ambulance diversion. ED crowding has financial implications causing costs per patient to rise because the average (inpatient) length of stay can be extended. Furthermore, it is assumed that inadequate care due to ED crowding might increase the probability of being readmitted to the ED which further contributes to rising health care costs. Since patients increasingly use EDs as point of entry into the health care system, ED crowding is not only an EM specific nuisance but rather a public health problem. Due to the relevance of ED crowding and the pressures it causes on healthcare systems, a remarkable number of studies on the topic have been published in the Operational Research (OR) literature recently. Singapore is no exception to the international trend of rising ED attendance and crowding. In 2016, the total population of Singapore amounts to 5.61 million with an average annual growth rate of 1.3%. In comparison, total ED attendance at public hospitals has grown at a disproportionately higher rate, that is, roughly 5.57% per year between 2005 and 2016. illustrates the evolution of both total ED attendance at public hospitals and population size in Singapore for the period from 2005 to 2016. Furthermore, EDs in Singapore and worldwide must cope with highly variable patient arrivals. Typically, patient arrival patterns are cyclic both during the course of a day and over the course of a week. This variable demand and the fact that patient arrivals are unpredictable and stochastic pose an additional burden on ED management teams. emphasizes the variability of patient arrivals by displaying the arrival patterns over the course of a peak day and a ‘quiet’ day in an ED in Singapore in one month. Although the number of emergency physicians (EPs) has risen, that is, 13.4% annually between 2005 and 2014, their workload has remained very high. In Singapore, on average, each EP sees between 6.4 and 8.5 patients per hour depending on the mode of calculation while previous studies have suggested that the optimal ED throughput lies between 2 and 2.8 patients per EP hour. Therefore, it can be assumed that there is still an undersupply of ED personnel and significant investments into training must be made. Considering rising patient numbers and the workload of ED staff, the current state of emergency medical care in Singapore might not be sustainable and has consequences for the well-being of both patients and ED professionals. The purpose of this paper is to develop a *virtual ED*, i.e., a simulation model that comprehensively reflects all major patient flows and medical resources of a hospital-based ED in Singapore, that is fully transparent (documented) and accessible for researchers and subject experts. Subsequently, the virtual ED is used to analyze the effectiveness of currently debated policies to streamline ED operations in Singapore. Specifically, we investigate the impact of (i) co- location of primary care services within the ED, (ii) increase in the capacity of doctors, (iii) a more efficient patient transfer to inpatient hospital wards, and (iv) a combination of policies (i) to (iii), on patients’ average length of stay (ALOS) in the ED. To that end, we use system dynamics (SD), an advanced simulation modeling approach that is currently underutilized in the modeling of ED operations. SD is a handy approach in this context because its main modeling elements, the so-called ‘stocks’ and ‘flows’, make it particularly easy to model aggregate patient flows and stock of patients in a health care setting. Consequently, SD is a useful approach to assess patient flow optimizing policies in an ED. There are only a handful of studies that analyze ED processes through an SD lens, indicating a gap in the literature (a more thorough discussion of previous modeling works follows in the literature review section). developed a model to simulate the effect of point-of-care testing on ED crowding but the model has not been made publicly available. focused on a specific subgroup of ED patients, that is, patients that were later admitted to general internal medicine in the hospital. Similarly, the model has not been made available. modeled the interplay between an ED and the associated hospital wards focusing on the trade- off between emergency admissions and elective admissions. Unfortunately, the precise sub-model referring to the processes within the ED has not been made transparent and so cannot be evaluated. created a model of hospital patient flows with the aim to define policies that reduce delays within the ED. The model has not been made available. and, two related studies, had a broader perspective and modeled emergency care systems instead of detailing a single ED. The respective models have not been made open-access. Finally, studied the acute bed blockage problem in the Irish healthcare system but refrained from modeling patient flows within the ED. The model is not available. To the best of our knowledge, there is no study using SD to create a virtual ED as we understand the term―a comprehensive representation of all major patient flows and corresponding medical resources in an ED―which is thoroughly documented and open-access. Complete model transparency and free access, however, are crucial if models shall be refined, validated, and reused by others. For that reason, in this paper, we put great emphasis on listing and explaining all model equations that are necessary to rebuild the simulation model. Furthermore, because many EDs are structured similarly having critical care (resuscitation care), isolation care, and ambulatory care areas, the model we present here can quite easily be translated into any hospital-based ED worldwide. (Currently, we are adapting the model to fit to the largest ED in Switzerland.) # 2. Literature review The public importance, the wait-for-treatment ethos and the clear structural layout of EDs have contributed to them being one of the most commonly modeled systems in OR healthcare. A recent and comprehensive literature review on simulation modeling methods applied to EDs identified in total 254 relevant publications, of which 209 used discrete event simulation (DES), 25 agent-based simulation (ABS), 18 SD, and 2 other modeling approaches. The dominance of DES in ED modeling seems justified considering the method’s strengths in handling individual patient flows and random variation of variables. We do not deny the suitability of DES in modeling ED processes and crowding. However, instead of focusing on DES alone, we argue for a diversity of simulation methods to be applied to ED operations. In our opinion, only such a multi-perspective approach can lead to new insights. In the following, we limit ourselves to reviewing the seven SD studies briefly touched upon in the introduction. Among the 18 SD works, we selected those of high-quality, written in English, and published in renowned international OR emergency medicine journals. developed an SD model to examine the effect of decreasing lab turnaround time on emergency medical services diversion, ED patient throughput, and total ED length of stay (LOS). Unfortunately, there is no information on model conceptualization. They concluded that compelling improvement in ED efficiency with decreasing lab turnaround time can be attained. constructed an SD model to study the impact of evenly distributing inpatient discharges over the course of a week on the bed occupancy rate. The model is limited to only include ED patients that are later admitted to general internal medicine (GIM) in the associated hospital. Model conceptualization entails three main components: (i) a patient category component, (ii) a hospital location component, and (iii) a feedback mechanism component. The interplay between the three components steers the movement of patients from hospital admission to discharge. They found that discharging patients evenly across the week can significantly reduce bed requirements and ED LOS. built an SD model to analyze the response of ED waiting times to reductions in bed capacity. To that end, they conceptualized the system in terms of two areas: (i) the community, and (ii) the hospital which is further subdivided into the ED, the management of elective patients, and the wards. The simulation model was subsequently used to assess the impact of changes in bed capacity and in ED demand on various key performance measures. The key finding was that reductions in bed numbers do not increase waiting times for emergency admissions because elective admissions fall sharply. So, the elective cancellation rate acts as a so-called ‘safety valve’ compensating for any change in bed capacity. developed a hospital-wide SD model to improve understanding of the causes of delays and length of stay variations experienced by patients in the ED. They tested the impact of altering nurse levels, delay reductions, and re-routing of patients on total ED length of stay, particularly for admitted patients. Overall, however, the main purpose of this study was to evaluate the applicability of SD to patient flow modeling. It was concluded that the quantitative approach to simulating ED delays and patient flows using SD is reasonable and that the resulting model is appropriately representative of the system under consideration. and adopted an SD modeling approach to describe the components of an emergency and urgent care system and to investigate ways in which patient flows and system capacity could be improved. The developed model was then used to test the effect of changes in emergency/elective admissions, ‘front door’ demand, patient discharge schemes, and bed capacity. They found that strengthening community care has the greatest potential to relieve pressure on the emergency and urgent care system. Finally, created an SD model that visualizes and simulates the dynamic flow of elderly patients in the Irish healthcare system to better understand the system’s dynamic complexity, i.e., the nonlinear interactions of system elements over time. The model focuses on general patient pathways of emergency admissions through the entire Irish healthcare system. Special emphasis is placed on post- acute care by including long-term care, care at home, convalescent care, and rehabilitation care in the model. Based on the simulation model, they evaluated various pre-acute, e.g., increasing general practitioners’ (GPs) access to community services, and post-acute, e.g., increasing discharge rates from long- term care facilities, policy interventions. They found that a mixed strategy of pre-acute and post-acute policy interventions is potentially very effective in reducing pressures on acute care provision. Based on this literature review, although not systematic, it can be said that the simulation model presented herein is the first attempt to model all relevant patient flows running through critical care, ambulatory care, and isolation care of a large interdisciplinary hospital-based ED using SD. The novelty of this work does not lie in the particular case study selected here (ED of the largest tertiary hospital in Singapore) but on the detailed and comprehensive representation of all major ED patient flows in an aggregated form. Furthermore, and equally important, the model presented below is described in such a way that interested parties can rebuild, test, and experiment with the model increasing the value of our work. # 3. Study setting ## 3.1. Methods SD is a computer-facilitated approach to policy analysis and design with a focus on modeling stocks (accumulations) and flows (rates) of systems. Typically, SD is applied to dynamic problems that are characterized by interdependence and mutual interaction of elements, information feedback, and circular causality. Virtual worlds, i.e., simulation models, created with SD can act as learning laboratories with the purpose of developing and testing strategies before they are implemented in practice. This is highly relevant for organizations nowadays considering the fact that many of them operate within increasingly dynamic environments and, therefore, strategies have to be evaluated and adjusted constantly. We chose SD as our modeling approach for the following reasons. First, it seemed important to us that ED operations are not only analyzed from one methodological viewpoint, that is, discrete event modeling, but tackled by a diversity of simulation methods in order to generate new insights. Second, agreeing with, we think that considering aggregated variables (e.g., aggregated flows of patients) which is the focus of SD encourages both a systemic view of the interactions of patient flows and information, and a more strategic perspective of the management of the system. Third, due to its accessible graphical iconography, SD is particularly useful to engage stakeholders both in the model building and in the model analysis phase. In SD, the model structure can be explained and presented in simple mathematical terms which facilitates communication with a non-technical audience. Additionally, SD models take high-level policies as inputs making them accessible for interpretation and fostering dialogue between hospital stakeholders and the modeling team. This was a key aspect to us because we intended to involve EPs, nurses, and ED managers throughout the entire modeling process. SD is still our method of choice when it comes to stakeholder involvement, despite recent efforts in facilitated discrete-event simulation modeling. ## 3.2. ED under study Singapore is a city state with a population of 5.61 million people. The study institution is the largest hospital in the country with 1’600 inpatient beds and provides tertiary care to a significant share of the population. The hospital is part of the Singhealth Regional Health System (RHS) which covers a population of more than 1.1 million people and handles more than 4 million patient visits yearly. The hospital-based ED cares for more than 140’000 patient visits annually, with about 350 visits per day in 2019. The ED is equipped with 25 specialist EPs who work an average of 180 clinical shift hours per 28 days, along with roughly 40 non-specialists who clock an average of 216 clinical shift hours in the same period. ## 3.3. Overall structure of the ED Patients come to the ED by ambulance or other forms of transportation (walk-in) from the community to seek care. Upon arrival, ED patients go through a brief registration before the triage processes commence. Triage refers to the categorization of ED patients for treatment in situations of scarce resources according to the patients’ medical conditions and established sorting plan. In Singapore, the Patient Acuity Category Scale (PACS) which prioritizes patients into four main priorities is used to triage patients at the ED. The priorities are: (1) Priority 1 are patients in a state of cardiovascular or imminent collapse. They are the most serious, time-critical patients who require immediate attention or resuscitation—examples of conditions are heart attack, severe injuries, severe bleeding, shock and severe asthma attack; (2) Priority 2 patients are non-ambulant patients with acute medical conditions who appear to be in a stable state with no immediate danger of collapse—examples of conditions are major limb fracture/dislocation, moderate injuries, severe abdominal pain and other severe medical illnesses; (3) Priority 3 refers to ambulant patients with acute symptoms who are in a stable condition. These patients could be treated by general practitioners, family physicians with acute care resources—example of conditions are sprains, minor injuries, minor abdominal pain, vomiting, fever, rashes and mild headaches; Finally, (4) Priority 4 are non-emergency patients with old injuries or conditions that have been present for a long time—examples include chronic joint pain, chronic skin rash, long- term nasal discharge, old scars, cataracts, removal of tattoos and sore throats. A trained nurse evaluates the patient’s condition, takes his or her medical history, initiates diagnostic measurements, and determines the priority for treatment, i.e., P1, P2, P3 or P4. Patients with fever, irrespective of treatment priority—P1, P2, P3, or P4—are sent to the *isolation area* to be seen by a physician to reduce the risk of infecting other patients in the ED. Non- ambulant or trolley-based patients in priority 1 and 2 are treated at the *critical care area*, while ambulatory patients, irrespective of their treatment priority are sent for treatment at the *ambulatory care area*. Each treatment area—critical care, ambulatory care, and isolation care—has a dedicated waiting area and allocated ED nurses and physicians. The average waiting time to consult an ED physician depends on the number of ED patients waiting for consultation and the number of ED physicians available. The higher the ED patient’s acuity or priority, the greater the average physician consultation time. For P1 and P2 patients receiving treatment in the critical care area, after initial consultation, almost all patients are admitted to the observation ward for observation. During observation, patients who require laboratory services undergo the investigation there and wait for the results. If there are no beds in the observation ward, patients are observed in the waiting area. For ambulatory patients, after initial consultation, laboratory services are provided. Those who require observation are admitted into the observation ward, while others wait for laboratory investigation results in the waiting area. Lastly for isolation patients, after initial consultation, patients are observed before a decision to discharge is made. For patients in the observation ward, a decision to send them home or admit them into the hospital is made after a further review by ED physicians. To that end, laboratory results (if available) are reviewed. Discharged ED patients proceed to the pharmacy for medication and payment. For those who require hospital admission, arrangement is made with the appropriate hospital ward for patient transfer. An overview of the principal processes in a hospital-based ED in Singapore is shown in. ## 3.4. Model structure SD models consist of an interconnecting set of differential and algebraic equations developed from a broad range of empirical data. SD models comprise of stocks, interconnected flows and auxiliary variables. A general mathematical representation of stocks and flows are: $$Stock(t) = {\int_{t_{o}}^{t}{\left\lbrack {inflow(t) - outflow(t)} \right\rbrack dt + Stock\left( t_{o} \right)}}$$ $$Inflow(t) = f\left( Stock(t),N \right)$$ $$Outflow(t) = f\left( Stock(t),M \right)$$ where *N* and *M* are the system parameters. The flows are the derivatives or rates of change of the associated stocks. Stocks create disequilibrium dynamics as they decouple flows. As a consequence, typically, inflows and outflows differ and are governed by different decision rules. The overall model structure of an ED in Singapore is presented in. For a list with all variables and their respective abbreviations see. ### 3.4.1. Registration and triage For ED patients, the journey begins when they arrive at the ED to seek care. New patient arrivals *a*(*t*) at any time (*t*) proceed for registration and quickly transition from registration to triage. The equation for patients waiting for registration *P*(*t*) at time (*t*) is: $$P(t) = {\int_{t_{0}}^{t}\left\lbrack a \right.}(t) - g(t)\rbrack dt + P\left( t_{0} \right)$$ where *P*(*t*₀) is patients waiting for registration at time (*t*₀) and $$a(t) = exogenous\ data$$ $$g(t) = \left\lbrack a(t)\left( {t - RT} \right) \right\rbrack$$ New patient arrivals *a*(*t*) is an exogenous input and is fed into the simulation model as historical time-series data; *g*(*t*) is patients moving from registration to triage and is represented herein as a pipeline delay function of new patient arrivals *a*(*t*) and average registration time *RT*. After registration at the ED, patients wait to be triaged. Patients are normally triaged into four treatment priorities (*j*)—P1, P2, P3, and P4; and three care areas—critical care, ambulatory care, and isolation care. *B*(*t*), that is, patients waiting for triage, increases as patients move from registration to triage *g*(*t*) and decreases as patient are triaged to critical care *cca*_*j*(*t*), ambulatory care *ab*_*j*(*t*), and isolation care *is*_*j*(*t*). The equation for patients waiting for triage *B*(*t*) is: $$B(t) = {\int_{t_{0}}^{t}{\left\lbrack {g(t) - {cca}_{j}(t) - {ab}_{j}(t) - {is}_{j}(t)} \right\rbrack dt + B\left( t_{0} \right)}}$$ where *B*(*t*₀) is patients waiting for triage at time (*t*₀) and $${cca}_{j}(t) = g(t - TT)*{fcca}_{j}(t)$$ $${ab}_{j}(t) = g(t - TT)*{fab}_{j}(t)$$ $${is}_{j}(t) = g(t - TT)*{fis}_{j}(t)$$ Patients triaged to critical care *cca*_*j*(*t*), ambulatory care *ab*_*j*(*t*), and isolation care *is*_*j*(*t*) are modeled herein as pipeline delay functions of patients moving from registration to triage *g*(*t*) and average triage time *TT*, adjusted by the fraction of patients sent to each care area; *fcca*_*j*(*t*) is the fraction of patients triaged to critical care, *fab*_*j*(*t*) is the fraction of patients triaged to ambulatory care, and *fis*_*j*(*t*) is the fraction of patients triaged to isolation care. All three fractions sum up to one. ### 3.4.2. Critical care pathways Patients triaged to the critical care area wait in queue for consultation. In total, there are four main patient pathways within the critical care area: 1. Waiting for consultation → consultation → discharge 2. Waiting for consultation → consultation → laboratory investigation → discharge 3. Waiting for consultation → consultation → observation → discharge 4. Waiting for consultation → consultation → laboratory investigation → observation → discharge The number of patients waiting for consultation *C*_*j*(*t*) increases by patients triaged to critical care *cca*_*j*(*t*) and new ambulance arrivals *nab*_*j*(*t*), and decreases as patients start consultation *cs*_*j*(*t*). Patient consultation *cs*_*j*(*t*) is initiated when an ED doctor becomes available and initiates consultation *nc*_*j*(*t*). The equation for patients waiting for consultation *C*_*j*(*t*) is: $$C_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {cca}_{j}(t) + {nab}_{j}(t) - \right.}{cs}_{j}(t)\rbrack dt + C_{j}\left( t_{0} \right)$$ where *C*_*j*(*t*₀) is patients waiting for consultation at time (*t*₀) and $${nab}_{j}(t) = exogenous\ data$$ $${cs}_{j}(t) = {nc}_{j}(t)*ppd$$ *nab*_*j*(*t*) is the exogenous historical ambulance arrival data; *ppd* is the patient per doctor ratio in the critical care area. Initiation of consultation requires an ED doctor. The number of ED doctors consulting *PCC*(*t*) increases as ED doctors initiate consultation *nc*_*j*(*t*) and decreases as consultation is completed *cc*_*j*(*t*). ED doctors available to initiate consultation *pa*(*t*) is the difference between the number of ED doctors allocated to critical care *NP*(*t*) and ED doctors consulting *PCC*(*t*). An average consultation time is assumed for each patient by treatment priority. P1 patients are assumed to require longer consultation time followed by P2, P3, and P4 patients. However, only P1 and P2 patients are triaged to the critical care area. The equation for the ED doctors consulting *PCC*(*t*) is: $$PCC(t) = {\int_{t_{0}}^{t}\left\lbrack {nc}_{j} \right.}(t) - {cc}_{j}(t)\rbrack dt + PCC\left( t_{0} \right)$$ where *PCC*(*t*₀) is ED doctors consulting at the critical care area at time (*t*₀) and $${nc}_{p1}(t) = MIN\left( pa(t),\frac{C_{p1}(t)}{AT} \right)$$ $${nc}_{p2}(t) = MIN\left( \frac{pa(t)}{AT} - {nc}_{P1}(t),\frac{C_{p2}(t)*dpp}{AT} \right)$$ $${cc}_{j}(t) = {nc}_{j}(t)(t - CT)$$ $$pa(t) = MAX\left( 0,NP(t) - {\sum{{PCC}_{j}(t))}} \right.$$ *C*_*p*1(*t*) and *C*_*p*2(*t*) are P1 and P2 patients waiting for consultation in the critical care area; AT is adjustment time—a model artifact to ensure unit consistency. The value of AT is 1. CT is consultation time; *dpp* is the doctor per patient ratio in the critical care area. A co-flow structure was used to model patients in consultation. As an ED doctor initiates consultation *nc*_*j*(*t*), a patient moves from the stock of patients waiting for consultation *C*_*j*(*t*) to the stock of patients in consultation *EP*_*j*(*t*). Hence, completion of consultation *cc*_*j*(*t*) decreases the number of patients in consultation *EP*_*j*(*t*)via to observation *co*_*j*(*t*), to laboratory and investigation *cl*_*j*(*t*) or to home *ch*_*j*(*t*). The equation for patients in consultation *EP*_*j*(*t*) is: $${EP}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {cs}_{j} \right.}(t) - {co}_{j}(t) - {cl}_{j}(t) - {ch}_{j}(t)\rbrack dt + {EP}_{j}\left( t_{0} \right)$$ where *EP*_*j*(*t*₀) is ED patients in consultation at time (*t*₀) and $${co}_{p1}(t) = MIN\left( avb(t),\ {cpo}_{p1}(t) \right)$$ $${co}_{p2}(t) = MIN\left( avb(t) - {cpo}_{p1}(t),\ {cpo}_{p2}(t) \right)$$ $${cpo}_{j}(t) = \left( {cc}_{j}(t)*ppd \right)*fb$$ $${cl}_{j}(t) = \left( {cc}_{j}(t)*ppd \right) - {co}_{j}(t) - {ch}_{j}(t)$$ $${ch}_{j}(t)\left. = \left( \left( {{cc}_{j}\left( \left. t \right)*ppd \right.} \right) - {co}_{j}(t) \right. \right)*fh$$ *avb*(*t*) is available beds in the observation ward; *cpo*_*p*1(*t*) and *cpo*_*p*2(*t*) are P1 and P2 patients requiring referral to the observation ward, *co*_*j*(*t*) is the patients from consultation to observation, *fb* is the fraction of patients who require observation, *fh* is the fraction of patients discharged home after consultation. After consultation, patients are either referred to the observation ward *co*_*j*(*t*), to laboratory and investigation *cl*_*j*(*t*) or discharged home *ch*_*j*(*t*) depending on their care needs The number of patients waiting for laboratory and investigation *PLI*_*j*(*t*), i.e., patients who have to go through the laboratory investigation process and wait for their results, increases as patients are referred to laboratory and investigation *cl*_*j*(*t*) and decreases as patients are either discharged after laboratory and investigation *ld*_*j*(*t*), transferred to the observation ward *lo*_*j*(*t*), or observed at the waiting area due to lack of beds in the observation ward *lw*_*j*(*t*). Patients under observation in the waiting area *POW*_*j*(*t*) are discharged *dw*_*j*(*t*) as their conditions improve or admitted to the hospital *cad*(*t*). The equations for patients waiting for laboratory and investigation *PLI*_*j*(*t*) and patients under observation in the waiting area *POW*_*j*(*t*) are: $${PLI}_{J}(t) = {\int_{t_{0}}^{t}\left\lbrack {cl}_{j}(t) - \right.}{lo}_{j}(t) - {ld}_{j}(t) - {lw}_{j}(t)\rbrack dt + {PLI}_{j}\left( t_{0} \right)$$ $${POW}_{j}(t) = {\int_{t_{0}}^{t}{\left\lbrack {{lw}_{j}(t) - {dw}_{j}(t) - {cad}_{j}(t)} \right\rbrack dt +}}{POW}_{j}\left( t_{0} \right)$$ where *PLI*_*j*(*t*₀) is patients waiting for laboratory and investigation at time (*t*₀), *POW*_*j*(*t*₀) is patients under observation in the waiting area at time (*t*₀), and $${lo}_{p1}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t),{al}_{p1}(t)*{fob}_{p1} \right)$$ $${lo}_{p2}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t) - {lo}_{p1}(t),\ {al}_{p2}(t)*{fob}_{p2} \right)$$ $${ld}_{j}(t) = {al}_{j}(t)*\left( {1 - fob}_{j} \right)$$ $${lw}_{p1}(t) = MAX\left( 0,\left( {al}_{p1}(t)*{fob}_{p1} - {lo}_{p1}(t) \right) \right)$$ $${lw}_{p2}(t) = MAX\left( 0,\left( {al}_{p2}(t)*{fob}_{p2} - {lo}_{p2}(t) \right) \right)$$ $${al}_{j}(t) = {cl}_{j}(t - LIT)$$ $${dw}_{j}(t) = {POW}_{j}(t)/wt$$ $${cad}_{j}(t) = {POW}_{j}(t)*fdd$$ *al*_*j*(*t*) is patients who have completed laboratory and investigation; *fob*_*j* is the fraction of patients who need to go to the observation ward after laboratory and investigation; *LIT* is the average waiting time for laboratory and investigation, *fdd* is the fraction of patients waiting in the waiting area admitted; and *wt* is the average observation time. ### 3.4.3. Ambulatory care pathways Patients triaged to ambulatory care, like other care pathways, wait in queue for consultation. In total, there are two main pathways for patients triaged to the ambulatory care area: 1. Waiting for consultation → consultation → laboratory investigation → discharge 2. Waiting for consultation → consultation → laboratory investigation → observation → discharge The number of patients waiting for consultation *CAB*_*j*(*t*) increases by patients triaged to ambulatory care *ab*_*j*(*t*) and decreases as consultation starts *csAB*_*j*(*t*). Patient consultation is initiated when an ED doctor becomes available and starts consultation *ncAB*_*j*(*t*). The number of ED doctors consulting *PCAB*(*t*) increases as an ED doctor initiates consultation *ncAB*_*j*(*t*) and decreases as consultation is completed *ccAB*_*j*(*t*). Available ED doctors to initiate consultation *paAB*(*t*) is the difference between ED physicians allocated to ambulatory care *NPAB*(*t*) and the number of ED physician consulting *PCAB*_*j*(*t*). The equations for ED patients waiting for consultation *CAB*_*j*(*t*) and ED physicians consulting *PCAB*(*t*) are: $${CAB}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {ab}_{j}(t) - \right.}{csAB}_{j}(t)\rbrack dt + {CAB}_{j}\left( t_{0} \right)$$ $$PCAB(t) = {\int_{t_{0}}^{t}\left\lbrack {ncAB}_{j} \right.}(t) - {ccAB}_{j}(t)\rbrack dt + PCAB\left( t_{0} \right)$$ where *CAB*_*j*(*t*₀) is the number of patients in ambulatory care waiting for consultation at time (*t*₀), *PCAB*(*t*₀) is the number of ED physicians consulting at time (t₀), and $${csAB}_{j}(t) = {ncAB}_{j}(t)*ppd$$ $${ncAB}_{p1}(t) = MIN\left( paAB(t),\frac{{CAB}_{p1}(t)}{AT} \right)$$ $${ncAB}_{p2}(t) = MIN\left( \frac{paAB(t)}{AT} - {ncAB}_{P1}(t),\frac{{CAB}_{p2}(t)*dpp}{AT} \right)$$ $${ncAB}_{p3}(t) = MIN\left( \frac{paAB(t)}{AT} - {ncAB}_{P1}(t) - {ncAB}_{P2}(t),\frac{{CAB}_{p3}(t)*dpp}{AT} \right)$$ $${ncAB}_{p4}(t) = MIN\left( \frac{paAB(t)}{AT} - {ncAB}_{P1}(t) - {ncAB}_{P2}(t) - {ncAB}_{P3}(t),\frac{{CAB}_{p4}(t)*dpp}{AT} \right)$$ $${ccAB}_{j}(t) = {ncAB}_{j}(t)(t - CT)$$ $$paAB(t) = MAX\left( 0,NPAB(t) - {\sum{{PCAB}_{j}(t))}} \right.$$ *ppd* is the patient per doctor ratio in the ambulatory care area; *dpp* is the doctor per patient ratio; *CT* is consultation time, and AT is adjustment time. Similar to the critical care area, a co-flow structure was developed to track patients in consultation in ambulatory care. As an ED physician initiates consultation *ncAB*_*j*(*t*), a patient moves from the stock of patients waiting for consultation *CAB*_*j*(*t*) to the stock of patients in consultation *EPAB*_*j*(*t*). Consequently, completion of consultation *ccAB*_*j*(*t*) decreases the stock of patients in consultation *EPAB*_*j*(*t*) via to laboratory and investigation *clAB*_*j*(*t*). The equation illustrating this dynamic is: $${EPAB}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {csAB}_{j} \right.}(t) - {clAB}_{j}(t)\rbrack dt + {EPAB}_{j}\left( t_{0} \right)$$ where *EPAB*_*j*(*t*₀) is ED patients in consultation at the ambulatory care area at time (*t*₀) and $${clAB}_{j}(t) = \left( {ccAB}_{j}(t)*ppd \right)$$ After consultation, patients proceed to laboratory and investigation. Patients waiting for laboratory and investigation *PHIAB*_*j*(*t*)–a procedure which includes various tests and examinations as well as waiting for test results to be discussed with the ED physician–increases as patients are referred to laboratory and investigation *clAB*_*j*(*t*) and decreases as patients are referred to the observation ward *ao*_*j*(*t*), discharged home *ldAB*_*j*(*t*) or transferred to be observed in the waiting area due to limited beds in the observation ward *lwAB*_*j*(*t*). Patients under observation in the waiting area *POWAB*_*j*(*t*) due to capacity constraints in the observation ward decreases via discharge *dwAB*(*t*) or hospital admission *aAB*_*j*(*t*). The equations for patients in laboratory and investigation *PHIAB*_*j*(*t*) and patients under observation in the waiting area *POWAB*_*j*(*t*) are: $${PHIAB}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {clAB}_{j} \right.}(t) - {ao}_{j}(t) - {lwAB}_{j}(t) - {ldAB}_{j}(t)\rbrack dt + {PHIAB}_{j}\left( t_{0} \right)$$ $${POWAB}_{j}(t) = {\int_{t_{0}}^{t}\lbrack}{lwAB}_{j}(t) - {dwAB}_{j}(t) - {aAB}_{j}(t)\rbrack dt + {POWAB}_{j}\left( t_{0} \right)$$ where *PHIAB*_*j*(*t*₀) is the initial number of patients in the ambulatory care area waiting for laboratory and investigation at time (*t*₀), *POWAB*_*j*(*t*₀) is the number of patients under observation at time (*t*₀), and $${lwAB}_{p1}(t) = MAX\left( 0,\left( {ala}_{p1}(t)*{foba}_{p1} - {ao}_{p1}(t) \right) \right)$$ $${lwAB}_{p2}(t) = MAX\left( 0,\left( {ala}_{p2}(t)*{foba}_{p2} - {ao}_{p2}(t) \right) \right)$$ $${lwAB}_{p3}(t) = MAX\left( 0,\left( {ala}_{p3}(t)*{foba}_{p3} - {ao}_{p3}(t) \right) \right)$$ $${lwAB}_{p4}(t) = MAX\left( 0,\left( {ala}_{p4}(t)*{foba}_{p4} - {ao}_{p4}(t) \right) \right)$$ $${ldAB}_{j}(t) = {ala}_{j}(t)*\left( 1 - {foba}_{j} \right)$$ $${dwAB}_{j}(t) = {POWAB}_{j}(t)/wt$$ $${aAB}_{j}(t) = {POWAB}_{j}(t)*famb$$ *ala*_*j*(*t*) is the number of patients who have finished laboratory and investigation, and *foba*_*j* is the fraction of patients who have completed laboratory and investigation and are admitted to the observation ward, while *famb* is the fraction of patients observed in the waiting area and are admitted to the hospital. ### 3.4.4. Observation ward and discharge The observation ward receives patients from critical care and ambulatory care areas—patients triaged to isolation care have a separate observation ward. The number of patients in the observation ward *OW*_*j*(*t*) increases as critical care patients are referred to it immediately after consultation *co*_*j*(*t*) or after laboratory and investigation *lo*_*j*(*t*), as well as referral of ambulatory patients after laboratory and investigation *ao*_*j*(*t*), and decreases as patients are admitted into the hospital *ah*_*j*(*t*) or discharged home via pharmacy and payment *do*_*j*(*t*). Admission into the observation ward depends on the available beds *avb*(*t*). Available beds *avb*(*t*) is the difference between observation bed capacity *bc*(*t*) and the number of patients in the observation ward *OW*_*j*(*t*). After observation, discharged patients go through pharmacy and payment *PHAB*_*j*(*t*) for payment and collection of prescribed medication. The number of patients in pharmacy and payment *PHAB*_*j*(*t*) increases as patients are discharged from the observation ward *do*_*j*(*t*), as patients are released from laboratory and investigations both in critical care *ld*_*j*(*t*) and ambulatory care *ldAB*_*j*(*t*), as well as patients are discharged from observation in waiting areas, both in critical care *dw*_*j*(*t*) and ambulatory care *dwAB*_*j*(*t*), as well as patients discharged after consultation from critical care *ch*_*j*(*t*) and decreases as patients leave for home *hab*_*j*(*t*). The equations for observation ward admission and discharge are: $${OW}_{j}(t) = {\int_{t_{0}}^{t}\lbrack}{co}_{j}(t) + {lo}_{j}(t) + {ao}_{j}(t) - {ah}_{j}(t) - {do}_{j}(t)\rbrack dt + {OW}_{J}\left( t_{0} \right)$$ $${PHAB}_{j}(t) = {\int_{t_{0}}^{t}\lbrack}{ldAB}_{j}(t) + {dwAB}_{j}(t) + {do}_{j}(t) + {ld}_{j}(t) + {dw}_{j}(t) + {ch}_{j}(t) - {hab}_{j}(t)\rbrack dt + {PHAB}_{j}\left( t_{0} \right)$$ where *OW*_*j*(*t*₀) is the initial number of patients in the observation ward at time (*t*₀), *PHAB*_*j*(*t*₀) is the initial number of patients in the stock pharmacy and payment at time (*t*₀), and $${ao}_{p1}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t) - {\sum{lo}_{j}}(t),\ {ala}_{p1}(t)*{foba}_{p1} \right)$$ $${ao}_{p2}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t) - {\sum{lo}_{j}}(t) - {ao}_{p1}(t),{ala}_{p2}(t)*{foba}_{p2} \right)$$ $${ao}_{p3}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t) - {\sum{lo}_{j}}(t) - {ao}_{p1}(t) - {ao}_{p2}(t),{ala}_{p3}(t)*{foba}_{p3} \right)$$ $${ao}_{p4}(t) = MIN\left( avb(t) - {\sum{co}_{j}}(t) - {\sum{lo}_{j}}(t) - {ao}_{p1}(t) - {ao}_{p2}(t) - {ao}_{p3}(t),{ala}_{p4}(t)*{foba}_{p4} \right)$$ $${ah}_{j}(t) = {aobw}_{j}(t) - {do}_{j}(t)$$ $${aobw}_{j}(t) = {OW}_{t}/ot$$ $${do}_{j}(t) = {aobw}_{j}(t)*df$$ $$avb(t) = \left( \left( {bc(t)*ppb} \right) - {\sum{ow}_{j}}(t) \right)/ttba$$ where *avb*(*t*) is the total available beds in the observation ward; *ala*_*j*(*t*) is the number of patients who have finished laboratory and investigation; *foba*_*j* is the fraction of patients who have completed laboratory and investigation and are admitted to the observation ward; *df* is the fraction of discharged patients from the observation ward; *ot* is the average observation time for patients in the observation ward; *ppb* is the patients per bed ratio; *ttba* is the time to make a bed available for patients to use. ### 3.4.5. Isolation care pathways Patients triaged to isolation care, like other care pathways, wait in queue for consultation. In total, there are two main pathways for patients triaged to isolation care: 1. Waiting for consultation → consultation → discharge 2. Waiting for consultation → consultation → observation → discharge The number of patients waiting for consultation *CIS*_*j*(*t*) increases by patients triaged to isolation care *is*_*j*(*t*) and decreases as consultation starts *csIS*_*j*(*t*). Patient consultation is initiated when an ED physician becomes available and initiates consultation *ncIS*_*j*(*t*). The number of ED physicians consulting *PCIS*(*t*) at the isolation care area increases as an ED physician starts consultation *ncIS*_*j*(*t*) and decreases as consultation is completed *ccIS*_*j*(*t*). Available ED physicians to initiate consultation *paIS*(*t*) is the difference between ED physicians allocated to isolation care *NPIS*(*t*) and ED physicians currently consulting with patients *PCIS*_*j*(*t*). The equations for patients waiting for consultation *CIS*_*j*(*t*) and ED doctors consulting in the isolated care area *PCIS*(*t*) are: $${CIS}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {is}_{j}(t) - \right.}{csIS}_{j}(t)\rbrack dt + {CIS}_{j}\left( t_{0} \right)$$ $$PCIS(t) = {\int_{t_{0}}^{t}\left\lbrack {ncIS}_{j} \right.}(t) - {ccIS}_{j}(t)\rbrack dt + PCIS\left( t_{0} \right)$$ where *CIS*_*j*(*t*₀) is patients in the isolation care area waiting for consultation at time (*t*₀), *PCIS*(*t*₀) is the number of ED physicians consulting at time (*t*₀), and $${csIS}_{j}(t) = {ncIS}_{j}(t)*ppd$$ $${ncIS}_{p1}(t) = MIN\left( paIS(t),\frac{{CIS}_{p1}(t)}{AT} \right)$$ $${ncIS}_{p2}(t) = MIN\left( \frac{paIS(t)}{AT} - {ncIS}_{P1}(t),\frac{{CIS}_{p2}(t)*dpp}{AT} \right)$$ $${ncIS}_{p3}(t) = MIN\left( \frac{paIS(t)}{AT} - {ncIS}_{P1}(t) - {ncIS}_{P2}(t),\frac{{CIS}_{p3}(t)*dpp}{AT} \right)$$ $${ncIS}_{p4}(t) = MIN\left( \frac{paIS(t)}{AT} - {ncIS}_{P1}(t) - {ncIS}_{P2}(t) - {ncIS}_{P3}(t),\frac{{CIS}_{p4}(t)*dpp}{AT} \right)$$ $${ccIS}_{j}(t) = ncis(t)\left( {t - CT} \right)$$ $$paIS(t) = MAX\left( 0,NPIS(t) - {\sum{{PCIS}_{j}(t))}} \right.$$ where *ppd* is the patient per doctor ratio in the isolation care area; *dpp* is the doctor per patient ratio; AT is adjustment time; and *CIS*_*P*1, *CIS*_*p*2, *CIS*_*P*3, *and CIS*_*P*4 are the stocks of patients waiting for consultation. Similar to other care areas, a co-flow structure was developed to model patients in consultation in the isolation care area. As an ED physician initiates consultation *ncIS*_*j*(*t*) a patient moves from the stock of patients waiting for consultation *csIS*_*j*(*t*) to the stock of patients in consultation *EPIS*_*j*(*t*). Hence, completion of consultation *ccIS*_*j*(*t*) decreases the stock of patients in consultation via to observation *coIS*_*j*(*t*) and for discharge *cpIS*_*j*(*t*). After consultation, patients referred to the observation ward *coIS*_*j*(*t*) are observed in the observation ward *OWIS*_*j*(*t*). After a period of observation time *otis*, patients in the observation ward are either discharged *oph*_*j*(*t*) or admitted into the hospital *ahis*_*j*(*t*). Likewise, patients discharged from the isolation care area, i.e., via observation ward *oph*_*j*(*t*) and or after consultation *cpIS*_*j*(*t*), proceed to the pharmacy and payment *PHIS*_*j*(*t*) for prescribed medicine and payment and then leave *his*_*j*(*t*). The equations illustrating the stock of patients in consultation *EPIS*_*j*(*t*), the stock of patients in observation *OWIS*_*j*(*t*), and the stock of patients in pharmacy and payment *PHIS*_*j*(*t*) are: $${EPIS}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {csIS}_{j} \right.}(t) - {coIS}_{j}(t) - {cpIS}_{j}(t)\rbrack dt + {EPIS}_{j}\left( t_{0} \right)$$ $${OWIS}_{j}(t) = {\int_{t_{0}}^{t}\left\lbrack {coIS}_{j} \right.}(t) - {oph}_{j}(t) - {ahis}_{j}(t)\rbrack dt + {OWIS}_{j}\left( t_{0} \right)$$ $${PHIS}_{j}(t) = {\int_{t_{0}}^{t}{\lbrack{oph}_{j}}}(t) + {cpIS}_{j}(t) - {his}_{j}(t)\rbrack dt + {PHIS}_{j}\left( t_{0} \right)$$ where *EPIS*_*j*(*t*₀) is the initial number of patients in consultation at the isolation care area at time (*t*₀), *OWIS*_*j*(*t*₀) is the number of patients under observation at time (*t*₀), *PHIS*_*j*(*t*₀) is the number of patients in pharmacy and payment at time (*t*₀), and $${coIS}_{j}(t) = \left( {{ccIS}_{j}(t)*ppd} \right)*fis$$ $${cpIS}_{j}(t) = \left( {{ccIS}_{j}(t)*ppd} \right)*(1 - fis)$$ $${ahis}_{j}(t) = {owd}_{j}(t)*fah$$ $${oph}_{j}(t) = {{owd}_{j}(t) - ahis}_{j}(t)$$ $${owd}_{j}(t) = {owis}_{j}(t)/otis$$ $${his}_{j}(t) = phis(t)/ppt$$ where *fis* is the fraction of patients who need to go to the observation ward; *owd*_*j*(*t*) is the number of patients who were observed and proceed to discharge or hospital admission; *fah* is the fraction of patients who were observed and are admitted to the hospital; *otis* is the average time patients were observed; *ppt* is the average time at the pharmacy and payment. ## 3.5. Data sources We parameterized a simulation model that runs for 24 hours. To that end, we had access to ED data for the period of June–August 2017. The key parameter values essential for estimating ALOS are *average registration time*, *average triage time*, *consult time* (for critical care, ambulatory care and isolation care), *average waiting time for observation ward*, *lab and investigation waiting time*, *pharmacy and payment waiting time*, *time to make bed available*, and *average time to admit patients*. The distribution of the key parameter values is assumed to follow a triangular distribution. The estimated values used for the lower limit *a*, upper limit *b*, and mode *c* are as follows: For registration and triage, the values are—*average registration time* \[Min = 3; Median = 5; Max = 7\], and *average triage time* \[Min = 4; Median = 5; Max = 7\]. For critical care the input values are—*consult time* \[Min = 10; Median = 15; Max = 20\], *average observation waiting time* \[Min = 30; Median = 60; Max = 90\], and *laboratory and investigation* \[Min = 35; Median = 45; Max = 60\]. For ambulatory care, the input values are—*consult time* \[Min = 10; Median = 15.5; Max = 20\], and *average observation waiting time* \[Min = 30; Median = 60; Max = 90\]. For isolation care, the values are—*consult time* \[Min = 20; Median = 30; Max = 45\], *average observation waiting time* \[Min = 30; Median = 60; Max = 90\], and *pharmacy and payment waiting time* \[Min = 10; Median = 15; Max = 30\]. For observation ward and discharge, the input values are—*time to make beds available* \[Min = 10; Median = 15; Max = 25\], and *average time to admit patients* \[Min = 100; Median = 120; Max = 150\]. Referring to patient arrivals, we picked the highest daily patient arrival pattern in the 3-months period from June to August 2017 because we intentionally wanted to stress-test the system. Operational problems in the ED become only visible when the workload is high, and the system is stretched to its limits. To supplement the data requirements, an observational study on other process timings that were not captured in the patient records was conducted over a 2 weeks period in November 2017 to estimate some of the model parameters. Finally, for all parameters that could not be observed nor estimated from the data, we had to rely on expert judgment. More information about model inputs, their values, units, and sources can be found in. The simulation model is provided in the supplementary file for review. ## 3.6. Model validation To ensure that the ED model developed herein is fit for purpose and robust and that the results could be used to inform ED policies, structure and behavior validation tests were conducted. Structure validation tests focused on engaging ED doctors with significant experience in the operations of the ED in Singapore to verify the model structure and its assumptions regarding causal relationships, feedbacks, time delays, and patient flows. This validation process was conducted in four different meetings—where the model structure was thoroughly reviewed—to ensure that the model structure is as close to reality as possible. Hence, we believe that the current model structure is firmly grounded in current operations of a hospital-based ED in Singapore. On behavior validation tests, the simulation results were compared to available data of selected outcomes (average length of stay across all venues—CCA, ambulatory, and isolation). In addition, a mean absolute percentage error (MAPE) and a Theil statistic analysis were conducted to check the behavioral validity of the model. The MAPE—which is a measure of prediction accuracy—for the selected outcomes were; 9.78% for ALOS in the critical care area, 12.5% for ALOS in the ambulatory care area, and 8.85% for ALOS in the isolation care area. Given that MAPE of 30% is considered to be good, our results—which have a maximum MAPE of 12.5%—indicate that the simulated model compares well with available data considering that available ALOS data used was an average across venues over an hour. For the Theil statistic, the error due to bias (U^M) for CCA, ambulatory, and isolation care areas were 11%, 4.5%, and 17.7% respectively; while that for unequal variance (U^S) for CCA, ambulatory, and isolation care areas was 2.4%, 7.1%, and 9.4% respectively; and the error for covariation component (U^C) for CCA, ambulatory, and isolation care areas were 87.1%, 88.4%, and 72.9% respectively. Thus, critical care, ambulatory care, and isolation care areas have most of the error within the covariation component (U^C) as compared to bias (U^M) and unequal variance (U^S). For Theil statistics, if majority of the errors comes from covariation components, it indicates that the simulated variables track the underlying trend well, but diverge when comparing point-by- point, indicating that majority of the errors are unsystematic with respect to the purpose of the model. # 4. Policy experiments Policy experimentation was conducted to explore the range of potential future directions on how to manage ED crowding as discussed with the ED physicians, as well as address the identified bottlenecks—significant waiting times for consultation, laboratory investigation and observation ward for admission—in the ED care processes. To that end, we tested the impact of four policies on the average length of stay (ALOS) of ED patients by venue of care, care pathway, and time period of the day. Policies were compared to the base case, where the status quo is simulated. We use ALOS as a proxy for ED crowding. The tested policies are: 1. **Business-as-usual (BAU)**: The business-as-usual (BAU) or base-case experiment assumes no change to key model inputs that may be affected by current or future policies. Under this policy experiment, patients reporting at the ED for care are triaged to CCA, ambulatory care, or isolation care. The current allocation of ED doctors across the venues remains unchanged and waiting time for patients in the observation ward is assumed to remain constant. This hypothetical scenario is unlikely in the current context as new policies are expected to change some of these key variables. However, it is included to serve as a reference point for evaluating alternative policies. 2. **Co-location** (policy 1): This policy experiment varies the fraction of P4 and P3 patients decanted from the ED to a GP clinic co-located in the ED from 10% to 30% to assess the impact on ALOS. The rationale of this policy is to relieve ED Operations by redirecting non-emergency patients to primary care services. In Singapore and other industrialized countries, there is a tendency to co-locate primary care services within EDs. These primary care services are clearly separated from ED operations where more severely ill patients are treated. In such a way, the flow of patients with only minor and non-emergency symptoms are cared for in a separate venue with own resources. This reduces the heterogeneity in acuity of patients in the other venues of the ED. Overall, this policy aims to increase efficiency and effectiveness of ED operations by diverting non-emergency patients to GP care. 3. **Capacity of doctors** (policy 2): This policy experiment stepwise increases the capacity of ED doctors by 10%, 20%, and 30%, across all venues of care, to evaluate its impact on waiting time and ALOS. This policy experiment aims to achieve the target ALOS of 4 hours for patients allocated to critical care and ambulatory care and 4.5 hours for patients triaged to isolation care. Singapore loosely follows the ‘4-hour rule’ from the UK where 98% of all ED patients must be seen and discharged or admitted within 4 hours of their arrival. In Singapore and other wealthy countries such as Switzerland, managers of EDs typically tried to meet waiting time targets by hiring more physicians and nurses. So, this policy simply reflects the attempt to balance the increased demand for emergency medical services by increasing the supply of these services. However, there is a clear financial limit to such a policy. As more and more health systems are forced to reduce spending and become more efficient, a policy of matching increased demand with increased supply might not be a viable one in the long-term. 4. **Observation ward and laboratory** (policy 3): This policy explores the impact of a 10%-30% reduction in waiting time at the observation ward, as well as waiting time for laboratory and investigation on ALOS across all venues of care. This policy experiment aims to achieve the target ALOS of 4 hours for patients allocated to critical care and ambulatory care and 4.5 hours for patients triaged to isolation care. The rationale of this policy is to analyze the consequences of an improved patient outflow from the ED. One of the most severe bottlenecks of a hospital-based ED is the outflow of patients who cannot be discharged but need to be admitted to the hospital. Because of limited hospital capacities these patients typically accumulate in the ED and gradually fill up the observation ward. Consequently, these ‘boarded’ patients block ED resources and create significant inefficiencies in the system. This policy analyzes the implications of an improved transfer of ED patients to hospital wards by reducing the waiting time at the observation ward (i.e., the boarding time). Furthermore, this policy tests the consequences of an increased turnover for laboratory and investigation (e.g., more efficient blood testing). 5. **Combined interventions** (policy 4): This policy experiment implements all the previous interventions—i.e., co-location, capacity of doctors, and observation ward and laboratory—simultaneously to assess it impact on ALOS across all venues of care. # 5. Results – Figs show the ALOS by venue of care―critical care, ambulatory care, and isolation care—care pathway and time of the day, for all the policy experiments. To present the result in a meaningful way (as shown in Tables –), we divided the day into three time periods—herein referred to as phases. *Phase 1* (ph1) is from 00:00 am to 08:00 am, *phase 2* (ph2) is from 08:00 am to 04:00 pm, and *phase 3* (ph3) is from 04:00 pm to 00:00 am. The results are presented as follows: ## 5.1. Business-as-usual (BAU) As shown in Tables, in the BAU case, a critical care patient that goes through critical care pathway 1 for emergency care is projected to experience an ALOS of **25 minutes** from *00*:*00 am to 08*:*00 am*; **39 minutes** from *08*:*00 am to 04*:*00 pm*; and **27 minutes** from *04*:*00 pm to 00*:*00 am*. The estimated ALOS for critical care pathway 2 patients is **72 minutes** from *00*:*00 am to 08*:*00 am*; **85 minutes** from *08*:*00 am to 04*:*00 pm*; and **73 minutes** from *04*:*00 pm to 00*:*00 am*. The ALOS for critical care pathway 3 patients is **149 minutes** from *00*:*00 am to 08*:*00 am*, **162 minutes** from *08*:*00 am to 04*:*00 pm* and **150 minutes** from *04*:*00 pm to 00*:*00 am*. Lastly, the ALOS for critical care pathway 4 patients is **195 minutes** from *00*:*00 am to 08*:*00 am*, 209 minutes from *08*:*00 am to 04*:*00 pm* and 197 minutes from *04*:*00 pm to 00*:*00 am*. For patients seeking emergency care at the ambulatory care venue, and who go through ambulatory care pathway 1 are estimated to experience an ALOS of **72 minutes** from *00*:*00 am to 08*:*00 am*; **166 minutes** from *08*:*00 am to 04*:*00 pm*; and **85 minutes** from *04*:*00 pm to 00*:*00 am*. For ambulatory pathway 2 patients, under the base-case, a projected ALOS of **196 minutes** from *00*:*00 am to 08*:*00 am*; **290 minutes** from *08*:*00 am to 04*:*00 pm*; and **208 minutes** from *04*:*00 pm to 00*:*00 am* is expected. Lastly, patients at the isolation care area following the isolation care pathway 1 are projected to experience an ALOS of **42 minutes** at all phases. For isolation care pathway 2 patients, ALOS is projected to be **101 minutes** from *00*:*00 am to 04*:*00 pm* and **102 minutes** from *04*:*00 pm to 00*:*00 am*. ## 5.2. Co-location (policy 1) As indicated in, under policy 1, where 10% to 30% of P4 and P3 patients from each venue of care are decanted from the ED to a GP clinic co-located in the ED to provide needed care, the ALOS for critical care patients going through critical care pathways 1 to 4 are projected to be the same as BAU. However, for patients going through ambulatory care pathway 1 and 2, ALOS was projected to reduce under policy 1. In the scenario where 10% of P4 and P3 patients were decanted, ALOS for ambulatory care pathway 1 is projected to decrease by **24.6%** from *08*:*00 am to 04*:*00 pm*, and by **4.7%** from *04*:*00 pm to 00*:*00 am*, compared to the BAU. The ALOS of patients in ambulatory care pathway 1 arriving in the time period between *00*:*00 am and 08*:*00 am* remains unchanged for the 10%, 20%, and 30% scenario; for ambulatory care pathway 2 ALOS is projected to decrease by **14.1%** from *08*:*00 am to 04*:*00 pm*, while that for *04*:*00 pm to 00*:*00* am is **1.44%**, compared to the BAU. The ALOS of patients in ambulatory care pathway 2 presenting themselves between *00*:*00 am and 08*:*00 am* remains unchanged for the 10%, 20%, and 30% scenario. In the scenario where 20% of P4 and P3 patients were removed from the ED, ALOS for ambulatory care pathway 1 is projected to fall by **36.1%** from *08*:*00 am to 04*:*00 pm*, while that for *04*:*00 pm to 00*:*00 am* is **8.2%**, compared to the BAU; for ambulatory care pathway 2 ALOS is forecasted to diminish by **20.7%** from *08*:*00 am to 04*:*00 pm*, while that for *04*:*00 pm to 00*:*00 am* is **3.4%**, compared to the BAU. Finally, in the scenario where 30% of P4 and P3 patients are redirected to an inhouse GP clinic, ALOS for ambulatory care pathway 1 is projected to fall by **42.2%** from *08*:*00 am to 04*:*00 pm* and by **11.8%** from *04*:*00 pm to 00*:*00 am*, compared to the BAU; for ambulatory care pathway 2 ALOS is forecasted to reduce by **24.5%** from *08*:*00 am to 04*:*00 pm* and by **4.3%** from *04*:*00 pm to 00*:*00 am*, compared to the base case. Lastly, policy 1 does not show an impact on isolation care pathways in our experimental set-up. On average, only 6% of all presenting patients are triaged to isolation care (i.e., fever patients) and so patient flow is smooth through this care venue. This means that fever patients typically do not need to wait until they see a doctor. Consequently, reducing the inflow of patients into isolation care (which is the effect of policy 1) does not change the ALOS of patients going through this venue of care. ALOS is driven by the waiting time for lab and investigation and if necessary, by the waiting time for a bed in the hospital. Both waiting times are not influenced by policy 1. ## 5.3. Capacity of doctors (policy 2) As shown in, under this policy where the number of ED doctors is gradually increased by 10%, 20% and 30%, a critical care patient going through critical care pathways 1 to 4 is projected to experience ALOS similar to that of the BAU from *00*:*00 am to 08*:*00 am* and *04*:*00 pm to 00*:*00 am*. However, during the time period of *08*:*00 am to 04*:*00 pm*, with a scenario where the doctors allocated to each venue is increased by 10%, ALOS is projected to decrease by **20.5%**, **8.2%**, **4.3%** and **3.8%** respectively for care pathways 1 to 4. In the scenarios where a 30% increase of doctor’s allocation was experimented, ALOS is projected to decrease by **28.2%**, **12.9%**, **6.7%** and **5.2%** respectively for care pathways 1 to 4. Similarly, under the ambulatory care pathways 1 and 2, ALOS is projected to be similar to that of the BAU from *00*:*00 am to 08*:*00 am*. However, ALOS is projected to decrease under the 10% increase in doctor’s capacity scenario by **34.3%**, and **20%** for ambulatory care pathways 1 and 2 respectively from *08*:*00 am to 04*:*00 pm*; while that for *04*:*00 pm to 00*:*00 am* was **7.05%** and **2.4%**. Under the 30% increase in doctor’s capacity scenario, ALOS is expected to reduce by **49.3%** from *08*:*00 am to 04*:*00 pm* and **12.9%** from *04*:*00 pm to 00*:*00 am* for ambulatory care pathway 1; and **28.2%** from *08*:*00 am to 04*:*00 pm* and **5.2%** from *04*:*00 pm to 00*:*00 am* for ambulatory care pathway 2. Lastly, ALOS for patients going through isolation care pathways are expected to experience ALOS like that of the BAU. The drivers of ALOS for patients in isolation care are the waiting times for lab and investigation and admission to the hospital. Both waiting times are not changed (reduced) by increasing the capacity of doctors in the ED. ## 5.4. Observation ward and laboratory (policy 3) As shown in, under the observation ward and laboratory policy, a critical care patient going through critical care pathways 1 is projected to experience ALOS comparable to that of the BAU across all time phases. But, for critical care pathway 2, 3 and 4, ALOS is projected to decrease as observation ward and laboratory and investigation waiting times are reduced. Under the scenario where observation ward and laboratory waiting times were reduced by 10%, compared to the BAU, ALOS is expected to decrease by **6.9%**, **4.7%** and **5.4%** for critical care pathway 2 across the three-time phases; **8%**, **7.4%** and **8%** for critical care pathway 3 across the three-time phases; while that for critical care pathway 4 was **8.7%**, **8.1%** and **8.6%** respectively across the three phases. Under the scenario where observation ward and laboratory waiting times were reduced by 30%, ALOS for critical care pathway 2 is projected to reduce by **19.4%** from *00*:*00 am to 08*:*00 am*, **16.4%** from *08*:*00 am to 04*:*00 pm* and **19.1%** from *04*:*00 pm to 00*:*00 am*; while that for critical care pathway 3 was **24.8%** from *00*:*00 am to 08*:*00 am*, **22.8%** from *08*:*00 am to 04*:*00 pm* and **24.6%** from *04*:*00 pm to 00*:*00 am*. Likewise, the ALOS for critical care pathway 4 is projected to reduce by **26.15%** from *00*:*00 am to 08*:*00 am*, **24.4%** from *08*:*00 am to 04*:*00 pm* and **25.8%** from *04*:*00 pm to 00*:*00 am*. Considering the ambulatory care pathways, under the 10% reduction in observation ward and laboratory waiting times ALOS is projected to decrease **6.9%**, **3%** and **5.8%** respectively across the time phases for care pathway 1; while that for care pathway 2 is projected to decrease by **8.6%**, **5.8%** and **8.1%** respectively across the time phases. Under the 30% reduction in waiting times, ALOS is projected to decrease by **19.4%**, **8.4%** and **16.4%** respectively for care pathway 1, while projections for care pathway 2 are reductions of **26%**, **18.6%** and **24.5%** respectively. For isolation care pathways, while isolation care pathway 1 is projected to remain unchanged relative to the BAU, isolation care pathway 2, under the 10% reduction in observation ward and laboratory waiting times is projected to decrease ALOS by **5.9%** across all time phases. However, under the 30% reduction in observation ward and laboratory waiting times, ALOS is projected to decrease by **16.8%**, **17.8%** and **17.6%** respectively across the time phases. ## 5.5. Combined interventions (policy 4) As indicated in, under the combined interventions where policies 1 to 3 are implemented simultaneously, under the 10% assumptions—where 10% of P4 and P3 patients are decanted to a GP clinic in the ED, 10% increase in doctors allocated to each care venue and 10% reduction in observation ward and laboratory waiting times—a critical care patient that goes through critical care pathway 1 for emergency care is projected to experience **20.5%** reduction in ALOS from *08*:*00 am to 04*:*00 pm*; while that for *04*:*00 pm to 00*:*00 am* is **3.7%**. Interestingly, ALOS rises by **4%** for patients arriving between *00*:*00 am and 08*:*00 am*. For critical care pathway 2 a reduction in ALOS of **6.9%**, **14.1%**, and **6.8%** are projected across the three-time phases. Patients going through the critical care pathway 3 are projected to experience **8%**, **12.3%** and **8.6%** reduction in ALOS across the time phases. Lastly, patients going through critical care pathway 4 are projected to experience **8.7%**, **11.9%** and **9.1%** reduction in ALOS across the time phases. Under the 30% assumptions, patients going through the critical care pathways 1 to 4 are projected to experience greater reductions in ALOS as indicated in. For critical care pathway 1 a reduction in ALOS of **28.2%** is projected from *08*:*00 am to 04*:*00 pm* and of **3.7%** for *04*:*00 pm to 00*:*00 am*. The ALOS of patients arriving between *00*:*00 am and 08*:*00 am* remains unchanged compared to the BAU. For critical care pathway 2 a reduction of **19.4%**, **29.4%**, and **20.5%** is projected across the three-time phases; while that for critical care pathway 3 are **24.8%**, **29.6%** and **25.3%** respectively across the time phases. Lastly, patients going through critical care pathway 4 are projected to experience **26.1%**, **29.6%** and **26.9%** reduction in ALOS across time phases under policy 4. For patients going through the ambulatory care venue, ambulatory care pathway 1 patients, under the 10% assumptions, are projected to experience a reduction in ALOS by **6.9%** from *00*:*00 am to 08*:*00 am*, **43.9%** from *08*:*00 am to 04*:*00 pm*, and **15.2%** from *04*:*00 pm to 00*:*00 am*; while the reduction in ALOS for ambulatory care pathway 2 patients is **8.6%** from *00*:*00 am to 08*:*00 am*, **29.6%** from *08*:*00 am to 04*:*00 pm* and **12%** from *04*:*00 pm to 00*:*00 am*. As is the case with all the policies, under the 30% assumptions, ALOS is projected to reduce much more compared to the 10% assumptions as indicated in. Finally, for patients going through the isolation care venue, ALOS for isolation care pathway 1 patients is projected to remain unchanged relative to the BAU across all time phases. For isolation care pathway 2, under the 10% assumptions, ALOS is projected to decline by **5.9%**, **5.9%** and **5.8%** respectively across time phases, whereas under the 30% assumptions, ALOS is projected to decline by **16.8%**, **17.8%** and **17.6%** respectively across time phases. # 6. Discussion ## 6.1. General remarks ED crowding is a multifaceted issue and so far, many solutions have failed because they ignored or underestimated the dynamic complexity of the problem. After thoroughly reviewing the literature on ED crowding, recommends to ‘research systems-wide solutions on the basis of existing evidence and operations theory, with the aim of mitigating the risk/problem of crowding.’ For this reason, in the present study, we analyzed ED crowding from a systems thinking perspective to explicitly account for the problem’s dynamic complexity caused by a web of interrelated influencing factors. More specifically, we used SD to map and simulate the interrelations among variables affecting and affected by ED crowding. The resulting simulation model helps to better understand the dynamic nature of this phenomenon and it can serve as an effective decision support system because of its capability to test policy proposals *in silico*. This has the pivotal advantage that ED mangers can experiment with policy proposals and study their consequences in a risk-free environment. Given the current arrival pattern of patients and the configuration of the hospital-based ED in Singapore, a patient seeking emergency care, with the exception of a few periods of the day, is highly probable to receive care within the target time of 4 hours if triaged to the critical care or ambulatory care units or within 4.5 hours if triaged to the isolation care unit. As expected, the longer the care pathway of the patient (which includes consultation, laboratory, and observation), irrespective of the care venue, the larger the ALOS. Patients triaged to the critical care unit have the shortest ALOS compared to ambulatory and isolation care patients. Policies that focus on decanting P4 and P3 patients to a GP clinic co-located within the ED are more likely to reduce the ALOS of ambulatory patients, since all the P4 and P3 patients are triaged to the ambulatory care unit. Likewise, a policy that increases the efficiency of patient transfer from the observatory ward in the ED to the hospital ward, as well as decreasing the waiting time of laboratory investigation is more likely to reduce the ALOS of patients whose care pathway includes the observation ward and laboratory investigations. The observed results can be explained by the interaction between the implemented policies (i.e., co-location policy, optimal allocation of doctor’s policy, and observation ward and laboratory policy) and available ED capacity. For example, as patients are decanted to GP clinics on arrival at the ED, the share of patients triaged for emergency care decreases; therefore, waiting time for consultation will decrease resulting in a drop in ALOS. The decline in consultation waiting time is due to the fall in the number of patients waiting for consultation. Thus, available resources (ED doctors) are able to care for fewer patients demanding ED services. In addition, raising the number of ED doctors (via the optimal allocation policy) increases the resources (ED doctors) available to provide care, hence a reduction in ALOS as consultation waiting time declines. Lastly, an efficient transfer of patients from the observation ward to the acute hospital wards, as well as reduction of waiting time for laboratory and investigation is expected to decrease the ALOS of patients whose pathway includes the observation ward and laboratory and investigation. These policies were implemented based on the identification of ED bottlenecks—waiting time for consultation, laboratory and investigation waiting time and observation ward waiting time. These were identified as bottlenecks due to the significant time patients spend in these venues, thus increasing the ALOS of ED patients. ## 6.2. Main findings of this study Ambulatory care patients with priority 3 and 4 make up the lion’s share of all attending patients in our ED under study. On average, 55% of all patients are triaged to the ambulatory care area. Consequently, even a small reduction in the number of incoming patients into the ambulatory care area has a significant impact on waiting times and ALOS. This key finding has policy implications. Overall, the finding suggests that a comprehensive ED system that anticipates inappropriate self-referral of patients and makes provisions for such patients—by transferring such patients to a GP clinic in the ED—is more likely to triage only appropriate ED patients for emergency care, thus reducing the pressure on available ED resources. The lesson from this finding is that, there is a significant proportion of ED patients who inappropriately self-refer to the ED. Policymakers should anticipate that behavior and either provide GP care co- located at the ED or incentive non-emergency patients—by reducing the out-of- pocket-cost—to seek GP care first before coming to the ED. In Singapore, a pilot intervention called *GP first*, which incentivizes patients with less serious conditions to see their GP first before going to the ED was shown to reduce the number of non-emergency patients seeking care at the ED. In addition, EDs should be incentivized (by sharing savings from non- emergency patients triaged to co-located GPs) to triage patients accurately—to prevent up-coding where patients are assigned to higher severity than the actual condition to justify their use of ED services—and to ensure that patients are right-sited for care, that is, patients receive care at the appropriate venue with the lowest cost. The implication of this finding is that if EDs are inappropriately incentivized referring to the triage function, e.g., punished for providing ED care to non-emergency patients (P4), the EDs are likely to up- code non-emergency patients preventing the opportunity to improve care efficiency and reduce cost. Lastly, it is important to emphasis that, a sustainable approach to reduce ED crowding will require a well-functioning enhanced primary care system that improves health outcomes of the population and significantly lessens ED care demand among non-emergency patients. This is vital for countries with an aging population where demand for healthcare services is expected to increase. If the primary care system is not strengthened to provide appropriate care for the elderly population with multiple chronic diseases, inappropriate demand for ED care is expected to increase with its consequences of ED crowding. Given that the main bottlenecks in most EDs are significant waiting times for consultation, laboratory investigation, and for hospital admission (mostly through the observation ward), it is important to ensure that proactive and innovative interventions are explored to reduce waiting times in these locations of ED care. Interventions that focus on the optimal allocation of ED doctors (typically increasing their numbers) should be explored to reduce consultation waiting time. In addition, efficient operating systems that ensure speedy transfer of patients from the observation ward in the ED to hospital wards should be implemented. For instance, interventions that focus on (i) categorization of wards to medical specialty; (ii) instituting a no reject policy; and (iii) performing ward level audits have been shown to improve waiting time for hospital admission. ## 6.3. Limitations of this study The model presented here has some limitations. First, the use of an SD modeling approach for modeling ED patient flows introduces patient mixing that makes it difficult to track each patient individually; however, there are other modeling forms—such as agent-based modeling (ABM)—that focus on simulating the actions and interactions of autonomous agents that address this limitation. Patient mixing and the assumption that patients triaged into the same category have similar characteristics is an oversimplification that may affect the results. Second, transfer of patients to other hospitals—which was not relevant in our case—was not included in the model; whereas the likely impact of nurses and other allied health workers on waiting time was not included. Third, modeling results are reported as single values (ALOS) without an indication of the statistical uncertainty for the different venues of care and time phases. This could be improved by deriving confidence intervals for the modeling results through Monte Carlo simulation. Despite these limitations, the model presented herein remains useful for policymakers to test and evaluate innovative policies. For instance, the model could be used as an exploratory tool to search for high- leverage policies and to evaluate the likely impact of alternative policies on specific outcomes of interest. In addition, the model could help policymakers to design and communicate policy insights to stakeholders to build consensus and inform policy implementation. # 7. Conclusions This paper provides a detailed simulation model structure of patients’ flows in a hospital-based ED in Singapore allowing for the exploration and evaluation of policies. The insights generated from the policy experiments suggest that to reduce ED crowding an enhanced primary care system is required. A strengthened primary care system has the potential to improve health outcomes of the population and, as a consequence, to reduce the demand for non-emergency care at the ED. In view of this result, policymakers should design a cost-effective way to enhance primary care, co-locate GP clinics in all EDs to decant non-emergency patients seeking care at the ED, as well as incentivize all EDs to accurately triage patients and to send patients to the appropriate venues for care. # Supporting information 10.1371/journal.pone.0244097.r001 Decision Letter 0 Kuo Yong-Hong Academic Editor 2021 Yong-Hong Kuo This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. # Transfer Alert This paper was transferred from another journal. As a result, its full editorial history (including decision letters, peer reviews and author responses) may not be present. 11 Aug 2020 PONE-D-20-14680 Modeling Emergency Department Crowding: Restoring the Balance between Demand for and Supply of Emergency Medicine PLOS ONE Dear Dr. Schoenenberger, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Sep 25 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Yong-Hong Kuo Academic Editor PLOS ONE Additional Editor Comments: This manuscript has been reviewed by two experts in this area. Overall, they believe that the manuscript has potential to be published at PLOS ONE. However, there are still major concerns to be addressed before the manuscript can be considered for publication. Please refer to the Reviewers' comments to Author for the detailed comments. Both reviewers also wish that the contributions of the research can be better presented. As the reviewers suggest, the literature on ED overcrowding is huge. Some recent significant related studies are suggested for review: Ghanes, K., Wargon, M., Jouini, O., Jemai, Z., Diakogiannis, A., Hellmann, R.... & Koole, G. (2015). Simulation-based optimization of staffing levels in an emergency department. Simulation, 91(10), 942-953. Hu, X., Barnes, S., & Golden, B. (2018). Applying queueing theory to the study of emergency department operations: a survey and a discussion of comparable simulation studies. International transactions in operational research, 25(1), 7-49. Kuo, Y. H., Chan, N. B., Leung, J. M., Meng, H., So, A. M. C., Tsoi, K. K., & Graham, C. A. (2020). An Integrated Approach of Machine Learning and Systems thinking for Waiting Time Prediction in an Emergency Department. International Journal of Medical Informatics, 104143. Kuo, Y. H., Rado, O., Lupia, B., Leung, J. M., & Graham, C. A. (2016). Improving the efficiency of a hospital emergency department: a simulation study with indirectly imputed service-time distributions. Flexible Services and Manufacturing Journal, 28(1-2), 120-147. Uriarte, A. G., Zúñiga, E. R., Moris, M. U., & Ng, A. H. (2017). How can decision makers be supported in the improvement of an emergency department? A simulation, optimization and data mining approach. Operations Research for Health Care, 15, 102-122. Vanbrabant, L., Braekers, K., Ramaekers, K., & Van Nieuwenhuyse, I. (2019). Simulation of emergency department operations: A comprehensive review of KPIs and operational improvements. Computers & Industrial Engineering, 131, 356-381. Vanbrabant, L., Martin, N., Ramaekers, K., & Braekers, K. (2019). Quality of input data in emergency department simulations: Framework and assessment techniques. Simulation Modelling Practice and Theory, 91, 83-101. Yousefi, M., Yousefi, M., & Fogliatto, F. S. (2020). Simulation-based optimization methods applied in hospital emergency departments: A systematic review. Simulation, 0037549720944483. Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. We note you have included a table to which you do not refer in the text of your manuscript. Please ensure that you refer to Table 2 and 3 in your text; if accepted, production will need this reference to link the reader to the Table. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Partly \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: I Don't Know \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: This paper presented a comprehensive simulation model built for a hospital-based emergency department (ED) in Singapore using system dynamics (SD). The model was then used to test the performance of three policies intended to streamline patient flow in the ED and reduce crowding. Policy 1 works by decanting lowest-priority patients from the ED to the general practitioner (GP) clinic. Policy 2 adds to Policy 1 by enforcing a 20% reduction on the waiting time in the observation ward of patients to be transferred to the main hospital. Policy 3 improves upon both Policy 1 and 2 by adding 20% staffing capacity. The main finding shows that Policy 1 most significantly decrease patient average length-of-stay (ALOS). The model has thoroughly captured all major patient flow mechanisms in the three primary divisions within the ED system, i.e., critical care area, isolation area, and ambulant area, and meticulously discussed in simple mathematical terms. However, I have several major concerns about the methodology as well as some of the assumptions adopted by the authors. First, the model assumed constant sojourn time for the majority of the service processes such as waiting time (for the providers), consultation time (with the providers), time spent waiting for lab results, and time spent in the observation ward. The authors frankly acknowledged these assumptions as limitations yet provide very little justification or ramification for them. In my opinion, it is critical to capture the time- and census- dependency of these sojourn times to accurately reflect the potential improvement of a policy on the real-world ED system. It was mentioned that the authors had access to patient arrival data and conducted an additional observational study to obtain the process timings. Thus, I would suggest them to utilize the arrival and departure time stamps in the data to impute empirical distributions for the corresponding sojourn times. This effort, if feasible, should increase the robustness and credibility of the results. Second, there seems to be room for improvement for the policies examined. To be more specific, Policy 2 imposed a 20% reduction on boarding time (time spent in the observation ward waiting to be transferred to the main hospital). I wonder where this value comes from – whether it is realistic, and if so, how it can be achieved. Also, instead of presenting an isolated value I would recommend testing a few more scenarios, e.g., from 5% to 50% with 5-10% increment. Although I hold doubt about the feasibility of reaching a time invariant reduction in boarding time, it would be great to see some explanations for the choice of this value, as well as some concrete examples of how it can be achieved in a real ED system. My concern with Policy 3 can be summarized in a similar vein. I.e., there seems to be a lack of rationale behind the single choice of 20% for the increase in physician capacity. Moreover, the authors addressed this policy as dynamic. I could be missing something here, but it seemed to be a static policy in that the staffing capacity was not dependent on the system state (time or census level). It would be interesting to see the authors incorporating a few more sophisticated policies targeting at unclogging the bottleneck of the queues, as well as providing some discussions around how the target reduction of demand or increase of supply can be obtained in reality. To my knowledge, the majority of the literature on ED simulation utilizes discrete-event simulation, the advantage of which over SD is that one can track the system state and obtain statistics at individual patient level. As a result, it is a drawback that one cannot evaluate the policies’ impact on ALOS by patient triage priority, especially given the fact that the bottleneck of ED crowding usually lies in the non-critical patients consisting the majority of the arrivals. Still, I think this paper does bring in novelty by exploring ED dynamics through the lens of SD. I personally have very limited knowledge of this simulation technique and have not used it for my own research. Yet, upon a short literature review, I see that SD does offer some appealing features: First, the model structure can be explained and presented in simple mathematical terms, which is suitable to be communicated to non-technical audience; Second, the model takes high-level policies as inputs, which again, makes it accessible for interpretation and necessitates dialogue between hospital stake holders and the modeling team. Furthermore, the major finding of this paper coincides with the landscape of policy reforms addressing ED crowding in Singapore. As the authors pointed out, the fact that a simple policy of decanting lowest priority patients from ED to GP clinics significantly reduces ALOS serves as a strong backbone for the country’s “GP first” movement. I appreciate that the authors are taking the pioneering step in tackling the problem of simulating EDs using a SD perspective and I think that after some refinement this work can serve as a motivation and basis for future work on diversifying ED simulation techniques and alternative policy development and implementation. Finally, I would like to outline the rest of my minor comments/recommendations as below: 1\. Add a section for input parameter analyses and present estimated values for the key processing times such as triage time, registration time, waiting time to see a physician, consultation with a physician, waiting time to be transferred, etc. 2\. Provide definition of AT (adjustment time) in the paper and short explanations for its use in the equations. 3\. In my humble opinion, it would be sufficient and possibly serve better visual purpose to present smoothed hourly average ALOS instead of the histogram of raw simulation data in Figure 10 and 11. 4\. In Figure 10 and 11, it would be good to use more distinctively different colors for base case scenario and target time horizontal line. 5\. Since in some scenarios policy 1 and 2 had no effect on reducing ALOS it would be less confusing to use the same legend or add a footnote to the bottom of Figure 10 to highlight this part of the results. I hope the authors find my comments helpful and wish the authors the best in progressing this research. Reviewer \#2: In this paper, a system dynamics simulation model of an ED is developed and extensively described. The simulation model is used to investigate three policies to reduce the impact of ED crowding on ED performance. Although the topic of the paper is interesting, the paper needs a major revision before it can be considered for publication in PLOS ONE. Special attention should be given to the positioning/communication of the scientific contribution of the paper. It is not clear whether the main contribution and novelty of this study lies in the development of a system dynamics simulation model, in the investigation of the three policies to improve ED performance, or both. In addition, the main part of the paper deals with the mathematical description of the system dynamics simulation model, while the sections dealing with the experiments and results are rather short. A more thorough explanation of the base case setting in the ED under study, the experiments and the results can provide added value to the paper. While I believe the authors have done a lot of relevant work, they fail at this moment to convince the reader of the main academic and practical contributions of their research. The following remarks and comments can be made for each of the paper sections: \- Reading the Abstract, the actual contribution and main conclusion of the paper is not clear. The authors should explicitly state the novelty and main insights of their work \- The Introduction section fails to clarify the academic and practical contribution of the paper. What has already has been done in the field? How does this paper differ from existing research? What is its specific added value, both for theory and practice? o p2, last paragraph: “Typically, patient arrival patterns are cyclic … This variable demand poses an additional burden on ED management teams.”. The fact that arrival patterns are cyclic doesn’t necessarily result in a more difficult management of EDs, as the same pattern can be repeated every time. However, in addition to the cyclic pattern, patient arrivals are unpredictable and stochastic, which makes ED management more difficult. o p2, last paragraph: “Although the number of emergency physicians (EPs) has risen, that is, 13.4% annually…”. Can it be explained why there is a high physician workload and ED crowding, while the number of ED physicians has risen more than double as fast as the number of ED patient visits? One of the investigated policies focuses on the capacity of ED physicians, but are they really the bottleneck resource? o Figure 2: The word “quite” should be replaced by “quiet” in the caption. o p4, first paragraph: “This allows identifying bottlenecks in the system, i.e., ED venues where patients accumulate and put a strain on ED performance.”. The authors define bottlenecks as areas in the ED, rather than specific ED resources. However, resolving a bottleneck which is a complete area in the ED is difficult to accomplish and requires further examination of the specific problem areas within the venue. In addition, this paragraph suggests that bottlenecks will be identified by means of the simulation model, which is not the case in the paper. o p4, first paragraph: The description of the three investigated policies is a little bit misleading. For example, it is stated that the paper will investigate “How ED manpower needs to be adjusted in order to dissolve congestions and smooth ED patient flows”, while only a 20% increase in physician capacity for all venues at all times is investigated, without a prior identification of the bottleneck resources/venues/times and accompanying required adjustments in ED manpower. o p4, last paragraph: “To the best of our knowledge, there is no study using SD on the development of a virtual ED, as we understand the term―a simulation model that comprehensively reflects all major patient flows and medical resources in an ED―that is fully transparent (documented) and accessible for researchers and subject experts.”. Is this the main contribution of the paper? And to what extent is the developed simulation model generic and reusable by other researchers, as the structure of the simulation model is based on the ED under study (e.g. care pathways, venues…)? \- The value of the Literature Review section is limited, for the following reasons: o Only six studies in the vast amount of ED simulation literature are described, while the authors indicate that a recent literature review paper (Salmon et al., 2018) identified 18 studies that applied system dynamics to EDs. Why only focus on these six studies? What is the link/difference between the described studies and this paper? What are the shortcomings of these studies and how are they dealt with in the current study? o Some references are wrongly placed between brackets in this sections, e.g. (Lane et al., 2000). o p5, first paragraph of Literature Review section: “… (the rationale for choosing SD is explained in the study setting section below)”. As this is the second time the authors refer to the study setting section for the rationale behind choosing SD, a short explanation can already be provided here to justify this choice and to clarify the contributions compared to existing literature. Why is it important to use SD in addition to all DES studies? What are the advantages? o It is not clear how this paper contributes to the existing (and large) body of literature. What is the contribution to existing literature and what is link between existing literature and this research? In case the main contribution of this paper lies within the use of system dynamics to model an ED, a more thorough discussion of the advantages of SD, existing SD models of EDs, and the novelties within this model is required (see previous comment). In case the main contribution are the investigated policies and accompanying insights, an overview of existing studies that examine the three policies investigated in this paper can provide useful information on the added value of this paper compared to existing studies. \- From the Methods section (Section 3.1), it does not become clear what the advantages of SD are. Why is it important to also use other methods than DES, and especially SD? Which new insights can SD provide? The use of a diversity of simulation methods is only valuable when all the different methods have their advantages (which is the case, but does not become clear from the text). \- In Section 3.2 on the overall structure of the ED, some more information regarding the ED under study and the characteristics of the ED would be interesting, as there exists no general structure that represents every ED, and the effectiveness of improvement policies depends on ED characteristics (e.g. yearly number of patient arrivals, number of physicians working in each venue/shift, number of beds in the ED, percentages of patients per venue…). This might also enhance the understanding of the experiments and results presented in sections 4, 5 and 6. o p8, first paragraph: “The higher the ED patient’s acuity or priority, the greater the average physician consultation time.”. How is this determined, based on real-life data, observations, literature? o Figure 3: This figure also contains ‘treatment’ and ‘follow-up consultation’ as part of the patient flow through the ED. Are these also included in the simulation model? Based on the care pathways discussed in Section 3.3, all patients seem to have only one consultation with a physician, which does not represent the actual patient flow in most EDs. \- The Model Structure section (Section 3.3) is difficult to read and understand because of the many equations and abbreviations. Depending on the main contribution of the paper (i.e. system dynamics or policy investigation), this section can be shortened and some equations can be placed in the appendix. For example, the four care pathways in the critical care area (Section 3.3.2) are all possible pathways in the ED (i.e. no new pathways are introduced in the other two zones). As the pathways and accompanying mathematical equations are discussed in great detail in Section 3.3.2, the description in Sections 3.3.3 and 3.3.5 can be shortened by indicating that the reasoning and equations are comparable to the pathways in the critical care area, but with other parameter values. In addition, the abbreviations make it difficult to understand the mathematical equations and model description. Would it be possible to provide an explanation of all the stocks and flows in words, in addition to the mathematical equations? Especially for non-experts (e.g. ED staff) this might be useful to better understand the model. o p9, first paragraph: “… developed from a broad range of empirical data.”. What empirical data is used in this study? As the appendix contains a table with all model parameters and their source, a reference to this table should be added to the text. o p13, second paragraph: “… AT is adjustment time …”. What is meant with adjustment time? o p14, equation (33): As POW refers to all patients under observation in the waiting area, shouldn't this equation only contain the fraction of patients that is discharged home, because POW also contains admitted patients and these are not discharged home? An explanation in words would enhance the understandability of the equations. o p14, Section 3.3.3: Why is ‘AB’ added to all abbreviations to refer to ambulatory care (and IS to refer to isolation care in Section 3.3.5), while no reference to critical care is added to the abbreviations in Section 3.3.2? Try to be consistent, as this will enhance the understanding of the many abbreviations. \- The Data sources section (Section 3.4) is very short, especially as very little information is provided within the other sections on the characteristics of the ED under study, the data used to determine model parameters and the values of these parameters. o p22, last paragraph: “The patient arrival data used in the simulation model was the highest patient arrival pattern during the period of June – August 2017 …”. Does this arrival pattern consist of just one day of patient arrivals? Or one month? What are the run parameters of the simulation model? And why only using the highest arrival pattern, is this representative for actual ED operations? o P23, last sentence of Section 3.4: “The simulation model is provided in the supplementary materials for review.”. There was no supplementary material available regarding the simulation model. \- In the Model validation section (Section 3.5), a more comprehensive explanation of the statistical tests is required to better understand the validation process. How to interpret MAPE, and how to determine whether the MAPE value indicates a good fit between the simulation model and actual data? The same for the Theil statistic, how to interpret the U-values and how to determine whether they represent a good fit? o p23, second paragraph of section 3.5: The behavior validation tests only look at the average length of stay across all venues. However, when the variability in LOS within and across venues (and time periods) is high, the average LOS of the simulation model and data can be comparable, while the simulation model poorly represents actual ED operations. o p23, last sentence: “This suggests that the simulated variables track the underlying trend well, but diverge when comparing point-by-point, indicating that majority of the errors are unsystematic with respect to the purpose of the model.”. Explain? And what does the difference in U-values between the critical care area, ambulatory care area and isolation care area mean? Does the model only provide a good fit for some of the areas? \- Regarding the Policy Experiments section, some information on the reason for choosing these policies is lacking. What is the current ED performance and what are the bottlenecks? How are the experiments determined in order to resolve these bottlenecks? o Base case scenario: In the text, this scenario is referred to as a ‘hypothetical scenario’, but isn’t this the current ED setting? o Policy 3: An increase by 20 percent seems a very high increase, given the fact that the number of patients in the ED decreases through co-location. Is an equal increase in the number of ED physicians required in all venues and at all times (e.g. the average LOS is higher in phase 2 than in the other phases)? In the results section, it appears that the introduction of additional physician capacity is not capable of reducing LOS for all patients in the ED. \- The Results section contains a lot of interesting results, but an explanation of the results is lacking. What explains the large differences in LOS between different care pathways and time intervals? Why are certain policies more effective to reduce LOS? And why do some care pathways benefit more from a certain policy? In addition, are the improvements in LOS statistically significant? o Based on the results in Table 1, policy 1 seems least effective. Why do the authors only focus on this policy in the conclusion section? o Figure 10: The base case scenario seems lacking on these graphs. o Figure 10 and 11: How can the large fluctuations in LOS be explained (even within a time period/phase)? o Figure 11, isolation pathways: The target time is not presented on the graph regarding isolation pathway 1, probably because of the scale on the y-axis. Why is the scale on the y-axis different on both graphs of the isolation pathways? \- In the Discussion and Conclusions sections, the focus is on the insights regarding the co-location of general practitioners in the ED. I agree that this is a very interesting improvement option to reduce ED crowding, but are there no main insights regarding the other policies and the SD model? The conclusions should provide an overview of the main contributions of the paper, and based on the current conclusions the only contribution seems to be the investigation of the co-location of general practitioners in the ED. If this is the main contribution, the introduction and literature review section should be rewritten to focus on this contribution (e.g. what literature does already exist on the use of GPs in an ED, and what is the added value of this study compared to existing literature?). Furthermore, a critical reflection on how the investigated policies can be implemented in practice, and on the generalizability of the results, is lacking. o p29, last paragraph: “In addition, raising the number of ED doctors (via the dynamic allocation policy) …”. The investigated policy seems no dynamic allocation policy, as only a 20% increase in physician capacity for all venues and time periods is investigated. o p30, first paragraph of Section 6.2: “The key finding that decanting non- emergency patients who seek care at the ED to a GP clinic co-located at the ED significantly decreases ALOS and improves patient experience has policy implications.” Is this the key finding? Based on the results in Table 1, this policy seems least effective to reduce LOS. o p30, first paragraph of Section 6.2: “The lesson from this finding is that, there is a significant proportion of ED patients who inappropriately self-refer to the ED.” Is this true? The policy investigates the effect of referring 50% of the P3 and P4 patients to a GP, but does 50% of these patients inappropriately go to the ED for care? How is this percentage determined? o Section 6.3: The advantages of using SD do not become clear from this section (e.g. SD results in an oversimplification of actual ED operations). o Section 7: A paragraph on future research opportunities is lacking. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0244097.r002 Author response to Decision Letter 0 31 Oct 2020 Dear editor, dear reviewers, we have uploaded our responses to your requests in a separate file. Thanks for your time and effort. Best 10.1371/journal.pone.0244097.r003 Decision Letter 1 Kuo Yong-Hong Academic Editor 2021 Yong-Hong Kuo This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 19 Nov 2020 PONE-D-20-14680R1 Modeling Emergency Department Crowding: Restoring the Balance between Demand for and Supply of Emergency Medicine PLOS ONE Dear Dr. Schoenenberger, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Jan 03 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Yong-Hong Kuo Academic Editor PLOS ONE Additional Editor Comments (if provided): The revision has been reviewed by one of the reviewers from the last round. The other reviewer was unable to accept the review invitation. I have gone through the revision and reviewer's comments. I suggest minor revision. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#2: I Don't Know \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#2: The authors are to be commended for revising the manuscript. The academic and practical contribution of the paper have been clarified, and the experiments and results sections are extended and rewritten, which enhances the quality of the paper. Following are some comments that still need to be addressed: Section 2 \- The contributions of the paper are clarified in the introduction, but the added value of the literature review section is still rather limited. As indicated by the authors, an extensive literature review is not the goal of the paper and also not required. However, it is indicated that 18 papers exist on the application of SD in an ED context, while only 6 papers are discussed. At least, an explanation should be provided for only discussing these 6 papers in detail: Why are these papers most relevant to be discussed? In addition, a short discussion on how the current paper relates to the existing (and discussed) literature should be included at the end of the section to position the paper within the state-of-the-art literature. Section 3 \- The term ‘adjustment time’ should be clarified within the paper. In the response to review comments (comment 19) an explanation is provided, but this explanation should also be included in the paper such that readers understand this term. \- Section 3.5, p23: “The patient arrival data used in the simulation model was the highest patient arrival pattern during the period…”. In the response to reviewer comments (comment 23), it became clear that a 24 hour arrival pattern corresponding to a peak day in patient arrivals was used as input to the simulation model, in order to clearly represent the operational problems in the ED. However, this is still not clear from the text, so it might be interesting to include the explanation of the response to reviewer comments in the text. \- Section 3.5: Patient records and observations are indicated as the two data sources, but expert opinion also seems an important source of input data based on S1 Table 5 and is currently not mentioned in this section. \- Section 3.7: This section clearly adds value to the paper, but it can be integrated with Section 3.5, since both sections are rather short and deal with the use of input data to determine simulation model parameters. \- Section 3.6: The explanation of the statistics makes this section more clear, but in my opinion an explanation for the difference in Theil statistic value between the different care areas is still lacking. Why is there a difference between the care areas? And since the majority of errors for the ambulatory care area does not come from the covariation component, is the simulation model a good representation of actual operations in the ambulatory care area? Section 4 \- p26, policy 3: “This policy experiment gradually increases the number of ED doctors currently allocated to CCA, ambulatory care and isolation care from 10% to 30%...” This sentence is misleading in two different ways: (1) Based on this sentence, it seems that the number of allocated doctors is currently at 10% and gradually increases to 30%, but the base case scenario is a 0% increase; (2) In the next paragraphs, it becomes clear that the capacity of doctors is gradually increased by 10%, 20% and 30%, but based on this sentence it seems that the number of allocated doctors varies between 10% and 30%, but then the reader might pose the question: 10% or 30% of which doctors is allocated to the ED? Of the total amount of doctors in the hospital?. In order to avoid confusion, ‘Optimal capacity of doctors’ might be a better name for the policy, since the capacity of doctors is increased (no change in allocation of doctors between hospital departments). Section 5 \- Sometimes not all time phases are discussed in the description of the results o Section 5.2: The results for the ambulatory care area are only discussed for time phases 2 and 3, not for time phase 1 (00:00am- 08:00am). o Section 5.5: Only 08:00am – 04:00pm is discussed for critical care pathway 1 under the 30% scenario, 00:00am-08:00am is lacking for ambulatory care pathway 1 under the 10% scenario. \- A lot of results are presented in this section, but an explanation of the results is sometimes lacking. An explanation can enhance the understanding of the results and support the main insights from the experiments. o Section 5.2: Why is there no impact of policy 1 on isolation care pathways, since there are also P3 and P4 patients in this care area? o Section 5.3: Why is there no influence on the ALOS of isolation care patients? The capacity of physicians also increased in this area? o Section 5.4: There seems to be only an influence on the ALOS for care pathways that include observation, not for care pathways that include laboratory examinations without observation. How can this be explained, as the waiting time for laboratory examinations is also reduced? \- Are the results regarding the comparison of the different policies with the BAU scenario statistically significant? In simulation there is a lot of randomness, which can for example be seen in the graphs of S1 Fig 10-13, so is it possible to compare the results of different policies only based on a mean value without statistical tests? Section 6 \- Section 6.2: From the text, it is still not clear why the insights regarding the co-location policy are the key findings of the paper. Other policies also seem efficient for several types of patients, sometimes even more efficient as they result in a higher reduction in ALOS. In the response to reviewer comments (comment 38), it is indicated that ambulatory care patients with priority P3 and P4 are the large majority of patients in the ED, and a small reduction in the number of patients from this category already highly impacts ED performance. This explanation should be included in this section to explain why this is indicated as the key finding of the study. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#2: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0244097.r004 Author response to Decision Letter 1 1 Dec 2020 Dear reviewer, please find our reply to your comments at the end of our submitted document. Thanks for your time and effort. Warm regards 10.1371/journal.pone.0244097.r005 Decision Letter 2 Kuo Yong-Hong Academic Editor 2021 Yong-Hong Kuo This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 3 Dec 2020 Modeling Emergency Department Crowding: Restoring the Balance between Demand for and Supply of Emergency Medicine PONE-D-20-14680R2 Dear Dr. Schoenenberger, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. Kind regards, Yong-Hong Kuo Academic Editor PLOS ONE Additional Editor Comments (optional): Based on the reviewer's recommendation and comments, the comments from the last round have been successfully addressed and so I recommend accept. Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#2: No 10.1371/journal.pone.0244097.r006 Acceptance letter Kuo Yong-Hong Academic Editor 2021 Yong-Hong Kuo This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 7 Dec 2020 PONE-D-20-14680R2 Modeling Emergency Department Crowding: Restoring the Balance between Demand for and Supply of Emergency Medicine Dear Dr. Schoenenberger: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. If we can help with anything else, please email us at <plosone@plos.org>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Yong-Hong Kuo Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction The importance of understanding and using science for public policy-making has long been recognised, but recent years have seen a growing debate over how this is best achieved. Still more recently, ‘evidence-based policy’ has become the desired norm in many fields (even if its meaning is still disputed), and this has led to a greater embedding of scientists, both natural and social, alongside other specialists in public policy–making processes. In many governments, scientists are engaged at a senior level. The US, for example, has the President's Council of Advisors on Science and Technology, while the UK has Chief Scientific Adviser posts in all government departments, in addition to a Government Chief Scientific Adviser with a place in some Cabinet Committees. In spite of their acknowledged importance however, relations between science and policy are sometimes troubled, and periodically erupt into controversy. Prominent examples include the acrimonious debate over scientific understandings of climate change, further inflamed by the ‘Climategate’ email controversy, disputes over the use of genetically modified crops and foods in Europe, the failure to acknowledge the risk of possible BSE transmission to humans, and conflict over stem cell research, which is particularly acute in the United States. In 2009, the public sacking of the Chair of the UK Advisory Council on the Misuse of Drugs began a row not only about appropriate policy (in this case for drugs classification), but also about the proper place of independent scientific advice in the policy-making process. Such troubles are symptomatic of the complexity of science-policy interactions, and suggest that there is still much to understand about the nature of scientific authority and processes of policy formation and change. Against this backdrop, this paper reports the results of an exercise that sought to identify the most important outstanding questions in this domain. Precedents for attempts to identify ‘key questions’ go back to the learned civic societies of enlightenment England and France. For example, the Royal Society for the Encouragement of Arts, Manufactures, and Commerce (founded 1752) and the French National Institute (1795–1983) identified specific policy-relevant questions for which they offered prizes to promote commercial and social applications of science. Other examples include Hilbert's famous set of mathematical questions , Paul Erdös' posing of mathematical questions with cash prizes for those who solved them and Steffen *et al's* listing of questions in the environmental sciences. Contemporary ‘top down’ examples include the US National Research Council, in its assessment of strategic directions for the geographical sciences, and the International Council for Science, with its Grand Challenges in Global Sustainability Research. We have adopted a rather different, bottom-up, approach, bringing together researchers, policy makers and practitioners with interests in relations between science and policy to identify priority, researchable questions in this field. The method is similar to that used in conservation biology – and agricultural science. Previous exercises have been remarkably influential. For example, two of the resulting papers, were the most downloaded ever from their respective journals, and one was explicitly cited as the basis for the priority research questions identified within the UK Marine Science Strategy. Our aim has been to identify key questions which, if addressed through focused research and enquiry, might not only help resolve important theoretical challenges but might also improve the mutual understanding and effectiveness of those who work at the interface of science and policy. The questions presented below were generated through a democratic, transparent and collaborative process similar to those used in previous exercises. There are interesting differences in this case, however, because the existence of a pre- determined research and policy community is much less evident. Participants were therefore selected to cover a wide range of academic disciplines (including the biological, environmental, medical, physical, and social sciences) as well as governmental and non-governmental organisations, consultancies and industry. Initially, each participant was invited to produce a list of questions, consulting widely if they wished to do so (see the section below). The 239 questions submitted at this first stage are presented in the. A process of voting, deliberation and further voting (the final stages of which took place at a meeting of participants over two days) subsequently reduced the initial list to a final set of 40 questions. During this process the questions were also redrafted and grouped thematically. They are presented in the following section, ordered by theme but not in rank order. The outcomes of an exercise such as this are inevitably influenced by the composition of the set of participants, as well as by the process. Clearly, therefore, the results are not ‘reproducible’ (in the sense that a re-run with different people could be expected to produce exactly the same set of questions). Nevertheless, if the exercise were to involve a similarly large and diverse group of participants, and were to be conducted, like this one, through several rounds of voting, deliberation and editing, we consider it highly likely that broadly similar general themes would emerge. This is, of course, an empirically testable proposition. # Results ## Understanding the role of scientific evidence in policymaking 1. How do different political cultures and institutions affect the acquisition and treatment of scientific evidence in policy formulation, implementation and evaluation? 2. How do scientists and policy makers recognise and convey the limitations of scientific advice? 3. At what stages during the development of policy does scientific evidence have the greatest impact on the decisions made? 4. Under what conditions does scientific evidence legitimise political decisions? 5. What roles have science and other forms of expertise played in international governance regimes, such as the World Trade Organisation? 6. Are there conditions under which scientific evidence may help resolve value-laden conflict and if so, what are those conditions? 7. What factors affect the utility and legitimacy of formal decision support, assessment and evaluation tools, and their adoption (or otherwise) by policy makers? 8. What influences the form and application of monitoring and evaluation practices in the development of policy informed by science? ## Framing questions, sourcing evidence and advice, shaping research 1. How do policy makers decide which questions they should ask their expert advisors and when in the policy cycle they should be asked? 2. What are the most effective mechanisms for identifying the evidence required to inform policy-making on new and emerging problems? 3. How, and with what consequences, have the sources of scientific evidence and advice used by policy makers changed over recent decades? 4. In what ways do different political cultures shape the frameworks through which evidence and advice are sourced? 5. In what circumstances are policy problems likely to require the inclusion of experts with conflicting views? 6. When is it considered appropriate to consult experts with conflicting views, and what mechanisms can ensure that this takes place? 7. What factors influence whether different disciplines are included effectively when defining and addressing complex policy problems? 8. What are the mechanisms by which budgetary pressures and societal constraints on policy-making influence the prioritisation and funding of research? 9. What is the effectiveness of different techniques for anticipating future policy issues requiring science input? ## Advisory systems and networks 1. How are national science advisory systems constructed and to what extent do different systems result in different outcomes? 2. How and why does the role of scientific advice in policy-making differ among local, regional, national and international levels of governance? 3. Which commissioning and operational arrangements lead to the most effective use of science in policy-making? 4. Policy makers typically use networks of experts, formal and informal. How does the structure and composition of such networks influence the outcomes of decision making? 5. How do different ways of using and organising in-house scientific expertise affect the quality and use of scientific evidence and advice in policy-making? 6. What are the consequences of different approaches to institutionalising, professionalising and building capacity in the exchange of knowledge between science and policy? 7. How can the effectiveness of knowledge-brokering be assessed? ## Policy making under conditions of uncertainty and disagreement 1. How is agreement reached on what counts as sufficient evidence to inform particular policy decisions? 2. How is scientific evidence incorporated into representations of, and decision-making about, so-called “wicked” problems, which lack clear definition and cannot be solved definitively? 3. Can distinctions be made in scientific advice between facts and values; to the extent that this is possible, how effective are policy makers in distinguishing them and what factors influence their effectiveness? 4. How can risks, and the associated uncertainties, complexities, ambiguities and ignorance, be effectively characterised and communicated? 5. How do policy makers understand and respond to scientific uncertainties and expert disagreements? 6. Do different approaches to building consensus, or illuminating lack of consensus, result in different consequences for policy and, if so, why? ## Democratic governance of scientific advice 1. What factors (for example, openness, accountability, credibility) influence the degree to which the public accept as trustworthy an expert providing advice? 2. What governance processes and enabling conditions are needed to ensure that policymaking is scientifically credible, while addressing a perceived societal preference for policy processes that are more democratic than technocratic? 3. How might the attitudes and values of diverse publics relating to science and technology, and their governance, be incorporated effectively into debates about the use of evidence in policy-making? 4. What has been the influence of scrutinising institutions, such as those of legislative bodies (e.g. Parliament, Congress, National Assembly or Bundestag) on the roles of science in policy-making? 5. What are the implications for their effectiveness of opening up expert advisory processes to different forms of transparency? 6. What are the implications for science-policy relations, and for the democratisation of science, of novel methods of engagement and dissemination (such as citizen science, and new media technologies, including social media)? ## How do scientists and policy makers understand expert advisory processes? 1. What factors shape the ways in which scientific advisors and policy makers make sense of their own and each other's roles in the policy process? 2. How and why have the conceptual models of science-policy relations held by policy makers, scientists and other stakeholders changed over time, and with what consequences? 3. How is guidance on the handling and communication of risk, uncertainty and ambiguity interpreted by policy makers, and what impact do their views have on the uptake and implementation of recommendations? 4. What impact has research on the relationship between science and policy actually had on science policy? # Discussion Although it may seem self-evident that policy should be informed by scientific understanding, and should therefore be evidence-based, this normative assumption is itself based on surprisingly weak evidence. Debates continue, for example, about what exactly constitutes good evidence, where and how such evidence should be sought, and at what stage in the policy process different forms of evidence might be more or less appropriate. That such debate persists reflects the fact that there are many open questions about the nature of science-policy interactions, as this exercise has revealed. In short therefore, we need to ask not just how science can best inform policy, but also how policy and political processes affect what counts as authoritative evidence in the first place. Jasanoff's seminal study of science advisers showed that the value of science in policy stemmed in part from its capacity for detailed engagement with practical policy problems. At the same time, the authority of science was seen to depend on maintaining its independence from politics through separation, in what has been referred to as ‘boundary-work’. Rhetorical commitments in the policy world to a clear distinction between facts and values were ever-present. Since then, however, experience in many different contexts, both national and issue-based, has brought about a much greater awareness of the processes of interconnection among science, politics, policy-making and publics. As Bijker *et al.* note, an appreciation of the limits of science as an impartial arbiter among policy options comes at exactly the moment when demands for scientific input to policy are increasing. This tension is reflected and articulated in many of the questions generated by the interdisciplinary exercise reported here. The six broad themes around which the questions have been organized constitute a potential framework for formulating research priorities, if we seek to develop better understandings of how science-policy interactions occur, and of evidence- based policy in practice. Beginning with a set of questions that consider the formal role that science might be expected to play in policy-making, we move on to two sets of more empirical questions about the ways in which science is selected and evaluated within the policy process, and how advisory processes actually work as an established system of governance; both sets of questions bear on the issues of expertise and authority. The following two themes then consider some of the limits to scientific knowledge, specifically in relation to inherent uncertainty and pervasive interdisciplinarity, and the roles of democratic participation and accountability in science-policy interactions. Taken together, these first five themes suggest a maturing appreciation of complexity and mutual interdependence in these relations; of the value and ubiquity of science in contemporary policy making; of the limits of ‘speaking truth to power’; and of the considerable effort that goes into the routine tasks of managing science policy. Perhaps most interestingly, the final theme opens up a series of questions about how reflection on, and better understanding of, the nature of science-policy relations might help to improving the ways in which scientific evidence and advice is commissioned, constructed and transmitted when developing forms of evidence-based policy. The exercise reported here may therefore be seen as a contribution to developing a broad research agenda for investigating this critical, complex and contested relationship, perhaps in ways that could enhance its capacity to bring the best available knowledge effectively to bear on twenty-first century problems. # Materials and Methods The methods used in this exercise are similar to those described in Sutherland *et al*. based on the experience of a series of attempts to identify priority questions,. The 52 participants were selected to cover a wide range of approaches to science and policy across government, non-governmental organisations, academia and industry. All participants are authors of this paper; the address list indicates their affiliations. Each participant was permitted to consult widely among their own colleagues in obtaining an initial list of questions. We asked participants how many people they had actively consulted (for example, in workshops, meetings or email discussions, but not including those who were sent details and did not respond). From the responses we know at least an additional 83 beyond the participants were involved in devising questions. In total, 239 questions were submitted. These questions were collated into twelve themes. They were then sent to all participants, who were asked to select around fifty that they considered to be the most important. 29 voted. 11 questions obtained no votes. Participants were also invited to suggest alternative wording. The final screening took place at a two-day workshop held in Cambridge in April 2011. On the first evening the process was discussed and potential misunderstandings and problems resolved. Prior to the meeting all participants had been provided with the number of votes for each question and any suggested rephrasing. On the following day, the workshop was divided into three 105 minute sessions, each with four groups meeting in parallel – twelve discussion groups in total, one discussing each question theme. Each group was charged with reducing one of the twelve-question themes to three priority questions plus a ranked list of three reserves. A rapporteur (from outside the team) was assigned to each session to incorporate changes to questions and capture the shortlist of the emergent top six; participants observed the editing process (projected onto a screen) as it was being carried out. Each group had a different, pre-allocated chair (three of whom had previous experience of chairing sessions in similar exercises). A guidance note for chairs suggested that early decisions could be made to drop questions with zero or very few votes from the initial voting round; and also that groups of questions that clearly addressed similar issues could be identified. This process was designed to assist the group in identifying priorities, removing redundancy, and rewording questions to eliminate overlap and improve clarity. The group then voted on the remaining questions in order to select those considered the most important. Chairs also needed to maintain structure and direction in what were invariably vigorous and challenging deliberations. In a final plenary session chaired by WJS, the top 36 questions (three from each of the twelve groups) were presented as a printed list to each participant to identify overlaps, problem questions and potential clarifications. Editing was again projected onto a screen and so was visible to all. When disagreement could not be resolved by discussion, decisions about inclusion or exclusion of questions, and about specific wording, were made by majority voting. Seven questions were removed by this process. The 12 top-ranking second-level questions were examined and the top 6 of these selected by voting (each participant having 6 votes). They were then discussed further to resolve any overlaps. The next 12 secondary questions were examined along with the remaining top ranked questions and the final five questions selected with each participant having five votes. Selected questions were then clustered into 6 categories by placing related questions together, and edited by the entire group to produce the questions set out in this paper. During this process, after discussion and another round of voting, one question was removed and one short-listed question was added. As with previous exercises most questions changed considerably from initial submission to final product. Forty-three participants made comments on or edits to the 64 successive versions of the paper that were circulated to all participants. We did not obtain ethics approval for this exercise, as it was agreed from the outset that all those participating in the voting and selection of questions were to become authors of the resulting paper. However, all submitted questions were treated anonymously; and an agreement was made to publish in an open-access journal, if possible, in order to facilitate general accessibility for those in policy communities. # Supporting Information This project was organised as an activity of the Centre for Science and Policy, University of Cambridge. We thank J. Ouchikh for efficiently dealing with the administration, L. Luheshi, I. Naicker, J. Palmer and E. Rough for transcribing the discussions, Miranda Gomperts for constructive help and three referees for useful and constructive comments. [^1]: Conceived and designed the experiments: WJS MB RD KSR SO CPT. Performed the experiments: WJS LB JRB JJB RMB MB VMC DDC AC ASC DRC AAD CD LDA SD RD NRD RJE WYF HCJG PH SEH AJH JH AH MH CI RCJ GSK PL TMM GM EPM WJN SO MMP SP JP RP ASP BP GR KSR JGR LS LS BGS DJS JS ACS CPT DEW RLZ. Wrote the paper: WJS LB JRB JJB RMB MB VMC DDC AC ASC DRC AAD CD LDA SD RD NRD RJE WYF HCJG PH SEH AJH JH AH MH CI RCJ GSK PL TMM GM EPM WJN SO MMP SP JP RP ASP BP GR KSR JGR LS LS BGS DJS JS ACS CPT DEW RLZ. [^2]: The authors have the following competing interest: Nicholas R. Dusic is employed by Pfizer Ltd. and Louise Shaxson by the Delta Partnership. There are no patents, products in development or marketed products to declare. This does not alter the authors' adherence to all the PLoS ONE policies on sharing data and materials.

# Introduction The auditory steady state response (ASSR) is an oscillatory brain response, which is phase locked to the rhythm of an auditory stimulus. ASSRs can be elicited by using repetition click, amplitude-modulated (AM) and frequency- modulated tones. ASSRs have been recorded in response to a wide frequency range of modulation and/or repetition, but the physiological features of the ASSRs are somewhat different depending on the modulation frequency. Comparison of ASSRs generated by stimuli modulated with a frequency greater than 70–80 Hz (80-Hz ASSR) and ASSRs generated by stimuli modulated with a frequency around 40 Hz shows that the amplitude of the 40-Hz ASSR is significantly greater than that of the 80-Hz ASSR in the awake condition. However, the 40-Hz ASSR is very sensitive to the arousal state, and is greatly decreased under the conditions of sleep or general anesthesia, whereas the 80-Hz ASSR is not so affected by the arousal state,. Moreover, the effects of contralateral noise on ASSRs are known to depend on the modulation frequencies. Contralateral white noise at a level which does not cause a significant psychophysical threshold elevation results in remarkable suppression of the 40-Hz ASSR, but no significant effects were observed in the 80-Hz ASSR. Since the same level of contralateral noise also does not cause any significant effects on the auditory brainstem response and N1 cortical response, these suppressive effects seem to be a characteristic property of the 40-Hz ASSRs. However, the effects of contralateral noise on the ASSRs elicited by modulation frequencies other than 40 Hz and 80 Hz are not known. Recently, in addition to 40-Hz and 80-Hz ASSRs, the 20-Hz ASSR has been investigated in the diagnosis of several psychiatric disorders, as different physiological properties have been observed for the 20-Hz and 40-Hz ASSRs in patients with bipolar disorder and/or schizophrenia. The 40-Hz ASSRs were significantly reduced, whereas the 20-Hz ASSRs showed no significant reduction in these patients. These findings appear to indicate that the 40-Hz and 20-Hz ASSRs have different properties. The major source of 20-Hz ASSRs is thought to be the auditory cortex, as for 40-Hz ASSRs, but the amplitude of 20-Hz ASSRs is usually slightly smaller than that of 40-Hz ASSRs,. However, other properties of 20-Hz ASSRs have not yet been clarified. The present study compared the effects of contralateral noise on the 20-Hz ASSR and the 40-Hz ASSR in normal subjects to gain a better understanding of the physiological properties of the 20-Hz ASSR. # Materials and Methods ## Subjects This study included 9 normal volunteers, 8 males and 1 female aged 31.2+/−3.42 years (mean +/− standard deviation), without histories of auditory diseases or neurological disorders. Audiometry revealed that all subjects had hearing level thresholds of 20 dB or better for frequencies from 0.125–8 kHz. All subjects were right-handed with scores above +90 on the Edinburgh Handedness Inventory. Written informed consent in accordance with ethical committee of Tohoku University Graduate School of Medicine and the Declaration of Helsinki (1991) was obtained from each subject. The present study was approved by the ethical committee of the Tohoku University Graduate School of Medicine. All parts of the present study were performed in accordance with the guidelines of the Declaration of Helsinki. ## Stimuli The test stimulus to record the ASSRs consisted of 1000-Hz long tone bursts (duration 5 s, rise-fall time 1 ms) with 100% AM at 39 Hz and 20 Hz with an exponential modulation envelope (resulting in the 40-Hz and 20-Hz ASSRs, respectively), which was produced using a digital signal processing platform (TDT System III, Tucker-Davis Technologies, Gainesville, FL) under the control of an IBM PC/AT computer. The inter-stimulus interval was 3 s. The sound pressure level of the AM tone was 80 dB SPL and was presented monaurally. Continuous noise (white noise) at a level of 70 dB SPL was applied to the ear not receiving the amplitude-modulated tone bursts, which is the same noise condition (type and presented level) used in our previous study to observe the contra-noise effects on the auditory cortical responses (40-Hz ASSR and N100m). The test stimuli (AM tone and tone bursts) and noise were presented to the subject through tube earphones (ER-3A, Etymotic Research, Elk Grove Village, IL). ## Recording and analysis Magnetoencephalography (MEG) recording of auditory evoked fields used a 200-channel whole-head type axial gradiometer system (MEG vision PQA160C, Yokogawa Electric, Musashino, Tokyo, Japan) in a magnetically shielded room. The sensors consisted of first-order axial gradiometers with a baseline of 50 mm, with each coil of the gradiometers measuring 15.5 mm in diameter. The sensors were arranged in a uniform array over a helmet-shaped surface at the bottom of the dewar vessel. The centers of two adjacent coils were separated by a mean distance of 25 mm. The field sensitivity of the sensors (system noise) was 3 fT/Hz within the frequency range used in the study. Auditory evoked fields were recorded only in the awake state as confirmed by real-time MEG monitoring of the occipital alpha rhythm. The MEG signal was band-pass filtered between 0.16 Hz and 100 Hz, and sampled at 1000 Hz. Coils were attached at 5 locations on the head surface. They acted as fiduciary points with respect to the landmarks (nasion and preauricular points) and the position of the head within the helmet by passing currents through the coils and measuring the magnetic fields. In addition to these fiduciary markers, the head shape of each participant was digitized using a three-dimensional digitizer (FastSCAN Cobra, Polhemus Inc., Colchester, VT) and co-registered with individual structural magnetic resonance (MR) images acquired using a 3T MR system (Achieva, Philips, Best, the Netherlands). The responses to stimuli without and with contralateral noise were recorded alternately at least twice. MEG signals were recorded for at most 600 s during the presentation of AM tone bursts with/without contralateral noise, and later analyzed (offline) using the built-in software in the MEG system (MEG Laboratory, Yokogawa Electric) to obtain the ASSRs. Data epochs of 1 s in duration, starting at the onset of the trigger signal synchronized with a certain phase of the amplitude modulation, were extracted from the serial recorded data after filtering with a digital band-pass filter (35–45 Hz for 40-Hz ASSR, 17–23 Hz for 20-Hz ASSR) and averaged in the time domain. Epochs with signals exceeding 3 pT were rejected on a single-channel basis, so that about 3000–4500 epochs (mean  = 3580.2, standard deviation  = 289.1) were usually used in the averaging process. The location of the signal source was estimated for the amplitude maxima of the responses for each hemisphere using an equivalent current dipole (ECD) model with the best fit sphere for each subject's head. We used a single ECD model based on Sarvas law in a spherical volume conductor for identifying the sources of the magnetic signals. Phase-lags were usually present between the amplitude maxima of the bilateral hemispheres, so the location of the signal source was separately analyzed in the right and left hemispheres. Dipole location and orientation of the ECDs were calculated for the amplitude maxima of the response using the data from channels from each hemisphere. ECDs with a goodness-of-fit value of 90% were accepted and the source was superimposed on the three-dimensional MR image of the individual subject using a MEG-MR image coordination integration system and the measured responses were verified to originate from the auditory cortex. In the present study, the effects of contralateral noise on the power of ASSRs were analyzed by focusing on the channels of the maximum signals measured over each hemisphere. The power of the ASSR measured at these channels was quantified by fast Fourier transform spectrum analysis using the built-in software in the MEG system (applied window function: Hamming window, length of FFT: 1 sec, spectral resolution: 0.488 Hz). The measured power of the ASSR obtained under the same sound conditions of signal and contralateral noise were averaged, and the effects of contralateral noise on the ASSRs were analyzed separately for the right and left hemispheres. Subjects were instructed to stay awake during recording (subjects usually watched silent movies during the measurements to prevent the need for artificial conversation with the subjects if they appeared to be sleepy) because the cortical ASSR tends to be significantly reduced during sleep. # Results Both 20-Hz and 40-Hz ASSRs were observed bilaterally in all subjects under all stimulus conditions, using all combinations of right or left ear stimulation with or without contralateral noise. and show examples of the effects of contralateral noise on the 20-Hz and 40-Hz ASSR waveforms, respectively, obtained from one subject (Case 3) mapped onto a flattened projection of the sensor position. and show the responses to the stimulus and noise presented to the left ear, and and show the responses to the stimulus and noise presented to the right ear., illustrate the largest amplitude waveforms in each hemisphere. Both 20-Hz and 40-Hz ASSRs were suppressed by contralateral noise in the bilateral hemispheres, but the magnitude of suppression appeared to be larger in the right hemisphere than in the left hemisphere. The effects of contralateral noise on the powers of ASSRs were analyzed focusing on the channels of the maximum signals measured over each hemisphere, as reported before. show the averaged effects of contralateral noise on the power of the ASSRs in the channels of the maximum responses measured over each hemisphere for all measurement conditions. Data were statistically analyzed using three-way analysis of variance (ANOVA) with the factors of stimulation side (right/left ear), measured hemisphere (right/left), and presence of contralateral noise (off/on). Significant suppression of power was caused by contralateral noise for both 20-Hz and 40-Hz ASSRs, detected as significant main effects of hemisphere (p\<0.01) and noise (p\<0.001), as well as interaction between stimulation side and hemisphere (p\<0.001), but the main effect of stimulation side was not significant. The ratios of ASSR power between the conditions with/without contralateral noise, calculated based on the data shown in, are plotted in. Significant main effects of modulation frequency (p\<0.05) and interaction between stimulation ear and hemisphere (p\<0.05) were observed by three-way ANOVA with the factors of stimulation side (right/left ear), measured hemisphere (right/left), and modulation frequency (20 Hz/40 Hz). The magnitudes of contralateral noise suppression were significantly larger in 40-Hz ASSRs than in 20-Hz ASSRs. Moreover, the suppression was greatest in the right hemisphere with right ear stimulation. # Discussion In the present study, significant suppression of ASSRs caused by contralateral noise was observed in both the 20-Hz and 40-Hz ASSRs, although the magnitude of suppression was significantly smaller in the 20-Hz ASSR than in the 40-Hz ASSR. Moreover, the greatest suppression of the 20-Hz and 40-Hz ASSRs tended to occur in the right hemisphere during presentation of the stimulus to the right ear. This finding that the 20-Hz ASSR can be remarkably suppressed by contralateral noise is important to recognize. ASSRs are thought to be generated throughout the central auditory system. However, the location of the highest activity depends on the modulation frequency,. In general, when the ASSR is measured by conventional electroencephalography (EEG), the ASSR generated by stimuli modulated with a frequency around 80 Hz (80-Hz ASSR) contains more components from the brainstem, whereas the ASSR generated by stimuli modulated with a frequency lower than 40–50 Hz contains more components from the upper auditory pathway. On the other hand, the source reflected in the recorded ASSR is affected by the recording methods as well. Comparison of 20-Hz and 40-Hz ASSRs detected by EEG found that most components of the 20-Hz ASSRs originated from the auditory cortices, whereas the 40-Hz ASSRs may contain components from lower brain locations than the cortex such as the medial geniculate body in addition to components originating from the auditory cortices. However, when the ASSR is measured by MEG as in the present study, almost all of the recorded ASSRs were assumed to consist of cortical components for both 20-Hz and 40-Hs ASSRs. The effects of contralateral noise on the ASSRs have so far been examined for the 20-Hz ASSRs by MEG (present study), 40-Hz ASSRs by EEG , and MEG, and 80-Hz ASSRs by EEG. Contralateral suppression was only observed in the 20-Hz ASSR (present study) and the 40-Hz ASSR, but not in the 80-Hz ASSR with EEG. Considering that the 80-Hz ASSR recorded with EEG is thought to mainly consist of signals originating from the brainstem, we suppose that contralateral noise might not so greatly affect the ASSR originating from the brainstem. On the other hand, contralateral noise suppression may be a phenomenon only observed in ASSRs originating from the auditory cortex. However, contralateral noise does not necessarily suppress other types of auditory cortical response such as N1, so this phenomenon may be quite specific to cortical ASSRs, and is presumably related to the auditory processing of the modulation. Identification of the origins of 20-Hz and 40-Hz ASSRs with MEG has found no substantial difference between the locations of the cortical sources of the dipole moments of 20-Hz and 40-Hz ASSRs. However, a recent in vitro study of brain oscillation suggested that the origin of gamma-band oscillation might be different to that of beta-band oscillation, as the fast rhythmic bursting neurons in layer II/III and the pyramidal cells in layer V are closely related to the generation of the gamma-band and beta-band oscillations, respectively. These different sources for the gamma-band and beta-band oscillations may be one of the factors causing the observed differences between the 20-Hz and 40-Hz ASSRs measured with MEG, such as selective reduction of 40-Hz ASSRs in patients with bipolar disorder and/or schizophrenia. Despite the probable differences in the physiological or pathophysiological properties between the 20-Hz and 40-Hz ASSRs, the temporal information regarding the modulation is likely to be analyzed in some common system, regardless of the modulation frequency. If so, the common features of ASSRs not depending on the modulation frequency, such as contralateral suppression of both 20-Hz and 40-Hz ASSRs, may reflect the general physiological properties of the common processing system for modulation in the cortex. Recently, the clinical importance of the 20-Hz ASSR has been emphasized, especially in psychiatry clinics, but the detailed physiological properties of 20-Hz ASSR are not so well known as those of the 40-Hz and 80-Hz ASSRs. The present study newly showed that 20-Hz ASSRs are suppressed by contralateral noise, which may be important both for characterization of the 20-Hz ASSR and for interpretation in clinical situations. Physicians must be aware that the 20-Hz ASSR is significantly suppressed by sound (e.g. masking noise or binaural stimulation) applied to the contralateral ear. For example, masking noise is often applied to the contralateral ear to avoid cross talk effects in auditory examinations. The effects of this contralateral masking noise are likely to be negligible in measurements of the auditory brainstem response, N1 cortical response, and 80-Hz ASSR. However, contralateral masking noise could significantly suppress the 20-Hz ASSR, as for the 40-Hz ASSR. We wish to thank Ms. Yuki Yamada for her continuous support of this work. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: TK AK NN RK. Performed the experiments: HU TK AK IY HM. Analyzed the data: HU TK HM YK. Contributed reagents/materials/analysis tools: RK NN. Wrote the paper: HU TK AK YK.

# Introduction Visual awareness of the physical properties of a to-be-touched or to-be-grasped target object does not limit effective interactions with that object. The most striking demonstration of this phenomenon is exemplified in individuals with action-blindsight: a deficit arising from lesions to the primary visual cortex (V1),. Persons with action-blindsight report being “unaware” of visual stimuli within their impaired hemifield; however, such individuals demonstrate preserved saccades, visually guided pointing and tracking within their scotoma. A more subtle demonstration of action without awareness is also observed following lesions to the lateral occipitotemporal cortex (LOC). In particular, the extensive studies of DF demonstrate that although impaired in identifying object shape and orientation (i.e., visual agnosia), she is readily able to scale her actions to the metrical properties of a target. Not surprisingly, the clinical neuropsychology literature has spawned interest in whether or not “action without awareness” can be observed independent of a *chronic* visual processing deficit. For example, Ro's work reports that transient V1 disruption via single-pulse transcranial magnetic stimulation impacts the perceptual identification of a remote distractor but does not diminish the extent to which the same distractor facilitates movement planning (i.e., the redundant target effect). Further, extensive work using a double-step paradigm demonstrates that automatic and online limb adjustments arising from a change in target location can occur in the absence of visual awareness. As well, work by our group – has shown that visual awareness of an intrinsic target property (i.e., size) is not necessary for appropriate size-scaling of reach trajectories. Taken together, the results described just above indicate that visual awareness is not a precursor to motor output and that action-blindsight is not a restrictive clinical deficit; rather, the phenomenon reflects a general visuomotor characteristic. The neural basis for the separation between conscious visual awareness and motor output is provided by Goodale and Milner's perception/action model (PAM). The PAM states that V1 or extrageniculate projections to the posterior parietal cortex of the dorsal visual pathway mediate motor output whereas projections from V1 to the inferotemporal cortex of the ventral visual pathway mediate conscious visual judgments. Thus, visuomotor processes are retained in the face of clinical or experimental disruption to early visual processing areas (i.e., V1) because extrageniculate projections to dorsal visuomotor networks can proxy for V1 inputs. Additionally, visuomotor processes are not influenced by disruption to the ventral visual pathway because dorsal visuomotor networks are not dependent on a conscious (i.e., top-down derived) visual percept. An interesting question related to dorsal visuomotor function is the timeframe by which unconscious and metrical information can be retained and used to support motor output (so-called memory-guided action). A strong view of the PAM asserts that dorsal visuomotor networks operate along an evanescent time frame; that is, unconscious and metrical information is available only on a moment-to- moment basis (real time processing). As such, introducing even the briefest of delay (i.e., 0 ms) between the occlusion of a visual stimulus and the onset of a response is proposed to nullify metrical reaching and grasping. Support for this view is garnered from reports that visually guided - but not memory-guided - responses are refractory to the context-dependent properties of pictorial illusions. According to the real time position of the PAM, such results demonstrate that in the absence of continuous visual contact with a target object, memory-guided responses are mediated by a context-dependent visual percept laid down and maintained by the ventral visual pathway. However, the degree to which pictorial illusions provide a systematic and reliable means to address the timeframe of dorsal visual processing is questioned by accumulating evidence that illusions “trick” both visually and memory-guided responses. An alternative to real time processing holds that visuomotor networks maintain a spatially enriched and temporally durable representation. Consistent with this assertion, recent work by our group has employed a variant of Di Lollo et al's four-dot object-substitution masking paradigm to demonstrate visuomotor memory in the absence of a conscious record. In our previous work, participants were asked to verbally report the size of a target circle or reach (with their right hand) to that same target circle under conditions wherein the target was primed (i.e., no-mask trials) or perceptually masked (i.e., mask trials). Importantly, reaches were cued concurrent with presentation of the target circle (i.e., immediate cuing) or following a visual delay of 1000 or 2000 ms. Consistent with previous implementations of this masking procedure, perceptual reports of target size were correctly identified during no-mask (mean accuracy of verbal report = 88%, d' = 1.66) but not mask trials (mean accuracy of verbal report = 54%, d' = 0.17). Most importantly however, for the reaching task, movement times and other representative kinematic measures scaled to veridical target size independent of mask condition (i.e., no-mask and mask) and across the immediate and delay (1000 and 2000 ms) conditions. Put another way, the absence of visual awareness did not preclude the veridical scaling of reach trajectories for up to 2000 ms of visual delay. A question arising from our previous work relates to how (and where) unconscious visual information is used to support the sensory-to-motor transformations underlying metrical memory-guided reaching performance. One scenario holds that a movement plan related to target size is pre-computed at stimulus presentation (via parieto-frontal networks) and stored in memory to support later motor output. An alternative scenario holds that visual target information is retained by dorsal visuomotor networks and subsequently accessed for conversion into a motor plan at - and not before - response cuing. Thus, the present investigation sought to determine if visual information for which we are not aware is immediately used to specify the kinematic parameters of a memory-guided response or whether such information is retained in sensory form and used to support response specification at the time of movement cuing. In order to test these accounts, we again employed the four-dot masking paradigm to manipulate participant's perceptual awareness of target size and included the maximal delay condition (i.e., 2000 ms) used previously. Importantly however, the hand performing the response was not specified until response cuing; unimanual left and right hand responses were randomly interleaved on a trial-by-trial basis (i.e., two distinct auditory imperative tones designated left and right limb performance). We reasoned that specifying the limb only at the time of response cuing would preclude advanced sensory-to-motor transformations. Thus, if a pre- computed motor plan supports memory-guided reaches to a perceptually masked target, then precluding limb specification during the 2000 ms delay condition should nullify the metrical scaling of reach trajectories. In other words, precluding the specification of response parameters at initial stimulus presentation should render reach trajectories that are refractory to the size differences of a perceptually masked target. In contrast, if dorsal visuomotor networks retain a visual target representation, then such information should be flexibly able to support delayed motor output. # Methods ## Participants Thirty individuals from the University of Western Ontario community participated in this experiment. Participants were right handed and had normal or corrected- to-normal vision. Participants gave written informed consent for a protocol approved by the University of Western Ontario's Office of Research Ethics (Review \#14041S), and this work was conducted in accord with the 1964 Declaration of Helsinki. ## Apparatus and Procedure Participants sat at a custom-built three-shelved aiming apparatus for the duration of this experiment. The top shelf supported an inverted computer monitor (30-inch Monitor, 14 ms response time; 60 Hz: Dell 3007WFP: Round Rock, TX, USA), the middle shelf was composed of a one-way mirror (96 cm wide by 65 cm deep), and the bottom shelf was composed of a solid surface (96 cm wide by 65 cm deep). The distance between each shelf was 34 cm; thus, the optical geometry of our setup created a situation in which stimuli projected onto the mirror appeared to participants as being located on the bottom shelf (i.e., the reaching surface) of the aiming apparatus. In addition, head position was restrained via a head-chin rest (ASL-6000: Bedford, MA, USA). The reaching surface contained a home position defined by a haptic cue (5×3×3 cm magnet) located at participant's midline and 10 cm from the front edge of the apparatus. Because the one-way mirror occluded direct limb vision, dual light emitting diodes (LEDs) were placed at the home position to allow for the pre-movement visual calibration of limb position. Computer events and all visual and auditory stimuli were controlled via MatLab (7.6: The MathWorks, Natick, MA, USA) and the PsychToolBox (ver. 3.0). Participants were presented with a central fixation cross (1 cm by 1 cm) for a randomized foreperiod (1000–2000 ms) after which an array of differently sized circles (1, 2, 3, and 4 cm) was presented for 13 ms (an array contained 5 randomly placed circles of each size). Within the array, one circle served as the “target” and was identified by four small red dots (i.e., the four-dot mask) arranged in an imaginary square (16 cm²) that surrounded but did not touch the target. Notably, the size of the four-dot mask was constant across all experimental trials. In the no-mask condition, the circles array and four-dot mask were extinguished simultaneously (i.e., after the 13 ms presentation) and replaced with a blank screen. In the mask condition, the four-dot mask remained visible for 320 ms after occlusion of the circles array (see for timeline of experimental events). Target circles were located 26.5 cm anterior to the home position and 8.5 cm left (i.e., left space) and right (i.e., right space) of participant's midline. All participants completed perceptual and reaching trials in separate and counterbalanced trial blocks. Half of the participants (n = 15) were cued to complete their perceptual and reaching trials in time with onset of the circles array (0 s delay; i.e., the D0 group). The remaining participants completed their perceptual and reaching trials 2000 ms following onset of the circles array (i.e., the D2000 group). ### Perceptual Task To avoid confusion with the naming of intermediate-sized circles, only the 2 and 4 cm circles were used as targets. These circles were identified in advance of the perceptual block and were labeled as “small” and “large” respectively. For each trial, participants were prompted (via auditory tone) to report (forced- choice binary decision) whether the target was small or large. The prompt was provided immediately (i.e., the D0 group) or 2000 ms (i.e., the D2000 group) after onset of the circles array. No-mask and mask trials were completed in separate and randomly ordered blocks. Within no-mask and mask blocks, small and large targets were randomly presented in left and right space on five occasions for a total of 20 perceptual trials. ### Reaching Task Participants were instructed to place their left and right index fingers on the home position in advance of each trial. At this position the fingers were spaced by approximately 2 cm. The goal of the reaching task was to point to the cued target as “quickly and accurately” as possible in response to an auditory tone. Because a trial could involve the performance of the left or the right hand, 300 Hz and 950 Hz, 13 ms tones were used to identify left and right hand performance respectively. Seven familiarization trials for each hand-tone combination were provided in advance of the reaching task. For the D0 group, the initiation tone was provided concurrent with onset of the circles array (see panel 2 of). For the D2000 group, the initiation tone was provided 2000 ms after onset of the circles array (following panel 4 of). Target sizes included 1, 2, 3 and 4 cm (i.e., each circle size within the array was used as a target) and produced respective index of difficulty (ID) values of 3.7, 4.1, 4.7 and 5.7 bits \[log₂(2A/W\] (where A = amplitude and W = target width). In line with the perceptual task, targets circles were located 26.5 cm anterior to the home position and 8.5 cm left and right of midline (i.e., resultant movement vector of 278 mm). As such, participants reached to targets in ipsilateral (e.g., left hand-left target) or contralateral (e.g., left hand-right target) space. No-mask and mask trials were completed in separate blocks and within each block hand (left vs. right hand), target location (left space vs. right space) and target ID (3.7, 4.1, 4.7 and 5.7 bits) were randomly ordered with five trials completed to each combination for a total of 160 reaching trials. Infrared emitting diodes (IRED) were attached to the nail of the left and right index fingers. IRED position data were sampled for 1.5 s at 200 Hz via an OPTOTRAK Certus (Northern Digitial Inc., Waterloo, ON, Canada). Offline, IRED position data were filtered at 10 Hz via a second-order dual-pass Butterworth filter. Instantaneous velocities and accelerations were computed via a three- point central finite difference algorithm. Movement onset was defined as the first frame that exceeded 50 mm/s for ten consecutive frames (50 ms) and movement offset was the first frame falling below 50 mm/s for ten consecutive frames. ## Dependent Variables and Statistical Analysis For the perceptual task, the percentage of correct responses in no-mask and mask trials was examined via 2 (group: D0, D2000) by 2 (stimulus presentation: no- mask, mask) mixed-design ANOVA. For the reaching task, we computed reaction time (RT), movement time (MT), times to resultant peak acceleration (tPA), velocity (tPV) and deceleration (tPD) and resultant error (RE). Dependent variables for the reaching task were examined via 2 (group: D0, D2000) by 2 (stimulus presentation: non-mask, mask) 2 (hand: left hand, right hand) by 4 (target ID: 3.7, 4.1, 4.7 and 5.7 bits) mixed-design ANOVA. Only significant effects are reported and we report regression equations and R² values as a means for interpreting significant effects/interactions. Means and between-participant standard deviations are reported in parentheses. Note that to streamline our results section, and in line with earlier work, we did not include target location (i.e., left versus right space) in our ANOVA model. It is, however, important to note that RTs and MTs for left and right hand responses in ipsilateral space were faster than contralateral space, F(1,28) = 5.16, and 283.07 respectively for RT and MT, ps\<0.05, and ipsilateral responses were always more accurate than contralateral ones, F(1,28) = 87.28, p\<0.001. Importantly, target location did not interact with stimulus presentation or group (Fs\<1.0). # Results ## Perceptual Task Target size was judged more accurately during no-mask (85% SD11, d' = 1.80 SD0.94) than mask trials (58% SD12, d' = 0.31 SD0.47), F(1,28) = 124.55, p\<0.001. It is also noteworthy to mention that group and stimulus presentation did not interact (F\<1.01). ## Reaching Task Examination of RT yielded an interaction involving group and stimulus presentation, F(1,28) = 4.49, p\<0.05. D0 group RTs during no-mask trials (438 ms SD89) were slower than mask trials (408 ms SD100) (t(14) = 2.62, p\<0.03); however, D2000 group RTs were comparable across no-mask (412 ms SD81) and mask (409 ms SD69) trials (t(14) = 0.44, p\>0.05). For MT, left hand reaches (429 ms SD80) were slower than right hand ones (404 ms SD76), F(1,28) = 41.04, p\<0.001, and MT scaled in relation to increasing target ID, F(3,84) = 7.42, p\<0.001. As shown in (see also), regression equations and R² values relating movement time to target ID across D0 and D2000 group mask and no-mask trials indicate a reliable and robust increase in movement time as a function of increasing target ID. Moreover, examination of indicates that MT elicited null effects for stimulus presentation by target ID, as well as group by stimulus presentation by target ID (Fs\<1.3). The time to achieve peak acceleration, velocity and deceleration for the left hand (tPA = 124 ms SD68, tPV = 290 ms SD59, tPD = 376 ms SD71) was longer than the right hand (tPA = 103 ms SD87, tPV = 268 ms SD67, tPD = 352 ms SD69), Fs(1,28) = 6.79, 23.90, and 32.14 respectively for tPA, tPV and tPD, ps\<0.02. In addition, each measure increased with increasing target ID, Fs(3,84) = 5.30, 7.97, and 5.65 respectively for tPA, tPV and tPD, ps\<0.03. As demonstrated in, regression equations and R² values indicate that the impact of target ID on tPA, tPV and tPD was such that increasing target ID resulted in an increase in the time to achieve each kinematic marker. Last, analysis of RE indicated that mask trials (1.5 mm SD28.7) were more accurate than no-mask trials (8.8 mm SD20.6), F(1,28) = 6.10, p\<0.03. # Discussion The goal of this investigation was to determine how visual information that is unavailable to conscious verbal report is used to support the scaling of memory- guided reaching. In particular, we sought to determine if unconscious and metrical information related to an intrinsic target property (i.e., size) is used to pre-compute the parameters of a memory-guided response or whether such information is maintained as a sensory (i.e., visual) representation for on- demand sensorimotor conversion at response cuing. To that end, participants completed immediate (i.e., D0) or memory-based (i.e., D2000) verbal reports and reaching (left and right hand) responses to perceptually masked targets using a variant of Di Lollo et al's four-dot masking paradigm. Importantly, a critical response parameter associated with the reaching task (i.e., the limb performing the movement) was specified only at the time of response cuing thereby limiting the pre-computing of an advanced motor plan. ## Re-entrant processing and the perceptual masking of target size In line with previous work, verbal reports during no-mask trials achieved a robust level of accuracy whereas mask trials operated at chance. According to Di Lollo et al's computational model of object substitution, the simultaneous offset of target and non-target items during non-mask trials allows for uniform decay of visual features and permits a stable visual percept to be laid down and accessed by high-level visuo-perceptual networks (i.e., the ventral visual pathway). In contrast, the asymmetric offset of target and non-target items during mask trials elicits non-uniformity of decay; that is, re-entrant processing of non-target features at low-level visual processing areas (i.e., V1) conflicts with a “visible persistence” of target features maintained by high-level visual processing areas. As such, re-entrant processing renders the original percept (i.e., target and non-target features) unavailable for conscious report. It is also worthy to note that in our study the D0 and D2000 groups showed equivalent performance across no-mask and mask trials. In particular, the equivalent findings for no-mask trials across the two groups used here indicates that when consciously perceived, the visuo-perceptual networks of the ventral visual pathway maintain a temporally durable representation of target size. ## Four-dot masking and the scaling of reaching trajectories Before addressing the principal issue of how perceptual masking and motor uncertainty impact the size-scaling of memory-guided reaching, we outline the general impact of our limb manipulation. First, specifying the limb at response cuing (i.e., left or right hand) resulted in longer reaction times (Grand Mean = 416 ms) than a similar experiment employing only right hand reaches (Grand Mean = 234 ms). Of course, the between-experiment difference represents an expected result owing to the increased stimulus response alternatives used here. Moreover, the longer planning times evidence that motor parameters were not pre-computed at the time of stimulus presentation; rather, the reaction times shown here indicate that selection of the limb performing the response in combination with specifying the movement parameters for that limb occurred in time – and not before - response cuing. Second, the use of left and right hand responses yielded an expected asymmetry in response execution such that the right hand elicited faster movement times and achieved representative kinematic markers sooner than left hand counterparts. We did not find that reaction time was sensitive to target ID and this null effect generalized across mask and no-mask trials for D0 and D2000 groups. In other words, results provide no evidence that reaction time scaled in relation to target size. We did, however, observe that D0 group no-mask trials exhibited slower reaction times than mask trial counterparts whereas D2000 group reaction times did not vary across no-mask and mask trials. Recall that D0 group no-mask trials involved the simultaneous blanking of the visual array and onset of the auditory imperative tone whereas in the other experimental conditions the imperative tone was provided in time with persistence of the four-dot mask (i.e., D0 group mask trials) or after all elements of the visual array were extinguished (i.e., D2000 group mask and no-mask trials) (see for timeline of experimental events). It is therefore possible that the double stimulus cue provided during D0 group no-mask trials delayed movement planning processes. Although reaction time did not scale to target ID, shows that movement times for D0 and D2000 groups increased as a function of increasing target ID for both mask and no-mask trials. also demonstrates equivalent slopes relating movement time to target ID across the different experimental conditions. Moreover, the times to achieve peak acceleration, peak velocity and peak deceleration for D0 and D2000 group no-mask and mask trials demonstrated a scaling effect with target ID. Thus, results from our experiment demonstrate that across all conditions lawful speed-accuracy trade-offs related to target size emerged during the response evocation stage of reaching. As noted by a myriad of studies, this effect is taken to reflect the need to devote longer movement durations to ensure that a response “hits” the desired target location ;. It is also worth mentioning that our study did not provide participants with online limb vision: a manipulation quite different from Fitts' original work, wherein participants were afforded continuous limb vision. Indeed, the fact that we observed speed-accuracy relations on par to that reported by Fitts indicates that speed-accuracy relations are not entirely determined by feedback-based limb corrections. Rather, our results are in line with accumulating evidence that speed-accuracy relations are in part determined by central planning mechanisms. The combined results of the perceptual and reaching task match previous work by our group and indicate that awareness of target size is not necessary to support the metrical scaling of immediate or memory-guided reaches involving up to 2000 ms of delay. Moreover, the current investigation adds importantly to the extant literature in demonstrating that unconscious and metrical information supporting memory-guided reaches reflects a sensory (i.e., visual) representation maintained by dorsal visuomotor networks. The basis for this assertion is predicated on the fact that our limb manipulation – and introduction of premovement motor uncertainty - prevented participants from pre-computing the kinematic parameters of their reach trajectories in advance of response cuing. In particular, the limb associated with any given trial for the D2000 group was specified well after extinction of the target object. Thus, the ability of the D2000 group to scale their reach trajectories to veridical target size mandated that a sensory representation be maintained in memory until the time of response cuing. In general, the present results support the PAM's assertion that dorsal visuomotor networks operate independent of an obligatory visual percept. However, the present results are inconsistent with the PAM's contention that dorsal visuomotor networks operate along an evanescent timeframe (i.e., real- time control). As mentioned in the Introduction, the real time nature of dorsal visuomotor function is supported by some work involving memory-guided reaching/grasping of pictorial illusions – and the studies of patient DF (i.e., visual agnosia) demonstrating a breakdown in her ability to scale reach and grasp trajectories following a memory delay. In a complementary manner, there exists some data involving an individual (i.e., GY) with action-blindsight to report null scaling between grip aperture and target size when a delay is introduced between target presentation and the onset of a movement within the impaired hemifield. It is, however, important to note that Weiskrantz et al's classic study of DB demonstrates preserved visuomotor function in the presence of a visual delay. In particular, Weiskrantz et al. presented a static visual target for a 2000 ms preview and the extinction of the target served as the experimenter's cue to verbally prompt DB to initiate his reaching response. Of course, the time required for the experimenter to perceive offset of the visual target and the time for the experimenter to produce the verbal imperative in combination with the time required for DB to plan and initiate his response would have introduced an appreciable period of visual delay (\>1000 ms). Thus, and although we are unable to offer specific insight into the nature of the discrepant literature provided above, we believe that findings from a clinical patient as well as the present and other work by our group – provides convergent evidence that unconscious and metrical visual information is retained as a sensory based representation and is available to support visuomotor processes for up to 2000 ms of visual delay. Indeed, future work is set to provide a systematic probe of the impact of increasing memory delays (i.e., immediate reaching, 0, 500, 1000, 1500, and 2000 ms of delay) on movement scaling in persons with documented action-blindsight and matched controls. The goal of this future work is to ascertain whether the persistence of unconscious and metrical information in the aforementioned groups is susceptible to differing decay properties. A final issue requiring redress relates to the impact of our experimental manipulations on endpoint accuracy. Similar to previous work, target ID did not influence the accuracy of reaching responses. That finding in combination with the temporal measures described above indicates that emergent speed-accuracy relations were defined by the timing, and not the spatial, properties of the movement goal. It was also observed that mask trials were more accurate than no- mask trials. In line with our previous work – we attribute such a finding to the improved ocular gaze anchoring and spatial landmarking afforded by the four-dot mask. More specifically, the four-dot mask provided additional spatial information allowing for more effective target localization. ## Conclusions Here we demonstrate that the scaling of memory-guided reaching movements to target size is not dependent on an obligatory visual percept. Moreover, by precluding the specification of a movement parameter during the delay interval used here, we establish that a persistent sensory (i.e., visual) representation supports the unconscious and metrical scaling of memory-based actions. Such findings indicate that the visuomotor networks of the dorsal visual pathway retain a spatially enriched and temporally durable sensory-based representation that is distinct from that subserving perception based activities. We thank Digby Elliott for his constructive comments on an earlier version of this manuscript. [^1]: Conceived and designed the experiments: MH GB. Performed the experiments: AM. Analyzed the data: AM. Contributed reagents/materials/analysis tools: MH BG GB. Wrote the paper: MH. [^2]: The authors have declared that no competing interests exist.

# Introduction The human immunodeficiency virus (HIV) can cross the blood-brain barrier early in the course of infection and trigger a cascade of functional and structural alterations. One of the primary loci of this damage is the deep grey matter of the basal ganglia which are reciprocally connected to a broad range of cortical regions. HIV infection has been reliably linked to injury of the fronto-striato- thalamo-cortical loops, and can thus adversely impact a variety of higher-order neurocognitive functions that rely on these circuits. For example, HIV- associated neurocognitive deficits have been observed in approximately half of infected persons in such ability areas as fine-motor skills, working memory, and executive function. Risk-taking and reward processing are important processes that influence behavior. Making a choice in a risky situation typically requires a choice between an option that is associated with a large outcome that may be either advantageous or disadvantageous versus an alternative with a smaller, more certain advantageous outcome. One formulation of risk is its econometric definition as the variance of the value of the possible outcomes. This conception, however, does not account for the influences of emotion since human decision-making is not always rational. More recent approaches attempt to bridge the gap between rational choice and emotions guiding decisions. For example, the somatic marker hypothesis, affect heuristic hypothesis and risk-as- feelings hypothesis posit that emotions are integrated with cognitive evaluation of choices to regulate the decision making process. Risky decision-making is commonly observed in individuals at risk for and infected with HIV. For example, on the Iowa Gambling Task (IGT), HIV-infected (HIV+) individuals made disproportionately more selections from “bad” decks as the task progresses, which may reflect poor inhibition response to the lure of high rewards even in the face of large penalties. Such risky decision-making is more common in individuals with HIV-associated Neurocognitive Disorders (HAND;) and has been specifically, but not consistently, linked to cognitive inflexibility and engagement in HIV transmission risk behaviors. Additionally, personality characteristics such as sensation seeking, that is a propensity to seek out novel, exciting and arousing stimuli, have been associated with risk behaviors for HIV infection. However, few studies have examined the neural substrates of risky decision-making in HIV. A number of neuroimaging studies have shown that HIV infection can alter brain function. In functional magnetic resonance imaging (fMRI) experiments, HIV+ individuals have been shown to exhibit deficits on tests of attention and working memory and altered responses within the associated neural substrates. In these tasks, HIV+ individuals exhibited greater activation and/or larger load dependent increases in the frontal and parietal cortical regions underlying task performance, , a pattern that is consistent with damage to fronto-striatal white matter tracts that has been attributed to HIV infection. Preliminary evidence suggests that hippocampal function can be affected by HIV infection. Reduced hippocampal activation during encoding and elevated hippocampal activation during the recall phase of a verbal memory task having been reported in a sample of HIV+ women. There has been a recent interest in the use of magnetoencephalography (MEG) in the study of HIV infection in part due to its high test-retest reliability in HIV infection. These studies have reported strong responses (in the 8–13 Hz band) in the dorsolateral prefrontal cortex (DLPFC) during a simple finger tapping task. Reduced levels of synchronization (in the 6–12 Hz band), recorded during a simple visual processing task, have been also been observed in a HIV+ compared to a control group. Furthermore, those HIV+ individual with desynchronization more similar to controls demonstrated better performance on a neuropsychological test of verbal learning. Notably, cortical thinning in the DLPFC has been linked to the severity of immune suppression in HIV infection. Finally, reduced functional connectivity has been reported in resting state MEG and in a fMRI task studying semantic event sequencing. Together, neuroimaging studies showing preferential involvement of fronto-basal ganglia brain systems in HIV-infection, and neuropsychological studies of the maladaptive use of feedback in risky choice, lay the foundation for fMRI studies to more directly link impaired behavior with abnormalities of the brain substrates of risky decision-making and reward in HIV infection. Extensive study of risk-taking in the context of decision-making has revealed a key network of cortical and subcortical brain regions. This network is composed of circuits, two of which, namely the ventral limbic circuit and the dorsal executive circuit, may be important to choice behavior. The ventral circuit includes several striatal structures: the nucleus accumbens, rostromedial caudate, rostroventral putamen, and ventromedial caudate. These regions receive extensive innervation from prefrontal cortical regions such as the orbitofrontal and ventromedial cortex, and insula. This circuit is implicated in the identification of rewarding and emotionally salient stimuli, integrating these with autonomic, visceral, and hedonic information, and generating affective responses to these stimuli. The dorsal executive circuit encompasses the dorsal caudate, dorsal anterior cingulate cortex (DACC), and DLPFC. In this circuit DACC is thought to play a role in performance monitoring – whereas DLPFC is thought to be important to the maintenance of goal-directed behavior. Indeed, in the context of risky-choice behavior, modulation of activity within the DLPFC can lead to different response styles under risk. The anterior cingulate cortex (ACC) and DLPFC are extensively interconnected with the DACC also projecting to the dorsal caudate. This circuit is important to selective attention, planning and effortful regulation of affective states, including task switching and inhibition. Acting in concert, these two circuits may code stimulus-reward value, maintain representations of predicted future reward and future behavioral choice, and transform decisions into motor output, playing a role in integrating and evaluating reward prediction to guide decisions. This study aimed to examine whether HIV alters brain processes underlying risk- taking decision-making. We hypothesized that greater activity for risky choices would be observed in grey matter basal ganglia regions. Additionally, we theorized that dorsolateral and anterior cingulate cortex would display greater activity in the HIV+ group relative to HIV- comparisons, and that this may be an adaptive functional response to compensate for aberrant information provided by other cortical and subcortical structures that may have been impacted by HIV infection. Finally, we explored whether nadir cluster of differentiation 4 (CD4) count, a measure of historical immune function that has been shown to predict neurocognitive impairment – and structural volumes in HIV+ individuals, would demonstrate any association with functional brain measures of risky decision- making. # Materials and Methods ## Ethics Statement The University of California, San Diego human research protection program approved this study. Participants gave informed written consent and were compensated for their time and effort. ## Participants Participants were recruited as part of the Translational Methamphetamine AIDS Research Center and included 21 HIV+ and 19 seronegative comparison adults. HIV status was confirmed by MedMira Multiplo rapid test (MedMira Inc., Nova Scotia, Canada). All participants were seronegative for Hepatitis C virus (HCV) as determined by the MedMira Multiplo rapid test. Current CD4 T lymphocyte counts (cells/ml) were determined by flow cytometry at a medical center laboratory certified by Clinical Laboratory Improvement Amendments (CLIA), or CLIA equivalent. HIV RNA levels were measured in plasma by reverse transcriptase PCR (Roche Amplicor, v. 1.5, lower limit of quantitation 50 copies/ml). CD4 nadir was obtained by self-report, with confirmation by documented prior measurements in a subset of individuals. Participants were excluded if they tested positive for illicit drugs (with the exception of marijuana (MJ)) or alcohol (urine toxicology screen or Breathalyzer respectively) on the day of scan; had contra- indications for MRI; had a lifetime history of schizophrenia or other primary psychotic disorders; had previous cerebrovascular events, determined by comprehensive neurological exam; head injury with loss of consciousness for greater than 30 minutes or resulting in neurologic complications; seizure disorder; demyelinating diseases from non-HIV neurological disorders; met Diagnostic and Statistical Manual of Mental Disorders (4^th Edition- Text Revision; DSM-IV-TR) conditions for substance (other than alcohol, MJ and nicotine) abuse in the prior year or dependence within the preceding five years. Participants who met criteria for lifetime dependence or abuse of MJ within the last 12 months were enrolled. Those who met lifetime criteria for alcohol abuse within the prior 12 months were enrolled but were excluded if they met criteria for dependence within the previous 12 months. Nicotine use was not exclusionary and participants were told not to alter their typical pattern of daily usage. They were asked to refrain from smoking during the break in the scanning session. Breath carbon monoxide levels and the presence of cotinine in urine were assessed on the day of the scan. ## Participant Assessment Assessment took place on two separate visits: an initial neurological and neuropsychiatric visit and a subsequent MRI scan (see for the interval between visits). Determination of relevant psychiatric diagnoses were achieved using the Composite International Diagnostic Interview (CIDI 2.0), a computer based structured interview administered by a trained research associate on the initial visit. This evaluation tool yields lifetime and current (within 1 month) diagnoses that are consistent with DSM-IV-TR criteria. The resultant measures of mood and substance use disorders were used to inform eligibility for study enrollment and characterizing the sample. Additionally, participants were asked their age, gender, handedness, and sexual orientation. Since HIV has been associated with cognitive impairment, all participants completed the oral word reading subtest of the Wide Range Achievement Test (WRAT-4) on the initial visit and a comprehensive neuropsychological test battery (described in detail), performance on which was summarized using the Global Deficit Score (GDS). Specifically, we assessed the following domains: speeded information processing, verbal fluency, learning, working memory, executive functions, and motor skills (more details can be found). As neuropsychiatric symptoms are common in many HIV+ individuals, we also administered the Beck Depression Inventory-II (BDI-II) and Profile of Mood State (POMS) on the day of the scan. BDI-II was also measured on the initial visit. Deficits in the IGT suggest that HIV+ individuals tend to choose larger immediate rewards over smaller rewards that result in longer term gains overall. This suggests an impulsive response style in HIV+ individuals. Consequently, we measured this using the Barratt Impulsiveness Scale-11 (BIS-11). Furthermore, since risky decision-making in HIV+ individuals may be related to sensation seeking and could represent a common pathophysiology, we chose to measure it using the Kalichman Sexual Sensation Seeking Scale (KS4) which assesses personality characteristics and high-risk sexual behavior known to be associated with HIV transmission risk. Estimates of smoking rates among HIV-infected individuals are in the range of 50–70%, with smokers being less likely to adhere to antiretroviral medication regimes. We therefore assessed nicotine usage using the Fagerström Test for Nicotine Dependence (FTND). ## Task Participants were administered the risky-gains task as previously described – and depicted in. The task consists of 102 trials during which the numbers 20, 40, and 80 are presented in ascending order for one second each. Our choice of numbers, where each one is twice the preceding choice, was motivated by the observation that people typically reject gambles unless the amount that could be gained is at least twice the amount that could be lost. The serial nature of the task, which required participants to make sequential fast judgments as to whether to accept or reject the amount displayed on the screen, was designed to capture the escalating tension that often accompanies naturalistic risky decision-making. If the participant made a button press within one second of stimulus onset, that amount (20/40/80) could be added to their total winnings along with immediate visual and auditory feedback. When a 40 or 80 appear there is a chance that it appears in a different color with immediate feedback indicating the loss of 40- or 80-cents, respectively. When this happens the trial ends immediately and the participant may not make any more responses. Each of the 102 trials lasted 3.5s regardless of the participants’ choices and whether punishment was scheduled or not. Participants were told that waiting for 40- or 80-cents was risky as, though it was possible to win more, it was also possible to lose that amount. Additionally, they were informed that though they would win less if they chose 20-cents, there was no risk of loss associated with this choice. That is, they could always win by choosing a 20-cents. However, participants were not told that there was no inherent advantage to choosing the risky (40, 80) choices over the safe choices (20). While the best response for each trial depended on whether there was a punishment scheduled for that trail or not, a strategy of selecting all 20 s would yield exactly the same winnings as selecting all 40 s or all 80 s. Three different trial types were presented in a predetermined pseudo-randomized order: (1) non-punished (20, 40, or 80; n = 54), (2) punished 40 (n = 24), (3) punished 80 (n = 18) and six null trials. The relative number of punished and non-punished trials was chosen to guarantee that the strategy of consistently choosing 20-cents or always selecting 40- or 80-cents would yield the exact same winnings. Punishment occurred on a punished trial only if the participant failed to respond to the previous numbers on that trial (i.e. when holding out for 80 they did not respond to either the 20 or 40 stimulus). The relative frequency of safe (20) to risky (40 and 80) was used to quantify baseline risk-taking behavior. To investigate the sensitivity to punishment, defined here as the propensity to alter choice pattern on a trial immediately following a losing trial, the relative frequency of risky responses was examined as a function of the outcome of the previous trial, that is, punished versus non-punished risky trials. This task has previously been used to assess risk-taking behavior in healthy volunteers. This investigation revealed activation differences in insula, DLPFC, and posterior parietal cortex. Greater activation was observed in the insula during risky than during safe choices. The insula also showed significant activation on punished trials with activation magnitude predicting subsequent safe choices after a punished trial. In a purely behavioral study of stimulant users, we observed greater propensity for risk taking in the stimulant users but a similar degree of sensitivity to punishment in both the users and a control group. Those with higher measures of sensation seeking and impulsivity showed greater propensity to risk taking. Finally, Lee et al. examined the effect of aging using the risky gains task. They reported greater levels of risk taking in younger people and more safe choices in older adults with faster reaction times to risky choices irrespective of group. They also observed greater insula and DLPFC activation for risky as compared to safe choices in the older adult group. ## MR Data Acquisition Functional images were acquired in bottom-up interleaved axial slices using T2\* weighted echo planar imaging (EPI). Images were acquired on two scanners: a 3T GE Discovery MR 750 (Milwaukee, WI) (252 volumes TR/TE = 2 s/30 ms, flip angle = 90°, 64×64 matrix, 40 axial slices, 3.75×3.75×3.0 mm voxels) and a 3T GE Signa HDx (Milwaukee, WI) (252 volumes, TR/TE = 2 s/30 ms, flip angle = 90°, 64×64 matrix, 40 3.0 mm (2.6 mm +0.3 mm gap) axial slices, 3.5×3.5 mm voxels). High-resolution T1-weighted fast spoiled gradient echo anatomical images (MR 750: TR/TE = 8.1 ms/3.17 ms, flip angle = 8°, 256×256 matrix, 172 sagittal slices, 1×1×1 mm voxels; Signa HDx: TR/TE = 7.77 ms/2.97 ms, flip angle = 8°, 256×256 matrix, 172 sagittal slices, 0.97×0.97×1 mm voxels) were acquired to permit subsequent activation localization and spatial normalization. Gradient echo field-maps were also acquired to permit compensation for geometric distortions caused by magnetic field inhomogeneity (MR 750: TR = 1 s, TE = 3.7/5.5 ms, flip angle = 60°, 64×64 matrix, 160 axial slices, 3.75×3.75×3 mm voxels; Signa HDx: TR = 1 s, TE = 3.5/5.5 ms, flip angle = 60°, 64×64 matrix, 160 3.0 mm (2.6 mm +0.3 mm gap) axial slices, 3.5×3.5 mm voxels). Stimuli were projected onto a screen at the participants’ feet and viewed with the aid of a mirror attached to the head coil. ## MR Data Analysis Analyses were conducted using AFNI and FSL. T1-weighted images were skull- stripped using 3dSkullStrip and transformed to MNI152 standard space using FLIRT, followed by nonlinear registration using FNIRT. Functional images were processed using FUGUE to compensate for B₀ inhomogeneity. Time series were motion corrected using an iterated linearized weighted least squares algorithm and aligned to the anatomical using a local Pearson correlation method before being subjected to global mean-based intensity normalization. These were resampled to 3 mm isotropic voxels and transformed to standard space using the warp fields derived from transforming the T1 anatomy to MNI152 space. Finally, the time series were spatially blurred with a 4.2 mm isotropic FWHM Gaussian filter kernel within a mask derived from the T1 anatomy. Data were visually inspected to assess the quality of the warping and alignment. Preprocessed time series were subjected to multiple linear regression. Time series of interest, derived from the behavioral data (described below), were convolved with a gamma variate function and subsequently normalized to a peak amplitude of 1. Decision phase regressors were created such that they started at trial onset and ended either when the participant responded or was punished. Five regressors were defined: (1) 20 (safe), (2) win 40, (3) win 80, (4) lose 40, and (5) lose 80. General linear tests (GLT) of the safe (20) versus the risky win (40, 80) choices were computed for each participant. The baseline comprised all other time-points not accounted for by the regressors of interest. Additionally, six nuisance motion-related regressors (three translational and three rotational) and a 3^rd order Lagrange polynomial, which accounted for slow signal drift, were included in the baseline. Brain activation was operationally defined as percent signal change relative to baseline. ### Task Related Group Analyses The within-participant general linear tests of safe versus risky choices were subjected to linear mixed effects (LME) analyses in the R statistical analysis package. As some participants had a lifetime diagnosis of major depressive disorder (MDD), a within-participant dichotomous variable indicating the presence/absence of MDD was also included in the model. Significant voxels were required to pass a voxel-wise statistical threshold (F_{(1, 37)} = 4.11, p = 0.05, uncorrected) and, to control for multiple comparisons, were required to be part of a cluster of no less than 1685 µL. The volume threshold was determined by a Monte-Carlo simulation that together with the voxel-wise threshold resulted in a 5% probability of a cluster surviving due to chance. The average percent signal change was extracted from the clusters so formed and a series of post hoc t-tests were conducted in R to examine the group and task interaction effects. Since the HIV+ group consisted of more nicotine users than the HIV- group (see below), we conducted an additional analysis to investigate whether there was an effect of nicotine usage (FTND total score) on differential brain activation to risky versus safe choices. This was performed within the HIV+ group and accomplished using linear mixed effects models on the average contrast of risky versus safe choices in each of the clusters identified by the task related group analysis. In this analysis participant was treated as a random effect. ### Controlling for Scanner Effects Several recent studies have examined the effect of including MR data from multiple sites within the same analysis. Overall, these studies reported that inter-participant variance was anywhere from 7–44 times greater than that generated by site variance, even when group membership was confounded by site. This suggests that scanner-induced variance is less likely to contribute to task or group-related effects. The inclusion of data from two scanners in our study effectively makes it a multi-site study and though the inclusion of site as a fixed-effect in the model examining group differences has been recommended, our study data was acquired in such a way that each participant was scanned only once on one scanner. It was therefore not possible to separately estimate the effects due to scanner and participant. Since scanner and participant are confounded in our study, we opted to include a dichotomous variable for scanner as the random effect in the linear mixed effects model described above. ### Neuropsychiatric and Neuromedical Measures and Task-Related Brain Activations Whole-brain voxel-wise Huber robust regressions, were conducted in R to examine the relationships between the neuromedical and neuropsychiatric measures and GLTs of safe versus risky win trials in the HIV+ group. We opted to perform regressions as some of the neuropsychiatric variables were confounded by diagnosis, and the neuromedical variable existed solely within the HIV+ group, precluding inclusion in the LMEs. Regressions were performed for the non-sexual sensation seeking and sexual compulsivity subscales of the KS4 based on our observation of between group differences on these measures (see below). A further regression examining the effect of nadir CD4 count was also conducted. Dichotomous variables for lifetime MDD diagnosis and scanner (for reasons outlined above) were included in these models. In the case of the nadir CD4 count, age and estimated duration of infection were also included in the model, as older participants or those with longer durations of HIV infection might have had lower nadirs and accumulated more damage due to HIV infection that may manifest in altered brain functioning. In all cases, regression coefficients and their corresponding t-values were split according to whether they demonstrated a positive or negative relationship with the GLTs. Thereafter, significant voxels were required to pass a voxel-wise statistical threshold (t(18) = 2.10, p = 0.05, uncorrected) and, to control for multiple comparisons, were required to be part of a cluster of no less than 1685 µL which resulted in a 5% chance of a cluster surviving due to chance. The volume threshold was determined in the same manner as above. ### Overlap between Group and Regression Regions of Interest To investigate whether between-group differences could potentially be attributed to neuropsychiatric and neuromedical variables, we assessed whether the regions identified from the task related analysis overlapped with those regions identified by the robust regression analysis. To accomplish this we computed the intersection of all those regions from the task related analysis with those from the regression analyses conducted solely within the HIV+ group. Since both of the maps included in this analysis included significantly different clusters, the resultant overlap maps can, therefore, also be regarded as statistically significantly different. ## Demographic and Clinical Scales Analysis All analyses were conducted in R. Between-group differences for demographics and clinical scales were assessed by means of Welch T tests for age, years of education, WRAT-4, BDI-II, POMS, KS4, BIS-11, FTND, speeded information processing, verbal fluency, learning, working memory, executive functions, and motor skills. Effect sizes were computed using Hedge’s g. A linear mixed effects analysis (where participant was treated as a random effect) was conducted to investigate whether there was an effect on the BDI-II score of group, visit (baseline, scanning day), and their interaction. Group differences in GDS and days between the initial visit and scanning were assessed using Wilcox rank sum test. Effect sizes for these two measures were computed using the probability of superiority. Group differences in gender, ethnicity, handedness, sexual orientation, the number of participants per group, the number of participants per scanner, and the number of individuals with a positive urine toxicology test for MJ were assessed using χ² test of equal proportions. Using Spearman’s rank correlation test, we tested for the presence of a relationship between BDI-II score and the mean percentage signal change from each of the clusters resulting from the task-based whole brain analysis. This correlation was performed both solely within participants with lifetime diagnosis of MDD and within the sample as a whole. ## Task Analysis The behavioral data gathered during task performance was subjected to a 3-way ANOVA in R. In this model the effect of response choice (2 levels: safe (20) and risky (40, 80)), group (2 levels: HIV- and HIV+), punishment (2 levels: non- punished and punished trial), and their interactions were examined. This permitted assessment of the effects of choice behavior and susceptibility to prior punishment and how these varied by group. # Results ## Demographics There were no significant differences between the serostatus groups in age (t(37.99) = −0.74, p\>0.1), handedness (*χ*²(1) = 0.18, p\>0.1), gender (*χ*²(1) = 1.00, p\>0.1), years of education (t(37.97) = −0.85, p\>0.1), or ethnicity (*χ*²(3) = 3.85, p\>0.1). The groups differed in terms of sexual orientation, with the HIV+ group composed primarily of men who have sex with men, whereas the HIV- group was predominantly heterosexual (*χ*²(1) = 13.81, p\<0.001). Three participants tested positive for MJ (2 HIV+, 1 HIV-), a proportion that did not differ between groups (*χ*²(1)≈0, p≈1). None of these participants met DSM-IV criteria for current or lifetime substance abuse or dependence. No participants tested positive for alcohol. ## Clinical Scales There was no difference between the groups on GDS (W = 151.50, p\>0.05). Three HIV- and seven HIV+ were classed as impaired (GDS \>0.5); the proportions did not differ between the two groups (*χ*²(1) = 0.84, p\>0.1). There were no between-group differences observed in any of the speeded information processing, verbal fluency, learning, working memory, executive functions, and motor skills domains (all p≥0.1). There was a trend for the HIV- group to exhibit a higher WRAT-4 score than the HIV+ group (t(35.81) = 1.85, p = 0.07, g = 0.60). On the subscales of the KS4, the groups did not differ on the sexual sensation seeking subscale (t(34.04) = −1.79, p\>0.05, g = −0.60) but the HIV- positives scored higher on sexual compulsivity (t(26.01) = −2.33, p\<0.05, g = −0.70) and lower on non-sexual sensation seeking (t(36.32) = 2.02, p\<0.05, g = 0.60). The groups showed no differences on the BIS-11 or any of its subscales (all p\>0.05) (See.) Three HIV- and 11 HIV+ participants had a lifetime diagnosis of MDD (*χ*²(1) = 4.37, p\<0.05). The HIV+ group was comprised of marginally more nicotine dependent individuals than the HIV- group (*χ*²(1) = 3.22, p = 0.07). There were no significant between- group differences on nicotine usage measured by the FTND (t(38) = −0.22, p\>0.1). There were no significant differences between the two groups on other substance use characteristics. Significant between-group differences were observed on the BDI-II scale with the HIV+ group endorsing significantly greater levels of depression than the HIV- group both at the initial visit (t(22.53) = −4.07, p\<0.001, g = −1.3) and at the time of scanning (t(25.71) = −3.16, p\<0.01, g = −1.00). When determining the stability of BDI-II score over time, the HIV+ group had significantly higher BDI-II scores than the HIV- group (F_{(1, 38)} = 14.06, p\<0.001), there was no effect of visit (F_{(1, 38)} = 1.81, p = 0.18), or an interaction between group and visit (F_{(1, 38)} = 1.79, p = 0.18). This suggests that the BDI-II score was stable over time. Finally, no significant between-group differences on the POMS were observed (t(29.02) = −1.75, p\>0.05). ## Behavioral Task Results A significant main effect of risk was evident (F_{(1, 214)} = 337.92, p\<0.0001) with safe (M = 0.65) responses more likely than risky (M = 0.16). No significant effects of punishment (F_{(1, 214)} = 0.05, p\>0.05, punished: M = 0.33, non-punished: M = 0.33) or group (F_{(1, 214)} = 0.21, p\>0.05; HIV-: M = 0.33, HIV+: M = 0.32) were observed. No significant interactions of punishment × group (F_(1,214) = 0.06, p\>0.05) or risk × punishment × group (F_(1,215) = 0.00, p\>0.05) were observed. A marginally significant effect for risk × group (F_{(1, 214)} = 3.09, p = 0.08) was observed with the HIV- group selecting risky choices marginally more than the HIV+ group (HIV-: risky M = 0.19, safe M = 0.62; HIV+: risky M = 0.14, safe = 0.68). Consistent with our prior observations on this task, a significant interaction of risk × punishment (F_(1,214) = 11.42, p\<0.001) was observed wherein participants were more likely to chose the safe option when the prior trial was punished. See for a complete list of the cell and marginal means. ## fMRI Task Results ### Task Related Brain Activation We identified eight regions where the HIV- and HIV+ groups differed ( and). Cortical regions were located in the right anterior cingulate gyrus, inferior parietal lobule, superior frontal gyrus, and bilaterally in the middle frontal gyri. Two subcortical regions, one in each hemisphere with centers of mass in the left lentiform nucleus and right claustrum were identified. The left cluster extended dorsally from the ventral striatum to include portions of the head and body of the caudate and further extended to include parts of the putamen and anterior insula. The right cluster predominantly included portions of the head of the caudate and extended laterally to include the anterior insula. An additional subcortical cluster in the left thalamus was identified. Post hoc analyses were conducted to identify the directionality of these effects. Within the HIV+ group, activation was greater for risky relative to safe choices in the right ACC (t(31.31) = 4.56, p\<0.001), left middle frontal gyrus (t(28.43) = 4.97, p\<0.001), right middle frontal gyrus (t(27.86) = 6.39, p\<0.001), left thalamus t(36.07) = 3.67, p\<0.001), right claustrum (t(26.66) = 4.58, p\<0001), and right superior frontal gyrus (t(28.76) = 3.21, p\<0.001. Within the HIV- group, there were no significant differences between risky and safe choices (all p\>0.1). For risky choices, the HIV+ group displayed greater activation than the HIV- group in left thalamus (t(70.83) = 2.66, p\<0.01), left lentiform nucleus (t(68.71) = 3.41, p\<0.01), and right claustrum (t(57.64) = 2.37, p\<0.05). For safe choices, the HIV+ group displayed less activation than the HIV- group in the right ACC (t(32.45) = −4.1, p\<0.001), left middle frontal gyrus (t(30.58) = −2.75, p\<0.01), and right middle frontal gyrus (t(36.48) = −3.23, p\<0.01). Within the participants with a lifetime diagnosis of MDD, none of these clusters showed a relationship with the BDI-II scores (all p\>0.05). Across the sample as a whole, none of the clusters showed a relationship with the BDI-II scores (all p\>0.05). Within the HIV+ group, there was no relationship between nicotine usage (FTND total score) and the contrast of risky versus safe choices in any of the aforementioned brain regions (all p\>0.1). ### Neuromedical and Neuropsychiatric Measures and Task-Related Brain Activations Significant associations, detailed in, between differential brain responses to risky versus safe choice were identified for nadir CD4 and the Kalichman sexual compulsivity subscale. Those individuals with a higher CD4 nadirs exhibited lower activation in several regions including anterior cingulate gyrus, bilateral inferior parietal lobules, and middle frontal gyrus. Moreover, those individuals with higher ratings on the sexual compulsivity subscale of the KS4 showed lower activation in cingulate gyrus, and medial and middle frontal gyri. One region in the right pyramis was negatively associated with the non-sexual sensation seeking subscale of the KS4. ### Overlap between Group and Regression Regions of Interest Several brain regions identified as showing task-related between-group differences overlapped with the regions identified in the regression analysis results just described. Regions that demonstrated a negative association with nadir CD4 overlapped with the between-group results in the right anterior cingulate, bilateral middle frontal gyri, left inferior parietal lobule, and right superior frontal gyrus. When the beta values for the risky and safe choices that contributed to this negative relationship were separated out, the difference between the two choice types appeared to be driven almost entirely by an increased response to the safe choice with greater nadir CD4 count. Regions showing a negative relationship with the sexual compulsivity subscale of the KS4 overlapped with the task related clusters in the right anterior cingulate gyrus, left thalamus, and right superior frontal gyrus. Similarly, when the risky and safe beta values were separated out, the negative relationship appeared to be largely driven by the relationship between safe beta values and the compulsivity measure. No overlap between the non-sexual sensation seeking subscale of the KS4 and the between-group differences was observed. # Discussion Risky decision-making is a common feature of HIV-associated neurocognitive disorders, but its neural substrates within persons living with HIV are poorly understood. Here, we examined risky choice behavior in HIV+ individuals compared to seronegative individuals using functional magnetic resonance imaging. We observed significant between-group functional activation differences in a number of regions (ACC, DLPFC, caudate, and insula) critical to risk and reward processing despite broadly similar task performance between the two groups. In the overlap between the task-related regions of interest and those resulting from the robust regression analysis, those HIV+ individuals with greater sexual compulsivity measured by the KS4 and higher nadir CD4 count displayed lower differential responses to safe versus risky choices in many of the regions that showed between-group task related differences. Taken together, these results support the hypothesis that HIV alters risk-related processing in the basal ganglia, among other structures. We observed significant between-group differences in left and right hemisphere clusters that included subcortical regions and small insular components. The subcortical constituents of this cluster included the ventral and dorsal striatum. These regions have been heavily implicated in reward related processing. In non-human primates, single unit recordings have revealed populations of neurons in the caudate and putamen that fire in proportion to the value of an action irrespective of whether the action was subsequently executed. In humans, fMRI studies have shown that the ventral striatum is important to judging reward value, and reduced striatal volume has been reported in HIV infection. Similarly, increased blood oxygenation level dependent (BOLD) responses have been recorded in the dorsal striatum in response to anticipation of both primary (e.g., food) and secondary (e.g., money) rewards. Additionally, BOLD responses in the dorsal striatum have been shown to predict expected value of actual choices in a risky context. This has lead some to suggest that the striatum is primarily involved in the prediction of reward value and that other brain regions (e.g., the insula) may be more important to quantifying risk. The insula is thought to be important to integrating autonomic, visceral, and hedonic information, and it has been suggested that it is a critical neural substrate for selecting between internally and externally available homeostatically relevant information that serves to guide behavior. Indeed, greater activation levels within the insula have be associated with risky choices in the task deployed here. Our results suggest that the differences in these clusters are predominantly driven by the HIV+ group who display progressively greater responses to the 20-, 40- and 80-cent choices compared to the HIV- group. Larger striatal activation for high gain/high risk trials has been observed in a study of decision making under risk. Our observation of increased activation with greater value of the potential gain may be significant insofar as it may be the functional neuroanatomical realization of over- valuation of the potential benefit of risky choices which has previously been reported in HIV+ individuals. Alternatively, it is possible that increasing activation to progressively higher valued risky choices in these subcortical clusters may be related to damage caused to the basal ganglia by HIV and may thus be an adaptive functional response to this injury. Disambiguation of these alternatives will require future studies. There were significant between-group differences in activation level in the right anterior cingulate between the HIV+ and HIV- groups. This dorsal region of the ACC is thought to be important to cognitive processes, reward-based learning and affective valence. Indeed, prior studies have indicated that the ACC is critical to judging the magnitude and likelihood of risky outcomes, and others have suggested that ACC activation may be related to encoding of action cost and action selection for uncertain rewards. It has also been suggested that the ACC may perform a cost benefit analysis to guide action selection. Activation of the ACC in a risky decision making paradigm has been associated with risk-aversive behavior, whereas deactivation was correlated with risk-seeking behavior. Finally, the DACC has been proposed to play a critical role in the detection of response conflict. In the context of decision making, tension between reward seeking and loss avoidance may naturally give rise to a state of conflict. The greater DACC activation for the risky choices in the HIV+ group may indicate that they are more sensitive to conflict between risk seeking and loss avoidance behavior or, alternatively, the cost of losing. Here, it may be the case for the HIV+ group that the possibility of losing on the risky options outweighs that benefit of winning. The opposite may be the case for the HIV- group: the cost of losing on the risky choices may not loom large and be reflected in the lower activation for the risky choices. This stands in contrast to the observations on the IGT where HIV+ individual make more selections from disadvantageous decks. However, recent evidence suggests that this effect may be more common in individuals with HAND. The small number of participants (n = 7) with HAND in the present study precluded examination of this possibility here. Therefore, further studies with larger numbers of participants with and without HAND are required to assess whether the present patterns of activation would vary by diagnosis. Finally, consistent with prior finding of this task in, a between-group activation emerged bilaterally in the middle frontal gyrus, a component of the DLPFC. DLPFC is thought to be one of the seats of higher executive brain function, and it has been suggested that DLPFC plays a key role in the maintenance of goal-directed behavior necessary for successful task performance when alerted to the presence of conflicting behavioral choices by the DACC. Several studies have examined the role played by the DLPFC during risky decision-making. In an fMRI study, activation of the DLPFC has been identified in decision making under uncertainty. Using repetitive transcranial magnetic stimulation, a technique that can transiently suppress neuronal function, it has been shown that interrupting activity in the right DLPFC can increase risk- taking behavior. Another technique, transcranial direct current stimulation, has also been employed to assess the role of DPLFC in risky decision-making. In this method, low-voltage direct current is passed through the brain using electrodes placed on the scalp. Using this technique, it has been shown that increasing excitability of the right or left DLPFC leads to risk-aversion and suggests that DLPFC may be critical to the suppression of riskier choices. DACC and DLPFC are extensively interconnected and are components of a dorsal executive circuit that is critical to performance monitoring – and maintenance of goal-directed behavior. Furthermore, this dorsal circuit interacts with the ventral circuit (consisting of insular and striatal regions) to predict stimulus-reward value and guide future behavior. The striatal regions of this circuit are known to be injured by the HIV virus, and documented cognitive deficits in HIV infection are consistent with fronto-striatal white matter damage that has been attributed to HIV infection. Our observation of increased DACC and DLPFC activation in the HIV+ group in the presence of broadly similar task performance may therefore be an adaptive functional response, wherein additional cortical resources are recruited to maintain task goals. This may be necessitated by aberrant information provided by other cortical and subcortical structures to which the dorsal circuit is connected and that may have been damaged by HIV infection. Indeed, additional functional recruitment in the presence of equivalent task performance has previously been observed in HIV+ individuals performing visual attention tasks, working memory tasks, and a simple finger tapping task. This has led some to suggest that functional brain differences, in the absence of behavioral changes, may precede clinical signs of cognitive impairment. Since, relative to the seronegative group, the HIV+ group displayed elevated sexual compulsivity – a factor that has been associated with risk of HIV infection – we investigated whether sexual compulsivity would show any relationship between differential activation to safe versus risky choices in the HIV+ group. Of those regions-of-interest (ROIs) identified as showing between- group task related differences, a subset of voxels in three of those ROIs also showed a relationship with sexual compulsivity. We separated out the risky and safe components of this relationship. The change in differential responses to risky versus safe responses appeared to be driven by an overall increase in activation to both risky and safe choices with increasing compulsivity. While there are, to our knowledge, currently no brain imaging studies that examine the relationship between functional activation and sexual compulsivity, neurobiological models of obsessive-compulsive disorder, however, have implicated excessive activity in fronto-striatal circuits, particularly in orbitofrontal, ACC, thalamus and caudate. This suggests that those HIV+ individuals with elevated sexual compulsivity may be characterized by on overall increase of activation in fronto-striatal regions. Within the HIV+ group, we also investigated whether nadir CD4 counts would show any relationship to differential activation to safe versus risky choices. We observed that subsets of the voxels identified as showing between-group task related differences also showed a relationship to nadir CD4 count. As depicted in and, greater nadir CD4 counts were negatively associated with decreased differences in response to risky versus safe choices in all of the task related ROIs. When the risky and safe components of this difference were separated out, the disparity between risky and safe responses appeared to be driven by increased activation to safe responses with greater nadir CD4 count. This suggests that differential activity to safe versus risky choices may be, in part, predicted by nadir CD4 count. Furthermore, it may be the case that those individuals with higher nadir CD4 counts may have activations patterns more similar to that of seronegative individuals than those with lower CD4 nadirs. This finding is consistent with the so-called “legacy events” hypothesis wherein historical immune-compromise increases the vulnerability of HIV-associated central nervous system injury. This study has several limitations. The data were acquired from the two groups of participants on two different scanners. Though we attempted to minimize the differences between the protocols on both scanners and to account for this source of variation in our models, future studies are required to replicate the results reported here in the absence of this potential confound. This is a cross-sectional study, and thus we cannot address whether the differences reported here arose as a consequence of HIV infection or whether they predated HIV infection. Future longitudinal studies are required to determine whether factors such as duration of infection or the use of anti-retroviral therapy may influence impairment. In light of the recent report that risky decision-making is more prevalent in individuals with HAND, future studies should examine whether the effects reported herein are driven by more impaired individuals. HIV infection has been associated with risky decision making in individuals at risk for and infected with HIV. Our observation in the behavioral analysis of a marginally significant interaction of risk and group with HIV- participants, counter-intuitively, choosing more risky options than the HIV+ group therefore warrants further investigation in a larger sample where the source of this observation may be more fully explored. Recent studies have reported differences in brain structure and function between homosexual and heterosexual men. The confounding of the serostatus groups by sexual orientation in the present sample prevented us from investigating whether HIV-infection status interacts with sexual orientation in risky decision-making. Future studies with a larger non- confounded sample are required to elucidate this issue. Depression is common in many HIV+ individuals and inclusion of such individuals arguably makes the present sample more representative of the individuals seen in clinics and thus improves the generalizability of our results. Nevertheless, future studies should be conducted in groups with equivalent levels of depression to determine the specificity of the results reported herein. Risk for HIV infection has been associated with substance use (cf. –) and, as with depression, our inclusion of participants with histories of such behaviors arguably makes our sample more representative of the HIV-infected population. Nevertheless, substance use has been independently associated with functional brain changes in many of the brain regions reported here. Investigating the specificity of the changes reported here in larger samples of HIV+ individuals with and without a history of substance use is therefore crucial. Given the preliminary evidence which suggests that anti-retroviral therapy (ART) can effect recovery of brain function to patterns typical of healthy controls, it remains unclear whether ART can influence risky choice behavior or the brain processes underlying it. Since our study was underpowered to examine this question, future studies with larger cohorts are required to examine this issue. Our sample of participants is almost exclusively male, limiting generalizability to the female population. Future studies with a larger sample of females are required to address this issue. Finally, our study concentrated on the time period prior to choosing between risky and safe options, future studies should investigate whether HIV+ individuals, compared to seronegative individuals, exhibit differences in sensitivity to the outcomes of these choices. # Conclusions In summary, the present study examined the functional neuroanatomy of risky decision-making in HIV+ individuals compared to seronegetative individuals. The HIV+ group displayed altered functional responses to safe and risky choices in several brain regions compared to the seronegetative group. Specifically, these regions included portions of the anterior cingulate, ventral and dorsal striatum, insula, and bilateral DLPFC. These results are consistent with and further support the role of these structures in risky decision-making,. We observed greater DACC and DLPFC activation to risky choices in the HIV+ group in the presence of broadly similar task performance between the two serostatus groups. This suggests an adaptive functional response, wherein additional cortical resources are recruited to maintain task goals. This may be in response to aberrant information provided by other cortical and subcortical structures to which these regions are connected and that may have been damaged by HIV infection. Within the HIV+ group, we observed increased activation in the right ACC, left thalamus, and right superior frontal gyrus as a function of increased sexual compulsivity. This suggests that those HIV+ individuals with elevated sexual compulsivity may be characterized by on overall increase of activation in fronto-striatal regions. Finally, we also observed that greater nadir CD4 count was significantly associated with greater activation to safe choices rather than risky options in all of the regions displaying between-group task-related differences. This suggests that HIV infection may alter risk-related neural processing. # Supporting Information The Translational Methamphetamine AIDS Research Center (TMARC) group is affiliated with the University of California, San Diego (UCSD) and the Sanford- Burnham Medical Research Institute. The TMARC is comprised of: Director – Igor Grant, M.D. ^1,3; Co-Directors – Ronald J. Ellis, M.D., Ph.D. ⁴, Scott L. Letendre, M.D. ⁵, and Cristian L. Achim, M.D., Ph.D. ¹; Center Manager – Steven Paul Woods, Psy.D. ^1,3; Assistant Center Manager – Aaron M. Carr, B.A. ³; Clinical Assessment and Laboratory (CAL) Core: Scott L. Letendre, M.D. (Core Director) ⁵, Ronald J. Ellis, M.D., Ph.D. ⁴, Rachel Schrier, Ph.D. ^{3, 6}; Neuropsychiatric (NP) Core: Robert K. Heaton, Ph.D. (Core Director) ^{1, 3}, J. Hampton Atkinson, M.D. ^1,3, Mariana Cherner, Ph.D. ^{1, 3}, Thomas D. Marcotte, Ph.D. ¹, Erin E. Morgan, Ph.D. ^1,3; Neuroimaging (NI) Core: Gregory Brown, Ph.D. (Core Director) ¹, Terry Jernigan, Ph.D. ^{1, 7}, Anders Dale, Ph.D. ⁴, Thomas Liu, Ph.D. ^{8, 9}, Miriam Scadeng, Ph.D. ^{8, 9}, Christine Fennema-Notestine, Ph.D. ¹, Sarah L. Archibald, M.A. ¹; Neurosciences and Animal Models (NAM) Core: Cristian L. Achim, M.D., Ph.D. (Core Director) ¹, Eliezer Masliah, M.D. ⁴, Stuart Lipton, M.D., Ph.D. ⁴, Virawudh Soontornniyomkij, M.D. ¹; Administrative Coordinating Core (ACC) – Data Management and Information Systems (DMIS) Unit: Anthony C. Gamst, Ph.D. (Unit Chief) ¹⁰, Clint Cushman, B.A. (Unit Manager) ^1,3; ACC – Statistics Unit: Ian Abramson, Ph.D. (Unit Chief) ¹¹, Florin Vaida, Ph.D. ¹², Reena Deutsch, Ph.D. ¹, Anya Umlauf, M.S. ¹; ACC – Participant Unit: J. Hampton Atkinson, M.D. (Unit Chief)^1,3, Jennifer Marquie-Beck, M.P.H. (Unit Manager) ¹; Project 1: Arpi Minassian, Ph.D. (Project Director) ¹, William Perry, Ph.D. ¹, Mark Geyer, Ph.D. ¹, Brook Henry, Ph.D. ¹; Project 2: Amanda B. Grethe, Ph.D. (Project Director) ², Martin Paulus, M.D. ^1,2, Ronald J. Ellis, M.D., Ph.D. ⁴; Project 3: Sheldon Morris, M.D., M.P.H. (Project Director) ⁵, David M. Smith, M.D., M.A.S. ⁵, Igor Grant, M.D. ^1,3; Project 4: Svetlana Semenova, Ph.D. (Project Director) ¹, Athina Markou, Ph.D. ¹, James Kesby, Ph.D. ¹; Project 5: Marcus Kaul, Ph.D. (Project Director) ¹³∶¹ Dept of Psychiatry, University of California, San Diego, La Jolla, California, United States of America. ² Psychiatry Service, VA San Diego Healthcare System, La Jolla, California, United States of America. ³ HIV Neurobehavioral Research Program, University of California, San Diego, San Diego, California, United States of America. ⁴ Department of Neurosciences, University of California, San Diego, San Diego, California, United States of America. ⁵ Department of Medicine, University of California, San Diego, San Diego, California, United States of America. ⁶ Department of Pathology, University of California, San Diego, San Diego, California, United States of America. ⁷ Department of Cognitive Science, University of California, San Diego, San Diego, California, United States of America. ⁸ Department of Radiology, University of California, San Diego, San Diego, California, United States of America. ⁹ Center for Functional Magnetic Resonance Imaging, University of California, San Diego, San Diego, California, United States of America. ¹⁰ Department of Biostatistics and Bioinformatics, University of California, San Diego, La Jolla, California, United States of America. ¹¹ Department of Mathematics, University of California, San Diego, La Jolla, California, United States of America. ¹² Department of Family and Preventative Medicine, University of California, San Diego, La Jolla, California, United States of America. ¹³ Immunity and Pathogenesis Program, Sanford-Burnham Medical Research Institute, La Jolla, California, United States of America. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ABG IG MPP RJE SPW. Performed the experiments: ABG CGC SJJ. Analyzed the data: CGC SJJ. Contributed reagents/materials/analysis tools: ABG CGC MPP. Wrote the paper: ABG CGC IG MPP RJE SPW.

# 1. Introduction Head and neck cancer is the sixth most predominant cancer in the world and \~90% of them arise from squamous cell epithelium. In India, head and neck squamous cell carcinoma (HNSCC) is a major concern, as it covers one third amongst all cancers compared to 4–5% in the developed world. Moreover, 60 to 80% of patients in India present with advanced disease (stage III/IV) as compared to 40% in developed countries. Currently, in India, HNSCC is predominantly managed by surgery, chemo-radiotherapy, targeted therapy and immuno-therapy. Huilgol et al, demonstrated the significant potential of hyperthermia therapy (HT) to enhance the efficacy of chemo-radiation in HNSCC in terms of increasing overall survival, disease-free survival and quality of life without increasing the risk of complication or additional toxicities. Thus, HT presents a promising approach for the treatment of HNSCC by improvement of clinical response and reducing the toxicities of radio- and chemo-therapy. In recent years, research has been aimed at developing targeted molecular therapeutics for the management of HNSCC. However, despite the improvement in the treatment modalities and development of targeted molecular therapeutics, there has been only modest improvement in the overall survival of HNSCC patients. This seems to be associated with the late diagnosis as well as the advanced clinical stage at the time of diagnosis. Thus, early diagnosis has become imperative for reduction in the current mortality rate in HNSCC patients. Several recent studies have examined plasma, serum, saliva and tissue biopsy samples to identify clinically significant bio-markers for HNSCC patients. Tissue bio-markers pose challenges owing to the heterogeneous nature of HNSCC involving different sites ranging from oral cavity to upper aero- digestive tract. Moreover, obtaining biopsy samples at different points of therapy is not feasible. On the contrary, serum or plasma based bio-markers are expected to have better diagnostic value due to ease of sample collection and better representation of tumor secretory profile. Heat shock proteins (HSPs) have been known to be elevated in the serum/plasma samples of several cancer patients, especially solid tumors of lungs and liver. HSPs are of particular interest due to their role in modulation of thermo- tolerance, which in turn affects the response of the tumors to HT. Tu et al, demonstrated the modulation of HSP70 and HSP90 expression in gastric tumors after transient hyperthermic intra-peritoneal chemoperfusion (HIPEC) treatment. Their findings suggested that the application of second round of HIPEC after a gap of 24 h could minimize the chemo- and thermo-tolerance induced by elevated serum levels of HSP70/90 after first round of HIPEC. Amongst HSPs, HSP90 (particularly the alpha and beta isoforms of HSP90) have been implicated for the diagnostic potential in case of lung and liver cancer. However, a definitive correlation of the two isoforms of HSP90 and their modulation in cancer patients has not yet been established. Besides, reports do not exist that study the level of HSP90 in HNSCC and its variation in response to HT. Our previous studies suggested the role of HSP90 in the mechanism of radio-sensitization and thermo- sensitization after magnetic hyperthermia therapy (MHT) in mouse fibrosarcoma tumor model. Thus, the aim of the present study is to validate the potential of HSP90 as diagnostic and prognostic marker in clinical scenario. For this serum samples of HNSCC patients were evaluated for the levels of HSP90. HNSCC cancer type was chosen as conventionally HNSCC patients are subjected to HT to improve the efficacy of chemo-radiotherapy (CRT), thereby making them a suitable model to evaluate the prognostic potential of HSP90 for predicting the response to HT. # 2. Methods ## 2.1 Tumor details and treatment plan in HNSCC During November 2016 to May 2019, total 15 HNSCC patients from Radiation Oncology Department, Dr Balabhai Nanavati Hospital, Mumbai were enrolled in this study with patients with curative intent. Necessary written consent from the patients and ethical approval from the Institute has been taken for the study. Healthy subject samples (age-matched) were obtained after appropriate ethical approval from BARC hospital. The patients with squamous cell carcinoma (SCC) comprised of the primary tumor sites of pyriform fossa (n = 4), tongue (n = 3), supra-glottis (n = 2), vocal cord (n = 1), uvula (n = 1), lung (n = 2), gingivobuccal sulcus (n = 1) and pharyngeal wall (n = 1). The SCC was confirmed by histopathology and the tumor staging status classified according to the eighth edition of American Joint Committee on Cancer (AJCC) classification, showed most of the tumors to be localized (T2-T4, N = 12). Three out of 15 tumors involved the lymph nodes with staging of T3N1, T4N2 and T3N2. Prior to HT, all the patients had been subjected to radiation therapy (average fractionated dose of 65 ± 5.2 Gy, each fraction of 2 Gy) and chemotherapy (cisplatin 60 mg/week for 6 doses). Radiation was delivered with conventional fractionation (2 Gy/day 10 Gy/week, total dose: 66–70 Gy) using Linear Accelerator with Intensity Modulated Radiation Therapy or 3 dimensional conformal therapy. Chemotherapy was infused once a week on any convenient day. Chemotherapy was not infused on the day of hyperthermia. Hyperthermia was delivered after pre-cooling for a period of 30–40 minutes, once every week \~1 h after the radiation therapy session. Hyperthermia always followed radiotherapy. HT was delivered to the patients (following a pre-cooling step) on Thermatron, a radiofrequency (RF) machine operating at 8.2 MHz. A pair of antennae was placed across the solid tumor, guided by visible tumor or anatomical landmarks. The power input was started after impedance matching input varied from 400 to 1000 kW. Power was gradually escalated unless the patients complained of unbearable pain, stress or discomfort. Power was then reduced and maintained till completion of the treatment. HT was stopped if the patient developed ≥ grade II thermal burns. Whenever feasible, the temperature attained during HT was measured by invasive thermometry with a thermistor probe. The average HT temperature to which the tumors were exposed was 41.87 ± 0.83°C. Patients received HT (1 to 9 weekly sessions) for 30 minutes after pre-cooling for 10 minutes. Patients’ response to the treatment was assessed at the end of the treatment cycle. Response of patients to therapy was categorized according to RECIST (Response Evaluation Criteria in Solid Tumors) criteria. Accordingly, disappearance of all tumor lesion was categorized as complete response, reduction of \>30% in tumor size as partial response (PR), \<30% reduction in tumor size as stable disease (SD) and growth of \>20% or occurrence of new lesions as progressive disease (PD). Patients with complete response were categorized as complete responders (CR), whereas, patients with PR, SD or PD were categorized as non-responders (NR) to treatment. The progression free survival ranged from 12 to 30 months with an average of 18.6 ± 3.41 months for CR as against 2.8 ± 0.91 months for NR (range: 1 to 6 months). Serum samples of healthy donors were collected after taking appropriate approval from BARC ethical committee. ## 2.2 Serum collection and storage For ELISA analysis, 5 ml of blood was collected and serum was isolated by centrifugation (2200 RPM for 15 min) within 1 h of blood collection. Serum samples were collected at two time points, viz., before HT (but after completion of CRT) and 24 h after completion of HT session. The serum samples were stored at -80°C till further analysis. Repeated freeze-thaw cycles were avoided. ## 2.3 ELISA analysis The levels of HSP90 beta were determined by using human specific ELISA kits (Cusabio, USA) following manufacturer’s instructions. Values lower than the minimum detectable dose (MDD) of the kit have not been considered and labeled as non-detectable (ND). Anonymized data for serum levels of HSP90 beta in HNSCC and HC is mentioned in ( and Tables). In case of HNSCC patients, 1 out of 15 patients showed serum level of HSP90 lower than the MDD. Similarly, in case of HC, 10 out of 42 Samples showed non-detectable levels of HSP90 beta in serum. Therefore, for statistical analysis 14-HNSCC and 32-HC samples have been considered. ## 2.4 Statistical analysis Statistical significance was determined by Mann-Whitney test and two sample t-test using Origin Pro 8.0 software. The paired comparison of ROC curves and determination of Youden index were performed using MedCalc Software ver 18.11.6 and web-based tool for ROC analysis (Epitools, [http://epitools.ausvet.com.au](http://epitools.ausvet.com.au/)). The optimum cut off value was determined by using the quantity corresponding to the maximum value of Youden’s index (Youden’s index = sensitivity+specificity-1). ## 2.5 Ethics approval All procedures performed in studies involving human participants were in accordance with the ethical standards of the Dr. Balabhai Nanavati Hospital, Mumbai and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. # 3. Results ## 3.1 Serum levels of HSP90 beta were found to be significantly higher in HNSCC as compared to healthy controls (HC) To establish the diagnostic potential of HSP90 in clinical scenario, we compared the serum levels of HSP90 alpha and beta in HNSCC (N = 14, Mean age: 59.2±12.7 years) and HC (N = 32, Mean age: 52.2±16.2 years) ****. Results showed significantly higher level of HSP90 beta (2.7 fold) in serum samples of HNSCC (65.6±13.08 ng/ml) compared to HC (23.5±3.8 ng/ml) with a P value of 0.002 **( & ).** However, the levels of HSP90 alpha did not show significant difference between the HNSCC and HC. An anonymized data of the HNSCC patients and healthy controls for HSP90 beta is being mentioned in ( **and Tables**). ## 3.2 Comparison of HSP90 beta levels between complete (CR) and partial/non-responders (NR) in serum samples of HNSCC patients subjected to CRT or CRT+HT To further determine the correlation between serum levels of HSP90 beta in HNSCC and their response to HT treatment, the HNSCC patients were categorized as CR or NR at the end of chemo-radiation therapy (CRT)+HT treatment session. Baseline serum samples were collected before HT (but after completion of CRT) and compared with the HSP90 beta levels in CRT+HT cohort, at 24 h after completion of HT session. Our results showed in-significant difference in the serum levels of HSP90 beta between CR and NR in the CRT cohort. However, 24 h after HT, significantly lower (\~ 5 fold) serum levels of HSP90 beta was observed in CR (25.62±9.04 ng/ml) compared to NR (130.51±34.23 ng/ml) **( and &)**. ## 3.3 Serum HSP90 beta showed significant efficacy for detection of HNSCC and their clinical response to HT ROC curve analysis showed an AUC of 0.78 (95% CI: 0.64–0.92) with a sensitivity of 84.6% and specificity of 62.5% at a cut-off value of \>22.6 ng/ml for distinguishing HNSCC from HC (P \<0.05) ****. Nevertheless, no significant difference (P = 0.17) was observed between CR and NR in case of CRT cohort **( & Tables)**. However, in CRT+HT cohort, a significant difference (P\<0.05) in the serum levels of HSP90 beta was observed between CR and NR **( & Tables)** with an AUC of 0.96 (95% CI: 0.86–1.06) and a sensitivity and specificity of 100 and 80%, respectively, at a cut-off value of \>35.4 ng/ml. Youden’s index was found to be 0.47 and 0.8 for distinguishing HNSCC versus HC and CR versus NR, respectively in serum samples of CRT+HT ****. # 4. Discussion Hyperthermia therapy (HT) has emerged as a promising approach for improving the efficacy of CRT in HNSCC. Despite of the advancements in the therapeutic strategies, the improvement in overall survival of HNSCC patients has been dismal. Major factors contributing to the poor prognosis of HNSCC patients has been the lack of diagnosis of disease at an early stage and the associated failure of treatment at the advanced stages of the disease. Delayed diagnosis of HNSCC has been associated with poor prognosis with a 5-year median overall survival of 42% at stage IV in Indian patients. With the evolving understanding of the molecular basis of human malignancies, there has been great interest in determining whether serum biomarkers might aid in early diagnosis and guide treatment decision for the cancer patients. Heat shock proteins (HSPs), especially, HSP70, HSP27 and HSP90 have been known to be released in the extra-cellular environment in various solid tumors. Elevated expression of HSP90 has been found to correlate with tumor cell proliferation, tumor stage and poor clinical outcome, suggesting potential use of HSP90 expression in cancer diagnosis and prognosis. Correspondingly, there are studies highlighting the role of HSP inhibitors used in conjunction with HT to improve its overall therapeutic efficacy in solid tumors. This effect is conjectured on the ability of the HSP inhibitors to potentiate the cyto-toxic and/or anti-proliferative effects of HT. Concurring with these reports, our previous studies in murine fibrosarcoma tumor models (both *in vitro* and *in vivo*) showed differential expression of HSP90 in cancer cells and tumor lysates. Moreover, the modulation in the levels of HSP90 correlated with the response of tumor to MHT. Therefore, to further evaluate the diagnostic efficacy of HSP90 in clinical settings, we studied the serum levels of HSP90 beta and alpha in HNSCC patients. HNSCC model was specifically chosen owing to the proven advantage of HT in HNSCC for the improvement of therapeutic efficacy of chemo- radiotherapy. In the present pilot study, HNSCC patients were found to have significantly higher expression of HSP90 beta but not HSP90 alpha **** as compared to HC. This is contrary to other reports demonstrating association of elevated plasma levels of HSP90 alpha with presence of lung or liver cancer, compared to HC. Interestingly, both the isoforms of HSP90, viz., alpha and beta have been implicated as diagnostic serum/plasma biomarkers for several solid tumors. The difference in our observation (insignificant change in HSP90 alpha) compared to these studies may be attributed to the different cancer types/nature of sample. Thus, a correlation between the HSP90 isoforms and their predictive efficacy in a particular tumor type is not well reported and needs further investigation. Furthermore, our results showed significantly higher levels of HSP90 beta in partial or non-responders (NR) as compared to CR in serum samples collected after 24h of completion of HT in CRT+HT cohort. ROC analysis showed a sensitivity of 84.6 and 100% and specificity of 62.5 and 80% for distinguishing the HC from the HNSCC patients (P-value \<0.05) and CR from NR (P\<0.05) in CRT+HT cohort, respectively. These values are comparable to another study by Yamashita et al., wherein a sensitivity and specificity of 57.3 and 85.3%, respectively, was observed for predicting HNSCC using serum midkine levels as a bio-marker. Thus, our preliminary results demonstrate the diagnostic efficacy of HSP90 beta for HNSCC and its predictive efficacy for response to HT in HNSCC patients. # 4. Conclusions Present pilot study suggests the potential of serum HSP90 beta as a diagnostic serum bio-marker for HNSCC and predicting their response to HT. However, more patients need to be incorporated in the study for improving the statistical power and clinical translation of the research. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Travel distance is a major barrier to the utilisation of all kinds of health services. We have previously shown that increasing distance from the population centroid to a casualty clinic is strongly associated with decreasing contact and consultation rates in a Norwegian out-of-hours district. The association was almost completely independent from important demographic and socioeconomic factors registered. Often no data on the exact location of individual patients are available. To examine the effect of geography on the utilisation of health services, several methods have been used to estimate the average travel distance for the entire population of an area. It has previously been shown that estimates of distance to the nearest cancer centre from population centroid is superior to estimates based on geometric centroid and population polygon methods when the exact addresses are unknown. E.g. if a casualty clinic for a population of one municipality is located in another municipality, the distance from the population centroid is fairly representative for the average travel distance. In municipalities hosting their own casualty clinic, however, the population centroid often almost coincides with the location of the clinic. In these cases the distance between the population centroid and the casualty clinic is an underestimation of the travel distance for the average inhabitant. In order to get more representative measures on distance and travel time at short distance between the casualty clinic and the population centroid of a municipality, we have developed a method that uses postcode coordinates that have been determined by crowdsourcing. The data are freely available for everyone, and since our method uses Google maps, there is no need for complicated and expensive geographic information system (GIS) software. In Norway, the alternative is to get estimates on population centroids or mean distances and travel times from Statistics Norway, the official statistics bureau. This is an expensive and time-consuming process. Crowdsourcing is an online, distributed problem-solving and production model that involves outsourcing of tasks to a distributed group of people. The contribution from thousands of voluntary people has been important for the development of geographic information systems (GIS) and related online resources. In this paper we describe the method and compare it with other ways of calculating distance and travel time to a casualty clinic when the exact address is unknown. We also apply our method to examine whether distances and travel times are associated with the utilisation of out-of-hours primary care services. # Methods ## Ethics Statements The data used in this study are from three sources. Estimates on mean distance, travel time and population centroids of the municipality are not freely available, and can only be obtained from Statistics Norway for a fee. The postcode database is freely distributed under the terms of the Creative Commons Attribution License 3.0, and is open to all from its website. This database does only contain data on post/ZIP-codes. The Watchtower project database is owned and managed by our institution, the National Centre for Emergency Primary Health Care. The database is not publicly and freely available. These data are anonymised, patient identity is not recorded at any time. The Watchtower project is approved by the Regional Committee for Medical and Health Research Ethics and by the Norwegian Social Science Data Services, the Data Protection Official for Research. ## Crowdsourcing Process In 2009 a joint weather service owned by the Norwegian Meteorological Institute and the Norwegian Broadcasting (NRK), wanted to provide weather forecasts for each of the 4 585 Norwegian postcodes. At that time, maps from Norway Post (Posten Norge) displaying the area covered by each postcode were available. Based on these data coordinates for the geometric centroids could be calculated. In a sparsely populated country like Norway, however, these coordinates are in many cases not representative for where people actually live. To get more representative postcode coordinates, a crowdsourcing project was launched in July 2009. Volunteers were recruited using NRKbeta, the technology website of the Norwegian Broadcasting. The head of the project, Erik Bolstad (EB), created a Google map with approximate locations of each postcode based on Norway Post’s downloadable postcode registry. Contributors were able to add more accurate coordinates directly by clicking in the map. Updated coordinates were received by Google forms and transferred to a Google spreadsheet. EB checked all submitted coordinates before updating the project webpage. Before the project started 39% of postcodes were located. After two days, this increased to 74%. Within the first month 4 590 feedbacks from approximately 600 persons were submitted. The capacity of the registration system was then exceeded due to the large amount of information. EB then had to find the location of the remaining 600 postcodes. A five point scale was used to assess the data quality of each location. Since the formal termination of the project, the project pages have been open for volunteers to improve the database further. The database also contains data from Statistics Norway on the number of people belonging to each postcode. ## Casualty Clinics In Norway, casualty clinics (“legevakt”) are emergency primary care centres that handle all kinds of medical out-of-hours inquiries, including life-threatening incidents. Some municipalities have their own casualty clinic (monomunicipal), but casualty clinics serving the population of two or more municipalities (intermunicipal) are also common. Of the 203 out-of-hours districts in Norway in 2012, 84 were monomunicipal and 88 were intermunicipal. The remaining districts were partially monomunicipal. We used the postcode database to estimate average travel distance to the casualty clinic for 18 municipalities participating in the “Watchtower project”, a database of a Norwegian sentinel network for monitoring emergency primary health care activity in Norway. The database was established by the National Centre for Emergency Primary Health Care in 2007. The Watchtowers are seven out- of-hours districts that have been selected to be representative for the entire primary health care emergency service in Norway. The development and implementation of this database have been described in detail elsewhere. ## Calculation of Postcode Based Travel Time and Distance The travel distance and travel time between each postcode coordinate within the Watchtower municipalities and the casualty clinic was calculated using Google maps. This was then multiplied by the number of inhabitants belonging to the postcode. The average distance to the casualty clinic was then calculated by dividing the sum of person-kilometres or person-minutes by total number of inhabitants in the municipality. Instead of using GIS software to calculate distances and travel times, Google maps was chosen primarily because it is freely available to all. It is also easy for people without knowledge of GIS to calculate distances and travel times since the postcode database website has integrated Google maps functionality. ## Validation of the Method The postcode based distances to casualty clinics for each municipality were compared with mean and median travel distances and travel times and distance from the population centroid of the municipality. Mean distance and travel time of each municipality to the casualty clinic as calculated by Statistics Norway was considered the main basis of comparison for our method. Statistics Norway used Excel to manually calculate mean and median distance and travel time from all “address points” in the municipality, taking number of persons living at each address point into account. These data were taken from Statistics Norway’s Address Points database January 1^st 2012. This not freely available database contains the exact location of all addresses in Norway. Population centroids were calculated with the mean centre function in ArcInfo (ArcInfo, ESRI, Redlands, CA). Distances and travel times from population centroids were estimated using OD cost Matrix in ArcInfo. These estimates were commissioned from Statistics Norway for a fee. In addition, we made comparisons with distances and travel times between casualty clinic and the town halls of the municipalities, calculated by use of Google maps. Comparisons of average absolute deviation from mean distance as calculated by Statistics Norway for each method were performed between mono- and intermunicipal out-of-hours districts, and between municipalities hosting a casualty clinic or not. ## Postcode Based Distance and Utilisation of Out-of-hours Services We used the postcode based distances to examine the association between distance to casualty clinic and the utilisation of out-of-hours services. One municipality (Austevoll) was excluded from these analyses because the out-of- hours service uses more than one casualty clinic. Similar analyses on population centroid based distances have previously been performed by our group in the ten Watchtower municipalities comprising one specific out-of-hours district. These ten municipalities were among the 17 municipalities that were examined in the present study. Utilisation was defined as rate of all first contacts with the communication centre of the out-of-hours district and rates of face-to-face doctor consultations in each of the Watchtower municipalities. Rates were calculated by dividing aggregated total number of contacts and consultations for each municipality in each year from 2007 to 2011 by population January 1st the same year. ## Statistical Methods IBM SPSS 20 and Excel 2010 software were used to analyse data. The correlations between mean distance and travel time to casualty clinics and postcode based and other methods were examined using Pearson correlation coefficient. The correlation analyses were repeated several times, including municipalities at different maximum distances and travel times. Short distances were defined as less than ten kilometres, and short travel times defined as less than 15 minutes. The magnitude of difference from mean distance calculated by Statistics Norway was expressed as mean absolute error (MAE) with 95% confidence intervals. Contact and face-to-face consultations rates are number of events per year, which are assumed to be Poisson distributed. The rates of our main outcomes were so high that a normal distribution will fit the data well. Correlation coefficients, r, and constants and coefficients with 95% confidence intervals of exponential functions (non-linear regression) were used to assess any association between distances from municipality (postcode based) to out-of- hours-centre and contact and consultation rates for the years 2007–2011 for each municipality. Exponential curve fit was chosen because this was found to best describe the association between distance and utilisation of out-of-hours services in a previous study. # Results 449 868 Watchtower contacts were registered from 2007 through 2011. 2 457 contacts were excluded because information about municipality was lacking. 135 770 contacts were excluded because the patient was a non-Norwegian citizen or was living in a municipality outside the Watchtowers out-of-hours districts. 311 641 contacts were eligible for analysis, aggregated to 90 municipality-year observations. Baseline demographic and socioeconomic data for the municipalities are given in. and display correlation analyses of the relationship of distance to the casualty clinic from different population centres and mean distance as computed by Statistics Norway. When analysing only municipalities with less than ten kilometres to a casualty clinic, only postcode based distances were significantly correlated with mean distances. shows the superior correlation of the postcode based method at distances less than 20 kilometres. For travel time to casualty clinic, as shown in and, there were strong and statistically significant correlations for all time measures at long travel times. At travel times less than 15 minutes, none of the methods were significantly correlated with mean travel time as calculated by Statistics Norway. At longer travel times, the postcode based method showed a correlation to mean travel time similar to the other methods. Three of the municipalities had very similar mean distances and mean travel times that explain the abrupt change in critical r at 9.5 kilometres and 11.5 minutes. For municipalities not hosting a casualty clinic, there were no statistically significant differences in MAE between the distance measures compared with mean travel distance calculated by Statistics Norway. For distances based on population centroids and town halls, the deviation from mean distance was more than twofold larger in the municipalities where a casualty clinic is located. No statistically significant differences were observed in deviation between hosting and non-hosting municipalities in neither postcode nor median based distances, but the variance of deviation to mean distance was much higher for median based distances. For median and postcode based distances, there was no difference in MAE between mono- and intermunicipal out-of-hours districts. For distances based on population centroid and town hall location, the deviation from mean distance was more than twofold higher in municipalities not taking part in an intermunicipal out-of-hours district. As shown in and, there were no significant differences in regression coefficients between different methods when analysing the association between distance and consultation and contact rates, but population centroid and town hall based distances differed more from mean distance, although not statistically significant. Exponential regression showed that increasing (postcode based) distance was significantly associated with lower contact rate and to a larger extent face-to- face consultation rate. There was no difference between municipalities participating in intermunicipal or monomunicipal out-of-hours districts. # Discussion We have described a method for estimating distance and travel time to a casualty clinic based on postcodes determined by crowdsourcing. The method showed good correspondence to the main basis of comparison, which was mean distance and travel time based on address points. The method was at least as good as the other examined methods on both short and long distances. At distances shorter than 20 kilometres, the postcode based measure was superior to distances based on town hall location and population centroid. In contrast to distance and travel time measures based on town hall location and population centroid, the postcode based method showed no difference between mono- and intermunicipal out- of-hours districts, or between municipalities hosting a casualty clinic or not. The association between postcode based distance and contact and consultation rates in the 17 Watchtower municipalities was in accordance with our previous analysis of the 10 municipalities comprising Arendal out-of-hours district. To calculate average travel distance and time from a municipality based on postcodes with the method described in this study, a database containing coordinates and population number for each postcode is necessary. The availability of such data is for most countries more limited than in Norway. There are some on-line services in UK that provide coordinates of postcodes, but to our knowledge information on the population of each postcode is not readily available. Some commercial providers, like Deutsche Post in Germany provide data based on addresses. Another limitation is that we used Google maps to calculate road distances and travel times. The reliability of Google Maps presumably varies from area to area because map data are acquired from different sources. Crowdsourcing projects with large numbers of contributors result in products that are complete, accurate and up to date, but they are vulnerable to errors and vandalism. Research on Wikipedia has shown that strict coordination of the editing process is crucial for quality. The thorough planning and coordination of the more than 600 contributors of the postcode project suggests that the database is trustworthy. ## Conclusion The results show that distances and travel times for the inhabitants of Norwegian municipalities to a casualty clinic can be easily calculated with standard Internet tools from freely available postcode based coordinates determined by crowdsourcing. The method described here proved valid at both long and short distances, and is more reliable than distance estimates based on population centroid at short distances. We would like to thank Erik Bolstad for all the effort he has put into the postcode project, Ole-Johan Eikeland for preparing the Watchtower data files, Narve Sætre for technical assistance, Åsta Haukås for proofreading the final manuscript, and the staff at the Watchtower casualty clinics for performing the registrations. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GR SH. Performed the experiments: GR. Analyzed the data: GR. Contributed reagents/materials/analysis tools: GR SH. Wrote the paper: GR SH.

# Introduction Karst ecosystem occupies approximately 7\~12% of emerged land globally, mainly distributed in southwest China, and is characterized by high habitat heterogeneity and high vegetation fragmentation with high soil erosion, rocky desertification, and barren vegetation nutrient deficiency. However, the karst vegetation retains a robust natural resilience even in harsh habitats. Initially, the pioneer herbaceous plants, mainly from Gramineae and Compositae, have high resistance to drought and barrenness, grow fast, and improve soil structure in the primary succession stage ecosystem restoration. In addition, soil microorganisms play an essential role in recovering degraded karst systems through promoting the growth and nutrient uptake by plants as well as increasing soil nutrient bioavailability. Thus, soil microorganisms in karst vegetation restoration cannot be ignored. AM fungi, a soil functional microorganism, can play critical roles in recovering degraded terrestrial ecosystems. AM fungi formed a symbiotic relationship with 80% of terrestrial plants, improve plant growth, nutrient accumulation, enhance drought stress tolerance and maintain soil structure, e.g. Guo et al. (2021) proposed that AM fungi differently affected the competitive ability of *Broussonetia papyrifera* and *Carpinus pubescens*; Xia et al. (2020) also showed that AM fungi increased nutrients of host plants by regulating the morphological development of karst plant roots. In addition, Shi et al. (2015) illustrated that AM fungi increased the biomass, N, and P content in shoots and roots of plants. Furthermore, AM fungi mycelium can transfer the photosynthetic carbohydrates from the host plants to the soil, which recruits soil microorganisms. However, we know relatively little about how regulation of plant growth and nutrient by AM fungi is affected by interaction with indigenous microorganisms. AM fungi via extensive extraradical hyphae interacting with indigenous microbial communities play crucial roles in plant growth in natural habitats. AM fungi and bacteria are ubiquitous in natural soil. Specifically, AM fungi regulate plant growth, and they are positively affected by cooperating with indigenous microorganisms or negatively affected by competing with indigenous microorganisms. Ortiz et al. (2015) suggested that the combination of AM fungi and specific bacteria could promote plant growth by minimizing drought- related stress effects. Artursson et al. (2006) also proposed that the co- inoculation of AM fungi and phosphorus-solubilizing bacteria positively promotes plant nutrient absorption. In addition, plant growth-promoting rhizobacteria could promote mycorrhizal fungal activity and establishment, which are called mycorrhiza helper bacteria. In contrast, the competition phenomenon between AM fungi and bacteria was also widely reported. Azcón-Aguilar et al. (1997) presented evidence of direct competition between AM fungi and indigenous microorganisms for photosynthetic products of the host plant. Indirectly, Doumbou et al. (2005) proposed that many *Streptomyces* sp. could exude antifungal compounds, which indicated that they are fungal competitors under the appropriate environmental conditions. Thus, the cooperation and competition between AM fungi and indigenous microorganisms are ineluctability in karst soil. In summary, AM fungi play important roles in improving plant growth and nutrient absorption. However, AM fungi inevitably interact with indigenous microorganisms in the vegetation restoration of the karst-degraded ecosystem. It remains unclear how indigenous microorganisms affect the benefits of growth and nutrition regulated by AM fungi for plants in karst soils. Because of the complexity and uncertainty of the interaction between AM fungi with indigenous microorganisms, it is necessary to assess the effect size of AM fungi and indigenous microorganisms, and their interaction, on plant growth and nutrition. The aim is to clarify how indigenous microorganisms affect the benefits of growth and nutrition regulated by AM fungi for plants in karst soils. We hypothesize that: (1) AM fungi can promote the growth and nutrients of karst plants (H1), according to that AM fungi increased plant biomass and nutrition accumulation. (2) Indigenous microorganisms can offset the benefits from AM fungi on plant growth and nutrient accumulation (H2), according to that indigenous microorganisms may negatively affect the AM benefits for the host plant through competition. # Materials and methods ## Experiment treatments A potting experiment was conducted by using four herb species: *Setaria viridis*, *Arthraxon hispidus*, *Bidens pilosa*, and *Bidens tripartita* in polypropylene plastic pots in a greenhouse of Guizhou University in Guiyang, China (E: 106°22′ E; N: 29°49′ N; 1,120 m above the sea level). Three different microbial conditions soil was created to explore the interaction of AM fungi with indigenous microorganisms in the regulation of plant growth and nutrient utilization. It included AM fungi inoculating into sterilized soil (*AMF* treatment), AM fungi inoculating into natural conditions soil containing indigenous microorganisms (*AMI* treatment), and the control soil by removing microorganisms with sterilization (*CK* treatment). In the beginning, limestone soil (Calcaric regosols, FAO) was collected from a typical karst habitat, from which approximately two-thirds of the soil was used for sterilization at 126°C, 0.14 Mpa for one hour to eliminate microbes, and one-third of the soil was retained for further experiments. Subsequently, a 2.5 kg soil subsample of the sterilized or unsterilized soil was put into each polypropylene plastic pot (180 mm × 160 mm, diameter × height). Five seeds of *Setaria viridis*, *Arthraxon hispidus*, *Bidens pilosa*, and *Bidens tripartita* were disinfected with a 10% H₂O₂ solution for 10 minutes and repeatedly washed with sterile water, and sown in each pot. After sowing seeds in each pot, seeds were covered with 200 g of the respective soil for promoting seed germination. In addition, the sterilized soil was inoculated with 10 g *Glomus mosseae* inoculum as the *AMF* treatment, and the original soil from field habitat was inoculated with 10 g *Glomus mosseae* inoculum as the *AMI* treatment, indicating the AM fungi interacting with the indigenous microorganisms in this experiment. Especially, *CK* treatment received an additional 10ml of the filtrate by weighing 10g of *Glomus mosseae* inoculum with sterile water using a double- layer filter paper, along with a 10 g of sterilized inoculum of *Glomus mosseae* was added in order to maintain the consistency of microflora except for the targeted fungus *Glomus mosseae* corresponding to *AMF* treatment. The inoculum propagated for four months with *Trifolium repens*, including approximately 100 spores per gram soil, hyphae, and colonized root pieces. There is mutual control between two of three treatments: the AM fungi effect through comparing *AMF* with *CK* treatment; the interactive effect of AM fungi with indigenous microorganisms through comparing *AMI* and *CK* treatment; and the indigenous microorganisms effect related to AM fungi through comparing *AMI* with *AMF* treatment. Of course, we had to admit that the unsterilized soil probably had native AM fungi under *AMI* treatment, even the targeted species *Glomus mosseae*. However, it was sure that the *Glomus mosseae* inoculum interacted with native AMF species and indigenous microorganisms; further, they jointly affected plants and soil for growth and nutrition when comparing *AMI* with *AMF*. All treatments were replicated five times, and four plant species contained 60 pots. The physicochemical properties of limestone soil (per kg) were measured by the methods from Tan (2005), the PH 8.2, total nitrogen (TN) 0.622 g, alkaline hydrolysis nitrogen (AN) 0.315 g, total phosphorus (TP) 1.274 g, available phosphorus (AP) 0.163 g, total potassium (TK) 37.79 g, and available potassium (AK) 0.532 g. All plant seeds were also collected from the same karst habitat used to collect soil. According to the primary field survey, these plants are successive pioneer species of karst communities as the herbaceous stage, which generally coexist in the same habitat as the main Gramineae and Compositae. Three weeks after seeds germination, only two seedlings were retained in the pot and cultured for five months. All growing seedlings were watered one time per day for maintaining field capacity, then harvested to determine the biomass, N, and P concentrations. The *Glomus mosseae* inoculum was initially purchased from the Institute of Nutrition Resources, Beijing Academy of Agricultural and Forestry Sciences (NO.BGA0046). ## Determinations of the root mycorrhizal colonization, biomass, and the accumulation of nitrogen and phosphorus The grid line-intersect method determined the root mycorrhizal colonization rate. The biomass of *S*. *viridis*, *A*. *hispidus*, *B*. *pilosa*, and *B*. *tripartita* were respectively determined by weighing tissue of root, stem, and leaves after drying at 80°C to constant weight. The nitrogen and phosphorus concentrations in plant tissue were determined by the traditional Kjeldahl method and the Molybdenum-antimony anti-colorimetric method, respectively. Additionally, the accumulations of nitrogen and phosphorus were calculated through nutrient concentration multiplying by biomass, respectively. Then the nutrient accumulation of plant individuals was accumulated by root, stem, and leaf. ## Calculation of effect size The effect size was calculated using the response ratio (ln*R*) of treatment groups to the control groups plant biomass referred from the proposition of regarding the plant response mycorrhizal fungi. The AM fungi effect (*AME*) by *AMF vs*. *CK*, the interactive effect of AM fungi with indigenous microorganisms (*AIE*) by *AMI vs*. *CK*, and the indigenous microorganisms effect related to AM fungi (*IME*) by *AMI vs*. *AMF* were calculated respectively, due to the mutual control between two of three treatments in this experiment. Therefore, the modified method was adopted according to Hoeksema et al. (2010) and Hedges et al. (1999) as follows: $${\text{ln}\mspace{360mu}{R =}}{\text{ln}{\mspace{360mu}(Xt/\mspace{360mu} Xc)\mspace{360mu}}}$$ Where *Xt* and *Xc* represent the biomass or nutrient accumulation of the plant in the values of the treatment group and control group, respectively, values\> 0 indicate positive effects promoting plant growth or nutrient accumulation, values \< 0 indicate negative effects suppressing plant growth or nutrient accumulation. ## Statistical analyses All of the statistical analyses were performed through SPSS 25.0 software. All of the data were tested for normality and homogeneity of variance before analysis. Two-way ANOVA was applied for assessing the effects of plant species (Ps; *Setaria viridis* vs. *Arthraxon hispidus* vs. *Bidens pilosa* vs. *Bidens tripartita*), soil microbial treatments (Ms; *AMI* vs. *AMF* vs. *CK*), and their interactions (Ms×Ps) on plant biomass, nitrogen accumulation, and phosphorus accumulation, N/P ratio and effect size by the ln*R*. The least significant difference (LSD) test was applied to compare significant differences in root mycorrhizal colonization, biomass, nitrogen, and phosphorus accumulations, and N/P ratio with effect size by the ln*R* among the three different conditions of soil microbial treatments with *AMI*, *AMF*, and *CK* or four plant species of *Setaria viridis* and *Arthraxon hispidus* and *Bidens pilosa* and *Bidens tripartita* at P≤0.05. All graphs were drawn on Origin 2018. # Results ## Root mycorrhizal colonization of four plant species under different microbial treatments A non-significant *AMI* \> *AMF* of root mycorrhizal colonization was observed in the four species. However, the root mycorrhizal colonization of *CK* treatment was zero; meanwhile, the AM fungus spore and mycelium were not discovered under *CK* soil substrate via microscopic detection. The root mycorrhizal colonization of *B*. *pilosa* and *B*. *tripartita* were significantly greater than that of *A*. *hispidus* and *S*. *viridis*, respectively, while for *A*. *hispidus*, it was also greater than *S*. *viridis*. Besides, there was no significant difference in root mycorrhizal colonization of *B*. *pilosa* and *B*. *tripartita* under *AMI* and *AMF* treatments. These results indicate root mycorrhizal colonization is species differences, and it provides evidence for host preferences of AM fungal. ## Biomass and its response ratio of four plant species under different microbial treatments The soil microbial condition treatments (Ms) significantly affected biomass. Significantly *AMF* \> *AMI* \> *CK* of biomass were observed in *A*. *hispidus*, *B*. *pilosa*, and *B*. *tripartita* seedlings except for *S*. *viridis*. Plant biomass was increased by AM fungus when comparing *AMF* with *CK* and *AMI* with *CK*, respectively. However, the plant biomass under *AMI* was significantly lower than under *AMF*. The plant species (Ps) also significantly affected individual biomass. Under *AMF* and *CK* treatments, the biomass of *A*. *hispidus* was significantly greater than the other three species. The biomass of *S*. *viridis* was significantly lower than the other three species under *AMF* and *AMI* treatments. In addition, there was a non- significant difference in biomass observed between *B*. *pilosa* and *B*. *tripartita* seedlings under any soil microbial condition treatments. Meanwhile, the interaction of Ms×Ps significantly affected the individual biomass for the four species. The results revealed that *AMF* and *AMI* treatments significantly increased the biomass accumulation of four karst pioneer species. Meanwhile, the biomass was significantly different between *A*. *hispidus* and *S*.*viridis* of Gramineae, except for *B*. *pilosa* and *B*. *tripartita* under *AMF*. Similarly, the soil microbial condition treatments (Ms), the plant species (Ps), and their interaction significantly affected the response ratio of biomass (ln*R*_*Biomass*). On the one hand, a positive effect (ln*R*_Biomass \> 0) of biomass was observed in the four species under *AME* and *AIE* conditions except for *S*. *viridis* in *AIE*. However, a significant *AME* \> *AIE* was observed in ln*R*_*Biomass*, indicating that AM fungus was beneficial for plant biomass, but the positive effect was decreased when AM fungi interacted with indigenous microorganisms. On the other hand, a negative effect (ln*R*_Biomass \< 0) was shown in the *IME* condition, indicating that indigenous microorganisms offset the AM fungi promotion in plant growth. Precisely, the results indicated that AM fungi significantly increased the biomass accumulation of four karst pioneer species; however, the ln*R*_*Biomass* reduction by comparing *AIE* to *AME* specified that the indigenous microorganisms offset the benefits of inoculated AM fungi in promoting plant biomass. ## Nitrogen accumulation and its response ratio of four plant species under different microbial treatments The soil microbial condition treatments (Ms) significantly affected N accumulation. Significantly *AMF* \> *AMI* \> *CK* of N accumulation were shown in *A*. *hispidus*, *B*. *pilosa*, and *B*. *tripartita* seedlings, except for *S*. *viridis*. Specifically, the N accumulation was enhanced by AM fungus when comparing *AMF* with *CK* and *AMI* with *CK*, respectively. At the same time, N accumulation under *AMI* was significantly lower than under *AMF*. The plant species (Ps) also significantly affected N accumulation. Under *AMF* and *AMI* treatments, N accumulation in *S*. *viridis* was significantly lower than other three species. For *CK* treatment, N accumulation in *A*. *hispidus* was significantly greater than the other three species. Moreover, there was a non- significant difference in N accumulation between *B*. *pilosa* and *B*. *tripartita* seedlings under any soil microbial condition treatments. Furthermore, the interaction of Ms×Ps significantly affected the N accumulation for the four species. These results showed that *AMF* and *AMI* treatments significantly increased the N accumulation of four karst pioneer species. Meanwhile, N accumulation was significantly different between *A*. *hispidus* and *S*. *viridis*, but not for *B*. *pilosa* and *B*. *tripartita* under *AMF*. Similarly, the soil microbial condition treatments (Ms), the plant species (Ps), and their interaction significantly affected the response ratio of N (ln*R*_*N*). One side has a positive effect (ln*R*_*N* \> 0) of N was observed in four species under *AME* and *AIE* conditions except for *S*. *viridis* in *AIE*. However, a significant *AME* \> *AIE* was observed in ln*R*_*N*, indicating that AM fungus was beneficial for plant N accumulation, but the positive effect was decreased when AM fungi interacted with indigenous microorganisms. Another side has a negative effect (ln*R*_N \< 0) obtainable in the *IME* condition, indicating that indigenous microorganisms offset the AM fungi promotion in N accumulation. Overall, the results indicated that AM fungi significantly increased the N accumulation of four karst pioneer species; however, the ln*R*_*N* reduction by comparing *AIE* to *AME* specified that the indigenous microorganisms offset the benefits of inoculated AM fungi in promoting N accumulation. ## Phosphorous accumulation and its response ratio of four plant species under different microbial treatments The soil microbial condition treatments (Ms) significantly affected P accumulation. Significantly *AMF* \> *AMI* \> *CK* of P accumulation was admissible in four species. Unambiguously, AM fungus enhanced P accumulation when comparing *AMF* with *CK* and *AMI* with *CK*; but the P accumulation under *AMI* was significantly lower than under *AMF*. The plant species (Ps) also significantly affected P accumulation. Under *AMF* and *AMI* treatments, the P accumulation in *S*. *viridis* was significantly lower than other three species. For *CK* treatments, the P accumulation of *A*. *hispidus* was significantly greater than the other three species. In addition, there was no significant difference in P accumulation between *B*. *pilosa* and *B*. *tripartita* seedlings under any microbial condition soil treatments. Meanwhile, the interaction of Ms×Ps significantly affected the P accumulation for four species. It shows that *AMF* and *AMI* treatments significantly increased the P accumulation of four karst pioneer species. Meanwhile, P accumulation was significantly different between *A*. *hispidus* and *S*. *viridis* of Gramineae, except for *B*. *pilosa* and *B*. *tripartita* of Compositae under *AMF*. Likewise, the soil microbial condition treatments (Ms), the plant species (Ps), and their interaction significantly affected the response ratio of P (ln*R*_*P*). Alternatively, it has a positive effect (ln*R*_*P* \> 0) of P on four species under *AME* and *AIE* conditions. However, a significant *AME* \> *AIE* was observed in ln*R*_*P*, indicating that AM fungus was beneficial for plant P accumulation, but the positive effect was decreased when AM fungi interacted with indigenous microorganisms. It also has a negative effect (ln*R*_*P* \< 0) in the *IME* condition and depicts that the indigenous microorganisms offset the AM fungi promotion in P accumulation. Therefore, the results consolidated that AM fungi significantly increased P accumulation of four karst pioneer species, then the ln*R*_*P* reduction by comparing *AIE* to *AME* designated that the indigenous microorganisms offset the benefits of inoculated AM fungi in promoting P accumulation. ## N/P ratio and its response ratio of four plant species under different microbial treatments The soil microbial condition treatments (Ms) significantly affected the N/P ratio, significantly greater N/P ratio between plant species ranked as the *CK* \> *AMF* ≈ *AMI* for *S*. *viridis*, the *AMI* \> *CK* ≈ *AMF* for *A*. *hispidus*, the *AMI* \> *AMF* \> *CK* for *B*. *Pilosa*, and *CK* ≈*AMI* \> *AMF* for *B*. *tripartita*. The plant species (Ps) also significantly affected the N/P ratio, and the N/P ratio for four plants showed species differences under different soil microbial treatments. Explicitly, there was a non- significant difference in the N/P ratio of the four species under *AMF* treatments. Under *AMI* treatments, the N/P ratio of *A*. *hispidus* and *B*. *tripartita* were significantly greater than *S*. *viridis* and *B*. *pilosa*, respectively. In the interim, the N/P ratio of the *B*. *pilosa* was greater than *S*. *viridis* seedlings. Under *CK* treatment, the N/P ratio of *B*. *tripartita* was significantly greater than the other three species, while the N/P ratio of *B*. *pilosa* was significantly lower than the other three species. Likewise, the interaction of Ms×Ps significantly affected the N/P ratio for four species. Therefore, AM fungi significantly reduced the N/P ratio of four species. Equally, the soil microbial condition treatments (Ms), the plant species (Ps), and their interaction significantly affected the response ratio of N/P (ln*R*_*N/P*). Overall, AM fungi significantly reduced the N/P ratio for the four-karst pioneer species, portraying that the AM fungi alleviate P limitation and promote plant growth in karst areas with low P. # Discussion ## AM fungi differently regulated the plant growth and nutrient accumulation AM fungi significantly increased biomass and N and P accumulation for the four karst pioneer species (Figs,). Consistently, the positive influence of AM fungi inoculation on host plant growth and nutrient accumulation was also observed in some previous studies. For instance, He et al. (2017) showed that AM fungi enhanced plant growth and nutrient absorption of *B*. *papyrifera* and *B*. *pilosa* in karst soil, which is consistent with our results that AM fungi significantly increased biomass and accumulation of N and P for the four plants. There are two main mechanisms that AM fungi promote plant growth and nutrient accumulation. One side is that AM fungi can extend the absorbing network beyond the rhizosphere nutrient depletion region and absorb a larger amount of soil mineral nutrients, thereby improving the ability of plants to obtain nutrients and ultimately benefit plant growth. Another is that AM fungi can secrete organic acids and soil enzymes to dissolve the insoluble nutrients and mineralize the organic nutrient, thereby promoting the availability of soil nutrients. Elbon and Whalen (2014) illustrated that AM fungi could increase the plant-available P concentration by secreting organic acids and phosphatase enzymes. Therefore, AM fungi facilitated the growth and nutrient accumulation of four karst pioneer plants, which can verify the hypothesis of H1. However, the specific mechanism of AM fungi affecting nutrient accumulation of karst pioneer species needs to be explored further. The N/P ratio can predict plant nutrient restrictions. A low N/P ratio (\< 14) indicates N limitation, whereas a high N/P ratio (\> 16) indicates P limitation, and both N and P limit plant growth when the N/P ratio is between 14 and 16. In our experiment, the N/P ratio of all species was greater than 16 under *AMI* and *CK* treatments, except for *S*. *viridis* under *AMI* and *B*. *pilosa* under *CK*, showing that plant growth was mainly limited by phosphorus in karst soil. However, the N/P ratio of the four species significantly decreased under *AMF* treatments compared with *AMI* and *CK* treatments for a whole. AM fungi reduced the N/P ratio of seedlings, representing that AM fungus is more effective in assisting plants in obtaining P than N by alleviating P limitation. These results were similar to those of Shen et al. (2020), who suggested that AM fungi alleviated the P limitation of plants via the mycorrhizal network in low-P karst soils. Consequently, the AM fungi play a vital role in alleviating the nutritional restriction of nutrient-deficient karst soils. AM fungi enhanced four plants’ biomass, N, and P accumulation differently. Meanwhile, the *A*. *hispidus*, *B*. *pilosa*, and *B*. *tripartita* obtained greater benefits than the *S*. *viridis* (Figs, and), demonstrating that the promotion effect of AM fungi on plants was different by host type. Besides, the mycorrhizal colonization of *A*. *hispidus*, *B*. *pilosa*, and *B*. *tripartita* was significantly higher than *S*. *viridis*. It was well proof of the different roles of AM fungi on different species, and these differences reflected that AM fungi had the selectivity for host plants. AM fungi showed host-specific growth response and induced differential growth responses in host plant species. It was similar to the research conducted by Liu et al. (2003), who proposed that *Nicotiana tabacum* was a more favorable host plant for *Glomus constrictum* and *Glomus multicaule* to the other hosts. Therefore, AM fungi are crucial for plant growth and nutrient utilization. However, the mutual selection between AM fungi and host plants cannot be ignored, and thus the specific mechanism of selective plant-AMF combinations of karst pioneer species needs to be explored in further study. ## Indigenous microorganisms affected the benefits of AM fungi on plant growth and nutrient accumulation In this experiment, the positive AM fungi effect on plant growth and nutrition was greater than the interactive effect related to AM fungi interacting with indigenous microorganisms for a whole (Figs,). It seems to imply that the indigenous microorganisms offset the benefits of AM fungi on plant growth and nutrient accumulation, signifying a negative relationship between AM fungi and indigenous microorganisms. Previous studies have demonstrated that AM fungi interact with a wide variety of indigenous microorganisms. Meanwhile, AM fungi regulated plant growth positively affected by cooperating with indigenous microorganisms or negatively affected by competing with indigenous microorganisms, which depended on the species of indigenous microorganisms that interact with AM fungi. Positively, Mortimer et al. (2012) presented a synergistic relationship between AM fungi and nitrogen-fixing bacteria showing additive benefits for the growth and nutrient accumulation in the *Acacia cyclops*. Artursson et al. (2006) illustrated that the plant growth-promoting rhizobacteria (PGPR) could enhance the activity of AM during a symbiotic relationship with the host plant. It is because of the stimulatory effects of PGPR on AM growth. Negatively, AM fungi can compete with indigenous microorganisms to produce different effects on plant growth. Some bacteria in the rhizosphere would compete for resources with AM fungi or inhibit the activity of AM fungi, thereby affecting plant growth. It is because indigenous microorganisms have great advantages in colonizing plant roots due to their priority in resources and allocating root space of the host plants compared with colonizers. In addition, Dąbrowska et al. (2014) presented that inoculation AM fungi promoted the growth of plants, but interactive effects of AM fungi with indigenous microorganisms inhibited plant growth. It was similar to our study that AM fungi positively affected plant growth and nutrient accumulation; however, indigenous microorganisms reduced this effect, indicating a negative relationship between AM fungi and indigenous microorganisms. It is possibly caused by the competition between AM fungi and indigenous microorganisms, mainly two sides. One side is interference competition, meaning that some microbes directly inhibit the function of AM fungi via exuding allelochemical substances and bacterial antibiotics. For example, Doumbou et al. (2005) proposed that numerous *Streptomyces* sp. could exude antifungal compounds, thereby inhibiting the function of AM fungi under certain environmental conditions. The other side is resources, and ecological niches competition, which was proposed by Leigh et al. (2011) who suggested that resource competition for decomposition products between AM fungi and bacteria, resulting in an antagonistic relationship between them. Niwa et al. (2018) suggested that the fungus inoculum mainly competed with the indigenous fungi, probably because their life-history strategy was identical to the inoculum fungus. All the above-mentioned can explain why the indigenous microorganisms relieved the benefits of AM fungi on plant growth and nutrient accumulation. It was consistent with Biró et al. (2000), who found the indigenous microflora greatly reduced the functioning of the functioning of the mycorrhizal inoculum. Collectively, indigenous microorganisms offset the benefits of AM fungi in this study, which illustrated the interactions between AM fungi and indigenous microorganisms in karst areas should be mainly a negative relationship, it verified the hypothesis of H2 that indigenous microorganisms offset the benefits of AM fungi on plant growth and nutrient accumulation. However, the specific mechanisms of the negative relationship between specific microorganisms and AM fungi in karst soil remain to be further studied. # Conclusions In this experiment, AM fungi significantly enhanced the biomass, N, and P accumulation for the four species but reduced the N/P ratio partly. AM fungi interacting with indigenous microorganisms increased plant biomass, N, and P accumulation, except for *S*. *viridis* seedlings. However, the benefits from interaction were lower than benefits from AM, indicating that the indigenous microorganisms offset the benefits of AM fungi for host plants. In conclusion, we suggest that the indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated AM fungi for pioneer plants in karst soil. Finally, it is necessary to understand the interactions of AM fungi with indigenous microbial communities to better apply mycorrhizal technology to the degraded ecosystem in karst areas. We thank Xinyang Xu, Lu Gao, Li Wang, Xiaorun Hu and Jingting Li for helping in this experiment. We are grateful to the Institute of Nutrition Resources, Beijing Academy of Agricultural and Forestry Sciences for providing *Glomus mosseae* (NO. BGA0046) for use in our experiments. 10.1371/journal.pone.0266526.r001 Decision Letter 0 Liu Jian Academic Editor 2022 Jian Liu This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 18 Jan 2022 PONE-D-21-40761Indigenous microorganisms relieved the benefits of growth and nutrition regulated by arbuscular mycorrhizal fungi for four pioneer herbs in karst soilPLOS ONE Dear Dr. He, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ============================== ACADEMIC EDITOR: The study is interesting and the manuscript is relatively well written. But the manuscript still have some problems as suggested by the reviewers. The authors should respond to the comments of the reviewers one by one and revise the manuscript accordingly. ============================== Please submit your revised manuscript by Mar 04 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Jian Liu Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at  <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and  <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match.  When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. 3\. Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: In this manuscript, the authors explored and discussed the interactions between Arbuscular mycorrhizal and indigenous microorganisms in regarding to their effects on plant growth and nutrient accumulation. The findings may help elucidate the role of AMF and other soil microorganisms in constructing the plant communities in the Karst area in China. Main suggestions; 1\. ‘Relieve’ usually refers to lightening the pressure, stress, weight, etc. on (something)(<https://www.collinsdictionary.com/us/dictionary/english/relieve>), which is the bad situation of something. Therefore, ‘offset’ is recommend here to replace ‘relieve’. The definition of ‘offset’ is something that counterbalances, counteracts, or compensates for something else; compensating equivalent (<https://www.collinsdictionary.com/us/dictionary/english/offset>). 2\. Th title is suggested as ‘Indigenous microorganisms relieved the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’，as the the native soil microbes and native AMF were not separated in this study. 3\. There are too many English grammar mistakes in the manuscript. It is strongly suggested that the English expresses should be checked through the whole manuscript. Some corrections were made in the manuscript. 4\. Line 353: the treatment description is not consistent with the Methods part. 5\. In the Discussion part, there are too much discussions on the effects of AMF, which were already intensively studied by other researchers. Furthermore, it is better to extend the findings of this study to the mechanisms of ecological processes in the Karst area or how this findings can be applied in the restoration of the vegetation in the Karst area. Reviewer \#2: This article entitled "Indigenous microorganisms relieved the benefits of growth and nutrient regulated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil", provides an interesting work about the effects of mycorrhizal fungi interacting with indigenous microorganisms on plants in degraded soil. The authors claimed that the indigenous microorganisms relieved the benefits of AM fungi in the growth and nutrient absorption of four plants in kast. The topic is very interesting and innovative. The experiment is well done and the writing is good. Some modifications are necessary before the consideration of publication. In general: 1\. How can you give the H2 “Indigenous microorganisms relieved the benefits of AM fungi on 98 plant growth and nutrient accumulation”? it is not enough based your literatures review to deduce this H2. 2\. Why you chose the four species to manipulate the experiment? Please give the reason. 3\. In the discussion you paid more attention on the effect of AMF on plant growth and nutrient absorption. However, I think the combined effects of mycorrhizal fungi and indigenous microorganisms is more important to explanation. Some details: 1\. Line 36: indigenous microbes are inconsistent with line 28. 2\. Line 106-108: do you sure the consistency of soil condition in physicochemical properties in AMF, AMI and CK? It is different in natural soil and sterilized soil in general cognition except for microbes. 3\. Line 118: 10 g Glomus mosseae should being 10 g Glomus mosseae inoculum. 4\. Line 118-119: Did the 10 g Glomus mosseae inoculum include the spore, hyphal and root piece? Please give the information. 5\. Line 152: are you sure this condition is the constant weight of drying? 6\. Line 173-174: Whether the data has been tested for normality and homogeneity of variance before analysis? In the best way, additional description is necessary to ensure the feasibility of statistical data. 7\. Line 154-156: how to calculate the accumulations, can you give us the details about it? 8\. Line 290-291: Change "in negative N/P of " for "in negative N/P ratio of "; Change "in positive N/P of " for "in positive N/P ratio of "; 9\. Line 322: Change " the N/P of " for "the N/P ratio of " 10\. Line 272-273: the indigenous microorganisms relieved the benefits of AM fungi on P accumulation. This sentence is unclear and contradicts the first part of the sentence (AM fungi improved P accumulation). 11\. Line 376: Change "streptomycetes " for " Streptomyces sp." Reviewer \#3: Soil microbial interactions play an important role for plant adaptation in natural habitat. As a kind of beneficial microorganisms, Arbuscular mycorrhizal fungi largely promote growth via the improvement of mineral nutrients for the host plant. This paper attempts to solve the interaction between AM fungi and indigenous microorganisms and explore the benefits of indigenous microorganisms on AM fungi promoting plant growth and nutrient utilization through four karst herbs, which were planted in three different microbial condition soil. The results indicated that the indigenous microorganisms relieved AM fungi's benefits in biomass and nutrient accumulation for plants. I believe this work is interesting and meaningful to apply mycorrhizal technology for restoring in degraded karst areas. However, it still needs to improve in some points as the potential publication of this paper, in detail as follows: 1.Line 92-95: This sentence of “Thus, an experiment was ……with indigenous microorganisms”, is not necessary in the Introduction section. It is better to take it into the Methods section. 2.Line 103: do the “1120m.a.s.l” represent elevation? Please correct it. 3.Line 104: “ soil microbial conditions” should be “ soil microbial condition soil”. 4.Line 110: specify limestone soil as International Soil Classification 5.Line 117-118: I confused the reason about promoting germination rate by yours treatment of 200g soil. Please check it and clear it. 6.Line 121-122: This does not makes sense at all. Are you saying that you added AMF inoculum to your treatment control? If so, that does not constitute a control at all. 7.Line 121-123 This part (starting from "Especially, a 10 g…" and ending on "… a double-layer filter paper") is not clear at all. Please make it clear. 8.Line 302: please correct the citation of He, Jiang et al.(2017). 9.Line 328: please correct the citation of Shen, Cornelissen et al.(2017). Check all reference citations in full text, I think it's not standard. 10.Line 332-335: This sentence was so long, I'm very confused with this result; please make it clear and shorten it. 11.Lin 399: change “we can say that” being “ we suggest that”, delete “Finally”. 12.The discussion needs further refinement and accuracy, comparing your results with previous researches for drawing relevant conclusions. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No Reviewer \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0266526.r002 Author response to Decision Letter 0 2 Feb 2022 Journal Requirements Q1: Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.or>g/ plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and <https://journals.plos.or> g/ plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf RESPONSE: Thank you for your comments. We revised the manuscript to meet PLOS ONE's style requirements. Q2: We note that the grant information you provided in the ‘Funding Information’ and ‘Financial Disclosure’ sections do not match. When you resubmit, please ensure that you provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. RESPONSE: Thank you for your comments. I checked and ensured that we provide the correct grant numbers for the awards you received for your study in the ‘Funding Information’ section. Q3: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. RESPONSE: Thank you for your comments. In the revised version: (1) In the materials and methods section, we added a reference \[43\], see Line106. \(2\) In the discussion section, we added four references \[78\], \[80\], \[81\] and \[82\], see Line 364-365 and Line 367-371. \[43\] He YJ, Cornelissen JHC, Wang P, Dong M, and Ou J. Nitrogen transfer from one plant to another depends on plant biomass production between conspecific and heterospecific species via a common arbuscular mycorrhizal network. Environmental Science and Pollution Research. 2019;26(9):8828-8837. <https://doi.org/10.1007/s11356-019-04385-x>. \[78\] Bender SF, Schlaeppi K, Held A, and Van der Heijden MGA. Establishment success and crop growth effects of an arbuscular mycorrhizal fungus inoculated into Swiss corn fields. Agriculture Ecosystems & Environment. 2019;273:13-24. <https://doi.org/10.1016/j.agee.2018.12.003>. \[80\] Niwa R, Koyama T, Sato T, Adachi K, Tawaraya K, Sato S, et al. Dissection of niche competition between introduced and indigenous arbuscular mycorrhizal fungi with respect to soybean yield responses. Scientific Reports. 2018;8. <https://doi.org/10.1038/s41598-018-25701-4>. \[81\] Hausmann NT and Hawkes CV. Order of plant host establishment alters the composition of arbuscular mycorrhizal communities. Ecology. 2010;91(8):2333-2343. <https://doi.org/10.1890/09-0924.1>. \[82\] Dąbrowska G, Baum C, Trejgell A, and Hrynkiewicz K. Impact of arbuscular mycorrhizal fungi on the growth and expression of gene encoding stress protein–metallothionein BnMT2 in the non‐host crop Brassica napus L. J. Plant Nutr. Soil Sci. 2014;177(3):459-467. <https://doi.org/10.1002/jpln.201300115>. Reviewer \#1: Q1: In this manuscript, the authors explored and discussed the interactions between Arbuscular mycorrhizal and indigenous microorganisms in regarding to their effects on plant growth and nutrient accumulation. The findings may help elucidate the role of AMF and other soil microorganisms in constructing the plant communities in the Karst area in China. RESPONSE: Thank you for your comments. According to your suggestions, we revised the title of the paper to ‘Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’, and revised the conclusion in abstract, result, and conclusion section. Q2: ‘Relieve’ usually refers to lightening the pressure, stress, weight, etc. on (something)(<https://www.collinsdictionary.com/us/dictionary/english/relieve>), which is the bad situation of something. Therefore, ‘offset’ is recommend here to replace ‘relieve’. The definition of ‘offset’ is something that counterbalances, counteracts, or compensates for something else; compensating equivalent (<https://www.collinsdictionary.com/us/dictionary/english/offset>). RESPONSE: Thanks a lot for your good suggestions. We have modified ‘relieve’ to ‘offset’ all in the revised manuscript. see Line36, Line40, Line93, Line216, Line219, Line242, Line245, Line268, Line271 and Line401 of the revision. Q3: The title is suggested as ‘Indigenous microorganisms relieved the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’，as the the native soil microbes and native AMF were not separated in this study. RESPONSE: Thank you for your good suggestions. We have modified the title of the article to ‘Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’, and revised the conclusion in abstract, result, and conclusion section. Q4: There are too many English grammar mistakes in the manuscript. It is strongly suggested that the English expresses should be checked through the whole manuscript. Some corrections were made in the manuscript. RESPONSE: Many thanks for your good comments. We further checked and modified the language and refined expression to the whole manuscript in the new version. Q5: Line 353: the treatment description is not consistent with the Methods part. RESPONSE: Thank you for your comments and questions. We checked and corrected it in Line 350 of the revised version. Q6: In the Discussion part, there are too much discussions on the effects of AMF, which were already intensively studied by other researchers. Furthermore, it is better to extend the findings of this study to the mechanisms of ecological processes in the Karst area or how this findings can be applied in the restoration of the vegetation in the Karst area. RESPONSE: Thank you for your good suggestions. We widely agree with your view that there are too many discussions on the effects of AMF, and it is better to extend the findings of this study to apply them in the restoration of the vegetation in the Karst area. We have checked and revised the Discussion section and Conclusion section carefully and deeply. In order to highlight the emphasis of this paper is not only on the effects of AMF, but we have also enriched the content of the interaction of AM fungi and indigenous microorganisms, and then compare with others, as follows: \(1\) AM fungi regulated plant growth positively affected by cooperating with indigenous microorganisms or negatively affected by competing with indigenous microorganisms. When we discussed‘Negatively affected’section, we added the sentence ‘AM fungi can compete with indigenous microorganisms to produce different effects on plant growth, and clarified possible reason for this occurrence. ‘It is because indigenous microorganisms have great advantages in colonizing plant roots due to their priority in resources and allocating root space of the host plants compared with colonizers’, and the compare with ours. See the specific explanation of Line 364-365 and Line 367-371 in the new version. \(2\) In addition, in the Conclusion section, based on the results of this study, we extended mycorrhizal technology to the degraded ecosystem in karst areas, see Line 402-404 of the revised version. Reviewer \#2: Q1: This article entitled "Indigenous microorganisms relieved the benefits of growth and nutrient regulated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil", provides an interesting work about the effects of mycorrhizal fungi interacting with indigenous microorganisms on plants in degraded soil. The authors claimed that the indigenous microorganisms relieved the benefits of AM fungi in the growth and nutrient absorption of four plants in kast. The topic is very interesting and innovative. The experiment is well done and the writing is good. Some modifications are necessary before the consideration of publication. RESPONSE: Thank you for your comments. We have completely revised the manuscript in the new version. In the revised version: (1) According to the suggestions of Reviewer \#1, we revised the title of the paper to ‘Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’, and revised the conclusion in abstract, result and conclusion section. \(2\) In order to better present the results of this paper, we added some examples about the relationship between AM fungi and indigenous microorganisms in the discussion section to combine the explanation, see Line 364-365 and Line 367-371 of the revision. Q2: How can you give the H2 “Indigenous microorganisms relieved the benefits of AM fungi on 98 plant growth and nutrient accumulation”? it is not enough based your literatures review to deduce this H2. RESPONSE: Thanks a lot for your good suggestions. In the revised version: (1) According to the comments of Reviewer \#1, we modified H2 to “Indigenous microorganisms offset the benefits of inoculated AM fungi on plant growth and nutrient accumulation”. (2) We have added appropriate discussion in revied version to deduce H2“Indigenous microorganisms offset the benefits of inoculated AM fungi on plant growth and nutrient accumulation”, by that AM fungi can compete with indigenous microorganisms to produce different effects on plant growth \[1\] ……. it is because indigenous microorganisms have great advantages in colonizing plant roots due to their priority in resources and allocating root space of the host plants compared with colonizers \[2,3\]. In addition, Dąbrowska et al. (2014) \[4\] presented that inoculation AM fungi promoted the growth of plants, but in the soil with indigenous microorganisms, growth inhibition after inoculation was observed compared to the control. It was similar to our study that……; see Line 364-365 and Line 367-371 of revised version. \[1\] Bender SF, Schlaeppi K, Held A, and Van der Heijden MGA. Establishment success and crop growth effects of an arbuscular mycorrhizal fungus inoculated into Swiss corn fields. Agriculture Ecosystems & Environment. 2019;273:13-24. <https://doi.org/10.1016/j.agee.2018.12.003>. \[2\] Niwa R, Koyama T, Sato T, Adachi K, Tawaraya K, Sato S, et al. Dissection of niche competition between introduced and indigenous arbuscular mycorrhizal fungi with respect to soybean yield responses. Scientific Reports. 2018;8. <https://doi.org/10.1038/s41598-018-25701-4>. \[3\] Hausmann NT and Hawkes CV. Order of plant host establishment alters the composition of arbuscular mycorrhizal communities. Ecology. 2010;91(8):2333-2343. <https://doi.org/10.1890/09-0924.1>. \[4\] Dąbrowska G, Baum C, Trejgell A, and Hrynkiewicz K. Impact of arbuscular mycorrhizal fungi on the growth and expression of gene encoding stress protein–metallothionein BnMT2 in the non‐host crop Brassica napus L. J. Plant Nutr. Soil Sci. 2014;177(3):459-467. <https://doi.org/10.1002/jpln.201300115>. Q3: Why you chose the four species to manipulate the experiment? Please give the reason. RESPONSE: Thanks a lot for your comments. In our primary field investigations, the Gramineae species Setaria viridis vs. Arthraxon hispidus and Compositae Bidens pilosa vs. Bidens tripartita are successive pioneer species of karst communities as the herbaceous stage, which generally coexist in the same habitat as the main Gramineae and Compositae. In addition, A. hispidus and S. viridis are of the same family but different genera, while B. pilosa and B. tripartita have a common family and genera. Therefore, we also wanted to investigate whether AM fungi have different effects on different or the same taxonomic level of species. Of course, our results show that the biomass and nutrients of N and P were significantly different between A. hispidus and S. viridis of Gramineae, but not for B. pilosa and B. tripartita of Compositae under AMF. Q4: In the discussion you paid more attention on the effect of AMF on plant growth and nutrient absorption. However, I think the combined effects of mycorrhizal fungi and indigenous microorganisms is more important to explanation. RESPONSE: Thank you for your good suggestions. Yes, the combined effects of mycorrhizal fungi and indigenous microorganisms are more important to explain. AM fungi regulated plant growth positively affected by cooperating with indigenous microorganisms or negatively affected by competing with indigenous microorganisms. In the original manuscript in Line 354-387, we reviewed some literature about the combined effects of mycorrhizal fungi and indigenous microorganisms, including positive and negative to lead to our results, and discussed that the offset of the role of AM fungi by indigenous microorganisms may be caused by competition. In order to better explain, we added some literature to complement. See Line 364-365 and Line 367-371 of revised version. Q5: Line 36: indigenous microbes are inconsistent with line 28. RESPONSE: Thank you for your comments. We have corrected it in Line 36 of the revised version. Q6: Line 106-108: do you sure the consistency of soil condition in physicochemical properties in AMF, AMI and CK? It is different in natural soil and sterilized soil in general cognition except for microbes. RESPONSE: Thank you for your comments and questions. Here, we measured the soil quality, see the description of that the PH 8.2, total nitrogen (TN) 0.622 g, alkaline hydrolysis nitrogen (AN) 0.315 g, total phosphorus (TP) 1.274 g, available phosphorus (AP) 0.163 g, total potassium (TK) 37.79 g, and available potassium (AK) 0.532 g. Yes, autoclaving sterilization satisfied the requirements of sterilization, but affect some of the basic properties, including organic matter, specific surface area, PH, cation-exchange capacity, free iron/aluminum oxides and zero point of charge of the soils \[1\]. However, these basic properties affected by sterilization are not affected the research content of our article. In our experiment, we studied the role of indigenous microorganisms in affecting the growth and nutritional functions of plants regulated by AM fungi. In addition, there were strictly controlled experiments by the AMF and AMI (with AM fungus) and CK treatment (without AM fungus). Therefore, we pay more attention to chemical properties, and it is consistent in AMF, AMI and CK. \[1\] Zhang H, Zhang J, Zhao B, Zhang C, and Zhang Y. Influence of autoclaving sterilization on properties of typical soils in China. Acta Pedologica Sinica. 2011;48(3):540-548. Q7: Line 118: 10 g Glomus mosseae should being 10 g Glomus mosseae inoculum. RESPONSE: Thanks a lot for your comments. We have corrected it in Line 115 and Line 116 of the revised version. Q8: Line 118-119: Did the 10 g Glomus mosseae inoculum include the spore, hyphal and root piece? Please give the information. RESPONSE: Thank you for your comments and questions. Yes, the 10g Glomus mosseae inoculum includes the spore (above 100 spores per gram of soil), hyphae and colonized root pieces. For clarity, we deleted Line 145-147 in the original manuscript and added the information of 10 g Glomus mosseae inoculum in Line 121-123 of the revised version. Q9: Line 152: are you sure this condition is the constant weight of drying? RESPONSE: Thanks a lot for your comments. It is our negligence. We have checked carefully and corrected it in Line 148 of the revised version. Q10: Line 173-174: Whether the data has been tested for normality and homogeneity of variance before analysis? In the best way, additional description is necessary to ensure the feasibility of statistical data. RESPONSE: Thank you very much for your comments. Here in Statistical Analysis, we added the description by the sentence of “All of the data were tested for normality and homogeneity of variance before analysis”, see Line 170-171 of the revised version. Q11: Line 154-156: how to calculate the accumulations, can you give us the details about it? RESPONSE: Thank you for your comments and questions. In fact, we have given the details about the calculation of the accumulations in the original manuscript. Specifically, the nutrient concentrations of nitrogen and phosphorus of plant tissues of root and stem and leaf were determined. Further, the plant tissue accumulations of nitrogen and phosphorus were calculated respectively using nutrient concentration multiplying by biomass, then plant individual accumulations were accumulated by root and stem and leaf. Here, we revised the details about the calculation of accumulations. See Line 151-153 in the new version. Q12: Line 290-291: Change "in negative N/P of " for "in negative N/P ratio of "; Change "in positive N/P of " for "in positive N/P ratio of "; RESPONSE: Thank you for your comments. We already corrected it, see Line 288 and Line 289 of the revision. Q13: Line 322: Change " the N/P of " for "the N/P ratio of " RESPONSE: Thanks a lot for your comments. We have corrected it in Line 320 of the revised version. Q14: Line 272-273: the indigenous microorganisms relieved the benefits of AM fungi on P accumulation. This sentence is unclear and contradicts the first part of the sentence (AM fungi improved P accumulation). RESPONSE: Thank you for your comments. In fact, this is not contradictory. AM fungi can promote P accumulation in four karst pioneer species, however, indigenous microorganisms offset the benefits of inoculated AM fungi in promoting P accumulation. Of course, in order to express clearer, we corrected this sentence, see Line 271 of the revision. Q15: Line 376: Change "streptomycetes " for " Streptomyces sp." RESPONSE: Thank you for your comments. We already corrected it, see Line 378 of the revision. Reviewer \#3: Q1: Soil microbial interactions play an important role for plant adaptation in natural habitat. As a kind of beneficial microorganisms, Arbuscular mycorrhizal fungi largely promote growth via the improvement of mineral nutrients for the host plant. This paper attempts to solve the interaction between AM fungi and indigenous microorganisms and explore the benefits of indigenous microorganisms on AM fungi promoting plant growth and nutrient utilization through four karst herbs, which were planted in three different microbial condition soil. The results indicated that the indigenous microorganisms relieved AM fungi's benefits in biomass and nutrient accumulation for plants. I believe this work is interesting and meaningful to apply mycorrhizal technology for restoring in degraded karst areas. However, it still needs to improve in some points as the potential publication of this paper, in detail as follows: RESPONSE: Thank you very much for your comments. We have completely revised the manuscript in the new version. Further, according to the suggestions of Reviewer \#1, we revised the title of the paper to ‘Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil’, and revised the conclusion in abstract, result and conclusion section. Q2: Line 92-95: This sentence of “Thus, an experiment was ……with indigenous microorganisms”, is not necessary in the Introduction section. It is better to take it into the Methods section. RESPONSE: Thanks a lot for your good suggestions. We have deleted Line 92-95 in the Introduction section in the original manuscript, and these contents have been presented in the Methods section in the original manuscript. Q3: Line 103: do the “1120m.a.s.l” represent elevation? Please correct it. RESPONSE: Thank you for your comments and questions. We have corrected it in Line 99 of the revised version. Q4: Line 104: “soil microbial conditions” should be “soil microbial condition soil”. RESPONSE: Thanks a lot for your comments. We already corrected it, see Line 100 of the revision. Q5: Line 110: specify limestone soil as International Soil Classification RESPONSE: Thank you for your good suggestions. The soil substrate was used by limestone in our experiment, according to your suggestion, we added the soil classification basis of FAO in Line 106, and changed the description of soil substrate. Q6: Line 117-118: I confused the reason about promoting germination rate by yours treatment of 200g soil. Please check it and clear it. RESPONSE: Thank you for your comments and questions. The three basic conditions for seed germination are appropriate temperature, appropriate water and sufficient air. Specifically, in our experiment, after covering the soil, the seeds can be kept germinating slowly and under a certain humidity. In order to express more clearly, we have modified this sentence, see Line 114 of the revision. Q7: Line 121-122: This does not makes sense at all. Are you saying that you added AMF inoculum to your treatment control? If so, that does not constitute a control at all. RESPONSE: Thank you for your comments and questions. In fact, we have given the details about added sterilized inoculum of Glomus mosseae to treatment control. The equal amount of sterilized inoculum and 10 ml of filtrate taken from sterilized inoculum were added in CK treatment, in order to maintain the consistency of microflora except for target fungi Glomus mosseae. Here, we revised the details about added sterilized inoculum of Glomus mosseae to treatment control. See Line 117-120 in the new version. Q8: Line 121-123 This part (starting from "Especially, a 10 g…" and ending on "… a double-layer filter paper") is not clear at all. Please make it clear. RESPONSE: Thanks a lot for your comments and suggestions. Regarding the long sentence, we have revised and made it clearer in the revised version, see Line 117-120. Q9: Line 302: please correct the citation of He, Jiang et al.(2017). RESPONSE: Thank you for your comments and questions. We already corrected it, see Line 300 of the revision. Q10: Line 328: please correct the citation of Shen, Cornelissen et al.(2017). Check all reference citations in full text, I think it's not standard. RESPONSE: Thank you for your comments and questions. We already checked and corrected all reference citations in full text, see Line 58, Line 60, Line 62, Line 71, Line 73, Line 78, Line 80, Line 162, Line 300, Line 309, Line 313, Line 326, Line 337, Line 359, Line 377, Line 380, Line 382 and Line 386 of the revision. Q11: Line 332-335: This sentence was so long, I'm very confused with this result; please make it clear and shorten it. RESPONSE: Thanks a lot for your comments and suggestions. Regarding the long sentence, we have revised and made it clearer and shorter in the revised version, see Line 330-332. Q12: Lin 399: change “we can say that” being “we suggest that”, delete “Finally”. RESPONSE: Thanks a lot for your good suggestions. We have deleted “Finally”, and have corrected it in Line 401 of the revised version. Q13: The discussion needs further refinement and accuracy, comparing your results with previous researches for drawing relevant conclusions. RESPONSE: Thanks a lot for your good suggestions. We have checked and revised the Discussion section carefully and deeply. On the whole, in order to make our points clear, the idea of revision was to explain and analyze the main research points directly, and then compare with others, as follows: \(1\) AM fungi regulated plant growth positively affected by cooperating with indigenous microorganisms or negatively affected by competing with indigenous microorganisms. In the original manuscript in Line 354-387, we reviewed some literature about the combined effects of mycorrhizal fungi and indigenous microorganisms, including positive and negative to lead to our results, and drew relevant conclusions that the indigenous microorganisms offset the benefits of inoculated AM fungi in biomass and nutrient accumulation for pioneer plants in the karst habitat. \(2\) In order to better clarify, we added some literature to complement in Negatively section, we added the sentence ‘AM fungi can compete with indigenous microorganisms to produce different effects on plant growth, and clarified possible reason for this occurrence. ‘It is because indigenous microorganisms have great advantages in colonizing plant roots due to their priority in resources and allocating root space of the host plants compared with colonizers’, and the compare with ours. See the specific explanation of Line 364-365 and Line 367-371in the new version. 10.1371/journal.pone.0266526.r003 Decision Letter 1 Liu Jian Academic Editor 2022 Jian Liu This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 28 Feb 2022 PONE-D-21-40761R1Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soilPLOS ONE Dear Dr. He, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ============================== ACADEMIC EDITOR: The revised version has been improved a lot.  But the manuscript still has some problems as suggested by the reviewer. ============================== Please submit your revised manuscript by Apr 14 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Jian Liu Academic Editor PLOS ONE Journal Requirements: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#2: All comments have been addressed Reviewer \#3: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: No Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: 1. Line 37-39. What is the purpose to compare the growth status of the species in this experiment? Is it essential for this topic? 2\. Line 93-94. What were the evidences to support this hypothesis before this research was conducted? 3\. Please add sub-headlines for the Discussion part. It is not clear what is the central topic for each paragraph. Still，there are too many discussions on the roles of AMF on plant growth, which were not the central topic of this study. 4\. Line 396-399. The first argument is self-contradictory with the following statement. 5\. Line 399-401. What is the significance of this finding? 6\. Some grammar mistakes and English expressions are corrected in the tracked PDF. Reviewer \#2: Thanks for the authors. I think all the comments have been addressed so far. I have no other questions. Reviewer \#3: All comments were addressed. In this edition, the results and discussion were reorganized and now are clear for readers. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: No Reviewer \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0266526.r004 Author response to Decision Letter 1 3 Mar 2022 Journal Requirements Q1: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. RESPONSE: Thank you for your comments. According to the suggestions of Reviewer \#1, we refined the discussion part and deleted five references \[55\], \[56\], \[57\] \[69\], \[70\] in the original manuscript, as follows: \[55\] Tinker PB and Nye PH, Solute movement in the rhizosphere. 2000: Oxford University Press. \[56\] Leigh J, Hodge A, and Fitter AH. Arbuscular mycorrhizal fungi can transfer substantial amounts of nitrogen to their host plant from organic material. New Phytol. 2009;181(1):199-207. <https://doi.org/10.1111/j.1469-> 556 8137.2008.02630.x. 557 \[57\] Shao YD, Hu XC, Wu QS, Yang TY, Srivastava AK, Zhang DJ, et al. Mycorrhizas promote P acquisition of tea plants through changes in root morphology and P transporter gene expression. S. Afr. J. Bot. 2021;137:455-462. <https://doi.org/10.1016/j.sajb.2020.11.028>. \[69\] Danuso F, Zanin G, and Sartorato I. A modelling approach for evaluating phenology and adaptation of two congeneric weeds (Bidens frondosa and Bidens tripartita). Ecol. Model. 2012;243:33-41. <https://doi.org/10.1016/j.ecolmodel.2012.06.009> 589 \[70\] Bartolome AP, Villaseñor IM, and Yang WC. Bidens pilosa L.(Asteraceae): botanical properties, traditional uses, phytochemistry, and pharmacology. Evid- Based Compl. Alt. 2013;2013. <https://doi.org/10.1155/2013/340215>. Reviewer \#1: Q1: Line 37-39. What is the purpose to compare the growth status of the species in this experiment? Is it essential for this topic? RESPONSE: Thank you for your comments. We checked and agreed with your view that it is not essential for this topic in the Abstract section, so we deleted Line 37-39 in the original manuscript. Q2: Line 93-94. What were the evidences to support this hypothesis before this research was conducted? RESPONSE: Thank you for your comments and questions. In the revised version: (1) we revised the original summary between AM fungi and indigenous microorganisms to “Thus, the cooperation and competition between AM fungi and indigenous microorganisms are ineluctability in karst soil” See Line80-81 of the revision. \(2\) we added the two sentences about previous studies as evidences to support the hypothesis, see Line 91-92 and Line 93-95 of the revision. Q3: Please add sub-headlines for the Discussion part. It is not clear what is the central topic for each paragraph. Still，there are too many discussions on the roles of AMF on plant growth, which were not the central topic of this study. RESPONSE: Thank you for your good suggestions. We added two sub-headlines for the Discussion part; see Line 297 and Line 340-341 of the revision. In addition, we further refined the Discussion section, please see the new version. Q4: Line 396-399. The first argument is self-contradictory with the following statement. RESPONSE: Many thanks for your comments and questions. We have checked and revised carefully, in order to express clearer, we modified “while the indigenous microorganisms offset the benefits of AM fungi foe host plants” to “However, the benefits from interaction were lower than benefits from AM, indicating that the indigenous microorganisms offset the benefits of AM fungi for host plants”. See Line 390 of the revision. Q5: Line 399-401. What is the significance of this finding? RESPONSE: Thanks a lot for your comments. We checked and agreed with your view that it is not essential for this topic in this manuscript, and we thought it is not the significance of this finding in this manuscript. Thus, we deleted Line 399-401 in the original manuscript. Q6: Some grammar mistakes and English expressions are corrected in the tracked PDF. RESPONSE: Thank you for your good suggestions. We further checked the language and refined expression, and modified some grammar mistakes and English expressions in the whole manuscript in the new revision according to your suggestions. 10.1371/journal.pone.0266526.r005 Decision Letter 2 Liu Jian Academic Editor 2022 Jian Liu This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 6 Mar 2022 PONE-D-21-40761R2Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soilPLOS ONE Dear Dr. He, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. ============================== ACADEMIC EDITOR: The revised version has been improved a lot. All the comments have been addressed. But the authors still need to polish the language and revise the language errors. For example: Line 390 : “the benefits form” should be “the benefits from”. ============================== Please submit your revised manuscript by Apr 20 2022 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Jian Liu Academic Editor PLOS ONE Journal Requirements: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Additional Editor Comments (if provided): The revised version has been improved a lot. All the comments have been addressed. But the authors still need to polish the language and revise the language errors. For example: Line 390 : “the benefits form interaction” should be “the benefits from interaction” \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0266526.r006 Author response to Decision Letter 2 14 Mar 2022 Journal Requirements Q1: The revised version has been improved a lot. All the comments have been addressed. But the authors still need to polish the language and revise the language errors. For example: Line 390 : “the benefits form interaction” should be “the benefits from interaction” RESPONSE: Thank you for your comments and questions. We revised this sentence of “the benefits form interaction” being “the benefits from interaction”, see Line 383 of the revision. In addition, we further checked and modified the language and revised the language errors. Please see the new version. 10.1371/journal.pone.0266526.r007 Decision Letter 3 Liu Jian Academic Editor 2022 Jian Liu This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 23 Mar 2022 Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil PONE-D-21-40761R3 Dear Dr. He, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. Kind regards, Jian Liu Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0266526.r008 Acceptance letter Liu Jian Academic Editor 2022 Jian Liu This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 14 Apr 2022 PONE-D-21-40761R3 Indigenous microorganisms offset the benefits of growth and nutrition regulated by inoculated arbuscular mycorrhizal fungi for four pioneer herbs in karst soil Dear Dr. He: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. If we can help with anything else, please email us at <plosone@plos.org>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Jian Liu Academic Editor PLOS ONE [^1]: NO. The authors have declared that no competing interests exist.

# Introduction For over 50 years vitamin K antagonists (VKAs) have been widely used, not only as (first choice) treatment for thromboembolism, but as primary and secondary prevention of (venous) thromboembolism as well. Warfarin is currently the most prescribed VKA followed by acenocoumarol and phenprocoumon. The predominant adverse effect of anticoagulant therapy is an increased risk of bleeding which can lead to morbidity and mortality. Annually approximately 1 to 4% of patients treated with VKAs suffer from major bleeding episodes. Clinically relevant bleeding occurs in up to 20% of patients. The risk of bleeding increases with age. Patients that are older than 75 years, experience major bleeding more frequently than younger patients: 5.1% versus 1% per year, respectively. This bleeding risk increases even more when VKAs are combined with antiplatelet therapy. In the past several attempts were made to more accurately estimate the bleeding risk of individual patients treated with VKAs. One of the commonly used clinical methods for the identification of patients with atrial fibrillation at risk for bleeding is the HAS-BLED score, which is a clinical decision score. The HAS-BLED score contains the risk factors hypertension, abnormal renal/liver function, stroke, bleeding history or predisposition, labile international normalized ratio (INR), elderly (age ≥ 65 years) and drugs/alcohol (ab)use concomitantly. Although the HAS-BLED score is developed and validated only in patients with AF, it would be reasonable to think that the score could be applied in patients with different indications for VKA use, considering the comparable risk factors for bleeding. Moreover, the HAS-BLED score has the highest predictive potential compared to other clinical prediction scores; however its accuracy differed based on the cohort used for validation. As of yet there are no laboratory methods that prospectively predict which patients are at risk for bleeding. Considering the INR, there is an increased risk of bleeding at higher INR levels, yet the majority of bleeding events occurs in patients that are within the therapeutic range. Thrombin generation, a method that detects the enzymatic activity of thrombin, has been shown to be able to detect both prothrombotic and bleeding phenotypes based on changes in the coagulation system. Additionally, thrombin generation has the capacity to detect the anticoagulant effect of many if not all anticoagulants, including VKAs and direct oral anticoagulants (DOACs). Until recently this method was only applicable in plasma due to quenching of the fluorescent signal by sedimentation of erythrocytes. Introduction of a porous matrix, preventing this sedimentation, and using a different thrombin-sensitive substrate enabled studying thrombin generation in whole blood. In this study we investigated whether thrombin generation, in plasma or whole blood, could be used to predict bleeding episodes in 129 patients taking VKAs and compared these parameters to the INR, the HAS-BLED score, fibrinogen levels and other factor determinations. # Materials & methods ## Study population Patients taking VKAs were randomly included in this study between March 2012 and October 2013. A sample size of 127 was sufficient to provide power of 80% with a two-sided α-level of 0.05. Patients were eligible for inclusion in this study when treated with VKAs for longer than three months and undergoing a venapuncture in order to determine their INR value at the Maastricht anticoagulation clinic. Patients under 18 years of age were excluded. All patients were informed and provided written consent. The study was approved by the local medical ethical committee (Medisch-ethische toetsingscommissie academisch ziekenhuis Maastricht/universiteit Maastricht (METC aZM/UM), approval number: 11-4-142.4/ccl). ## Blood samples Patients were included in the anticoagulation clinic during one of their routine checkups for VKA treatment. After giving informed consent on this day of inclusion, blood was drawn for both INR determination and further laboratory assays necessary for this study, such as thrombin generation. Blood was collected through antecubital venapuncture using citrate tubes (1 volume of trisodium citrate 3.2% to 9 volumes of blood), (BD Vacutainer system, Roborough, Plymouth, UK). Platelet rich plasma (PRP) was prepared by centrifuging the blood at 250g for 15 min. Platelet poor plasma (PPP) was prepared by double centrifugation at 2,000g for 10 min. PPP for thrombin generation was used immediately, the remainder was stored at -80°C until bulk analysis in other assays was possible. ## Follow-up Bleeding episodes were recorded during the follow-up period (minimum four months, mean follow-up time: 15.5 months) at the anticoagulation clinic in Maastricht according to the definitions and criteria defined by the Dutch Federation of Anticoagulation Clinics. These criteria are based on the guidelines by the Scientific and Standardization Committee of the International Society on Thrombosis and Haemostasis. The definitions of major bleeding are: any intracranial hemorrhage, any objectively diagnosed intra-articular hemorrhage and bleeding leading to death, transfusion, surgery, and/or hospital admission. Minor bleeding was defined as all other clinically relevant bleeding not meeting the definition for major bleeding. Clinically relevant bleeding was defined according to the guidelines of the Dutch Federation of Anticoagulation Clinics. Bleeding events, when detected by a general practitioner or at the hospital, were systematically reported to the anticoagulation clinic. Minor bleedings were mostly reported by patients themselves, during routine visits and only rated if they were clinically relevant. The specifications of the types of bleeding (minor/major) in our study are listed in. We also recorded age, sex, type of VKA, time in therapeutic range (TTR) and indication for VKA use. The use of antithrombotic therapy as well as other co- medications affecting coagulation (e.g. non-steroidal anti-inflammatory drugs (NSAIDs)) was also documented to assess potential confounding effects. Bleeding episodes were recorded until the end of the study or until discontinuation of treatment. Other parameters were determined on the day of inclusion. ## Reagents Synthetic phospholipids were obtained from Avanti Polar Lipids Inc. (Alabaster, AL, USA). Recombinant tissue factor (TF) known as Innovin (Dade-Behring, Marburg, Germany) was used. Z-Gly-Gly-Arg-aminomethylcoumarine (ZGGR-AMC) was purchased from Bachem (Basel, Switzerland). Rhodamine substrate (P₂Rho) was a gift of Diagnostica Stago (Asnières sur Seine, France). Recombinant human thrombomodulin (TM) was a kind gift of Asahi Kasei Pharma (Japan). The calibrator, α2-macroglobulin-thrombin complex, was prepared as described previously. Hepes buffers containing 5 mg/ml or 60 mg/ml bovine serum albumin were prepared as described by Hemker et al.. ## Thrombin generation CAT was performed in plasma as described earlier. TF was used to initiate the reaction at a final concentration of 1 pM in PRP and at both 1 pM and 5 pM with 4 μM phospholipids in PPP. The effect of TM was also tested in PPP (1 pM TF) at a concentration of 2.5 nM. In PPP (5 pM TF) and PRP 20 nM of TM was used (around IC₅₀). The whole blood CAT technique was performed according to our group’s earlier specifications. Thirty microliters of blood were mixed with 10 μl of P2Rho substrate (1.8 mM) and 20 μl of TF/CaCl₂ solution were added, which initiated thrombin generation. In the calibration wells, 20 μl of reagents were replaced with calibrator (final concentration: 100 nM). Instantly after the activation, 5 μl of the mixture were pipetted on paper disks (Whatman 589/1, Whatman GmbH, Dassel, Germany) in a flat bottom 96-well polystyrene plate and covered with 40 μl of mineral oil (Affymetrix, USB, Cleveland, Ohio, USA). The final TF concentration was 1 pM and TM (20 nM) was added as well. Fluorescent signals were measured using the Fluoroskan Ascent software (Thermo Labsystems, Helsinki, Finland). Measurements were performed in triplicate and fluorescent signals were transformed into thrombin concentrations as described by Hemker et al.. The thrombin generation parameters which were analyzed were: lag time, which is the time until the first traces of thrombin are formed; ETP, the area under the thrombin generation curve, peak level of thrombin formation and the time-to-peak or the time until the thrombin peak is reached. ## Additional analyses Hematocrit and hemoglobin concentration determinations were performed in citrated blood with a Coulter Counter analyzer (Beckman Coulter, Woerden, the Netherlands). Levels of clotting factors, including fibrinogen were assessed using the STA-R Evolution analyzer (Stago, Asnières sur Seine, France). Factor II, V, VII, VIII, IX, X levels were determined with clotting assays triggered by either a thromboplastin based reagent (FII, FV, FVII, FX) or a kaolin based reagent (FVIII and FIX). Fibrinogen levels were measured using the Clauss method. Protein C activity was determined by an aPTT based assay, activated by Agkistrodon c. contortrix venom. Protein S was tested in a clotting assay in which the activity of protein S as a cofactor of protein C is measured by its effect on factor Va. Antithrombin (ATIII) was determined by a chromogenic measurement. INRs were determined in the local anticoagulation clinic. The prothrombin time was determined in citrated plasma with an automated coagulation analyzer (Sysmex CA 1500, Siemens Diagnostics, the Netherlands) using Innovin® (Dade-Behring) as the thromboplastin reagent. The INR value was expressed as the ratio of the subject’s PT to a normal (control) sample raised to the power of the International Sensitivity Index (ISI); (PT_test/PT_normal)^ISI. HAS-BLED scores of each patient were calculated post-hoc by a blinded physician using medical records, allotting one point for each risk factor. ## Statistical analysis All analyses were performed using Graphpad Prism version 5.00 (Graphpad Software Inc., La Jolla, CA, USA). Patients with missing data were not excluded from the analysis. Correlation analysis of whole blood versus plasma CAT parameters was performed using the Pearson correlation test. Patients were divided into two groups (with bleeding and without bleeding). Differences between groups were analyzed via the Mann Whitney U test and represented by medians with interquartile ranges (IQR), range from minimum value to maximum value and 95% confidence intervals (CI). Receiver operating curves (ROC) were used to investigate the ability of WB CAT and the HAS-BLED score to discriminate between bleeding and non-bleeding patients. The area under the ROC curve (95% CI) quantified the predictive value of parameters. Differences between two groups (other than bleeding versus non-bleeding) were also analyzed using the Mann Whitney U test. A two-sided p-value of ≤ 0.05 was considered statistically significant. # Results ## Patient characteristics One hundred and fifty patients were eligible for the study and 21 patients had to be excluded for several reasons: failed blood collection, not fulfilling the inclusion criterion of using VKA for at least three months, technical problems during measurements or other reasons. The demographics of the remaining 129 patients are listed in. The average duration of VKA treatment until the inclusion date was approximately 5 years. Therapeutic ranges consisted of INR’s from 2.0–3.5 (n = 103) or 2.5–4.0 (n = 26), depending on the indication. ## Bleeding episodes In our study we found that 26 patients (20.2%) suffered from 40 clinically relevant bleeding episodes during a mean follow-up of 15.5 months after inclusion. The mean time between inclusion (including testing) and bleeding was 9.8 months. The bleeding rates in male (n = 121) and female (n = 28) patients were comparable (20% and 21% bleeding, respectively). Seventeen patients had one bleeding episode, six patients experienced two bleeding events, two patients had three bleeding episodes and one patient suffered five times from a clinically relevant bleeding. Patients experienced different types of bleeding. In our population patients mainly experienced conjunctival eye bleeds. Four major bleeding episodes occurred during the follow-up period. Two patients experienced severe digestive tract bleedings and one patient suffered twice from severe nosebleeds. ## Thrombin generation Analyzing samples with whole blood CAT a significantly lower ETP and peak was found in the patients that suffered from bleeding compared to patients that did not have this adverse effect. Differences in ETP and peak remained statistically significant in the presence of TM, although to a lesser degree (median \[IQR\] (CI) ETP: 134.9 \[104.7–193\] (126.4–169.9) versus 174.3 \[129.5–222.9\] (169.0–195.7), p = 0.009; peak: 20.18 \[13.74–30.39\] (18.0–28.9) versus 27.72 \[18.08–37.49\] (27.1–37.0), p = 0.033 (n = 25 and n = 102)), in bleeding versus non-bleeding patients respectively). The lag time and time-to-peak did not differ significantly between bleeding and non-bleeding patients. A receiver operating curve (ROC) was constructed for ETP and peak determined in whole blood. Assessment of the area under the curve (AUC) of the ROC demonstrated that both ETP and peak were significantly associated (AUC (CI) ETP: 0.700 (0.584–0.816), p = 0.002 (n = 25) and AUC (CI) peak: 0.642 (0.516–0.767), p = 0.029 (n = 102), respectively) with the bleeding tendency. In plasma no significant differences were detected for the thrombin generation parameters between patients that suffered from bleeding and those without bleeding events, although a trend for a decreased ETP and peak height was observed in the bleeding population. Similarly, no significant differences were found in the lag time and time-to-peak between both populations. The addition of thrombomodulin did not change these results. We assessed whether co-medication affecting platelets could have an effect on the ETP and peak in whole blood. Twenty-nine patients used co-medication that affects platelets (P2Y12-inhibitors, acetylsalicylic acid and/or non-steroidal anti-inflammatory drugs) of which 12 (41.4%) suffered from bleeding and 17 patients did not. From it can also be calculated that 13.4% displayed bleeding when not using co-medication. We found that the use of these co-medications did not influence the ETP and peak measured in whole blood. Therefore our analyses were not corrected for this type of co-medication. The parameters in whole blood CAT (1 pM TF) were compared to the parameters (ETP, peak, lag time and time-to-peak) in the plasma CAT (PRP at 1 pM TF and PPP at 1 and 5 pM TF). All thrombin generation parameters displayed a significant correlation between plasma and whole blood measurements and the highest correlation coefficients were established for the ETP and peak. The best correlation for the four parameters was found in PPP using 5 pM TF. Regarding the INR, similar correlations were observed with the whole blood CAT as compared to plasma CAT (inverse, hyperbolical correlation of ETP and peak with INR, linear correlation of lag time and time-to-peak with INR). ## Additional analyses Bleeding events were not associated with a difference in age (median \[IQR\]: 70 \[60–80\] for both groups (CI bleeding: 60–70 versus non-bleeding: 70–70), p = 0.4074 (n = 26 and n = 103)). Patients with and without bleeding events did not significantly differ regarding the INR, hematocrit, hemoglobin levels, or fibrinogen concentration ( information). Differences in levels of other coagulation and anticoagulant factors were determined, but none of them reached statistical significance. ## HAS-BLED HAS-BLED scores in patients with bleeding episodes were significantly higher than in patients that did not bleed (median \[IQR\] (CI): 3 \[2–3.25\] (2.3–3.1) versus 2 \[1–3\] (1.9–2.3), p = 0.003 (n = 26 and n = 103)). A ROC curve was constructed resulting in an AUC of 0.682 (CI: 0.571–0.792) (p = 0.004). Additionally, the HAS-BLED score did not correlate to the whole blood ETP (r = 0.051, p = 0.573) and peak (r = 0.075, p = 0.401). # Discussion Thrombin generation tests in plasma have been shown to provide a more complete overview of coagulation, since it encompasses both the extrinsic and intrinsic pathway. Clotting times (e.g. PT, aPTT) only include one of the pathways and measure the time until fibrin is formed, representing just 5% of the thrombin formation. It is known that an assessment of the complete thrombin generation is better related to clinical outcomes. More recently, the whole blood CAT was developed, bringing thrombin generation one step closer to physiology paving the way for point of care thrombin generation assays. In this study, we provide a first clinical validation of this technique in patients using VKAs. Patients with and without clinically relevant bleeding episodes significantly differed in whole blood CAT ETP and peak. As expected, the INR, which is the standard follow up test for patients taking VKAs, was not predictive for bleeding on the long term; since the treatment of patients is adjusted to the measured INR. In a population of patients using VKAs the whole blood CAT suggests to be the first laboratory test that is related to a bleeding risk in a prospective set- up. The AUC of ROC is a measure for the discrimination of a test/score between patient with and without the disease. The ETP and peak in whole blood have an AUC of ROC of 0.700 and 0.642, respectively and therefore particularly the ETP can be considered as a fair predictor of bleeding. Although these parameters cannot fully discriminate between bleeding and non-bleeding patients, their discriminative value is at least comparable to that of the HASBLED score. In contrast, the plasma-based CAT did not discriminate between bleeding and non- bleeding patients. Other studies showed that plasma CAT was indicative of bleeding in different patient populations. However, this finding could not be repeated in our study. The presence of platelets alone (PRP) did not improve the outcome of the test considering discrimination between patients with and without bleeding. This is in accordance to earlier findings indicating that recurrent bleeding in patients with a stable INR cannot be explained by changes in platelets or von Willebrand factor function. In contrast to plasma CAT, whole blood CAT includes erythrocytes that can directly contribute to thrombin generation, e.g. as a source of procoagulant phospholipids on the cell membrane. Interplay between the coagulation system and these red blood cells may provide an explanation that whole blood CAT, but not plasma CAT enables discrimination between bleeding and non-bleeding patients in patients on VKA. Earlier studies by Whelihan et al. show that a sub-fraction of red blood cells express phosphatidylserine and this might serve as a surface for thrombin and other phospholipid-bound coagulation factors. Apart from erythrocytes, whole blood differs from plasma in the presence of leukocytes. The release of TF by these cells (in particular monocytes) may also contribute to the differential activation of coagulation in bleeding versus non-bleeding patients. However, the exact mechanism resulting in an improved discrimination of whole blood thrombin generation between patients with and without bleeding episodes needs further research. Bleedings that occurred during the study were mostly spontaneous, minor bleedings; none were caused by surgical intervention or major trauma. Although this increases the likelihood of similarity in etiology, the bleedings in this study population can still result from multiple factors, e.g. VKA treatment, co- medication, the coagulation system (e.g. clotting factors), age, gender or TTR. Co-medication is one of the known factors which can increase the risk of bleeding. In our population we found that 41.4% of the patients who were taking co-medication affecting platelets suffered from bleeding, whereas this was only 13.4% in patients that did not use co-medication. In this study whole blood ETP and peak were not influenced by the use of co-medication. This confirms previously reported results suggesting that the presence or absence of platelets and by extension platelet agonists or antagonists, had no effect on thrombin generation in whole blood. Moreover, we detected no differences between bleeding and non-bleeding patients in PRP. Therefore our analyses were not corrected for the use of co-medications. Furthermore, clotting factor deficiencies or increased levels of natural anticoagulants could result in an increased bleeding risk. However, both clotting factor and anticoagulant factor levels proved to be similar in the bleeding and non-bleeding patients. Additionally, an increased age or difference in gender distribution may confer an increased risk for bleeding. Yet in our study no significant difference in age or gender distribution was found between the patients with and without bleeding episodes. The TTR (which is also included in the HAS-BLED score) was lower in patients with bleeding. It is known that patients with a low TTR are at higher risk of complications (either bleeding or thrombosis). A significantly higher HAS-BLED score was detected in patients with bleeding events compared to those without. The ROC AUC of the HAS-BLED score was comparable to that of the whole blood ETP and peak. The HAS-BLED score was developed for AF patients; however other patient groups taking VKAs have comparable risk factors for bleeding. In the present study 72% of the population was treated for AF and 28% of the patients had other indications. Even within the latter group, in spite of its small number, a significant difference in HAS-BLED scores was found between the bleeding and the non-bleeding group, indicating that the HAS-BLED could be of value in these patients. Since bleeding may be provoked by many different factors, it seems unlikely that one single test would be able to predict a patients bleeding risk. Within this reasoning, a combination of laboratory tests with existing bleeding scores, may lead to a more accurate prediction of the bleeding risk. On the other hand, it could also be suggested to include the whole blood CAT ETP and peak in the HAS-BLED score, since this was the coagulation assay with the best clinical association. The INR, which is currently incorporated in the HAS-BLED score, did not prospectively discriminate between bleeding and non-bleeding patients. Additionally, there was no correlation between whole blood TG parameters and the HAS-BLED score. Therefore, we speculate that the predictive value of the HAS-BLED could be improved by replacing the INR by whole blood CAT parameters. This would then open the possibility to also use this score in assessing the bleeding risk for patients using DOACs. We expect this might be important in the future with the increasing use of DOACs. Inevitably, this study had some limitations. Firstly, the observed risk associations are with all clinically relevant bleeding complications, whereas it could be argued that only major bleeding complications matter. Previous studies showed that patients with minor bleeding events are at increased risk (\>2.5-fold) for major bleedings. Therefore, we believe that the inclusion of minor, clinically relevant bleedings is a rational and clinically relevant choice. Secondly, it can be argued whether this study featured a large enough sample size to discover a relevant effect of bleeding predisposition on thrombin generation, but in accordance with our power calculation, the observed rate of bleeding complications allows for drawing conclusions based on the presented data. It is obvious that in order to assess the utility of whole blood CAT testing in patients on anticoagulation a focus on major bleeding complications will be important, which warrants repetition in a larger study to corroborate our findings. Finally, it is known that some biomarkers and risk factors associated with bleeding have also been linked to stroke and systemic thromboembolism. Therefore, it would be interesting to explore this topic. Unfortunately we were unable to do so because there were no episodes of recurrent VTE or stroke during the follow-up period. In conclusion, clinically relevant bleeding in patients taking VKAs in our study is associated with a diminished whole blood ETP and peak, but not with INR. Consequently, whole blood ETP and peak could be of value in assessing patients who suffer from recurrent bleeding with an INR in therapeutic range. In our population an augmented HAS-BLED score was associated with bleeding as well. Implementing whole blood CAT might improve the sensitivity of bleeding scores such as the HAS-BLED, enabling further tailoring of therapy for individual patients. # Supporting information The authors would like to thank Hilde Kelchtermans for critically reading the manuscript and the employees of the Maastricht Anticoagulation Clinic for their assistance with patient inclusion and blood collection. [^1]: I have read the journal's policy and the authors of this manuscript have the following competing interests: HTC reports personal fees from Stago and a position as chair of the Federation of Dutch Anticoagulation Clinics. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: **Conceptualization:** SB HTC ATCH BDL. **Data curation:** SB SZ. **Formal analysis:** SB SZ. **Investigation:** SB SZ. **Methodology:** SB SZ ATCH. **Project administration:** SB ATCH. **Resources:** BDL. **Supervision:** ATCH BDL. **Writing – original draft:** SB SZ BDL. **Writing – review & editing:** SB SZ HTC ATCH BDL.

# Introduction: Pathways, interventions and emergent properties Relationships between poor socio-economic conditions and a wide range of adverse health outcomes have been observed in numerous contexts. These relationships arise from direct health risks, stress from being at the lower end of the income/wealth distribution, and higher levels of risk behaviour. Here we explore: (a) aspects of why prevalence of risk behaviours is higher among people living in adverse environments, and (b) why the relationship between socio- economic status and risk behaviour is not linear. There is considerable evidence of associations between risky environments and increased risk behaviour. Although these associations are well established, the evidence is chequered regarding the *pathways* through which contexts influence health. This is an important *lacuna* as a better understanding of those pathways could facilitate the design of better interventions. However, the notion of “pathway” is itself problematic, adding a layer of largely unresolved complexity to any discussion of contextual influences on health status and outcomes. In this paper we discuss pathways to risk behaviours among adolescents in South Africa while noting theoretical difficulties implicit in using “pathway” as a framing concept. In the absence of critical appraisal, the concept is all too often deployed in a mechanistic and linear manner implying a mechanistic theory of social, cultural and economic influences on the health status and rates of disease acquisition by individuals. Such linearity is a conceptual simplification which may contribute to the design of simplistic interventions. This research suggests that the nature of the association glossed in the term “pathway” may be clarified by introducing the theoretical idea of *emergent properties* and that this clarification could contribute to improved design of interventions. Different kinds of evidence suggests that local history, politics and culture are important components to be considered in the design of interventions intended to reduce the frequency of risk behaviours. Generalised interventions which do not take account of such factors are less likely to succeed. It is likely that while such contextually sensitive approaches may appear expensive, cheaper generic interventions may waste money because, even if effective, they are sustainable only if and while funding is available. Interventions which take account of contextual factors as experienced by the target population may turn out to be more cost effective and efficacious in the medium to long term, both because they may be more effective at changing behaviour, and because such changes may not require as much ongoing reinforcement. Such interventions may be said to take three broad forms: (a) major socio-economic restructuring via social movements and political changes; (b) social and economic reform of health and welfare systems; and (c) context specific manipulation of individual choice as in campaigns to achieve behaviour change. The 1958 Cuban revolution and the so-called “Washington Consensus” of the 1990s are contrasting examples of the first strategy; the mid-twentieth century health and welfare reforms in the United Kingdom, known as the “Welfare State”, are examples of the second; and (c) limited contextual interventions which demonstrate an understanding of the *choice architecture* within which individuals are able to make health choices are examples of the third. In this paper, we discuss changes enabling people to hope as a health intervention strategy engaging all these approaches in different degrees. If young people are to embrace hope for their individual and collective futures it is likely that interventions along all these dimensions are required but that interventions designed with choice architecture in mind may prove to be a pragmatic and feasible first step which can contribute to reforms at the other two levels. Two deep intellectual influences frame our analysis. First are ideas from the work of the nineteenth century sociologist Emile Durkheim. He theorised that *social currents* were important in relation to the prevalence and incidence of suicide, an apparently individual act which he attempted to demonstrate was in fact intensely social, the outcome of the influence of *social currents* on the lives and life chances of individuals. Second is the very long tradition of thought which recognises that, in simple terms, the whole is greater than its parts–thus a higher-level system has qualities which are possessed by none of its individual parts. While a common example of such emergence is the snowflake where the crystalline structure is quite different from the molecular structure of water, so social distributions of power, income, wealth, taste, fashion and hope reflect processes of emergence. The framing theoretical idea of this research is recognition that individual behaviour is simultaneously limited by social structure and constitutes that structure. Sennett and Cobb’s *The Hidden Injuries of Class* is an example of the application of this idea to class structure in the United States. The idea appears in sources as ancient as Aristotle and since then and contemporarily in most fields of study from evolutionary theory to physics, biology and ecology. To investigate the usefulness of this approach we examined how well our exemplar emergent property, hope, explains variance in alcohol use among adolescents in rural KwaZulu-Natal. We conceptualise hope as a distinct and operationalizable construct providing a way of measuring, via individuals and groups of individuals, indications of where ecologies of risk may be mapped as they affect individuals. The concept of hope may lead us to a better understanding of the pathways between individuals’ perspectives on their life worlds, social and economic conditions, and risk-taking behaviours. This perspective is dynamic, concerned with socio-economic and socio-cultural processes and offers a generalised entry point into diverse risk environments without homogenising this diversity into simple “pathways”. Our study includes a qualitative component investigating young people’s understanding of hope and how it relates to risk. This informed the development of a measure of hope, which was then administered and its association with self- reported alcohol consumption examined. Measures of happiness and life- satisfaction were included as comparators. This allowed us to investigate how measures which reflect aspects of how individuals experience and interpret their environments compare, as predictors of risk behaviour, to a measure of socio- economic status, which provides a description of their environment, with no room for individual interpretation. It is the potential that the interpretation, which occurs because of interactions between individual, group and context, could help explain differences in risk behaviour, which underlies the emergent properties approach. Hope is of interest because it speaks to how the individual sees their future from within their social structural position, their “life world”, a possibly important determinant of their present-day behaviour. Hope, happiness and life- satisfaction are likely to be linked to both individual and group narratives. That is, to how the social conditions are seen and understood through an interaction of individual and group. In this study, we are interested in the extent to which these measures provide an indication of how social determinants of health link to risk behaviour via these individual and group interactions, and to differences in interpretation. To clarify this problem, we posed the theoretical question: do variables such as hope, happiness and life satisfaction which seek to capture elements of individuals’ perceptions and experience of their structural contexts explain behaviour better than variables that seek to describe that context, for example an objective measure of socio-economic status? # Study design We adopted a mixed methods approach: (a) a qualitative component investigating risk behaviours and risk environments in relation to one example of an emergent property, hope; (b) a quantitative component, which investigated the relationship between a measure of hope, developed from the qualitative findings, and differences in self-reported alcohol consumption. In the qualitative component, discussions with adolescents focused on the meaning of hope, as understood by the target group, and to what extent they thought it captured aspects of their life-worlds relevant to explaining risk behaviour. This component included focus group discussions and individual interviews, details of which are provide below. In the quantitative component, we gathered data on 500 adolescents to investigate the possibility that hope, measured initially as an index, was associated with self-reported risk behaviour. The design of this measure was informed by the qualitative work. The qualitative results informing the design and the survey methods are outlined below. To provide a comparison we measured two other variables which capture how individuals see themselves and how they feel: these are happiness and life satisfaction. Both were included in the qualitative and quantitative components because, like hope, they may be “summary” indicators of individuals’ experience at the intersection of individual agency and social structure Barnett, Fournié. We also compared the explanatory power of these variables with a measure of socio-economic status, an asset index. ## Study site The study was conducted in KwaZulu-Natal in the Republic of South Africa at a site within the Demographic Surveillance Area (DSA) of the Africa Health Research Institute. The DSA is located near the market town of Mtubatuba, 250 km north of Durban in uMkhanyakude, a sub-district of Hlabisa. The population is largely Zulu-speaking and the area is predominantly rural, with an urban township and informal peri-urban settlements. Data were collected in 2017. A total of 8 research assistants employed an active recruitment strategy which included identifying potential participants during road shows (presentation of study to the community), in places where young people congregate such as the playgrounds, outside the shops and water taps. Snowball technique was also employed as individuals referred their friends or told the research assistants of other young people meeting the inclusion criteria. Before the consent and assent process, potential participants where asked where they wanted to be interviewed. All participants preferred to be interviewed in the privacy of their home with only the researcher and participant present in the space during the interview. Participants in the quantitative component were identified from the DSA database and interviewed at home. ## Ethics approach The study was approved by the Biomedical Research Ethics Committee (BREC) of the University of KwaZulu-Natal (BE549/16) and the Ethics Committee of the Human Sciences Research Council (4/17/02/16). Informed consent was obtained from parents/primary caregivers for the participation of children in their care. Children were then asked if they agreed to participate. Both consent and assent were in writing. Information sheets, in Zulu, were read to caregivers and children, discussed and questions taken. These sheets were left with participants. They also contained contact details of the ethics committee, study PI, local coordinator and sources of support for adolescents. ## Qualitative methods We investigated young people’s understandings and experiences of hope in relation to their perceptions and enactment of risk behaviours. We began with the Snyder Hope Scale as an exploratory starting point. While the scale had been validated for use in South Africa, we did not want to use it in the quantitative work without first ensuring that it captured a measure of hope relevant for our purposes. Concern has been raised that this measure frames hope as an individual construct rather than as a collectively shared concept, and may therefore may not be appropriate for use in South Africa. Abler, Hill have similarly raised concerns regarding the use of existing measures of hope in the South African context and have even proposed an alternative. Their suggested measure appears to perform well, but seems to capture mood rather than a driver of behaviour. A guide, framed by the constituent items of the Snyder scale, was used to introduce discussions of hope with key informants in 30 Key Informant Interviews (KIIs) and four Focus Group Discussions (FGDs). The total sample comprised 53 young people, aged 15–17 years of age. We initially conducted 22 KIIs as per our sampling criteria, however the data and interaction with participants indicated some confusion regarding what we meant by hope. There was a challenge in translating the word ‘hope’ into Zulu as there are a number of alternative words. A further 7 KIIs were conducted using a revised topic guide to probe these different understandings and establish which of them respondents linked to risk behaviours. Participants for the KIIs were purposively recruited to include a minimum of four 15–17 year-olds in each of the following categories: i) out of school youth’ ii) school going, iii) with resources such as flushable toilets/ piped water/ electricity, iv) with limited or no resources, v) parents with income, and vi) with parents with no income or with unstable income. For the group discussions, we sought four groups (6–8 participants) of young people also aged 15–17 years, two groups for males and two for females. The semi-structured discussion guide was prepared in English and then translated into Zulu (Foxcroft and Roodt. A participatory approach was adopted, engaging with young people using a lifeline drawing to promote discussion during the group discussions and interviews. This was used to unpack group ideologies and the social construction of the key concepts as well as the more subjective experiences and narratives of these concepts. Interviews and FGDs were conducted in Zulu, audio-recorded, transcribed verbatim and translated. The transcripts were analysed using thematic analysis, facilitated by Atlas ti software (version 5). This study component of qualitative interrogation of the constructs and the measures provided us with an opportunity to revise the constructs for use in our survey. The results of the qualitative analysis informed development of the quantitative measure of hope. ## Quantitative methods We conducted a survey to examine the association between hope, happiness, life satisfaction and risk behaviours in a sample of adolescents, resident in the study site. Eligible participants were drawn at random from an existing database. We limited eligibility to a sub-region to limit costs. In keeping with ethical protocols, the guardians of eligible participants were approached at home and asked if they were willing to consent to the research team’s request to interview the child (15–17 years) in their care. Children were then approached and asked if they would agree to be interviewed. The survey included age, gender, and a range of questions related to hope, happiness and life satisfaction all rated on a Likert scale. The questions on hope, happiness and life satisfaction were selected based on the results of the qualitative component. Details of the measures are, therefore, outlined below alongside the qualitative results. The survey also included questions on whether respondents had ever used alcohol, *dagga* (cannabis), *tik* (methamphetamine), *buttons* (Mandrax), *whoonga* (heroin), ecstasy, or other drugs, whether the respondent was still using these substances, age at first use, and how regularly they used these substances. Self-reports of substance use were very rare and as a result we only used the data on alcohol consumption. Data from this survey were linked to data from the DSA on household and individual socio-economic characteristics, including data on sexual risk behaviour, an index constructed from household asset ownership, and the highest school grade attained. The data on sexual risk behaviour was of poor quality, with very high levels of non-response, and not used in the final analysis. The asset index was constructed using the first component extracted from principal component analysis (PCA), based on household ownership of a list of durable assets and the type of water source, cooking fuel, access to electricity, and toilet facilities. A variable indicating whether a respondent was two or more school grades behind where they should be for their age was derived. This was based on their highest grade attained and their year of birth: the variable was selected with a view to capturing peer influences. Following data collection and cleaning, analysis was undertaken in three steps. The first step involved basic descriptive analysis of the survey responses. The measures of hope, life satisfaction and happiness were summarised by individual items, as it was not obvious prior to data collection how items should be combined into a scale. The second step of the analysis addressed this question, examining the approaches to using the hope, happiness and life satisfaction data, before choosing the third (as we explain below). The first involved simply summing the items. This is appropriate if there is a relatively common pattern across responses, such as when the Cronbach’s Alpha was high (above 0.7). The second approach considered was to construct an index using polychoric PCA (Poly-PCA); essentially a PCA for non-binary ordinal variables. The third option was to use the single item which most directly referred to the construct of interest. The final stage of the analysis was a regression analysis with alcohol consumption as the dependent variable. Hope, happiness and life satisfaction were each included in two regressions without the other two, one with the asset index included and one without. This allowed us to examine whether they are associated with risk behaviour, and whether that association is independent of socio-economic status. Finally, a regression was conducted including hope, happiness, life satisfaction and assets. All regressions controlled for age and gender and whether the child was more than a year behind peers in school (as a control for peer effects on behaviour). In each case risk behaviour was a binary variable, therefore a logit regression was used, and odds ratios reported. All regressions were conducted in Stata. # Results ## Qualitative results and the selection of quantitative tools Hope, happiness and life satisfaction were all discussed with respondents in the qualitative phase. Respondents’ understanding of happiness and life satisfaction linked closely with the items in the associated scales. However, when we reviewed our findings on hope, and compared them to the Snyder scale, we found notable divergences. Moreover, analysis of the qualitative data revealed the complexity of the concept of hope, indicating that measurement may be a challenge, and that identifying the correct language around hope would be key. The focus of the Snyder scale is on hope as an *individual* characteristic (part of the tradition of “positive psychology”), a belief in what you can do. It did not appear to capture the idea of hope respondents linked to risk behaviour, which was a belief that things will improve. In particular, the scale focuses on how past experiences prepares individuals for current challenges, whereas respondents often focused on the future, on how to make it better than the past, and on their ability to influence that future, as factors shaping risk related decisions. The adolescent respondents reported they were generally introduced to the notion of hope by teachers and parents during early schooling years. Hope was often thought of in relation to an object, person or goal; having hope in someone or a hope for something (a wish or a dream). Hope was also considered to be a belief or mind-set that was associated with positive future changes often articulated as a source of encouragement in the face of adversity, or a coping mechanism. Some also considered hope to resemble a character trait associated with self- belief, agency, goal setting and, to some extent, decision-making. For example, hopefulness, and conversely hopelessness, was perceived to be embedded within young people’s behaviours. Those who engaged in risk behaviours such as substance misuse were considered to have “no hope” and lack future aspirations. The absence of hope was seen to be a hindrance to goal attainment and achieving future success. The two conceptualisations of hope, Snyder’s emphasis on individual capacity and the respondents’ emphasis on probable futures are related but sufficiently distinct that we did not include the Snyder Scale in the survey instrument. Rather, we developed a new hope scale derived from the qualitative data. This included eight questions related to hope. lists the items and indicates the construct they were intended to capture. The items aimed to address the issues raised in the qualitative work relating to hope and risk behaviour and included risk related to individuals not considering the future, i.e. living in the moment; hope providing a way to cope with current adversity, preventing risk behaviour; and hope related to a future orientation and an associated sense of control. These were scored in interviews using a Likert Scale. The qualitative work also provided us with the appropriate language, in Zulu, to discuss hope in relation to each of these constructs, something we had initially struggled with in the focus groups and interviews. ## Quantitative results ### Descriptive statistics Data on 503 respondents were collected between 11 May 2017 and 1 September 2017: 261 males (52%) and 242 females (48%). Respondents ranged in age from 15 to 18 years, with an average age of 16.48 (i.e. 16 years and 174 days). A total of 121 (24.40%) respondents were more than two grades behind the appropriate grade for their age in school. provides a summary of the basic demographic characteristics. summarises the responses to the eight hope items. The results indicate a high level of agreement with the positively framed items and a high level of disagreement with the one negatively framed item. The responses to the life satisfaction scale, by item, are reported in ; the responses are positively skewed for the first three items. The final two items do not follow the same pattern. It is not immediately clear how to reconcile being satisfied with life with not having got what you want and indicating that if you had the chance you would change the past. The first of the negative responses could be interpreted as reflecting the age of the respondents–‘So far I have gotten the important things I want in life’ being disagreed with because they see these things in their future. However, the final item suggests that many respondents are not satisfied with how their life has gone. This is difficult to reconcile with their reported satisfaction overall. summarises the item responses to the happiness questions. Most respondents reported being happy or extremely happy. When asked to compare themselves to their peers a larger majority reported the same. However, less than half reported being happy or extremely happy in the last month. There was very little self-reported drug use, and low self-reported alcohol use, see. It is possible that respondents may not have been comfortable admitting to these behaviours. ## Regression results: The association between hope and alcohol consumption We conducted a regression analysis to examine the association between reporting alcohol use and the exemplar emergent property, hope. For comparison, the association with happiness and life satisfaction is also examined. In each case, the strength of that association is compared to the association between socio- economic status, as measured by the asset index, and alcohol use. Before conducting the regression analyses, we had first to select the most appropriate of the three approaches to using the hope, happiness and life satisfaction data. The first approach, simply summing the responses, would have been appropriate had there been relatively common pattern across responses. The Cronbach Alphas suggest that this is not the case; none was above 0.7 (Hope 0.65; Life Satisfaction 0.64; Happiness 0.53). The second approach, constructing indices using Poly-PCA, was conducted. The eigen values suggested that it was appropriate to use only the first component as the index for each construct. However, given the patterns of responses, and the associated component loadings the indices are difficult to interpret. It appears that not all the items in each measure are measuring elements of a common underlying construct. The third option, i.e. to use the single item which most directly referred to the construct of interest, is more straightforward to interpret than the Poly-PCA. We do not have to make sense of component loadings in the indices. We therefore settled on the use of a single item. While we focus on the results using this third option, all analyses are repeated using the indices created with poly-PCA and these regressions are reported in the supporting information. The following single items in the regression analysis. - Hope: I generally feel hopeful about my future, scored 1–5 by strongly disagree, disagree, neither agree nor disagree, agree, strongly agree. - Life satisfaction: I am satisfied with my life, scored 1–5 by strongly disagree, disagree, neither agree nor disagree, agree, strongly agree. - Happiness: In general, I consider myself (extremely unhappy, unhappy, neutral, happy, extremely happy), scored 1–5. The use of a direct single item measure has positive and negative implications. It allows for variation in individual interpretations of hope, happiness and life satisfaction. Thus, it may better capture how individuals feel. However, we are no longer imposing an interpretation, as would be the case if we used a combination of multiple items, which makes interpreting the results more difficult. The regression results using the single item measure of hope are reported in. Three regressions were conducted. For each test the dependent variable is whether the respondent reported any alcohol use. All regressions include age, gender and whether the adolescent is more than a year behind in school as controls. The difference between the regressions is that the first includes the asset index, the second includes the measure of hope, and the third assets and hope. The table reports odds ratios. A ratio of more than 1 indicates a positive correlation with the probability of reporting alcohol consumption, less than 1 a negative correlation and 1 indicates no association. In all three regressions, being older and male are found to be positively associated with alcohol use, as would be expected. Being two or more grades behind in school was negatively associated (odds ratio less than 1) with alcohol consumption, as expected, reflecting a younger peer group. However, the relationship was weak and only significant in models 1 and 3 and then only at the 10% level. The result of primary interest is the significant negative correlation between hope and alcohol use (an odds ratio of less than 1). By comparison, there was no significant association between the household asset index and alcohol consumption. Including both hope and assets in the regression does not change the result for either, suggesting that they are not measuring similar things, i.e. hope is not simply a correlate of socio-economic status. reports the results of five regressions. The first regression is the same as the first regression in the previous table, it is repeated to facilitate comparison. Regressions two and three include life satisfaction, and life satisfaction and assets respectively. Regressions four and five include happiness, and happiness and assets, respectively. Age and gender were again significant in all regressions with the expected relationship: older males more likely to report alcohol consumption. The asset index was not significant in any of the regressions. The single item measures of happiness and life satisfaction both led to significant odds ratios below one, implying a negative association between them and the probability of alcohol consumption. The final regression conducted is reported in. It includes the standard controls plus assets, hope, life satisfaction and happiness. Again, the associations between the probability of reporting alcohol use and age and gender were as expected. The asset index again was not associated with alcohol use. With the measures of hope, happiness and life satisfaction all included in the same regression, only happiness showed a significant association with probability of alcohol use. # Limitations The study is hindered by the low rates of reporting of risk behaviours. In the case of sexual behaviour, there was too much missing data for the variable to be used. In addition, the response rates for drug use were too low for the variable to be used. It is possible that these rates are correct but given established rates of substance use among adolescents in KwaZulu-Natal, this is unlikely. The low rates of reporting left us with usable data on only one risk behaviour, alcohol use. There is a concern that this too suffered from reporting bias. If there is reporting bias, and it is associated with any of the measures, then our results are biased. However, for the bias to have led us to erroneous conclusions, it would have to be because those who were more hopeful (happy and satisfied), were more likely to conceal alcohol consumption, which seems unlikely. Similarly, if those with lower SES systematically concealed alcohol consumption more often, the bias would have implications for our results. There is no reason to expect this to be the case. If the misreporting was not associated with our variables, the bias is not a concern for our conclusions. This is a cross-sectional study and there is a possibility of reverse causation. It is possible that alcohol consumption makes people less hopeful or unhappy. Alternative longitudinal approaches to data collection will be required to address this concern in the future. The study population was relatively homogeneous in terms of socio-economic status. The lack of variance in socio-economic status may explain the weakness of the association between the asset index and alcohol consumption. However, it is interesting that despite the population being homogeneous, there was sufficient variation in hope, happiness and life satisfaction measures to identify a relationship with alcohol consumption. # Discussion Studies of the social determinants of health have repeatedly shown that there are strong associations between adverse social conditions and certain risk behaviours. Little attention, however, has been paid to the role of individual decision-making in linking context and behaviour. There are discussions of empowerment, and the extent to which adverse conditions reduce people’s sense that they can do something about their situation. However, even these are based on a thin conceptualisation of the people involved. Little or no space is given to trying to understand how individuals and groups understand their environment, how they interact with and change their environment, and how these understandings and interactions shape behaviour. In this analysis we recognise that “the social” and “social structure” consists of a complex set of fluid interactions between emergent properties of the socio- economic-cultural spaces within which we lead our lives. It is this complexity which engages with the idea that: (a) a “pathway” is not a constant, mechanical relationship, but rather reflects the generation, alignment and interaction of emergent properties with particular places and times; and thus (b) any interventions should engage with the shifting nature of the terrains in which they are located. A view of individuals which focuses on the importance of their (fluid) interpretation of the environment, and how this shapes their behaviour, requires a focus on what characterises that environment and what influences people’s understanding of those characteristics. This in turn highlights the importance of both group influences and individual variation. The environment and different understandings of that environment are shaped by groups. Yet not all members of groups share a common understanding. These are complex conceptual and logical issues. The approach through emergent properties is a step towards a fuller conceptualisation of risk behaviours and their link to the social individual. It does this by providing a way to investigate the role of interpretation in linking environment to behaviour. It allows us to ask if different interpretations help explain different responses, i.e. why the relationship between environment and risk behaviour is not linear or uniform. It can be used to examine group interpretations of the environment, and individual variation within groups. In this study we focused on a relatively homogeneous group. Therefore the analysis, has focused on individual variation within the group. To facilitate this examination of individual variation, we first had to examine the ways in which the group understands the links between context and behaviour. Using hope as an exemplar of an emergent property, we worked with young people to develop a fuller understanding of what hope means to them, and whether they saw links between hope and risk behaviour. The qualitative analysis suggested that young people in the study area believed that there was a link between hope and risky behaviours. On one hand, they emphasised a link between giving up hope and engaging in risky health-related behaviours; on the other hand, they emphasised the link between hope and future orientation and low-risk health behaviours. Through our qualitative component we found that the concept of hope was difficult to translate into Zulu, as it could be understood in several different ways. As a result, we had to refine the initial scale significantly. The quantitative results suggest that within our study population the level of hope reported by an individual is more strongly associated with their risk behaviour than is their socio-economic status. Thus, regression analysis showed a statistically significant negative relationship between hope and risk behaviour. We did not find the same for the asset index. However, neither measure explained much of the variance in risk behaviour. This lack of explanatory power suggests that other factors which explain individual differences in alcohol consumption were not captured. This result is grounds for cautious optimism: optimism stems from the exemplar out-performing a well-established measure of socio-economic status, known to be associated with risk; caution, from how weakly both hope and the asset index were associated with alcohol consumption. The regression results indicated a negative relationship between hope and the probability of alcohol consumption, i.e. higher levels of reported hope were associated with a lower probability of reporting alcohol consumption. Similarly, the direct questions on happiness and life satisfaction had significant negative associations with alcohol use. The asset index, which was used as a measure of socio-economic status was, however, not associated with alcohol use. The lack of association between assets and alcohol consumption may reflect homogeneity of socio-economic status within the sample. It is noteworthy that even among adolescents living in similar contexts, there was enough variance in reported levels of hope, that an association with alcohol use could be identified. This result is cause for optimism regarding the potential use of emergent properties in explaining variations in risk behaviour. Improving measurement of emergent properties is perhaps the biggest challenge facing this approach. The regression results suggest that there is significant overlap between hope, happiness and life satisfaction. When all are included in the same regression only happiness is significantly associated with alcohol consumption. More work is needed to take further the task of identifying emergent properties capable of distilling the influence of lower level variables into single measures useful for analysis and policy purposes. A feasible next step may be to examine the association between risky health behaviours and combinations of emergent properties. A single emergent property may not capture enough to adequately explain differences in risk. Combinations are likely to cluster, leading to emergent narratives, with these narrative framing and influencing risk behaviours. The possible importance of narratives was evident in the patterns of response to the various scales. It helps to explain how individuals may be happy and satisfied, yet not happy with the last month or satisfied with how things have gone. Firstly, people may need time to process recent experiences into their narratives. Secondly, once assimilated, people’s current narratives may differ from their past experiences. Understanding the development of emergent narratives as accounts of how people experience the influence of factors constraining their agency which may be summarised through emergent properties like hope may be important for further understanding of how emergent properties can be used in developing policies relating to health risk behaviours. # Supporting information We thank the community for their continued support for, and participation in, AHRI research projects, and the staff at AHRI for their support and hard work. We would like to acknowledge the support of the KZN Department of Health. We thank all the young people who took part in this project and shared their data with us. Thank you. [^1]: The authors have declared that no competing interests exist.

# Introduction Temperature is a key environmental variable in lakes and ponds, since it accelerates biochemical reactions, and therefore, increases the rates of many biological and ecological processes including photosynthesis and respiration, organic carbon mineralisation organisms’ growth, biomass production, organisms’ size, and ecological processes since it influences thermal niche and species distribution, and trophic cascades. More specifically, accumulated degree-days (ADD), which is the temperature integrated in time over a determined threshold, explain organisms’ development, while maximum water temperatures limit growth rate and warming tolerance. Temperature range is also essential in explaining life-history traits, as some specialist organisms, occupying a narrow temperature range, present a higher performance. Temperature oscillations, which can operate daily or within different days, are essential because they can increase poikilotherms growth, or can affect them negatively when temperatures are too high. Temperatures have been increasing since the beginning of the industrial revolution, and they are expected to increase between 0.3°C and 4.8°C globally, depending on the scenarios of greenhouse gas emissions, for the 2081 to 2100 period. The air temperature rise translates into a lake water temperature increase, although water temperatures trends can vary depending on climatic variables, such as radiation, cloudiness and wind speed, and regional characteristics of the lakes, and may differ through the water column of the lakes, showing high increasing trends in shallow waters and low warming trends in deep waters. In this sense, water transparency plays an important role in the configuration of the thermal structure of lakes. Warming is expected to be greater in both polar regions and high mountain areas. High mountain areas are globally biodiversity hotspots due to the encapsulation of different climate zones in very short distances. Despite the strong altitudinal gradient, the high spatial heterogeneity of high mountain areas creates a complex mosaic of local conditions, which make the upscaling of the local factors affecting surface water temperatures at regional scales not straightforward, and a missing gap in understanding future scenarios of climate change effects on high mountain species’ distributions. In this regard, the Pyrenees constitute an excellent case study of a high mountain region, containing over 3,000 lakes and ponds. Past air temperatures have been reconstructed in this region at some sites using statistical extrapolations for recent centuries or using lake sediment records. Also, there have been comparisons between the Pyrenean lakes and other mountain regions in Europe. However, there is not a comprehensive assessment of the variation within the Pyrenean lake district and their associated factors, which could be the basis for a spatial upscaling to the whole region as a reference for ecological studies, particularly of littoral and amphibian organisms. Amphibian populations in Pyrenean lakes need a minimum temperature threshold to develop. Moreover, they are known to be sensitive to water temperature increases. They include species such as the ubiquitous *Rana temporaria* L. 1758 and the endemic species *Calotriton asper* (Dugès, 1852). Surface water warming can also affect macroinvertebrates such as aquatic beetles (Coleoptera, Dytiscidae), which are both sensitive to temperature and predation by introduced salmonids. Fish species as *Salmo trutta* L. 1758, which are invasive to these lakes, have both a minimum and maximum temperature threshold for their development. Both native and invasive species may see their potential habitat altered by increasing temperatures. Therefore, an assessment of thermal conditions of high mountain ecosystems such as Pyrenean lakes are necessary for future studies on ecology and conservation in these ecosystems. Many studies have dealt with thermic variables monitoring and modelling of lake surface water temperature (LSWT) at different temporal scales: daily, monthly or seasonal, maximum annual temperature, daily minima and maxima and diel temperature range (DTR). Also, accumulated degree-days, and temperature oscillations have been used for physiological studies of freshwater organisms, and accumulated degree-days have also been modelled. However, no studies have attempted to develop models of a handful of thermic variables using the same methodology. The physical processes that drive the heat balance in lakes are essentially known. However, the available physical models require the input of driving variables which are hardly available for a large number of lakes. Therefore, the upscaling of thermal conditions to large lake districts is difficult. Statistical modelling is an alternative that requires a sufficiently large calibration set and the identification of the main landscape factors that primarily constrain the actual physical drivers. We used surface water temperature 9-year series from 59 lakes and developed mixed regression models. We included explicative factors that constrain the atmosphere-water thermal interaction due to the location (latitude, longitude and altitude), the system inertia (lake area), the advective heat flow (water renewal and hydrological complexity of the catchment), and the incoming radiation (topography). These constraints do not consider interannual variation but mean conditions; therefore, we also used spring and summer air temperatures from a weather station in the Central Pyrenees to account for this temporal variation. We also modelled some thermal niche features that are commonly affecting a variety of organisms, namely ice-free period mean and maximum temperature, accumulated degree-days, mean diel temperature range, and temperature oscillation; the latter defined as the difference in maximum temperature between a time-lapse of three days. Warming can increase habitat availability (e.g., accumulated degree- days in cold lakes) and limit species survival (e.g., maximum temperature in warm lakes). # Methods ## Temperature and environmental parameters A set of 59 lakes and ponds were selected to deploy minilog-thermistors with attached dataloggers (Vemco Minilog-T) with a precision of ±0.1°C spread along the Central—Eastern Pyrenees (42 and 43° N and between 1° W and 3° E;), to consider a wide range of water temperature variability. The lakes and ponds were selected following key environmental variables, which can affect high mountain lake water temperatures, such as altitude, lake and catchment size, residence time, and radiation (See and for a description of their ranges and comparison with the Pyrenean water bodies). Temperature thermistors were deployed in the lakes of the National Park of Aigüestortes i Estany de Sant Maurici, the Natural Park of Alt Pirineu and Natural Park of Posets-Maladeta with permissions from the park authorities. Other lakes belonging to public domain did not require other specific permissions (those belonging to protected sites are detailed). No endangered or protected species were involved in this study. Thermistors recorded temperature continually, all year long, and were replaced before batteries depleted. They provided us with complete summer water temperatures for 2001, 2002, 2004 just in Lake Redon, and from 2009 to 2014 for the whole set of water bodies, constituting 9 years of recorded summer water temperatures. Thermistors were deployed from the lakeshore at 1.5 m depth and separated from the lake bottom using a fishing rod. Temperature measurements were taken at 90-minute intervals. The thermic variables were calculated for the ice-free periods, defined as the periods over 4.0°C. Mean temperature (Tmean) was calculated as daily mean temperatures averaged over that period. Maximum temperatures (Tmax) was the maximum recorded each year. Diel temperature range (DTR) was calculated as the average of diel temperature ranges over the ice-free period, and temperature oscillation (Tosc) was calculated as the difference in maximum temperatures in a three-day lapse, averaged over the same period. The whole set of thermal variables for each lake and year are detailed in. The lakes and ponds with measured water temperature are represented in white. All the mapped lakes and ponds of the region are in blue and the automatic weather stations are in yellow triangles. ADD were calculated over two different temperature thresholds (T_T), 4.0°C, and 7.6°C, the first (ADD4) representative of the ice-free period, and the latter (ADD7.6) as an animal development threshold, as it was the minimum temperature necessary for the development of *Rana temporaria* larvae. For ADD, but not for the other thermic variables, the variables were transformed to ADD at 0.1m depth as it is more representative of the habitat of these organisms, the transformation was performed using a linear regression between ADD at 1.5m and ADD at 0.1 m in 5 lakes. See for a detailed description of the ADD calculation procedures. Mean spring (March–May) and summer (June–August) air temperatures were calculated for years 2000 to 2015 from available daily data of the Catalan Meteorological Service in 14 altitude automatic weather stations located in the Catalan Pyrenees. Air temperatures used in the models were those of Lake Redon meteorological station, which is located in a central location in the Pyrenees and has a long temperature series, given that all temperature series were highly correlated and were representative of each year seasonal air temperatures in the region. Other meteorological variables, which are known to affect water temperatures, such as wind, precipitation or cloud cover, were not considered in this study, as they are less coherent spatially in a mountain region as the Pyrenees. Lake surface area (Larea), direct catchment area (Dcatchment), which is the area where water flows directly to the water body (i.e. not passing through another lake), and total catchment (Tcatchment) were digitalised and calculated following existing topographical cartography and aerial photography from the Catalan Cartographic and Geologic Institute, the French National Geographic Institute and the Spanish National Geographic Institute, as it was done by Casals-Carrasco et al.. From these variables, we calculated two ratios describing the geomorphology of the catchments; Dcatchment/Tcatchment, which is inversely related to water retention in the upper lakes of the catchment, and Tcatchment/Larea, which is a surrogate of water renewal time, given comparable precipitation among sites and time, since runoff is related to catchment size and lake depth is related to lake area. So lake volume is proportional to lake area. Since water renewal time is defined as the ratio between runoff and lake volume it can be represented by Tcatchment/Larea. Geographic coordinates were expressed as X and Y coordinates in ETRS89 UTM coordinate system in the zone 31N. Theoretical radiation data for each lake and its catchment were calculated from a digital elevation model (15 m of pixel resolution) of the Pyrenees. We calculated the monthly radiation using the Area Solar Radiation algorithm. This analysis tool calculates insolation across a landscape for specific locations based on methods from the hemispherical viewshed. Global, direct and diffuse radiation and the sun hours were calculated for both the lake surface and the whole catchment. Radiation was calculated considering a solar constant of 1,367 W m^-2, considering solar track, atmosphere attenuation and topography (defined with the digital elevation model). Radiation variables were highly correlated among them. Therefore, we chose the sun hours in the total catchment as a representative of them all, because in a preliminary analysis it was the radiation variable with a better statistical performance. # Statistical modelling Mixed models were performed in R language version 3.6.1. They were constructed relating thermic variables (ADD7.6, ADD4, Tmean, Tmax, DTR, and Tosc) to altitude, lake area, geomorphology, X and Y coordinates, radiation (as the sun hours in the total catchment) and air temperatures of spring (Tspring) and summer (Tsummer), since they are the previous and current season of the ice-free season. The geomorphology variables included the ratio between direct and total catchment (Dcatchment/Tcatchment) and the ratio between total catchment and lake surface area (Tcatchment/Larea) as a surrogate of lake water renewal time. These two ratios and Larea were normalised by logarithmic transformation. We checked that correlation between pairs of explicative variables was not too high (r \<0.71) to avoid collinearity effects. All variables were standardised to z-scores subtracting the mean and dividing by the standard deviation of these variables in the 59 sampled lakes during the 9 studied years. Random variables considered in the models were year and waterbody type (classified as pond \<0.5 ha and lake \>0.5 ha). The model selection used followed the ten-step protocol described by Zuur et al.. It consisted of a first selection of the variables in the fixed part, first adjusting ordinary multiple linear regression models, then, selecting variables by stepwise selection, choosing the model with the minimum Akaike Information Criterion (AIC). We considered two different model selection methods for each variable, using forward or backward selection. We also considered interactions between the most relevant variables (Altitude, Larea, Tspring and Tsummer). Later, we adjusted mixed models with random structures, by restricted maximum likelihood (REML), using the “lme” package and we compared them with the ordinary multiple regression model. The null random structure was tested with three different structures: the year, the water body type and the year nested in the water body type. The structure with the lowest AIC was selected. The following step was to test the optimal fixed structure using the likelihood ratio test with the mixed model estimated with maximum likelihood (ML) to determine the significance of the variables and drop the non-significant ones. Finally, the model was refitted with REML and tested for homogeneity of variance and independence. From the four models calculated for each variable, the one with the lowest AIC was kept (all models tested are). Summary statistics of R² marginal (R²_m) and R² conditional (R²_c) were calculated with the package MuMIn, where R²_m concerned the fixed part of the model, while R²c referred to both the fixed and the random part. When the resulting model had no random part ordinary R² (R²_o), and adjusted R² (R²_adj) were calculated, the latter penalised by the number of coefficients. RMSE calculation was performed using k-fold cross- validation of the models (k = 5). As the modelled water temperature variables are of great interest for the biology and ecology of lake organisms along the entire region, we used the models presented here to make present projections along the Pyrenees in a wide dataset of 2,267 lakes and ponds. To do so, we multiplied the coefficients of the mixed models with the spatial data of these water bodies and the seasonal air temperature in Lake Redon AWS for the period between 2006 and 2015. # Results ## Thermic characteristics of Pyrenean lakes and ponds Thermic variables showed a wide range in the monitored Pyrenean water bodies during the ice-free period. Tmean ranged between 5.9 and 15.3°C, being 11.0°C the average temperature for all water bodies. Tmax reached a value of 27.3°C, while the lowest Tmax registered was only 8.4°C, having a mean Tmax of 17.9°C. DTR had a mean of 1.5°C and ranged between 0.6 and 4.6°C. Tosc (maximum: 2.8°C) was smaller than the maximum DTR, but was similar for the mean and minimum (1.1 and 0.5°C respectively). Some water bodies had low ADD, having a minimum of 18.1°C day ADD7.6 and 41.6°C day ADD4, while mean ADD (697.2 ADD7.6 and 1,161.8°C day for ADD4) and maximum ADD (2,126.5°C day for ADD7.6 and 2,931.5°C day for ADD4) were much greater. ## Models of thermic variables Mixed models of thermic variables showed good fits (; R²_c, 0.53–0.84). Only the model for DTR fitted better without the random part in the model, and therefore, it was adjusted as a multiple regression model and resulted in a lower fit (R²_o 0.45). The rest of the models included the year as a random variable except Tmean, which only included the water body type (lake or pond). The water body type was also included in the model for ADD7.6 and ADD4 together with the year. The models for ADD7.6 and ADD4 showed the best performance (R²_c 0.84–0.82). They were mostly explained by the fixed part of the model; R²_m was the highest of all the models (0.75). The Tmean model had the lowest adjustment for the fixed part (R²_m 0.52), while Tmax had the most important effect of year and water body. Tosc had the lowest performance among the mixed models (R²_c 0.53). The primary variable explaining the variation in thermic variables was altitude, whose mixed models coefficients were between -0.27 and -0.81 depending on the response variable: Altitude showed an inverse relationship with all thermic variables; it affected more ADD7.6, ADD4 and Tmax and it had a lower effect on DTR and Tosc. In contrast, lake area was more important than altitude in explaining DTR and Tosc (-0.61 and -0.69 respectively). It also had an inverse relationship with all variables. Larea was also important in explaining the maximum temperature (-0.42). Geomorphology had its most important role in mean temperature (-0.71), but it also affected ADD7.6, ADD4 and Tmax (-0.21 - -0.38). Greater Tcatchment/Larea and Dcatchment/Tcatchment produced a decline in water temperature, and therefore, an inverse relationship with ADD7.6, ADD4, Tmean and Tmax but not on DTR or Tosc. Tcatchment/Larea can be considered as a surrogate for water renewal time. This variable explained most of the effects of geomorphological variables. Radiation had a similar positive effect on all thermic variables (0.16–0.26). Geographic coordinates had a limited effect (-0.09–0.15), with an increase of ADD7.6, Tmax and DTR to the west and a decrease of ADD4 to the north. Interannual variability in air temperatures, represented by the seasonal air temperatures in Lake Redon, had a substantial positive influence on ADD7.6, ADD4 and Tmean positive interaction with altitude on Tmax, and no effect on DTR and Tosc. Of the two seasons considered in the models, spring and summer, spring showed the strongest effects of air temperature on the thermic variables. Interannual variability in air temperature accounted for a lower effect than the sum of the other variables which depended on the lake characteristics, such as altitude, lake area, morphology or radiation. Nevertheless, it has to be taken into account that the variability attributed to the year inside the random fraction of the models represents interannual variation due to other undetermined meteorological causes out of mean air temperature, which would include wind, precipitation, and cloudiness. ## Projections of thermic variables Projections in the lakes and ponds for the measured years using k-fold cross- validation provided a value of RMSE for the models. Predicted values were in good agreement with the observed ones. Predictions for ADD had lower errors than for Tmax and Tmean while DTR and Tosc showed higher prediction errors. Predicted time series followed the same temporal pattern as observed thermic variables and generally observed values were within the error of the predicted thermic variables or close to it. Biases of the predictions were related to extreme events and to the fact that DTR and Tosc were not explained by interannual variability in air temperatures. These models allowed us to project the thermic variables for the whole set of lakes and ponds along the Pyrenees (n = 2,267) during the time of this study, using measured seasonal air temperatures in Lake Redon AWS from 2000 to 2019 and the spatial information about these water bodies. Mean seasonal air temperature in Lake Redon was 0.77°C in spring and 10.55°C in summer. Maximum mean temperature in spring was 2.62°C in 2005 and, in summer, 13.47°C in 2003. Projections for lakes and ponds for these two decades are included in S5 and. In average, for all these water bodies Tmean was 11.36°C, ADD7.6 was 820°C days, while ADD4 was 1,315°C days, Tmax was 19.21°C, DTR was 2.08°C and Tosc was 1.52°C. Maximum ADD7.6 projected in Pyrenean lakes was 3,09°C days and ADD4 was 3,670°C days. Maximum Tmean in those lakes was 21.55°C, while maximum Tmax was 40.78°C, which was far beyond measured maximum temperatures. These extreme values were found in six small (8–63 m²) and low altitude ponds (1479–1694 m). Maximum projected DTR was 5.58°C and Tosc was 3.01°C. Minimum Tmean was 5.64°C, and minimum Tmax was 7.57°C. There were 107 lakes with negative ADD7.6, which were at altitudes between 2362 and 2978 m.a.s.l. and one outlier with negative ADD4, which was the highest lake. Conceptually, negative ADD are not possible, but they are a result of extrapolations in extremely cold lakes, which would have ADD close to zero. Also three lakes showed negative DTR which were the three big lakes, with areas between 41 and 88 ha. Minimum Tosc was 0.017°C. Histograms of the projections are represented for ADD7.6, ADD4, Tmean, Tmax, DTR and Tosc. Median values are represented with red lines. See for the description of the variables. Here the variables are not transformed. # Discussion Among all the predictor variables, altitude was the factor that explained more variability in lake surface water temperature in the Pyrenean mountain range, since lake surface water temperature decreases linearly with altitude as the air temperature adiabatic lapse rate, but at a higher rate, as found in the Alps. This was the case for accumulated degree days, mean and maximum water temperatures. This relationship has been previously described in lake districts situated in mountain ranges at daily and monthly scales. In the case of maximum temperature, a weak relation with altitude was found in Canadian lakes, probably because they considered a reduced altitude range compared to ours (1053–2978 m; ). In contrast, we found that altitude was the main explicative variable for maximum temperatures in high mountain lakes. Previous studies have successfully modelled accumulated degree days in lowland areas (Midwest) of North America. In this study, we show the relevance of altitude on accumulated degree-days in mountain areas. In the case of diel temperature range and temperature oscillation, lake area was the most explicative variable. This was also found by Woolway et al., and it is a result of lake area being proportional to lake depth and mixing depth, which increase thermic inertia, and therefore, reduce diel temperature range. Besides, we have found that, in lakes located in high mountain ranges, altitude was also a significant variable explaining differences in diel temperature range together with other less importance variables, as longitude and insolation. Lake area was also relevant to model accumulated degree-days, mean and maximum temperatures. Similarly to our results, but at daily scales, lake area smoothed water temperature comparing to that of air, as the lake morphometry is related to the heat exchange between water and air and the heat storage. In addition to lake area, morphological variables related to lake catchment features were found to be important for modelling water temperature. The total catchment to lake area ratio can be used as a proxy of the inverse of water renewal time, and consequently, higher ratios corresponded to lower accumulated degree-days, mean and maximum temperature of the water bodies, as faster flushing and increased water coming from snowmelt would prevent a warming up of the epilimnion of the lakes and ponds. The ratio between direct and total catchment may be inversely related to water retention in upstream lakes and ponds and increased heating in these water bodies, and thus higher ratios corresponded to lower water mean and maximum temperatures. Taking into account these variables may help us to understand the potential effects of precipitation on lakes’ water temperatures, as catchment morphology variables have not been considered in the development of empirical models of water temperature. In the case of precipitation, it is commonly used in mechanistic models of water temperature, but it is rarely used in empirical ones, and effects of precipitation were not even found. In the Mediterranean region, precipitation is expected to decrease (IPCC, 2013), which could lead to lower water renewal times resulting in even higher warming in lakes. Testing the effects of precipitation in conjunction with the catchment morphometries could help us understand the thermic characteristics of lakes. This would require a good understanding of the precipitation spatial and temporal distribution. Solar radiation, as sun hours reaching the water bodies’ catchment, had a notable positive effect on lake temperature. In our study, we focused on the local differences in solar radiation instead of seeking temporal variations. Spatial modelling allowed us to find the effects of latitude, altitude and the shading by topography on incoming solar radiation into the lakes and catchments, and thus on thermic variables of the lakes. It is already known that the effect of shading by topography was of great importance in high mountain ranges, where shaded lakes were cooler than expected by their altitude. Different approximations have been considered when introducing solar radiation into water temperature models. For instance, theoretical clear-sky solar radiation has been used to model daily water temperatures in Greenland, where altitude and latitude were used, but not the shading by topography. Another option is the use of radiation downscaled from global climate models (GCMs), in such an approximation, spatial differences can be assessed although topographical shading may not be considered. Using solar radiation in empirical lake surface water temperature models showed contrasting results. Whereas solar radiation was not a good predictor for maximum temperature in Sharma et al., spatial and temporal changes in solar radiation were related to lake warming trends worldwide using satellite data. The spatial variability in incoming solar radiation can vary from one study to another, as latitude, altitude and shading depend on the scale and the region studied. In the case of the Pyrenees, latitude might not be very relevant as the mountain range is oriented W-E, but altitude and topographical shading have an important effect on spatial variability of incoming solar radiation. Whereas radiation derived from GCMs, or from satellite data are a good option to account for temporal and local variability in solar radiation, our work shows the important effects of topography on mountain regions, when modelling water temperatures, as it has a great spatial precision, and it is commonly neglected. Integrating methods to account for more precise temporal and spatial calculations of radiation arises as an excellent option to further improve models like the ones presented here. Interannual variability in air temperatures was represented by the seasonal temperature data series of Lake Redon, which represented air variation at a regional scale. This temporal change in air temperatures explained accumulated degree-days and mean and maximum temperatures in higher altitudes, whereas it did not explain diel temperature range or temperature oscillation. This may be related to a higher unpredictability of the last three variables, and due to a probable increase in minimum water temperatures during summer, for diel temperature range and temperature oscillation. Spring air temperatures explained more variability of the thermic variables than summer temperatures. The latter was only significant for mean water temperature. This may be due to spring temperature effect on thawing and ice-off and the onset of lake surface water temperature warming, as in the Tatra mountains, where the lake surface water temperature began to show an altitudinal gradient in late spring. Mean annual temperature, in addition to summer temperature, was significant to explain water maximum temperatures in Canada, indicating that air temperature is influencing lake surface water temperature also beyond summer, especially during spring. Other variables which can contribute to the interannual variability of the thermic variables are wind, precipitation and cloudiness, which were not explicitly included in these models, but are represented by the random variability due to the year. This variability is especially high in the case of maximum temperatures. Future research can be conducted to disentangle the effects of these meteorological variables, as their degree of coherence is lower than for air temperatures, so their spatial variation has to be well known before considering them in predictive models. Lakes and ponds at low altitudes would show a greater increase in accumulated degree-days over the 7.6°C threshold in response to spring temperature increase, as we found a negative interaction between altitude and spring air temperature on accumulated degree days, which meant contrasting effects of interannual air temperature on higher and lower lakes and ponds. This was likely a result of a greater advance in ice-off date at lower than at higher altitude, causing high altitude lakes to be disconnected during more time from spring air temperature, as ice-cover has an insulating effect. In contrast, maximum water temperatures, which take place at the middle of summer, are expected to increase more at lakes from higher altitude, since they depended on the positive interaction between interannual air temperature and altitude. This may be the consequence of a greater increase in air temperatures in altitude, as described in a review by Pepin et al.. Also, accumulated degree-days over 7.6°C threshold in ponds (surface areas \< 0.5 ha) had a steeper decreasing slope with altitude than in lakes (\> 0.5 ha), as lake area interacted positively with altitude on accumulated degree-days, an interaction which was also found for mean and maximum temperatures and diel temperature range. Low altitude ponds may accumulate more heat because of earlier ice-off and a quicker response to air temperature rise in comparison with low lakes. High altitude ponds, in turn, may be more affected by cooling in autumn, whereas lakes may remain warmer at high altitudes because of higher thermic inertia. Besides, differing warming trends between small and big lakes have been described, as small lakes benefit of higher wind sheltering, smaller fetch, and thus they may show stronger stratification and a warming trend in the epilimnion and smaller or opposed trend in the hypolimnion. Furthermore, the transparency of the lakes has a primary role in conforming the thermal structure of the lake, as the radiation is absorbed in the surface of the lake enhancing the stratification strength, and increasing radiative loss to the atmosphere. This process is especially relevant in small lakes, where wind stirring has less importance. Differences between thermal characteristics of lakes and ponds are relevant as these habitats can have different species composition, as has been found between lakes and pond from the Pyrenees. Lake thermal structure is also relevant as it will affect differently the species depending at which depth they inhabit. Therefore, a deeper knowledge on different warming trends is fundamental to assess potential effects on the organisms of these ecosystems. In the Pyrenees, further research on the thermal sturcture, the hypolimnion trends, the differences in warming between ponds and lakes and the processes involved should be done to completely assess the effects of temperature changes on the whole ecosystem. We have shown that, for a nine-year period, the effects of variables which depend on the spatial distribution of the lakes (altitude, lake area, catchment morphology and radiation) were more important than the effects of interannual air temperature differences for all thermic variables. These results may vary in a wider time window, as the measured period was of slight temperature increase in comparison to the last half-century, when temperature had a steeper increase, and may continue to do so with the current scenario of climate change, thus rising upper temperatures. However, mean summer air temperatures from this study ranged between 8.2 and 11.3°C, while it was found to be between 6.1 and 11.6°C from 1781 to 1997 in the Pyrenees, derived from instrumental records, comparable to the range found in our study. On the other hand, air temperatures were not significant in explaining diel temperature_range and temperature oscillation, which indicates an increase also in minimum water temperature during summer. Therefore, they would not be expected to change in the future. Temperatures increase can drive changes in water bodies’ communities, such as functional traits, composition, biomass or abundance. For instance, in Canadian lakes, it was found that water temperatures were negatively related to zooplankton body size, a similar effect to the one produced by fish predation, and these combined effects would be non-additive. Therefore, lakes warming would favour smaller zooplankton species. Combined fish predation with warming may increase small zooplankton by predation on large zooplankton, causing an increase in producers’ abundance due to less efficient consumption of small zooplankton. Moreover, higher altitude lakes would be more sensitive to warming since they have less functional diversity than lower ones. This functional diversity would move upwards, as species may change their distribution ranges to higher elevations. At high elevations, cold stenothermal species are more vulnerable as they have a restricted distribution range, and they could be lost as a consequence of climate change. A temperature increase can also advance zooplankton phenology. Accumulated-degree-days can be modelled to predict recruitment and abundance of fish, which may be benefited by increasing accumulated degree-days. In this sense, in Pyrenean lakes, species, both natural and introduced, could develop in higher altitude lakes as a result of the accumulated-degree-days increase. The present paper opens the possibility to define the potential thermal habitat of lacustrine species in this mountain range. In contrast, maximum temperatures may induce thermal stress to certain species, influencing thus species composition. These effects would have a more significant impact on higher water bodies where we have foreseen a greater increase in maximum temperatures. In addition, small increases of maximum temperatures in low and warm lakes may cause thermal stress in organisms, as their maximum temperatures are already high. Consequently, lower distribution of species in the Pyrenean range may be limited, while seeing their upper potential habitat increased. The knowledge about thermic variables and the models developed here enable making spatial extrapolations of water thermic variables, which are fundamental for disentangling ecological and conservation issues in a global change context. # Supporting information We want to acknowledge all the people who have helped during the fieldwork, Danilo Buñay and many others, Guillermo Carrasco and Joan Garriga, who helped with the programming. We would also like to acknowledge the Parc Nacional d’Aigüestortes i Estany de Sant Maurici and the Parc Natural de l’Alt Pirineu for their support. Servei Meteorologic de Catalunya provided air meteorological data, and Institut Cartogràfic Geològic de Catalunya, Instituto Geogràfico Nacional and Institut Géographique National provided GIS data. [^1]: The authors have declared that no competing interests exist.

# Introduction The cervix is a dense fibrous tissue that is located at the lowest part of the uterus. It is cylindrical in shape with average dimensions of 3 cm long and 2.5 cm in diameter. The mechanical function of the cervix is crucial for a term pregnancy (defined as a pregnancy that extends beyond 37 weeks of gestation). Cervical mechanical function has two roles: 1) prior to term it must remain closed and resist the increasing mechanical load from the growing pregnancy and 2) at time of parturition it must be soft to deform and dilate to allow for delivery of the fetus. To accommodate this drastic dilation of the cervix at time of delivery, the extracellular matrix (ECM) of the tissue must drastically remodel, reorganize, and soften during gestation. The timing and characteristics of this remodeling behavior is currently an active research focus because it is hypothesized that premature remodeling in pregnancy can lead to a preterm birth, a leading cause of neonatal death or significant neonatal morbidity. In an effort to characterize the remodeling behavior of human cervical tissue, the objective of this study is to measure and quantify the collagen fiber orientation and dispersion (e.g. ultrastructure) of non-pregnant and term pregnant cervical tissue using optical coherence tomography (OCT). The premature change in mechanical properties of the cervix induced by alterations of cervical tissue ECM content and ultrastructure is thought to contribute to cervical failure leading to preterm birth. Cervical collagen exists as fibers in a hierarchical network embedded in a viscous ground substance of negatively charged glycosaminoglycans (GAGs) and other proteins. The collagen (types I and III) makes up 34 to 77% of the dry weight, with evidence from human tissue studies showing that this dry weight content remains constant during gestation. The cervical collagen fiber ultrastructure has been studied using X-ray diffraction, magnetic resonance diffusion tensor imaging (MR DTI), second harmonic generation imaging (SHG) and optical coherence tomography. In general, the collagen fiber network is reported to be anisotropic with different preferred orientations in distinct anatomical regions within the cervix. X-ray diffraction studies found three radial zones of preferentially aligned collagen fibers, where collagen fibers along the outer edge and next to the inner canal predominately run parallel to the canal and collagen fibers in the mid-stromal area run circumferentially around the canal. The dispersion of these fibers was found to be aligned in a narrow range of 50 to 70 degrees within each zone. MR DTI confirmed the inner and middle zones but the outer zone was not resolved. More recent preliminary SHG data from Feltovich *et al*. and OCT studies from our group reveal a large band of circumferential fibers extending to the outer edge of the tissue. To date, it is unclear how this ultrastructure evolves with pregnancy. In this study, we use OCT, a non-invasive imaging technique based on the principle of low coherence interferometry, to image and characterize the cervical fiber ultrastructure. Using OCT, the sample morphological information in depth is obtained by interfering backscattered photons from a sample irradiated with a broadband low coherence source with a reference beam. A typical OCT system can achieve a high axial resolution at the micron level, a penetration depth up to 2 mm, and video rate data acquisition, and thus emerges as a promising image modality to image a variety of organ systems. In particular, efforts have been made to image cervical tissues for cancer detection by analyzing the layered structure of the epithelium, the basement membrane, and the stroma. In addition to cancerous structure, OCT can be also used to image the collagen fiber network. Our previous work has demonstrated the feasibility of imaging entire axial cervical slices. This was enabled by an image-stitching algorithm to increase the field of view (FOV), encompassing entire axial slices (\~3 cm × 3 cm), which allowed for assessment of collagen fiber orientation trends and for identification of unique anatomical regions. Our previous methods for fiber orientation estimation targeted extracting dominant fiber orientation in a subregion, however it could not provide detailed information about non-dominant orientations, resulting inaccuracy in the evaluation of fiber dispersion. Building on our previous OCT investigation of the human cervix and the work of others, we report collagen fiber orientation maps of whole, unfixed, axial cross-section slices of non-pregnant and pregnant human cervical tissue to visualize anatomically-relevant trends. Due to the complex structural environment of the cervix, we hypothesize that the fiber orientation and local dispersion is heterogeneous with regions determined by anatomic location. We postulate that the homogeneity of fiber orientation and local dispersion will depend on four anatomic quadrants of the cervix (posterior, anterior, left and right) and will depend on the radial location from the inner canal. In this paper, we present the methodology of OCT imaging on human cervical tissue, and describe the fiber orientation maps and collagen fiber orientation distribution and dispersion across four quadrants and different radial distances from the inner canal. # Methods ## Sample collection and preparation Thirteen human cervices were collected from consented hysterectomy patients by an IRB approved protocol at Columbia University Medical Center. Among the cervices, 11 were from non-pregnant (NP) patients undergoing hysterectomy for benign indications and 2 were from pregnant (PG) patients undergoing cesarean hysterectomy due to abnormal placentation. Patient age ranges from 36 to 49 and parity number from 0 to 5. The cervices were sliced perpendicular to the inner canal immediately after hysterectomy using a custom-built slicer. The thickness of each slice was 3–5 mm. Axial slices within the upper half of cervix were excised. In this study, we analyzed the slice that is closes to uterus for each cervix. All samples were kept on dry ice and then stored at -80°C for later imaging. This study was approved by the Columbia University IRB, with an IRB protocol Number: IRB-AAAL4005. Study participants gave their consent by signing a written consent form that was approved by the Columbia University IRB. A more detailed protocol used for sample collection and preparation is described in our earlier work. ## OCT scan and fiber recognition algorithm Before OCT imaging, cervical slices were thawed in phosphate buffered saline (PBS) overnight at 4°C, and the surface closer to the internal os was microtomed. During the imaging procedure, the cervical slice was laid on top of a gauze soaked in PBS to keep the tissue hydrated. Samples were imaged using a commercial OCT system, Telesto I (Thorlab GmbH, Germany). The system is an InGaAs based system with its source centered at 1325 nm and a bandwidth of 150 nm. The axial and lateral resolutions are 6.5 μm and 15 μm in air, respectively. In our experiments, each volume consisted of 900 × 900 × 512 voxels, corresponding to a tissue volume of 4.5 mm × 4.5 mm × 2.51 mm (in air). Samples were placed in a linear translation stage underneath the objective. For each sample, we obtained multiple volumes. There was an overlap proportion of at least 10% between two adjacent volumes. A white light camera obtained an image of the sample corresponding to the OCT FOV. The camera images and OCT images were calibrated by the default factory setting. Volumetric data was stitched based on the shift invariant feature in camera image within the *en face* plane and surface information of the OCT data in the axial direction. Upon generating the three dimensional data, parallel *en face* images were obtained 245 μm beneath the surface to perform 2D fiber directionality and dispersion analysis. Fiber orientations were extracted for each pixel (5 μm × 5 μm) by optimizing a pixel-wise fiber orientation method for OCT image datasets. In each *en face* image, the collagen fiber region was masked based on the signal to noise ratio. Then, the image was enhanced through histogram stretching. The image was sharpened by second order Butterworth high pass filter and subsequently denoised by a median filter. A weighted summation scheme was utilized to determine the fiber orientation at each pixel over the entire region. For a pixel of interest, *p*₀, there were multiple candidate directions *α*_j towards its neighboring pixels, *p*₁ and *p*_2. A weight was assigned to each candidate direction as following: $$w_{j} = w_{i} \times w_{d}$$ $$w_{i} = \frac{1}{3} - std\left( {p_{1},p_{0},p_{2}} \right)$$ $$w_{d} = \frac{1}{dist\left( p_{0},p_{2}\ or\ p_{1} \right)}$$ The weight was determined by two factors, *w*_*i* and *w*_*d*. The first factor, *w*_*i*, was the intensity variations between the pixel of interest (*p*_*0*) and its neighboring pixels (*p*_*1* and *p*_*2*) along a particular direction. The second factor, *w*_*d*, was the corresponding distance between the pixel of interest (*p*_*0*) and the neighboring pixel (*p*_*1* or *p*_*2*). The direction, *α*, of target pixel *p*₀ is determined by the weighted circular mean of all direction candidates as described in : $$\alpha = arg\left( {\sum_{j = 1}^{N}{w_{j} \times}}exp\left( i\alpha_{j} \right) \right)$$ Where N is the number of direction candidates around pixel *p*₀. Given the direction information of each pixel, we generate the directionality map of the whole OCT image. Based on the pixel-wise orientation information, we obtained the directionality map of collagen fibers within the *en face* image. The directionality map was further divided into sub-regions of 400 μm × 400 μm along the radial direction in the four anatomical quadrants from inner canal to outer edge. In each 400 μm × 400 μm subregion, a 2D von-Mises probability density function, $$P(x) = \frac{e^{b{\cos{({x - \theta})}}}}{2\pi I_{0}(b)},$$ was fit to the pixel-wise orientation data to determine the fiber direction *θ* and the concentration parameter *b*. The two parameters were estimated by a least squares method using MATLAB (MathWorks, R2014b) function (fit). *I*₀(*b*) is a modified Bessel function of the first kind of order 0. Here, *θ* ∈ \[0,2*π*) is the dominant fiber direction and *b* \> 0 is the concentration parameter. The concentration parameter *b* describes the dispersion level of Von Mises distribution. When *b* approaches 0, the distribution gets closer to isotropic (circular in 2D case), and as *b* increases to infinity the distribution gets closer to perfectly aligned fibers. In other words, *b* is inversely related to fiber dispersion where a low *b* describes a high fiber dispersion and a high b describes a low fiber dispersion. ## Statistical analysis A group of analysis of variance (ANOVA) tests were performed in MATLAB using one-way ANOVA function (anova1) and multiple comparison function (mulitcompare) to compare the von-Mises fiber ultrastructure parameters (b and θ) between NP and PG specimens and among NP specimens with different parity. The data normality was verified by Kolmogorov–Smirnov test in MATLAB (kstest function) before performing the ANOVA analysis. The homogeneity of these fiber ultrastructural parameters within individual sample slices were assessed by comparing results between circumferential quadrants, inner and outer radial zones. In the circumferential direction, the cervical slice was divided into four anatomical quadrants. In the radial direction, the cervical slice was divided into inner and outer zones. The border between the inner and outer radial zones was manually determined by differentiating the distinct patterns of fiber orientation of the two zones. The radial direction was also subdivided into 400 μm × 400 μm subregions as described by the pixel-wise fiber tracking method above. Parity is the number of times that a woman has given birth. We divided our patients into three parity groups: nulliparous patients (*n* = 2) who have never given birth, primiparous patients (*n* = 4) who have given birth once, and multiparous patients (*n* = 5) who have given birth two or more times. When *b* and *θ* were compared between different samples, averages were taken of the results from all 400 μm × 400 μm subregions within the quadrant and radial zone. When comparing *b* and *θ*, the variance of *b* and *θ* along radial direction within each quadrant and zone were also measured by calculating the standard deviation. All ANOVA tests were performed in MATLAB using the anova1 function where a *p*-value of 0.05 was considered statistically significant. # Results ## OCT *en face* images and fiber orientation maps The regional collagen fiber architecture of the upper half of the human cervix is depicted in 2D image a fixed axial distance below the surface in using the method in. In 9 NP tissue samples out of the 11 NP samples imaged, two radially zones are found with distinct fiber orientation characteristics. In these tissue slices there is an inner zone with collagen fibers preferentially aligned in the radial direction and an outer zone with collagen fibers preferentially aligned in the circumferential direction. The shape of the inner zone and the fiber orientations in this zone are highly affected by the shape of the inner canal. The inner zone can be relatively wide or narrow, where shows the widest inner zone which is around 30% of the slice radius and shows a narrower inner zone. In the 4 remaining slices, including both pregnant samples, there is no inner zone and the whole slice is dominated by circumferentially aligned fibers. For many slices), the inner canal opening aligns from left to right. Others have the inner canal opening aligning from anterior to posterior, or the inner canal is round or has an irregular shape. Based on the samples we examined, there is no clear relationship between the inner zone size and patient parity. We validated our pixel-wise fiber recognition algorithm on synthetic data in, where our algorithm accurately estimates the directionality of segments oriented at various orientations (A-B) and circular shape (C-D). The new pixel-wise fiber recognition algorithm is superior to the gradient-based method because the pixel-wise method is able to capture the existence of distinct fiber families at different orientations, especially non-dominant orientations. A directionality map using the pixel-wise fiber orientation algorithm of a cervical sample is shown in , in comparison with the OCT *en face* image in. The fiber distribution obtained from three subregions using the pixel-wise method and the gradient- based method in were compared in. In general, the estimated dominant direction using two methods approximate each other within each subregion. However, the gradient method is unable to capture the actual fiber distribution of the probability of fiber existence at each angle. A typical example of the pixel-wise method on a stitched OCT cervix image comprised of 24 OCT volumes is shown in. The original OCT image is an *en face* image 245 μm parallel to the cut surface as shown in. From the pixel-wise directionality map, such as, we observe a circumferential trend of fiber in the outer zone. From a zoom-in box in corresponding to a 4 × 4mm region, it shows fiber directions can vary dramatically within a small region. Similar circumferential trends and direction variation patterns are observed in all other cervical samples. 2D von-Mises distribution provides a close fit to the raw fiber dispersion data. Concentration parameter *b* can be as high as 0.820 and as low as 0.010 as shown in. For certain subregions, more than one family of fibers can be observed where the current 2D von-Mises analysis cannot capture these distinct fiber families. The fitting for multiple families of fibers will be discussed in discussion. ## The posterior and anterior cervix contains regions of preferentially-aligned collagen fibers The upper half of the cervix contains zones of preferentially-aligned collagen with distinct fiber directionality and dispersion properties in the posterior, anterior, left, and right quadrants. The dominant fiber directionality data *θ* for each of the 13 specimens averaged across radial subregions in the outer radial zone within each anatomical quadrant are represented in. For all 13 specimens, including NP and PG, the dominant fiber directionality in the posterior and anterior quadrants of the cervix is in the circumferential direction, with fibers circling around the inner canal. When comparing these anterior and posterior quadrants between specimens, the averaged directions *θ* are within a small range (≈35°) with one exception of the anterior of one PG sample. In the left and right quadrants, although the average dominant direction for all specimens is circumferential, dominant directions themselves are scattered within a larger range (≈140°), which indicates a higher variability between specimens in the left and right quadrants. The data normality is verified in every radial quadrant and inner/outer region since Kolmogorov-Smirnov test results accept the null hypothesis (lowest p = 0.28). When comparing 400 μm × 400 μm radial subregions within a single slice in the outer zone for NP specimens, the standard deviation (SD) of the dominant direction *θ* is higher in the left and right quadrants comparing to that in posterior and anterior quadrants. This higher SD means fiber changes orientation more dramatically along radial direction and the tissue is more heterogeneous. This difference is very significant in the outer zone (*p* \< 0.001) but not significant in the inner zone (*p* \> 0.983). The SD of *θ* in the outer zone of posterior and anterior quadrants of NP samples is different than the rest regions in NP samples as well as all regions in PG samples. If we group them this way, the difference is very significant (*p* = 3.4 × 10⁻⁶). ## The posterior and anterior of the outer radial zone of NP specimens have the lowest collagen dispersion The posterior and anterior quadrants of the outer zone in NP specimens have the lowest fiber dispersion in the 400 μm × 400 μm subregions (i.e. highest concentration parameter *b*) compared to other quadrants in NP samples and to all quadrants in the PG samples. The left and right quadrants of the NP samples have similar fiber dispersion properties compared to all quadrants of the PG specimens. Similar to the SD of *θ*, if we group the outer zone of posterior and anterior of NP samples together and group the remaining regions in NP samples as well all regions in PG samples together, the difference in *b* is statistically significant (*p* = 2.1 × 10⁻⁷, see). Within the PG samples there is no significant difference among quadrants. The variance in fiber dispersion, represented by the SD of concentration parameter *b*, between 400 μm × 400 μm radial subregions within a single slice are not significantly different. ## There is difference in dispersion between NP and PG but no difference is found between NP samples with different parity We found a statistically significant difference in *b* between NP and PG specimens in posterior and anterior in the outer zone. In further detail, among four quadrants, *b* of PG specimens has distinctly lower mean values in posterior and anterior than NP specimens, which suggests PG specimens have more dispersed collagen fibers. The difference is significant if we compare the combined posterior and anterior region and combined left and right regions (*p* between 0.006 and 0.045). Such difference is not found with the variance of *θ*. Among parity groups in NP specimens, there is no difference found for either *b* or *θ* (Figs). # Discussion In this paper, we present a regional OCT collagen fiber orientation and dispersion analysis of 13 fresh, unfixed human cervical slices. To measure local fiber orientation and dispersion, a new pixel-wise fiber orientation algorithm is developed for cervical tissue analysis. Based on this method, fiber orientation maps are generated to visualize and measure the tissue-level architecture of the upper cervix. In all of the cervical fiber orientation maps, there is a dominant outer radial zone of preferentially aligned collagen fibers circling around the inner canal where the posterior and anterior quadrants are more aligned than left and right quadrants. In 9 out of the 11 non-pregnant samples, there is an additional inner radial zone with fibers preferentially aligned in the radial direction that is perpendicular to inner canal opening direction. In this inner radial zone, though, the trends are difficult to study because this zone also includes mucous glands located around the inner canal opening, which cannot be differentiated from dense collagen fibers. The NP cervical tissue samples measured in this study have two regions with distinct fiber directionality and dispersion properties. The posterior and anterior of the outer zone is labeled Region 1 and the remaining parts of the cervix (left and right of outer zone and all inner zones) are labeled Region 2. For a NP cervix, Region 1 and Region 2 have different fiber dispersions between Regions and similar dispersions within each Region. However, when a NP cervix becomes a PG cervix, Region 1 will have a shift in the fiber dispersions so that the properties are similar to Region 2 while Region 2’s properties do not shift. In other words, Region 1 is more sensitive to pregnancy status and remodels more dramatically than that happened in Region 2 during pregnancy. The arguments above are verified by ANOVA test in Result section by comparing Region 1 in NP with Region 2 in NP and all Regions in PG. We believe the regional differences in collagen fiber properties within a single sample and between samples are influenced by the anatomical and loading environment of the cervix in the pelvic region. The cervix is the lower portion of uterus. The upper portion of the cervix, or the portio supravaginalis, lies above the vaginal attachment to the cervix. Cardinal ligaments attach this portion laterally (i.e. left and right), and the bladder lies anterior to the cervix separated by loose connective tissue. In pregnancy, the upper cervix is substantially loaded by the growing fetus. The positioning, symmetry, and shape of the uterus and cervix drive the patterns of cervical stress and stretch and can be vastly different for each person. Often in pregnancy the cervical axis is angled posteriorly from the uterine axis. This positioning leads to increased tissue loads and stretching in the anterior and posterior sections of the cervix. The angle of the cervix with the uterus can be a potential cause of the increased anisotropy in Region 1 of the cervix, and the fact anatomical factors vary widely between patients can explain the variability between samples. Related research of finite element analysis of human uterus and cervix also supports the heterogeneity of fiber dispersion we find between quadrants. The FEA analysis demonstrates that the collagen directionality and dispersion play a role in resisting physiological relevant deformation during pregnancy. Further studies with larger patient populations must be conducted to understand the mechanical loads on the cervix and cervical tissue remodeling behaviors during pregnancy. Many imaging modalities that have been used to study collagen fiber orientation and dispersion of hydrated soft tissues, with each method suited for different length scales and tissue sample preparations. Our OCT analysis presented here interrogates the tissue within 400 μm × 400 μm subregions across whole, hydrated, and unfixed axial tissue slices. In comparison, second-harmonic generation (SHG) microscopy is a standard method for the characterization of hydrated biological tissue at sub-cellular resolution. The FOV of SHG systems can be tens of microns to hundreds of microns, with a resolution of better than 2μm. With these features, SHG captures information on fibril-level structure and the “crimping” or waviness of fibrils. Small angle light scattering (SALS) uses laser light to image fiber orientation and dispersion. Collagen fibers can be distinguished from striated muscle, smooth muscle, and elastic fibers from the resulting angular distribution of scattered light because striated muscle, smooth muscle, and elastic fibers have different birefringence properties. Each reading covers \~ 4 × 4 μm area which is the size of a single or small group of collagen fibers. In the future, functional extensions of OCT can also allow for interrogating the anisotropy and birefringence properties of cervical samples using polarization sensitive OCT. We carefully compared our conclusion with a similar paper, which used x-ray diffraction to study the collagen ultrastructure in human cervix. The study presents detailed cervical collagen orientation and dispersion data from NP human 1 mm³ cube samples that have been scanned in three orthogonal directions. There are similarities and differences between this X-ray diffraction study and our OCT study. First, they found that near the internal os there are three zones of preferentially aligned collagen fibers—the inner zone, the middle zone, and the outer zone. The X-ray data found that the both inner and outer zone have longitudinally aligned fibers while the middle zone has circumferentially aligned fibers. Our study indicates a similar inner zone. Since we cut and scan axial slices, it is not possible to see longitudinal fibers in our OCT images. It may be possible that the radially aligned fibers seen in the OCT data are skewed longitudinal fibers that not perfectly aligned with the inner canal. The middle zone indicated in the X-ray study is similar to our outer radial zone of circumferential fibers. However, our OCT images show these circumferential fibers extend to the outer edge of the sample. They also measured the radial thickness of each zone, the inner and outer zones are 3–5 mm and the middle zone is 5–12 mm. The inner zone in our study ranges from 0 mm to 6 mm and the middle zone ranges from 6 mm to 14 mm, which is comparable to the results presented by them. Second, the fiber dispersion in is much tighter than that in our study. In’s finding, almost all the fiber are aligned within 90° while in our study there are always fibers throughout 180°. We believe reason for the difference lies in the difference of method including sample preparation (our samples were not fixed) and fiber recognition. This research presented in this paper has the following limitations. First, as discussed earlier, the collagen fiber network is three-dimensional but fiber orientation and dispersion were only studied in two dimensions. Longitudinal fibers cannot be verified in this research because OCT images were stitched in the plane that is perpendicular to the inner canal. Second, only the available slice that is closest to the internal os had been studied. We selected the first slice to start our research because the internal os is the location of premature funneling and maximum stress during pregnancy. The premature funneling is often followed by opening up from the rest of the cervical inner canal and preterm birth. As we found in different quadrants, it is highly possible that the cervix is heterogeneous in the longitudinal direction since the percentage and type of biological and chemical components have been found to be different along longitudinal direction and the inner zone was found to disappear as we approach to the external os. Third, due to the limited number consented patients, we have a smaller database comparing to research that uses animal tissue. Currently, our preliminary study on pregnancy is based on 2 specimens. We plan to collect more specimens and draw a more conclusive comparison in our future study. Also, we will keep increasing the number within nulliparous, primiparous, and multiparous groups. Fourth, PG slices were not necessarily at the upper cervix because both of our PG samples were from patients with accreta and it is difficult to distinguish the location of the internal os. In order to avoid tissue with accreta, cervical slices of PG patients were obtained at the most proximal location available so the slices could be from mid-cervix. Fifth, although our pixel-wise analysis can capture the patterns of multiple fiber families, the von-Mises based distribution fitting is not efficient for subregions with two dominant families. Since the only case (two families) that cannot be efficiently captured by von-Mises distribution accounts for about 5% of all subregions, von- Mises is a good distribution model overall. In the future, we will improve our method and develop a more generalized method for all fiber family patterns. # Conclusions In this study, we measured the heterogeneity of local fiber orientation and dispersion in human tissue slices from the upper cervix using an OCT pixel-wise fiber orientation algorithm. We found that human cervical tissue has a distinct collagen fiber ultrastructure where collagen fiber orientation and dispersion vary according to anatomical quadrants. We found that in non-pregnant tissue, the anterior and posterior quadrants have highly aligned circumferential collagen fibers that are less dispersed than the left and right quadrant. Overall, we found that the non-pregnant samples examined here had more aligned and less dispersed collagen fibers than pregnant tissue, and that there was no difference in collagen properties between non-pregnant samples of different parity. The OCT imaging and tracking algorithm presented here is suited for our application because it offers tissue fiber ultrastructure characteristics at a length scale appropriate for implementation into a previously developed fiber- based continuum material model for human cervical tissue. Additionally, the whole sample fiber maps inform the implementation of tissue architecture into large-scale finite element models of pregnancy. Lastly, OCT is a nondestructive technique, which allows for ultrastructural, biochemical, and mechanical analysis to be conducted on a single sample. In future work, multiple slices from internal os to external os will be analyzed to look for trend of fiber dispersion along longitudinal direction, mechanical tests will be conducted to determine corresponding material behavior, and the structural importance of the regional ultrastructural properties of the cervix will be explored in finite element models of human pregnancy. # Supporting Information Research reported in this publication was supported by the National Science Foundation CAREER award (CMMI 1454412) to KMM and the National Institutes of Health (DP2) to CPH (1DP2HL127776-01). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Science Foundation nor the National Institutes of Health. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** KMM JYV RJW. **Data curation:** WY YG. **Formal analysis:** WY KMM CPH. **Funding acquisition:** KMM CPH. **Investigation:** WY YG. **Methodology:** WY YG KMM CPH. **Project administration:** KMM CPH. **Resources:** KMM JYV RJW CPH. **Software:** YG CPH. **Supervision:** KMM JYV RJW CPH. **Validation:** KMM CPH. **Visualization:** WY YG. **Writing – original draft:** WY YG KMM CPH. **Writing – review & editing:** WY YG KMM JYV CPH.

# Introduction The filamentous fungus *Aspergillus fumigatus* is one of the most common human fungal pathogens found infecting a large population of immunodepressed patients. This group includes individuals with hematological malignancies, those with genetic immunodeficiencies, patients infected with HIV, and cancer patients treated with chemotherapy. This immunodepressed population is currently increasing due to the higher number of organ transplants performed, immunosuppressive and myeloablative therapies for autoimmune and neoplastic diseases, and the HIV pandemic. The mortality rate resulting from *A. fumigatus* infections in immunodepressed patients ranges from 40% to 90%. In a previous report we demonstrated that the global regulatory *velvet* gene *veA* controls *A. fumigatus* production of conidia, the main inoculum during infection, and production of gliotoxin, a compound with immunosuppressive properties also found to inhibit phagocytosis in macrophage and to induce apoptosis. *veA* orthologs have been identified and characterized in other fungi including other *Aspergillus* species, such as *A. flavus*, *A. parasiticus* and the model filamentous fungus *A. nidulans*. These previous studies provided abundant evidence of the role of *veA* as a regulator of both fungal morphological development and secondary metabolism. In 2003 our group described for the first time the role of *veA* as a global regulator of secondary metabolism in *A. nidulans*, including production of the mycotoxin sterigmatocystin. *veA* also regulates the biosynthesis of other mycotoxins, including aflatoxin, cyclopiazonic acid and aflatrem in *Aspergillus flavus*, the synthesis of trichothecenes in *F. graminearum*, and the production of fumonisins and fusarins in *Fusarium* spp, specifically *F. verticillioides* and *F. fujikuroi*. However, *veA* also controls the synthesis of other secondary metabolites known for their beneficial medical applications, for example, the beta-lactam antibiotic penicillin in *A. nidulans* and *P. chrysogenum* as well as cephalosporin C in *Acremonium chrysogenum*. In-depth studies of *A. nidulans veA* and its gene product also revealed mechanistic details of its mode of action. For instance, it is known that the VeA protein is transported to the nucleus by the KapA α-importin, and that this transport is promoted in the absence of light. In the nucleus, VeA interacts with light-sensing proteins that also affect secondary metabolism and fungal differentiation, such as the red phytochrome-like protein FphA, which interacts with the blue- responsive proteins LreA-LreB. In the nucleus VeA also interacts with VelB and LaeA. VelB is another protein in the *velvet* family, and LaeA is a chromatin modifying protein that, like VeA, is required for the synthesis of numerous secondary metabolites. In addition, a LaeA-like putative methyltransferase was also described to interact with VeA. A microarray-based transcriptome study showed that *A. fumigatus laeA* affects the expression of 13 secondary metabolite gene clusters; however, at that time the extent of *veA* regulation of the activation of secondary metabolite gene clusters was mostly unknown. With the goal of elucidating the full extent of *veA*-regulation of the *A. fumigatus* genome, particularly with respect to genes involved in secondary metabolism, we performed RNA sequencing analyses and chemical characterization of *A. fumigatus* cultures, obtaining results consistent with a global regulatory pattern. This study also contributes to uncovering the regulation of novel secondary metabolite gene clusters in *A. fumigatus*. For example, an important discovery in our study is that *veA* and *laeA*, both of which encode *velvet* complex components, regulate the recently discovered gene cluster responsible for the synthesis of fumagillin. Fumagillin has been intensely studied due to its potential in the treatment of amebiasis, microsporidiosis and most recently, for its anti-angiogenic activity as inhibitor of the human type 2 methionine aminopeptidase (MetAP2). # Materials and Methods ## Strains and culture conditions *Aspegillus fumigatus* strains used in this study are listed in. Fungal strains were grown on Czapek Dox media (Difco), unless otherwise indicated, and supplements for the corresponding auxotrophies as needed. Solid medium was prepared by adding 15 g/liter agar. Strains were stored as 30% glycerol stocks at -80°C. ## RNA extraction Total RNA was extracted as previously described. Briefly, conidia from the wild type, deletion *veA* (∆*veA*), complementation and over-expression *veA* (OE*veA*) strains were inoculated in Czapek-Dox (approximately 10⁷ spores/mL) and grown as liquid stationary cultures at 37°C in the dark. Mycelia were collected 48 h and 72 h after inoculation and RNA was extracted using TRIzol (Invitrogen) following the manufacturer’s instructions. RNA samples were further purified using QIAgen RNeasy mini kit as previously described. ## Transcriptome analysis ### Genome and transcriptome sequence versions All *A. fumigatus* Af293 sequences used are from sequence version s03-m04-v01 from the Aspergillus Genome Database (AspGD). ### Library preparation and RNA sequencing RNA-Seq libraries were constructed and sequenced at the Vanderbilt Genome Sciences Resource using the Illumina Tru-seq RNA sample prep kit as previously described. In brief, total RNA quality was assessed via Bioanalyzer (Agilent). Upon passing quality control, poly-A RNA was purified from total RNA and the second strand cDNA was synthesized from mRNA. cDNA ends were then blunt repaired and given an adenylated 3’ end. Next, barcoded adapters were ligated to the adenylated ends and the libraries were PCR enriched, quantified, pooled and sequenced an on Illumina HiSeq 2000 sequencer. ### Read alignment and quantification of gene expression Illumina TruSeq adapters were trimmed from the 3’ end of reads using the scythe software package (available from Buffalo, V. at https://github.com/ucdavis- bioinformatics/scythe), and low-quality bases were trimmed using the sickle software package (available from Joshi, N. at https://github.com/ucdavis- bioinformatics/sickle). Reads were aligned to the transcriptome using the bowtie read alignment software for single-end reads, with the maximum mismatches per read set at 2 and a seed length of 28. Read count per gene was calculated using SAMtools idxstats software. For each sample, gene expression was quantified using the reads per Kilobase of exon per million mapped reads (RPKM) metric. Differential expression was calculated between ∆*veA* and wild type, between OE*veA* and wild type, and between the complementation and wild-type strains. Two cutoffs were used to determine differentially regulated genes. The first cutoff compared the fold difference between genes by calculating the relative RPKM (rRPKM = RPKM_sample1/RPKM_sample2) for each gene. The second cutoff compared the proportion of reads mapping to a gene in different samples using Fisher’s exact test with Bonferroni’s correction for multiple comparisons. A gene was considered differentially regulated if the log₂ rRPKM value was equal to or greater than 2 and the Bonferroni- corrected Fisher’s exact *p*-value was less than 0.05. ### Gene ontology categorization The gene ontology (GO) categorizations of differentially regulated genes were compared against the set of non-differentially regulated genes to identify GO categories that were specifically enriched in differentially upregulated or downregulated gene sets in the three strain comparisons. GO categorizations for each gene were obtained from the AspGD’s GOSlim mapper for *A. fumigatus* Af293. AspGD’s GOSlim mapper contains higher-order GO terms for the process, component, and function sections of GO. All comparisons were performed using Fisher’s exact test with Bonferroni’s correction for multiple comparisons. ### Sliding window analysis We used a sliding window analysis to determine clusters of genes that were upregulated or downregulated in the *A. fumigatus* genome in ∆*veA* versus wild type and OE*veA* versus wild type comparisons. Briefly, each gene was encoded as upregulated, downregulated or not significantly differentially regulated according to our specified differential regulation cutoffs. Then, we calculated the cumulative binomial probability of observing every window of 24 genes along each chromosome. To account for multiple comparisons, we used an empirically derived false discovery rate (FDR) by randomly permuting the expression data 1,000 times and running the sliding window analysis on each permuted dataset. Our FDR cutoff was set conservatively at 0.01. After all clusters below this FDR cutoff were found, genomically overlapping windows were collapsed into a larger cluster. Due to the window size, regions at the beginning or end of these master clusters may contain stretches of non-differentially regulated genes. We have reported all clusters from the first to last differentially regulated gene found for each cluster and the start and end genes of all significant clusters. ## Quantitative RT-PCR analysis One microgram of total RNA was treated with RQI Dnase to remove possible DNA contamination. Then cDNAs were obtained by reverse transcription using Moloney murine leukemia virus (MMLV) reverse transcriptase (Promega). Quantitative Real Time PCR was performed with an Agilent MX3000p thermocycler using SYBR green Jump Start *Taq* (Sigma). The primers used for gene expression analysis are listed on. ## Metabolome analysis ### Extractions of secondary metabolites Secondary metabolites were extracted as previously described. Briefly, liquid Czapek-dox stationary cultures of the wild type, ∆*veA*, complementation and OE*veA* strains were grown as described. Supernatants were collected by filtration through sterile Miracloth™ (Calbiochem, USA) from 72 h and 120 h cultures. Fifteen mL of the culture filtrate was extracted with same amount of chloroform. Extracts were allowed to dry and were resuspended in 500µL of methanol; 10 µL aliquots were used for LC-MS analysis. ### LC-MS All solvents and other chemicals used were of analytical grade. All LC-MS analyses were performed on a Shimadzu 2010 EV LC-MS (Phenomenex® Luna, 5μ, 2.0 × 100 mm, C18 column) using positive and negative mode electrospray ionization with a linear gradient of 5–95% MeCN-H₂O (0.1% formic acid) in 30 minutes followed by 95% MeCN for 15 minutes with a flow rate of 0.1 mL/min. The value of area under curve was observed by EIC (extracted ion chromatogram). ## Generation of the *fumR* (Afu8g00420) deletion strain Fusion Polymerase Chain Reaction (Fusion PCR) was used to create the deletion cassette of *fumR* as previously described. First, *Aspergillus parasiticus pyrG* was PCR amplified from *A. parasiticus* genomic DNA using primers AparapyrGF-Linker and AparapyrGR-Linker. The *Aspergillus parasiticus pyrG* fragment was ligated into pJET (Fermentas) yielding plasmid pSD38.1 Then, 1.5kb of 5’ UTR and 3’ UTR of *fumR* was amplified from *A. fumigatus* genomic DNA using primer pairs 420P1 & 420P2 and 420P3 & 420P4, respectively. *Aspergillus parasiticus pyrG* was then amplified from pSD38.1 using primers 420P5 and 420P6. Three fragments were fused using primers 420P7 and 420P8. Protoplast mediated fungal transformation was done as previously described using CEA17ku80 (gift from Robert Cramer) as the host strain*. Aspergillus parasiticus pyrG* was utilized as selectable marker, resulting in a complete gene replacement of *fumR* in CEA17ku80. Transformants were first screened using PCR (data not shown) and Southern blot analysis. Other DNA manipulations were done as previously described. ## Generation of the *laeA* deletion strain The *laeA* deletion DNA cassette was also generated by fusion PCR. Briefly, a 1.5 kb 5’ UTR fragment was first amplified from *A. fumigatus* genomic DNA with primers laeA_p1 and laeA_p2. A 1.3 kb 3’ UTR fragment was also amplified from genomic DNA with primers laeA_p3 and laeA_p4. *Aspergillus parasiticus pyrG* was amplified from pSD38.1 using primers laeA_p5 and lae_p6. The three fragments were fused using primers laeA_p7 and laeA_p8 as previously described. The *laeA* deletion cassette was transformed into CEA17ku80. Transformants were first screened using PCR (data not shown) and Southern blot analysis. # Results ## Hundreds of genes are differentially regulated by *veA* Of the 9,784 genes in the *A. fumigatus* genome, 453 were upregulated and 1,137 were downregulated in the ∆*veA* strain when compared with the wild-type strain. A similar pattern was observed in the OE*veA* versus wild type comparison, where 335 genes were upregulated and 908 genes were downregulated in the OE*veA* strain. In sharp contrast, comparison of the complementation strain and wild type showed that the two strains present very similar expression patterns. ## Differentially regulated genes are dramatically enriched for secondary metabolism-related processes Different GO process, component, and function categories are significantly enriched for both upregulated and downregulated genes in the two comparisons. Importantly, enrichment analysis using GOSlim categories showed that differentially regulated genes in the ∆*veA* versus wild type and OE*veA* versus wild type comparisons have significant functional overlap; specifically, 13 of the 19 GO categories that are enriched for either upregulated (3 categories) or downregulated (16 categories) genes in the ∆*veA* versus wild type comparison are also enriched and in the same direction in the OE*veA* versus wild type comparison. For example, both secondary metabolic process (GO:0019748) and toxin metabolic process (GO:0009404) GO function categories are enriched in downregulated genes from both comparisons. ## Differentially regulated genes are non-randomly distributed across the *A. fumigatus* genome We used a sliding window analysis to determine clusters of genes that were upregulated or downregulated for the ∆*veA* versus wild type and OE*veA* versus wild type comparisons. In total, 31 downregulated gene clusters and 6 upregulated gene clusters were identified. Ten downregulated clusters were found independently in both the ∆*veA* versus wild type and OE*veA* versus wild type comparisons, suggesting some similarity in phenotype between the α*veA* and OE*veA* strains. Twelve of the 31 downregulated clusters and 2 of the 6 upregulated clusters overlap with known or predicted secondary metabolic gene clusters. For example, the gene clusters encoding for the secondary metabolites fumagillin, fumitremorgin G, and fumigaclavine C are all differentially regulated in at least one of the two strain comparisons. All 11 genes in the fumigaclavine C biosynthetic gene cluster are downregulated in both the ∆*veA* versus wild type and OE*veA* versus wild type comparisons. In both the ∆*veA* versus wild type and OE*veA* versus wild type comparisons, 13 of the 15 genes in the fumagillin gene cluster are downregulated. All genes involved in fumitremorgin G are upregulated on the OE*veA* versus wild type comparison, but none of these genes are differentially regulated in the α*veA* versus wild type comparison. ## *veA* regulation profile of secondary metabolite gene clusters in *A. fumigatus* partially differs from the *laeA* profile Previous studies of the model fungus *Aspergillus nidulans* demonstrated that the *veA* gene product, VeA, interacts with other proteins in cell nuclei, among them LaeA. Additionally, Park et al. used tandem affinity techniques to show that the *A. fumigatus* LaeA co-purifies with *A. nidulans* VeA, suggesting that *A. fumigatus* VeA might also interact with *A. fumigatus* LaeA. LaeA is a putative methyl transferase that affects chromatin conformation. This VeA- interacting protein has also been described as affecting expression of secondary metabolite gene clusters in *A. fumigatus*. Our results indicate that although the regulation patterns of the two proteins overlap, they are not identical. Specifically, five out of the nine clusters that Perrin et al. report as being under full *laeA* regulation are found to be in windows of differentially regulated genes in both the ∆*veA* versus wild type and the OE*veA* versus wild type comparisons. Of the remaining four clusters, two are not found to be differentially regulated and two have different expression patterns in the two comparisons. The Afu6g12040–2080 *laeA*-regulated cluster was found to be part of a window of downregulated genes in the ∆*veA* versus wild type comparison but not in the OE*veA* versus wild type comparison. However, all five genes in this cluster are also downregulated in the OE*veA* versus wild-type comparison, and its lack of detection by the sliding window analysis is due to the gene cluster’s small size and the test’s stringency. Finally, of the four clusters Perrin et al. describe as being partially regulated by *laeA*, two are differentially regulated in both *veA* comparisons, one cluster is differentially expressed in the OE*veA* versus wild type but not in the ∆*veA* versus wild type comparison, and one cluster is not differentially expressed in either comparison. ## *veA* regulates the synthesis of fumagillin, fumitremorgin G, fumigaclavine C and glionitrin A *Aspergillus fumigatus* has the potential to produce 226 bioactive secondary metabolites, and the genes responsible for their synthesis are commonly associated in the form of gene clusters. Recently, Inglis et al. described 39 secondary metabolite gene clusters in the *A. fumigatus* genome, some of which are experimentally characterized and some of which are computationally predicted. The production of a number of secondary metabolites has been shown to be under the control of *veA* orthologs in different fungal species. In *A. fumigatus* we recently reported that the expression of gliotoxin genes and gliotoxin production are regulated by *veA*. Additionally, in our current study, our RNA-seq data indicates that the expression of many secondary metabolite gene clusters is also *veA*-dependent, strongly suggesting that *veA* affects the synthesis of other natural products in *A. fumigatus*. For this reason, we also used LC-MS to analyze the production of other compounds in wild type, ∆*veA*, complementation strain and OE*veA* cultures. Our data revealed that production of four additional secondary metabolites was also dependent on *veA* under the experimental conditions assayed, specifically fumagillin, fumitromorgin G, fumigaclavine C and glionitrin A. The production of these four compounds was notably decreased in the ∆*veA* strain. Relative amounts of fumagillin, fumitremorgin G, fumigaclavine C and glionitrinA in the ∆*veA* strain were 18%, 23%, 22% and 18% compared to the wild type levels, respectively. The production of these compounds was also reduced in the OE*veA* with only 0.3%, 0.1%, 0.7% and 0.8% as compared to wild-type levels, respectively. ## *veA* controls the fumagillin gene cluster and fumagillin production by regulating the expression of *fumR* (Afu8g00420) Due to the medical applications of fumagillin and fumagillin-related compounds for their potential use in the treatment of amebiasis, microsporidiosis, and for their anti-angiogenic properties, we further characterized the genetic regulation of *veA* on the fumagillin gene cluster. RNA-seq analysis revealed that *veA* controls the fumagillin gene cluster, including *fumR* (Afu8g00420), a gene encoding a putative C6 type transcription factor that our previous bioinformatics analysis identified within the fumagillin gene cluster, located in chromosome 8. We further validated these results by qRT-PCR analysis. The expression levels of *fumR* were 12% in ∆*veA* and 5% in OE*veA* with respect to the levels in the wild type strain. Expression of Afu8g00370, which encodes a polyketide synthase (PKS) in the fumagillin cluster, was also evaluated. We recently showed that expression of Afu8g00370 is necessary for the production of fumagillin, confirming the predicted role of Afu8g00370 as an indispensable PKS in fumagillin biosynthesis. Our data indicated that expression levels in ∆*veA* and OE*veA* strains were only 3% and 0.1%, respectively, compared to wild-type levels. ## *fumR* is necessary for the expression of other genes in the fumagillin gene cluster To gain insight into the function of the transcription factor encoded by *fumR*, this gene was deleted by gene replacement techniques using the *A. parasiticus pyrG* marker as described in the materials and methods section. The ∆*fumR* strain was confirmed by PCR (data not shown) and Southern blot analysis. Deletion of *fumR* did not change growth rate or development in *A. fumigatus* (data not shown). Our experiments showed that expression of the PKS gene, Afu8g00370, is regulated by *fumR*. The expression of Afu8g00370 in the Δ*fumR* strain and control strain was analyzed by qRT-PCR at 48 h and 72 h post inoculation. At both time points examined, there was only negligible expression of Afu8g00370 in the ∆*fumR* mutant as compared to the wild-type levels. Furthermore, our analysis revealed that the expression of other genes in the fumagillin cluster is also under the control of *fumR*. We examined the expression of Afu8g00520, the terpene cyclase involved in fumagillin biosynthesis and Afu8g00380, an essential acyltransferase for fumagillin production. In addition to the characterized genes in the fumagillin cluster, we analyzed the expression of other predicted genes in the cluster, including Afu8g00510, Afu8g00500, Afu8g00480, Afu8g470, Afu8g440 and Afu8g00430. Our results indicate that all genes were minimally expressed in the ∆*fumR* mutant compared to the wild type at both time points tested. Only Afu8g00470 showed slightly higher expression than the other genes in the cluster in ∆*fumR* (20% and 36% at 48 h and 72 h post inoculation, respectively) as compared to wild-type levels. ## *fumR* is required for fumagillin biosynthesis Extracts from wild type and ∆*fumR* 72 h and 120 h cultures were subjected to LC-MS analysis as described in the material and methods section. The chemical analysis showed a peak corresponding to fumagillin in the wild type, with a retention time of 29.1 minutes. However, this peak was completely absent in the ∆*fumR* strain, indicating that *A. fumigatus* is unable to produce fumagillin in the absence of *fumR*. The amount of fumagillin production in the wild type increased over time. ## *laeA* regulates expression of *fumR* and the PKS gene Afu8g00370 The identity of the cluster involved in fumagillin biosynthesis was not elucidated at the time of the microarray study previously carried out with a ∆*laeA* mutant. In a recent study we described the *A. fumigatus* fumagillin gene cluster, and in the present study we have demonstrated that *veA* regulates *fumR*, and that expression of *fumR* is necessary for the activation of other genes in the cluster. We also examined whether the expression of *fumR* was also dependent on *laeA*, analyzing the expression of this gene in a ∆*laeA* mutant and corresponding control strain. The ∆*laeA* strain was constructed as described in material and methods, and the strain was verified by Southern blot analysis. qRT-PCR results indicated a near complete loss of *fumR* expression in ∆*laeA* under conditions that allowed its expression in the control strain (3% and 8% at 48 h and 72 h as compared to wild-type levels). Expression of Afu8g00370 was also downregulated in Δ*laeA* (9% and 7% at 48 h and 72 h as compared to wild-type levels). # Discussion Invasive aspergillosis is a disease caused by the ubiquitous opportunistic invasive mold *Aspergillus fumigatus*. In immunocompromised patients, the intraepithelial immune system of the lung is unable to properly eliminate the inhaled conidia, which then germinate. Despite the prevalence of aspergillosis infection and the improvements in diagnosis, novel effective strategies to reduce *Aspergillus* infections are still needed. Deciphering the genetic mechanism controlling *A. fumigatus* cellular processes might provide the basis for the development of new strategies to prevent or treat aspergillosis. Our study clearly established that *veA* is a global regulator of the *A. fumigatus* genome, affecting the expression of hundreds of genes, many of which are involved in secondary metabolism related processes, in non-random genomic locations. Secondary metabolites, also known as natural products, are part of the fungal chemical arsenal important for habitat adaptation. Some of them are considered virulent factors playing an important role in the pathogen-host interaction. The most studied of these secondary metabolites is gliotoxin, known for its immunosuppressive properties, for inhibiting phagocytosis in macrophage, and for induction of apoptosis. We recently reported that expression of gliotoxin genes and concomitant gliotoxin production is dependent on *veA* in *A. fumigatus*. In this study we demonstrate that *veA* controls the expression of 14 secondary metabolite gene clusters whose metabolite products are known and an additional 23 putative secondary metabolite gene clusters. Although *veA* exercised negative regulation on some gene clusters, overall *veA* acted as a positive regulator. The *veA*-dependent gene clusters regulatory pattern was not identical to that described for *laeA*, which encodes a VeA-interacting protein in the *velvet* complex, presenting some differences in their regulatory output. Among the gene clusters regulated by *veA* are those involved in the synthesis of fumitremorgin G, fumigaclavine C and fumagillin. Production of most of these compounds correlates with the *veA* regulatory pattern observed for the respective gene clusters, with the exception of fumitremorgin, suggesting that in this case other *veA*-dependent factors might be needed for production of this compound at wild-type levels. Fumitremorgin G, fumigaclavine C, and fumagillin, together with production of glionitrin A, currently an orphan compound without an associated gene cluster, have been described to be relevant in the *A. fumigatus* infection process and in other pathologies. Fumitremorgins are associated with dysfunction of the nervous system causing tremors, seizures and abnormal behavior in animals. The alkaloid fumigaclavine also causes nervous system damage as well as alteration of the reproductive system. Fumagillin has been associated with invasive aspergillosis due to its effect in slowing ciliary beat frequency and inhibition of endothelial proliferation. Some of these compounds, however, are bioactive molecules with potential or current applications, particularly anti-tumoral compounds such as glionitrin A, and the well-known fumagillin and related compounds, with applications against amebiasis, microsporidiosis, and with known anti-angiogenic activity as inhibitors of the human type 2 methionine aminopeptidase (MetAP2). Our study shows that the expression of most of the genes in the fumagillin gene cluster was negatively affected by either deletion or over-expression of *veA*. This corresponded with a decrease in fumagillin production in these two strains compared to the wild type. Further insight into the mechanism regulating the secondary metabolite gene cluster in *A. fumigatus* may contribute to decreasing the detrimental effects of this fungus as well as increasing the production of valuable secondary metabolites, such as fumagillin. Our study indicated that the fumagillin gene cluster is regulated by a C6 transcription factor gene, now denominated *fumR*, located within the boundaries of this cluster. Other C6 type transcription factor genes have been found within other secondary metabolite genes clusters, such as the well-known *gliZ* in the *A. fumigatus* gliotoxin gene cluster and *aflR* in the *A. nidulans* sterigmatocystin cluster, which have been demonstrated to regulate such clusters. *fumR* positively regulated the expression of the recently characterized PKS gene, Afu8g00370, the terpene cyclase gene, Afu8g00520, and the acyltransferase Afu8g00380. Furthermore, *fumR* also regulates all the other predicted genes in this cluster. Deletion of *fumR* resulted in complete absence of fumagillin production. Our study showed that *fumR* is under the control of *veA*. Both deletion and over-expression of *veA* downregulated *fumR* transcription, suggesting that *veA* influences the activation of the fumagillin gene cluster through regulation of *fumR*. Over-expression of *veA* also had a marked negative effect on the expression of many *A. fumigatus* secondary metabolite gene clusters. We previously described a similar effect in the *veA* regulation of the gliotoxin gene cluster, showing a decrease of gliotoxin production in both deletion and over-expression strains with respect to wild-type levels. Our RNA sequencing data provide strong evidence supporting this pattern. The same pattern has also been observed in other *Aspergillus* species; for example, the production of penicillin has been demonstrated to be negatively affected by either deletion or over-expression of *veA* in *A. nidulans*. Since VeA is part of a protein complex or complexes, we hypothesized that a balanced stoichiometry between VeA and other possible partners might be necessary for proper function, including the activation of secondary metabolite gene clusters. It is possible that other VeA-interacting proteins might also modulate the expression of the fumagillin gene cluster. In this current study we also examined whether *laeA* affects the expression of genes in this cluster. Our results revealed that this is indeed the case; the absence of *laeA* greatly decreases *fumR* and Afu8g00370 expression. This indicates that both VeA and LaeA, components of the fungal *velvet* protein complex, are indispensable for normal expression of the fumagillin gene cluster and fumagillin production in the opportunistic pathogen *A. fumigatus*. # Conclusion In this study we have demonstrated that *veA* is a global genetic regulator in the opportunistic human pathogen *A. fumigatus*, controlling the expression of hundreds of genes. Among the genes governed by *veA* are numerous secondary metabolite gene clusters, some of them responsible for the synthesis of natural products considered to be virulent factors during *A. fumigatus* infection. Interestingly, we also showed that some of these *veA*-dependent gene clusters are associated with the production of important medical drugs, such as fumagillin, known for its anti-angiogenic properties among other relevant medical applications. All the genes in the fumagillin gene cluster are under the control of the endogenous regulator *fumR*, which is regulated by *veA* and *laeA*, both encoding interacting components in the *velvet* complex. The findings presented here provide further insight into the regulatory dynamics of the *A. fumigatus* genome, contributing to settinga basis for novel strategies to decrease the negative effects of *A. fumigatus* while increasing its potential to produce beneficial compounds. # Supporting Information We wish to thank Scott Grayburn and Xue-Huan Feng for technical support, as well as Travis Clark, Chelsea Baker, and the Vanderbilt Technologies for Advanced Genomics for Illumina library preparation and RNA sequencing. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AMC AR YT SD. Performed the experiments: SD ALL HCL. Analyzed the data: AMC SD AR ALL YT HCL. Contributed reagents/materials/analysis tools: AMC AR YT. Wrote the manuscript: AMC SD AR YT ALL HCL.

# Introduction Mesenchymal stem cells (MSCs) are multipotent stem cells. Accumulating evidence suggests that MSCs have profound therapeutic potential for a variety of diseases such as myocardial infarction, neural diseases and wound healing. Due to encouraging preclinical results, a large number of clinical trials for various diseases are underway. MSCs are distributed in a variety of tissues such as the bone marrow and adipose tissue, but represent a rare cell population in tissues. For example, MSCs account only approximately 0.001% to 0.01% of the nucleated cells in the bone marrow. Recently, it has been demonstrated that MSCs are also present in umbilical cord and placenta. This profoundly increases the availability of MSCs, but *ex vivo* expansion remains an indispensable procedure to obtain sufficient amounts of MSCs for cell therapies and tissue engineering. MSCs have long been considered as expandable stem cells. However, recent studies indicate that MSCs age rapidly and undergo considerable changes in cell morphology and production of paracrine factors during culture expansion. Oct4, Sox2 and Nanog are main transcription factors that govern embryonic stem cells self-renewal and pluripotency. They are also expressed in MSCs and are involved in their multipotency. Associated with morphological changes, rapid down- regulated expressions of these genes have been detected in MSCs during culture expansion. Previous studies suggest that limited culture expansion of MSCs does not cause alterations in their genetic DNA sequences. However, the epigenetic status of MSCs appears to be unstable in culture. Previous studies indicate that culture expansion of MSCs caused deacetylation of histone H3-K9 and 14 at promoters of pluripotent genes, which was associated with the appearance of aging signs. Meanwhile, no evident changes in DNA methylation were found in the promoter regions of the pluripotent genes. In this study, we attempted to use a histone deacetylase (HDAC) inhibitor trichostatin A (TSA) to suppress the reduction of histone acetylation in human MSCs (hMSCs) during culture expansion thus maintaining their primitive properties. We found that low concentrations of TSA significantly inhibited morphological changes of hMSCs that otherwise occurred during cell passaging. In addition, TSA-treated MSCs grew much faster. Moreover, TSA stabilized the expression of pluripotent genes and their histone H3 acetylation levels in lysine (K) 9 and K14 in the promoter regions. Our results suggest that TSA may be used as an effective agent to maintain the primitive feature of MSCs in culture expansion. # Materials and Methods ## Cell isolation and culture Human MSCs were isolated from human placenta as described previously. Briefly, term (38–40 weeks’ gestation) placentas from healthy donors were harvested with written informed consent and the procedure was approved by the Ethics Committee of Xili Hospital. The placental tissue was washed several times with cold phosphate-buffered saline (PBS) and then mechanically minced and enzymatically digested with 0.25% trypsin for 30 minutes at 37 °C in a water bath. The digest was subsequently pelleted by centrifugation and resuspended in a growth medium consisting of Dulbecco’s modified Eagle’s medium (DMEM, Gibco-Invitrogen) supplemented with 10% fetal bovine serum (FBS; Gibco-Invitrogen) and antibiotics. Cells were seeded on uncoated polystyrene dishes and incubated in the growth medium at 37 °C with 5% CO₂. Medium was replaced every 2 days. When reaching 80% confluence, the cells were lifted by incubating with 0.25% trypsin/EDTA and sub-cultured. ## TSA Treatment of hMSCs To obtain optimal concentrations of TSA for hMSCs, TSA at concentrations of 0, 6.25 nM, 12.5 nM, 25 nM, 50 nM, 100 nM, 200 nM and 300 nM (dissolved in dimethyl sulfoxide, DMSO) was added to the growth medium. Equal volumes of DMSO alone were used as control. Human MSCs were cultured in 24-well plates at a concentration of 1x10⁴ cells per well in the presence of TSA or DMSO alone and incubated for 3 days. Then the cells were collected and counted with a hemacytometer. 6.25 nM TSA was chosen as an optimal concentration to stabilize histone acetylaton of hMSCs in culture for subsequent experiments. ## Cell proliferation and cell cycle analysis 1x10⁵ cells per well of passage 1 hMSCs were seeded in six-well plastic tissue culture plates in triplet wells in the growth medium in the presence of 6.25 nM TSA or vehicle DMSO and incubated. Medium was changed every 2 days. When one of the wells reached 80% confluence, cells in all wells were harvested individually after trypsinization, counted with a hemacytometer and 1x10⁵ cells per well were re-seeded to a new six-well tissue culture plate. Cumulative cell numbers from passage 2 to passage 10 were calculated. Cells grown to full confluence (passages 6 and 10) and at the first day after passaging (passages 7 and 11 in ^\~30% confluence) were harvested for cell cycle analysis. Cells were fixed with 70% ethanol chilled at -20°C for 2 h, washed with PBS and re-suspended in a buffer containing 100 µg/mL propidium iodide and 10 µg/mL RNase A for at least 30 min in dark. Cells were then analyzed by flow cytometry (BD Biosciences). ## Western blotting Western bloting was performed to assess the expression levels of cell cycle proteins using a method as previously described. Briefly, cell lysates were prepared using a lysis buffer containing 1% Triton X-100 and proteinase inhibitors (Sigma-Aldrich). Equal amounts of total protein were separated on a 12% SDS-polyacrylamide gel and transferred to nitrocellulose membranes. Membranes were incubated overnight at 4°C with corresponding antibodies against cyclin D1, cyclin B1 and p21 (Santa Cruz), respectively. The bound antibodies were visualized using an ECL kit according to the manufacturer’s instructions. ## MSC differentiation assays Passage 1 hMSCs were grown in the presence of TSA (at 6.25 nM) or equal amount of vehicle DMSO to passage 6. Then the cells were incubated in adipogenic, osteogenic and chondrogenic induction media, respectively, for 3 weeks. The adipogenic induction medium contained 10^–6 M dexamethasone, 10 µg/mL insulin and 100 µg/mL 3-isobutyl-L-methylxantine (Sigma). Cells were finally stained with Oil Red-O to detect lipid. The osteogenic medium contained 10^–7 M dexamethasone, 50 µg/ml ascorbic acid and 10 mM β-glycerophosphate (Sigma). Cells were finally stained using Alzarin Red for calcium deposition. For chondrocyte differentiation, pellet hMSCs were cultured in DMEM (high glucose) containing 10^–7 M dexamethasone, 50 µg/ml ascorbate-2-phosphate, 100 µg/ml pyruvate (Sigma), 10 ng/ml TGF-β1 (R&D Systems) and 50 mg/ml ITS Premix (BD Biosciences, 6.25 µg/ml insulin, 6.25 µg/ml transferrin, 6.25 ng/ml selenious acid, 1.25 mg/ml bovine serum albumin and 5.35 mg/ml linoleic acid). Medium was changed every 2 days for 3 weeks. The pellet was fixed, embedded and sectioned for H&E and toluidine blue (Sigma) staining, respectively. ## RNA extraction and Real-Time PCR Total RNA was extracted from hMSCs with TRIzol (Invitrogen) following the manufacturer’s instructions. First-strand cDNA was prepared by reverse transcription using Superscript II reverse transcriptase (Invitrogen) and oligo(dT) primers and stored at minus 20°C until use. Real-Time PCR was performed using SYBR Premix Ex Taq II (TaKaRa) on an ABI 7300 QPCR System for the expression of Oct4, Sox2, CD133, TERT, REX1, Nanog, alkaline phasphatase (ALP) and osteopontin (OPN) using primers previously described. As an internal control, levels of glyceraldehyde-3- phosphate dehydrogenase (GAPDH) were quantified in parallel with target genes. Normalization and fold changes were calculated using the ΔΔCt method. ## Chromatin immunoprecipitation Chromatin immunoprecipitation (ChIP) assay was performed using an Acetyl-Histone H3 Immunoprecipitation Assay Kit (Millipore) following the manufacturer’s protocol. 1x10⁶ cells were used for each reaction. Histone acetylation was determined using specific antibodies against acetylated histone H3 at K9 and K14, respectively. After chromatin immunoprecipitation, DNA was extracted with a standard procedure (phenol/chloroform/isoamilic alcohol 25:24:1), and subsequently measured by quantitative fluorescent PCR analysis using SYBR Premix Ex Taq II (TaKaRa) on an ABI 7300 QPCR System. Primer targets were within 500 bp upstream of the gene transcription start site and primer sets were as follows: TERT forward 5’- GGCTCCCAGTGGATTCGC-3’, reverse 5’-GGAGGCGGAGCTGGAAGG- 3’; Sox2 forward 5’-AGTTGGACAGGGAGATGGC- 3’, reverse 5’-AACCTTCCTTGCTTCCACG-3’; Oct4 forward 5’-CTTCCACAGACACCATTGCC-3’, reverse 5’-AGTCCCACCCACTAGCCTTG- 3’. Data were analyzed using Percent Input Method (Invitrogen) following the instruction. # Results ## Low concentrations of TSA promotes proliferation and retains primitive features of hMSCs With successive passages of hMSCs in plastic tissue culture dishes as monolayer, the shape of hMSCs became larger and fatter. In accordance with the morphological changes, Real-Time PCR analysis showed marked decreases of expression levels of pluripotent genes Oct4, Sox2, CD133, TERT, REX1 and Nanog, and increased levels of osteogenic genes ALP and OPN in passage 10 hMSCs compared to hMSCs in passage 1. We proposed that TSA could inhibit the decline of histone acetylation in pluripotent genes and thus retained the primitive properties of hMSCs. To test this, hMSCs were incubated in the growth medium supplemented with TSA at 0, 6.25, 12.5, 25, 50, 100, 200 and 300 nM for 3 days. We found that low concentrations of TSA (6.25 nM and 12.5 nM) increased the cell number by 2 folds (*P*\<0.01), and did not cause detectable changes in cell morphology; however, excessive amounts of TSA (200 or 300 nM) decreased hMSC proliferation and lead to significant changes in cell morphology, such as larger and flatter cell body in culture (, *P*\<0.01). Similar results were obtained with hMSCs derived from three donors. ## TSA Stabilizes Histone Acetylation and the Expression of Pluripotent Genes in hMSCs We then analyzed the long term influences of TSA on hMSCs. Human MSCs were cultured in the presence of TSA (at 6.25 nM) or an equal amount of DMSO (the dissolvent of TSA) in the growth medium from passage 1 to passage 10. Progressive changes in cell morphology with successive cell passages as described earlier were observed in hMSCs treated with DMSO alone. However, the morphological changes did not occurred in hMSCs cultured in the presence of TSA. Meanwhile, there was a profound increase in the cumulative cell number of hMSCs in culture in the presence of TSA (\<0.01). To investigate whether transformation occurred in TSA-treated cells, we examined cell contact inhibition in cell growth in hMSCs after successive TSA treatment. Similar to DMSO-treated hMSCs, TSA-treated hMSCs stopped proliferating when they reached full confluence and no multi-layer foci were found in the culture. Cell cycle analysis of passage 10 hMSCs treated with DMSO or TSA showed similar percentages of cells arrested in G1 phase (76% versus 77%, n=3, *P*\>0.05) when they reached full confluence. However, when cells were passaged to new culture plates (passage 11) and incubated in the growth medium for 24 hours (in ^\~30% confluence), a higher percentage of TSA-treated hMSCs entered S phase compared to DMSO-treated cells (67% versus 61%, n=3, *P*\<0.05). We further examined the expression levels of cell cycle proteins in passage 6 and passage 10 hMSCs in full confluence by Western blot, and the results showed similar amounts of cyclin D1, cyclin B1 and p21 in DMSO- and TSA-treated cells. We also examined the multipotent differentiation potential of hMSCs into adipocytes, osteoblasts and chondrocytes, which has been considered as a typical feature of MSCs, and found that similar differentiations into these three cell lineages occurred in TSA-treated hMSCs, compared to DMSO-treated hMSCs. Next, we examined the expression of pluripotent genes in hMSCs in the above cultures. We found that TSA significantly inhibited the down-expression of Oct4, TERT, Sox2, Nanog, REX1 and CD133 genes from passage 1 to passage 6 hMSCs, which occurred in hMSCs treated with vehicle DMSO alone (\<0.01). Finally, we examined histone H3 acetylation in K 9 and K14 in the promoter regions of TERT, Sox2 and Oct4 genes in hMSCs. Compared to hMSCs in passage 1, hMSCs cultured in the presence of DMSO alone in passage 6 showed significantly decreased histone H3 acetylation levels of the pluripotent genes in K 9 and K14 (\<0.01). In the presence of TSA (at 6.25 nM), the acetylation levels of histone H3 in K 9 and K14 of these genes in passage 6 hMSCs showed no significant decreases (\>0.05). # Discussion To achieve maximum therapeutic effects of MSCs in tissue repair/regeneration, it is a pre-requirement to retain their primitive properties during *ex vivo* expansion. However, several previous studies have indicated that MSCs age quickly and gradually lose their multipotent differentiation potential in culture. Therefore, it is a crucial issue to develop optimal culture conditions to maintain the quality of MSCs for clinical uses. Previous studies indicate that the acetylation of K9 and K14 in histone H3 is crucial for gene transcription. It is required for the recruitment of transcription factor II D (TFIID), one of general transcription factors that bind the TATA box in the core promoter to initiate gene transcription. In our previous study, we found that decreases of acetylation levels in K9 and K14 of histone H3 were closely associated with the aging and spontaneous differentiation of hMSCs. In this study, we showed that TSA at low concentrations was potent in maintaining the histone acetylation states of K9 and K14 in histone H3 of pluripotent genes in hMSCs, thus stabilizing the expression of pluripotent genes and retaining the primitive features of the cells during culture expansion. Moreover, hMSCs cultured in the presence of low concentrations of TSA grew faster with consistent cell morphology. TSA, which was initially used as an antifungal antibiotic, has recently been found to be a potent and specific inhibitor of HDAC activity. It selectively inhibits the class I and II, but not class III, mammalian HDAC families of enzymes. Previous studies suggest that TSA modulates a wide variety of cellular activities such as cell differentiation and proliferation depending on cell types and their functional states. TSA at concentrations of 200^\~300 nM has been found to exhibit pronounced suppressive effect on breast cancer cells with immeasurable toxicities. In this study, however, TSA at concentrations of 200^\~300 nM caused evident morphological changes in addition to growth inhibition. It appears that MSCs are very sensitive to TSA. Previous studies suggest that histone acetylation is critically involved with *ex vivo* aging of MSCs. Culture expansion of MSCs caused deacetylation of histone H3-K9 and 14 in promoters of pluripotent genes. Meanwhile, no significant changes were found in levels of DNA methylation in promoters of the genes. In this study, preventive application of low doses of TSA markedly prevented the deacetylation of histone H3-K9 and 14 and the appearance of aging signs. These results suggest that low dose TSA may serve as an effective supplement to hMSC culture to stabilize their histone acetylation, and thus keep their primitive features. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YW PL. Performed the experiments: BH JL ZL LG SW. Analyzed the data: BH JL LG SW YW. Contributed reagents/materials/analysis tools: YW PL. Wrote the manuscript: YW.

# Introduction The emergence of new viruses is a constant challenge to the well-being of the human race and its food supply. New viruses or viral strains are produced from existing forms as a consequence of two processes: mutation and recombination or reassortment, which occur in both plant and animal hosts. The potential for recombination and reassortment is greatly enhanced in persistent and chronic infections, in which multiple genotypes of a single viral species, or multiple viral species, are introduced into a single host through repeated infections. Co-replicating viral genotypes create an environment conducive for RNA recombination to generate potentially new combinations of genes or protein domains that are exponential to the number of genotypes in the mixed infection. These recombinants may evolve subsequently into new and emerging viruses,. However, the extent to which such long-term infections result in genotypic variants remains largely unexplored. *Citrus tristeza virus* (CTV) represents an example of a virus that causes persistent infections in a long-lived, economically important hardwood perennial crop plant, so that with time a single host plant may become infected by multiple, phylogenetically distinct CTV genotypes. CTV is a member of the genus *Closterovirus* within the family *Closteroviridae*, and is the most important and destructive virus of citrus. CTV virions are flexuous rods, 2000 nm in length and 12 nm in diameter, consisting of one single-stranded, (+)-sense RNA genome encapsidated by two species of coat proteins (97% CP and 3% CPm). The 19.2 to 19.3 kb genome contains 12 open reading frames, is the largest of the plant RNA viruses, and is one of the largest of all RNA viruses –,. The 5′ half of the genome (∼nt. 1–11,000) encodes proteins (RNA-dependent RNA polymerase, helicase, methyltransferase, and proteases) that are required for viral replication and are thought to be translated directly from the genomic RNA. The 3′ half encodes proteins which are thought to interact with host plants – and are expressed from ten 3′ co-terminal subgenomic RNAs. The global CTV population is very diverse, with numerous, disparate strains, many inducing different types and degrees of disease symptoms on different citrus species and varieties. Often in natural infections in the field, CTV exists as a complex comprising multiple strains or genotypes, due to the longevity of individual citrus trees and the extensive use of vegetative propagation of budwood. Continual vertical transmission coupled with repeated horizontal transmission mediated by aphids throughout the history of citrus cultivation has led to the complexity of the CTV population increasing over hundreds of years, resulting in the co-existence of multiple CTV genotypes in a single host. The presence within a host of multiple replicating CTV genotypes and the relatively long periods of co-replication create opportunities for recombination between the genotypes, leading to extensive viral diversity. In this report, we characterized a persistent infection by multiple CTV genotypes by genome-wide microarray resequencing analysis and deep sequencing analysis of selected genomic regions. Our results demonstrate an extraordinary amount of viral variability generated by promiscuous recombination between multiple genotypes, and provide evidence for subsequent divergence of the recombinants within a single host plant. # Results ## Resequencing analysis of FS2-2 reveals presence of multiple CTV genotypes To study the CTV genetic complexity of CTV in detail at the sequence level, we designed and validated an Affymetrix resequencing microarray that queries entire genomes of multiple, phylogenetically distinct CTV genotypes. Sequences tiled on the microarray include full-length sequences of four CTV type strains, T3 (Hilf, unpublished), T30, T36, and VT, three-quarters of the strain T68-1 genome (Hilf, unpublished), as well as unique genomic sequences identified from five other CTV isolates. Together, over 117 kb of CTV sequences representing a genetic diversity equivalent to ten full-length CTV genomes were tiled on the resequencing microarray. Using known CTV isolates as the source for target cDNAs, the CTV resequencing microarray yielded call rates of ∼99.7–99.8% and call accuracies of ∼99.9–100%, performing comparably to, or better than, several reported resequencing microarrays. We subsequently employed the microarray for analysis of the genetic complexity in a natural CTV isolate, FS2-2. The isolate, collected originally from a citrus grove in Florida, was associated with an unusual stem-pitting symptom in the Hamlin sweet orange. The isolate was initially suspected to possess multiple CTV genotypes on the basis of PCR-amplification with genotype-specific primers (data not shown). The multiple CTV genomes present in FS2-2 were amplified in their entirety by long-range RT-PCR with a high fidelity polymerase from total RNA prepared from infected tissues, using four sets of universal CTV primers capable of amplifying all known CTV genomes. DNA fragments of expected sizes were amplified from the total RNA extracted from FS2-2-infected tissues. Furthermore, each of the three unique 5′ end PCR primers, designed specifically for CTV genomes with very different 5′ terminal sequences, successfully amplified DNA fragments of the predicted sizes in RT-PCR, indicating the presence of multiple genotypes in the FS2-2 CTV complex. The three 5′ DNA fragments and the three DNA fragments amplified from the remaining portion of the CTV genomes were pooled proportionally, and then were processed for microarray hybridization. The existence of multiple CTV genotypes in the FS2-2 complex was clearly evident in the scanned images of the hybridized resequencing microarray. Hybridization with target DNA prepared from a single genotype generated a single block of intensive hybridization corresponding to the location on the microarray of the tiling path for that genome (e.g. hybridizations with the T30 and the T36 target DNA). In comparison, hybridization with the FS2-2 target DNA yielded strong signals in multiple microarray blocks corresponding to the tiled genomes of VT, T30, and T36, indicating that the FS2-2 isolate contained at least three genotypes. Hybridization intensities of each probe on the resequencing microarray were subsequently processed using the Affymetrix GeneChip Operating Software version 1.4. Base calls were made using the ABACUS algorithm with a haploid model as implemented in the Affymetrix GeneChip Sequence Analysis Software (GSEQ) version 4.0. A total of 252 sequence fragments corresponding to full length CTV genomes and genomic fragments tiled on the microarray were produced by the resequencing analysis. These fragments and the quality scores associated with each base call were used in contig assembly by the Phrap program as implemented in the CodonCode Aligner program. Three consensus contigs of complete CTV genomes (fs2_2\_vt, fs2_2\_t30, and fs2_2\_t36), corresponding to the three visually identified CTV genotypes, were assembled. In addition, two partial consensus contigs were also assembled (fs2_2\_t3 and fs2_2\_t68). Both were placed as intermediates between fs2_2\_vt and fs2_2\_t30 by Bayesian phylogenetic analysis, and therefore might represent minor components of the CTV complex or variants generated by recombination between the major genotypes. ## Deep sequencing analysis of the 5′ 1 kb of isolate FS2-2 confirms existence of multiple genotypes To verify the presence of multiple CTV genotypes in the FS2-2 complex, a fragment of approximately 1 kb in size, corresponding to the 5′ termini of the CTV genomes, was amplified by RT-PCR and cloned. The 5′ terminal region was targeted because of its unusually high sequence variability in the CTV genomes. Previous comparative sequence analysis revealed an interesting and unusual distribution of sequence variation across the CTV genome. The 3′ halves of the CTV genomes were highly conserved, with sequence identity being at or above 90%, as would be expected of strains of the same virus. However, sequences of the 5′ halves were much more divergent, with pairwise comparisons yielding sequence identities of 70% or less for some isolates. Sequence identities as low as 48% was observed at the 5′ untranslated regions (UTRs). The 5′ half of the CTV genome encodes a polyprotein that contains motifs characteristic of RNA- dependent RNA polymerases, helicases, methyltransferases, and proteases, and is required for viral replication. Generally speaking, viral proteins required for replication are highly conserved within viral species and genera, and occasionally this conservation may extend across an entire family of viruses. The 5′ genomic fragment of CTV targeted for RT-PCR amplification includes the highly divergent 5′ UTRs of approximately 100 nucleotides, and approximately 830 nucleotides from the 5′ open reading frame that encodes for the polyprotein required for CTV replication. RT-PCR amplification of the fragment was accomplished using a combination of three 5′ end primers and a universal, 3′ primer that were designed for the amplification of all known CTV sequences at the 5′ end. A total of 70 clones were randomly selected and sequenced using an ABI 3730XL DNA Analyzer. Among the clones sequenced in this manner, 15 contained sequences nearly identical (98.7∼99.8% identity) to the reported T36 sequence. Sequences of 20 clones were nearly identical (99.3∼99.6% identity) to the published T30 sequence, while another 34 clones produced sequences highly similar (97.6∼97.9% identity) to that of VT. These results unequivocally identified three CTV genotypes in the FS2-2 complex that corresponded to the three major genotypes revealed by the resequencing microarray. Within each of the T30-like and the VT-like genotypes, the clones were nearly homogeneous, with a sequence identity of 99.3∼100%. However, sequences within the T36-like genotype were found to be slightly more diverse, with a sequence identity ranging from 98.6% to 100%, suggesting that this genotype might have evolved faster and started to diversify within the population. Indeed, Bayesian phylogenetic analysis showed that three sequences (fs22_78, fs22_79, and fs22_83) of the T36-like genotype emerged as a distinguishable subclade within the T36-like clade. Different degrees of divergence were also evident in the number of identical sequences obtained from clones of each of the three major genotypes. Nearly one half (15) of the clones from the VT-like genotype were completely identical, indicating that genomes of the VT-like genotype were quite uniform with little divergence. In contrast, only three of the 15 sequenced clones of the T36-like genotype contained the same sequence, suggesting the genotype has diverged more significantly. Sequence divergence of the T30-like genotype appeared to be between those of the T36- and VT-like genotypes. Among the 20 T30-like clones sequenced, one group of four clones and another group of five clones produced identical sequences with a single nucleotide difference between the groups. More interestingly, one sequenced clone, fs22_28, occupied an odd place between the T30-like clade and the VT-like clade in the phylogenetic tree. This placement was similar to that of the two minor genotypes (fs2_2\_t3 and fs2_2\_t68) identified by the resequencing analysis. Therefore, fs22_28 likely represented one of the minor genotypes identified in the resequencing analysis. Further analysis revealed fs22-28 to be recombinant: its 5′ region comprised 282 nucleotides identical to the sequences of 7 independently sequenced clones from the T30-like genotype, and its 3′ region consisted of 574 nucleotides identical to sequences in 22 sequenced clones of the VT-like genotype. The recombination crossover site in fs22_28 comprised 41 nucleotides identical to sequences within both parental genotypes. Although the recombinant clone represents one of the minor genotypes detected by the resequencing microarray, direct cloning and sequencing of the 5′ 1 kb of the CTV genomes in the FS2-2 complex failed to identify all the minor genotypes. ## CTV p33 ORF exhibits promiscuous genotypic recombination To search for additional CTV genotypes identified by resequencing analysis in the FS2-2 complex, an additional 1 kb region containing the entire p33 ORF was selected for further RT-PCR cloning and deep sequencing analysis. The p33 ORF of CTV has no homologue in other closteroviruses. The protein it encodes has been shown to be required for CTV movement in specific hosts (Dawson, unpublished data). The p33 ORF, located about 11 kb from the 5′ end of the genome, is more conserved among CTV isolates than is the 5′ end 1 kb region, yet it retains significant sequence divergence for differentiation of strains and isolates. The higher sequence conservation made it possible to design a set of universal RT- PCR primers capable of amplifying all known CTV genotypes non-selectively. A total of 84 RT-PCR clones derived from the p33 ORF of isolate FS2-2 were chosen randomly for deep sequencing analysis. Sequencing data from the 983 nucleotides of these clones provided a detailed portrait of the CTV complex, and of on-going, promiscuous recombination among various major CTV genotypes. Among the sequenced clones, the three major genotypes again were well represented with 51 VT-like clones, twelve T30-like clones, and six T36-like clones. Several groups of clones shared identical sequences (groups of nine, seven, five, five, two, two, and two clones for the VT-like genotype; and groups of two and two clones for the T30-like genotype) while all T36-like clones had unique sequences. These confirmed that the VT-like genotypes were more homogeneous and the T36-like genotypes were more divergent in FS2-2, an observation arising earlier from analysis of the 5′ terminal CTV fragments. The large number of identical sequences obtained from these clones, as well as from clones derived from the 5′ 1 kb region, also demonstrated that the RT-PCR employed in this study maintained a high fidelity during viral sequence amplification and that errors introduced during RT-PCR were negligible. A surprisingly large proportion of the clones (15 clones, 17.9%) were recombinants. Eight were recombinants between fs2_2\_vt and fs2_2\_t36, five were recombinants between fs2_2\_t30 and fs2_2\_vt, and two were recombinants between fs2_2\_t30 and fs2_2\_t36. Further, four recombinants contained two crossovers, resulting from either a double-crossover or two independent recombination events. The parental sequences for each recombinant were readily identified among the three major genotypes. The crossover sites appeared throughout the p33 ORF without an apparent recombination hotspot. One common feature of these recombinants was that recombination crossover sites contained a stretch of sequence that was identical in both parental molecules. The identical sequences at the crossover sites varied in length, ranging from as short as 3 nucleotides to as long as 25 nucleotides. This feature of shared parental sequences suggested that the recombinants were most likely formed by a template- switching event during viral RNA replication. Some of the crossover sites were characteristic of classical recombination hotspots with a stretch of AU-rich sequence that promotes dissociation and switching of RNA polymerase from one template to another. However, others lacked such sequences. These recombinants represented a remarkable amount of genetic variability. A large number of recombinants were placed as intermediates between the VT clade and the T30 clade by Bayesian analysis. These recombinants could represent the minor genotypes detected by the resequencing microarray. Since the resequencing microarray is not sensitive enough to differentiate individual recombinants, the minor genotypes identified by the resequencing analysis could very well represent composites of these recombinant genomes. Although the crossover sites were scattered throughout the p33 coding region, all but one of the recombinants maintained the correct reading frame, suggesting that the recombinant p33 protein was also expressed. Therefore, this genetic diversity was expressed at both nucleotide and protein levels. To eliminate the possibility that the recombinant clones were artificially generated by PCR amplification from a pool of DNA template representing three different genotypes, a control PCR amplification was carried out. Equimolar amounts of two cloned DNA molecules, representing the p33 ORF cloned from the VT-like genotype and from the T36-like genotype respectively, were combined to mimic a mixture of different genotypes, and this was used as the template for the control PCR amplification with the same set of PCR primers and identical PCR conditions. The DNA fragments amplified from this artificial mixture were cloned, and 75 random clones were sequenced. Sequence analysis showed that 26 clones were derived from the T36-like genotype and 49 were derived from the VT genotype. More importantly, none of the 75 sequenced clones was recombinant between the two genotypes present in the template DNA. This result strongly suggests that the extraordinarily large number of recombinant viral molecules detected in the FS2-2 CTV complex represent natural recombinants. Although generation of artificial recombinant during PCR has been reported under specific conditions, our results clearly showed that the template DNA molecules and the PCR conditions used in these experiments did not favor the formation of artificial recombinants. Further analysis of the recombinants shed light on the emergence of new CTV genotypes through recombination and subsequent divergent evolution. This was particularly evident in two pairs of recombinants. The first pair (fs22-05 and fs22-36) contained an identical recombination crossover site but four divergent nucleotides located both 5′ and 3′ of the crossover site. A plausible explanation is that they originated from a single recombination event and that progenies of the recombinant diverged subsequently. The second pair of recombinants (fs22-15 and fs22-100) contained an identical crossover site at the 5′ end, but fs22-15 possessed an additional crossover site toward the 3′ end. One probable explanation is that both were the progeny of a single recombination event at the 5′ end with fs22-15 subsequently undergoing a further round of recombination. In the identical region of the two recombinants, three divergent nucleotides were found spanning the 5′ crossover site, indicating that the regions had diverged after the first recombination event. These data suggest an event line of recombination, divergence, and further recombination, leading to a distinct CTV genotype. This scenario also explains the chimeric nature of the sequenced SY568 CTV genome as a consequence of recombination between two CTV genotypes. # Discussion The complexity of a natural CTV isolate consisting of multiple genotypes was revealed through a combination of genome-wide resequencing analysis using a CTV Affymetrix GeneChip resequencing microarray and deep sequencing of selected genomic regions in this study. Furthermore, deep sequencing analysis illustrated an unparalleled level of promiscuous recombination between multiple, co- replicating genotypes in a persistent infection. Viruses have been a favorite target for resequencing analysis using oligonucleotide microarrays because of their small genome sizes,. However, most of these applications either targeted viruses with relatively little sequence variability, such as severe acute respiratory syndrome virus, or only selected regions of viruses having larger sequence variability. With increased tiling capacity, the Affymetrix resequencing microarrays can now be used to simultaneously detect divergent viruses, as we have demonstrated in our work, where three major genotypes and additional minor genotypes of CTV were successfully identified from isolate FS2-2 by resequencing analysis. The three major CTV genotypes investigated by the resequencing analysis were quite divergent with genome-wide nucleotide differences of 11% between the VT- and the T30-like genotypes, 20% between the VT- and the T36-like genotypes, and 19% between the T30- and the T36-like genotypes. Successful resequencing analysis of these divergent CTV genomes with a single resequencing microarray chip is largely attributed to the tiling strategy of selecting representative full- length CTV genomes guided by phylogenetic analysis and of selecting unique sequences identified by pair-wise comparisons from other CTV genomes. This strategy allows the CTV resequencing microarray with a tiling capacity of 117 kb to encompass a sequence diversity equivalent to ten full-length CTV genomes. With its ability to query concurrently entire genomes of multiple CTV genotypes as demonstrated in this work, the CTV resequencing microarray will find many uses in future CTV studies. CTV has been known to exist as a complex in nature, – likely because of the longevity of the host plant, continuous vertical viral transmission by grafting, and repeated horizontal viral transmission by aphids. In previous studies, characterization of CTV complexes was limited to analysis of genotype-specific PCR markers, or to sequencing analysis of selected regions representing a small percentage of the CTV genome. In this study, a genome-wide approach was used to characterize multiple genotypes in a CTV complex for the first time. The availability of complete genomic sequences of co-replicating and interacting CTV genotypes should facilitate detailed and sophisticated analyses of genotypic interactions in any given CTV complex in future studies. The presence of multiple, co-replicating CTV genotypes was expected to promote recombination. However, the extraordinarily large proportion of recombinant viral molecules as a consequence of promiscuous recombination in the FS2-2 isolate was surprising. We found that 17.6% of cloned and sequenced molecules from the p33 ORF were recombinants derived from the three predominant genotypes in the FS2-2 complex. In contrast, a previous study of another CTV complex, SY568, identified less than 4% of the cloned molecules as being recombinants between two genotypes. It is not clear whether the large proportion of recombinants in FS2-2 is a consequence of an elevated recombination rate in the presence of more than two genotypes or a result of recombinants accumulating over time. Likely though, as a result of current horticultural practices, the sweet orange tree from which FS2-2 was derived was planted free of CTV and the CTV population is a result of ingress of the T30, T36 and VT genotypes by aphid transmission, followed by recombination to produce the recombinants found in this study. Subsequent transmission of the FS2-2 complex by aphids or by vegetative propagation to a similar or a new host may perpetuate the same complexity or alter it by generating new and additional recombinant genotypes. There was a substantial difference in the proportions of the recombinant sequences obtained upon cloning the 5′ 1 kb (one out of 70 clones) and the p33 ORF (15 out of 84 clones), suggesting that the sequence most proximal to the 5′ end of the CTV genome was more recalcitrant to recombination or that the terminal 1000 bases are subject to more stringent purifying selection, due perhaps to decreased fitness of recombinants, so consequently fewer recombinants are recovered. This difference perhaps can explain the unusual distribution of sequence variability in the CTV genome, with the 5′ half being more variable and the 3′ half being less variable. Recombination in RNA viruses has been extensively documented, as a powerful driving force for generating new sequences which appear as emerging viruses, since recombination can rapidly generate new genotypes by swapping genes or protein domains to reconstitute proteins with novel host-colonizing and pathogenicity traits. The unprecedented scope of recombination between multiple CTV genotypes within a single, persistently infected host as revealed in this study further underscores the importance of recombination in RNA virus evolution, and may explain the extraordinary diversity observed in CTV today. The large number of genetic variants generated through recombination can potentially evolve on their own and become an emerging viral isolate when, and if transmitted to a new host or into a new environment. The influenza virus responsible for the 1918 Spanish flu pandemic was hypothesized to have acquired, through recombination, a portion of the hemagglutinin gene, a key virulence gene, from a swine-lineage influenza. In this regard, it is interesting to note that all but one of the recombinants in this study maintained the correct open reading frames, and consequently generating a functional recombinant protein. Further studies are needed to determine the viability of these CTV recombinants in isolation from the parental sequences and to evaluate the likelihood that any of the recombinants will emerge as a new CTV strain. CTV shares similarities in its genome organization and gene expression strategies with the largest animal RNA viruses, corornaviruses (which includes the viral agent of SARS), so it is conceivable that similar processes may also operate in global populations of coronaviruses to generate genetic diversity. # Materials and Methods (See Supporting Information for experimental details) ## CTV isolates and genomic sequences CTV isolate FS2-2 was collected from a citrus grove in Florida in 2004 and maintained on Madam Vinous sweet orange in an insect-proof greenhouse. Full- length genomic sequences of CTV isolates NUagA, Qaha, SY568, T30, T36, T385, and VT were retrieved from GenBank. In addition, full-length unpublished sequences of CTV isolates T3 (M.E. Hilf, unpublished) and H33 (T.E. Mirkov, personal communication) and a partial sequence (13,585 nt) of the CTV T68-1 isolate (M.E. Hilf, unpublished) were included. ## Amplification of the CTV genome by RT-PCR Full-length genomic equivalents of CTV from each sample were amplified from each sample as four DNA fragments ranging from 4.5 to 5.5 kb by RT-PCR, using four sets of RT-PCR primers. Total RNA was extracted from CTV-infected tissue using the Trizol reagent (InVitrogen, Carlsbad, CA). Reverse transcription was carried out using the ImProm II reverse transcriptase (Promega, Madison, WI). CTV genomic fragments were then amplified by 35 cycles of long range PCR using the Stratagene EXL DNA polymerase (Stratagene, La Jolla, CA). ## RT-PCR amplification, cloning, and sequencing of CTV genome fragments The 5′ 1 kb fragments of the CTV genome were amplified using the three different 5′ PCR primers and a 3′ conserved universal primer. The 1 kb fragments containing the p33 ORF were amplified using a set of universal RT-PCR primers. PCR products were purified using the Qiagen MinElute PCR Purification Kit (Qiagen, Valencia, CA) and were employed directly for TA-cloning using a pUC18-based vector containing twin *Xcm* I restriction sites²⁵. Plasmid DNA from randomly selected clones was purified and sequenced in both directions using an ABI 3730XL DNA Analyzer. ## Microarray hybridization and base-calling Equimolar amounts of each PCR fragment were pooled, and labeled with biotin-dNTP by terminal deoxynucleotidyl transferase. Hybridization of the labeled target DNA to the microarrays, washing, and subsequent staining in GeneChip Fluidics Station 450 were performed in strict accordance with the instructions provided by Affymetrix. The stained microarray was then scanned at a resolution of 1.563 µm/pixel using a GeneChip Scanner 3000 (Affymetrix, Santa Clara, CA). The final probe intensity data were analyzed with the Affymetrix GeneChip Sequence Analysis Software (GSEQ) to extract sequencing information. Base calls were made using the ABACUS (adaptive background genotype calling scheme) algorithm. ## Contig assembly of sequence fragments generated by resequencing analysis Sequence fragments and the associated quality scores generated by GDAS were converted into fasta-format files and used to assemble full and partial CTV genomic contigs using the Phrap program implemented in the CodonCode Aligner (CodonCode, Dedham, MA). ## Sequence analysis Sequences of CTV genomes or genomic fragments were aligned using the default parameters of the ClustalX program. Bayesian inference of phylogenetic relationships was carried out using the general time reversal model with gamma- shaped rate variation and a proportion of invariable sites (GTR+I+G) as implemented in MrBayes 3.12. Phylogenetic trees were then visualized using the TreeView program. Recombinant molecules and their cross-over junctions were determined by RDP2, a recombination detection program that deploys 10 published methods to detect recombinant sequences and recombination breakpoints. # Supporting Information We thank T. E. Mirkov for providing unpublished CTV H33 genomic sequence and L. S. Pierson for critical reading of the manuscript. [^1]: Conceived and designed the experiments: DG ZX ZW RB WD. Performed the experiments: ZX ZW RB SG MH. Analyzed the data: DG ZX ZW WD. Contributed reagents/materials/analysis tools: ZX WD SG MH. Wrote the paper: DG ZX ZW. [^2]: The authors have declared that no competing interests exist.

# Introduction Preterm delivery (i.e., delivery at \<37 completed weeks' gestation) is one of the main causes of infant mortality in the U.S., and it is associated with substantial morbidity. Its prevalence shows substantial variability geographically, for example ranging from 9–17% across U.S. states. It also varies widely by race-ethnicity. Black women are twice as likely as white women to have preterm deliveries and three times more likely to have very preterm deliveries (\<32 weeks), which are the most vulnerable to mortality and long- term morbidities. Many studies have tried to determine what factors explain individual-level risk of preterm delivery; fewer have focused on what explains prevalence (i.e., population-level risk)., As Geoffrey Rose eloquently articulated several decades ago, the determinants of individual risk may not be the same as the determinants of prevalence, but both are important to understand from a prevention as well as an etiologic standpoint. Cullen et al. recently reported that most of the county-level variation in premature adult mortality (i.e., death before age 70) in the U.S. – as well as black-white disparities – was explained by 22 sociodemographic, socioeconomic, environmental, and health-related variables that were measured at the county level. The study demonstrated an innovative approach to understanding geographic variability and health disparities, in that it incorporated multiple variables into a single model and focused on their combined ability to explain disparities. Our objective here was to apply that approach to preterm delivery, given the commonalities between preterm delivery and premature adult mortality – namely that both have substantial geographic variability, black-white disparity, and are likely affected by a complex array of factors related to the social context, environment and health-related behaviors. Specifically, we examined the ability of a set of social, environmental and health-related factors to explain geographic variability in the risk of preterm delivery among live births to black and white women in the US and whether these factors explain black-white disparities in preterm delivery. # Materials and Methods We examined U.S. singleton births from 1998–2002, using birth certificate data from the National Center for Health Statistics final natality data. Following the methods of Cullen et al., we examined births in counties whose total population included at least 100,000 people and in US Census-defined Public Use Metadata Areas (PUMAs) for counties with \<100,000 people. PUMAs represent contiguous groupings of counties, such that the resulting population includes at least 100,000 people. We used PUMAs rather than single counties to increase the stability of estimates within sparsely populated counties. For convenience, we refer to PUMAs as ‘counties’ here. There were 957 counties available for analysis; 382 were single counties, 575 were PUMAs. All analyses were conducted separately for blacks and whites. This approach follows that of Cullen et al., it avoids the assumption that associations with risk factors are the same for blacks and whites, , and it enabled us to examine models that focused on variables specific to either the black or white population, rather than some combination or average. The outcome for analyses was prevalence or risk of preterm delivery, from 20–31 or 32–36 gestational weeks. For initial descriptive analyses, we examined prevalence of preterm delivery, defined as the number of preterm deliveries divided by the total number of deliveries with non-missing gestational age (0.7% of births in the study counties were excluded due to missing gestational age). For regression analyses, we refined the outcome to be 1) the number of deliveries at 20–31 weeks divided by that number of deliveries plus the number at 37–41 weeks or 2) the number of deliveries from 32–36 weeks divided by that number plus those from 37–41 weeks. We used this refined definition because we consider early and moderately preterm delivery as two distinct but potentially related adverse outcomes. As such, excluding one preterm group from the denominator when considering the other preterm group avoids dilution of the observed associations due to including a related outcome in the denominator. We thus refer to the refined outcome measure as ‘risk’ of preterm delivery, since it does not follow the traditional definition of prevalence. We restricted analyses to counties with at least 20 preterm deliveries during the study period (at either 20–31 weeks or 32–36 weeks), in an effort to enhance stability of the estimates. All analyses were conducted separately for black and white women. We focused analyses on the 468 counties that had at least 20 deliveries from 20–31 weeks of gestation among black and among white women and that had complete data on all covariates, to enable comparability across models. These counties encompass 61.1% of U.S. births to white mothers (n = 6,986,984) and 90.9% of U.S. births to black mothers (n = 2,607,150) during the study period. We conducted multivariable population-weighted ordinary least squares linear regression analysis for each preterm outcome, stratified on race-ethnicity. To adjust for differences in the size of the birth population in each county and their potential influence on model parameter estimates and variances, the weight applied to each stratified model was the total number of white or black live births in each county, respectively. Independent variables reflected a variety of exposures that may be related to reproductive health; models for black women included variables specific to black women, and white models included variables specific to white women, whenever possible. First, we included variables from Cullen et al. 's analysis of premature adult mortality, which were primarily derived from the 2000 US Census and represent county-level sociodemographics, socioeconomic level, and environmental exposures. Variables from the Census describe black or white adults (females when appropriate) aged 30–59, age- adjusted by the direct method. In addition, we included several health-related variables derived from birth certificates, to reflect these characteristics among all black or white women giving birth in a given county during the study period. Preliminary analyses included crime as an indicator of stress. Specifically, we examined violent crime per capita (i.e., murder and non-negligent manslaughter, forcible rape, robbery, and aggravated assault) as derived from the FBI Uniform Crime Reports (<http://www.fbi.gov/about-us/cjis/ucr/ucr>). Crime was not significantly (p\<0.05) associated with preterm delivery in any of the preliminary regression models. Given this lack of association and that a substantial number of counties were missing crime data (about 9%), we excluded this variable from further analyses. California was the only state that did not include maternal smoking on the birth certificate during the study period. Given the importance of smoking to preliminary results, we chose to exclude California rather than exclude smoking from our primary analyses. However, given that California contributes over 10% of all US births, we conducted sensitivity analyses that included the 17 eligible California counties but excluded smoking. We also conducted sensitivity analyses that included the maximum number of counties possible for each model, rather than just the 468 common counties. Given the relatively large number of independent variables, their inter- relatedness, and somewhat modest number of PUMAs, we were concerned about the stability and precision of the regression coefficients. We thus used forward stepwise selection to reduce the full models; specifically, with entry and stay criteria at p\<0.15. Second, we conducted a principal components analysis of all of the variables and then ran regression models that included the top factors as independent variables, separately for blacks and whites. We did the following to assess the degree to which the distributions of the independent variables explain racial-ethnic differences in preterm delivery. We calculated the predicted county-level risk of preterm delivery among black women, using regression coefficients from the step-wise models for black women and inserting the corresponding values of the independent variables among black women. We then recalculated predicted risk after inserting the (‘counterfactual’) corresponding *white* values for each of the variables in the models for black women. # Results The numbers of counties with at least 20 deliveries at 20–31 or 32–36 weeks and complete covariate data were 468 and 619, respectively, for black women, and 907 and 913 for white women. illustrates the striking difference in the distributions of preterm delivery for black and white women within the 468 counties that had at least 20 early preterm deliveries to black *and* white women. Among whites, the mean prevalence was 1.2% (range 0.7–2.4%) for delivery at 20–31 weeks and 8.2% (range 5.3–13.2%) for delivery at 32–36 weeks. Among blacks, the respective prevalences were 3.6% (range 1.9–7.1%) and 12.6% (range 7.4–18.7%). Among births to black women, the stepwise reduced models explained 46% of the variability in county-level risk of delivery at 20–31 weeks and 55% for delivery at 32–36 weeks (based on model R-squared values). The respective percentages for whites were 67% and 71%. Percentages were similar for the full models that included all variables and for models that included all counties with at least 20 preterm deliveries (depending on the particular analysis) rather than just the 468 counties. We also examined models that only contained the census variables, the environmental variables, or the birth certificate variables. Models that only included census or birth certificate variables explained similar percentages of variability in preterm delivery (32–64% for census variables, 33–62% for birth certificate variables). Models that only included the environmental exposure variables explained considerably less of the variability (2–38%), but these models were based on the fewest variables. Models that only included the factors from the principal components analysis explained somewhat less variability than models that included the actual independent variables. For all the models that were tested, the amount of variability explained consistently went from lowest to highest in this order: deliveries to black women at 20–31 weeks, blacks at 32–36 weeks, whites at 20–31 weeks, and whites at 32–36 weeks. A variety of variables were retained in the final regression models. Three variables were not retained in any of the four gestation/race-ethnicity models: high occupation, availability of fast food, and air pollution. Three variables were retained in all four models: warm climate, smoking and late/no prenatal care. Most variable associations were in the expected directions (e.g., among whites, more women with low education was associated with higher preterm risk), although some were not (e.g., higher income was associated with higher risk of delivery at 32–36 weeks among whites). Among whites and blacks, there were many common correlates of preterm delivery at 20–31 and 32–36 weeks. Chronic hypertension was significant in both models for blacks and had one of the largest coefficients; i.e., a 1% absolute change in prevalence of hypertension among women giving birth was associated with a 0.21% absolute change in risk of delivery at 20–31 weeks and a 0.39% change in risk of delivery at 32–36 weeks. The set of variables in the final models tended to be different for blacks and whites. Patterns of results were generally similar when we excluded smoking but included California counties and when we included the maximum number of counties possible for each model, rather than just the 468 common counties (data not shown). After inserting values of the independent variables among whites into the models for blacks (i.e., using the regression coefficients from the models for blacks), the mean predicted black-white difference in county-level risk of preterm delivery was 2.0% *higher* for delivery at 20–31 weeks (i.e., 3.0% versus 3.1%) and 14.9% lower for delivery at 32–36 weeks (i.e., 5.2% versus 4.5%). Thus, the hypothetical substitution of the values of the independent variables for whites into the step-wise models for blacks did not result in substantial explanation of the black-white disparity in preterm delivery. Given that maternal smoking was the only variable that seemed to have a substantially more ‘favorable’ distribution among blacks than whites, we re-did these calculations using models that excluded smoking. The substitution of white values of the variables resulted in a predicted risk of preterm delivery that was 32% lower for deliveries at 20–31 weeks (i.e., 3.0% versus 2.0%) and 48% lower for deliveries at 32–36 weeks (i.e., 5.2% versus 2.7%). # Discussion The prevalence of preterm delivery varied two- to three-fold across U.S. counties, and the prevalence distributions were strikingly distinct for black and white women. Together, the selected factors reflecting county-level socioeconomic, demographic and environmental exposures and health explained a substantial amount of the variability in risk of preterm delivery – close to 50% among black women and 70% among white women. However, these factors were less effective at explaining black-white disparities in preterm delivery. That is, underlying relationships in these factors appear relevant to preterm delivery in general but not to the disparity in preterm delivery between blacks and whites. When analyzed separately, variables related to socioeconomic and demographic characteristics of the population (primarily from the U.S. census) and variables related to health from the birth certificate explained about the same amount of variability in the occurrence of preterm delivery. Associations with the majority of the birth certificate-derived variables were in the expected directions (e.g., more smoking or hypertension was associated with higher preterm birth). Some associations with socioeconomic and demographic variables were in expected directions, but several were not (results for high education, income, poverty, wealth and home ownership). In models that included only the socioeconomic and demographic variables, these associations were in the expected directions, except for home ownership (data not shown). Thus, even though we conducted stepwise selection to reduce collinearity, it may have affected results for these variables in the final models. Variables related to environmental exposures did not explain as much variability in preterm birth as the other variables, especially for blacks. Climate was the only environmental variable retained in final models. In all four analytic groups, warmer climate was associated with higher preterm delivery, which agrees with previous literature. When examining all variables together, differences in their distributions did not explain much of the black-white disparity in preterm birth. This is despite the fact that most of the variables had a less favorable distribution for black than white women (e.g., black women had more hypertension and lower socioeconomic level than white women). Smoking is an exception, being lower among black than white women. When smoking was excluded from these analyses, however, the models explained 32% of the black-white disparity in preterm birth at 20–31 weeks and 48% at 32–36 weeks. Our interpretation of these results is that if smoking were as prevalent among black women as it is among white women, the black-white disparity in preterm birth might be considerably greater than it is. As noted above, most studies of preterm delivery have focused on individual- level risk factors rather than what factors explain population-level variability. One previous study examined racial-ethnic variability in the prevalence of preterm birth in the U.S. but was restricted to metropolitan areas and focused more on descriptive differences in preterm delivery by race- ethnicity than on multivariable modeling. Another study focused on racial segregation and county-level prevalence of preterm delivery. Many of the variables we examined were included in a previous analysis of premature adult mortality, except the health-related variables from the birth certificate. The percentage of variability that was explained was higher, however, for premature mortality (72% and 79% among black and white females, respectively, and 86% and 79% among black and white males) than what we observed here for preterm delivery. In addition, differences in the distributions of the variables between blacks and whites explained the bulk of the black-white disparity in premature mortality, which was not true for preterm delivery. Our study was limited in several ways. As an ecologic study, the results cannot be used to make individual-level inference, but they can be used to derive clues about what drives population-level variability in the occurrence of preterm delivery. We designed our analysis to parallel the analysis of premature adult mortality by Cullen et al. as closely as possible. In the future, it would be useful to expand the analytic framework, for example to include other race- ethnicities (e.g., Hispanics, Asians), more recent data years, further refinement within very large counties (e.g., Los Angeles), more environmental exposures (e.g., air pollution), and multi-level analyses that compare area- level with individual-level results. We used data from 1998–2002 to parallel the Cullen et al. paper; it is possible that the observed associations may have changed over time. We consider early and moderately preterm delivery as two distinct but potentially related adverse outcomes. Individual-level studies that examine multiple degrees of preterm delivery typically restrict their comparison group to term deliveries. We used an analogous approach for our analysis; i.e., we excluded one preterm group from the denominator when considering the other preterm group. This was particularly important for analyses of early preterm delivery, because moderately preterm delivery is relatively common and its inclusion in the denominator could thus potentially ‘dilute’ associations with early preterm delivery. A limitation, however, is that this approach does alter the interpretation of our results somewhat, because the outcome does not translate to a traditional estimate of prevalence. This study has illustrated that much of the geographic variability in preterm delivery can be explained by socioeconomic, demographic and health-related characteristics of the population, but less so for blacks than whites. Importantly, however, differences in the distribution of these characteristics between blacks and whites did not explain the marked black-white disparities in preterm delivery. Additional area-level studies are needed to determine what factors explain the remaining variability in prevalence of preterm delivery. Areas of inquiry that we believe are particularly important to explore further are environmental stressors, quality of health care, and more detailed indicators of racial-ethnic and socioeconomic disparity. As has been the case with individual-level risk of preterm delivery, it seems that explaining variability in its prevalence is also a complex challenge. Despite such difficulties, area-level studies provide clues that can be further investigated at the individual level, and they are also important to the development of effective population-level policies aimed at reducing preterm delivery. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SLC MRC. Analyzed the data: JAM. Wrote the paper: SLC MRC JAM JBG PL DKS PHW GMS. Supervised the project: SLC.

# Introduction Due to more advanced and targeted treatments, breast cancer survivors are living longer. Five-year survival among younger women with breast cancer increased from 74% to 89% between 1975 and 2015. As survival rates improved, survivors’ health became a major focus, and a concerning pattern emerged regarding the cardiovascular health of breast cancer survivors: Seven years after recovery, cardiovascular morbidity was twice as high among breast cancer survivors as those without a cancer diagnosis. In one study, cardiovascular disease (CVD) accounted for 35% of the mortality in survivors over 50. Indeed, CVD has become the leading non-cancer cause of mortality among breast cancer survivors. Therefore, identifying breast cancer survivors’ risk factors for cardiovascular morbidity has notable clinical implications. ## The centrality of treatment type Treatment type is a known risk factor for cardiovascular morbidity and mortality. Chemotherapy, particularly anthracyclines (e.g., doxorubicin), is a standard adjuvant treatment for breast cancer, with about one in three women receiving anthracyclines. Despite their effectiveness in fighting cancer, they have a well-documented cardiotoxic consequences, including both systolic and diastolic dysfunction. Similarly, radiation is a risk factor for CVD. For example, one study of breast cancer survivors found the risk of a major coronary event increased linearly with the mean dose of radiation, with an overall rate of 7.4% increase in major coronary events per unit of absorbed ionizing radiation. Therefore, cancer treatment type is a fundamental consideration when evaluating cardiovascular risk. ## Depression and poorer cardiovascular outcomes Depression disproportionality affects breast cancer survivors. A meta-analysis of 72 studies across 30 countries found that almost one in three breast cancer survivors suffered from clinically elevated depressive symptoms, compared to only 13% in the general population. Among non-cancer populations, both clinical depression and self-reported, continuously-measured depressive symptoms are risk factors for cardiovascular morbidity and mortality. One meta-analysis showed depressed individuals had a 30% increased risk of coronary heart disease and myocardial infarction compared to their nondepressed peers. A more recent study found depressed individuals without coronary heart disease were 1.5 times more likely to be hospitalized for diastolic heart failure over a median of nine years’ follow-up. Additionally, longitudinal evidence among non-cancer populations shows that those with a history of manic and hypomanic episodes were also at a higher risk for cardiovascular disease (odds ratio: 2.97, 95% CI: 1.40–6.34). Poorer health behaviors, physiological dysregulation (e.g., autonomic imbalance, systemic inflammation), higher incidence of relevant physical comorbidities, or psychotropic medication usage may help to explain why depressive disorders and self-reported depressive symptoms predict poorer cardiovascular outcomes. ## The current study The current study investigated whether self-reported depressive symptoms and mood disorder history were risk factors for faster cardiopulmonary aging, as assessed via a novel index derived from an exercise stress test (EST), detailed below. Additionally, we looked at the effects of cancer treatment type on the relationship between mood disorder history or self-reported depressive symptoms and cardiopulmonary aging. We hypothesized that higher self-reported depressive symptoms or a mood disorder history would be associated with higher ABEST scores cross-sectionally and longitudinally, and that cardiotoxic cancer treatments (i.e., anthracyclines or radiation) would enhance this effect. We also predicted that cardiopulmonary age would increase between post-surgery and post-adjuvant treatment (Visit 1 and Visit 2), due to the combination of cancer- and treatment-related stress, the previously established cardiotoxic effect of some cancer treatments, and the effect of aging. # Methods ## Participants Women (*N =* 80) were diagnosed with stage I-IIIA breast cancer and were recruited post-surgery and pre-adjuvant therapy after staging was known and treatment decisions were made. They were recruited through the Stefanie Spielman Comprehensive Breast Center at The Ohio State University. Exclusions included age (\< 21 or \> 75), distance from laboratory (\> 100 miles), a prior history of any malignancy except basal or squamous cell skin cancer, neoadjuvant chemotherapy or radiation treatment, stroke, diabetes, current heart disease or uncontrolled hypertension, peripheral vascular disease, prior heart attack, heart failure, heart transplant or other major cardiovascular surgery, liver disease, autoimmune and/or inflammatory diseases, alcohol or drug abuse, steroid use, recent (\<3 months) initiation of antidepressant medication, and other medical conditions that would limit participation (e.g., dementia). See for the flow of participants throughout the study. Note that although 150 participants completed the full Visit 1 protocol, only 80 had all the data necessary for the current analyses. Overall, 14 women were lost to follow-up, and 56 did not have the requisite data for these analyses. For 12 of these 56 women, the COVID-19 pandemic prevented completion of the second EST due to hygienic concerns related to the EST protocol (i.e., wearing the face mask). Other reasons that some individuals did not complete the second EST include: non-cancer medical or orthopedic issues (*n =* 9), cancer recurrence or metastasis (*n =* 3), life circumstances such as moving or caregiving (*n =* 7), or refusal to complete a second EST (*n =* 15). Also, 10 women were either missing covariate data or EST data from Visit 1. The Ohio State University Institutional Review Board approved the study, and participants provided written informed consent. ## Procedure Between 2014 and 2021, participants completed two visits–one soon after surgery and the other two to four years later (*M =* 32 months, *SD =* 6, range: 24–51). During each visit, participants reported their depressive symptoms, engaged in a structured psychiatric interview, and completed an EST. ### Exercise stress test ESTs can identify subtle changes in cardiopulmonary age even in asymptomatic populations. A trained exercise physiologist conducted a graded cycle ergometer test using an Excalibur cycle ergometer (Lode B.V., Groningen, The Netherlands). This test is appropriate for people of all ages and exercise experience. Prior to the test, participants’ oncologists cleared them for testing. During the test, the exercise physiologist instructed participants to cycle at a constant speed (between 50 and 60 revolutions per minute), and resistance was increased every two minutes. Participants continued exercising until they felt that they could no longer do so or reached an exercise time of 15 minutes. In this subsample, the mean exercise time was 9.26 (*SD =* 2.01) minutes. This test provided an index of cardiopulmonary fitness via participants’ peak oxygen consumption (Vo₂max). Also, participants completed the EST while wearing a Firstbeat BodyGuard 2 device (Firstbeat Technologies Ltd, Jyväskylä, Finland) and a breathing mask. The heart monitor collected heart rate data throughout the test including peak heart rate and heart rate one-minute post- exercise. ### ABEST score The Age Based on Exercise Stress Test (ABEST) incorporates three widely used, objective indicators of EST performance: (1) metabolic equivalent of task (MET), the objective measure of energy expenditure relative to body mass; (2) chronotropic reserve index (CRI), a standard measure of heart rate increment during exercise; and (3) heart rate recovery (HRR), the difference between peak heart rate during exercise and heart rate one minute post-exercise. Low levels of all three are associated with higher mortality risk. Compared to chronological age, ABEST was a better predictor of all-cause mortality among a middle-aged, non-cancer sample. Also, ABEST scores may be a helpful tool to use among patients to promote health behavior engagement, as people understand ABEST scores better than the Framingham risk score and other common metrics. Data obtained from the EST were used to calculate ABEST scores. Individual components were calculated as described in the initial ABEST publication. ABEST scores were derived via the established formula: ABEST = 63.637–1.861 METs + 4.858 CRI + 3.633 (abnormal HRR) + 1.316 (beta- blocker use) + 2.226 (non-dihydropyridine calcium antagonist use). ### Depressive symptoms and mood disorder history The 20-item Center for Epidemiological Studies–Depression (CESD) assessed frequency of depressive symptoms in the past week, ranging from not at all ‘0’ to nearly every day ‘3’ for a maximum score of 60 (Visit 1 Cronbach’s α =.74, Visit 2 α =.85). Well-trained research personnel administered the Structured Clinical Interview for the DSM -5 (SCID-V) to assess current and lifetime mood disorder history. The diagnostic team reviewed each SCID-V interview in consensus meetings to obtain diagnoses. If a discrepancy between the interview and the rest of the diagnosticians was present following the meeting, the licensed clinical psychologist provided a recommendation. Lifetime history of Major Depressive Disorder, Persistent Depressive Disorder, Bipolar Disorder, Cyclothymia, and Other or Unspecified Depressive Disorder were collapsed to index mood disorder history. ### Covariates Covariates were selected a priori following a review of the literature. A whole- body dual x-ray absorptiometry (DXA) scan provided data on trunk fat, as central adiposity increases risk for cardiovascular morbidity. Cardiotoxic treatment history was attained via medical chart review (doxorubicin: yes/no, radiation: yes/no). Other relevant covariates included age and education as a proxy for socioeconomic status (some or all of high school, some college, college graduate, graduate or professional training). To be sure that any observed relationships were not due to antidepressant medication, we additionally controlled for antidepressant usage, which was obtained from a self-reported medication list at each visit and coded as a single dichotomous variable (yes/no). ### Physical activity We measured physical activity due to our interest in whether it mediated effects on cardiopulmonary aging. At each visit, participants reported their physical activity levels on the Godin-Shephard Leisure-Time Physical Activity Questionnaire, which is an easy-to-administer, three-item assessment of participants’ physical activity. The measure is weakly correlated with percentile VO2max (*r =* 0.24, *p \<*.001), and percentile body fat (*r =* 0.13, *p \<*.01), and can classify individuals as fit or unfit with 69% accuracy. The questionnaire asks participants how often and for how long in the past week they did strenuous, moderate, and mild intensity exercises, with examples provided for each category. Based on physical activity guidelines, Godin proposed an updated scoring system to utilize only the moderate and strenuous (but not mild) components of the Godin because of the known health benefits of exercise that is at least of moderate intensity; accordingly, we computed the met adjusted activity score based on the vigorous and moderate (but not mild) activity. ### Analytic method Primary models only include participants with ABEST scores at both visits. After completing the EST at Visit 1, 42 participants opted to complete a partial Visit 2 protocol, which did not include the EST. Fourteen others were lost to follow- up, and an additional 14 were missing other relevant data. Therefore, only 80 women were included in our models. We performed unequal variance t-tests and Fisher’s exact tests to compare those who were included versus excluded from models. Those who were excluded from the analysis sample were more likely to have a mood disorder history at Visit 1 (44% vs 26%, *p* =.03) and had higher Visit 1 ABEST scores (*t*(122.3) = 3.31, *p* =.001). In line with our hypotheses, we first used paired t-tests to test whether ABEST scores and its constituent components changed over time. We then used a linear mixed effects model with an unstructured covariance matrix for the repeated visits and Kenward-Rogers adjustment to the degrees of freedom to model ABEST scores at both visits and used as predictors 1) continuous depressive symptoms and 2) mood disorder history, in separate models. Next, we used linear regression models controlling for Visit 1 ABEST scores to predict change in ABEST scores from Visit 1 to Visit 2 using depressive symptoms at Visit 1 and mood disorder history at Visit 1 or anytime between Visit 1 and Visit 2 as the primary predictors of interest in separate models. Lastly, we added two interaction terms between the predictor of interest and cardiotoxic treatment type (doxorubicin or radiation treatment, as separate interaction terms) to examine whether the effect of depression on changes in ABEST scores depended on treatment type. Primary models controlled for age (time-varying), doxorubicin treatment, radiation treatment, education, trunk fat (time-varying), antidepressant use, and time between visits (longitudinal models) or time since surgery (cross-sectional models). Of note, only one of our participants had a bipolar disorder history, and none had a history of cyclothymia, so we conducted sensitivity analyses with this participant excluded, but our pattern of results remained unchanged. As a post- hoc test, we used the PROCESS macro to test whether change in physical activity mediated significant relationships. PROCESS uses bootstrapping to estimate the indirect effect as a test of mediation, and we used 5,000 bootstraps to generate an indirect effect and confidence interval. For all analyses, alpha levels were set at.05. All analyses were conducted in SAS version 9.4 (Cary, NC). Data are available from the corresponding author on reasonable request. # Results ## Demographic information Women included in the analysis sample ranged in age from 26 to 72 (*M =* 51.36, *SD =* 9.79) and tended to be highly educated with 61% earning at least an undergraduate degree. The sample was mostly White (84%) and non-Hispanic (95%). Of the four individuals who identified as Hispanic, three identified as White and one identified as mixed race. In terms of treatment, 39% had chemotherapy, 13% had doxorubicin, 60% had radiation, 46% had a lumpectomy, and 54% had a mastectomy. Doxorubicin was the only anthracycline used among our sample. The surgeries occurred an average of 53 days (SD = 26) prior to Visit 1 (range: 22 to 139). In terms of staging, 48% of women had Stage 1 cancer. Just over a quarter of our sample had a mood disorder history, 19% were taking antidepressant medication, and the average CESD score was 8.70 (SD = 5.51, range: 0–29) at Visit 1. The same number of individuals took antidepressant medication at Visits 1 and 2 (*n =* 23), but only 12 reported taking antidepressant medication at both visits, three discontinued the medication between visits, 11 began taking medication after Visit 1, and 54 did not report antidepressant usage at either visit. Of the 21 who had a mood disorder history at Visit 1, 18 had major depressive disorder (with one also meeting criteria for persistent depressive disorder), one had bipolar disorder, one had persistent depressive disorder and had a major depressive episode between Visits 1 and 2, and one had other specified depression and also had a major depressive episode between Visits 1 and 2. In total, nine of these individuals had a recurrent episode between Visits 1 and 2. Five additional individuals who did not have a mood disorder history at Visit 1 had a major depressive episode between Visits 1 and 2. The average ABEST score (*M =* 57.17, *SD =* 3.23, range: 45.11–64.27) was higher than the average age. Also, 19% of women had abnormal heart rate recovery, defined as fewer than 12 beats per minute. A paired t-test revealed that across the sample, there was no change in physical activity from Visit 1 (Visit 1: *M =* 19.84, *SD* = 19.23) to Visit 2 (Visit 2: *M =* 21.29, *SD =* 20.79, *p =*, 53). See for further demographic information. provides the zero- order correlation matrix for the Visit 1 variables of interest. ## Primary analyses ### Cross-sectional results In terms of the ABEST’s individual components, MET (*p =*.75), CRI (*p =*.32), and HRR (*p* =.11) did not change from Visit 1 to Visit 2. ABEST scores were not significantly different at Visit 2 (mean (SD) = 56.84 (3.65)) compared to Visit 1 (mean (SD) = 57.17 (3.23), *p =*.25). Cross-sectionally, depressive symptoms (*p =*.68) and mood disorder history (*p =*.75) were unrelated to ABEST scores. In terms of covariates, older women (*B =* 0.10, *SE =* 0.028, *F(*1, 73.3) = 13.62, *p =*.0004), those with more trunk fat (*B =* 0.31, *SE =* 0.043, *F(*1, 101) = 52.86, *p \<*.0001), and those with a shorter time since surgery (*B* = -0.31, *SE* = 0.10, *F*(1,95) = 8.61, *p* =.004) had higher physical ages based on exercise testing, but none of the other covariates predicted ABEST scores cross-sectionally (*p*s\>.06). ### Longitudinal results On average, women who had a mood disorder history had an increase in ABEST scores from Visit 1 to Visit 2 (n = 26, mean change (SD) = 0.58 (1.69), 95% CI =.10 to 1.26) while women who did not have a mood disorder history had a decrease in ABEST scores (n = 54, mean (SD) = -0.76 (2.67), 95% CI = -1.49 to.03). This difference was significant even after controlling for Visit 1 ABEST scores and other relevant covariates (*B =* 1.18, *SE =* 0.58, *t(*67) = 2.03, *p =*.046), see. Visit 1 depressive symptoms did not predict change in ABEST scores (*p =*.80). Neither treatment with doxorubicin or radiation modulated the relationships between Visit 1 mood disorder diagnosis or self-reported depressive symptoms and change in ABEST scores over time (*p*s\>.42). In terms of covariates, in models without depressive symptoms or mood disorder history, those with lower Visit 1 ABEST scores (*B =* -0.38, *SE =* 0.11, *t(*68) = -3.48, *p* =.0009) and older individuals (*B* = 0.065, *SE* = 0.030, *t*(68) = 2.15, p = 0.04) had greater change in ABEST scores across time. None of the other covariates predicted ABEST change (*ps*\>.08), though there was a trend for individuals more trunk fat to have greater change in ABEST scores (*B* = 0.098, *SE* = 0.057, *t*(68) *=* 1.72, p = 0.09). ## Ancillary analysis To corroborate our primary significant finding–the longitudinal relationship between mood disorder history and change in cardiopulmonary aging–we used the same modeling strategy to test whether mood disorder history predicted Vo₂max change. On average, women who had a mood disorder history had a decrease in Vo₂max from Visit 1 to Visit 2 (mean change (SD) = -1.02 (2.73), 95% CI = -2.12 to.08) while women who did not have a mood disorder history did not experience this decline (mean (SD) =.30 (3.93), 95% CI =.78 to 1.37). This difference was marginally nonsignificant after controlling for relevant covariates (*B* = -1.57, *SE* =.93, *t(*67*) =* -1.69, *p* =.095). Treatment type did not moderate this relationship (*ps*\>0.22). ## Post-hoc mediation analysis Change in physical activity, as measured by the Godin, did not mediate the relationship between mood disorder history and cardiopulmonary aging (95% bootstrapped CI: -0.16–0.51). # Discussion Among an age-diverse sample of breast cancer survivors evaluated roughly two months post-surgery as well as over two years post-adjuvant treatment, we used a novel index of cardiopulmonary aging—the ABEST and found that mood disorder history predicted ABEST trajectories from pre- to post-adjuvant treatment. Whereas women without a mood disorder history recovered post-adjuvant treatment (i.e., significant declines in ABEST scores), women with a mood disorder history experienced faster cardiopulmonary aging. Ancillary analyses corroborated our primary finding, demonstrating that those with a mood disorder history also exhibited a trend toward greater declines in Vo₂max over time. The primary finding is particularly notable when considered in light of our non- significant predictors of cardiopulmonary aging trajectories, including established cardiotoxic cancer treatments (i.e., radiation, anthracyclines). Additionally, the relationship between mood disorder history and cardiopulmonary age only emerged longitudinally–not cross-sectionally–suggesting that a mood disorder history may become physiologically relevant over time across particularly stressful and taxing life events. Also, only mood disorder history, rather than current self-reported depressive symptoms, predicted ABEST trajectories, which shows that symptoms must be of adequate severity and duration to predict cardiopulmonary aging trajectories in cancer survivorship. ## Mood disorder history and cardiopulmonary aging A mood disorder history may impact post-adjuvant cardiopulmonary aging through physiological and behavioral pathways. For example, clinical depression as well as mildly elevated depressive symptoms can enhance inflammatory responses to stress, which, over time, may lead to the heightened basal inflammation that is observed in a significant subset of depressed patients. Heightened levels of inflammation, in turn, correspond with blood vessel damage and plaque buildup, ultimately increasing risk for CVD. In terms of behavioral mediators, depressed people are significantly less likely to adhere to lifestyle recommendations, such as physical activity guidelines, and half as likely to follow proper medication management. The problematic health behaviors common in depression are associated with an increased risk for cardiovascular disease, stroke, and heart failure. Even so, in the current study, declines in physical activity did not mediate the relationship between mood disorder history and cardiopulmonary aging, suggesting a need to explore other mechanisms that may connect a mood disorder history with faster cardiopulmonary aging. ## Possible support for the scarring hypothesis Our findings provide some support for the scarring hypothesis, which asserts that a mood disorder’s physiological and psychological effects may linger even after remission, increasing vulnerability for future depressive episodes. To date, empirical evidence is mixed. Of note, scars can wax and wane throughout the lifespan, and failing to account for this dynamicity may lead to mixed and inconclusive results. Specifically, scars may reemerge during the stress of breast cancer, increasing the psychological and physiological toll of breast cancer. For example, those with a mood disorder history may have heightened physiological responses to stress, which, over time, may promote the poorer physiological aging trajectories observed in this study. In essence, mood disorders may leave physiological scars that emerge during disease onset or treatment or other particularly stressful times. ## EST tolerance As mentioned above, we had a much smaller sample size for the longitudinal analyses than initially expected. There were many reasons for this attrition, such as cancer recurrence or metastasis, orthopedic issues, or moving. Another issue we repeatedly encountered was that ten percent of our participants refused to complete their second EST because they found the first to be aversive. This percentage is likely even higher because we lost 14 additional participants to follow-up, and it is possible that some of these women discontinued participation because they did not want to complete the second EST. This unexpected phenomenon is notable in and of itself, and it is a critical consideration for future research among breast cancer survivors. Generally, breast cancer survivors have poorer cardiovascular fitness than their age- matched peers, which can make exercise more aversive, and survivors remain physically inactive after cancer treatment. Indeed, in our sample, the average Godin activity score at each visit was below the cut score of 24 that corresponds with American College of Sports Medicine physical activity guidelines. Even so, one small study found that survivors may tolerate treadmill ESTs better than cycling ESTs. Future longitudinal studies conducting repeated ESTs among breast cancer survivors should factor in the possibility of a substantial refusal rate for repeated testing. ## Strengths and limitations This novel study examined depression and cardiopulmonary age before and after adjuvant cancer treatment–a potentially sensitive time for breast cancer survivors. A major strength of the study is the unique timing of measurements, and, in particular, the pre-adjuvant timepoint when breast cancer survivors are recovering from surgery and anticipating what is to come. The study included both self-reported depressive symptoms, which captured the dimensional nature of depression, as well as the SCID-V to assess lifetime mood disorder history–other strengths. Additionally, this is the first study to use the ABEST in a cancer population, allowing us to probe changes in cardiopulmonary age even among patients who are otherwise relatively healthy. Of note, given our strict exclusionary criteria, this breast cancer sample primarily had low-grade breast cancer, did not have CVD, and did not have the higher prevalence of current mood disorders often seen in cancer populations. Also, many of our participants, particularly those who were more depressed and had poorer ABEST scores, did not complete the second EST and therefore were excluded from analyses. Our observations may be even more pronounced in a sicker, more depressed breast cancer sample. Along the same lines, this sample was comprised of mostly White, well-educated women, which is not representative of the general population–a limitation; results should be replicated among a more diverse sample. Lastly, one nuance of our sample that does not align with the larger literature was that those with more advanced cancer were less likely to have a mood disorder history than those with earlier stages of cancer (*r = -*0.23). That said, all but one of our participants had Stage I or II cancer, so due to this restricted range, it is problematic to assign too much weight to this finding. Nonetheless, it is notable that those with a mood disorder history had faster cardiopulmonary aging despite having less advanced cancer, indicating that the observed result is unconfounded by more advanced disease. ## Clinical implications Assessing mood disorder history prior to initiating adjuvant treatment can help to identify those who are at risk for faster cardiopulmonary aging following treatment. Current clinical guidelines recommend depression screening at diagnosis, treatment initiation, and at regular intervals during and after treatment. The current findings underscore the importance of doing so prior to adjuvant treatment initiation as an important predictor of post-treatment cardiopulmonary aging. [^1]: The authors have declared that no competing interests exist.

# Introduction Evidence is growing, that inflammatory processes play an important role in atherogenesis, promoting the risk of cardiovascular diseases. Possible pathophysiological links between inflammation and vascular damage were previously described. One of the most investigated mechanisms is the oxidative modification of LDL, which leads to foam cell formation and development of lesions in the vascular wall. Wang et al. observed a stimulated arterial cell apoptosis and cytokine expression in humans and mice by elevated serum IgE levels. This might be preceded by decreased serum-levels of low-affinity IgE receptor-positive B cells, as observed after coronary artery bypass graft surgery. A possible link between chronic inflammatory activity caused by atopic sensitization and atherosclerosis might be an elevated activity of mast cells, which leads to multiple effects in the vascular wall, promoting development and vulnerability of atherosclerotic lesions. The influence of childhood exposure to several pro-inflammatory risk factors on vascular health in adult life has previously been shown, as well as the relevance of childhood exposure to cardiovascular risk factors for the later development of atherosclerosis. Repeated bacterial or viral infections, obesity and diabetes mellitus are strong promoters of increased carotid intima-media thickness (cIMT) in children by means of chronically elevated inflammatory activity. So far, a possible association of chronic systemic inflammation related to atopic sensitization with cIMT and other biomarkers of early vascular ageing has been analyzed mainly in adult populations. However, a few studies suggested that allergic diseases might contribute to early vascular ageing already in early childhood. Atopic sensitization, independent of its clinical penetrance, involves chronic systemic hyperinflammation. Hence, we hypothesized that atopic sensitization might contribute to early vascular ageing in young people. The primary aim of this study was to investigate a possible association of atopic sensitization, independent of its clinical significance, with the distensibility coefficient (DC) of the common carotid arteries and cIMT, which are indicators of early vascular ageing. Our second aim was to investigate whether the clinical manifestation of atopic sensitization might be associated with these parameters. Hence, we also analyzed the association of allergic disease with DC and cIMT. # Methods ## Study population All Norwegian offspring aged 10 to 18 years of ECRHS Bergen participants were invited to participate in the prospective RHINESSA generation study (Respiratory Health In Northern Europe, Spain and Australia, see [www.rhinessa.net](http://www.rhinessa.net/),). Of the 285 offspring 125 had parental consent for clinical investigation and were screened for eligibility. Exclusion criteria were a recent operation, an acute infection, diabetes mellitus or other chronic inflammatory diseases unrelated to atopy, severe heart disease or pregnancy. Overall, we excluded two candidates because of diabetes mellitus type 1. Of the remaining 123 participants, 21 had no test of immunological total or specific IgE, because they had not agreed for blood analysis. Furthermore, seven participants’ ultrasound images did not meet the predefined quality criteria (see and for further details). Therefore, 95 participants were available for main analysis. Data analysis was performed in accordance with the Declaration of Helsinki and approved by the Regional committee for Medical and Health Research Ethics, Western Region (REC West 2012/1077). Written informed consent was retrieved prior to participation (from parents if the offspring was below 16 years of age, from the offspring themselves if 16 years or older). ## Questionnaires Extensive information on respiratory health, allergic diseases, general health and environmental exposures were assessed by a web-based questionnaire, covering all possible covariates for the analysis: parents’ atopy status, physical activity, frequent exposure to smoking (either active smoking or exposure to regular parental smoking at home), modality of birth (caesarean section versus natural birth) and preterm birth. An interviewer-led questionnaire before clinical examination assessed respiratory symptoms during the last months and specifically during the last three days and current medication. Participants were asked to bring any regular or emergency medication to the study centers. ## Clinical examinations Based on interview data on hours since having smoked or consumed food and drinks, medication use, and current infections, no participants were excluded from specific examinations. Afterwards, we conducted spirometry, FeNO analysis, analysis of total and specific IgE and anthropometric measures (see for detailed information). ## Main predictors: Atopic sensitization and allergic disease Atopic sensitization was defined as a positive total or specific IgE towards inhalant allergens (house dust mite, cat, Timothy (grass, birch, and Cladosporium). Allergic disease was defined as atopic sensitization plus two of the following clinical criteria: allergic rhinitis, atopic eczema, food allergy, allergic bronchial asthma or frequent use of doctor-prescribed antihistaminic medication (see for further details). ## Ultrasonographic examination and main outcomes: DC and cIMT ECG derived heart rate as well as systolic and diastolic blood pressure were obtained simultaneously during the ultrasonographic assessment of DC and cIMT. Blood pressure was measured on the left upper arm with an OMRON 705 IT-IS Automatic-IS device just before the examination was started and after ten minutes of rest in a sitting position. The appropriate cuff size was determined by measuring the upper arm circumference. The procedure of DC and cIMT measurement was performed by two trained field workers, using an ultrasound instrument (UF-870, Fukuda Denshi Co. Ltd., Tokyo, Japan) with a LA38 5–16 MHz linear probe. Temporal resolution was 10.47 ms per frame. Data assessment was conducted using an automated wall-detection software, as previously described and in accordance with current recommendations to ensure acceptable data quality (see for further details;). In the pre-study training examinations intraobserver variability was 7.2 and 7.9%, respectively, and, thus, similar to typical values. Interobserver variability was 10.7% and, thus, slightly higher than previously reported values. Near- and far-wall cIMT as well as end diastolic and peak systolic outer lumen diameter were obtained from four examination planes (bilateral common carotid artery horizontal plane and ear-to- ear plane, respectively). Afterwards, a mean DC \[10⁻³/kPa\] and mean cIMT \[mm\] were calculated for each participant and used for further statistical analysis (see for further details). Validity, reliability and clinical predictive value of DC and cIMT in children and adolescents have been shown previously. ## Statistical analyses Data analysis was performed using SPSS version 25.0 for Windows (SPSS Inc., Chicago, Illinois, USA) and R version 3.5.0 for Windows (R Foundation for Statistical Computing, Vienna, Austria). Descriptive analysis included means, standard deviations (SD), minimum and maximum values. The level of significance was set at *p* ≤ 0.05; estimated effects were reported with 95% confidence intervals (95% CI). Unadjusted (crude) linear regression models, as well as age- and sex-adjusted multiple regression models were applied to analyze associations of DC and cIMT with atopic sensitization and allergic diseases. Complete data were available for 95 (81.1%) of 117 participants. In the remaining 22 cases parental or participants´ consent for IgE analyses was not given. We used multiple imputation by chained equations using the “mice” package in R (version 3.0.3) to impute the missing data. Specifically, we imputed 300 datasets with 10 iterations each. Convergence and distribution of imputed values were assessed graphically. We applied predictive mean matching for imputation of continuous variables, logistic regression for binary variables and ordered logistic regression for ordered categorical variables. The results from the regression models based on the imputed datasets were pooled using Barnard- Rubin adjusted degrees of freedom for small samples. We repeated all statistical analyses with height-related standard deviation scores (SDS) of DC and cIMT, because height seems to be a strong determinant for vascular wall properties in childhood and adolescence. Sensitivity analyses were performed including factors, which are more or less controversially discussed in literature as possibly being relevant for early vascular ageing and also the risk of atopic sensitization. These factors are current exposure to smoking (passive and active), physical activity, preterm birth and delivery by caesarean section. None of these additional analyses resulted in significantly different outcomes and, therefore, they are not presented. # Results ## Basic characteristics of the study population General empiric and vascular characteristic were comparable in participants with and without atopic sensitization. Atopic sensitization was found in 33 (34.7%) of all participants with available blood samples. Of these participants, allergic disease was found in 11 (33.3%) individuals. Nine of them took antiallergic medication on a regular basis at the time of examination. Mean DC was 46.99±8.07\*10⁻³/kPa in the group with atopic sensitization and 51.50±11.46\*10⁻³/kPa in the group without atopic sensitization. Mean cIMT was 0.50±0.04mm in both groups. ## Association of atopic sensitization with biomarkers of early vascular ageing Neither crude comparison of DC and cIMT, nor the age- and sex-adjusted multivariate regression model, indicated a significant association of atopic sensitization with these parameters. However, DC tended to be lower in participants with atopic sensitization than in those without. ## Association of allergic disease with biomarkers of early vascular ageing Neither crude comparison of DC and cIMT, nor the age- and sex-adjusted multivariate regression model, indicated a significant association of allergic disease with DC and cIMT. # Discussion Mean DC tended to be lower in participants with atopic sensitization than in those without. However, atopic sensitization revealed no significant association with DC and cIMT in this study population of Norwegian adolescents. Further, no significant associations of clinically apparent allergic diseases with DC and cIMT were identified. In 2015, Evelein et al. found an increased cIMT in five-year old children with several clinical forms of allergies but no changes in arterial distensibility and elasticity. They concluded that allergies are associated with arterial changes in young children. However, our data do not support their cIMT findings. Possibly, age as well as timing and severity of clinical manifestations of atopy might play a role. Yet, the tendency towards a lower DC in our participants with atopic sensitization might be a very early sign for a chronic subclinical impact of atopic sensitization on vascular ageing. Helpful markers of the severity of systemic inflammatory activity (i. e. oxLDL, high-sensitivity C-reactive protein, soluble interleukin-2 receptor, eosinophil cationic protein) were not available in both studies. However, we analyzed total and specific IgE, which are valid markers for qualitative assessment of atopic sensitization but do not give information about the severity of systemic inflammatory activity. We assume that the inflammatory activity in our study population might have been somewhat heterogeneous, which would explain the lack of association of atopic status and allergic disease with DC and cIMT. Whether the strength of such an association might depend on the severity of systemic inflammatory activity needs to be investigated in larger prospective studies including more information about current and cumulative lifetime systemic inflammatory activity. Another study found an adverse influence of repeated episodes of common childhood infectious diseases on cIMT. This has also been suspected by Liuba et al. in 2005. Their multiple hit theory states that repeated episodes of acute infections might enhance oxidized modification of LDL, which plays an important role in the development of atherosclerosis by fostering foam cell accumulation and subsequent thickening of the vascular wall. This association of inflammation and formation of atherogenic oxidized LDL (oxLDL) has been described in autoimmune disorders as well and a comparable pathomechanism is imaginable in allergic diseases. Acute allergic reactions cause IgE-triggered mast cell activation and an acute-phase reaction, both leading to increased oxidative stress. Activated mast cells degranulate cytokines, leukotrienes, prostaglandins and histamine, leading to endothelial activation and facilitated intracellular penetration of LDL. Endothelial oxidative modification of LDL might be enhanced in children with allergic diseases, due to increased oxidative stress and decreased antioxidative capacity. Furthermore, several authors suggested that IgE-triggered mast cell activation during acute allergic reactions might lead to facilitated presentation of LDL to macrophages subsequently enhancing formation of foam cells. It was also suggested, that atherogenic complexes of CRP and oxLDL with or without β2-glycoprotein-I might be present during acute phase reactions. Accordingly, the severity of clinical penetrance of atopic sensitization in terms of cumulative lifetime load of number and severity of allergic bouts might be a key factor to produce a relevant atopy-associated effect on vascular ageing. The results of our analyses on participants with allergic diseases do not support this hypothesis. However, we did not have information about the severity of past clinical allergic episodes in our participants and the lifetime load of antiallergic treatment. Yet, clarification of this question might be of importance for clinicians, as they would be encouraged to control allergic diseases very carefully in order to avoid adverse effects on long-term vascular health in affected children and adolescents. Future studies should therefore obtain detailed information about the number and severity of acute clinical exacerbations of allergic diseases of their participants in the past and address this question. Finally, it would be intriguing to compare the levels of oxLDL during chronic and acute hyperinflammatory activity in atopic children and adolescents and investigate their association with accelerated arterial stiffening and increased intima- media thickness. ## Strengths and limitations The standardized, high-quality measurements of DC and cIMT in a cohort of adolescents was a main strength of our study. Although early structural changes due to chronic systemic inflammation have been reported in children and adolescents, one should be aware, that the assessment of functional instead of structural alterations (e.g. by flow-mediated dilation), might lead to earlier detection of increased vascular risk related to atopic sensitization. Furthermore, our study does not allow establishing causality, due to its cross- sectional nature. Yet, there was a tendency towards an association of atopic sensitization with decreased DC and the non-significance of the results might be, at least in part, due to the relatively small sample size. We performed analysis of total and specific IgE, which are qualitative markers of the degree of atopic sensitization on the immunological level. Further, based on interview and clinical measurements, our cohort had a relatively thorough phenotyping with regard to allergy. Further detail about the number and severity of clinical episodes of allergic disease, earlier medical treatment and assessment of further inflammatory markers (i. e. oxLDL, high-sensitivity C-reactive protein, soluble interleukin-2 receptor, eosinophil cationic protein) would have been helpful for a characterization of the lifetime load of hyperinflammatory activity in our study population. This may be especially relevant, as the relatively small number of participants with symptomatic allergic disease might be a major limiting factor for the non-significance of our results. Future studies should aim towards a detailed characterization of the extent and severity of inflammatory activity in their participants, as well as a larger sample size. A selection bias due to the exclusion of the 21 participants, due to denied parental consent for blood tests, seems also very unlikely, as their empirical, vascular and clinical allergy characteristics were not different from those included in the study. ## Conclusions and perspectives To the best of our knowledge, this study is the first to analyze a possible association of atopic sensitization and allergic diseases with DC and cIMT in adolescents. Whereas evidence points towards an impact of systemic hyper- inflammation due to atopic sensitization on the vascular endothelium, our results do not support this assumption in adolescents. Better knowledge about the impact of the clinical character, main determinants and potential role of disease control of either atopic sensitization or allergic diseases might be valuable. Further studies should investigate, whether the number and severity of repeated acute clinical bouts of allergic diseases might be a predictor of early vascular ageing, rather than chronic low-grade inflammatory activity mediated by atopic sensitization. # Supporting information We would like to wholeheartedly thank the study participants and support staff that facilitated the study. We would also thank Professor Andreas Bircher (Department of Dermatology, Allergy Unit, University Hospital Basel, Switzerland) for his valuable expert advice. The authors declare that they have no conflicts of interest or sources of income relating to the research. The data are presented clearly, honestly, and without fabrication, falsification or inappropriate manipulation. [^1]: The authors have declared that no competing interests exist. [^2]: ‡ These authors are last authors on this work.

# Introduction ## Background Lung cancer is the most common cancer and leading cause of cancer death in the United States with an estimated 224,000 new cases and 159,000 deaths in 2014. Historically, tumors have been divided into small cell carcinoma (15%) and non- small cell carcinoma/other sub-types (85%) for treatment purposes. Recent advances in treatment of non-small cell carcinomas, including bevacizumab, premetrexed, erlotinib, and crizotinib, have brought hope for prolonged survival and quality of life in advance stage disease. These recent advances have also required pathologists to accurately sub-classify non-small cell carcinomas and conserve tissue for molecular/advanced testing. Non-small cell carcinomas can be sub-classified into multiple histologic sub- types, but are primarily classified as squamous cell carcinoma, adenocarcinoma, or large cell carcinoma. This classification is typically done by morphology and immunohistochemistry (IHC) as needed. Bevacizumab is contraindicated in the treatment of squamous cell carcinoma due to risk of life threatening pulmonary hemorrhage. Pemetrexed treated adenocarcinomas show improved outcomes. EGFR mutated tumors comprise approximately 11% of lung adenocarcinomas and have a 70% response rate to erlotinib with prolonged survival. Finally, ALK and ROS-1 mutated tumors (typically adenocarcinomas or large cell carcinomas) have high response rates (61–72%) and prolonged survival when treated with crizotinib. Only 30% of non-small cell carcinoma lung cancer patients are surgical candidates. Therefore, the diagnostic biopsy may be the only material available to diagnose and guide treatment for 7 out of 10 patients. Optimal conservation and utilization of biopsy material is critical for sub-classification and identification of specific mutations, which may be responsive to targeted therapies. Consensus guidelines and multi-organizational reports have stated the importance of a multidisciplinary strategy for optimal collection and utilization of diagnostic biopsy tissue, but few specific recommendations have been given to put into practice. The 2011 the International Association for the Study of Lung Cancer/American Thoracic Society/European Respiratory Society (ATS/ERS/IASLC) classification review was one of the first prominent publications to emphasize a multidisciplinary approach, but to our knowledge there is no documentation in the medical literature quantifying the effect of a multidisciplinary strategy to increase the volume of diagnostic tissue for molecular/advanced testing. The purpose of this study was to evaluate the effects of such a strategy developed and implemented at a community hospital (Saint Bernards Medical Center, Jonesboro, AR) in late 2008 for CT-guided biopsies. ## Multidisciplinary Strategy Development As demands for advanced testing on small biopsy samples have significantly increased for EGFR mutations, and more recently ALK and ROS-1, the need for an effective strategy from a systemic standpoint was recognized, and a multidisciplinary team (pathology, interventional radiology, and oncology) was assembled to develop a practical strategy to optimally handle and allocate tissue for current and potential future testing. Prior to 2008 interventional radiologists would perform a CT-guided biopsy and prepare touch preparations of the core tissue, which would be reviewed by a pathologist as part of a rapid onsite evaluation (ROSE) to determine adequacy. If tissue was inadequate, then additional biopsies (passes) would be performed by the radiologist until there was confidence diagnostic tissue had been obtained, at which time the procedure would be terminated. Biopsy material would then be placed in 10% formalin solution and typically submitted in a single cassette for histologic processing. IHC, fluorescence in-situ hybridization (FISH) or molecular diagnostic testing would be performed on tissue available in the remaining block or unstained slides when needed. The purpose of the multidisciplinary strategy was to increase diagnostic tissue volume and maximize tissue available for ancillary testing. Input was obtained from interventional radiologists, oncologists, and pathologists to better understand each specialty’s needs and limitations, which resulted in a consensus three-pronged strategy for CT-guided biopsies. This strategy focused on the following: Once the radiologist is confident diagnostic tissue has been obtained, and if safe for the patient based on the radiologist’s clinical assessment, additional passes (biopsies) are performed to increase the volume of diagnostic tissue available for ancillary testing. A specific number of total biopsy passes or total tissue volume was not recommended, but at least two passes were implied. The consensus of the radiologists was to perform in the range of 2–4 biopsy passes depending on the location and characteristics of the lesion. Biopsy material is placed in multiple tissue cassettes (typically two) to minimize loss of diagnostic tissue if IHC or special stains are performed on one of the two blocks (leaving the second block available for ancillary testing). All tissue ribbons obtained during cutting at the microtome would be saved on unstained slides between levels, which may be used for IHC stains, FISH, or molecular studies. The hypothesis for this study is that the change in approach to CT-guided biopsies increased material available for advanced testing. This study evaluates the effectiveness of the above described multidisciplinary team strategy for CT- guided lung biopsies and compares it to historic methodology with an emphasis on quantification of changes in diagnostic tissue volume. Additional effects including complications, tumor volume, need for ROSE, and rate of malignant diagnoses were also examined. # Materials and Methods This study is a retrospective analysis comparing CT-guided lung needle core biopsies performed in 2007 and 2012 to evaluate the sustained effect of the multidisciplinary strategy for obtaining and processing CT-guided lung biopsies, which was implemented in late 2008. Internal review board (IRB) approval was obtained through Saint Bernards Medical Center, Jonesboro, AR (Study \#140801). Patient consent was waived, and patient data were anonymized and de-identified prior to analysis. CT-guided biopsy reports were obtained for 209 consecutive biopsies in 2007, and 208 consecutive biopsies in 2012 at a single institution (Saint Bernards medical center) to represent steady-state practices before and after implementation of the multidisciplinary strategy. For the purpose of this study, only CT-guided lung biopsies were evaluated. The inclusion criteria were patients with a lung lesion who underwent a CT-guided biopsy resulting in a pathology report and had slides available for review. Exclusion criteria included cases from radiologists who did not perform CT-guided lung biopsies in both 2007 and 2012 at the study institution. Patient medical records were reviewed and included imaging studies, demographics, procedural notes, and pathology reports. Lung biopsy microscopic slides were retrieved and analyzed for number of tissue fragments, minimum and maximum tissue fragment size, and total biopsy sample size by pathologist \#1 (PEF). Total biopsy sample size was estimated to the nearest 0.1 cm using a transparent ruler containing tick marks to the nearest 0.05 cm (Ted Pella, Inc., Cat. No. 54480, Redding, CA). Pathologist \#2 (EWS) estimated tumor percentage (area/volume estimate) visually on each slide in 5% increments. It was assumed for the purpose of the study that the volume of tissue available for testing was directly proportional to specimen length on the microscope slide. $$Volume = \pi\left( r^{2} \right) \times Length$$ Technical aspects of obtaining CT-guided lung biopsies were unchanged between 2007 and 2012. Bard (C.R. Bard, Inc., Tempe, AZ) biopsy needles (19-gauge) were used exclusively for both 2007 and 2012 data sets maintaining a constant biopsy radius/diameter. Fluoroscopic CT-guidance was introduced in the radiology department 2009, which shortened the length of the procedure, but did not alter the technical methodology of obtaining biopsy material. Complications for the purpose of this study were defined as secondary events caused by the biopsy procedure (e.g. pneumothorax or pulmonary hemorrhage), which resulted in additional treatment or prolonged observation (e.g. chest tube placement or admission for observation), and were identified through review of the interventional radiology biopsy procedure notes. ## Statistical Analysis Statistical analysis was performed using GraphPad Prism Version 6 (GraphPad Software, Inc., La Jolla, CA). Ordinary one-way analysis of variance (ANOVA) was performed on both 2007 and 2012 data sets to determine if significant differences were present between radiologists. Unpaired t-test with Welch’s correction was used to evaluate changes in biopsy size between 2007 and 2012 for individual radiologists. The Mann-Whitney U test was used to evaluate the overall total biopsy length, and the Fisher exact test was used to evaluate statistical significance for the rate of malignant diagnoses, use of rapid onsite evaluation (ROSE), number of blocks submitted, and complication rate. # Results One-hundred and fifty CT-guided lung biopsies were performed in 2012 for mass lesions compared to 130 in 2007. Nine cases from 2007 were excluded from analysis. Four cases did not have microscopic slides available, and 5 cases were performed by radiologists who did not perform CT-guided lung biopsies at the study institution in 2012 for comparison. Thirty eight cases were excluded from the 2012 data analysis as they were performed by a radiologist who did not perform CT-guided lung biopsies at the study institution in 2007 for comparison. As a result, there were 121 cases from 2007 and 112 cases from 2012 being included in the study data set for analysis. The average total biopsy sample size increased from 1.0 cm (0.9–1.1 cm) in 2007 to 2.5 cm (2.3–2.8 cm) in 2012, which correlates to a 2.5 fold increase in average diagnostic tissue volume highlighted in and. shows the biopsy sample size distribution frequency as a function of biopsy sample size in 0.5 cm increments, and compares biopsy sample size for individual radiologists between 2007 and 2012. Ordinary one-way ANOVA analysis of the 2012 data set showed no significant difference in biopsy sample size between all three radiologists (P = 0.1175), but the same analysis of the 2007 data set showed a significant difference among the radiologists (P\<0.0001). This difference was identified to be between radiologist \#3 and both radiologists \#1 and \#2 using Tukey’s multiple comparison test. The increase in biopsy size from 2007 to 2012 was statistically significant for each individual radiologist using an unpaired t-test with Welch’s correction. The percentage of cases diagnosed with malignancy increased from 62.0% (n = 75) in 2007 to 76.8% (n = 86) in 2012 (P = 0.0104). Average portion of biopsy material involved by tumor (volume estimation) was 28% (23–33%, 95% CI) and 35% (30–40%, 95% CI) for 2007 and 2012, respectively. The spectrum of CT-guided lung biopsy diagnoses is shown in. The rate of non- small cell carcinoma, not otherwise specified (NOS) diagnoses decreased substantially from 29.3% to 4.6% between 2007 and 2012 (P\<0.0001). Miscellaneous neoplasms included large cell carcinoma with neuroendocrine features (1), lymphoma (2), carcinoid (1), and pulmonary hamartoma (1). Non- diagnostic samples included necrosis without identifiable lesional tissue and benign lung, which was not representative of a radiologic abnormality. The only clinically significant complication identified from the biopsy procedure was pneumothorax requiring chest tube placement. In 2007, 15% (n = 18) of biopsies required chest tube drainage, but this decreased to 7% (n = 8) in 2012 (P = 0.065). The number of biopsies submitted in at least two cassettes increased from 11% to 96% between 2007 and 2012 (P\<0.0001). # Discussion ## Biopsy and Tumor Volume Average total biopsy sample size, which is directly proportional to volume, increased from 1.0 cm (0.9–1.1 cm) in 2007 to 2.5 cm (2.3–2.8 cm) in 2012 after implementation of the multidisciplinary strategy. The increased volume appears consistent for all radiologists. A total of 3 radiologists submitted lung biopsy material in both 2007 and 2012 with similar results. Radiologist \#3 difference was statistically significant compared to both radiologists \#1 and \#2 for 2007 by one-way ordinary ANOVA analysis with Tukey’s multiple comparisons test. However, the same analysis of 2012 data did not show statistically differences between the 3 radiologists, implying better consistency among the interventional radiologists after implementation of the multidisciplinary biopsy strategy. The proportion of biopsy material involved by tumor was similar at 28% (2007) and 35% (2012) while the biopsy volume increased 2.5 fold. Therefore, increasing the biopsy volume appears to have the desired and expected effect of increasing tumor volume available for testing. These results further emphasize the importance of maximizing the biopsy volume from the procedure to have maximum potential tumor volume available for advanced testing. ## Tissue Adequacy for Molecular Testing Few studies have been published which evaluate the effect of tumor volume on the sensitivity of molecular studies for CT-guided lung biopsies. One study comparing molecular analysis on both CT-guided lung biopsy and the subsequent surgical resection specimen, demonstrated a failure rate of 11% in CT-guided biopsies to detect EGFR or KRAS mutations in a series of 18 patients. In a recent study evaluating the effectiveness of crizotinib in ROS-1 mutated lung tumors, 4 of 26 (15%) FISH proven ROS-1 mutated tumors failed to have detectable mutations by next generation sequencing (NGS) when testing the remaining available biopsy material. Large tertiary referral medical centers have documented an adequacy rate for molecular studies from CT-guided biopsies ranging from 67% to 91%. Unfortunately, characterization of selection bias for testing or remaining tumor volume available for testing was not described in these studies. Studies have typically documented procedural strategy in terms of “biopsy passes” and not in the form of biopsy or tumor volume. This issue raises concern that insufficient tumor may have been present for molecular testing. Future studies including biopsy and tumor quantification may be helpful to define goals for capturing adequate tissue during the pre-analytical tissue acquiring phase. A recent article by Ferretti et al. demonstrated an increase in biopsy length of 15.6% from 10.9-mm to 12.6-mm after publication of the ATS/ERS/IASLC guidelines in 2011. Their adequacy for molecular testing of EGFR pyrosequencing increased from 85% to 98% over the same time period. Needle size used was an 18-gauge Angiotech BioPince (Medical Device Technologies Inc., Gainesville, FL), which has 49% more volume per unit length than a 19-gauge Bard biopsy needle used in the current study. There is no description in their study of any institution- based educational or multidisciplinary strategy intervention during the time between the two data sets. The multidisciplinary-based strategy used in this study to obtain and share input from physician stakeholders in the patient’s care may be an important difference between the 15.6% increase in biopsy volume (Ferretti, et al.) compared to the 250% increase in this study. No known studies using CT-guided lung biopsies have evaluated tumor volume and molecular testing sensitivity with correlation of testing in subsequent lung resection specimens. Such an analysis will be important, knowing the growing dependence of molecular tumor analysis in therapeutic decision making with 70% of lung cancer patients not being surgical candidates, and their entire treatment plan being dictated by information present in the biopsy material. Advanced testing methodologies like next generation sequencing (NGS) are becoming more routine and show promise to be front line in the testing of lung carcinoma because the number of genes tested is very large at a relatively low cost. This study takes the initial steps in a methodology strategy to increase tumor volume available for testing. An important future direction will be to verify the effect of strategies, such as this one, with regards to minimizing false negative results with advanced molecular testing in small biopsies. ## Rapid Onsite Evaluation A secondary observation in this study was that the use of immediate cytologic assessment by a pathologist, commonly referred to as rapid onsite evaluation (ROSE), decreased significantly from 85% of cases in 2007 to 0% in 2012 among the radiologists included in this data set (P\<0.0001). ROSE is commonly performed on cytology and needle core biopsy samples to determine if they contain diagnostic material from which to make a definitive pathologic diagnosis. If diagnostic tissue is not present, then additional tissue may be obtained immediately to increase the likelihood of a definitive diagnosis, and potentially preventing an unnecessary repeat biopsy. After implementation of the multidisciplinary strategy in this study, interventional radiologists included in this analysis gained tacit knowledge that the overall diagnostic rate was at least equally sensitive after obtaining at least two needle core biopsies irrespective of whether a pathologist had performed ROSE or not. Subsequently, they changed from their historic practice and began to submit biopsies without ROSE. The perceived benefit by the radiologists was a shorter and possibly safer procedure. The diagnostic malignancy rate increased between 2007 and 2012, while the complication rate decreased. If a standard biopsy size or volume recommendation could be determined, which maximizes the likelihood of adequacy for diagnosis and advanced testing, then the use of ROSE could be significantly lowered and would result in significant cost savings. Additionally, a recent article by Rekhtman, et al. demonstrated that performing touch preparations on needle core biopsy tissue can decrease the total DNA content on average 15–50% depending upon how vigorous the touch preparations are made. These combined factors raise the question whether ROSE by a pathologist is necessary when an adequately sized biopsy is obtained, and if it may contribute to unnecessary expense, loss of valuable tissue for advanced testing, and increasing the procedure time that could contribute to complication risk. Additional study is needed in these areas. ## Number of Tissue Blocks Submitted After implementation of the multidisciplinary strategy, the number of specimens with biopsy material placed in at least two blocks increased from 11% in 2007 to 96% in 2012 (P\<0.0001). Tumor was present in both tissue blocks in 91% of the cases diagnosed with malignancy in 2012 (n = 86). Dividing biopsy tissue into two separate cassettes allows for less tissue to be consumed by immunohistochemistry than if all diagnostic tissue was submitted in a single cassette. Therefore, this strategy will on average increase the volume of tumor tissue remaining available for ancillary studies after the primary diagnosis has been made. In theory, if one estimates half of the diagnostic biopsy tissue is consumed in making the primary diagnosis including IHC stains, and there is on average equal tumor present in both blocks, then submitting tissue in two cassettes would increase the amount of tumor available for molecular/advanced testing by 50% compared to submitting the biopsy sample in a single cassette. Combining this with the 2.5 fold increase in average tumor volume in this study would result in increasing tissue available for advanced testing 3.7 fold compared to the historic (pre-2008) strategy. ## Diagnosis highlights the distribution of diagnoses between 2007 and 2012. The most significant change over time was the decrease in the non-small cell carcinoma, NOS diagnosis rate from 29.3% to 4.6% (P\<0.0001). This decrease is related to both increased utilization of immunohistochemistry by pathologists to assist in sub-classification, and in response to the clinician’s need to determine therapy given recent advances in chemotherapy (bevacizumab and pemetrexed), which are utilized in lung adenocarcinomas, and targeted therapies (crizotinib and erlotinib) for which mutations are most common in lung adenocarcinomas. also shows the rate of malignant (neoplastic) diagnoses increased from 62.0% in 2007 to 76.8% in 2012 (P = 0.0104). The increase in biopsy sample size may be the significant factor. The incidence of small cell carcinoma (4.6–5.3%) is below the expected incidence of approximately 13%. The reason for this apparent discrepancy is not clear from the data analyzed and is beyond the scope of this study, but could be related to choice of diagnostic methodology from the ordering physician. Small cell carcinomas tend to be more centrally located and may be more commonly biopsied through bronchoscopy or endobronchial ultrasound guided fine needle aspiration compared to CT-guided biopsy. Granulomatous inflammation was a common diagnosis in this analysis. Histoplasmosis and blastomycosis are endemic to the geographical region in this study, which relates to the prominence of this finding. The non-diagnostic / inadequate rate ranged between 5.8% in 2007 and 1.8% in 2012. This may be an underestimate of the false negative rate because some of the other benign diagnoses (e.g. inflammation/fibrosis) may represent un-sampled neoplasms. A significant portion of the decrease from 2007 to 2012 may be due to the increased neoplastic diagnosis rate from 2007 to 2012. This is a secondary observation of the data and not the primary endpoint of the study, and requires further study. ## Complications Similar to previous interventional radiology studies, aside from chest tube placement for pneumothorax, complications are uncommon for CT-guided lung biopsies. Some reports noted hemorrhage, but no cases of clinical significance were identified in our study. Pneumothorax not requiring chest tube placement was not considered clinically significant for the purpose of this study. The rate of pneumothorax requiring chest tube placement decreased from 15% (n = 18) to 7% (n = 8) from 2007 to 2012 (P = 0.065). While this study was not designed to evaluate the underlying causes of this reduction, two main changes in workflow occurred during the intervening time between 2007 and 2012. First, fluoroscopic CT-guidance was installed in the CT-suite, which allows the radiologist to more quickly evaluate needle biopsy placement in near real-time at the bedside without having to walk between the patient and control room. Second, after changing the radiologist’s procedure to obtain more biopsy passes, the perceived tissue adequacy rate was high enough that radiologists did not routinely request ROSE by a pathologist. These two factors likely combined to reduce the length of time the biopsy needle is in the patient, which may be responsible for the decreased incidence of pneumothorax requiring chest tube placement. Higher pneumothorax rates are associated with emphysema, smaller lesions, lesion depth, and operator experience. Most studies did not correlate the length of time the biopsy needle is in the patient with risk of pneumothorax or chest tube placement. One study demonstrated that use of fluoroscopic CT-guided biopsies cut procedural time by 50% compared to conventional CT-guided biopsy. Our data may indicate that the length of time the biopsy needle is placed in the patient could be another important risk factor of pneumothorax requiring chest tube placement. Unfortunately, this finding did not reach statistical significance for this study and lacks the data to further evaluate the possible association. There is a wide range of reported incidences of pneumothorax for patients undergoing CT-guided lung biopsy, which has ranged from 9–54%. In one study, up to 31% of patients undergoing CT-guided biopsy of small pulmonary nodules required chest tube placement. The largest known study, which evaluated 15,865 cases from 4 different states for 2006, showed an overall average of 15% \[CI, 0–50%\] for any pneumothorax and 6.6% \[CI, 0–25%\] for pneumothorax requiring a chest tube. Our 2012 pneumothorax chest tube placement rate (7%) was consistent with large-scale published data. Based on this data, there does not appear to be an increased risk of pneumothorax requiring chest tube placement after implementation of the multidisciplinary strategy, which should be reassuring to clinicians hesitant to perform additional biopsy passes out of fear of such complications. # Conclusions The findings of this study support the general consensus recommendations from multiple authors and organizations on the importance of having a multidisciplinary strategy for collection and handling of diagnostic lung biopsy tissue. Such an approach to CT-guided lung biopsies, which focuses on additional biopsy passes and submission of tissue in multiple cassettes, was effective in significantly increasing biopsy and tumor volume for potential molecular/advanced testing. Laboratories may want to consider strategies similar to those outlined in this study for small diagnostic biopsy material. As molecular/advanced testing expands and becomes more commonplace, it will become important to develop best practice recommendations for obtaining and handling small biopsy specimens. Future studies correlating tissue adequacy for molecular testing with biopsy size and tumor composition may be helpful in defining best practices for tissue acquisition. The multidisciplinary strategy implemented in this study does not appear to increase risk of complication, which in fact was actually decreased. There may also be additional benefits including higher rate of malignancy detection and less need for ROSE, which could have significant cost savings. Additional study of these specific areas is needed. The authors appreciate the assistance of Ben Ball in the histology department of St. Bernards Medical Center. [^1]: Dr. Philip Ferguson (PEF) has employment affiliations with PathMD as outlined in the Financial Disclosure, and consultancy relationships with Leica Biosystems and Pfizer. None of these relationships had any direct or indirect influences on this study. These affiliations do not alter the authors’ adherence to all PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: PEF. Performed the experiments: PEF EWS CMS DCH. Analyzed the data: PEF CMS DCH. Contributed reagents/materials/analysis tools: PEF CMS DCH EWS. Wrote the paper: PEF CMS DCH EWS.

# Introduction Infections by multi-resistant microorganism are a problem increasing in our clinical practice. Specially, extended-spectrum beta-lactamase producing enterobacterales (ESBL-EB) are a priority. They are part of the group posing the highest public health risk according to the WHO statement published in 2018. These microorganisms are a frequent cause of urinary tract infection, with a high incidence in the hospital setting. In Europe, *E*. *coli* resistance to third generation cephalosporins in 2018 was 15,1% and *K*. *pneumoniae* resistance to third generation cephalosporins was 31,7%. In United States, there are high levels of antimicrobial resistance in enterobacterales in hospitalized patients and rates of ESBL-EB increased between 2013 and 2017. Several strategies have been developed to reduce the incidence of infections by these microorganisms. It is essential to optimize antibiotic therapy in accordance to the need, the drug choice and the duration of the treatment. There is evidence on the effectiveness of short course of antibiotic therapy in the treatment of urinary tract infections. However, there are no specific studies on the duration of antimicrobial treatment in complicated urinary tract infections caused by ESBL-EB. The aim of this study was to evaluate the clinical outcome in patients who received short (≤ 7 days) versus long courses (\>7 days) of antimicrobial therapy for complicated ESBL-EB urinary tract infections. # Material and methods This is an observational and retrospective study. Positive urine cultures for ESBL-EB in our hospital between March 2015 and July 2017 were identified. The database of the Microbiology department of the hospital was used. Patients who met criteria for complicated urinary tract infection according to the IDSA guidelines were included. A second analysis was performed including only cases that met the recommended criteria for conducting clinical trials by the FDA: clinical syndrome characterized by pyuria and a documented microbial pathogen on culture of urine or blood, accompanied by local and systemic signs and symptoms, including fever (i.e., oral or tympanic temperature greater than 38 degrees Celsius), chills, malaise, flank pain, back pain, and/or costovertebral joints pain or tenderness, which occur in the presence of a functional or anatomical abnormality of the urinary tract or in the presence of catheterization. Exclusion criteria were: age under 18 years old, asymptomatic bacteriuria, uncomplicated urinary tract infections, recurrences in the first week after the end of treatment, polymicrobial urine cultures, inadequate treatment according to antimicrobial susceptibility results and death before the end of the antibiotic treatment. The study was approved by “Galicia Research Ethics Committee”. All data were deidentified before we accesed them. Both Specialist and Primary Care electronic medical records were reviewed and a database was created with the following variables: - **Demographic and anthropometric variables:** Age, sex, weight, height and BMI. - **Origin of infection:** Community, nosocomial (defined as occurring after the second day of admission or within 10 days after discharge) or healthcare- associated (defined as admission to the hospital in the previous 90 days, institutionalized patient, treatment in day units, dialysis or home hospitalization). - **Devices within 7 days prior to collection of urine culture:** Central and urinary catheter or gastrostomy - **Comorbidities:** Hypertension, heart failure stage C (rated by ACC / AHA), heart disease of any etiology (including hypertensive, ischemic or valvular heart disease, heart failure, arrhythmias and cardiomyopathy), diabetes mellitus with or without target organ damage (neuropathy, nephropathy, retinopathy), estimated glomerular filtration rate calculated by the MDRD modified formula (categorizing the degree of renal failure as severe \< 30 ml/min and moderate between 30–60 ml/min according to the KDOQI guidelines of the National Kidney Foundation), COPD confirmed by spirometry, peripheral arterial disease (intermittent claudication, acute arterial ischemia, aortic aneurysm \> 6 cm, peripheral by-pass), cerebrovascular disease (transient ischemic attack or ischemic stroke), cognitive impairment, solid tumor, leukemia or lymphoma, connective tissue disease, mild chronic liver disease (non portal hypertension) or severe with portal hypertension, AIDS, neutropenia (\< 500 neutrophils at the time of urine culture), corticosteroids or immunosuppressive treatment. Charlson comorbidity index was calculated for each patient. - **Barthel index** - **Clinical and analytical features of each episode:** Measure of systolic and diastolic blood pressure, heart rate, temperature, serum leukocyte, hemoglobin, platelets, urea and creatinine in the urine culture collection day or in the nearest time. Presence of sepsis or septic shock criteria were evaluated. - **Microbiological features:** Urine samples were cultured in blood agar and CPS agar and they were incubated a minimum of 24 hours at 35–37°C in aerobic atmosphere. The microorganism was identified and sensitivity tests were carried out in the Biomerieux VITEK 2 system. For the identification of ESBL production, phenotypic methods were used according to the CLSI recommendations. - Adequate empirical antibiotic treatment, adequate definitive treatment according to antibiogram and antibiotic treatment duration. The primary outcome analyzed was all cause 30-day mortality. The secondary outcome was a combined item of all cause mortality and reinfection by the same enterobacteria at 30 days. ## Statistical analysis Differences in baseline characteristics between treatment groups were analyzed using the Chi-square test or Fisher's exact test for the analysis of dichotomous variables and the Student's T test in the quantitative variables with normal distribution or the Mann-Withney U test in the quantitative variables without normal distribution. Quantitative variables with statistically significant differences were dichotomized using their median as a cut-off point. Variables clinically relevant and with P \<0.05 were included in a binary logistic regression model, calculating a propensity index for assigning each patient to short treatment. For the bivariate survival analysis, Kaplan-meier curves were performed and the Log-rank test was used to assess the differences between curves. Finally, a multivariate survival analysis was performed using Cox regression, using a temporary variable (number of days to the event) and a dichotomous variable (occurrence of the event). All the variables that presented a P \<0.1, as well as the Charlson comorbidity index, sex, age and the previously calculated propensity index were included in the multivariate analysis. Stratified analysis by sex was subsequently performed. Statistical significance was established with P\<0,05. The statistical program SPSS 21 was used. # Results 273 urine cultures were positive for ESBL-EB during the study period. 75 episodes in 70 patients met the inclusion criteria and were included, 40 in the long treatment group and 35 in the short treatment group. The median age was 79 (ST 16.1) years and 43 (57.3%) were women. The median Charlson index was 2 (ST 2.02). Proportion of male sex was higher in long treatment group (55% vs 28,5%; OR 3; 95% CI 1,1–7,9; P = 0,02). In bivariate analysis there were more patients with hypertension in the short treatment group (47,5% vs 71,42%; OR 2,76, 95% CI 1,05–7,22; P = 0,03). Although in the multivariate analysis the only significant difference between groups was male sex. There were no other differences in baseline characteristics between groups. The most frequently isolated microorganism was *Escherichia coli* (61 cases, 81.2%), followed by *Klebsiella spp* (11 cases, 14.7%) and *Proteus mirabilis* (2 cases, 2.7%). 36 patients (48%) presented urinary tract infection without fever, 26 patients (34.6%) febrile urinary tract infection and 13 patients (17.4%) pyelonephritis. Blood cultures were performed in 16 cases (21.3%) and 4 patients (5.3%) presented bacteremia due to the same enterobacteria. Sepsis or septic was observed in 8 patients (10,7%). 93.1% of patients presented a temperature 37.8ºC or less during the first 48 hours and 94.8% hemodynamic stability during the first 48 hours. There were no differences in the type of infection, the severity of infection or vital signs at day 2 between both groups. 55 patients (73.3%) were hospitalized at the time of the urine culture. Community-acquired urinary tract infections were observed in 32 patients (42.7%), healthcare-associated in 34 patients (45.3%) and hospital-acquired in 9 patients (12%). 22 patients (29.3%) had an urological abnormality and 12 (16%) had permanent urinary catheter. In 36 patients (48%) adequate empirical antibiotic therapy was performed. Definitive treatment with carbapenem was used in 36 patients (48%), followed by fosfomycin in 13 (17.3%), quinolones in 12 (16%), cotrimoxazole in 7 (9.3%), beta-lactam plus betalactamase inhibitor in 5 (6.7%), furantoin in 2 (2.7%) and aminoglycosides in 1 (1.3%). Mean treatment duration in short and long treatment groups was 6,1 and 13,8 days respectively. Mortality at 30 days was 5.7% in the short treatment group and 5% in the long treatment group without significant differences (P = 0,8). Mortality or reinfection by the same microorganism at 30 days was 8.6% in the short treatment group and 10% in the long one (P = 0.8). Mortality at 30 days in the male subgroup was 0% in the short treatment group and 10% in the long treatment group (P = 0.1). Mortality or reinfection at 30 days in the male subgroup was 10% in the short treatment group and 9.1% in the long treatment group (P = 0.9). For the analysis of patients following the FDA criteria for complicated urinary tract infection, we excluded 24 patients. Finally, 51 cases met the inclusion criteria and were included, 22 in the long treatment group and 29 in the short treatment group. There was no difference in 30-day mortality (4.5% in short treatment vs 3.4% in long treatment, P = 0,8) nor in 30-day mortality or reinfection (9.1% in short treatment vs 6.9% in long treatment, P = 0.7). In the multivariate analysis, the factors associated with higher mortality at 30 days were the presence of leukopenia at the time of diagnosis (HR 16.4; 95% CI 1.4–181.1; P = 0.02) and the history of metastatic cancer (HR 1.6; 95% CI 1.09–2.4; P = 0.01). The factors associated with higher mortality or reinfection were the history of lymphoma (HR 9; 95% CI 1.004–81.2; P = 0.05) and definitive treatment with beta-lactam associated with beta-lactamase inhibitor (HR 8.4; CI 95% 1.5–46.1; P = 0.01). # Discussion We have not found differences in mortality and reinfection between long and short treatment groups in patients with complicated urinary tract infection caused by ESBL-EB. There are no clear recommendations in guidelines regarding the duration of antibiotic treatment in complicated urinary tract infections. However, there are studies that also support a short course of antibiotic therapy in these types of infections. A prospective, non-inferiority trial undertaken at 21 centres of infectious diseases in Sweden showed that 7 days of ciprofloxacin was not inferior to 14 days of treatment in women with acute pyelonephritis, including older women and those with a more severe infection. The clinical and bacteriological cure rates were high for both regimens. A randomized, placebo- controlled trial conducted in women and men with febrile urinary tract infection showed that 7 days of antimicrobial treatment in women were non-inferior to 14 days of therapy. In men, 7 days of antibiotic treatment was inferior to 14 days during short-term follow-up (10–18 day post-treatment) but it was non-inferior when looking at longer follow-up (70–84 days post-treatment). A meta-analysis revealed that seven days of treatment for acute pyelonephritis is equivalent to a longer treatment in terms of clinical failure and microbiological failure, including bacteremic patients. However, in patients with urogenital abnormalities, the evidence suggested that a longer treatment is required. In the randomized trial by Dinh *et al*., the efficacy of 5 days of fluoroquinolone treatment does not seem different from 10 days of treatment for acute uncomplicated pyelonephritis. Similar results were observed in the study published by Peterson *et al*. They compared 5 days of levofloxacin 750 mg once daily to 10 days of a standard dose of ciprofloxacin among patients with complicated UTI and pyelonephritis in a double-blind randomized trial including 619 patients. In a retrospective study, the treatment with a 7-day antibiotic course in men with urinary tract infection without evidence of complicating conditions was not associated with increased risk of recurrence. To our knowledge, there are no studies that evaluate the duration of antimicrobial therapy in patients with urinary tract infections caused by ESBL- EB. There is a tendency to perform longer antibiotic treatments in patients with infections by these microorganisms. In a retrospective study that evaluated treatment with ertapenem administered through outpatient parenteral antibiotic therapy (OPAT) in patients with urinary tract infections caused by ESBL-EB, the mean duration of antimicrobial treatment was 11.2 days. A prospective and observational study that included patients with complicated urinary tract infection caused by ESBL-producing *E*. *coli* evaluated the treatment with meropenem or imipenem for 14 days compared to 3 doses of 3 gr fosfomycin tromethanol every other day. Clinical and microbiological success in the carbapenem and fosfomycin groups was similar. Other retrospective study compared ertapenem to oral fosfomycin for outpatient treatment of ESBL-EB urinary tract infections. Total antibiotic duration was 10 days (IQR 7–12 days) in the fosfomycin group and 15 days (IQR 12–16 days) in the ertapenem group, without differences in infection-related hospital readmissions within 30 days. Our study is, to our knowledge, the first one conducted in patients with complicated urinary tract infection by ESBL-EB. No worse evolution was observed in patients treated with short antibiotic courses. We have not observed a worse evolution using different criteria to define complicated urinary tract infection or in patients more susceptible *a priori* to needing a longer antibiotic treatment, which could be the case of men or patients with urological abnormalities. There are, however, some limitations. First of all, it is a retrospective study, so there may be uncontrolled confounding factors despite the different statistical analysis performed. Secondly, it has been done in a single center, so the results may not be generalizable to different population groups. # Conclusions Patients with complicated urinary tract infections caused by ESBL-EB can be treated with a short course of antimicrobial therapy. Shorter duration of antibiotic treatment may lead to decreased risk of antibiotic resistance, fewer adverse effects, and lower costs. It is necessary to conduct randomized, controlled trial to establish the adequate duration of antibiotic treatment in these types of infections. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Prosocial behavior refers to a broad category of actions that benefit other people or the society that we lives in, such as helping, comforting, sharing, cooperation, philanthropy, and community service. A wide range of factors, including individual differences, situational variables, and outcome-related variables of prosocial behavior, have been found to influence prosocial behavior. Previous research has primarily focused on either the demographic and individual characteristics of the helper or situational factors; less attention has been paid to the roles of outcome-related variables. A number of theories and studies have suggested that the expected outcomes influence individual motivation and behavior. For example, the theory of planned behavior (TPB) states that personal evaluations, perceived social pressure, and perceived control should be considered as predictors of the intention to perform a given behavior. Later, researchers developed the standard TPB model by including anticipated affective consequences (i.e., anticipated regret and donation anxiety) in predicting behavioral intentions. And anticipating affective consequences was found to play an important role in the formation of intentions to donate blood. Other studies also showed that appeals for donation could differ due to their presentation format. However, most theories and studies only addressed the role of expected outcome in term of the helper. Little research has investigated the role of expected outcome relative to the beneficiary in predicting prosocial intention and behavior. In general, there are two types of expected outcomes for the beneficiary, obtaining positive outcomes and avoiding negative outcomes. Although both expected outcomes have equal positive valence for the beneficiary, their effects on helping willingness and behavior may be different. This hypothesis is derived from the theory and studies about loss aversion. ## Loss Aversion and Prosocial Outcomes Loss aversion, which was originally proposed by Kahneman and Tversky in their prospect theory, demonstrates that negative events are more intense in terms of their objective magnitude than positive events. In other words, people tend to weigh losses more heavily than gains with the same magnitude. Subsequently, researchers have distinguished and compared four types of outcomes that include losses, gains, non-losses, and non-gains. Losses and non-gains represent negative valence, whereas gains and non-losses represent positive valence. Previous studies on loss aversion have either primarily focused on the comparison between losses and gains or the two types of negative valence. The comparison between the two types of positive valences has received little attention. Tversky indicated that, in negotiations, eliminating losses (i.e., non-losses) were more effective than increasing gains. Following this perspective, some researchers have hypothesized that, when the principle of loss aversion applies to positive valences in the same manner as it applies to negative valences, non-losses should be evaluated as more positive than gains. Briefly, people are more likely to weigh non-losses heavily than gains with the same magnitude. However, this hypothesis was not supported when people made decision for themselves. How about when people make decision on behalf of others? Numerous studies have discovered evidence of loss aversion, suggesting that loss aversion reflects a long-held, fundamental phenomenon. These studies have primarily focused on risky decisions in the economic domain. However, much less is known about loss aversion in other areas. The present study aimed to examine whether the phenomenon of loss aversion occurs in the domain of prosocial behavior. With regard to prosocial behavior, as discussed above, two types of outcomes need to be distinguished: attaining positive outcomes and avoiding negative outcomes. Combining the two types of prosocial outcomes and the relevant concepts of loss aversion, prosocial gain is defined as a situation, in which helper A performs prosocial behavior in order to attain positive outcome for beneficiary B; prosocial non-loss is defined as a situation, in which helper A performs prosocial behavior in order to avoid negative outcome for beneficiary B. Based on the rule of loss aversion, the first hypothesis for the present study is that prosocial non-loss outcome would induce more help than prosocial gain outcome, which is defined as prosocial loss aversion. Based on the above discussion, the current study examined the phenomenon of prosocial loss aversion and explored its motivational mechanism. The motivational mechanism was addressed from the perspective of regulatory focus theory, examining the moderating effect of regulatory focus on the relationship between expected prosocial outcomes and prosocial performance. ## Regulatory Focus Regulatory focus theory proposes that regulatory focus is a principle of self- regulation, which interprets the motivational differences of behaviors. People’s regulatory focus can be categorized into two subsets of motivational orientation. The promotion focus emphasizes nurturance and reward, and the prevention focus emphasizes security and safety. The theory also suggests that people with distinct regulatory focus differ in their sensitivities to gains and losses. A promotion focus involves sensitivity to the presence and absence of positive outcomes. In contrast, the prevention focus is characterized by sensitivity to the presence and absence of negative outcomes. Theoretically, people who are prevention focused are more likely to exhibit greater prosocial loss aversion than those with promotion focus. Empirical evidence supported the links between the two types of regulatory focus and the levels of sensitivity to losses and gains. For example, Idson et al. found that losing was experienced more intensely by participants in a state of prevention focus than those in a state of promotion focus, while the opposite was found in the case of winning. Although the above studies indicate that individuals’ promotion and prevention focus are related to different levels of sensitivity to losses and gains respectively, no prior research has examined how regulatory focus may influence loss aversion in prosocial contexts. Prosocial behavior is different from behavior motivated by self-interests, referring to action that benefits others. Both gains and non-losses directed at others are prosocial outcomes. Polman suggested that loss aversion was reduced when promotion-focused people made choices for others compared to making choices for themselves, whereas prevention-focused people showed the same loss aversion in both circumstances. The study of Polman suggested that the effect of regulatory focus on loss aversion directed at others might be different. A clear distinction has to be made between prosocial loss aversion and helpers’ regulatory focus, as the former is beneficiary-centered but the latter is helper-centered. Therefore, the second aim of the present study was to investigate the effect of the helpers’ regulatory focus on prosocial loss aversion. ## The Present Research The present research consists of two studies. In Study 1, we investigated the differences of prosocial performances in helping others between attaining positive outcomes (defined as prosocial gains, e.g., enhancing other’s access to clean water) and avoiding negative outcomes (defined as prosocial non-losses, e.g., protecting other from suffering from dirty water). We hypothesized that prosocial non-loss outcomes would induce higher levels of prosocial performance compared with prosocial gain outcomes (defined as prosocial loss aversion). Unlike existing work that merely focused on the effect of expected outcomes on prosocial behavior from the perspective of helpers, the present research aimed to examine the effects of two types of expected outcomes on prosocial performance from the perspective of beneficiaries. In addition, the present research investigated the effect of expected prosocial outcome not only in terms of prosocial willingness but also prosocial behavior. In Study 2, we shed light on the moderating effect of the helpers’ regulatory focus on the relationship between the expected prosocial outcome and prosocial performance. Based upon the theories and research about loss aversion and regulatory focus, we expected that promotion-focused individuals should help more when they anticipate gain as an outcome of their behavior; whereas prevention-focused individuals should help more when they anticipate non-loss as an outcome of their behavior. Although several studies have examined individual differences and situational factors, scant attention has been paid to the factors related to prosocial outcomes and possible interactions between the different factors. Therefore, examining the interaction between regulatory focus and expected prosocial outcomes in the current work allowed us not only to explore the motivational mechanism of prosocial loss aversion, but also to examine the interaction between individual differences and outcome-related factors on prosocial performance. Furthermore, a moderation by dispositional and situational regulatory focus was examined. # Study 1 In Study 1, we designed two mini-studies to investigate the phenomenon of prosocial loss aversion. In Study 1a, the participants were instructed to decide whether they preferred to help others attain positive outcomes or avoid negative outcomes. In Study 1b, participants were asked not only to decide whom they were more likely to help but also to rate the degree of willingness to help for each of the two expected prosocial outcomes. We expected that participants would display greater prosocial performance when they were asked to help others avoid negative outcomes than attain positive outcomes. The present study was approved by the ethical committee of the School of Psychology, Beijing Normal University. Participants provided written informed consent before the study. ## Study 1a Study 1a provided an initial investigation of the phenomenon of prosocial loss aversion. A survey was used to investigate individuals’ preferences of expected prosocial outcomes. ### Materials and methods Participants: Hundred-and-six sophomores (73 females and 33 males) of China Youth University of Political Studies in Beijing completed a survey. These students participated in the study for credit toward a course requirement. Procedure: All participants were provided with a sheet that indicated two victims were in a stricken area and had the same demands. The participants were told that they had the ability and opportunity to help victim A attain positive outcomes (e.g., enhancing victim A’s access to clean water) and to help victim B avoid negative outcomes (e.g., protecting victim B from suffering from dirty water). Then, they were asked which victim they would be more likely to help. The answer (help A or B) was recorded. The order of the presentation of the two expected prosocial outcomes was counterbalanced. The detailed scenario can be seen in. ### Results and discussion A chi-square test was conducted to examine the differences in the number of participants choosing for two prosocial outcomes. The result showed that participants preferred the prosocial non-loss outcome (*N* = 67) compared to prosocial gain outcome (*N* = 39), *χ*² (1) = 7.40, *p* =.007, suggesting a greater tendency toward prosocial non-losses than prosocial gains. This result suggests individuals tend to help others avoid negative outcomes rather than to attain positive outcomes. Given that this survey only investigated the preference for expected prosocial outcomes, Study 1b was performed to further examine the differences between the degrees of individuals’ willingness to help for prosocial gain and prosocial non-loss outcomes. ## Study 1b In Study 1b, both the preference for and the exact degrees of prosocial willingness for prosocial gain and prosocial non-loss outcomes were measured and compared. ### Materials and methods Participants: The participants were an additional sample of 60 sophomore students (38 females and 22 males) recruited from China Youth University of Political Studies in Beijing, China. They participated in the experiment for credit of a psychology course. Procedure: The procedure and materials provided to the participants were similar to those used in Study 1a with the exception that participants were also required to indicate the degree to which they wanted to help for each of the two expected prosocial outcomes on a bipolar scale (see). The larger number indicated that the participants would be more likely to help, and the victims would get more help. The positions of the two expected prosocial outcomes on the bipolar scale were counterbalanced. The detailed scenario can be seen in. ### Results and discussion A chi-square test revealed that the difference between the number of each of two choices was statistically significant, *χ*² (1) = 4.27, *p* =.039. Thirty-eight participants indicated that they would like to help the victim avoid negative outcomes, and 22 participants indicated that they would like to help the victim attain positive outcomes. To further examine the difference in the degree of prosocial preference, a paired-samples *t*-test was performed on the willingness to help for the prosocial gain and non-loss outcomes. Among the 60 participants in the experiment, 55 participants responded to both outcomes. The results revealed that the participants expressed greater willingness to help the victim avoid negative outcomes (*M* = 5.73, *SD* = 1.50) than attain positive outcomes (*M* = 5.09, *SD* = 1.49), *t* (54) = 2.38, *p* =.021, *d*_*z* = 0.32. These findings suggested that prosocial non-losses evoked greater prosocial willingness than prosocial gains did. In other words, loss aversion occurred in the prosocial domain, which is in line with previous research results. Although the main findings were in line with our predictions, one potential limitation was that both Studies 1a and 1b were conducted in a classroom context, and the situation described in the materials was unfamiliar to the participants. To address the limitation, prosocial performance measured in Study 2 was conducted in a more realistic prosocial situation. In addition, the present order of two expected prosocial outcomes was only controlled, but the effect was not reported in Study 1. This order effect would be tested in Study 2. # Study 2 Although differences were found between participants helping others avoid negative outcomes and helping others attain positive outcomes, the reasons for these differences have not been fully understood. Despite a growing body of research focusing on the motivational mechanism of prosocial behavior, little research, if any, has examined motivational differences between the two types of prosocial outcomes. To fill this gap, Study 2 aimed to explore whether the differences in prosocial performances between prosocial gain and prosocial non- loss outcomes can be explained by regulatory focus theory. According to the regulatory focus theory and related research, we hypothesized that potential helpers with a prevention focus would be more likely to help others avoid negative outcomes than help others achieve positive outcomes. This would magnify the degree of loss aversion. In contrast, it is assumed that potential helpers with a promotion focus would be more prone to help others achieve positive outcomes than avoid negative outcomes, which might cause a reduction or even a reversal of loss aversion. In Study 2, the focus on promotion versus prevention was manipulated either as a dispositional individual difference or a situational difference by priming. Two mini-studies were designed to examine the moderating effects of regulatory focus. In Study 2a, the regulatory focus at the dispositional level was measured by the regulatory focus scale. The goal of Study 2a was to investigate individual differences in prosocial loss aversion in terms of dispositional regulatory focus, and to preliminarily examine the moderating effect of regulatory focus on the relationship between expected prosocial outcome and prosocial behavior. In Study 2b, the regulatory focus at the situational level was primed by helping a mouse out of a maze. The goal of Study 2b was to investigate the motivation of prosocial loss aversion by directly testing the moderating effect of situational regulatory focus. Meanwhile, to address the limitation about material, prosocial performance measured in Study 2 was conducted in a more realistic prosocial situation. The fictitious victims were not sufferers in a stricken area but were similarly aged peers of the participants. The prosocial gains were helping a stranger attain money, and the prosocial non-losses were helping a stranger avoid losing money. The participants were instructed to indicate that the extent to which they were willing to help for each of the two expected prosocial outcomes. Furthermore, two qualifying tests measured the participants’ real prosocial behaviors for each of the two expected prosocial outcomes. The participants were told that they would be able to help the strangers in need only on the condition of passing the qualifying test. The underlying logic of the qualifying test was that if the participants truly wanted to help, they would do their best to pass the qualifying test, thereby performing at a high level. Thus, the performances in two qualifying tests could be used as indices of their prosocial behaviors for the two expected prosocial outcomes. Additionally, the difference between their performances in two qualifying tests served as an additional index of their prosocial loss aversion. ## Study 2a To indirectly explore the motivational mechanism that underlies the occurrence of prosocial loss aversion, Study 2a examined the interaction effect between individual differences and outcome-related factors on prosocial performance from the perspective of the dispositional regulatory focus. ### Materials and methods Participants: Sixty-two undergraduate students (45 males and 17 females, *M*_age = 20.68, *SD* = 1.08) were recruited from Tianjin Chengjian University in China. Each participant received a gift for participating after the experiment. The experiment was approved by the Ethical Committee of the School of Psychology, Beijing Normal University. Consent forms were given and signed by the participants prior to the experiment. Procedure and Measures: Participants were randomly assigned to six groups with 10–12 members. Upon arriving at the laboratory, the participants were asked to independently complete a battery of paper-pencil questionnaires concerning regulatory focus and prosocial loss aversion for prosocial willingness and behavior. *Dispositional regulatory focus*. The Regulatory Focus Scale (RFS, 33-items,) was used to measure dispositional regulatory focus. The participants rated the items on a scale from 1 (*definitely untrue for me*) to 7 (*definitely true for me*). The RFS is divided into two subscales: Promotion focus (17 items) and Prevention focus (16 items). The promotion subscale (Cronbach’s *α* =.67) consists of statements that reflect a focus on achieving positive things (e.g., “It is very important to me to develop myself further and to improve myself”). The prevention subscale (Cronbach’s *α* =.78) consists of statements that reflect a focus on avoiding negative things (e.g., “I often think about how I can avoid failures in my life”). This scale has been shown to be reliable and valid in previous research examining individual regulatory focus orientations. *Prosocial loss aversion for prosocial willingness and behavior*. The participants were provided with a sheet that contained information about a prosocial scenario. In the scenario, two strangers had not received all of the money owed them from another experiment because they failed to complete all of the experimental trials. If participants wanted to help the strangers, they needed to complete trials for the strangers. The independent variable was two types of expected prosocial outcomes: prosocial gains (i.e., helping a stranger attain the remaining money) and prosocial non-losses (i.e., helping a stranger avoid losing money they had already received due to the uncompleted trials). First, participants were asked to choose the number of trials, from 0 to 10, that they would be willing to complete for each of the two strangers separately. After finishing these choices, the participants were told that they would have the opportunity to help the strangers only if they passed the qualifying tests. If the participants truly wanted to help, they would do their best in the qualifying tests, and their performances in these tests would represent their prosocial behavior. The qualifying tests were adapted from the Number Cancellation and Letter Cancellation tests which were used to measure selective attentional abilities. The Number Cancellation Test was one page in length with 25 rows and 40 numbers (from 0 to 9) in each row. The Letter Cancellation Test was one page in length with 25 rows and 40 letters (from A to I) in each row. All the numbers or the letters were randomly interspersed. The participants were instructed to cross out all the target numbers and letters respectively. Specifically, for the target number, the participants were asked to choose the number between 3 and 7 or 7 and 3. For the target letter, the participants were asked to choose the letter between D and H or H and D. The target numbers and target letters were equal in number. Participants had 4 minutes for two tests and had the freedom to allocate time to each test. Each qualifying test corresponded to one type of expected prosocial outcomes. For half of the participants, the qualifying test under the condition of the prosocial gain outcome was the Number Cancellation Test; the qualifying test under the condition of the prosocial non-loss outcome was the Letter Cancellation Test. While for the other half of the participants, the qualifying test under the condition of the prosocial gain outcome was the Letter Cancellation Test; the qualifying test under the condition of the prosocial non- loss outcome was the Number Cancellation Test. The order and correspondence of the two qualifying tests with two expected prosocial outcomes were counterbalanced. The dependent measures were the numbers of the trials selected for two expected prosocial outcomes and the performances in the two cancellation tests. After completing the experiments, the participants were thanked and debriefed by the experimenter. When asked, none of them reported any hypotheses relevant to the true purpose of the experiment. The detailed scenario can be seen in. ### Results and discussion Basic descriptive data on regulatory focus, prosocial willingness, and prosocial behavior is presented in. Two independent *t* tests of prosocial loss aversion on order and counterbalance were conducted. The effects of order (*p* =.756) and counterbalance (*p* =.095) were not significant. Two hierarchical regression analyses of prosocial willingness on expected prosocial outcome (prosocial gains vs. prosocial non-losses) and regulatory focus (promotion vs. prevention) were conducted. The type of expected prosocial outcomes was coded into a dummy variable (“0” for prosocial gains and “1” for prosocial non-losses). The continuous measures of promotion and prevention were centered prior to computing the interactions and used as the main effects respectively. Model 1 tested the effects of promotion focus, the type of expected prosocial outcomes, and their interaction on prosocial willingness. In the hierarchical regression, centered prevention focus as a controlled variable was entered in the first step, centered promotion focus and the type of expected prosocial outcomes were included in the second step, and the interaction between centered promotion focus and the type of expected prosocial outcomes were included in the third step. Model 2 tested the effects of prevention focus, the type of expected prosocial outcomes, and their interaction on prosocial willingness. In the hierarchical regression, centered promotion focus as a controlled variable was entered in the first step, centered prevention focus and the type of expected prosocial outcomes were included in the second step, and the interaction between centered prevention focus and the type of expected prosocial outcomes were included in the third step. The analyses revealed two main effects of the types of expected prosocial outcome. No significant main effect of promotion focus, prevention focus, or interaction was found. Two similar separated hierarchical regression analyses concerning prosocial behavior were conducted. Similarly, regressions on promotion focus and on prevention focus were conducted separately. The results revealed two main effects of expected prosocial outcomes. No significant main effect of promotion focus, prevention focus, or interaction was found. No other significant effect was found. Because dummy variable 1 represented prosocial non-losses, the main effects of type of expected prosocial outcome indicated that the more prosocial non-losses, the more prosocial willingness and behavior. Therefore, both analyses of prosocial willingness and behavior exhibited a significant main effect of expected prosocial outcome, indicating that the prosocial performances for the prosocial non-loss outcome were greater than those for the prosocial gain outcome. Furthermore, this experiment indicated that only the expected prosocial outcome affected prosocial willingness and prosocial behavior, whereas individuals’ dispositional regulatory focus did not affect the relationship between expected prosocial outcome and prosocial performance, which was inconsistent with our hypothesis. The non-significant moderating effect of dispositional regulatory focus may due to the non- significant difference between dispositional promotion and prevention focuses. The sample size was not large enough to distinguish between participants who reported high and low promotion and prevention focus orientations, respectively. Study 2a made a primary examination of the moderated effect of dispositional regulatory focus. Study 2b distinguished two regulatory focuses by situational priming and tested the hypothesis about the moderating effect of situational regulatory focus. ## Study 2b Following Study 2a, in order to directly identify the motivational mechanism that underlies the occurrence of prosocial loss aversion, Study 2b manipulated situational regulatory focus via experimental priming and examined the interaction between situational regulatory focus and expected prosocial outcomes. ### Materials and methods Participants: One hundred and six undergraduates (65 males; *M*_age = 20.86, *SD* = 1.36) from Tianjin Chengjian University in China participated in this experiment. The participants were compensated with a gift for their participation. The present experiment was approved by the Ethical Committee of the School of Psychology, Beijing Normal University. Consent forms were presented and signed by the participants before the experiment. Procedure and Measures: Upon arriving at the laboratory, the participants were instructed to independently complete a series of ostensibly unrelated essay tasks. The participants were randomly assigned into three conditions. Two groups were assigned to one pair of mazes that have previously been demonstrated to elicit either a promotion or prevention focus. In the promotion-focused-priming condition, the participants were instructed to help the mouse in the maze move toward the cheese. In the prevention-focused-priming condition, the participants were instructed to help the mouse in the maze escape from a hawk. After completing one of the mazes, each participant received the sheet used in Study 2a to measure their prosocial willingness and behavior. The participants in non- priming condition only finished the measure of prosocial willingness and behavior. ### Results and discussion The participants’ responses regarding prosocial willingness were submitted to a 2 (expected prosocial outcome: prosocial gains vs. prosocial non-losses) × 3 (regulatory focus: promotion vs. prevention vs. non-priming) mixed factorial design ANOVA, which yielded no significant effects, *F*s \< 1.71, *p*s \>.05. An analogous 2 × 3 repeated measures ANOVA on prosocial behavior revealed a significant main effect of expected prosocial outcome (*F* (1, 103) = 14.93, *p* \<.001, *η*² =.13) and no significant main effect of regulatory focus (*p* =.251). The interaction between expected prosocial outcome and regulatory focus was not significant, *p* =.063. Based on the strong hypotheses, simple effect analyses were conducted and revealed that participants whose prevention focus was primed showed a better performance for prosocial non-loss outcome than prosocial gain outcome. Similarly, participants in the non-priming condition also performed better for the prosocial non-loss outcome. However, no significant difference was found between the performances of prosocial non-loss and gain outcomes for participants whose promotion focus was primed. These results are presented in. Meanwhile, the prosocial loss aversion in the promotion-focus-priming condition was significantly lower than those in the prevention-focus-priming and non-priming conditions. And the difference in prosocial loss aversion between the prevention-focus-priming and non-priming conditions was insignificant. The findings of Study 2b only revealed that participants in the promotion-focus- priming condition tend to display lower prosocial loss aversion than that in prevention-focus-priming condition. A lack of significant interaction between expected prosocial outcome and situational regulatory focus might due to the weak priming manipulation of regulatory focus, especially for prevention focus. Another reason for the insignificant interaction might be the small sample size. # General Discussion The purpose of the present research was to examine the effect of expected prosocial outcomes on prosocial performance. Four experiments provided evidence that people were more likely to help others avoid negative outcomes than attain positive outcomes, which supported the occurrence of prosocial loss aversion. Additionally, the interaction effect of expected prosocial outcomes and regulatory focus on prosocial performance was examined. We only found that the prosocial loss aversion in the promotion-focus-priming condition was significantly lower than in the prevention-focus-priming and non-priming conditions. And the difference in prosocial loss aversion between the prevention-focus-priming condition and non-priming conditions was insignificant. This finding that non-loss outcomes induced more help than gain outcomes is consistent with the principle of loss aversion that losses exert greater influences on choice and predicted feelings about an outcome than do gains of the same magnitude. The current research shed light on loss aversion framed in the prosocial context in which prosocial loss aversion was reflected by the fact that prosocial non-losses induced greater prosocial performances than prosocial gains did. However, strictly speaking, the outcomes of events can be classified into losses, gains, non-losses, and non-gains. As mentioned before, gains and non- losses represent positive valences, while losses and non-gains represent negative valences. In previous studies of loss aversion, particular attention has been paid to gain-loss comparisons. Losses have been found to exert twice as much influence on decisions as equivalent gains. Kahneman and Miller suggested that the natural comparison should be a set of events with the same valence. Therefore, it is important to distinguish gains from non-losses and losses from non-gains. For example, Liberman et al. proposed that, according to the principle of loss aversion, losses should be perceived as more intensely negative than non-gains, while non-losses should be perceived as more positive than gains. However, their studies confirmed the first prediction, but failed to corroborate the second one. The current study showed that an outcome that resulted in a non-loss induced more help than an outcome that resulted in a gain. One reason for the discrepancies between the findings of the previous and the current research might be attributable to the differences in the experimental situations. Unlike the present research conducted in prosocial situations, previous studies were mainly conducted in the economic domain. Additionally, loss aversion examined in previous studies merely applied to the participants themselves, whereas in the present research, the decisions applied to others to some extent. Prosocial behavior refers to behavior that benefits other people or society. Both prosocial gains and prosocial non-losses incur benefits directed at other people. Although some studies have shown that loss aversion is reduced when people make choices for others, compare to making choices for themselves, others have reported that cognitive biases are stronger when decision makers choose on behalf of others than they choose on their own behalves. As decisions in prosocial scenarios are more likely to be made on behalf of others, prosocial scenarios induce significant prosocial loss aversion. The other purpose of the present studies was to examine the interaction effect of expected prosocial outcome and regulatory focus on prosocial performance. The result partly supported the hypothesis: prosocial loss aversion in the promotion-focus-primed condition was significantly lower than in the prevention- focus-priming and non-priming condition. These findings are supported by the regulatory focus theory and the regulatory fit theory. In addition, the present studies also showed that neither dispositional nor situational regulatory focus significantly interacted with expected prosocial outcome on prosocial performance. These results were not consistent with the study of Fransen et al.. Their study reported that individuals in a state of prevention focus donated more money when the goals of a charity were described as preventing a negative outcome. And individuals in a state of promotion focus donated more money when the goals were presented as encouraging positive outcomes. The inconsistent might due to the fact that the difference between promotion and prevention focus in the current study was not large enough. For the dispositional regulatory focus, the scores of participants’ promotion focus showed in Study 2a were similar to those of prevention focus. For the situational regulatory focus, the priming might be weak. Alternatively, the sample size in either study was not large enough. Despite the valuable findings of the current research, several limitations should be noted. First, as mentioned before, the outcomes of events can be classified into losses, gains, non-losses, and non-gains. Further research on prosocial loss aversion is necessary to include both positive-valence helping outcomes (gains and non-losses) and negative-valence non-helping outcomes (losses and non-gains) to allow for a clearer examination. Second, although we consistently observed prosocial loss aversion across four experiments, more diverse prosocial domains with different risk should be included to examine the phenomenon of prosocial loss aversion in future studies. Third, more detailed information should be considered. For example, the present order of two expected prosocial outcomes in Study 1 should be recorded and analyzed. The balance between the speed and the accuracy of the cancellation task, and the relationship between dispositional regulatory focus and the balance should be considered in future study. Fourth, for the non-significant dispositional regulatory focus on prosocial loss aversion, future studies may include large enough sample size to distinguish apparent promotion and prevention focus individuals. For the non-significant situational regulatory focus on prosocial loss aversion, future studies may use stronger priming condition and add operational check after priming. Finally, the current study only focused on the effect of the beneficiaries’ outcomes. However, prosocial behavior has impacts on both helpers and beneficiaries. The expected outcome of helpers’ prosocial behavior also affected their prosocial performance. The interaction between expected outcomes to helpers and beneficiaries and the relationship between helpers and beneficiaries (in- vs. out-group) is worthwhile considering in a follow-up study. # Conclusions The present studies explored how expected prosocial outcomes affected people’s prosocial performances. The findings indicated that the phenomenon of loss aversion also occurred in the prosocial domain, in which expected prosocial non- loss outcomes induced higher levels of prosocial performance than expected prosocial gain outcomes did. The present studies also showed that prosocial loss aversion in the prevention-focus-priming condition was significantly higher than in promotion-focus-priming and non-priming condition. Prosocial loss aversion might be lessened when promotion focus was primed, whereas prosocial loss aversion was not significantly increased when prevention focus was primed. The non-significant interaction between expected prosocial outcomes and regulatory focus might due to the shortcomings of the current studies. # Supporting Information The authors would like to thank Longfeng Li, Meng Zhang, and Bixi, Zhang for their help editing the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** FX YC. **Data curation:** FX YC. **Formal analysis:** FX ZZ YC. **Funding acquisition:** YC. **Investigation:** FX ZZ YC YL. **Methodology:** FX YC ZX. **Project administration:** YC. **Resources:** FX. **Supervision:** YC. **Validation:** YC. **Visualization:** FX. **Writing – original draft:** FX ZZ YC. **Writing – review & editing:** HZ ZX.

# Introduction Muscle atrophy in critically ill children may influence illness progression and functional recovery, yet it remains understudied. Only three prospective studies to date have examined intensive care-acquired muscle weakness (ICU-AW) or atrophy (ICU-AA) in children. Banwell et al. used neuromuscular exam to identify a 1.7% incidence of weakness among all children admitted to the PICU for \>24 hours (n = 830). Recently, however, Valla et al. used ultrasound to measure quadriceps femoris in critically ill children receiving invasive mechanical ventilation and identified 59% (10/17) with ≥10% decrease in thickness within 5 days of endotracheal intubation. Glau et al. found that diaphragm thickness in children receiving mechanical ventilation decreases on average by 3.4% per day. The two latter studies concur with robust adult data which indicate that ICU-AA, including diaphragm atrophy, and ICU-AW affect 30–70% of patients. A comprehensive assessment of muscle wasting and identification of potential risk factors during pediatric critical illness remains lacking. In particular, critically ill adults receiving mechanical ventilation exhibit diaphragm thinning 24 hours after intubation. Diaphragm thickness may decrease by 3–6% per 24 hours of mechanical ventilation. Diaphragmatic and other skeletal muscle atrophy and weakness are associated with difficulty weaning from mechanical ventilation, prolonged ICU stay, worse functional outcomes and increased mortality. Specific risk factors identified in critically ill patients include prolonged mechanical ventilation, neuromuscular blockade, sepsis, multi- organ dysfunction syndrome (MODS), and use of glucocorticoids. We examined how muscle thickness and electrical properties change over time in mechanically-ventilated critically ill children. Both muscle thickness and electrical impedance correlate with muscle strength, and thus may predict physical disability and need for rehabilitation after PICU discharge. We used ultrasound to serially measure thickness of the diaphragm, biceps brachii/brachialis, quadriceps femoris and tibialis anterior. These muscles are easily examined with bedside ultrasound equipment routinely available for vascular access. We also performed serial electrical impedance myography (EIM) on limb muscles to examine EIM utility in the PICU. Finally, we examined whether demographic and clinical variables thought to correlate with atrophy in adults influence muscle loss in critically ill children. # Methods ## Study design We performed a single-center, prospective, cohort study in a 35-bed medical- surgical PICU at a free-standing, academic, tertiary care children’s hospital (St. Louis Children’s Hospital (SLCH), St. Louis, MO). Washington University School of Medicine (WUSM) physician faculty staffs the SLCH PICU. The WUSM Institutional Review Board approved the study (IRB \#201504013). We recruited patients between June 2015 and May 2016. We enrolled 34 children aged 1 week—18 years who satisfied the following enrollment criteria: 1) normal neurologic development and gross motor function if \<15 months old or independently ambulatory before hospitalization if \>15 months old; 2) no known neuromuscular disease; 3) respiratory failure requiring endotracheal intubation; 4) intubated \<72 hrs before enrollment; and 5) expected to remain intubated \>48 hrs. We obtained informed consent from parents and/or guardians at enrollment and, whenever feasible, patients’ assent. The PICU attending physician directed all clinical interventions, including sedation management. Consistent with the current standard of care, continuous neuromuscular blockade is used sparingly and for the shortest duration possible, as dictated by the patient’s clinical status. Sedation protocol for patients on mechanical ventilation generally involves a continuous infusion of fentanyl with sequential additions of either midazolam or dexmedetomidine infusions. Sedation is titrated to a State Behavioral Score (SBS); SBS goal is specified on daily morning rounds and depends on the patient’s clinical status and clinical trajectory (e.g. worsening, improving, titrating towards extubation). The objective goal is to minimize sedative use and facilitate recovery without impeding ongoing care. ## Muscle measurements We measured muscle thickness with bedside ultrasound (US) (SonoSite Edge II, FUJIFILM SonoSite Inc., Bothell, WA) using a 13–6 MHz 6 cm linear probe (L25). We also measured electrical impedance using a commercial EIM device (Skulpt AIM, Boston, MA). Thickness was measured in the right diaphragm, biceps brachii/brachialis, quadriceps femoris, and tibialis anterior. For EIM, we examined biceps, quadriceps, and tibialis in children \>1 yr of age. In infants \<1 yr old, EIM device size limited observations to quadriceps only. EIM could not be performed on the diaphragm. We obtained US and EIM measurements at enrollment and at 5–8 day intervals until PICU discharge. Observers were blinded to prior measurements. For US reproducibility, we temporarily marked and regularly reinforced the following skin landmarks: diaphragm: lower ribs along anterior axillary line at end expiration; biceps: 2/3 from acromion to antecubital crease; quadriceps: 1/2 from anterior superior iliac spine to patella’s superior edge; tibialis: 1/3 from patella’s inferior edge to lateral malleolus. We measured thickness with limbs at rest without active flexion/resistance. For biceps, quadriceps and tibialis, the US probe was oriented strictly perpendicular to the skin to minimize error. Cross-sectional images were obtained using thick US gel layer without skin compression. Diaphragm thickness was measured at passive exhalation end with US probe oriented longitudinally along the ribs. We aimed for a simple and rapid assessment, which, in clinical practice, would permit a minimally trained PICU practitioner to obtain reliable measurements without interfering with ongoing care. Consequently, we did not measure diaphragm thickness at end- inspiration, precluding calculation of the thickening fraction. Measurements were obtained in triplicate and averaged for analyses. Intra-rater reproducibility coefficient (*κ*) was 0.99 for biceps, quadriceps and tibialis, and 0.8 for diaphragm, supporting recent suggestions by our group and others that serial US may reliably assess muscle properties in children. For EIM, we used a commercial handheld device (Skulpt AIM, Boston, MA). Per manufacturer’s instructions, EIM device was applied to moistened skin overlying the target muscle with all surface electrodes contacting the skin. As with thickness measurements, we measured EIM while avoiding active flexion/resistance. The EIM device uses proprietary algorithms to report two scores: muscle “quality” and “fat percentage”. Higher muscle “quality” score and lower “fat percentage” score are considered desirable outcomes. ## Outcomes Primary outcome measure was percent change in thickness between the 1^st and 2^nd US exams. For patients with \>2 US exams, the measurement with the greatest difference from baseline (either increase or decrease) was used. Secondary outcomes were changes in EIM-derived scores. Change in thickness was expressed both in absolute terms and as percent change, because muscle size in children changes substantially during development. In addition, we calculated percent change/day in order to normalize the data for different time intervals between US exams; this normalization does not imply that the change occurs linearly over time. EIM-derived scores–muscle “quality” and “fat percentage”–were compared using change in absolute values. ## Clinical covariates Demographic and clinical variables that may influence incidence and severity of muscle loss in critically ill children were defined *a priori* and abstracted from the electronic medical record. Non-modifiable variables included age (years), age group (infants \<1 yr old *vs* children \>1 yr old), PRISM score, SOFA scores on days 1 and 7, traumatic brain injury (yes/no), sepsis as recorded in the medical record by the treating physician (yes/no), PICU and hospital length of stay (days), ventilator-free days at 28 days, body mass index (BMI), Down’s Syndrome (yes/no) and extubation failure. Potentially modifiable variables included glucocorticoids (total dose expressed as hydrocortisone equivalent/kg), neuromuscular blockade (none, single dose, or continuous), hyperglycemia (serum glucose \>200 mg/dL), and vasopressors/inotropes (yes/no). We also calculated the Pediatric Risk of Mortality (PRISM) III and daily Sequential Organ Failure Assessment (SOFA) scores. For each organ system, a SOFA organ-specific score \>2 defined failure. We defined multi-organ failure as failure of ≥2 systems. ## Statistical analyses Statistical analyses were performed with SAS Statistics, v 9.4 (SAS Institute Inc., Cary, NC). Kolmogorov-Smirnov test was used to assess normality for continuous variables. Normally distributed variables were compared to using paired Student’s *t*-test. Non-normally distributed variables were compared using Wilcoxon Signed-Rank Test (WSRT). Bivariate relationships between patient characteristics (including risk factors) and continuous outcome variables were examined using simple linear regression. Violations of linear regression assumptions (linearity, multivariate normality, no multicollinearity, no auto- correlation and homoscedasticity) were verified before interpretation of the model. Statistical significance level was set at 0.05. Variables with a *p value* ≤0.20 in bivariate analyses were considered for multivariate model building. In the final multivariate model, *p* ≤0.05 was considered statistically significant. We defined muscle atrophy *a priori* as ≥10% decrease in thickness. We chose this cut-off because in prior studies, critically ill adults lost \~10% of their muscle mass within one week of hospitalization. Therefore, we additionally examined presence or absence of atrophy in each muscle group as a dichotomous outcome variable. We used bivariate and multivariate logistic regression to assess the relationship between patient characteristics and muscle atrophy. Similar to thickness analyses, risk factors for atrophy with a *p value* ≤0.20 on bivariate analyses were entered into multivariate logistic regression. Results are reported as mean (95% CI) for normally distributed variables and as median (interquartile range (IQR)) for non-normally distributed variables. # Results ## Participants We enrolled 34 patients between June 2015 and May 2016. Thirty patients had ≥2 assessments a median of 6 (IQR 6–7) days apart. Six patients had three assessments. Four patients were transferred (n = 3) or discharged (n = 1) before the second assessment and were excluded from final analyses. shows pertinent patient characteristics. The mean age was 5.42 years. Twelve (40%) were infants \<1 yr old, and 18 (60%) were children \>1 yr old. The most common primary diagnoses were airway/respiratory (33%), central nervous system (24%) and trauma (15%). Mean PRISM III score was 14.72, and mean day 1 SOFA score was 7.85. At least two US examinations of diaphragm, biceps, quadriceps and tibialis were completed in 100%, 83%, 100% and 97% of patients, respectively. The two most significant factors limiting assessment were 1) presence of peripherally- inserted central catheters (PICCs) which limited access to biceps landmarks and 2) liberation from mechanical ventilation and associated sedation in infants prior to second assessment, which precluded passive extension of the extremities. Two EIM assessments of biceps, quadriceps and tibialis were completed in 50%, 70% and 53% of patients, respectively. The main factor limiting EIM device use was patient size. Among infants \<1 yr old, we successfully conducted quadriceps EIM in only 4 subjects. Among children \>1 yr old, however, we completed two EIM assessments of biceps, quadriceps and tibialis in 83%, 89% and 83% of subjects, respectively. Therefore, we limited EIM analyses to children \>1 yr old. ## Change in muscle thickness In the entire cohort, diaphragm thickness decreased 11.1% (\[95%CI, -19.7% to -2.52%\]; p = 0.013) or by 2.2%/day. Sixteen subjects had their 2^nd diaphragm US a median of 1 day after extubation \[IQR 1–4\] and 14 –either prior to or on the day of extubation \[median 3 days prior, IQR 0–7\]. The timing of the 2^nd US exam relative to extubation did not affect the decrease in diaphragm thickness (WRST, p = 0.44). Quadriceps thickness decreased 8.62% (\[95%CI, -15.7% to -1.54%\]; p = 0.0187) or by 1.5%/day. We did not detect a decrease in either biceps (mean = -1.71%; \[95%CI, -8.15% to 4.73%\]; p = 0.5885) or tibialis (mean = 0.52%; \[95%CI, -5.81% to 3.40%\]; p = 0.4762) thickness ( and). We defined muscle atrophy *a priori* as ≥10% decrease in thickness. Approximately half of all children examined experienced diaphragm (47%, 14/30) or quadriceps (53%, 16/30) atrophy. Among the entire cohort, 83% (25/30) experienced atrophy in ≥1 muscle group and 47% (14/30) in ≥2 muscle groups. Among children \>1 yr old with all 4 muscle groups measured at least twice (n = 16), 94% (15/16) experienced atrophy in ≥1 muscle group and 56% (9/16)—in ≥2 muscle groups. We conducted bivariate and multivariate analyses to identify variables associated with muscle loss. Bivariate analyses revealed that increasing age and being a child (*vs*. infant) predicted greater decrease in the thickness of biceps, quadriceps and tibialis but not diaphragm. Indeed, children \> 1 yr of age on average lost 6.8% of their biceps (95%CI \[-14%, 0.8%\]), 6.5% of their tibialis (95%CI \[-11%, -2%\]), and 15% of their quadriceps (95% CI \[-22%, -7.2%\]) thickness. Increasing BMI correlated with greater loss in biceps and quadriceps muscles. Patients with TBI (n = 4) experienced significantly greater decreases in biceps and tibialis thickness than either children or infants without TBI. On average, children with TBI lost 21% and 17% of biceps and tibialis thickness, respectively. In comparison, children without TBI lost 3.7% and 3.8% of biceps and tibialis thickness, respectively. Predictor variables with p ≤0.2 on bivariate analysis (bold) were included in the multivariate model. For quadriceps, no statistically significant associations were found on multivariate analysis. For biceps, step-wise multivariate linear regression revealed that increasing age and presence of TBI were associated with greater decrease in thickness. Adjusting for TBI, each additional year of age was associated with an additional 1.46% decrease in biceps thickness. Adjusting for age, presence of TBI was associated with an additional 18.1% decrease in biceps thickness. In addition, hospital LOS correlated positively with biceps thickness. For tibialis, increasing age also predicted a greater decrease in thickness. Although the effect of TBI on tibialis thickness was not statistically significant, there was an interaction between TBI and age. Adjusting for TBI, each additional year of age was associated with an additional 1.20% decrease in tibialis thickness. Adjusting for age, presence of TBI was associated with an additional 18.9% decrease in tibialis thickness. ## Change in EIM parameters Muscle “quality” decreased in all muscle groups examined (biceps: - 6.5 \[IQR -19 to -2.0\], p = 0.040; quadriceps: - 11 \[95%CI, -16 to -5.4\], p = 0.001; tibialis: -8.1 \[95%CI, -12 to -4.1\], p = 0.001; all tests WRST;). Fat percentage increased by \~3–5% in all muscle groups examined (biceps: +5.2 \[IQR 2.2 to 7.0\], n = 14, p = 0.036; quadriceps: +3.5 \[95%CI, 0.0 to 7.1\], n = 17, p = 0.052; tibialis: +2.8 \[IQR -0.3 to 8.8\], n = 14, p = 0.009; all tests WRST;). We also correlated change in EIM parameters and change in thickness. For biceps, increased fat percentage and decreased muscle “quality” correlated with greater decrease in thickness (Spearman ρ = - 0.63, p = 0.017 and Spearman ρ = 0.44, p = 0.11, respectively). We did not observe an association between EIM parameters and thickness for quadriceps or tibialis. Similar to its effect on thickness, TBI was associated with worsening of EIM parameters. On bivariate analyses, only TBI was associated with significant decrease in quadriceps muscle “quality” (Kruskal-Wallis χ² = 4.40, DF = 1, p = 0.036). # Discussion The current study used ultrasound and EIM to measure changes in muscle over time in critically ill children with respiratory failure and correlated the extent of muscle loss with potential risk factors. Results indicate that muscle atrophy (≥10% decrease in thickness) in this population is common and rapid, occurring within 5–7 days. Eighty three percent of children experienced atrophy in ≥1 muscle group, and 47% − in ≥2 muscle groups. Importantly, in our study population of critically ill children receiving invasive mechanical ventilation for \>2 days, diaphragm atrophy affected almost half. We also found that increasing age and presence of TBI appear to increase muscle loss severity. Study limitations include single-center nature of the study and relatively small sample size, which may hamper generalization to other centers. Additionally, infants in our PICU do not represent the entire population of infants receiving critical care, as pre-term and full-term infants in the neonatal ICU and infants with congenital heart disease in the cardiac ICU were not included. The SLCH PICU, however, is a tertiary care referral center. Patient population mix and acuity are representative of other large, multidisciplinary PICU’s. One of the inclusion criteria–the requirement that patients were considered likely to remain intubated for at least 48 hours after enrollment–led to selection of a sicker patient cohort than the general PICU population. Consistent with selecting a subset of sicker patients, the median duration of intubation in the current sample is 5 days longer than that in our PICU population as a whole (7.4 *vs*. 2.4 days). Another limitation is lack of long-term follow-up, preventing characterization of functional deficits and/or recovery. The study was designed to investigate the incidence of muscle wasting in critically ill children and to identify potential risk factors. These data are required for larger observational or interventional studies to correlate muscle loss with long-term outcomes. Finally, we used a commercial rather than a research EIM device in an attempt to provide an easy bedside tool for the clinician. The commercial device limited available information to derived values, and size discrepancy between the EIM device and infants’ limbs precluded EIM in infants. Although US data suggest that loss of limb muscle mass may not be a major feature of muscle wasting in critically-ill infants \<1 year of age, it is possible that EIM- derived measures would have revealed additional abnormalities. Our data suggest that diaphragm atrophy in intubated critically ill children occurs rapidly and irrespective of age. Similarly, in critically ill adults, diaphragm atrophy occurs within days, and perhaps hours, of initiating invasive mechanical ventilation. In animal studies, specific force generated by diaphragm muscle fibers (maximum force normalized to fiber cross-sectional area) decreases 25% six hours into neuromuscular blockade and mechanical ventilation. Diaphragm atrophy in adults predicts difficulty separating from the ventilator, lengthened ICU stay, worse functional outcomes and mortality^9-17. Similar studies in children do not yet exist, although a retrospective analysis of a large database revealed that ICU-AW was associated with longer mechanical ventilation and with increased risk of discharge to a chronic care or rehabilitation facility. Our data indicate that future studies in children may need to investigate prospectively how diaphragm atrophy affects outcomes such as respiratory muscle strength, duration of mechanical ventilation and extubation failure. Indeed, a recent observational study showed that children with diminished respiratory muscle strength, as evidenced by lower airway pressure generated during airway occlusion, were significantly more likely to require reintubation within 48 hours of extubation. In addition, mechanical ventilation modes that require diaphragm contraction (e.g. Neurally Adjusted Ventilatory Assist, NAVA) may prevent diaphragm atrophy. In limb skeletal muscles, muscle atrophy in critically ill children appears to depend on age. On average, children \>1 year old lost muscle mass in arms and legs, whereas infants \<1 year old did not. Indeed, 89% of older children (16/18) lost thickness in ≥1 limb muscle. In contrast, only 41% of infants (5/12) lost thickness in limb muscles. US may be less likely to detect limb muscle loss in infants due to smaller muscle size. Age, however, has been noted as a potential factor in ICU-AW previously. Banwell et al. noted that 11/14 patients with muscle weakness in their cohort were \>10 years old, and none was \<18 months. One possible explanation is that older children, unlike infants, bear weight. Hence, supine positioning during mechanical ventilation may suddenly decrease limb muscle loads in older children but not in infants. Muscles atrophied more in children with TBI in our small sample. The association remained significant when adjusted for age. TBI often results in a hypermetabolic state with negative nitrogen balance (although see Mtaweh et al.), which may contribute to exaggerated muscle breakdown. In animal studies, TBI-associated muscle wasting occurs regardless of nutritional status, suggesting that brain injury may initiate specific signaling cascades that alter muscle function. Experimental TBI increased expression of atrophy markers atrogin-1 and m-calpain and altered muscle contractile properties. TBI may also decrease resting muscle tone early after injury, which may contribute to muscle wasting. Our sample of children with TBI is small, and our findings require confirmation in a larger study. If confirmed, the finding that TBI increases severity of acute muscle loss may highlight the need for early interventions to prevent secondary deterioration and improve functional outcomes. We included EIM to assess its utility in predicting muscle atrophy in critically ill children. EIM is noninvasive and requires minimal training and patient cooperation. Our data show that EIM, as assessed with a commercial device, is feasible in older children but not in infants. Both markers of muscle “quality” and fat percentage deteriorated during the study. The correlation with thickness, however, is relatively weak. It is unknown whether EIM predicts functional outcomes after critical illness. Interestingly, intermittent electrical muscle stimulation may prevent muscle atrophy in critically ill adults. It remains to be determined whether EIM could predict the efficacy of such intervention in critical illness. The current findings generate several questions and provide impetus for future studies of acute muscle loss in critically ill children. First, does diaphragm atrophy predict difficulty weaning from mechanical ventilation? Second, do mechanical ventilation modes that require sustained diaphragm activity during inspiration prevent diaphragm atrophy? Third, does atrophy severity correlate with functional disability after discharge? Fourth, do interventions such as passive and active motion, early mobilization or electrical stimulation preserve limb muscle mass and improve outcomes in older children? Fifth, how does nutrition affect muscle loss? Sixth, does severe TBI predispose children to acute muscle atrophy and, if so, through what mechanism? Future studies to address these questions in critically ill children are needed. [^1]: The authors have declared that no competing interests exist.

# Introduction Words of natural language along with idioms and phrases are used in speech and writing to communicate conscious experiences, such as thoughts, feelings, and intentions. Each meaningful word, considered without any context, is characterized by a set of semantic connotations. These connotations are a product of, and correlate with experiences communicated with the use of the word. Stated differently, communicated word semantics are behavioral correlates of experienced semantics. Therefore, the scientific characterization of word semantics can shed light on semantics of human experiences. In particular, if word meaning can be measured based on a metric system, the same metric system might be useful to measure the meaning of experiences. Thus, a precise metric system for the semantics of words could be a key in developing empirical science of the human mind. To build a metric system for the semantics of words means to allocate words in a metric space based on their semantics, i.e., to create a semantic map of words. There are multiple ways to generate such maps based on the representation of semantic dissimilarity as geometrical distance. Word semantics have multiple, possibly complementary aspects. Semantic maps created with distance metrics that emphasize different aspects may have different properties. One aspect of word semantics determines the likelihood for the word to appear in a particular topic or document. Most of the previous studies devoted to allocating words in space based on their meaning, including Latent Semantic Analysis (LSA:) and related techniques, focused on this aspect of word semantics, resulting in domain- specific semantic maps. Here we develop an alternative approach based on the separate aspect of word semantics that determines whether two words are synonyms or antonyms (we will generally refer to a word that is either a synonym or an antonym as an *onym*). This aspect of word semantics, when expressed parsimoniously, is in many cases domain-independent, as may be illustrated with the following example. The term *short-term memory* belongs to the domains of cognitive, computational and neuro-sciences, together with its antonym: *long-term memory*. At the same time, the general sense of the parsimoniously expressed antonymy relation, “short vs. long”, is applicable to virtually any domain. This semantic aspect relates to the basic “flavor” of experience captured by generally applicable antonym pairs, : e.g., *big* vs. *small*, *abstract* vs. *concrete*, *material* vs. *spiritual*, *whole* vs. *part*, *central* vs. *peripheral*, *one* vs. *many*, *rich* vs. *poor*, etc. Interestingly, we determined that the seemingly enormous variety of possible semantic directions is reducible to a small number (estimated as four) of main semantic dimensions that are in a definite sense orthogonal to each other. We found that these main semantic dimensions can be approximately characterized as (1) “good” vs. “bad”, (2) “calming” vs. “exciting”, (3) “open” vs. “closed”, and (4) “basic” vs. “elaborate”. # Materials and Methods ## Linguistic Corpora and Core Dictionaries This study was conducted using the dictionaries of synonyms and antonyms extracted from the thesaurus of Microsoft Office 2003 and 2007 Professional Enterprise Editions, further referred to as MS, in English, French, Spanish, and German, as well as the dictionary of English synonym and antonyms available as part of the Princeton WordNet 3.0 resource, further referred to as WN English or simply WN. The MS corpora have independent origin for different languages (e.g., the English thesaurus was developed for Microsoft by Bloomsbury Publishing, Plc., while the French thesaurus is copyrighted by SYNAPSE Development, Toulouse, France). These MS dictionaries of synonyms and antonyms were acquired automatically with the following recursive procedure (see below for hardware and software details). 1. Step 1. Start in the thesaurus with the seed word “*first*”, or its translation in other languages. Alteration of the initial word never changed the resultant core dictionary by more than a few words. 2. Step 2. Add all synonyms and antonyms of the word to the dictionary, avoiding duplicates; repeat step 2 using each of these onyms sequentially as a new word. 3. Step 3. Take the next word from the thesaurus in alphabetical order, and repeat steps 2 and 3. After the last alphabetical word, resume with the first one and continue until the entire thesaurus is processed. Next, we extracted the subset of the dictionary corresponding to the largest component of the graph of synonym and antonym links truncated to nodes (words) with a minimum of two links, including at least one antonym link, per node. In particular, the MS dictionaries of synonyms and antonyms, and the equivalent WordNet dataset downloaded from the zipped files available online (<http://wordnet.princeton.edu> on 3/29/07), were further processed in the following ways. 1. Step 4. Symmetrize the onym relation by making all synonym and antonym links bi-directional. In other words, if word *A* is a synonym of word *B*, then *B* is synonym of *A*. This symmetrization is necessary to define the energy function. 2. Step 5. Eliminate onym inconsistencies: if word *A* is listed at the same time as synonym and antonym of word *B*, both onym relations between *A* and *B* are removed. 3. Step 6. Identify the largest connected cluster in the graph of onym relations. Remove all words that do not belong to this main cluster. 4. Step 7. Eliminate all words with no antonyms or fewer than two synonym/antonym links. The remaining dictionary of synonym and antonyms is referred to as the “core” dictionary. The different core dictionaries had widely differing characteristics. The MS English core has 15,783 words, with an average of 11 synonyms and 2.7 antonyms per word. The WN English core has 20,477 words, with an average of 3.8 synonyms and 4.2 antonyms per word. The MS French core has 65,721 words, with an average of 6.5 synonyms and 10 antonyms per word. The MS German and Spanish cores have 93,887 and 259,436 words, respectively. The total size of each corpus is above 200,000 words, and in all cases, the extracted cores were a small part of the entire thesaurus. However, the next largest connected cluster was typically several orders of magnitude smaller than the core. For example, in WN the second largest connected cluster only contained 34 words. ## Construction of the Semantic Map Our approach to constructing a cognitive map by self-organization of a distribution of words in a multidimensional vector space is inspired by statistical physics. At the beginning, we randomly allocate all *N* words of a given core dictionary as points in a high-dimensional unit ball, i.e. as vectors with length ≤1. The specific results described here were obtained with a dimension of 26, but they remained essentially identical when using the lower and higher dimension values of 10 and 100, respectively. Next, we minimize an “energy” or cost function *H* of the distribution, thereby finding a minimum or “ground state” of the system. The energy function of the word configuration **x**, was defined precisely as follows: Here **x***_i* is the 26-dimensional vector representing the *i*^th word (out of *N*) in the configuration **x**. The *W_ij* entries of the symmetric relation matrix equal +1 for pairs of synonyms, −1 for pairs of antonyms, and zero for all non-onym pairs. Intuitively, maximizing the first sum moves synonyms towards the same hemispaces, while minimizing the second tends to align antonym pairs on opposite sides of the origin, reflecting their semantic relations. The fourth-power norm provides a soft limit to the absolute distance from the center. More specifically, the first term of the equation is the simplest analytical expression that captures the intent of aligning synonym vectors in parallel and antonym vectors in opposite directions. The last term is the lowest symmetric power term that is necessary to keep the distribution compact. This general approach and specific selection were empirically validated by their successful reconstruction of a map whose meaning was known a priori, that of color space, as illustrated at the end of the section. This process may be illustrated with an example. In the initial random distribution of all words, before minimizing the energy function (\*), the angles between word vectors in multi-dimensional space tend to be close to 90°. For instance, one specific simulation run using MS Word English data started from the following angles for a sample of word pairs: *right*/*wrong*, 63°; *excited*/*hectic*, 71°; *right*/*excited*, 91°. During the optimization process, words move from the initial random allocation based on their synonym/antonym relations, such that synonyms would “attract” each other and antonyms would “repel” each other. After the optimization is completed, the angles between the same word vectors become: *right*/*wrong*, 178° (almost opposite directions); *excited*/*hectic*, 12° (almost parallel); *right*/*excited*, 95° (almost orthogonal). These final angles do not depend on the initial angles. The adopted optimization procedures included a second-order Newton algorithm using analytic expressions for derivatives of the energy function, and a zero- order steepest-descent algorithm with time-dependent “thermal noise” or simulated annealing. Convergence of the optimization was assessed by measuring the norm of the gradient of the energy function, as well as the relative change of the energy function itself and word coordinates in one iteration (see below for hardware and software details). In particular, the process was terminated whenever any of these monitored parameters fell below the threshold of 2·10⁻⁶ (dictated by the precision of calculations), which was achieved in all cases in less than 10⁶ steps. When the optimization is completed, we rotate the resultant distribution to its principal components (PCs) by single value decomposition. Since the cost function and optimization procedure are symmetric with respect to the origin, the final sign of any PC coordinate is not meaningful by itself and can be considered a random outcome. Thus, upon completion of optimization, we flip each axis as needed to standardize its semantics for consistency among simulation runs. We selected the axes orientation arbitrarily once and for all maps, pointing the positive ends toward “good”, “exciting”, “open” and “elaborate”, respectively. Moreover, we normalized word coordinates by the average square length of all word vectors, effectively scaling the entire distribution to the unit variance. These post-processing operations of rotation, selective axis inversion, and rescaling, do not change the intrinsic shape of the optimized distribution, but are convenient and necessary for quantitative comparison of corpora. The final distribution appeared to be systematically invariant with respect to the choice of initial random coordinates over multiple trials, suggesting that the global minimum of *H* (\*) was reached in each case. ## Psychometric Data and Word Frequency Databases The Affective Norms for English Words (ANEW) database, developed by the Center for the Study of Emotion and Attention (CSEA) at the University of Florida, was kindly provided by Dr. Margaret M. Bradley. The ANEW database contains 1,034 words and was created using the Self-Assessment Manikin to acquire ratings of *pleasure*, *arousal*, and *dominance*. Each rating scale in ANEW runs from 1 to 9, with a rating of 1 indicating a low value (low pleasure, low arousal, low dominance) and 9 indicating a high value on each dimension. Two word frequency databases were used. The first is the demographic (conversational) set from the British National Corpus (BNC), a 100 million word collection of language samples from a wide range of sources, representative of contemporary English. The XML Edition (2007 release), maintained by the University of Oxford (United Kingdom), was downloaded from <http://www.natcorp.ox.ac.uk/corpus>. The raw dataset distilled so to exclude those items occurring five or fewer times included 14,736 words, of which 2,453 were common with the MS English core and were used in our study. The second word frequency database we employed is the Sydney Morning Herald Word Database, which contains frequency and density figures from one full year (1994) of newspaper publication, amounting to more than 23 million words in 38,526 articles. This “Australian” database, maintained by the University of Queensland, was downloaded from <http://www2.psy.uq.edu.au/CogPsych/Noetica/OpenForumIssue4/SMH.html>. The curator's filtering to exclude items that occur in only one article yield 97,031 words, of which 8,807 were common with the MS English core. ## Software and Hardware The algorithm to acquire the MS dictionaries of synonym and antonyms (Steps 1–3 above) was based on COM (Component Object Model) automation, and implemented in MathWorks Matlab (v7.5, R2007b) following published examples. The programs to extract the core dictionaries (Steps 4–7), to construct the semantic map (as described above), and to analyze the results, were custom implemented using a combination of GNU C (GCC 4.2) and Matlab along with their standard libraries and functions (all code is available upon request). These programs ran under the Windows XP Professional, Linux Fedora 7 and 8, or SunOS operating systems, either on a Dell Optiplex GX620 workstation or on a Sun Fire V890 server. # Results ## Construction and Geometric Characterization of the Semantic Map Starting from the synonym/antonym matrix extracted from the widely-employed English thesaurus of Microsoft Word (MS English), optimization converges to a definite stable state that is macroscopically independent of the initial random conditions (details in the section above). Upon rotation to principal components and normalization to unit variance, the resulting spatial distribution of words displays distinct geometric features associated with corresponding word meanings, i.e. it constitutes a semantic map. A first semantic interpretation of the principal components was derived by examining the word sorted along each axis. The top and bottom of these lists indicated that the first principal component captures the notion of good/bad (‘valence’), the second of calming- exciting (‘arousal’), and the third of open-closed (‘freedom’). A more detailed semantic analysis is provided below. The maximum spread planar projection exhibits a prominent “U-shape” resulting from a bimodal distribution along the first dimension and a unimodal distribution along the second. Subsequent components are all unimodal with a systematic increase in the “peakedness”, or kurtosis. The first three and four components encompass 95% and \>99.9% of the spatial variance, respectively, irrespective of the dimensionality of the initial embedding (R¹⁰–R¹⁰⁰). Qualitatively similar features emerge when adopting an independent dictionary of synonyms and antonyms, Princeton's WordNet, and different languages, including French, German, and Spanish ( section below). ## Qualitative and Quantitative Semantic Characterization of the Map A key issue in the analysis of the constructed semantic map is the assignment of clearly recognizable semantics, if any, to each of the significant principal components, which are all geometrically orthogonal to each other. Such identification of the principal semantic components demonstrates the suitability of this approach to establish a metric to measure meaning and the content of mental states. The relative locations of words in the map consistently match the content of their meaning. Specifically, the projection of words onto the first principal component of the map systematically lines up along the “good-bad” dimension (‘valence’). More precisely, the sign of this coordinate robustly predicts the “positive-negative” content of each word, and the numerical value along this axis accurately orders words according to that aspect of their meaning. To illustrate this feature with an example, we ranked words describing mood (from best to worst) based on an independent psychometric measure of “pleasure” derived from a large number of human raters, namely the first of the Affective Norms for English Words (ANEW). Traversing the resulting list in the MS English map yields a quantitative “mood scale”, from *happy* (1.96), *confident* (1.50), *merry* (0.99), and *untroubled* (0.78), to *bored* (−0.57), *helpless* (−1.01), *hurt* (−1.33), *depressed* (−1.59), and *sad* (−1.89). These words follow the exact same order in the map derived from WordNet (WN), and the quantitative values between the two are tightly correlated (R = 0.95, p\<10⁻⁴). This characterization of the first component generalizes to all words of the dictionary, demonstrating a highly significant correlation both between corpora (MS and WN) and with the ANEW ‘pleasure’ scale. The second component of the map similarly orders terms based on a connotation of “calming-exciting” or “easy-difficult” (vertical axis in). Both the sign and the relative value of this coordinate are again consistent semantic predictors, as in the examples of *relax* (−1.55 in the MS map, −1.05 in WN), *troubling* (0.62, 0.95), and *excite* (0.99, 1.16). Since principal components are by construction orthogonal on the map, the values of word coordinates in these first two dimensions (PC#1 and PC#2) are mutually independent. In particular, words with negative ‘arousal’ value can be either good or bad, as in *soothing* (first principal component 0.69, second −1.19 in MS) and *boring* (−1.31, −0.94), and the same holds for positive arousal terms such as *thrilling* (0.88, 0.74) and *shocked* (−0.50, 0.76). More generally, while the positions of words in the maximum spread projection (first two components) are highly consistent among MS English, WordNet, and the map derived from MS French thesaurus, they bear no implication on the values of subsequent components. The precise semantics of a component is given by the entire distribution of words on the map. For practical purposes, however, these semantics may be approximately described by the most representative words. In particular, the projection of a word on a given axis reflects its semantic amount along the corresponding component. At the same time, the alignment of a word with an axis provides an indication of the semantic specificity for that component. Thus, every semantic component of the map can be intuitively characterized by the words with both the largest projection on, and the best alignment with, each axis in either direction. For a given *i^th* component, these words can be found as follows. We divide the *i^th* coordinate of each word by the square root of their individual vector length, and sort all words according to the result. The projection of a word on each axis simply equals the value of the corresponding coordinate, while the alignment with an axis is measured by that coordinate value divided by the word vector length; thus the coordinate divided by the square root of the vector length is the geometric mean between the projection on, and the alignment with, a given axis. The top and bottom words of the sorted list are taken to represent the meaning of that component. A similar process can be applied to antonym pairs. In particular, antonym pairs can be sorted by dividing the difference of the two words in the given coordinate by the square root of their vector distance. The top antonym pairs in the sorted list are also taken to represent the meaning of that component. Both approaches based on individual words and antonym pairs reveal definitive and consistent semantics for all four significant PC's in MS English. For example, the top individual words for the first component (*clear*, *well…*, *improve*) all have positive valence, while the bottom ones (*decline*, *poor…*, *bad*) all have negative valence. Similarly, the sorted antonym pairs (e.g. *happy/sad*, *well/badly*, etc.) have opposite meaning relative to valence. The semantics of the third and fourth orthogonal dimensions can be summarized as “open/closed” (‘dominance’) and “copious/essential”, respectively. The first three components, but not the fourth, are also consistent with the corresponding semantics of both the WN English corpus and the MS French corpus, after automatic translation into English with the Google translator tool (<http://translate.google.com>). In particular, a large number of terms repeated in the same components across corpora and languages, reflecting general semantic agreement in matching PCs. As demonstrated in the next section, this correlation can be quantified and is statistically significant across these and several other languages and corpora. Words that ‘jumped’ components across corpora (*austere*, *bound*, *demolished*, *destroyed*, *dry*, *old*, *overcame*, *release*, *severe*, *slacken*, *subjected*, *subjugated*) always involve PC4, except one word (*smooth*) occurring in PC2 and PC3. Moreover, PC4 has no within-column cross-corpora repetitions, and in general shows lower consistency compared to the first 3 PCs. The general essence of each word can be thus quantitatively represented as a set of coordinates corresponding to its values along each of the principal components of the map. For example, the meaning of the word *serenity* has “good” valence (+0.59 on component 1), a major “calm” term (−1.08 on component 2), and a sense of “closure” (−0.21 on component 3). In this case, there is a clearly dominant component (the second). On average, by construction, the first components tend to have higher amplitudes than later components. This means that, broadly, the most informative element of a word is how “good” or “bad” it is, followed by how “calming/exciting”, etc. It is also interesting to compare the principal semantic components of a given word on a relative scale after filtering this general trend. This renormalization can be achieved by dividing each coordinate by the average amplitude of the corresponding component. In the *serenity* case, the third component becomes nearly as prominent as the first one on this relative scale (68% vs. 72%, respectively:). ## Predictive Power of the Semantic Map As expected based on the form of the energy function *H*, words with similar meanings (synonyms) have similar proportions on the principal components of the map, i.e. small angles between their vectors. In contrast, words with opposite meanings (antonyms) tend to have anti-parallel vectors. In particular, synonyms and antonyms in MS English had median angles of 13° and 170°, respectively (means of 21° and 165°). Less than 3% of synonym pairs have angles greater than 90°, and less than 1% of antonyms have angles smaller than 90°. Upon checking, these exceptions revealed rare instances of questionable assignments in the source dictionary, which the map effectively “corrects”. For example, *opposite* and *harmonizing* are listed as synonyms in MS English, but their angle on the map, 145°, suggests otherwise. Although in most usage cases opposite and harmonizing would be considered antonyms (as predicted by the map) the assignment as synonym in the source dictionary may still be appropriate in specific contexts (such as in describing power balance, or musical tones). As an alternative example, hot and cool are typically antonyms (referring e.g. to weather or beverages), except when used idiomatically to describe an idea, a videogame, or a classmate. Overall, given a pair of synonyms or antonyms in the dictionary, their dot product identifies the correct “onyms” relation with 99% accuracy. In particular, four real numbers associated with each word contain all essential information to identify antonyms among related terms: all semantic flavors of antonymy are reducible to four principal semantic dimensions. In contrast, random pairs (i.e., typically unrelated words) have an average angle of 90°, with less than 3% of values below 13° or above 170°. It is tempting to extrapolate these considerations and assume that proximity of two words in the map is sufficient to ensure a similarity of their meanings. However, this is not the case. Unrelated word pairs vastly outnumber synonyms (∼1500∶1) and antonyms (∼7400∶1). The majority of unrelated words pertains to separate semantic domains, and could not possibly be considered synonyms or antonyms. Even the tail ends of their angle distribution constitute a disruptive confounder of the semantic relations. Stated differently, given a particular word, it is fair to assume that, among all *related* terms, synonyms will be concentrated in the neighborhood and antonyms in the antipodes. Nevertheless, unrelated words will still constitute the majority of terms even close to 0° and 180°. These unrelated words randomly end up in the proximity of a given term by virtue of their large number in the self-organizing reduction of the high number of initial dimensions into the low-dimensional principal component space. Therefore, the constructed semantic map of words differs from the high- dimensional semantic spaces typically obtained with other existing approaches, in which the distance between any pair of embedded symbols reflects the whole semantic dissimilarity for a restricted contextual domain. Our low-dimensional map complements those local approaches by observing global semantic properties. Here, the distance between locations selectively measures the aspects of the dissimilarity broadly applicable to any context, without distinguishing between domain-specific semantic flavors. A pool of terms likely related to a given word is constituted by all synonyms of synonyms or, more generally, “onyms of onyms” of that word. In particular, words which are onyms of onyms are usually in overlapping semantic domains, but not all words in overlapping domains are onyms of onyms. Having a synonym or antonym in common does not guarantee, but strongly indicates, that two words pertain to overlapping semantic domains. Thus, within the pool of onyms of onyms, one could expect angular information to be a powerful predictor of semantic content. To test this hypothesis, we sampled 20 words from MS English and WN English, and computed the cosines of their angle with each of their onyms of onyms. We then assigned the binary values of +1 and −1 to the onyms of onyms that were also reported as synonyms or antonyms, respectively. The correlation between the cosines and binary values was statistically significant in all 40 cases. In addition to finding systematically significant numerical values in all 40 cases examined, this compilation reveals the consistent ability of the map to identify, based on the dot products, “new” synonyms and antonyms not explicitly listed as such in the dictionary. A specific example may constitute a useful illustration. In WN, the term *antonym* has 22 onyms of onyms. Among these, the two terms with the largest positive dot products are the only listed synonyms, namely *opposite_word* (1.000) and *opposite*. Similarly, the words with the largest negative dot products are the only two listed antonyms, namely *equivalent word* (−0.999) and *synonym* (−0.998). The two onyms of onyms with the positive and negative dot products closest to zero lack any synonym/antonym content: *cyclic* (0.190) and *secondary* (−0.066). These qualitative observations are reflected in an *R* value of 1.00 and a *P* value of 3.3·10⁻⁸. The term *antonym* is not part of the MS English core, but the word *opposite* is, and has 306 onyms of onyms. In this case, however, the same analysis returns relatively weaker *R* and *P* values of 0.44 and 0.025, respectively. A closer inspection to the list of onyms of onyms explains this apparent inconsistency and further corroborates the predictive value of the semantic map. The top ranking positive dot products correspond to terms listed as synonyms, namely *dissimilarity* (1.00), *the other extreme*, and *contra* (both 0.98). Next in the list, while not reported as synonyms, are nonetheless correct predictions: *heretical*, *heterodox*, *competing*, and *contrary to accepted belief* (all 0.98), followed by *contending* and *hotheaded* (both 0.97). Interestingly, the next terms at similar values are again listed as synonyms: *inverse*, *opposing* (both 0.97), *deviating*, and *contrary* (both 0.96). The lower correlation value for *opposite* in MS is due to a few outliers, such as *harmonizing* (dot product of −0.80, but listed as synonym). As discussed above (see footnote 1), even in these cases the map intuitively appears to be robust enough to actually “correct” mistaken assignments (i.e., *harmonizing* is more akin to an antonym than a synonym of *opposite*). To quantify this impression, we computed the correlation for the subset of the onyms of onyms that are listed as synonyms or antonyms of the word *opposite* in the independent WN dictionary, but not in MS. In other words, we “tested” the predicted assignment of the MS semantic map based on the available data in the WN dictionary. The resulting *R* and *P* values (0.99 and 0.005) were statistically significant, and the identified terms were consistent both among the new synonym (*different*, dot product of 0.94) and new antonyms (*like*, *similar*, and *same*, at −0.74, −0.93, and −0.94, respectively). Furthermore, the words with even more extreme negative dot products, although not explicitly listed in either dictionary, were all consistent with antonym meanings: *resemblance*, *congruence* (both −0.96), *analogy* (−0.97), *equivalence*, and *similarity* (both −0.99). A potential practical application of the described semantic map consists of specifying the connotation as well as the general meaning (denotation) of words. An illustration of considering connotation is provided in, where onyms of onyms of two words (*control* and *delicate*) are plotted in the plane of the first two principal components. In general, terms are located in the proper octant according to the connotation of their meaning (“good”, “good/exciting”, “exciting”, “bad/exciting”, “bad”, “bad/calming”, “calming”, “good/calming”). For instance, the term *control* can be substituted with a “good” connotation by *organize*, or with a “bad” connotation as *curb*. Likewise, *delicate* can connote a “calming” semantic as *soft* or an “exciting” semantic as *personal*. Moreover, the vector representation of words in this map has both absolute and relative meanings. For example, the terms *okay* and *good* lie in the same quadrant of the map with an angle of 10° between them and can be considered “absolute” synonyms. In particular, they both have a positive value in the first component (1.36 and 2.13, respectively). However, with respect to the position of *fine*, these two terms lie on opposite sides (i.e., the angle between the vector connecting *fine* and *okay* and that connecting *fine* and *good* is greater than 90°). Relative to *fine* (whose value in the first component is 1.70), the term okay has actually a negative valence (−0.34), whereas the term good has a positive one (0.43). The length of the vector can also be interpreted as a measure of the semantic component of a word measured by its main map dimensions, i.e. the aspect of the word meaning that distinguishes between antonym and synonym relations across most contexts. For example, the term *relevant* has greater vector length (1.33) than the term *pertinent* (1.15), but smaller than the term *important* (1.90). The word closest to the center is *emigrant* (vector length 0.36). Despite its definite meaning, this word is relatively neutral with respect to the main semantic dimensions of the map. The distribution of lengths over the whole dictionary shows a median meaning of 0.93 (μ±σ = 0.98±0.23). In contrast, the average semantics of the dictionary computed as the vector mean of all words nearly coincides with the origin of coordinates, i.e. the point of “no meaning” (first three components: −0.033±0.006, 0.080±0.004, and 0.004±0.002). However, words have different usage frequency in language. For example, the term *doctor* (which is used on average every 5511 words) is 26 times more common than the term *professor*. It is thus possible to compute an overall “concept mean”, as the frequency-weighted average position of all words in the semantic map. Such measure captures the most representative meaning composed across a particular language. In English, this vector has a significant length (close to 0.5) and a non-uniform contribution of principal components. In particular, the significantly positive projection on the first axis (\>0.5) corresponds to a “good” semantic, while all other dimensions have non-significant values. The same holds for the difference between the frequency-weighted and the absolute vector means. Thus, positive words are used more frequently in English than negative words (*P*\<10⁻¹⁸), while there is no significant preference in the other semantic dimensions (all three *P*\>0.3). ## Statistical Cross-Corpus Semantic Comparison The semantic characterization of the map principal components also enables a direct comparison across corpora, languages, and data types. As mentioned earlier, the first three components, but not the fourth, demonstrate high consistency across independent corpora (MS English vs. WN English) and languages (cf. MS French). To extend the comparison of principal semantic components to a quantitative measurement across additional corpora and languages, we also Google-translated the MS German and MS Spanish dictionaries into English. For the scope of this analysis, each corpus (after translation as applicable) was limited to the set of words that overlapped with the MS English core dictionary. For example, the 15,783 MS English core words and the 20,477 WN English core words have 5926 terms in common. For MS French, the overlap was 4704 English words, mapped onto from 19,944 French terms, representing approximately 30% of the MS French core dictionary. Many French words projected onto single English words, because word inflections are listed separately in the MS French thesaurus; the same occurred in German. We then extracted several correlation measures between the word coordinates from each of the separate semantic maps (WN English, MS French, MS German, and MS Spanish) and the MS English map. First, for each pair of corpora, we computed a matrix of PC-to-PC correlation coefficients. demonstrate a systematic two-way semantic correspondence of the first three PCs for all compared pairs of corpora. In particular, each of the first three PC in every corpus displays the highest correlation coefficient with the corresponding PC of the other corpus in the pair. These values are all statistically significant (p\<0.001). Such correspondence only holds for the fourth component between MS English and WN English, but not across different languages. Dimensions beyond the fourth are not statistically significant in MS English and are thus not represented in this table. Moreover, we compared the first three PCs of MS English with the three original dimensions of ANEW, whose semantics are identified as *pleasure*, *arousal*, and *dominance*. In this case, the two-way semantic correspondence was only revealed on the first two components. This is not surprising given that the coordinates of the ANEW dataset are not internally orthogonal. In fact, the first and third coordinates are highly correlated within the ANEW sample. We also computed the correlation of the first 4 MS English PCs with each of the 32 Paivio norms and of the 51 Rubin properties, which constitute, to the best of our knowledge, the largest available collections of psychometric measures. However, none of these attempts resulted in higher correlation coefficients than those found for ANEW. Next, we subjected each pair of corpora to canonical correlation analysis (CCA). CCA finds the basis vectors for two sets of multidimensional variables such that the correlations between the projections of the variables onto these basis vectors are mutually maximized. The first four CCA coefficients are reported in for each pair of corpora. CCA rotates two distributions of points so as to align them for maximal correlation. Thus, the first CCA correlation must be, by construction, higher than (or equal to) the correlation between the first principal components independently obtained in the two sets. The fact that these values are extremely close between MS English and each of the other corpora (e.g. 0.78 vs. 0.73 for WN English, 0.75 vs. 0.74 for MS French, 0.83 vs. 0.80 for ANEW) suggests an excellent alignment of their intrinsic principal components. Moreover, the fact that the number of statistically significant canonical correlations (7 for WN English, French, and Spanish, and 6 for German) systematically exceed the number of significant dimensions in MS English (4) is a further indication of geometric consistency across corpora, even if the semantics no longer strictly correspond beyond the fourth dimension. Finally, as an additional method of quantifying the linear relationships between pairs of corpora (i.e., two multidimensional variables), we defined an “overall correlation” *OC* (\*\*) based on the norms of the covariant matrices, which are the natural generalization to higher dimensions of the concept of the variance of a scalar-valued random variable. The covariance matrix or dispersion matrix is a matrix of covariances between elements of a vector, and naturally generalizes to higher dimensions the concept of the variance of a scalar variable. The correlation coefficient for a pair of scalar variables is the ratio of their covariance to the product of their standard deviations. Our formulation (\*\*) is a natural extension to variables in multiple dimensions. The formula is analogous to that of the Pearson correlation coefficient, and coincides with it in one dimension: This measure characterizes the alignment of two distributions of points, each independently rotated to their internal principal components, throughout all of their dimensions. The overall correlation coefficient consistently assumed high values (between 0.68 and 0.80), always intermediate between the first canonical correlation and the correlation between first principal components. This result of the cross-corpus comparison, as well as the qualitative assessment of the semantic content of the significant principal components, also proved to be generally robust with respect to alterations of the cost function parameters and/or the initial conditions in optimization. These findings indicate overall consistency and reliability across languages, datasets, and variations of the technique. ## Validation in Color Space To verify the general applicability and robustness of our approach, we designed a simple simulation of color mapping. The model semantic space *X_color* was defined as a sphere *S²*, in which each point was associated with a unique color, using the three Cartesian coordinates as RGB values. A number *n* of points (initially set to *n* = 1000) were randomly sampled from *X_color*. For each sampled point, a list of “synonyms” and “antonyms” was generated by stochastically selecting neighbors within a certain ‘threshold angle’ as synonyms and neighbors within that threshold angle from the antipode as antonyms. The initial values for the threshold angles and the average number of onyms per point (the ‘degree’ of the graph) were set to 20° and 3.5, respectively, consistently with the parameters of the available linguistic corpora, and later allowed to vary as described below. The points were then embedded in a *d*-dimensional space (with a default value of *d* = 10) with random initial coordinates. Their coordinates were optimized by minimizing the above-described energy function *H* of locations and synonym- antonym connections, using the same convergence criteria adopted for the main language study (see ‘Construction of the Semantic Map’). Finally, the resultant distribution was rotated to principal components. The resulting accurate reconstruction of the coloring of the sphere indicates that the topology and geometry of this cognitive map (whose semantic was in this case known by construction) could be reconstructed from a sparse subset of synonym and antonym relations. Specifically, after reconstruction, the amplitudes (standard deviations) of the first three PCs are each close to 1, while the remaining 7 are negligible, resembling the situation observed in MS English. The semantics of the reconstructed map are also consistent with the original map, as intuitively seen from comparison of the two color projections. This intuition is confirmed by numerical measures of the above defined overall correlation (\*\*) between the original and reconstructed maps. In particular, altering the dimensionality of the embedding space *d*, the average number of “onyms” per color node (i.e., the average node degree), the threshold angle between “onyms”, as well as the number of color nodes, did not affect the quality of the reconstruction in a wide range of parameters. In other words, the results of this approach are robust with respect to alteration of the corpus parameters: the dimension of the embedding, the number of “onyms” per “word”, the number of “words”, and the maximal/minimal distance or angle between “synonyms”/“antonyms”. # Discussion In his 1946 “Man's Search for Meaning”, neurologist and psychiatrist Viktor Frankl maintained that life has meaning under any imaginable circumstance, that the search for this meaning is the core human drive, and that personal freedom consists of the individual choice of such meaning. Although internal meaning may be viewed as the most (or arguably, the only) important matter of human existence, its scientific characterization has so far resisted the otherwise seemingly unstoppable strides of technological progress. This topic has been at times dismissed as metaphysical due to the perceived impossibility to reconcile the individual, first-person perspective of the very meaning of any concept, and the scientific requirements for objective validation, unambiguous communication, systematic reproducibility, and empirical falsifiability. Recently, however, the need, potential, and importance of extending traditional research paradigms to include subjective experience have been recognized with increasing urgency. One of the missing foundations is a precise measure of the content of mental states. The present study is a step toward bridging this gap. ## Major Conclusions This study demonstrates the possibility to derive a precise metric system for semantics of human experiences objectively from data collected without using human subjects. More generally, the new technical approach we presented may have practical implications for multiple fields. Previous studies that resulted in semantic maps either relied on subjective human judgments (e.g. ANEW, semantic differential) or were not explicitly related to human experiences (e.g. LSA, Latent Dirichlet Allocation: LDA). In contrast, we constructed a prototype general metric system for semantics from all-purpose dictionaries, and validated its applicability to human experiences by available psychometric data. The significant correlation between the affective space of ANEW and our semantic cognitive map establishes a strong, novel, and unexpected connection between results in experimental psychology and computational linguistics. Self-organizing semantic maps have been described before, and numerous methods exist to construct spatial representations of lexical knowledge. However, to our knowledge, this is the first objective approach to construct, based on available data, a simultaneous quantitative representation of synonymy and antonymy in a continuous metric space, whose dimensions have clearly identified general meanings. The low dimensionality of this semantic map indicates that, although thousands of distinct categories of meanings are conceivable, only very few apply to all contexts without a substantial domain-specific alteration of their semantic content. This limited number of general meanings is consistent with recent independent linguistic dimensional analyses and contrasts with the extensive lists of semantic categories represented in Roget's thesaurus and related or similar endeavors. At the same time, the remarkable consistency of the significant principal components of our map across dictionaries and languages, as well as with previous psychometric data obtained with very different methods (such as factor analysis and word ranking), suggests that they may be rooted in the fundamental laws of the human mind. The three dominant semantic categories revealed in our study (“good-bad”, “calm- excited”, “open-closed”) are consistent with earlier psychometric, cognitive, and linguistic theories and findings, including Osgood's semantic differential and Leary's interpersonal Circumplex (cf.). In particular, semantic differential rating was devised as a scale to measure the affective meaning of objects, events, and concepts. Subjects evaluate the semantic content of a term as a relative position between two bipolar words, such as warm-cold, bright- dark, beautiful-ugly, sweet-bitter, fair-unfair, brave-cowardly, meaningful- meaningless. Through factor analysis of large collections of semantic differential scales, Osgood characterized three recurring attitudes: evaluation, potency, and activity. These dimensions, mostly corresponding to the adjective pairs “good-bad”, “strong-weak”, and “active-passive”, respectively, were found to be cross-cultural universals. There is a clear resemblance between these connotations and the principal semantic components of language that emerged in our approach. Similarly, the interpersonal Circumplex is a two-dimensional representation of personality based on agency, or power (status, dominance, and control), and communion, or love (solidarity, friendliness, and warmth:). The possibility to objectively define a quantitative scale for the major categories of general semantic content, capturing both synonym and antonym relations, has practical applications to linguistic data mining and sentiment analysis. The main scientific value of the constructed map, however, is to lay the foundation of a precise metric system for meaning that goes far beyond the current practice of qualitative assessment, with important implications for artificial intelligence and cognitive neuropsychology. In fact, a rigorous science of mind may require a precisely defined, universal metric system for mental state semantics. Similarly, in cognitive architectures representations need to be sorted by their semantics. ## Related Works and Novelty of the Contribution of This Work Low-dimensional vector-space representations of word meaning were constructed previously at least in two fields, namely computational linguistics and experimental psychology. In the former case (e.g. LSA, probabilistic latent semantic analysis, or pLSA, LDA, Isomap) the purpose is often to improve information retrieval systems by indicating which documents are similar and which are not. Efforts in experimental psychology (Semantic Differential, ANEW, Circumplex) aim to describe aspects of human semantic memory and affective states. The present work connected results of these two fields by establishing a correspondence between the objectively constructed semantic cognitive map and ANEW. Previous semantic maps created with different techniques did not demonstrate similar features. The observation that positive words are used more frequently in English than negative words provides additional evidence for the usefulness of the map as a metric system for human experiences. The semantic similarity of our map with ANEW in the first two dimensions was quantitatively confirmed by canonical correlation analysis, based on the map locations of words that are common for the two maps. However, the two maps are not equivalent to each other. The map constructed in the present study contains more dimensions and more words, including words that do not belong to affective stimuli. Most importantly, this map differs qualitatively from previous data as it was not constructed based on given semantic dimensions. Instead, semantics of our map dimensions are emergent and defined by the locations of all words together. The constructed semantic cognitive map provides one geometrical representation for two relations: synonymy and antonymy. Most existing automated methods infer synonymy from word co-occurrence and do not explicitly account for antonymy. Thus, the ability to represent antonymy, which may capture a vital aspect of meaning, constitutes an essential feature of our approach. Previous semantic cognitive mapping studies involving dissimilarity metric, had problems to find a geometric representation of antonymy. This limitation of known approaches could be due to the non-trivial relation between antonymy and the traditionally used dissimilarity metric. For example, *king* and *queen* could be synonyms, as in head of the royal family, or antonyms, as in gender (see also footnote 1 above). Our choice of energy function (\*) departs from the current paradigm. The principal components of the resulting map uniquely capture the general aspects of antonymy, i.e. those that apply to most contexts. Accordingly, the notions of synonymy and antonymy used in our analysis differ from the concepts of similarity and dissimilarity as defined by co-occurrence, as illustrated by the king/queen or hot/cool examples mentioned above. Many definitions of antonymy were proposed over the years, and none of them is reducible to a notion of (dis)similarity. Unlike with LSA and related techniques, were the low-dimensionality of the map results from manual truncation of higher dimensions, in our case this property emerged naturally. This may have broad implications. In the foundational hypothesis of a set of categories as generators of language, the number of necessary categories was believed to be large. The idea that such large variety of antonymy senses used in natural language is reducible to relatively few basic notions was actually discussed in the previous century, but is no longer considered in modern linguistics. It is therefore surprising that this reduction can be achieved with only three or four basic dimensions. Unlike most previous studies, our model was not tailored for a special practical purpose, but was constructed starting from basic principles. Our energy function was selected as the most parsimonious analytical expression corresponding to the concept of synonym and antonym vector alignment. The first term is the simplest analytical expression that attempts to align synonym vectors in parallel and antonym vectors in opposite directions. The last term is the lowest symmetric power term that is necessary to keep the distribution compact. This conceptual framework significantly differs from the frameworks mentioned above, including LSA, LDA, Multidimensional Scaling (MDS), etc. Semantics of the principal dimensions of our map are reproduced across databases and languages. This is not a characteristic of any previously constructed vector semantic map in computational linguistics. Even though dimensions of the earlier constructed maps have identifiable semantics, those semantics are domain-specific, and there is no visible semantic similarity between our map and various vector representations of semantics of words constructed using LSA, LDA and other approaches. ## Limits and Applications of the Semantic Map Although the constructed semantic map reveals definitive semantics in each of its significant principal components, the vector associated with every word in the map should be interpreted as a “noisy” measure rather than an exact set of numerical values. This cautious interpretation is motivated by two considerations. First, the positions of individual words on the map depend on the selection of available synonym-antonym links, which only constitute a small subset of all possible synonym-antonym links. Adding or deleting a link changes map coordinates of the corresponding words. Stated differently, any dictionary of synonyms and antonyms only provides sparse sampling of the onym graph. The quantitative extent of this sparse sampling can be estimated by comparing two independent thesauri, such as MS English and WN English. Limiting the respective dictionaries of synonyms to the pool of their 5,926 words in common leaves 30,922 links for MS and 12,188 for WN, with 6,576 overlaps. Assuming that synonyms in each of the two dictionaries are sampled randomly and independently from the “comprehensive” set of all true synonyms, the cardinality of the true synonym set can be computed as (30,922·12,188/6,576) = 57,311. Thus, the MS and WN English dictionaries only represent at most ∼54% and 21%, respectively, of all synonyms. However, the assumption of independent random sampling is unlikely to be realistic, because more usual synonyms may have a greater chance to be listed in both dictionaries, thus increasing the number of overlaps. Therefore, these values should be considered coarse overestimates, and the real representation is likely to be even sparser. The second major source of noise in the constructed map is that each word is associated with a number of potentially very different meanings, or “senses”. For example, the word *mean* can assume the distinct meanings of “average”, “nasty”, and “indicate”. Therefore, the word vector may be forced to find a compromise orientation that does not match precisely any of the word meanings. From this perspective, the constructed map crudely approximates meanings with words. Semantics of individual words may not match precisely semantics of their map locations, and therefore should not be taken as literal definitions of the latter. Although the map was constructed based on relations among individual words, precise numerical definitions of its semantics only apply to large subsets of words, as in the analyses involving word frequency data. More generally, individual map locations can be viewed as representing unambiguous, topographically organized semantics defined by the entire distribution of all words on the map rather than by one word. In particular, the map location of a specific meaning could be computed precisely as the center of mass of the group of all its representing words. Two meanings with close/opposite centers of mass would be more likely to be synonym/antonym than two individual words separated by the same distance on the map. The accuracy of the map location of a meaning would increase with the number of its representing words. Ideally, in order to precisely allocate meaning on the map, the center of mass of all dictionary words should be computed with appropriate weights measuring their semantic agreement with the given meaning. As a result of these two limitations, namely sparse sampling and approximation of meanings with words, individual word coordinates are subject to considerable noise, the relative amplitude of which can be roughly estimated as 10–20%. Nevertheless, the map is robust with respect to the assignments of synonyms and antonyms, and their connotation, from sets of related words. In particular, within all onyms of onyms, constituting a pool of terms likely related to a given word, dot product is a powerful predictor of semantic content. Moreover, when global map characteristics are derived from all word coordinates, as in cross-corpus map correlations the noise effectively averages out. This means that the map can be used as a precise semantic scale, even if individual words cannot. In addition, our map does not capture the whole semantics of a word, but only the aspect that distinguishes between synonyms and antonyms in a context- independent query. The domain-specific part of meaning, including the aspect that determines the likelihood for a word to appear in a particular topic or document, is missed equally for all words. Words that fall near the origin (like “emigrant”) do not have a significant measure of the “semantic flavor” that this map represents. This is also why finding unrelated words next to each other on the map does not indicate an inconsistency. ## Relating the Constructed Word Map to Semantic Space Semantic space, or the set *X* of all meanings, by assumption can be mapped into a high-dimensional Euclidean space (left). Selected relations among meanings represented by words are shown as vectors connecting points of *X* (colored arrows). These relations have each their own domain of applicability in *X*. Dashed lines of corresponding colors show the domain boundaries. For example, the word *hot* can be viewed as a label for the relation among two meanings represented by points in *X*, one of which can be considered *hot* as compared to the other: the red color is hot compared to the blue color, the weather in Mexico is hot compared to Canada, the housing market in Manhattan is hot compared to that in Detroit. The relation *hot*, however, has a limited domain of applicability. For instance, this concept does not make sense in general when referred to pairs of elementary geometrical shapes. As a particular example, a triangle can be said to be *sharp*, but not *hot*, compared to a circle. Domains of applicability of two relations labeled by words may be overlapping or disjoint. For example, domains of applicability of *hot* and *sharp* overlap, e.g. in the food domain, while the domains of applicability of *differentiable*, a mathematical term, and *charismatic* appear to be disjoint. Two relations labeled by words within an overlap of their domains are synonyms, if their vectors point in the same or similar directions (e.g., *hot* and *sharp* in the food domain). They are antonyms, if their vectors point in the opposite or nearly opposite directions (e.g., *hot* and *cold*). These notions of synonymy and antonymy have a clear geometrical interpretation in *X* locally. However, they may or may not be globally consistent. For instance, *good* and *bad* are in general globally consistent antonyms, i.e. they point in nearly opposite directions in all of their overlapping domains of applicability. In contrast, *hot* and *cool* are often antonyms but occasionally point in similar directions, i.e. are synonyms, as in the example of “a *hot* videogame” and “a *cool* videogame” (cf. footnote 1). The vectors representing relations labeled by words, when translated to a common origin, span a vector space *V*. Here they can be further rotated to reduce the dimension of *V*, respecting the following rule: global synonyms should remain nearly parallel and global antonyms nearly anti-parallel. However, the converse may not be true. For example, if red and brown arrows have overlapping domains in *X and* represent synonyms, then they should be nearly parallel in *V*. Blue and purple arrows have disjoint domains and therefore cannot be called global synonyms or antonyms, despite the fact that they are nearly parallel in *V*. Thus, their mutual orientation in the embedding of *X* could be any. Red and purple arrows have overlapping domains and are nearly anti-parallel in *X* (antonyms), therefore, they have to keep this property in *V*. However, brown and purple arrows cannot be antonyms, because their domains are disjoint. Red and green arrows have overlapping domains in *X* and are orthogonal in their common domain in *X*: they are neither synonyms nor antonyms. While in principle according to the above rule they can be oriented at any angle in *V*, our numerical experiments show that they are more likely to be nearly orthogonal to each other in *V*, if other angular relations within the overlap of their domains are satisfied. The above rule to translate and rotate vectors from *X* to *V* is captured by the energy function described in (\*). As a consequence of the optimization process, the dimension of *V* can be smaller than the dimension of the Euclidean space into which *X* is mapped. However, because metrics in *V* respect consistent synonym and antonym relations among all vectors defined at any given location in *X*, the dimension of *V* is unlikely to be smaller than the dimension of *X* itself. Therefore, the dimension of *V*, which in our analysis is ∼4 provides an approximate upper bound on the dimension of *X* and a lower bound on the dimension of the Euclidean space into which *X* is mapped. According to this interpretation, the results of our work can be restated as the following. There are only a small number (∼4) of independent (“orthogonal”) semantic relations that generally apply in a consistent manner to almost all possible domains of applicability. In order of importance, or of the amount of meaning they express, as measured by the captured variance, they can be identified as good/bad (valence), calm/excited (arousal), open/closed (freedom), and copious/essential. The first three of these dimensions are consistent across corpora and languages. An alternative, simplistic view of the semantic space *X* is a connected graph *G* (right), where nodes are words now interpreted as corresponding to broad categories in the set *X*. Edges of *G* represent relations among words, namely synonymy (black) and antonymy (colored). Because each meaning of a word, and in most cases each word, typically has at most one antonym in the dictionary, words again can be associated with directions of their antonym links and therefore can be embedded as vectors in *V*, as described above. The above analysis suggests that equivalent semantic properties of *V* will result from interpretation of either individual words or pairs of antonyms as vectors in *V*. We thank Drs. Rebecca F. Goldin, Harold J. Morowitz, and James L. Olds for valuable discussions and feedback. [^1]: Conceived and designed the experiments: AS GAA. Performed the experiments: AS. Analyzed the data: AS GAA. Wrote the paper: AS GAA. [^2]: The authors have declared that no competing interests exist.

# Introduction Honey bee (*Apis mellifera* L.) workers dedicate most of their foraging phase specializing in the collection of nectar, pollen, propolis, or water. Floral nectar provides the carbohydrates needed for a colony’s energetic needs, while pollen, the main source of protein, provides bees with ten essential amino acids that are critical for brood rearing and queen feeding. Pollen is conserved during storage by mixing it with nectar and glandular secretions from workers to create what is known as “bee bread”. Nurse bees consume bee bread to develop their hypopharyngeal glands, which produce a protein-rich jelly that is used to feed developing larvae. Although a colony typically demands more carbohydrates than proteins, pollen can often become a limiting nutritional factor due to low resource availability or quality at certain times of the year. For instance, deficiencies in the availability of particular amino acids can create a bottleneck in brood rearing, and without adequate amounts and types of pollen, colonies can quickly deplete their protein reserves, leading to a reduction in brood rearing and even brood cannibalism. The average amount of pollen that a colony with 10,000–15,000 workers needs is estimated at 13.4 to 17.8 kg per year. While having a sufficient amount of pollen is important for colony maintenance, having access to diverse pollen is equally critical to colony nutrition because pollen varies across plant species in the type and amount of amino acids it contains. Not surprisingly, polyfloral diets increase worker immunocompetence and overall colony tolerance to pathogens. For example, colonies fed a polyfloral diet exhibit longer worker lifespan by decreasing their susceptibility to the microsporidian gut pathogen *Nosema* spp.. While a few studies suggest that honey bees tend to forage pollen based on the proximity to available floral resources, others suggest that foragers are capable of displaying pollen preferences based on their colony’s nutritional requirements. Nevertheless, it is clear that having access to a consistent flow of pollen from diverse floral sources is beneficial to honey bees. The recent surge in public awareness regarding the role of honey bee pollination in agriculture has led to an increase in the number of small-scale beekeepers across the U.S., particularly in developed urban and suburban areas. While large-scale commercial beekeeping operations (i.e., those that manage 500 or more colonies) still provide the majority of pollination services to agro- ecosystems, backyard and sideline beekeepers (i.e., those that manage up to 50 or 500 colonies, respectively) represent almost 99% of the beekeeper population in the country. Urban and suburban environments present a different system for colony management compared to rural or agricultural landscapes, given that taxonomic plant diversity can be affected in various ways depending on the degree of land development. For example, heavily developed urban areas are mainly covered with pavement and buildings, resulting in loss of green spaces and reduced availability of plants. In contrast, moderate levels of urbanization may actually increase diversity through irrigated public parks and private gardens. Regardless of the degree of development, many of the small vegetative patches in developed areas often contain few or no native plants because they have been replaced by non-native ornamental plants, which are selected for their aesthetic value rather than for their benefit to pollinators. Urban environments are dominated by few species of ornamental plants that are either competitively dominant or favored in those settings. Although many parks and private gardens in suburban environments contain native and non-native plant species, overall, urban development and habitat fragmentation have drastically altered resource availability and diversity for pollinators. Furthermore, landscape transformation is expected to continue, as urbanization is predicted to expand worldwide, which will likely have impacts for pollinators due to changes in local flora. To investigate the floral sources collected by honey bee foragers in any landscape, bee-collected pollen can be analyzed with melissopalynology techniques that are typically used to identify pollen in honey to determine its floral sources of nectar. However, pollen foragers do not visit the same plants that nectar foragers from the same colony visit, resulting in differences in the types of plants foraged by a colony depending on its nectar and pollen needs. The few studies that have analyzed the composition of pollen pellets collected from foragers have done so using colonies located in predominantly undeveloped or agricultural landscapes. Consequently, we have a limited understanding of the pollen foraging preferences of colonies located in urban and suburban environments. In this study, we identified the major pollen types collected by foragers in urban and suburban environments in California, Texas, Florida, and Michigan during the spring, summer, fall, and (when possible) winter months. We used this information to estimate species richness and diversity indices for each site, and compared those values across states and seasons, expecting to find differences in pollen diversity due to spatial and temporal variation between sites. Moreover, because studies in Europe found that colonies placed in agricultural landscapes collected the highest diversity of pollen in the summer months, we hypothesized that floral diversity in our survey would be highest in the summer season. This study provides insights regarding the foraging ecology and floral preferences of honey bee colonies throughout the year in urban or suburban environments, and serves as a foundation for future work focusing on honey bee foraging ecology and floral preferences in developed landscapes. # Materials and methods ## Site selection Beekeepers who managed colonies located in urban and suburban areas were identified through local and state beekeeping organizations in California (CA; Bay Area and Sacramento), Texas (TX; Austin and College Station), Florida (FL; Orlando and Tampa), and Michigan (MI; Detroit and East Lansing). Each owner granted permission to conduct the study on their site. We had a total of 20 sites in CA, 18 sites in TX, 19 sites in FL, and 15 sites in MI, for a total of 394 samples across all sites. Each site had at least two colonies so that a focal colony could be replaced in the event that it was later considered unsuitable for acquiring pollen for use in the study. The GPS coordinates for each site were collected and mapped on ArcGIS (Version 9.3 or 10.4). Concentric circles were plotted on the map around each site with radii of 0.8 km, 1.6 km and 4.0 km. Assuming honey bees forage in areas closer to the hive, a 0.8 km radius around a site was expected to represent the colony’s primary foraging range, while a 4.0 km radius was expected to encompass close to the maximum area where most of the foraging occurs for a typical colony under most circumstances. To account for regional differences in classifying land cover data, we used the 2011 National Land Cover Database (NLCD) to classify the types of land cover surrounding each site (<https://www.mrlc.gov/nlcd2011.php>). The NLCD classifies developed areas as those with open space, as well as those with low, medium, and high intensity of development or urbanization. The percentages of land cover for each of these four classes were summed to calculate the total percent of developed land around each site. Undeveloped landscapes consisted of those covered by aquatic systems, shrublands or grasslands, forests and woodlands, croplands, and riparian or wetland systems. Large bodies of water were categorized as unavailable foraging areas so that coastal areas did not underrepresent the level of development in a given area. Agricultural areas were separated into those covered by pasture and hay, and those with cultivated crops. Aerial images of the areas around each colony provided a visual depiction of each site and helped identify landscape features such as residential areas, golf courses, open water, and undeveloped areas. The aerial images were also used to supplement the NLCD to capture recent land development by overlaying the aerial images over the NLCD-classified features. Areas that could be identified as residential or developed using aerial images were manually re-classified as being developed. We then calculated the percentage of developed area out of the total area surrounding the concentric circles around each site. All sites in each state had to be separated from one another by at least four km. ## Pollen collection We sampled pollen from CA and FL monthly from July 2014 to June 2015 (12 months), MI from July to September 2014 and March to June 2015 (7 months) and TX from March to October 2015 (8 months). We used the Northern meteorological calendar to categorize the data by season based on the month in which they were collected. Spring sampling was done from 1 March to 31 May, summer sampling was done from 1 June to 31 August, fall sampling was done from 1 September to 30 November, and winter sampling was done from 1 December to 28 February. Each colony had a pollen trap (Brushy Mountain Bee Farm’s item ND#464 for ten- frame hives and item \#509 for eight-frame hives) placed at the hive entrance. Participating beekeepers were instructed to activate the pollen traps up to one week prior to the sampling date to collect sufficient pollen from a variety of floral sources. Pollen was passively collected when foragers passed through the activated trap, causing a subset of the pollen pellets on their corbicula to fall into the collecting tray. We sampled pollen within one week or less of activating the traps to get fresh samples that represented the surrounding foraging area from a hive at that particular time. A minimum of 1 g of pollen was collected for palynological analysis from each sampling site at each time point. In total, we collected 146 samples in CA, 79 samples in TX, 144 samples in FL, and 87 samples in MI. The samples were kept frozen at -20°C and shipped on dry ice to Texas A&M University in College Station, TX, for pollen analysis. ## Pollen analysis Each pollen sample was processed using standard acetolysis procedures. Acetolysis is used to remove the protoplasm of a pollen grain so that its morphological characteristics can be visible under a light microscope for taxonomic identification. Subsets of at least 0.25 g of pollen from each sample were homogenized in a test tube with 10 mL of 95% glacial acetic acid, which was added under a fume hood to remove water in the sample, thus preventing a potentially dangerous exothermic reaction between any water residues and the acetolysis mixture. The samples were vortexed and centrifuged at 1060 x *g* for 2 min. After the glacial acetic acid was decanted, 10 mL of the acetolysis mixture containing a 9:1 ratio of acetic anhydride and sulfuric acid was gradually added to the sample. The mixtures were stirred occasionally and were allowed to react on a heating block at 80°C for 10 min. The samples were then topped off with glacial acetic acid and allowed to cool before being centrifuged again at 1060 x *g* for 2 min. After the supernatant was discarded, the samples went through a series of washes with glacial acetic acid and distilled water, and were then stained using Safranin O (Sigma-Aldrich, MO) and rinsed with 95% ethanol. Safranin O is commonly used in palynology because it stains pollen grains in a way that increases the image contrast for microphotography. The samples were transferred to Eppendorf tubes containing glycerin and left open for 24 h to dry the ethanol. Slides were made the following day by evenly spreading a drop of pollen residue across the slide and sealing it with a coverslip using clear nail polish. Pollen grains were identified to the family, genus, or species level using a Nikon E200 light microscope. A Nikon DS-L3 stand-alone microscope camera controller was used to measure the size of the pollen grains and to take digital images of unique pollen types at 200× to 600× magnification. Pollen identification was based on the ornamentation, orientation, size, and surface structure of a grain. The specificity of pollen identification depended on the grain's diagnostic characteristics and the reference collection available at the Palynology Laboratory in the Department of Anthropology at Texas A&M University. Various regional pollen atlases were also used and yet, some pollen types were only identified down to the family level. For example, pollen in the Asteraceae, Liliaceae, Poaceae, Rhamnaceae, Rosaceae, and Ericaceae families is not easily identified to the genus level without Scanning Electron Microscope (SEM) images because of the morphological similarity among plants in each of those families. Most members of the Asteraceae family were separated by spine length and differentiated as high spine (HS) Asteraceae (spines \> 2.5 μm long), low spine (LS) Asteraceae (spines \< 2.5 μm long), or liguliflorae (LF) Asteraceae in the tribe Lactuceae. Taxa were also classified as tree, shrub, herb, combination, or unknown based on the USDA plant database. We identified at least 200 pollen grains per sample. This number was recently determined to be sufficient to obtain an accurate representation of the plant taxonomic groups present in a forager-collected pollen pellet. Each identified pollen type was placed into one of four frequency categories, as proposed by Louveaux et al.. The “predominant” pollen category included taxonomic groups that were represented in \>45% of the pollen grains counted per sample. The “secondary,” “important minor,” and “minor” category included taxonomic groups that were represented in 16–45%, 3–15% and \< 3% of the pollen grains counted per sample, respectively. ## Taxonomic diversity Floral taxonomic diversity in each sample was calculated using the Shannon- Weaver diversity index to characterize taxonomic richness and evenness for each season in each state. The Shannon-Weaver diversity index (*H’*) was calculated using the equation: $$H^{\prime} = - {\sum\limits_{i = 1}^{S}{p_{i}{\ln p_{i}}}}$$ where *p*_*i* is the proportion of each pollen type (*i*) in the sample and *ln* is the natural logarithm. A greater *H’* value indicates greater taxonomic diversity. Shannon-Weaver diversity indices were calculated for each individual site for each season in every state. We used this index to compare taxonomic diversity at the national and state scale. We also calculated the Effective Number of Species (ENS) from the Shannon-Weaver diversity index for each site to perform biodiversity comparisons between different plant communities, even though the ENS has a tendency to rely on abundant taxa and minimize the input of relatively rare taxa. Unknown pollen types were accounted for by providing each one with a unique “unknown” identification code. ## Statistical analyses We tested for normality in the Shannon-Weaver diversity indices and the ENS values using a Shapiro-Wilk test. Since the diversity indices for each season and state did not meet the assumptions for normality, we used the nonparametric Kruskal-Wallis test to identify significant differences among diversity values within each season or within each state. We made multiple comparisons for diversity within seasons and states using the Dunn’s test. Because each site had a minimum of two colonies available for sampling, we collected pollen from either the first or second colony at a site despite the possibility that each colony was foraging from different resources. Therefore, each sample was treated as an independent point for our analysis. All descriptive statistics are reported as the mean ± the standard error of the mean (S.E.M.). We set the level of statistical significance for all tests at α = 0.05 and used JMP Pro 14 (SAS Inc., Cary, NC) for all statistical analyses. # Results ## Verification of developed land cover We used the geographic position of each site and the NLCD information regarding the land cover composition surrounding the site to calculate the percent of land covered by each category outlined previously (i.e., aquatic systems, shrublands or grasslands, forests and woodlands, croplands, and riparian or wetland systems). Using this information, we estimated that across all sites, on average, 81.26% of the land within a four-km radius around each site was developed. The average developed land cover (excluding large bodies of water) surrounding each site was 74.69% in CA, 84.95% in FL, 84.71% in MI, and 80.72% in TX. Assuming a four-km primary foraging radius for foragers in a typical honey bee colony, the total surveyed area of potential foraging activity for pollen collection was approximately 3,619 km² across the four states. ## Overall plant species diversity We found a significant difference in the total overall species diversity across all four states (H = 34.69, *P*\<0.001). Post-hoc tests revealed higher overall species diversity in CA and lower overall species diversity in TX compared to other states (Dunn’s all pairs test, *P*\<0.05;). Nationally, total diversity was also significantly higher in the spring across all locations (H = 33.47, *P*\<0.001) compared to other seasons (Dunn’s all pairs test, *P*\<0.05;). There were also seasonal differences in species diversity within each state. For example, honey bees in CA collected pollen from significantly fewer (H = 23.67, *P*\<0.001) plant species in the fall than in the other seasons (Dunn’s all pairs test, *P*\<0.05). In TX, species diversity was significantly higher (H = 12.94, *P* = 0.002) in the spring (H = 12.94, *P* = 0.002) than in the summer (Dunn’s all pairs test, *P*\<0.05). Likewise, species diversity in FL was higher (H = 11.66, *P* = 0.009) in the spring than in any other season (Dunn’s all pairs test, *P*\<0.05). Although we found a significant effect of seasonal species diversity in MI (H = 6.25, *P* = 0.04), we found no pairwise differences in species diversity among seasons in that state (Dunn’s all pairs test, *P*\>0.05). ## Floral diversity in California In spring, foragers from colonies in CA collected pollen from 65 plant taxa belonging to 39 families. The average Shannon-Weaver diversity index for each site in the spring was 1.21 and the ENS was 3.75. None of the pollen collected belonged to the “predominant” category, as proposed by Louveaux et al.. That is, no individual pollen type collected represented \>45% of the pollen grains counted in a sample (see “*Site selection*” above for an explanation of each abundance category). The pollen samples consisted of “secondary” and “important minor” plant taxa. Colonies collected 37.5% of the sampled pollen from herbs and 58.0% from trees and shrubs. We identified pollen from 48 plant taxa belonging to 34 families collected in the summer. This included a “secondary” pollen taxon and “important minor” taxonomic groups. Colonies collected 41.3% of the pollen from herbs, while 58.5% came from trees and shrubs. The Shannon-Weaver diversity index for each site in the summer was 1.03, with an ENS of 3.09. In the fall, bees collected pollen from 33 taxa belonging to 24 families. The samples included “secondary” and “important minor” taxa. Colonies collected 27.1% of the pollen from herbs and 72.4% from trees and shrubs. The average Shannon-Weaver diversity index for each site in the fall was 0.69 and the ENS was 2.14. Finally, there were 33 taxa belonging to 23 families in the winter , which belonged to, there was a “secondary” pollen taxon and a few “important minor” taxa. Colonies collected 18.8% of the pollen from herbs and 69.7% from trees and shrubs. The Shannon- Weaver diversity index for each site was 1.14 and the ENS was 3.46. ## Floral diversity in Texas We identified pollen belonging to 49 taxa representing 36 families in the spring. There was one “secondary” taxon in the Anacardiaceae family, which was also the most abundant plant family found in all the samples. We also found “important minor” plant taxa, as well one important unknown “minor” taxon, which had grains with a tricolporate structure. Colonies collected 9.5% of the pollen from herbs, while 78.5% came from trees and shrubs. The Shannon-Weaver diversity index for each site in the spring was 0.82, while the ENS was 2.27. We found pollen from 20 plant taxa belonging to 15 families in the summer. We found one “predominant” taxon, as well as “important minor” taxa. Colonies collected 4.2% of the pollen from herbs and 90.1% from trees and shrubs. The Shannon-Weaver diversity index value for each site in the summer was 0.43 and the ENS was 1.61. In the fall, we found pollen from 15 plant taxa belonging to 14 families. These consisted of one “predominant” plant taxon, one “secondary” taxon, and one “important minor” taxon. Colonies collected 27.6% of the pollen from herbs and 72.4% from trees and shrubs. The Shannon-Weaver diversity for each site in the fall was 0.61 and the ENS was 1.98. We did not collect any pollen from November to February in Texas. ## Floral diversity in Florida We identified 36 plant taxa belonging to 30 families in the spring, consisting of one “secondary” taxon, and a few “important minor” taxa. Colonies collected 12.4% of the pollen from herbs and 83.3% from trees and shrubs. The average Shannon-Weaver diversity index for each site in the spring was 1.14, while the ENS was 3.57. Pollen identified in the summer belonged to 29 plant taxa and 22 families. These consisted of one “predominant” and one “secondary” plant taxon. Colonies collected 5.0% of the pollen from herbs, while 94.3% came from trees and shrubs. The Shannon-Weaver diversity index for each site in the summer was 0.76 and the ENS was 2.36. In the fall, we found pollen from 34 plant taxa representing 27 families. There were no “predominant” plant taxa represented in the pollen samples, but we did find “secondary” and “important minor” taxa, as well as one “minor” taxon. Colonies collected 32.3% of the pollen from herbs and 62.4% from trees and shrubs. The Shannon-Weaver diversity index for each site in the fall was 0.76 and the ENS was 2.14. Finally, there were 31 different plant taxa in 23 families in the winter , which belonged to “secondary” and “important minor” taxa. Colonies collected 6.9% of the pollen from herbs, while 90.6% came from trees and shrubs. The Shannon-Weaver diversity index for each site was 0.72 and the ENS was 2.28. ## Floral diversity in Michigan We identified pollen from 39 taxa and 27 families in the spring, which belonged to “secondary” and “important minor” plant taxa. Colonies collected 30.3% of the pollen from herbs and 61.3% from trees and shrubs. The Shannon-Weaver diversity index for each site in the spring was 0.96 and the ENS was 2.61. We found 29 plant taxa from 20 families in the summer, which consisted of “predominant” and “important minor” taxa. Colonies collected 77.4% of the pollen from herbs and 19.6% from trees and shrubs. The Shannon-Weaver diversity index for each site in the summer was 0.76, while the ENS was 2.32. Pollen collected in the fall belonged to 20 taxa from 12 distinct families. All samples belonging to the Asteraceae family were grouped together because we could not identify them to a lower taxonomic level. Combined, all pollen in the Asteraceae family (55.23%) was considered a “predominant” plant taxon. There were also “important minor” taxa as well as one “important minor” taxon belonging to an unknown taxonomic group with a tricolporate structure, which resembled pollen in the genus *Salix*. Colonies collected 89.9% of the pollen from herbs, while 3.2% came from trees and shrubs. The Shannon-Weaver diversity index for each site in the fall was 0.69 and the ENS was 2.07. Pollen was only sampled in September during the fall, and no pollen was collected in the winter. # Discussion We conducted a one-year survey of the floral sources of pollen foraged by managed honey bee colonies in urban and suburban environments in four different regions of the U.S. Using standard melissopalynological techniques, we identified pollen pellets collected by foragers in these environments at spatial and temporal scales. Overall, the “predominant” and “secondary” pollen sources collected by foragers across all states originated from plants belonging to various genera in the Fabaceae (legumes), Anacardiaceae (sumac), Lythraceae (loosestrife), Arecaceae (palm), Asteraceae (daisies and asters), Fagaceae (oak), Sapindaceae (soapberry), Rhamnaceae (buckthorn), Salicaceae (willow), Myrtaceae (eucalyptus), Rosaceae (rose), and Brassicaceae (mustard) families. There was also a large representation of pollen from trees and shrubs for colonies in TX, FL, and CA, including those in the families Lythraceae, Arecaceae, Fagaceae, Salicaceae, Myrtaceae, and Rosaceae. These types of trees and shrubs have previously been shown to be important pollen sources for pollinators in urbanized areas. Interestingly, the majority of pollen collected in MI in the summer and fall came from herbaceous plants rather than trees and shrubs, indicating the importance of herbs as honey bee forage in this region of the U.S. California was the state with the highest number of taxa identified in a given season (n = 64 in the spring). Despite this taxonomic abundance, the relatively high Shannon-Weaver diversity index for that season (1.21) suggests that there was low species evenness at every sampling period in each colony sampled. This indicates that foragers did not collect pollen evenly during a particular sampling period, and instead, likely favored collecting pollen from a few floral resources in each collection. This is consistent with previous work showing that, even in cases when there is a large diversity of plant resources available, honey bees tend to focus their foraging efforts on a few species \[, \] because the preferred sources are more abundant, or because they provide specific nutrients that colonies need at a particular time. Furthermore, honey bees are known to survey their surrounding area and collectively forage from a few species of plants at a given time until the resource is near exhaustion. The range in the observed Effective Number of Species (ENS) across all states and seasons ranged from 1.61 plant taxa in MI during the summer, where pollen in the family Fabaceae was the most “predominant” type, to 3.75 plant taxa in CA during the spring, where most of the pollen collected came from the families Rosaceae, Brassicaceae, Myrtaceae, and Fagaceae. Our study found that the highest diversity of plant taxa foraged occurred in the spring. This result was different from studies in Europe, which found that the highest plant diversity around agricultural landscapes occurs in the summer.These differences can be attributed to the season, differences in landscape characteristics among study regions, the number of samples collected per site, the lowest floral taxonomic level achieved during pollen identification, or the overall foraging activity of workers. In particular, warmer months in Europe do not occur until later in the year compared to the sites in our study. For example, temperatures in Germany range between 8 ˚C and 18 ˚C in May, which is similar to the temperature in the majority of our sites in the late winter through early springs months. The average latitude across those European studies, was approximately 48.35° N, whereas the average latitudes for our four states in the U.S. was 35.32° N. In comparison, the amount of pollen collected by honey bee colonies in Israel, which is at a similar latitude as the U.S. (around 31.77° N), was highest from early spring to mid-summer. The low plant taxonomic diversity observed during the summer months in MI is likely due to the abundance of plants grouped in the Fabaceae family, especially in the *Trifolium* and *Melilotis* genera. Even though a colony as a whole may be collecting pollen from several sources of plants over a longer period of time, individual workers typically exhibit temporary specialization and floral constancy for a specific pollen source, with 52–79% of the pollen that a colony collects in a week belonging to a single plant species. With a maximum pollen collection period of one week, our methods provided only a limited preview of what colonies are collecting at a given time, as bees could have returned with pollen from completely different plants in the days or weeks before or after sample collection. At times, volunteer beekeepers turned on their traps only a few days before the pollen was sampled. Despite the short window for pollen sampled each month, as well as a colony’s temporary specialization in collecting pollen, we still observed spatial and temporal differences in the diversity of plants foraged by our focal colonies. However, seasonality may have a larger effect on the diversity of pollen collected by foragers compared to landscape level effects, such as the level of urban development in the surrounding landscape of a colony. The spring provided the highest taxonomic richness of pollen in every region, with plants in the families Anacardiaceae, Rosaceae, Rhamnaceae, and Arecaceae representing the majority of pollen collected in TX, CA, MI and FL, respectively. Summer pollen in MI was predominantly represented by genera in the Fabaceae family, particularly plants in the *Trifolium*, *Melilotus*, *Lotus*, *Gleditsia* and *Mimosa* genera, which are common invasive herbs and shrubs found in urban and semi-urban areas. The plant genera found in MI are highly attractive to honey bees, but are considered invasive plants in that state. As the season progressed into the fall, colonies in MI predominantly foraged for pollen in the Asteraceae family, which was also significantly represented as a “secondary” pollen type in CA, TX, and FL. A similar trend of honey bees in the Northern Great Plains transitioning their foraging effort from plants in the Fabaceae family to plants in the Asteraceae family was observed previously by Smart et al.. In CA and FL, where colonies were sampled in the winter, foragers collected pollen mostly from early-blooming trees and shrubs in the Fagaceae, Aceraceae, Myrtaceae, Rosaceae, and Salicaceae families. Only a few plant groups provided pollen reliably throughout the year in all states. For example, plants in the Arecaceae and the Myrtaceae (*Eucalyptus* spp.) families provided a reliable pollen source for colonies the entire year in FL and CA. Plants in the *Eucalyptus* genus and the Arecaceae family are known to be commonly foraged by honey bees and stingless bees in other regions, including Brazil and the West Indies. Pollen from plants in the Arecaceae family tends to contain high levels of protein, essential amino acids and minerals. However, the antioxidant activity of Arecaceae pollen tends to be low when compared to that of other pollen types. Meanwhile, *Eucalyptus* is generally regarded as a nutritionally poor pollen type because it has a low content of lysine (an essential amino acid) and in omega-3 fatty acids, which are positively associated with improved honey bee learning and memory. Crape myrtle, which belongs to the genus *Lagerstroemia* (Lythraceae) was represented as a “predominant” plant type foraged in TX and FL, and was a “secondary” plant type in CA. A deciduous shrub native to India, crape myrtle is a common ornamental plant in urban environments that displays dimorphic pollen with two distinct whorls: one that provides food to pollinators, and one that is used for plant fertilization. Interestingly, although crape myrtle pollen has been previously found in honey from a few states, its presence in honey is likely a spillover effect resulting from “cross-contamination” from pollen foragers, given that this plant lacks nectaries and thus, it is not a source of nectar for bees. Crape myrtle has been documented as a pollen source for native and non-native bees in other studies. In fact, crape myrtle may be undervalued for its contributions to honey bee nutrition in the late summer months, a period when colonies can experience severe resource dearth. Therefore, the “predominant” presence of crape myrtle pollen we observed in the summer suggests that some of our study colonies relied on the presence of this ornamental plant in urban environments for pollen acquisition. Maintaining a reliable flow of pollen when brood rearing is critical for nutrition. In a heterogeneous environment containing food sources varying in quality and quantity, honey bees adopt different foraging strategies to regulate their nutritional state. For example, honey bees prefer certain essential amino acids and fatty acids over others, as well as certain diets that complement previous nutritionally deficient diets. In our study, we qualified the types of plants foraged by bees in urban environments using a method of categorization previously described, and addressed the spatial and temporal differences in pollen diversity. However, since we did not quantify the relative amount of pollen brought in from each floral source, our results can neither infer a colony’s nutritional state nor determine whether or not colonies in urban and suburban environments are pollen limited. Avni et al. found spatial and temporal differences in pollen quantity and content by measuring colony pollen intake and quantifying the macronutrients therein for an entire year. Despite these differences, the authors did not find any nutritional effects of pollen quantity and quality on overall colony growth on a yearlong basis. As generalist foragers, honey bees collect resources from a wide range of plants and can compensate for nutritional deficiencies by collecting complimentary diets. Future studies to address this question further should incorporate the use of The Geometric Framework, which explores how organisms (including individuals and colony-living “superorganisms”) balance their nutrient intake on a multidimensional nutrient space. Our main objective in this study was to identify the “predominant,” “secondary,” and “important minor” plant taxonomic groups from which honey bees collect pollen in urban and suburban areas. One limitation was that not all pollen types were identified to the species level, in part because the traditional light microscopy method we used to identify pollen based on grain morphology is a time-consuming process that requires a high level of expertise. Depending on the microscope and time constraints of a desired study, Scanning Electron Microscopy (SEM) could be used in conjunction with light microscopy to visualize certain morphological features that are unclear with light microscopy alone, providing better taxonomic resolution to the genus or species level. Newer technologies such as DNA meta-barcoding, which screens pollen samples for specific plant genome sequences, are currently being explored for pollen identification. However, this method is still in development, as it does not adequately provide the relative abundance of each pollen type, and it can provide “false positives” in the plant identification results. A combined use of microscopy and molecular techniques would likely provide the most thorough and reliable information when performing a palynological analysis of floral origins and pollen grain abundance in a given sample. Although we only sampled pollen in a limited time frame within each month, our results are representative of the plant sources of pollen preferred in developed environments across different seasons and geographic regions. With the ever-changing composition of landscapes, our results can help us better understand honey bee nutritional ecology in urban and suburban environments, and can aid in promoting the use of plants that provide appropriate pollen resources to honey bees in developed areas. In addition, the data generated in this study could be used by public and private urban gardeners to aid them in the selection of pesticides used on ornamental and landscaped plants, especially in terms of the timing and application rates of products used around homes and gardens. For example, pesticide treatment regimens should be planned to avoid chemical application when bees are primarily foraging from a given plant for pollen. Similar studies in other regions of the country are encouraged to help us better understand, and potentially improve, the foraging conditions available to honey bees in urban environments throughout the year. # Supporting information We would like to thank the collaborating beekeeping groups in California, Florida, Michigan and Texas for helping us collect pollen samples from their colonies. We thank Alexandria Payne and Vaughn Bryant’s staff for lending their palynological identification expertise to identify pollen types. We thank Brandi Simmons and Savannah Nease for laboratory help and pollen collection in Florida, Xianbing Xie and Shudong Luo (both MSU) for pollen collection in Michigan. [^1]: We have the following interests: This study was funded in part by Bayer Crop Science and Syngenta Crop Protection LLC and by the Texas Beekeepers Association. Daniel R. Schmehl and Ana R. Cabrera are employed by Bayer CropScience LP. Joseph Sullivan is employed by Ardea Consulting. There are no patents, products in development or marketed products to declare. This does not alter our adherence to all the PLOS ONE policies on sharing data and materials.

# Introduction India is the highest tuberculosis (TB) burden country in the world accounting for one fifth of global incidence. In 2009, 1.5 million patients were placed on anti-TB treatment in India, of whom 366,381 (24%) were registered as new sputum smear-negative pulmonary TB cases. According to current Indian guidelines, a patient is labelled as sputum smear- negative pulmonary tuberculosis (PTB) when he/she has clinical features of PTB, has two consecutive sputum smear examinations negative for acid-fast bacilli (AFB) and has radiographic abnormalities consistent with active PTB, as determined by a medical officer. A patient whose sputum smears are negative for AFB but who has *Mycobacterium tuberculosis* identified on culture is also designated as sputum smear negative PTB. Such patients who are placed on treatment have follow up sputum smear examinations performed twice during the entire course of anti-TB treatment. This includes a first follow-up smear at the end of the intensive phase of treatment (2 months) and a second smear at the end of treatment (6 months). The purpose of these follow up smear examinations is to assess the response to therapy, check on whether there has been a misdiagnosis, and to determine if there is disease progression due to non-adherence or drug resistance. All these factors could result in an initially smear-negative PTB patient becoming smear-positive at the end of the intensive phase or on completion of treatment. The Indian guidelines on following up sputum smear-negative PTB patients differ from the current World Health Organization (WHO) guidelines in that the latter recommends only one follow up sputum microscopy examination, which is done at the end of the intensive phase.This made us to ask; is there added value in performing an additional sputum smear examination at the end of treatment within the context of a national TB program? In a cohort of newly registered sputum smear-negative PTB patients, we determined i) the proportion of patients who had sputum smear examinations done at the end of the intensive phase and at the end of treatment, ii) the number and percentage of patients with positive sputum smears at these time points and iii) the treatment outcomes of those patients who were found to be sputum smear- positive. # Materials and Methods ## Design This was a descriptive study using routine programme data. ## Study setting The study was conducted in the state of Delhi, the National Capital of India. Delhi has an area of 1483 sq km, with a total population of 17 million and population density of 11,000/sq km, with 40% of the population living in slums or similar contexts. The Revised National TB Control Programme (RNTCP) has been implemented in all districts of Delhi since 1997. For the management of RNTCP, the state has been divided into 24 Chest Clinics. Under each Chest Clinic, there is one Tuberculosis Unit (TU) for half a million population having a Designated Microscopy Centre for every 0.1 million population. The 24 Chest Clinics under RNTCP fall within the nine Revenue Districts of Delhi State. Patients were registered in the programme in one of three treatment categories: details of these categories and the treatment regimens are presented in. At the end of treatment, one of several mutually exclusive treatment outcomes is given to patients as shown in. Most of the new sputum smear-negative pulmonary TB patients were treated with a category 3 regimen, but some with serious illness were treated with a category 1 regimen. The patients were treated using the Directly Observed Treatment Short course (DOTS) approach, and were supervised regularly thrice a week in the intensive phase and once a week during the continuation phase. They were followed with sputum smear examination at the end of the intensive phase (2 months) and on treatment completion (6 months). Sputum smears were examined in quality assured laboratories and results were recorded in the TB register. If patients failed on treatment (i.e., were found to be sputum smear positive at 5 months or later), they were changed to a retreatment category 2 regimen. Such patients were also considered to be suspects of MDR-TB (Multi-Drug Resistant TB-resistant to rifampicin and isoniazid) and their sputum specimens were sent to an accredited laboratory facility for culture and drug sensitivity testing. ## Study population From each of the nine revenue districts of Delhi, if there were more than one tuberculosis units, the TB unit that registered the maximum number of patients during the study period was selected for inclusion in the study. All consecutive new sputum smear-negative pulmonary TB patients registered in these nine TB units from 1^st January 2009 to 31^st December 2009 were included in the study. ## Data collection and analysis Data were obtained from the TB registers. A data extraction sheet was used by the study investigators to capture details on sputum smear examination and treatment outcome status at the end of the intensive phase and on treatment completion. Cross validation of the records was undertaken for 10% of smear- negative Pulmonary TB patients by examining patient treatment cards and TB Laboratory registers. Data were entered and analyzed using a Microsoft Excel spreadsheet Version 2007. ## Ethics approval Ethical approval was obtained from the Institutional Ethics Committee of Public Health Foundation of India, New Delhi and the Union Ethics Advisory Group, Paris. The study was a retrospective review of routine programme records and reports, and permission was obtained from the programme managers at the state and national levels to access these data. Individual patient consent was deemed un-necessary by both the ethics committees. Electronic databases created for the analysis were stripped of personal health identifiers and maintained securely. # Results ## Sputum examinations during and at end of treatment A total of 2,567 new sputum smear negative PTB patients were registered in the 9 TB units in Delhi. The management of these patients and the numbers who were eligible for and had sputum smear examination performed from diagnosis until the end of treatment is shown in. ## Smear status at the end of the Intensive phase of anti-tuberculosis treatment Of all newly registered patients, 367 (14%) were not eligible for sputum smear examination at end of the intensive phase due to various reasons shown in, Of those eligible for sputum smear examination at this point, 227 (10%) failed to have the examination. There were 1973 patients whose sputum smears were examined, of whom 36 (2%) were sputum smear-positive. Of the latter, 25 were receiving category 1 treatment and 11 receiving category 3 treatment. The treatment outcomes of patients who were smear-positive at the end of the intensive phase are shown in. Amongst the category 1 patients who were found to be smear-positive (25), all continued their same treatment regimens and the majority successfully completed treatment. There were two patients who were changed to a retreatment category 2 regimen, of whom one was later diagnosed with multi-drug resistant TB (MDR-TB, i.e., TB resistant to both isoniazid and rifampicin) and switched over to category 4 treatment for its management. Of the patients who were receiving category 3 treatments and were smear-positive at the end of the intensive phase (11), all were changed to a re-treatment category 2 regimen. The majority of these patients had a successful treatment outcome. ## Smear status at the end of anti-tuberculosis treatment During the continuation phase of treatment, 111 (5%) patients were not eligible for sputum smear examination at the end of treatment (6 months) due to various reasons shown in. Of those eligible for sputum smear examination, 323 (15%) did not have the examination done. There were 1766 patients whose sputum smears were examined at the end of treatment, of whom 16 (0.9%) were sputum smear-positive. Amongst these smear-positive patients, 6 were receiving category 1 and 10 were receiving category 3 treatment regimens. Of the 16 smear-positive patients, 12 were changed to a category 2 retreatment regimen. Two other patients from each category were lost to follow up after their treatment was completed and their outcome was registered as treatment completed. The outcomes of the 12 patients who were changed to a retreatment regimens shown in. Three of these 12 patients were later diagnosed with MDR-TB and were switched to a MDR-TB treatment regimen for its management. ## Sensitivity analysis and costs for sputum smear at end of anti-tuberculosis treatment A sensitivity analysis was performed to check what the estimate would have been if all of the patients who did not get their last follow up sputum smear examination turned out to be smear-positive or smear-negative. If all the missing patients had been smear-positive, the smear-positivity rate at the end of treatment would have been 16%. Conversely, if all these patients had been smear-negative, then the estimate would have been 0.8%. Even with the low estimate of 0.8%, from a public health perspective, the absolute number in terms of cases of potentially resistant TB cases identified is still significant. The cost of undertaking a single sputum smear examination in our setting is US \$ 0.90. In this study, 1766 patients at the end underwent sputum examination with 15 patients experiencing a change in either treatment category, yielding a Cost Benefit Ratio (CBR) of US \$ 106/ treatment regimen changed. If all eligible patients in the study cohort (2089) had undergone the last sputum smear examination and all the missing patients were smear positive resulting in a change in treatment regimen, the CBR would have been US \$ 5/ treatment regimen changed. On the other hand, if all the missing patients were smear-negative, CBR would have been US \$ 125/ treatment regimen changed. # Discussion This is one of the first studies evaluating the added value of sputum smear examinations at the end of the intensive as well as the end of treatment in new sputum smear negative PTB patients. The majority of patients who were eligible to submit sputum specimens at the end of the intensive phase of treatment and at the end of treatment had their sputum smears examined. Smear-positivity was 2% for the smear examination at the end of the intensive phase and 0.9% at the end of treatment. These percentages may seem low, but in India with an estimated 366,381 annual new smear negative PTB patients registered in 2009, this translates to over 7000 cases with positive smears at 2 months and 3297 cases of smear positive PTB at the end of treatment. For patients on the category 3 regimen, all of those with smear-positive sputum at 2 months were changed to a retreatment regimen with a generally successful treatment outcome, so action was taken on the result. Action was not so clear for those found to be smear- positive at the end of treatment. A quarter of patients were incorrectly registered as treatment completed. The remainder was changed to a re-treatment regimen with mixed results, but three patients were identified with MDR-TB and changed to the appropriate treatment, an action which would not have occurred if the smear examination had not been performed. There are important public health implications from these findings. First, persistent smear positivity at the end of treatment completion might herald misdiagnosis, failure on current drug regimens or primary or acquired drug resistance. In particular, the possibility of persistent smear positivity heralding MDR-TB is of major concern as this form of drug resistance needs early identification in order to limit its spread to households and the community at large. Such cases will add to the “difficult to manage” TB cases and the already high background burden of disease. Second, from a patient perspective, this failure of treatment is likely to impact on individual survival. It is not known what happened to the patients with smear-positive sputum at the end of treatment who were incorrectly declared successfully treated or what happens in general to these patients if the smear examination is not performed. However, it is likely that outcomes are compromised. One consideration is that our study group included patients in category 3 with a three-drug treatment regimen in the intensive phase in addition to patients in category 1 who were treated with a four-drug treatment regimen in the intensive phase. The category 3 regimen is considered a relatively inferior regimen and no longer exists in the country. It is possible that the country-wide shift to a category 1 regimen for all smear-negative PTB patients might affect our study estimate of 1%. It is thus urgent that RNTCP audits current data under the revised national guidelines to fully validate these findings. The strengths of this study are that a large number of patients were studied, and we used routine programme data. The findings are therefore likely to reflect the operational reality on the ground. We have also adhered to the STROBE guidelines on reporting. The study has a number of limitations. First, the use of record reviews does mean that sometimes the data are inaccurately recorded. Second, our estimate of the added value of 0.9% positive sputum smears at the end of treatment might be affected by the 15% of patients who for various reasons did not undergo the last follow up smear examination at the end of treatment completion. A large number of patients not undergoing follow-up sputum smear examination is not unique to this study and/or region and has also been highlighted in another study from a different region in the country. Even with a low estimate of 0.8%, obtained through the sensitivity analysis presented in the results, from a public health perspective, the absolute number in terms of cases of potentially resistant TB cases identified is still significant. Third, sputum smear microscopy as a tool of diagnosis is challenged by the issue of false positive and false negative results, which may occur even within quality assured programmes. However, sputum smear microscopy is the mainstay of diagnosis programmatically in India, and is used to identify patients who may be at risk of drug-resistant disease during follow up and who may need additional culture and drug sensitivity testing. Finally, we were unable to ascertain the number of patients who were sputum smear positive in follow up smears and were concomitantly HIV-seropositive as during the study period HIV testing was not routinely performed. This is the first report of its kind coming from a high TB burden country like India, but like many studies it raises additional questions. First, the RNTCP has stopped giving category 3 treatment, and therefore a similar study with the same research question should be conducted for another cohort of patients receiving the changed treatment regimens. Second, a prospective study should be done for a similar cohort of patients as this would obviate a number of the limitations inherent in this retrospective record review study. A prospective study could also include sputum culture and drug sensitivity testing on all specimens as well as a detailed cost-analysis on all aspects. The decision to continue sputum smear examination at end of continuation phase will also be dependent on CBR of the additional sputum smear examination. Given the political commitment by the Government of India to support the RNTCP with the aim of improving its performance over several years, the current costs of performing the additional end-of-treatment smear as presented in the results section, do not seem to impose an economic challenge to the programme. The end of smear examination in patients with smear-negative pulmonary TB must be assessed in relation to other strategies for improving case detection as part of the country's TB control efforts; for example, comparisons made with contact investigation of index patients where the yield may be in the region of 2–5%. In light of above and the evidence generated through this study, there is need for further evaluation on this research question. End-of-treatment smear is a low-yield strategy for detection of smear-positive TB cases, although further studies are needed to determine its population-level impact and cost, particularly in relation to other TB control interventions. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SM SZ SS ADH. Performed the experiments: SM SS SC RPV. Analyzed the data: SM SZ SS SC RZ ADH. Contributed reagents/materials/analysis tools: SM SZ SS SC RZ ADH. Wrote the paper: SM SZ SS SC RZ ADH.

# Introduction The 2009 IUCN Red List includes 65 species of plants and animals that are officially extinct in the wild, many of which continue to persist in captivity. These captive relicts of species lost from their native ranges are increasingly common, subject to intensive conservation management to prevent outright extinction. The Galápagos tortoises represent a group of 11 extant species (*Chelonoidis spp.*;, see for description of recognized taxonomy), many of which are imperiled and the object of extensive *in situ* and *ex situ* conservation efforts ranging from control of poaching, protection of habitat, head-starting of *C. ephippium* on Pinzon Island, and captive breeding and repatriation of *C. hoodensis* to Española Island. Previous genetic surveys investigating the origin of captive individuals of unknown ancestry provided managers critical historical information for maintaining the integrity of distinct lineages. These studies collectively examined 156 individuals of unknown ancestry held in captive populations on three continents, assigning them to the species level and, in many cases, to their population of origin. Not surprisingly, the majority of individuals assigned to tortoise populations on Santa Cruz and Isabela islands that are easily accessible and have been historically harvested. Fifteen individuals, however, were assigned to critically endangered species (e.g. *C. ephippium*, Pinzón Island) or to natural populations of known mixed ancestry (e.g. Volcano Wolf, Isabela Island). Yet, the power of population assignment approaches is fundamentally linked to the underlying reference population database. If the population of origin of an individual is not represented in the sampled set of reference populations, the assignment algorithms will still designate a population of origin, albeit an incorrect one. When additional reference population data become available either through expanded sampling across space (e.g. broader geographic coverage of contemporary distribution) or time (e.g. recently extinct species or extirpated populations), reanalysis of population assignments for individuals of unknown ancestry may be warranted. Such reanalyses may be particularly important when research questions have direct relevance to on-going conservation strategies. It has been well-publicized that the Pinta Island tortoise *C. abingdoni* is extinct in the wild (currently represented by a solitary male in captivity, Lonesome George), yet another species endemic to Floreana Island (*C. elephantopus*) was already extinct at the time of Van Denburgh's taxonomic revision in the early 1900's. A recent study reported living descendents of the extinct *C. elephantopus* on the neighboring island of Isabela, and suggested that the removal of multiple individuals may aid in the establishment of a captive breeding program and eventual reintroduction to Floreana. The population-level mitochondrial DNA (mtDNA) haplotypic and microsatellite genotypic data collected for *C. elephantopus* by way of historical DNA analysis of museum specimens added a critical reference population to the existing database of extant species for investigating the origin of individuals of unknown ancestry. In this study, we reanalyzed mtDNA haplotypic and microsatellite genotypic data for 156 captive individuals relative to the expanded reference population database that now includes the extinct *C. elephantopus* from Floreana to test hypotheses of ancestry set forth in Burns et al. and Russello et al.. Here we report the identification of individuals of recent Floreana ancestry that currently reside in a captive population in Galápagos. We further examined the relatedness of these individuals and discussed their utility for serving as a nucleus for re-establishing tortoises on Floreana Island that have now been absent for over a century. # Results Our sample of 156 captive individuals were assigned to their population(s) of origin based on mtDNA haplotypic and microsatellite genotypic data relative to reference databases including all extant species and the extinct species from Floreana. As revealed in earlier studies by Burns et al. and Russello et al., all but two reanalyzed individuals possessed haplotypes originally sampled from species on Isabela (62.8%) or Santa Cruz (35.9%) Islands. Two individuals (PRZ01, CDRS037) exhibited haplotypes from Pinzon and San Cristóbal Islands, respectively. Interestingly, 13 individuals possessed northern Isabela haplotypes sampled at the Puerto Bravo and Piedras Blancas sites previously shown to cluster phylogenetically with haplotypes from other *Chelonoidis* species on Española, San Cristóbal and southern Isabela as well as Floreana. The genotypic assignment tests of Rannala and Mountain and Pritchard et al. exhibited a high degree of overlap, yielding consistent species assignments for 78.8% of the individuals sampled. Overall, the genotypic assignments corroborated the results obtained from the mtDNA analyses, with 129 individuals (82.6%) consistently assigned to the same locality by both datasets. The other 27 individuals exhibited patterns of mixed ancestry. Specifically, nine of 27 such individuals were alternatively assigned to different species on Isabela Island. The remaining individuals were all assigned to the La Caseta *C. porteri* population on Santa Cruz or *C. becki* populations on Volcano Wolf by way of mtDNA, but assigned to different species according to their multi-locus genotypes. The high degree of mixed ancestry detected was not surprising, as 36 individuals from the Santa Cruz breeding facility were sampled from a known “progeny” pen. These individuals are direct descendents of founders from multiple Galápagos tortoise species that were housed together in a group enclosure prior to knowledge of their origin and taxonomic assignment. Of immediate interest, nine captive individuals exhibited congruent signatures of Floreana ancestry, one of which (CDRS047) also possessed a “Floreana-like” mtDNA haplotype. The remaining eight individuals with nuclear DNA assignment to Floreana, including the two females (CDRS106 & 107) currently housed with Lonesome George, possessed an “Española-like” haplotype only sampled in Puerto Bravo on northern Isabela Island. The Puerto Bravo population hosts the living descendents of the near-extinct *C. abingdoni* (Pinta) and extinct *C. elephantopus* (Floreana) previously detected by Russello et al. and Poulakakis et al.. These findings were consistent with historical records and anecdotal accounts, as at least two of the nine individuals of Floreana ancestry (CDRS106, CDRS107) were originally captured from the wild population in Puerto Bravo, while no less than two additional females were collected from unspecified locations on Isabela in 1966 and subsequently housed in the parental pen (M. Castro, pers. com). The triangle plot in depicts a fine-scale examination of the history of mixed ancestry in the nine captive individuals that assigned to Floreana, obtained through *q*-value distributions of 500 simulated genotypes each of parental, F1 hybrids, F2 hybrids, and B2 and B3 backcrosses for all pairwise comparisons between Puerto Bravo *C. becki* (Isabela), *C. hoodensis* (Española), and *C. elephantopus* (Floreana). One individual (CDRS 40) falls distinctly within the Floreana parental *q*-value distribution, with five others exhibiting strong Floreana ancestry within the Española-Floreana F1 hybrid distribution. Three additional individuals clustered within the Puerto Bravo-Floreana F1 hybrid *q*-value distribution with varying affinities to Floreana. Although these results are clearly indicative of some degree of Floreana ancestry for all nine individuals, additional loci will be necessary to further discriminate between F1 and higher-order hybrids and backcrosses for many of them. There is a high degree of relatedness among the CDRS individuals exhibiting signatures of Floreana ancestry \[mean pairwise relatedness (*r_xy*) = 0.15\]. Overall, the observed distribution of pairwise relatedness values among the CDRS individuals of Floreana ancestry overlaps substantially with simulated second order (half-sibling) and first-order (full- sibling, parent-offspring) distributions. At the individual level, three of the females (CDRS042-044) housed in the CDRS “parental” pen exhibit pairwise relatedness values consistent with full-sibling relationship (*r_xy* = 0.40−0.61), while a fourth (CDRS040) appears to be their half-sibling (*r_xy* = 0.31−0.37). None of the females of Floreana ancestry housed in the CDRS “parental” pen possess genotypic profiles consistent with maternity for any living individuals in the program. Yet, three of them (CDRS042-044) are likely grandmothers, exhibiting *r_xy* ranging from 0.23–0.31 with at least one individual of Floreana ancestry in the CDRS “progeny” pen. Of particular note, CDRS047, the male with congruent mtDNA and nuclear DNA assignment to Floreana, is the likely half-sibling of CDRS044 co- housed in the “parental” pen, consistent with genotypic pairwise relatedness (*r_xy* = 0.21) and the discrepancy in mtDNA haplotypes. # Discussion Broad application of DNA analysis of archival material (e.g. museum specimens) has provided critical spatial and temporal components to ecological, evolutionary, taxonomic and conservation-related research. A particularly powerful application of historical DNA analysis for informing *in situ* conservation has been enabling direct incorporation of extinct taxa in comparative studies with extant forms, whether involving “rediscovery” of presumed extinct species, refinement of evolutionary relationships or identification of cryptic diversity. Rarely has historical DNA analysis helped inform *ex situ* species recovery efforts as has been demonstrated here with the identification of the extinct *C. elephantopus* already in captivity. In the current study, we identified six females and three males of mixed ancestry that exhibited high assignment probabilities to the extinct Floreana species. All of these individuals are currently housed at a single breeding facility on Santa Cruz Island in Galápagos, allowing them to play a critical role as founders of a selective captive breeding program for resurrecting *C. elephantopus* without additional transport or disease transmission concerns. Backcrossing as a species restoration technique has long been considered but rarely implemented, especially in long-lived organisms such as Galápagos tortoises. Although time consuming and resource intensive, there is precedent for successful breeding and repatriation in another species of Galápagos tortoise (*C. hoodensis*) endemic to Española. Since the program's inception in 1975, over 2000 individuals have been repatriated to Española originating from 15 initial founders, assisting in population recovery with demonstrated *in situ* breeding. In addition to the nine captive individuals identified in the current study, a recent field expedition to Vólcan Wolf on northern Isabela in December 2008 sampled and tagged over 1600 individuals in an attempt to identify individuals of pure or mixed Floreana ancestry to further populate a breeding and repatriation program. Additional founders will be important for maintaining the genetic health of a Floreana breeding program given the high degree of relatedness among existing CDRS individuals. Like the recent rediscovery of the Tasman booby, this work generally demonstrates the benefits of integrating historical DNA data with more conventional population genetic approaches for elucidating evolutionary patterns and processes. For example, in the absence of population genetic data from the recently extinct *C. elephantopus*, nine Galápagos tortoise individuals of substantial conservation value were previously misassigned to extant species of varying conservation status. This enhanced ability to collect and analyze genetic data from recently extinct species represents a continued expansion of the conservation biologist's toolbox, in this case within an *ex situ* context, to inform strategies for recovering species diversity. # Materials and Methods ## Taxonomy The taxonomy of Galápagos tortoises has changed repeatedly since they were first described formally in 1824. Pritchard provides a thorough account of the history of Galápagos tortoise taxonomy. Currently, fifteen formally described taxa of Galápagos tortoises are generally recognized, 11 of which are extant and threatened by human activities and introductions of non-native species. These taxa have been described as full species of *Geochelone*, as well as subspecies of *Geochelone nigra*. A recent taxonomic revision recognizes all Galápagos tortoise taxa as subspecies of *Chelonoidis nigra*, a genus that now includes all South American tortoise species. Here, we continue to recognize the full species status of many of these taxa that is most consistent with the overwhelming morphological and molecular evidence, but adopting the genus-level revision to *Chelonoidis*. ## Sampling The 156 individuals reanalyzed were originally sampled in Burns et al. and Russello et al., all of which were of unknown ancestry at the time of collection from the following institutions: Caloosahatchee Aviary and Botanical Garden, Florida, USA (CABG; n = 25); Galápagos National Park Service Breeding Facility, Santa Cruz, Galápagos (CDRS; n = 60); mainland Ecuador hotels, universities, zoological and private collections (ECU; n = 29); former Wittmer Collection on Floreana, Galápagos \[WCF ; n = 29\]; Prague Zoo, Czech Republic (PRZ; n = 2); San Diego Zoo, USA (SDZ; n = 7); and Zurich Zoo, Switzerland (ZUZ; n = 4). The CDRS sampling includes 58 individuals originally analyzed by Burns et al., 23 of which were sampled from the “parental” pen with the remainder housed in “progeny” pens. Two CDRS females that are currently housed with Lonesome George (CDRS106 & CDRS107) and all other individuals were originally analyzed in Russello et al.. ## Mitochondrial DNA Analysis A 695 base pair fragment of the mtDNA control region was reanalyzed for all 156 captive individuals (see for GenBank Accession numbers). Degree of sequence similarity was assessed using stand-alone Basic Local Alignment Search Tool ([ftp://ftp.ncbi.nlm.nih.gov/blast/](http://ftp://ftp.ncbi.nlm.nih.gov/blast/)) relative to a database of 119 haplotypes recovered from over 1000 individuals sampled from all extant species of Galápagos tortoises, , as well as museum specimens from the near extinct *C. abingdoni* from Pinta and the extinct *C. elephantopus* from Floreana. All haplotypes are unique to one of the currently described species with the following exceptions: 1) thirteen haplotypes are shared among two or more southern Isabela taxa ; 2) one haplotype is shared between *C. becki* and *C. darwini*; and 3) nine haplotypes that are more closely related to haplotypes sampled in other species on other islands than the populations from which they were sampled in the wild (originally termed “aliens”). See Ciofi et al. for a comprehensive review of previous studies regarding genetic divergence and phylogenetic distinctiveness among the Galápagos tortoises. ## Microsatellite DNA Analysis Genotypic data at nine microsatellite loci \[GAL45, GAL50, GAL73, GAL75, GAL94, GAL100, GAL127, GAL136, GAL263\] previously collected and characterized for all captive individuals, as well as for 332 individuals sampled from all extant species, the near extinct *C. abingdoni* from Pinta and the extinct *C. elephantopus* from Floreana were reanalyzed in the current study. As new analytical approaches for assessing microsatellite data quality have emerged since much of these data were originally collected, we screened the dataset for null alleles using MICRO-CHECKER. Five out of 153 (i.e. nine loci for 17 populations) comparisons showed evidence of null alleles. Given this very low frequency, data at loci exhibiting null alleles in identified populations were removed prior to population genetic analyses. Following this culling, missing data remained minimal throughout the data set (2.8%). Captive individuals were assigned to island, species and, in many cases, population based on their multi-locus genotypes using two different approaches. First, the Bayesian model-based clustering method of Pritchard et al. was employed as implemented in Structure 2.3. Run length was set to 1,000,000 MCMC replicates after a burn-in period of 500,000 using correlated allele frequencies and prior population information following Russello et al.. Membership coefficients (*q*) of the captive individuals in one or more of the reference populations represent the fraction of its sampled genome that has ancestry in that population. In addition, the exclusion-simulation test of the partial Bayesian assignment method of Rannala and Mountain was used to assign individuals to the two closest natural populations where the likelihoods of its genotype occurring were the highest (*L₁* and *L₂*) as implemented in GENECLASS. The exclusion threshold was set to 0.01, relative to a distribution estimated from 10,000 randomly generated genotypes. To evaluate the validity of population assignments and to identify the possible range of *q*-values for potential purebreds and different hybrid classes, a series of simulations were conducted for parental, hybrid, and backcrossed genotypes. Specifically, 500 individuals were simulated for each parental population, as well as for all pairwise combinations of F1 hybrids, F2 hybrids, and B2 and B3 backcrosses. In this case, multi-locus genotypic data collected from population samplings on Floreana Island, Española Island, and Puerto Bravo on Vólcan Wolf on Isabela Island were used as the parental populations for genotype simulations. These simulated datasets were analyzed in STRUCTURE 2.3 using the previously described parameters. Pairwise relatedness \[*r_xy*; Lynch and Ritland \] values were calculated for all CDRS individuals of detected Floreana ancestry using the software iRel implementing the “leave one out” option and using starting allele frequencies based on putatively unrelated individuals in the “parental” pen only. The estimator of Lynch and Ritland was chosen as it has been demonstrated to minimize type II error (ex. full-siblings misclassified as unrelated) relative to other estimators such as Queller and Goodnight, an important consideration when using marker-based relatedness within *ex situ* population management programs aimed at avoiding inbreeding. To visualize the distribution of relatedness among the CDRS individuals of detected Floreana ancestry, the frequencies of observed pairwise *r_xy* estimates for all possible comparisons were plotted with those calculated from simulated distributions of known relatedness (unrelated, half-sibling, full-sibling and parent-offspring) following the approach in Russello and Amato. Specifically, iRel was used to simulate 1000 pairs each of unrelateds, half-siblings, full-siblings and parent- offspring using starting allele frequencies based on putatively unrelated individuals in the “parental” pen only. Lastly, allele transmission patterns were directly examined for CDRS individuals of detected Floreana ancestry to investigate putative maternity and paternity. # Supporting Information We gratefully acknowledge the Parque Nacional de Galápagos, Charles Darwin Foundation, San Diego Zoo, Zurich Zoo, Prague Zoo, Caloosahatchee Aviary and Botanical Garden, and all the participating institutions on mainland Ecuador for originally providing samples. Chaz Hyseni assisted in the lab. George Amato provided valuable insights that informed the manuscript and Sebastian Cruz coordinated sample collection from mainland Ecuador. [^1]: Conceived and designed the experiments: MAR AC. Performed the experiments: MAR NP EB. Analyzed the data: MAR. Contributed reagents/materials/analysis tools: JPG WT JRP AC. Wrote the paper: MAR JPG AC. [^2]: The authors have declared that no competing interests exist.

# Introduction Expert evaluation of brain development is mainly carried out by analysing age- related patterns in sleep electroencephalograms (EEG), represented by different characteristics such as waves, amplitude distributions, and variations over sleep stages, that reflect the non-stationary nature of EEG, see e.g.. For quantitative analysis, EEG data are split into segments within which changes are not significant and EEG can be considered as quasi-stationary signals. The duration of such intervals is typically between 2 and 20 sec. Despite the wide variability of sleep EEG, there have been identified patterns for newborns at different post-conception weeks (ages), that allow experts to evaluate EEG maturity with the accuracy of ±1 week, see e.g.. When brain development is normal, the EEG evaluation typically matches the newborn’s age, whilst in pathological cases the EEG evaluation mismatches the age. The results of evaluations however can be heavily affected by EEG artefacts, noise as well as by the variability of the age-related patterns. One of important patterns for EEG evaluation is the *discontinuity* that is represented by amplitude and frequency changes. An EEG pattern is defined discontinuous if an interval with a voltage above the normal value is interchanged with a period of a low voltage. The discontinuity in EEG of newborns between 28 and 30 weeks contains high-amplitude bursts visible as waves of mixed frequencies. These bursts are interchanged by long low-voltage periods. After 30 weeks, the variability of amplitudes decreases and periods of an uninterrupted EEG activity become longer, and the discontinuity is progressively decreased, see e.g. In practice of EEG evaluation, reference guidances have not been established as the discontinuity is difficult to measure quantitatively, see e.g. Automated estimation of the discontinuity has been attempted with a threshold segmentation technique proposed in. However, a threshold required for such segmentation is heavily dependent on EEG characteristics that widely vary between patients as well as during sleep hours. Adaptive segmentation has been proposed in order to find pseudo-stationary intervals in EEG, suitable for representation and evaluation, see e.g.. A technique that is based on such segmentation has been proposed in to extract a discontinuity feature from sleep EEG. Within this technique detected pseudo- stationary intervals were used for estimating the average amplitudes which then form an Amplitude Vector (AV). Statistics derived from distributions of AV were found correlated with the EEG maturation of newborns between 25 and 35 weeks post-conception. However, these statistics varied largely between patients. An alternative approach, proposed in our previous work, aimed at estimating the EEG discontinuity as a *rate* of pseudo-stationary segments. This technique detected EEG intervals within which the statistics of spectra powers were changed insignificantly. The calculated statistics were compared in adjacent intervals of EEG. The new feature was correlated with newborn age and shown to be capable of increasing the accuracy of classification between pre-term and full-term newborns, respectively. The above work was undertaken within a methodology of Bayesian Model Averaging (BMA) aimed at estimating the full predictive posterior probability distribution that is required for accurate estimation of uncertainty intervals, see e.g.. The use Decision Tree (DT) models within BMA provides selection of predictors that are important for classification, see e.g.. DT models provide experts with new insights into data and interpretation of decision making. A single DT model can be selected for interpretation purposes as shown in our work. The Bayesian averaging over DT models is practically implemented with the Markov Chain Monte Carlo (MCMC) method aimed at exploring a posterior density of model parameters by making random walk proposals, see e.g.. The MCMC methods have been recently applied for modelling and simulation problems in biomedicine, see e.g. including Bayesian analysis of EEG. In this paper we explore the EEG discontinuity feature used along with the spectral power characteristics within the Bayesian classification of newborn development in 10 age groups between 36 and 45 weeks. The proposed technique is compared with the conventional discontinuity techniques based on the threshold and adaptive segmentations in terms of correlation with newborn age, classification accuracy and uncertainty. We also compare our technique with the adaptive segmentation that is based on autoregressive modelling. The rest of the paper is structured as follows. We discuss the techniques of extracting EEG discontinuity features and describe a new approach. Then we describe our methodology and experiments and explore the correlation of the conventional and new discontinuity features with newborn brain maturity. We show that the new features are more strongly correlated with brain maturation. We also compare the new features for Bayesian classification of EEG obtained in 10 age groups in terms of age classification accuracy. Finally we show that the new features provide more accurate assessments of EEG maturation. The provides details of the Bayesian method. # Extraction of EEG features In this section we analyse the feature extraction methods based on adaptive segmentation, that were developed for detecting boundaries of pseudo-stationary EEG intervals. Finally we describe our approach to feature extraction. ## Adaptive segmentation for extracting EEG features In, boundaries of quasi-stationary intervals in a signal *x*(*n*) are detected by using an autoregressive (AR) model given with parameters *ω* for modelling homogeneous parts of the signal *x*. It has been shown that changes in parameters *ω* that are adjusted to different intervals define boundaries of interest. A given AR model generates the outcome $\hat{y}(n,\omega)$ as follows $$\begin{array}{r} {\hat{y}(n,\omega) = \sum\limits_{k = 1}^{p}\omega(k)x(n - k) - x(n),} \\ \end{array}$$ where *ω*(*k*) are the coefficients and *p* is the order of AR model. Signal *x*(*n*) is modelled in the reference and test windows. The modelling errors $e(n) = \hat{y}(n) - x(n)$ are hypothesised to be a white noise process on a homogeneous part of *x*(*n*). Based on the above approach, the errors *e*(*n*) calculated in a window are hypothesised to be distributed as white noise. Such a hypothesis is tested with *Z*-statistic as describe in. The overall *Z*-statistic is combined over the reference, *R*, and test, *T*, windows as follows $$\begin{array}{r} {Z = Z\left( I \middle| J \right) + Z\left( J \middle| I \right),} \\ \end{array}$$ where *Z*(*I*\|*J*) are the statistics of cross-validation errors calculated for windows *I* ∈ {*T*, *R*} and *J* ≠ *I*. The statistics *Z*(*I*\|*J*) are defined as follows $$\begin{array}{r} {Z\left( I \middle| J \right) = \left| {\frac{1}{2N_{I}}\sum\limits_{n = 1}^{N_{I}}\left( {\frac{e_{I}(n)^{2}}{\sigma_{J}^{2}} - 1} \right)} \right|,} \\ \end{array}$$ where *N*_*I* is the size of window *I*, *e*_*I*(*n*)² is the residual error, and $\sigma_{J}^{2}$ is the variance of estimated noise in the window *J*. The cross-validation error *e*_*I*(*n*)² in is calculated for an AR model with coefficients *ω*_*J* fitted to the window *J*, so that *e*_*I*(*n*)² is $$\begin{array}{r} {e_{I}(n)^{2} = \left( {\hat{y}}_{I}\left( n,\omega_{J} \right) - x_{I}(n) \right)^{2}.} \\ \end{array}$$ In the above the variance $\sigma_{J}^{2}$ is calculated for an AR model with parameters *ω*_*J*, so that $\sigma_{J}^{2}$: $$\begin{array}{r} {\sigma_{J}^{2} = \frac{1}{N_{J} - p}\sum\limits_{n = p + 1}^{N_{J}}\left( {\hat{y}}_{J}\left( n,\omega_{J} \right) - x_{J}(n) \right)^{2}.} \\ \end{array}$$ The *Z*-statistic defined in is calculated for each pair of the reference and test windows and then compared with a critical value, *Z*_*cr*. For EEG signals, *Z*_*cr* is found empirically. A similar approach was adopted in for extracting EEG features. In particular, the adaptive segmentation is used for generating an amplitude vector, proposed in, in order to extract the discontinuity feature. The above techniques were implemented for our experiments as Matlab scripts included in the Supporting Information. ## Extraction of discontinuity feature from amplitude vector According to, the discontinuity feature is extracted from an amplitude vector (AV) generated from a segmented EEG as follows. First, the mean *μ*_*i* of absolute amplitudes is computed for each pseudo-stationary segment, *i* = 1, 2, …, including *L*_*i* samples. The value *μ*_*i* is then repeated *L*_*i* times. For examples, given *L*_*i* = 600, the value *μ*_*i* is repeated 600 times. At the second step, a distribution of the generated AV is estimated and then approximated with a log-normal distribution. Finally, the location *μ* and scale *σ* of this distribution represent the features of interest. illustrates how discontinuity features are changed during sleep of a newborn at age of 44 weeks. Here the feature is represented by a location *μ* and a scale *σ* computed in a 10-min window sliding with a 1-min step over a 120-min recording. The intervals between 10 and 40 min as well as between 80 and 110 min, identified as the quiet sleep phase, are with a high discontinuity value. In contrast, the active phase, that is between 40 and 80 min as well as between 110 and 120 min, is with a low discontinuity value. ## Proposed feature extraction technique In, a technique proposed for estimating the stationarity of EEG signals has employed the spectral density function calculated in two separate intervals. The spectral densities estimated in these intervals are then compared within a 2-sample Kolmogorov-Smirnov (KS) test. This technique was used to estimate the stationarity of intervals when their lengths varied between 1 and 64 sec. A similar approach, based on a statistical test, is adopted in our technique in order to extract the discontinuity feature. The proposed technique based on the Spectral Power Statistics (SPS) is described below. Algorithm 1 describes the main steps of the proposed segmentation technique. The reference *W*₁ and test *W*₂ windows are sliding along a signal *X*. The length of both windows is given by *L*. For each position of the windows *W*₁ and *W*₂, Fast Fourier Transform (FFT) computes the spectral powers *S*₁ and *S*₂ within a given frequency band *S*. These powers are used for testing a hypothesis that EEG signals in the reference and test windows are from the same quasi-stationary process within a given critical level *d*₀. For given signal *X*, length *L*, band *S*, and value *d*₀, the Algorithm 1 finds boundaries of interest and returns their indexes as a vector *T*. At lines 9 and 10 the indexes of reference *W*₁ and test *W*₂ windows are assigned. At the next lines 11 and 12, the spectral powers *S*₁ and *S*₂ are calculated for windows *W*₁ and *W*₂, respectively. If a distance *d* of the KS test exceeds the critical value *d*₀, the EEG signals in windows *W*₁ and *W*₂ have different characteristics, and the line 15 assigns a boundary of the pseudo-stationary segment to the output vector *T*. In our experiments we achieved the best segmentation with the following parameters: length *L* = 200 samples, that is a 2-sec duration given a sampling frequency *F* = 100 Hz, a value *d*₀ = 0.15, and a frequency band *S* = (0, 13.5) Hz. Given *F* = 100 Hz, the band *S* is represented by 28 spectral lines that is a sufficient sample size for the statistical KS test. **Algorithm 1** Adaptive segmentation using Spectral Power Statistics 1: **Inputs**: *X*, *L*, *S*, *d*₀ 2: **Initialise**: 3: *i*₁ ← 1                 ▷ Reference window index 4: *i*₂ ← *i*₁ + *L*                  ▷ Test window index 5: *L*₁ ← *L* − 1 6: *K* ← *floor*(*length*(*X*)/*L*) − 1          ▷ Number of segments 7: *T*\[1, *K*\] ← 0                 ▷ Segmentation vector 8: **for** *k* ← 1, *K* **do** 9:   *W*₁ ← \[*i*₁, *i*₁ + *L*₁\]              ▷ Reference window 10:  *W*₂ ← \[*i*₂, *i*₂ + *L*₁\]                ▷ Test window 11:  *S*₁ ← *Sum*(FFT(*X*(*W*₁)), *S*)          ▷ Spectral powers 12:  *S*₂ ← *Sum*(FFT(*X*(*W*₂)), *S*) 13:  *d* ← StatTest(*S*₁, *S*₂)              ▷ Statistical test 14:  **if** *d* \> *d*₀ **then** 15:   *T*\[*k*\] ← *i*₂             ▷ A new segment boundary 16:  **end if** 17:  *i*₁ ← *i*₁ + *L*                 ▷ Moving windows 18:  *i*₂ ← *i*₂ + *L* 19: **end for** 20: **return** *T* ## New discontinuity feature Having recorded the locations of segment boundaries in the vector *T*, we can consider a rate of pseudo-stationary intervals as a discontinuity feature and introduce a segmentation rate, *sr*, as follows: $$\begin{array}{r} {sr = K\left\lbrack \frac{\parallel X \parallel}{L} \right\rbrack^{- 1},} \\ \end{array}$$ where *K* is the number of pseudo-stationary segments detected in a signal *X* and stored in *T*, $\left\lbrack \frac{\parallel X \parallel}{L} \right\rbrack$ is the maximal number of segments that can be detected in signal *X* by using a window of length *L*, and ‖*X*‖ is the length of *X*. According to, the larger the *sr* value, the larger is the number *K* of segments and, therefore, higher is the discontinuity of sleep EEG. shows the results of the proposed segmentation technique, where the boundaries of pseudo- stationary segments are labelled by the vertical bars in Red. The *sr* is higher for the EEG recorded at 36 and 38 weeks, shown on plots a) and b). For the EEG recorded at 41 weeks shown on plots (c) and (d), the variations in EEG activity are smaller and so segment rate *sr* is decreased. # Experiments with EEG data In this section we present results of our experiments on the EEG data recorded during sleep hours from newborns in 10 age groups. We explore the correlation of the proposed discontinuity feature with the newborn ages. Finally we compare the proposed and existing discontinuity features in terms of classification and uncertainty estimation accuracy. ## Description of EEG data In our experiments we used 1,110 EEG recorded from newborns in 10 age groups from 36 to 45 weeks, with approximately 100 recordings in a group. The data were recorded during the project on automated EEG assessment of newborn brain development, see e.g., conducted at the University of Jena, Germany. The recordings were made with the C3-T3 and C4-T4 electrodes with a sampling rate *F* = 100 Hz. The electrodes were positioned according to the standard 10–20 electrode system. Raw EEG were filtered to remove slow drifts with frequencies below 0.1 Hz and noise along with high-frequency interference above 30 Hz. The EEG segments with amplitudes that exceeded a threshold found as ±1.5 standard deviation of amplitudes in a 2-min sliding window were removed as artefacts. Segments with the spectral power below 10% of the average power were also removed as “lost” signal. The average rate of artefacts was around 20%. The EEG were analysed in the standard frequency bands that are typically used for analysis of sleep EEG. shows the six standard bands and their frequency ranges. ## Methodology of experiments The above data were used in our experiments for comparison of the proposed and existing techniques described in the previous section. The features extracted from segmented EEG were compared, first, in terms of correlation with newborn ages and, second, in terms of accuracy of age classification and uncertainty estimation. ### Correlation with brain development The AR model based segmentation technique described in the above section was run with the reference and test windows being set with 2-sec duration and a 2-sec moving step similar to the SPS technique. In our experiments we applied *Z*_*cr* ∈ (4.0, 9.0) and obtained almost the same correlation with ages, *ρ* ≈ 0.60. A threshold (TR) segmentation technique, proposed in, calculates a difference, *d*_*k*: $$d_{k} = \max\limits_{1 \leq n \leq N}\left( x_{n} \right) - \min\limits_{1 \leq n \leq N}\left( x_{n} \right),k = 1,\ldots,K,$$ where *N* is the number of EEG samples in the *k*th interval of 2-sec duration, and *K* is the number of the intervals in EEG. Differences *d*_*k* are calculated for all *K* intervals and then compared with a threshold *d*₀ ∈ {25, 50}*μV*: $$\begin{array}{r} {d_{k} - d_{0}\begin{cases} {> 0,} & {T_{k} = 1,\text{continuity},} \\ {\leq 0,} & {T_{k} = 0,\text{discontinuity.}} \\ \end{cases}} \\ \end{array}$$ Then, finally, a ratio of the continuous intervals, $\sum_{k = 1}^{K}\left( T_{k} \middle| T_{k} = 1 \right)/K$, is considered as a discontinuity feature along with the segmentation vector *T*. If *T*_*k* + 1 ≠ *T*_*k*, then a boundary is assigned between segments *k* and *k* + 1, otherwise the segments are considered to be similar. The proposed SPS technique was run with the reference and test windows of 2-sec duration, each including *L* samples. The windows were set to be moving with a 2-sec step. The frequency band *S*, in the Algorithm 1 was set in the range (0, 13.5) Hz, that includes the standard bands Subdelta to Alpha, shown in. The critical value *d*₀ for the KS test was given 0.15. This value enabled the algorithm to assign a segment boundary if the spectral powers *S*₁ and *S*₂, that are considered to be sampled from the same stationary process, are different with a *p*-value, *p* \< 0.9. ### Classification of EEG maturity In experiments we used Bayesian method to compare the assessment accuracy that can be obtained with the proposed and conventional EEG features. Bayesian methods are known for accurate estimation of predictive posterior probabilities, *P*_*ij*, for each input *i* and each class *j*. This enables practitioners to reliably estimate the uncertainty intervals for each patient. The provides details of the Bayesian method. The above predictive posterior probabilities are calculated in our experiments with different feature extraction techniques in order to estimate and compare uncertainties of age classification. Following, the uncertainty is estimated in terms of Entropy, *E*, as follows $$\begin{array}{r} {E = - \sum\limits_{i = 1}^{T}\sum\limits_{j = 1}^{C}P_{ij}log_{2}\left( P_{ij} \right),} \\ \end{array}$$ where *T* is the size of test data that are used for analysing the predictive accuracy, and *C* = 10 is the number of age groups. The EEG were recorded from newborns in 10 age groups between 36 and 45 weeks. Each group was represented by approximately 100 EEG recordings. Because of physiological variability, sleep EEG are difficult to distinguish, and assessments are made within ±1 week of the post-conceptual age. The accuracy of such assessment provided by EEG experts, known from, is 65.0%, that is the baseline for our comparison. ## Experimental results shows performances of the SPS, AR and TR segmentation techniques in terms of correlation observed between the extracted features and post-conceptional ages. The correlation was estimated with Spearaman’s rank correlation coefficient, *ρ*. The columns *AV*_*μ* and *AV*_*σ* show correlations *ρ* obtained by the AV technique when the EEG were segmented by the SPS, AR and TR techniques. The results achieved with the TR techniques were obtained for 25*μV* and 50*μV* threshold and denoted TR(25) and TR(50), respectively. The columns *AV*_*μ* and *AV*_*σ* in show the correlation obtained with the location *μ* and scale *σ*, that were estimated by the AV technique, respectively. The last column, *sr*, shows the results obtained with the feature *sr*, defined by, for all the segmentation techniques. In we see that the proposed SPS technique has extracted the new feature with the strongest correlation, *ρ* = −0.734. The second result, *ρ* = −0.598, was obtained with the AR segmentation technique. The TR(50) techniques, applied for segmentation with a 50*μV* threshold, provided the weakest correlation, *ρ* = 0.245. At the same time, the ratios of segments with EEG activity exceeding a given threshold are correlated with age, delivering *ρ* = 0.344 and *ρ* = 0.302 for 25*μV* and 50*μV* thresholds, respectively. All results were statistically significant with *p*-value, *p* \< 0.01. Observing the correlations *ρ* in the column *sr*, we see that the rates of segments are decreased with post-conceptional age for the SPS, AR, and TR(25) techniques, and *ρ* \< 0. For the TR(50) segmentation with a 50*μV* threshold the tendency is opposite and *ρ* \> 0. This can be explained by a higher EEG activity allowed in segments that reflects the fact of increasing EEG activity with newborn age. shows the correlation between newborn age and *sr* obtained with the proposed SPS and AR techniques. shows the performance, *P*, and entropy *E*, calculated by, for the Bayesian classification using EEG features extracted with the SPS, AR, and AV techniques. The average performance and 2*σ* intervals were calculated within the 10-fold cross validation. We observe that the average performance of the SPS technique is 69.2% that is higher than that provided by the AR techniques. Moreover, the new feature provides a smaller classification uncertainty, giving an entropy *E* = 199.3. # Conclusion EEG discontinuity is known in the literature as an important feature for evaluating brain development of newborns in weeks between 28 and 42 weeks of post-conceptional age. The conventional approach is based on discontinuity features that can be extracted from segmented EEG. In our research we found that the discontinuity features, extracted within the existing approaches, become weakly correlated with brain maturity at 36 and 45 weeks, that affects the assessment accuracy. This observation inspires us to assume that more accurate results can be achieved with a new discontinuity feature estimated as a rate of pseudo-stationary intervals which can be detected by a new adaptive segmentation technique. We hypothesised that such a feature will be more strongly correlated with brain maturation. Our assumption was based on the observation that during brain development the continuous EEG patterns become longer, while the discontinuous patterns become shorter, and this increases a correlation between the proposed feature and age-related changes. The proposed and conventional features were compared on the EEG data recorded from newborns in 10 age groups from 36 to 45 weeks. In our experiments we found that the new features provide a stronger correlation with ages. The new EEG features were explored within the Bayesian assessment of brain development. The new features have improved the assessment accuracy achieving 69.2%, whilst the accuracy of the baseline expert evaluation known from the literature is 65.0%. The existing feature extraction techniques were incapable of exceeding the baseline accuracy. It is also important to note that predictive distributions generated by the Bayesian method are used to provide an accurate approximation of uncertainty intervals within which a prediction is distributed. This becomes critically important when technologies assist practitioners to avoid fatal errors. # Supporting information The authors are grateful to the reviewers for constructive and useful comments. The research has been largely supported by the UK Leverhulme Trust. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** VS LJ. **Data curation:** LJ VS. **Formal analysis:** VS. **Funding acquisition:** VS. **Methodology:** VS LJ. **Project administration:** VS. **Resources:** VS. **Software:** VS LJ. **Supervision:** VS. **Validation:** VS LJ. **Visualization:** LJ VS. **Writing – original draft:** LJ VS. **Writing – review & editing:** LJ VS.

# Introduction Wheat (*Triticum aestivum* L.) and other related temperate cereal species are able to grow under a wide range of agro-climatic regions. A key factor underlying this successful wide adaptation is the variability in timing of important biological events that provides stress avoidance capabilities during different seasons. For example, wheat can optimally coordinate flowering time with changing season to avoid freezing temperatures, heat stress and drought stress that could potentially damage the floral organs. Local breeding programs can take advantage of the genetic variability governing these adaptive mechanisms to select for cultivars that suit their existing growing environment; thereby, developing cultivars resilient to future climate changes. Flowering time in wheat is determined with two basic and well-described environmental cues: low temperature and photoperiod, which categorize wheat genotypes into winter or spring (with or without vernalization requirement) and photoperiod insensitive and sensitive. The major genetic factors influencing such phenological characteristics in wheat are vernalization response genes (*VRN*), controlling the requirement of a cold period to switch from the vegetative to reproductive phase, as well as photoperiod sensitivity genes (*PPD*), determining plant response to day length. Three major genes, *VRN-1*, *VRN-2* and *VRN-3* control the vernalization requirement in wheat. *VRN-1* and *VRN-3* induce flowering when dominant, with *VRN-1* having the main influence on the transition of the apex from vegetative to reproductive phase, while recessive mutants of *VRN*-2 accelerate flowering. The *VRN-1* gene series include three homeologous loci *VRN-A1*, *VRN-B1*, and *VRN-D1*, on the long arm of chromosomes 5A, 5B, and 5D, respectively. Greater polymorphism within the promoter, exon1 and intron1 regions has been reported for VRN-A1 compared to *VRN-B1* and *VRN-D1*. Notably, mutations at the *VRN-A1* promoter region (*Vrn-A1a*) and intron 1 (*Vrn-A1b* and *Vrn-A1c*) and large deletions in intron 1 of the *VRN-B1* and *VRN-D1* genes have been associated with spring growth habit, whereas the presence of intact homozygous recessive *vrn-A1* allele confers winter growth habit. Once vernalization requirement of winter wheat is fulfilled, photoperiod response will control the flowering time. Photoperiod response in wheat is mainly controlled by the *PHOTOPERIOD1 (PPD-1)* loci located on the short arms of chromosomes 2A, 2B, and 2D (8). *PPD-1* genes identified in wheat and barley are members of the pseudo-response regulator (PRR) family. Each locus has alleles that determine whether the plant has a photoperiod sensitive (long day) or insensitive (day-neutral) phenotype. Genotypes containing a photoperiod insensitive allele (suffix “*a*”, e.g., *Ppd-A1a*) will flower regardless of the duration of daylight, however, genotypes that are photoperiod sensitive (suffix “*b*”) will be delayed during short days and switch to reproductive stage when day length is increasing. Photoperiod insensitivity is mainly controlled by the dominant *Ppd-D1a* allele, followed by the *Ppd-B1a* and *Ppd-A1a* alleles. Combinations of alleles at the *VRN-1* and *PPD-1* loci have been reported to result in variation in agronomic traits and physiological development such as flowering time, tillering, spikelet number, and plant height. The influence of allelic variation at the *VRN-1* and *PPD-1* loci on heading date has been studied in Western regions of Canada for spring wheat collections and several spring and winter wheat collections in other countries. Most of the variation in heading date in spring wheat collections can be explained by allelic variation at *VRN-1* loci. In contrast, *Ppd-D1*, *Ppd-B1* and their interaction were responsible for the variation in heading date for winter wheat collections studied in U.S. Great Plains. A study of 683 lines from Europe, Asia, Africa, America, and Australia, reported that the winter and spring alleles at the *VRN-D1* locus had no significant effect on heading. The study also examined the effect of photoperiod sensitive and insensitive alleles at the *PPD-B1* and *PPD-D1* loci, which had significant effects on heading date, with the photoperiod sensitive lines exhibiting later heading. Due to the importance of these loci for their influence on flowering time, studying the allelic variation at these loci in relation to plant phenology will help determine which allelic combinations are most beneficial in a particular growing region or in optimizing plant phenology for the changing climate. Determining the optimal allelic combinations will facilitate marker assisted selection within breeding programs to avoid advancing undesirable allele combinations. The objectives of this study were to (i) determine the genetic variation at the *VRN* and *PPD* loci in a diverse set of wheat genotypes and (ii) study the plant phenology as influenced by *VRN/PPD* genotypes in fall- grown crop in high latitude winter wheat growing regions in Ontario, Canada. These analyses will provide a better understanding of the role of vernalization and photoperiod genes in determining the flowering and maturity times of winter wheats in northern latitudes. # Materials and methods ## Plant material The plant material consisted of a diversity panel of 203 winter wheat genotypes in two sub-panels. The first sub-panel, designated VRN144, consisted of 144 entries, including lines that are currently grown as winter wheats adapted to Canada (70 entries), and other parts of the world (3 entries), as well as elite lines (66 entries) currently in advanced stages of testing in breeding programs or in regional performance trials in Ontario, Canada. The second panel, designated VRN64, consisted of commercial winter wheat cultivars from the Canadian Prairies (39 entries) and 25 cold-tolerant spring wheat genotypes developed at the Lethbridge Research and Development Centre (LeRDC), Agriculture and Agri-Food Canada (AAFC) in Alberta, Canada. The cold-tolerant spring genotypes were developed directly from winter by spring crosses or by intercrossing previously characterized cold-tolerant spring wheat lines. Selection for spring growth habit and superior survivability occurred over several generations. Five genotypes were common between the two sub panels, containing a total of 203 unique genotypes. ## Environments and experimental design In 2014–15, the entries were evaluated at the University of Guelph Elora Research Station near Elora, Ontario, Canada (43°38′N 80°25′W). The VRN144 sub panel was set up in a 12×12 partially balanced square lattice design with two replicates. The VRN64 sub panel was planted in the same field in an 8×8 partially balanced square lattice design with two replications. Each experimental unit was a two-row plot planted at a density of 400 seeds m^-2,1.5m long with 17.8 cm row spacing and a 0.5 m alley separating plots, with a 35 cm space between adjacent plots. During the 2015–2016 season, the diversity panel was planted in two locations; the Elora Research Station and the Woodstock Research Station (43°15′N 80°78′W) near Woodstock, Ontario, Canada, respectively. The same statistical designs were used as the 2014–2015 growing season. Each experimental design was a six-row plot 4m long, with 17.8 cm row spacing and with a 2 m alleyway separating plots. ## Phenotypic evaluation Winter survival data were recorded in April on a 0 to 10 scale, where 0 indicates no plants survived and 10 indicates 100% of the plants survived. Crop developmental stages (booting, heading and anthesis) were determined for each field plot using Zadoks’ scale. Number of days to booting (stage 41) was recorded when 75% of the tillers in the plot had the flag leaf fully expanded, the flag leaf sheath started opening and the head became visible inside the sheath. Number of days to heading (stage 59) was recorded when 75% of the tillers in the plot had complete head emergence. Number of days to anthesis (stage 61) was recorded when 75% of the tillers in the plot had anthers extruded from the florets. Number of days to physiological maturity (stage 87) was recorded when peduncles of 75% of the tillers in the plot turned color. Grain filling period (GFP) was measured by subtracting the days to anthesis from days to maturity. Plots were harvested with a Wintersteiger combine (Wintersteiger, Ried im Innkreis, Austria). Grain yield, test weight, and moisture content of each plot was collected at harvest using HarvestMaster’s Grain Gage (Juniper Systems, Inc., Logan, UT). Regardless of the range of maturity that was present among the genotypes, all plots were harvested at the same time, as there was no instance, in which an early variety may shatter. Plot yield data was then adjusted to 14% moisture content for every experimental unit according to the moisture measured at the time of harvest. ## Genotyping ### PCR amplification of allele-specific markers DNA was extracted from freshly grown seedlings. Four plants of each genotype, grown for a week in 96 well trays with cotton balls, were watered daily for five days. Five days after germination leaf tissue from the four plants was cut into smaller pieces and ground using an Eppendorf blue micro-pestle (Eppendorf, Hamburg, Germany) in a 1.5 ml tube. DNA extraction followed the Cetyl Trimethyl Ammonium Bromide (CTAB) protocol. The quality and quantity of DNA were assessed by determining A260nm and A280nm using a NanoDrop (ND-1000) spectrophotometer (NanoDrop Technologies, Wilmington, DE, USA). Genotyping was conducted using 7 allele specific PCR markers. The primer set VrnN_FP3/R3 was designed by Primer Express Software (Applied Biosystems). PCR assays were carried out in Fisherbrand 96-well semi-skirted PCR plate 96-well Plates (Thermofisher, Mississauga, ON, Canada) using a Mastercycler Pro (Eppendorf, Hamburg, Germany) with 25 μl reactions consisting of 3–4 μl of 50 ngμl^-1 template DNA, and 1X Taq PCR Master Mix (Qiagen, Maryland, USA). PCR cycles were conducted according to the conditions stated in the published reports. The PCR products were size fractionated and visualized using the QIAxcel Advanced System (QIAGEN GmBH, Hilden, Germany). ## Data analysis ### Cluster analysis Genotypes that had heterogeneous or undetermined alleles at any locus were removed from the cluster analysis. Genotypic data for the 203 genotypes was subjected to cluster analysis. Genotypic data was imported into the software Graphical Genotype (GGT 2.0;), in which a matrix of the pair-wise genetic distances were computed. This matrix was then saved as a MEGA file. The dissimilarity matrix was then exported to the MEGA 7.0 software where an unweighted pair group method with arithmetic mean dendrogram was calculated and a dendrogram was generated using”Construct/Test UPGMA Tree” command under the phylogeny tab. ### Growing degree days Cumulative growing degree-days (CGDD) was calculated as the number of daily growing degree days received to reach various phenological stages (booting, heading, anthesis, and physiological maturity) using the formula explained by McMaster and Smika, 1988. The equation for calculating GDD is: $$GGD = {\sum_{i = 1}^{p}{\left\lbrack {\left( {Daily\ Max\ Temp + Daily\ Min\ Temp} \right) \div 2} \right\rbrack - 4}}$$ where *i* is the number of the day ranging from the first frost-free day in each season (April 10^th in Elora 2015 and April 15^th in 2016 trials) to the *p*^th day, in which the respective phenological stage was recorded. ### Analysis of variance Analyses of variance of the phenotypic data were conducted using the PROC MIXED procedure in SAS 9.4 (SAS Institute, Inc., Cary, NC). The mixed model analysis was conducted on the raw data to determine significance of fixed (genotype) and random (block, incomplete bloc within block, environment, and their interactions) factors. Least squared means (Lsmeans) for the genotypes were computed using LSMEANS statement. Tests of normality of residuals were performed using Shapiro Wilk, Kolmogorov-Smirnov, Cramer-von Mises, and Anderson-Darling in PROC UNIVARITA procedure of SAS. The scatter plot of studentized residuals against predicted values were generated in PROC GPLOT to examine the random and independent distribution of residuals. Statistical significance of the genotypic groups at each of the *VRN* and *PPD* loci were examined using a linear mixed model in the PROC MIXED procedure of SAS, in which the phenotypic value was examined as a linear function of the genotype at each locus. If genotypes were heterogeneous or unidentified for a certain locus then they were removed from the analysis of the locus. The loci *PPD-B1*, *VRN-B1*, *VRN-D1* and *VRN-B3* were not tested against phenological traits due to a lack of variation at these loci. Pearson’s coefficients of correlations of phenological traits were computed using the PROC CORR Procedure in SAS. Principle component analysis was performed to examine the relationships between the observed traits using the PROC PRINCOMP and PRINQUAL in SAS. Biplots were created using PC1 and PC2 values as the x and y-axis, respectively. Box plots were created in Sigma-plot software (Systat Software Inc., Richmond, CA), using the box plot function. The whiskers of the box plot extend out to the 10^th and 90^th percentiles. Data points that extend past the whiskers are considered outliers. # Results ## Genotyping ### Allelic frequency at the *VRN-1* and *PPD-1* loci Analysis of the diversity panel (n = 203) with 7 gene-specific markers revealed that *PPD-D1* locus had the most allelic variation among the three photoperiod sensitivity loci, with 127 genotypes carrying the photoperiod-sensitive *Ppd-D1b* allele and 68 genotypes carrying the photoperiod insensitive *Ppd-D1a* allele. The locus with the second largest variation was *PPD-A1*, in which 170 genotypes carried the photoperiod sensitive *Ppd-A1b* allele and 31 genotypes carried the photoperiod-insensitive *Ppd-A1a* allele. The locus *PPD-B1* did not have any allelic variation, such that all 203 genotypes carried the photoperiod sensitive *Ppd-B1b* allele. *VRN-A1* showed more variability compared to other major vernalization loci tested, with 182 genotypes carrying the recessive *vrn-A1* winter allele, and 18 genotypes carrying the *Vrn-A1a* spring allele. With only an exception of one genotype carrying spring *Vrn-B1* allele, the rest of the germplasm (201 entries) carried the recessive *vrn-B1* allele. Low variation was found for *VRN-D1* locus. Most of the genotypes (202 entries) carried the recessive *vrn-D1* winter allele; only one genotype carried the spring *Vrn-D1* allele. The germplasm did not have any variation at the *VRN-B3* locus, with all 203 genotypes carrying the recessive *vrn-B3* allele. ### Cluster analysis Cluster analysis of the 203 genotypes identified six distinct clusters, based on their genotypes at the major *VRN* and *PPD* loci, using allele-specific marker data. The largest cluster (shown in green, ; n = 85) consisted of genotypes with photoperiod-sensitive alleles at all three major *PPD* loci, *Ppd-A1b/Ppd-B1b/Ppd-D1b*, and the recessive winter alleles at all three *VRN-1* loci, *vrn-A1/vrn-B1/vrn-D1* as well as the winter allele *vrn-B3*. The second largest cluster (shown in purple, ; n = 55) had the photoperiod-insensitive allele *Ppd-D1a*, the photoperiod-sensitive *Ppd-A1b* and *Ppd-B1b* alleles, and vernalization-sensitive *vrn-A1*, *vrn-B1*, *vrn-D1*, and *vrn-B3* alleles at the *VRN-1* loci. The next cluster (shown in blue, ; n = 18) had the photoperiod-insensitive allele *Ppd-A1a*, and the sensitive alleles *Ppd-B1b* and *Ppd-D1b*, as well as the recessive winter alleles; *vrn-A1*/*vrn-B1*/*vrn-D1/ vrn-B3*. Another cluster (shown in yellow, ; n = 12) had the photoperiod-insensitive alleles *Ppd-A1a* and *Ppd-D1a*, but the sensitive *Ppd-B1b* allele and the winter- sensitive alleles *vrn-A1*, *vrn-B1*, *vrn-D1*, and *vrn-B3*. Most of the Canadian winter wheat varieties are grouped in these four clusters; all carried recessive winter alleles at the *VRN*-*A1*, *VRN-B1*, *VRN-D1*, and *VRN*-*B3* loci in addition to photoperiod-sensitive alleles at either one, two or all three major *PPD* loci. The remaining genotypes carried at least one spring allele at one of the three major *VRN* loci. The largest cluster of this kind that contained spring alleles (shown in red, ; n = 18), had the photoperiod-sensitive alleles *Ppd-A1b*, *Ppd-B1b*, and *Ppd-D1b*, and winter-sensitive alleles *vrn-B1*, *vrn-D1*, and *vrn-B3* but the spring allele *Vrn-A1a*. These genotypes are exclusively the winter-hardy spring lines from Alberta, Canada, which were selected for maintenance of fall vegetative growth and spring growth habit, along with winter hardiness. The sixth and smallest cluster (shown in pink, ; n = 1) is a spring wheat check in the test, which contained the photoperiod sensitive alleles *Ppd-D1b* and *Ppd-A1b*, as well as the winter alleles *vrn-A*1 and vrn-B3, but had the spring alleles; *Vrn-B1* and *Vrn-D1* loci. ### Effect of the allelic variation at the *VRN-1* and *PPD-1* loci on phenotypic traits Significant differences (P = \<0.001) were observed among genotypes for yield, plant height, GFP, Thousand-Kernel Weight (TKW) and for the number of growing degree days required to reach booting, heading, anthesis, and maturity. The interaction effect of genotype-by-environment (G×E) was significant (P = \<0.001) for all traits, suggesting that genotypes responded differently to environments. The phenotypic differences of the genotypes with different alleles at the *VRN* and *PPD* loci were examined in separate mixed model analyses. *Ppd-D1b* and *Ppd-A1b* were associated with 38.78 and 36.52 GDDs later booting in combined analysis across environments as compared to the photoperiod insensitive lines carrying *Ppd-D1a* and *Ppd-A1a*, respectively. This pattern repeated itself for GDDs to heading and anthesis for entries with insensitive alleles at the *PPD-D1* and *PPD-A1* loci. The average effect of *Ppd-D1b* was 36.29 GDDs later to heading and 34.21 GDDs later for entries with *Ppd-A1b* compared with entries with insensitive alleles. In contrast, genotypes with *Ppd-D1b* reached anthesis 35.68 days later, while entries with *Ppd-A1b* reached anthesis 31.19 days later than genotypes with the insensitive alleles. Genotypes with different alleles at the *VRN-A1* locus, however, were not different for the number of GDD to booting, heading and anthesis. The effect of allelic variation at the *PPD-D1* locus was not significant for GDD required for GFP. Genotypes with different alleles at the *PPD-A1* locus, on the other hand were significantly different for GDD required for GFP. The average GFP of the genotypes with *Ppd-A1a* or *Ppd-A1b* were 516.47 and 497.44 GDD, respectively. Similarly, genotypes with different alleles at the *VRN-A1* locus were significantly (P = \<0.0001) different for GDD required for GFP (Tables and). The average GFP for genotypes with *Vrn-A1a* or *vrn-A1* were 469.15 and 504.09 GDD, respectively. Genotypes with different alleles at the *PPD-D1* (P = \<0.0001), *PPD-A1* (P = 0.01), and *VRN-A1* (P = \<0.0001) loci were significantly different for yield. The average yield of genotypes with *Ppd-D1a* or *Ppd-D1b* were 5.63 and 4.87 t ha^-1, respectively, while genotypes with *Ppd-A1a* or *Ppd-A1b* on average yielded 5.53 and 5.03 t ha^-1, respectively. The average yield of genotypes with *Vrn-A1a* or *vrn-A1* were 3.98 and 5.23 t ha^-1, respectively. Genotypes with different alleles at *PPD-D1* (P = \<0.0001), *PPD-A1* (P = 0.0004), and *VRN-A1* (P = \<0.0001) were significantly different for plant height. The average plant height of genotypes with *Ppd-D1a* or *Ppd-D1b* allele were 77.35 and 85.40 cm, respectively, while the average plant height of genotypes with *Ppd-A1a* or *Ppd-A1b* alleles were 77.57 and 83.54cm, respectively. The average plant height of genotypes with *Vrn-A1a* or *vrn-A1* alleles were 90.86 and 81.73cm, respectively. ### Correlations and principle component analysis The GDD to anthesis was very closely associated with GDD to booting (r = 0.94) and GDD to heading (r = 0.96) in the combined year biplot. Anthesis and maturity were associated (r = 0.73), but GDD to anthesis and GFP had no association (r = 0.00). Maturity was associated with GFP (r = 0.68), but had no association with yield (r = 0.01). GFP and yield were closely associated (r = 0.35). Yield had a positive correlation with TKW (r = 0.3), but it had a negative association with plant height (r = -0.29) and GDD to anthesis (r = -0.32). Principle component analysis (PCA) with GFP, TKW, height, yield, winter survival and phenological traits of 203 genotypes was performed to assess whether these variables could be used to differentiate varieties with different alleles at *PPD-D1* and *PPD-A1*. The first two principle components accounted for 66.13% of the variation. Component 1 was positively correlated with GFP, TKW, height and phenology traits and negatively correlated with yield and winter survival. In contrast, component 2 was only correlated with yield, GFP, TKW and maturity. The biplot of the first two principle components separated the effect of PPD sensitivity along the PC1 axis, with all photoperiod sensitive genotypes grouping togeather (with some overlap), clustering towards booting, heading, anthesis, maturity, GFP, plant height and TKW. In contrast genotypes with at least one photoperiod-insensitive allele clustered mainly in the left side of biplot tending towards higher yield, and away from phenological traits. Within genotypes with photoperiod insensitive alleles, fully insensitive ones showed more tendency towards yield. # Discussion Using diagnostic molecular markers for the most common alleles of the *VRN* and *PPD* loci affecting the vernalization requirement and photoperiod response, we characterized a germplasm collection of Canadian winter wheat genotypes grown or developed for the higher latitude regions of North America. The first objective of this study was to evaluate the allelic variation at the important *VRN* and *PPD* loci. Marker analysis revealed a high prevalence of photoperiod sensitivity alleles at the *PPD-D1* (62.5%), *PPD-A1* (84%) and *PPD-B1* (100%) loci. High frequency of *Ppd-D1* sensitive alleles in higher latitudes was previously reported by Guo et al., in lines collected from Canada and U.S. in a worldwide wheat collection (n = 492). The photoperiod insensitive allele *Ppd-D1b* was present at levels similar levels to that detected by Kiss et al. in winter wheat lines from America and Africa compared to those from other continents. In the germplasm studied in this investigation, we observed a lack of variation at the *VRN* loci, which was comparable to low frequency (2%) of *VRN*-*A1* in a European winter wheat collection of 521 winter wheat genotypes. The *Vrn-B1* and *Vrn-D1* spring alleles in their study were present at slightly higher frequencies compared to the Canadian winter wheat genotypes. Lack of variation at the major *VRN* loci was expected, considering that the diversity panel was almost exclusively winter wheat genotypes adapted to Canadian conditions. The low variation at the *VRN-1* locus was also observed in a winter wheat collection (n = 299) from U.S. Great Plains. However, they found higher frequency of photoperiod-insensitive alleles at *PPD-B1* (53%), rare occurrences of *Ppd-A1a* (2%), and comparable *Ppd-D1a* (35%). They indicated that varieties from the northern Great Plains had greater incidences of the photoperiod sensitive alleles than germplasm from central and southern breeding programs. Many genetic studies demonstrated that winter and spring wheat cultivars grown in northern latitude countries usually carry photoperiod sensitive alleles at higher frequencies. In contrast, genotypes grown in southern latitudes normally carry photoperiod-insensitive alleles. This is likely attributable to less variation in day length in southern latitude regions where photoperiod insensitivity could have clear adaptive significance. The typical Canadian winter wheat material carried winter alleles at all major *VRN* loci and photoperiod-sensitive alleles at either three or two of the major *PPD* loci. Most genotypes had the *Ppd/Vrn* allelic combination of *Ppd-A1b*, *Ppd-B1b*, *Ppd-D1b*, *vrn-A1*, *vrn-B1*, *vrn-D1*, and *vrn-B3*, which accounted for 43% of genotype combinations. The second largest group was similar except it carried *Ppd-D1a* instead of *Ppd-D1b*, and accounted for 28% of genotypes. Genotypes with all three *VRN-1* winter alleles and either one or more of the *Ppd* insensitive alleles account for 42% of the total genotypes. This indicates that selection in the Canadian winter wheat breeding programs has favored selection for photoperiod-sensitivity allele(s) at the major *PPD* loci. Grogan et al. also reported an increase in selection of photoperiod-insensitive alleles throughout the U.S. Great Plains after the year 2000. They found higher frequency of photoperiod sensitive alleles *Ppd-A1b*, *Ppd-B1b*, and *Ppd-D1b* northern plains compared to central and southern plains. On the contrary, Kamran et al. reported that in Western Canadian spring wheat, *Ppd-D1b* is being replaced with the photoperiod-insensitive *Ppd-D1a* allele in recent germplasm. Our second objective was to study the implications of *VRN* and *PPD* variation on plant phenology in a diverse set of fall-sown winter and spring wheat lines that represent the genetic variation in winter wheat in higher latitudes (\>40°N) of North America. We found that day length sensitive photoperiod genes play a major role in determining flowering time and adaptability of Canadian winter wheat. Grogan et al. found heading date in winter wheat from the U.S. Great Plains is strongly affected by photoperiod loci. Similar results were observed in present study, where genotypes with varying *Ppd-A1/D1* allele(s) had different time to booting, heading, anthesis, GFP, and height. As expected, photoperiod insensitivity resulted in earlier flowering compared to photoperiod sensitivity. On average, day length insensitive genotypes required 41.8 growing degree-days less than the genotypes with photoperiod-sensitive alleles at all three *PPD* loci to reach anthesis. The results of earlier flowering of photoperiod- insensitive genotypes is consistent with results from previous studies reporting 3.7 and 4.3 days earlier booting and heading, respectively, 1.6 to 8 days earlier flowering, conditioned by the presence of *Ppd-D1a* allele, when compared to photoperiod-sensitive genotypes. Despite earlier anthesis, genotypes with the photoperiod-insensitive allele *Ppd-D1a* in general yielded 13.5% higher than the photoperiod-sensitive genotypes, when compared to the genotypes with *Ppd-D1b* allele. The difference in yield is consistent with other reports; for example, in southern Europe, increased yields of up to 35% were reported for the photoperiod-insensitive genotypes carrying the *Ppd-D1a* allele. Similarly, 7.7% higher yield in Germany and 30% in the former Yugoslavia were reported for genotypes with the *Ppd-D1a* allele. The higher yield in *Ppd-D1a* genotypes can be explained by escape from hot and dry summer days due to earlier flowering. In this study, higher yield for *Ppd-D1a* genotypes may be linked to avoidance of biotic stresses associated with powdery mildew (*Erysiphe graminis*) and Fusarium head blight (caused mainly by *Fusarium graminearum*) which were present in 2015, and stripe rust (caused by *Puccinia striiformis*) in 2016. The difference in the genotypic groups at the *VRN-A1* locus was significant for GFP, height and time to maturity. The extension of GFP was due to a delay in maturity among the *vrn-A1* genotypes, while the two groups reached anthesis at the same time. The accelerated time to maturity of *Vrn-A1a* genotypes is consistent with a study of Canadian hard red spring wheat genotypes. *Vrn-A1a* also induced earlier flowering, compared to *vrn-A1*; however, in this study days to anthesis was not significantly different between the two allelic groups. The lower yield of Vrn-A1a genotypes (cold tolerant spring wheats) may be due to the panel having a high percentage of high-yielding winter wheat genotypes. However, based on the influence of time to maturity, *vrn-A1* genotypes with a delayed maturity may have demonstrated yield high due to a prolonged GFP. In addition, the cold-tolerant spring wheat genotypes were primarily selected on the basis of winter survival, with little regard for grain yield. There have been several theories as to why the spring alleles at the *VRN* locus and the photoperiod insensitive alleles at the *PPD* loci induce earlier flowering. Davidson et al. proposed that the accelerated flowering was due to developmental acceleration from emergence to floral initiation. Photoperiod insensitivity is associated with suppression of *PPD1* and up regulation of VRN3, which in turn promotes flowering by inducing meristem identity genes. The earliness may also be induced by shortening the required heat unit accumulation between emergence and stem elongation. Based on the results of the present study, a combination of the described theories can explain the differences of the time needed to reach booting, heading, and anthesis between photoperiod insensitive and sensitive genotypes. # Conclusions We studied a panel of 208 winter wheat genotypes representative of modern and historic Canadian winter wheat to evaluate allelic diversity and effects of vernalization and photoperiod loci on heading time as well as other physiological stages. We found that most of the variation in the phenology of winter wheat crop in higher latitudes can be explained by allelic variation at the *PPD-D1*, *PPD-A1*, and the interaction between these loci. Selecting for photoperiod insensitivity in the presence of winter alleles at the vernalization loci for fall-seeded wheat-growing regions in the high latitudes of the northern hemisphere may provide wider environmental adaptation. This is a result of early flowering due to photoperiod insensitivity, which results in higher yield due to avoiding late season biotic and /or abiotic stress factors. Increased frequency of photoperiod insensitive alleles in the winter wheat in the higher latitudes may become more prevalent with more frequent occurrence of milder winters due to climate change. Our study also indicates that breeding winter-hardy spring wheat genotypes seems achievable for milder winter areas by selecting for one spring allele at one of the *VRN-1* loci (*VRN-B1*, or *VRN-D1*) in combination with photoperiod sensitive alleles at all major *PPD* loci, which provides some protection of the floral meristem during the fall, along with an acceptable level of cold hardiness. # Supporting information The authors are grateful to staff at the University of Guelph Wheat Breeding Program, Nicholas Wilker, Katarina Bosnic, Melinda Drummond and Ryan Costello for their technical support and the Ontario Ministry of Agriculture, Food, and Rural Affairs (OMAFRA)-University of Guelph Partnership, the Grain Farmers of Ontario, *SeCan* and the National Scientific and Engineering Research Council of Canada (NSERC) for financial support of the project. [^1]: The authors have declared that no competing interests exist.

# Introduction In the past few decades, deer have become increasingly abundant worldwide; this population increases have contributed to the degradation of plant communities and ecosystems. In general, the population dynamics of animals are affected by birth, mortality, and migration rates. Large ungulates are able to breed under low food availability, therefore, the birth rate of deer would not largely decrease even in a degraded forest; however, the density-dependent decline in the birth rate of deer occurs at a later period of the outbreak stage. Furthermore, the survival rate of adult deer was high even in a poor nutritional environment. Thus, deer is a species that can live in high densities and low- nutrient environments. If predators (e.g. wolves) are absent, hunting is one options to control deer populations under these conditions. In snow-covered area and, in particular, during heavy snowfall, the survival rate of sika deer (*Cervus nippon*) decreases. Kawase et al. (2014) projected that winter precipitation including snowfall would decrease in broad regions of Japan due to the ongoing climate change. This climate change may mitigate the mortality of deer and cause further increases in deer populations in the future. Therefore, it is indispensable to estimate the effect of snow on the dynamics of deer populations. While some of the effects of global warming on population dynamics of ungulates have already been reported, models constructed in recent studies to describe deer population dynamics have not yet explicitly considered the effects of snow. From the viewpoint of plant communities and ecosystems, it is important to clarify not only annual trends but also seasonal trends in the deer population density. Plant fitness could be affected differently depending on whether deer browse on them before or after they have reproduced sexually. The timing of browsing could also affect the fitness of pollinators such as bumblebees. Therefore, in order to assess the effects of deer browsing on ecosystem levels it is important to, at least, estimate the seasonal deer abundance. However, in many areas, the annual census of ungulates is held during a season that, although offers good visibility to track ungulates, is not suitable for plant growth. In recent years, generalized linear models, generalized additive mixed models (GAMM,), density surface models, and Bayesian state-space models were used to estimate deer abundance based on field data. Among these models, the Bayesian state-space model can be a powerful tool for estimating deer population dynamics because it can easily handle time series data with temporal autocorrelation and can explicitly distinguish errors following measurement of data with uncertainty about population dynamics. However, there are still limited applications of this model when it comes to the effect of snow and seasonal fluctuations on deer population dynamics. In this study, we estimated deer population dynamics in a cool-temperate forest in Japan using a Bayesian state-space model. The model was based on data collected from block count surveys, road count surveys by vehicles, mortality surveys during the winter, and nuisance control over 12 years. The seasonal and annual fluctuation of the sika deer population and the effects of snowfall and nuisance control on population dynamics are discussed based on the results we obtained by the model and the parameters estimated in the model, respectively. # Materials and methods ## Study site The study site was located at the Ashiu Forest Research Station, Field Science Education and Research Center, Kyoto University, Japan (35°20′N, 135°45′E; 355–959 m a.s.l., 41.86 km²) and the surrounding area (46.12 km² in total). The mean annual temperature and precipitation in this area are 13.1 C and 2,333 mm, respectively. The maximum snowfall during each winter at 356 m elevation was 31.0–141.7 cm between 2007 and 2018. The forest is usually closed from January to early April because the roads in the forest must be blocked with snow. This forest is located in the transition part between the temperate deciduous forest zone and the warm temperate forest zone. This area is well known for being highly diverse in plant species and existing phylogeographically important populations of some species in the forest. Though the forest is one of the wildlife protection areas in Japan, forest vegetation has been steadily degraded by the browsing of *C*. *nippon*; thus, nuisance control started in 2008 using guns, traps, and cages. The last known Japanese wolf (*Canis lupus hodophilax*) was caught in the Nara prefecture in 1905 and there have been no sightings of it in Japan since. Thus, we considered that potent predators of deer such as wolves had been extinct all over Japan, including in our study site. ## Road count We selected three route sectors (A: 4.7 km, B: 3.3 km, E: 0.7 km;) to record the numbers of deer sighted. The investigators of this study were mainly researchers and technical staff employed by the Forest Research Station, including non-specialists in deer. They recorded the date, weather, sector name, and time when they began driving through each sector, whenever they drove through a whole sector by vehicle during the period from May 1, 2007 to December 31, 2018. Then, they recorded the number of deer in each sector. If they found no deer, they recorded the number as zero. The details of the survey are described in a previous study. We excluded records that lacked information about the number of deer sighted, sector name, year, and date. We also excluded records from January to April because few records were available from these periods due to snow accumulation and driving speed was different from other seasons. Furthermore, data within 15 minutes before and after were excluded from later analysis because data independence could not be guaranteed. After this data cleaning, we used 8,616 records for later analysis. ## Block count Block counts were conducted in two sites (north: 86.9 ha, south: 111.7 ha) of the Ashiu Forest Research Station in December, from 2001 to 2018 except for 2017. The sites were divided into 14 and 19 blocks (5–7 ha per block depending on the terrain), respectively. Each block was thoroughly surveyed by an observer walking in a zig-zag motion along the terrain in order to guarantee good visibility. When an observer spotted deer, they informed the observers of adjacent blocks using transceivers to avoid duplicate counting. Occasionally, we did not survey some blocks due to sudden snowfall and lack of observers; however, the total surveyed area was 181.27 ha in most years. Because it was a missing value only in 2017 and values did not change so much in the previous (2016) and next year (2018), we used the mean of 2016 and 2018 as the value of 2017 in the model described later. ## Number of deer carcasses at spring thaw We counted the number of deer carcasses found in forest during the thawing period from April to early July, for the years 2005–2018. We needed to find deer carcasses emerging from the snow before animals preyed on them. However, we could not distinguish their age and sex because the parts of carcasses bodies were sometimes scattered around. We covered a 1–21 km distance per survey and repeated the procedure for 10–18 times per year to look for deer carcasses across the forest ( and). In addition to looking for dead deer, we also relied on our sense of smell and detected carcasses based on the odor they emitted. ## State–space model We analyzed field observation data of relative abundance indices of deer with state–space models, based on a hierarchical Bayesian framework. The state–space model divided the observation data into a system model, representing “true” but unknown population size, and an observation model that accounts for error in counts caused by ability of observers. Because most observers were not specialists for animals, they sometimes missed the count. The state–space models allowed us to permit potential errors in the count data. In most past studies in deer population dynamics, the analysis was performed on a yearly basis. However, we set the time interval to 2 months, excluding the period from January to April (*t* = 1 in May and June 2007, *t* = 2 in July and August 2007, *t* = 3 in September and October 2007, *t* = 4 in November and December 2007, *t* = 5 in May and June 2008, etc.). This was because we were able to use the road count data from all year round except from January to April (when the forest was covered by snow). We wanted to know the seasonal in addition to the annual fluctuation. ### System models Expected deer abundance at time *t* (*N*_*t*) in the forest depended on expected deer abundance at time *t* −1 (*N*_*t*−1); the number varied with the effect of population growth (*r*_*t*) including birth, natural mortality, immigration, migration at time *t* (*r*_*t* did not include the effects of hunting and mortality due to snowfall), and the effect of hunting at time *t* (hunting rate: *h*_*t*). It can be expressed as follows: $$N_{t} = N_{t}{}_{–1} \times r_{t} \times \left( {1 - h_{t}} \right)$$ During the season when the forest was covered by snow (January to April), the deer sometimes got stuck or starved, due to lack of food as a result of the heavy snow. We defined the mortality rate during the seasons when the forest was covered by snow, just before the time *t*, as *d* _*t*. Then, *N*_*t* (*t* = 5, 9, 13, …45) can be expressed as follows: $$N_{t} = N_{t}{}_{- 1} \times r_{t}{}^{2} \times \left( {1 - h_{t}} \right) \times \left( {1 - d_{t}} \right)$$ Although we set the time interval to two months during May to December, we set it to 4 months during January to April. Thus we squared *r*_*t* in (2). If we calculate the logarithm of the two aforementioned equations, then the process follows a linear structure. Then, Eqs and can be re-written as follows: $$NL_{t} = NL_{t}{}_{- 1} + rl_{t} + \log\left( {1 - h_{t}} \right)\mspace{90mu}\left( {t = 2,3,4,6,7,\ldots,48} \right)$$ $$NL_{t} = NL_{t}{}_{- 1} + 2 \times rl_{t} + \log\left( {1 - h_{t}} \right) + \log\left( {1 - d_{t}} \right)\mspace{54mu}\left( {t = 5,9,13,\ldots,45} \right)$$ We introduced stochasticity into the deer population dynamics. Then, Eq can be expressed as follows: $$NL_{t} \sim Normal\left( {\mu_{t},\sigma_{1}{}^{2}} \right)\qquad\left( {t = 2,3,4,6,7,\ldots,48} \right)$$ $$\mu_{t} = NL_{t–}{}_{1} + rl_{t} + \log\left( {1 - h_{t}} \right)$$ Eq can be expressed as follows: $$NL_{t} \sim Normal\left( {\mu_{t},\sigma_{2}{}^{2}} \right)\qquad\left( {t = 5,9,13,\ldots,45} \right)$$ $$\mu_{t} = NL_{t–}{}_{1} + 2 \times rl_{t} + \log\left( {1 - h_{t}} \right) + \log\left( {1 - d_{t}} \right)$$ For the time interval we skipped four months every eight months, because we did not use the data collected from road count surveys by vehicles from the winter season (January to April). Thus, we defined different standard deviations of posterior distribution for deer abundance in the logarithmic scale (*σ*₁ and *σ*₂). The prior probability distribution of the log of expected deer abundance in the first year (*NL*_*1*) was determined as follows: $$NL_{1} \sim Normal\left( {0,100^{2}} \right)$$ Because time interval was short, the population growth rate (logarithmic scale) at time *t* (*rl*_*t*) depended on those at time *t* -1 (*rl*_{*t* − 1}). Thus it modeled as follows: $$rl_{t} \sim Normal(rl_{t - 1},\mspace{2mu}\sigma_{3}{}^{2})$$ We did not include a density-dependence parameter in the population growth rate. The density dependence in population growth of sika deer within only 25.3 km² in open ecosystem (4,465 km²) was found. However, it was largely depended on the habitat environment. Because our study sites were small (46.12 km²), we did not consider habitat heterogeneity in the model. Therefore, we did not consider the density-dependence parameter in this study. The hunting rate (*h*_*t*) was the inverse logit transform of the hunting rate in logit scale (*hl*_*t*). *hl*_*t* modeled as follows: $$hl_{t} \sim Normal(\varphi,\sigma_{4}{}^{2})$$ $$\varphi = hm + rho \times Ef_{t}$$ where *φ* is the mean hunting rate (logit scale) and *σ*₄ is the standard deviation of posterior distribution of the logit hunting rate, and *hm* is the hunting rate at the forest. Because hunting rate was assumed to increase when the hunting effort (*Ef*_*t*: the product of the number of hunters and days for hunting in time *t*) increases, we considered the effect of hunting effort on hunting rate in logit scale (*rho*). The prior probability distribution of *hm* and *rho* was determined as follows: $$hm \sim Normal\left( {0,\mspace{2mu} 100^{2}} \right)$$ $$rho \sim Normal(0,\mspace{2mu} 100^{2})$$ The mortality rate during the seasons when the forest was covered by snow (*d*_*t*) was the inverse logit transform of the mortality rate in logit scale (*dl*_*t*). *dl*_*t* was modeled as follows: $$dl_{t} \sim Normal\left( {\varepsilon_{t},\sigma_{5}{}^{2}} \right)\qquad\left( {t = 5,9,13,\ldots,45} \right)$$ $$\varepsilon_{t} = b + a \times Sn_{t}$$ Where *ε*_*t* is the mean mortality during the winter with severe snowfall in time *t* at logit scale and *σ*₅ is the standard deviation of the posterior distribution of mortality during the winter with snowfall, in the logit scale. Because the mortality may increase in severe snowfall conditions, it was assumed to increase linearly with the number of days with a snow depth of \> 50 cm (*Sn*) before time *t*. To consider the different effects of snowfall, we also used the maximum snow depth instead of the number of days with snow depth of \> 50 cm. The *b* and *a* were the intercept and coefficient, respectively. The prior probability distributions of *b* and *a* were as follows: $$b \sim Normal\left( {0,\mspace{2mu} 100^{2}} \right)$$ $$a \sim Normal\left( {0,\mspace{2mu} 100^{2}} \right)$$ We assigned weakly informative priors for scale parameters, σ₁ to σ₅ as *Cauchy*(0, 10). ### Observation models We modeled the number of deer seen in road count surveys (*C*_*t*,*m*) in time *t* in route *m* (*m* = a, b, e) as follows: $$C_{t,m} \sim Poisson(\delta_{t,m})$$ $$\delta_{t,m} = N_{t} \times R_{m} \times rS_{t} \times O_{t,m} \times A_{c,m}$$ where *R*_*m* is the observation rate per survey in route *m* that converts *N*_*t* to *δ*_*t*,*m*, *rS*_*t* is the seasonal observation rate per survey in time *t*, *O*_*t*,*m* is the number of survey occasions conducted over two months for each route, and *A*_*c*,*m* is the ratio of the study area in each drive count route (we assumed the census width to be 15m) per that of forest (a: 0.153%, b:0.107%, c: 0.023%). We assumed *C*_*t*,*m* followed a Poisson distribution. More precisely, the probability distribution of *C*_*t*,*m* is a Poisson/log-normal mixture because *N*_*t* is assumed to follow a log-normal distribution. Ideally, *C*_*t*,*m* should be modeled to follow a binomial distribution with the population size and the observation rate for each route and time. However, replicated measurements are typically required to estimate these parameters explicitly. Unfortunately, our data did not have such a structure, so we only estimated the expected population size in this model. This was the same for *B*_*t*, *H*_*t*, and *D*_*t*, mentioned below. The prior probability distribution of *R*_*m* were as follows: $$R_{m} \sim Uniform\left( {0,\mspace{2mu} 1} \right)$$ The seasonal observation rate (*rS*_*t*) was the inverse logit transform of the seasonal observation rate in logit scale (*rsl*_*t*). *rsl*_*t* would be affected by leaf phenology of understory vegetation and braches of trees and deer activity including their lactation, mating and so on. The fluctuation of this seasonality in observation rate would have periodicity and the sum of them would be small. Thus, we modeled these seasonal effects as follows; $$rsl_{t} = - \sum_{(l = 1)}^{3}rsl_{(t - l)} + \omega t\mspace{54mu}(t = 4,5,6,\ldots,48)$$ $$\omega_{t} \sim Normal(0,\mspace{2mu}\sigma_{6}{}^{2})$$ where *ω*_*t* is the noise term and *σ*₆ is the standard deviation of posterior distribution of *ω*_*t*. We assigned weakly informative priors for scale parameters, σ₆ as *Cauchy*(0, 10). We modeled the number of deer seen by block count (*B*_*t*) as follows: $$B_{t} \sim Poisson\left( \theta_{t} \right)\mspace{54mu}\left( {t = 5,9,13,\ldots,45} \right)$$ $$\theta_{t} = N_{t} \times bc \times A_{b,t}$$ where *θ*_*t* is the mean number of deer seen by block count in time *t*, *bc* is the observation rate per unit area that converts *N*_*t* to *θ*_*t*, and *A*_*b*,*t* is the survey area of the block count in time *t*. We assumed *B*_*t* followed a Poisson distribution, although Poisson/log-normal mixture is more accurate description we have already mentioned. The prior probability distribution of *bc* was as follows: $$bc \sim Uniform\left( {0,\mspace{2mu} 1} \right)$$ We modeled the number of deer hunted by nuisance control (*H*_*t*) as follows: $$H_{t} \sim Poisson\left( \lambda_{t} \right)\qquad\left( {t = 2,3,4,\ldots,48} \right)$$ $$\lambda_{t} = N_{t} \times h_{t}$$ where *λ*_*t* is the mean number of deer hunted in time *t*. We assumed *H*_*t* followed a Poisson distribution, although Poisson/log-normal mixture is more accurate description we have already mentioned. We modeled the number of deer carcasses found after thawing (*D*_*t*) as follows: $$D_{t} \sim Poisson\left( \eta_{t} \right)\qquad\left( {t = 5,9,13,\cdots,45} \right)$$ $$\eta_{t} = N_{t} \times d_{t} \times rD \times A_{d,t}$$ where *η*_*t* is the number of dead deer found after thawing in time *t*, *rD* is the detection rate per unit area that converts *N*_*t* to *η*_*t*, and *A*_*d*,*t* is the survey area of dead deer surveyed after thawing in time *t*. We assumed *D*_*t* followed a Poisson distribution, although Poisson/log-normal mixture is more accurate description we have already mentioned. The parameter estimation was performed by the Markov Chain Monte Carlo (MCMC;) calculation using RStan 2.18.2. We ran four parallel MCMC chains and retained 60,000 iterations after an initial burn-in of 30,000 iterations. We thinned sampled values to 1.0%. Convergence of MCMC sampling was judged by the criterion that the potential scale reduction factor on split chains, $\hat{R}$ was smaller than 1.1 and by a check of the MCMC trace. The predicted total deer abundance for each time was drawn from a Poisson distribution with the mean as *N*_*t*. To evaluate our models, we compared observed data to simulated data from the posterior predictive distribution. We generated 1,000 data used for posterior predictive checks which we simulate from the posterior predictive distribution. # Results The $\hat{R}$ values of our estimated parameters were all under 1.1. The estimated deer abundance had a sharp peak (September–October 2010) during the 12-year period. From 2011 to 2018, the estimated deer abundance was stable compared to the other periods. The models indicated seasonal patterns in deer abundance; deer abundance gradually increased from April to December during 2013 and 2018. The mean of observation rates in route E was higher than that in routes A and B. The 95% credible interval (CI) of *hm* and *rho* was −8.20 to −6.80 and 0.01 to 0.15, respectively. On the other hand, the 95% CI of *a* included 0. Even when maximum snow depth was used instead of the number of days with snow depth of \> 50 cm, the 95% CI of *a* included 0. The models was able to simulate new data that was similar to the observed values of number of deer seen in road count surveys in route A and B (*C*_*a*, *C*_*b*, respectively) and number of deer seen by block count (*B*). Compared to them, models were less able to simulate new data that are similar to the observed value of number of deer seen in road count surveys in route C (*C*_*e*). # Discussion Using a Bayesian state–space model, we were able to estimate annual and seasonal fluctuations of deer abundance with data collected from block count surveys, road count surveys by vehicles, mortality surveys during the winter, and nuisance control. The models was able to simulate new data that was similar to the observed values of *C*_*a*, *C*_*b*, and B, though they were less able to simulate new data that was similar to the observed value of *C*_*e* ( and). It suggested that our model was evaluated as good fit. However, we did not measure detectability of each survey though the distinction of abundance and detectability is very important in the estimation of wildlife abundance. Therefore, we need to adopt the robust design to estimate detection probability in future study. However, to improve the accuracy of the estimation, we tried to combine the multiple surveys because the uncertainty can be mitigated by using multiple indicators. We found a sharp peak of deer abundance during the 12-year study period (September–October, 2010). The estimated deer abundance at the autumn of 2010, in particular, was the highest (71.0 individuals per km²). The peak in 2010 could be considered an outbreak; this is also reported in other populations and deer species. In 2011, the estimated deer abundance decreased drastically, and has remained at a low level since then. By 2003, most shrubs, herbs, and dwarf bamboo in the forest had already been overgrazed. Therefore, the 2010 irruption and the 2011 decrease could not be due to food shortage of the understory vegetation. When the understory vegetation was poor, deer may have been depending strongly on nuts from canopy and sub-canopy species as food sources during the autumn. In the autumn of 2009, nut production was synchronously very high in three dominant masting Fagaceae species (*Fagus crenata*, *Quercus crispula* and *Quercus serrata*) in the Hyogo prefecture that lies next to the prefecture the study forest belongs to; then, nut production was synchronously very low in the autumn of 2010. Although we did not collect any masting data from our study site, a nut shortage may have affected the drastic deer population decrease of 2011. From 2011, the aforementioned three tree species did not produce nuts synchronously. This asynchronous nut production might have led to low deer population stability starting from 2011 onwards. The carrying capacity of deer might change not only spatial heterogeneity of habitat (the ratio of grassland, deciduous forest, and evergreen forest), which was reported in, but also temporal heterogeneity of habitats. The estimated deer abundance was 4.4 to 71.0 individuals per km² in this study. It is within the range of the estimated carrying capacity of sika deer (1.34 to 98.4 individuals per km²) in Yamanashi Prefecture in central Japan. Even in the open ecosystems, they found density dependent decline in the population growth rate. Therefore, in 2010, the density dependence in the population growth rate might occur in the study forest. We also need to consider the density-dependence in the population growth rate based on the habitat environment in future. In this study, the seasonal fluctuation of deer abundance was obscure. It is a little bit different from the past results obtained from road count surveys by vehicles. In the model, we considered the seasonal observation rate. It would be affected by leaf phenology of understory vegetation and braches of trees and deer activity including their lactation, mating and so on. It would purge the apparent seasonal fluctuation. However, the seasonal fluctuation of deer abundance gradually increased from April to December during 2013 and 2018. Though some deer exist in forests even during the winter, they migrate seasonally to avoid snow accumulation in heavy snow-covered areas. Therefore, the seasonal variation we detected may be due to the seasonal migration pattern in addition to the population recruitment through fawn births in early summer. The potential browsing pressure increase in the plant community during the summer may have negative effects on herbaceous plants, especially the one that grow in the summer and flowered in the autumn. In this area, as the plants that flower after midsummer are herbaceous and are more severely browsed compared to trees, the fitness of pollinators working from summer to autumn may critically decrease due to a shortage in their flower resources. The 95% credible interval (CI) of *hm* and *rho* ranged from −8.20 to −6.80 and 0.01 to 0.15, respectively. These results suggest that nuisance control could be useful in decreasing deer populations and are similar to past results. On the other hand, a previous study pointed out the difficulties of increasing hunting pressures because Japanese hunters were getting older. To establish an effective deer abundance management program under this circumstance, the development of simple and inexpensive capture methods is urgent. Late snowfall substantially affects the mortality of *C*. *nippon*. In *Cervus elaphus* in Norway, winter harshness affects first-year survival but not the survival of adults. In this study, the 95% CI of *a* included 0. This suggests that snowfall may have slightly affected deer mortality during the winter in the present study. This is similar to results obtained from studying the alpine ungulate *Rupicapra rupicapra*, though their population dynamics are largely affected by summer temperature. At first glance, our results seem to suggest that the mortality rate during the winter will not change even if snowfall decreases due to global warming. However, as we mentioned earlier, deer inhabiting regions with heavy snowfall, migrate to safe areas during the winter and go back to their initial habitats after snowmelt. Thus, snowfall decrease due to global warming may decelerate the winter migration of deer and, subsequently, deer that remain on-site may intensively forage evergreen perennial plants during the winter season. In route E, the observation rate was higher than that in routes A and B. In this study, we did not consider the spatial pattern of deer. While route A is close to a village, route E is remote and located deep in montane forest. Therefore, human activity may have affected the observation rate. The topographic pattern could have affected route visibility, though we uniform ranges of observation 15 m width in all routes. Landscape characteristics such as evergreen forests and artificial grasslands affect deer abundance in local areas. As shown in, this study site consists of steep slopes and deep valleys. The differences in observation rates among routes may also be due to the differences in landscape characteristics in a local scale in the forest. However, our model did not fit well in road counts at route E. We need to treat the results carefully. In conclusion, we clarified the population dynamics of deer not only annually but also seasonally. Snowfall accumulations did not affect population dynamics of deer in this study irrespective of higher mortality of deer during the winter. However, we need to pay attention to the effect the winter migration of deer has on plant communities because many deer migrated to another area during the winter and came back before the summer. Although we could not grasp the population dynamics during the snow accumulation season, in warmer winters, more deer may remain in the forest. Thus, a warmer winter may lead to degradation of evergreen perennial plant communities during the winter and early spring. Additional investigation on evergreen perennial plants could help examine the effect of deer browsing during the winter. In contrast to snowfall accumulations, nuisance control had an effect on the population dynamics of deer. Even in wildlife protection areas and national parks where hunting is regulated, nuisance control could be effective in buffering the effects of excessive deer browsing on forest ecosystems as well as plant communities, under the absence of potent predators. # Supporting information The authors would like to thank all the members who participated in this monitoring study. [^1]: The authors have declared that no competing interests exist.

# Introduction Ammonia is a commonly found pollutant in aquatic environments around the world. This compound can be found naturally, but there is also an additional contribution from sewage effluents, industrial waste, and agricultural run-off. The presence of ammonia in freshwater has been associated with the acidification of rivers and lakes, eutrophication, and direct toxicity to aquatic organisms. The toxicity of this compound on aquatic organisms will depend on the chemical form of ammonia, pH, and temperature. Furthermore, it will depend on the time of exposure. This compound damages the gills, liver, kidney, spleen and other organ tissues of fish, therefore causing breathing difficulties. This may lead to physiological alterations and, eventually, exhaustion or death. Ammonia can cause cell damage and can also affect the antioxidant defence system, thus altering the levels of oxidative stress in fish. Ammonia can also alter fish behaviour. Fish exposure to sub-lethal concentrations of ammonia can reduce swimming activity, foraging behaviour, and the ability to flee from predators. Behavioural analyses are commonly used in ecotoxicology as indicators of sub- lethal toxicity in aquatic animals, and an increasing body of evidence has demonstrated the effectiveness of this approach in a wide range of exposure scenarios. Fish exposed to increased ammonia concentrations experience difficulty in eliminating this metabolite from the body and, therefore, prolonged exposures to ammonia promotes its accumulation in fish. Several studies indicated that fish pre-exposed to episodes of pollution by inorganic nitrogen compounds and heavy metals could be more tolerant to these pollutants by acclimation. In these studies, it was shown that fish pre-exposed to sub-lethal concentrations of a pollutant exhibited increased tolerance to exposure to high concentrations of the same pollutant. Fish pre-exposed to sub- lethal concentrations of ammonia pollution could tolerate high concentrations of this compound by increasing the ammonia excretion rate as well as by favouring the evolution of adaptive mechanisms. These mechanisms have also been shown to work with other types of stressors such as hypoxia, salinity, and temperature changes. All these studies analyse the effect of fish pre-exposure from a biochemical and physiological point of view. The aim of this study was to analyse, under experimental conditions, the effect of sublethal ammonia concentrations on the swimming activity and feeding behaviour of wild-caught fish that had been pre-exposed for a long time to this compound. The species selected for this study was the Mediterranean barbel, *Barbus meridionalis* (Risso 1827), a freshwater fish endemic to the northeast Spain and southeast France. Fish from a long-term polluted river and fish from a pristine stream were exposed to sublethal ammonia concentrations in the laboratory. Fish stress responses were complemented using biomarkers. The analysis of biomarkers may provide valuable information by assessing the activity of enzymes/markers involved in energy metabolism, detoxification, antioxidant defences, and oxidative stress. In this study, biomarkers included lactate dehydrogenase (LDH), which is involved in anaerobic metabolism; glutathione *S*-transferase, a xenobiotic that metabolizes II enzyme response; glutathione (GSH) levels, which aid maintenance of the cell redox equilibrium as well as being a powerful antioxidant; catalase (CAT EC 1.11.1.6—reduces H₂O₂ to water) an antioxidant enzyme involved in detoxifying reactive oxygen species and markers of oxidative tissue damage such as lipid peroxidation. It has been suggested that *B*. *meridionalis* is relatively tolerant to organic pollution and, globally speaking, more tolerant to pollution than other cyprinid species. It was hypothesized that fish previously exposed to ammonia in the wild should have a higher tolerance to this compound than fish coming from unpolluted waters. # Materials and methods ## Study area and fish sampling Two sites (polluted and unpolluted) were sampled in the Besòs River basin (NE Spain). In both sites, there was prior knowledge about the existence of a population of *B*. *meridionalis*. The polluted site was located in the Congost River, a 43 Km long tributary in the Besòs basin, 50 m downstream the Granollers WWTP (41°56’97.31” N, 2°27’15.66” E). The unpolluted site was located in the Castelló stream, a pristine 3 Km long tributary inside the San Llorenç del Munt i l’Obac Natural Park (41°65’16.97” N, 2°06’11.18” E). The concentration of total ammonia nitrogen (TAN) in the polluted site (Congost River) ranged from 0.54 mg/L to 24.70 mg/L between 2011 to 2015 (data provided for Granollers Town Council). In the unpolluted site (Castelló stream), the concentration of TAN ranged from 0.00 mg/L to 0.02 mg/L during the same period. In this stream, there is no urban nucleus or any type of agricultural or industrial activity. Although ammonia is not the only pollutant present at these two sites, it is one of the most frequently found, not only in this river but in all rivers of NE Spain. shows the physical-chemical parameters analyzed at the two sites for the sampling month for Granollers Town Council (polluted site) and Fortuño et al. (unpolluted site). Other contaminants, such as contaminants of emerging concern (CEC) (pesticides, metals, industrial solvents, pharmaceuticals, and personal care products), could be found in other sites across these basins. Fish were sampled by electrofishing using a portable unit which generated up to 200 V and 3 A pulsed D. C. Fish collection was approved by the Department of Agriculture, Livestock, Fisheries and Food of the Autonomous Government of Catalonia, Spain (Permit number AP/003). A total of 72 individuals (40 in the polluted site and 32 in the unpolluted site) ranging from 5.5 to 10.8 cm were caught in January 2016. No differences in furcal length (FL, mean ± SD = 7.79 ± 1.31 cm) were found between the fish of the two sites. Once in the laboratory, fish of each site were acclimatized separately in 260 L aquaria over 21 days in clean dechlorinated water (there were 10–12 fish per aquaria). Chlorine elimination was achieved by storing water from the drinking supply net in 200 L containers for 48 h. According to Kroupova et al. fish affected by nitrite poisoning that were placed in clean water for over six days recovered the normal haematological parameters. Therefore, a period of 21 days seemed sufficient for the fish from the polluted site to recover normal physiological parameters. Aquaria were set in an acclimated room (20°C) under a 12 h light: 12 h dark photoperiod. All 260 L aquaria had the same equipment (biological filter and air diffusor), substrate (mix of sand, gravel, and coral with a proportion 2:2:1), and enough artificial refugees (PVC tubes and plastic plants) for reducing fish stress. Fish were fed “*ad libitum*” twice a day with frozen red chironomid larvae. A periodical cleaning of aquaria and partial water renovation (one-third of the volume) were carried out every 24 h. Physiochemical water conditions (mean ± SD) were controlled daily in the 260 L aquaria (water temperature = 21.97 ± 0.98°C, pH = 8.30 ± 0.27, NO₃^- = 5.63 ± 1.70 mg/L, NO₂^- = 0.00 ± 0.00 mg/L, NH₄⁺ = 0.00 ± 0.00 mg/L, and water hardness = 10.50 ± 4.36). These parameters did not show significant differences between aquaria during fish acclimatization. ## Experimental design After the acclimatization period, fish pre-exposed to ammonia pollution in the wild (hereafter, pre-exposed fish) and fish from the unpolluted site (hereafter, non pre-exposed fish) were exposed to four TAN treatments (0, 1, 5, and 8 mg/L) as follows: each fish was placed in individual 20-L aquaria (40 cm large x 20 cm height x 25 cm deep) and transferred to the room where the experiment was carried out. The aquaria were divided into four groups, and a treatment was randomly assigned to each group. For the Congost river (pre-exposed fish), there were ten aquariums per treatment (n = 40), while for the Castelló stream (non pre-exposed fish), there were eight aquariums per treatment (n = 32). Aquaria were positioned in two rows, side by side, within each group. In order to reduce fish stress, the lateral walls between neighboring aquaria were left transparent. To avoid fish interaction with the environment, the external and frontal walls as well as the bottom of aquaria, were covered by blue acetate sheets. Before starting the experiment, each fish was acclimatized to its 20 L aquaria for four days and fed daily with red chironomid larvae. During these four days of fish acclimatization, partial water changes were carried out every day, and TAN concentrations were measured with indophenol blue spectrophotometric method (in all aquaria TAN concentration was maintained at 0 mg/L; mean ± SD = 0.00 ± 0.00 mg/L). Next, fish were exposed to the assigned TAN treatment for eight days. The experiment was first carried out with the pre-exposed fish. After eight days, the experiment was repeated with the non pre-exposed fish. The TAN concentrations per experimental aquaria were achieved by adding analytical grade ammonium bicarbonate solutions (NH₄HCO₃, Sigma-Aldrich, Barcelona, Spain). These solutions were dispensed with automatic pipettes after water changes. Daily cleaning of aquaria and a two-third of the water volume renovation were carried out with dechlorinated water to guarantee the experimental conditions. TAN concentrations were measured daily by the indophenol blue spectrophotometric method. Once the absorbance values had been recorded for each sample, NH₄⁺ concentration was calculated using the equation of the calibration curve, and the proportion of the NH₃ form was calculated following Thurston et al. procedures. During the experiment, the aquaria group of each TAN treatment was visually isolated from the researchers with opaque curtains. In order to observe the activity of fish, a PVC tube (4 cm diameter x 13 cm length) was placed in each 20 L aquaria as a fish refuge. In order to observe feeding activity, fish were fed above satiation requirements (20 red chironomid larvae per fish were sufficient to quantify satiety). Fish behaviour was recorded with an overhead shot for each group of aquaria (TAN treatment) using a Sony HD (HDR-SR1E) camera. The experiment lasted for eight days, and recordings were made on alternative days (four days) between 9:00 and 12:00 AM. Every day, the recording order of each group of aquaria (TAN treatment) was established at random. Fish were only fed during the recording days. Two behavioural variables per individual were analyzed from video recordings: swimming activity (during 10’) and feeding behaviour (until fish stopped eating). The swimming activity was analysed by three variables: (1) “Swimming”, amount of time during which fish make displacements of the body using body or fin movement as propulsion (s), (2) “Not visible”, amount of time during which the fish was not visible because it was remaining inside the shelter (s), and (3) “Resting”, amount of time fish spent lying motionless on the bottom of the aquaria (s). Total swimming activity was expressed as a percentage of the total observation time. Feeding behaviour was analysed by measuring: (1) “Latency”, defined as the amount of time the fish took to start touching the food (s); (2) “Voracity”, defined as the number of chironomid larvae the fish ate in one minute and (3) “Satiety”, defined as the total number of red chironomid larvae fish eaten until they either stopped eating or they started spitting out the food. The concentration of NH₃ (mean ± SD, mg/L) for each TAN treatment was not significantly different between pre-exposed and non pre-exposed fish (GLM): \[0 mg/L\] = 0.007 ± 0.010, \[1 mg/L\] = 0.139 ± 0.077, \[5 mg/L\] = 0.534 ± 0.218, \[8 mg/L\] = 0.645 ± 0.237). Physiochemical parameters were controlled daily for each 20 L aquaria during the experiment. No differences in physiochemical parameters (mean ± SD) were found during the experiment between fish from the two sites and between the aquaria of each TAN group (GLM) (water temperature = 21.27 ± 0.45°C, pH = 8.33 ± 0.18, NO₃^- = 4.81 ± 0.68 mg/L, NO₂^- = 0.00 ± 0.00 mg/L, and water hardness = 15.05 ± 3.76;). All applicable international, national, and/or institutional guidelines for the care and use of animals were followed. The scientific procedure of this work was approved by the Animal Ethics Committee of the University of Barcelona (registration N° 9296), which follows European Directive 2010/63/UE on the protection of animals used for scientific purposes. One of the co-authors holds a category C FELASA certificate that regulates the use of animals for experimental and other scientific purposes. ## Biochemical determination For the biochemical determinations, fish were anesthetized on ice at the end of the experiment and euthanatized by decapitation. Biomarkers were analysed in the liver tissue for each individual fish, according to Faria et al.. The following reagents were purchased from Sigma-Aldrich (St. Louis, MO, USA): potassium phosphate dibasic (K₂HPO₄); potassium phosphate monobasic (KH₂PO₄); potassium chloride (KCl); ethylenediamine-tetraacetic acid, disodium, salt, dihydrate (EDTA); hydrogen peroxide (H₂O₂); reduced glutathione (GSH); sodium azide; 1-chloro–2,4–dinitrobenzene (CDNB); glutathione S-transferase, from equine liver (GST) (EC 2.5.1.18); monochlorobimane (mCB); sodium pyruvate; β-Nicotinamide adenine dinucleotide, reduced dipotassium salt (NADH), 2,6-di-tert- butyl-4-methylphenol (BHT); 1-methyl-2-phenylindole (MPI); 1,1,3,3-tetramethoxypropane (TMP) and Bradford reagent. All the other chemicals were analytical grade and were obtained from Merck (Darmstadt, Germany). Except for catalase activity, where a cuvette assay was used (Life Science UV/Vis Spectrophotometer DU® 730, Beckman Coulter–Fullerton, CA, USA), all the bioassays were performed in microplates (Synergy 2 Multi-Mode Microplate Reader, BioTek® Instruments–Vermont, USA). Liver tissue was homogenized in ice-cold 0.1M phosphate buffer with 150mM KCl and 0.1mM ethylenediamine-tetraacetic acid, disodium, salt, dihydrate (EDTA), then centrifuged at 10 000xg, 4°C for 30 minutes. The supernatant was collected, aliquoted, and stored at -80°C for biomarker determination. CAT activity was measured by estimating the decrease in absorbance at 240 nm due to H₂O₂ (50 mM H₂O₂ in 80 mM phosphate buffer, pH 6.5) consumption (extinction coefficient 40 M^-1cm^-1) according to Aebi. Reaction volume and time were 1 mL and 1min, respectively. GST activity towards CDNB was measured as described by Habig et al.. The reaction mixture contained 0.1M phosphate buffer (pH 7.4), 1 mM CDNB and 1 mM GSH. The formation of S– 2,4 dinitro phenyl glutathione conjugate was evaluated by monitoring the increase in absorbance at 340 nm during 5 minutes. Enzyme activity was determined using GST’s extinction factor coefficient of 9.6 mM^-1 cm^-1. Results were normalized by tissue total assay protein content. Reduced glutathione (GSH) quantification was adapted from zebra mussel digestive gland according to Kamencic et al.. It consists of adding 0.1mM of mCB along with 1U/ml of GST to each sample. Then the GSH present in the cells forming a GSH-mCB complex is measured fluorometrically at excitation: emission wave-length of 360:460 nm, after an incubation period of 90 minutes at room temperature and protected from light. The total content of GSH was then extrapolated from a GSH standard curve determined under the same physical and chemical conditions as the samples, and the results were normalized by the tissue wet weight (g ww). Lactate dehydrogenase (LDH) activity was determined according to Diamantino et al., by monitoring the absorbance decrease at 340 nm due to NADH oxidation. The reaction contained 100 mM phosphate buffer (pH 7.4), 0.15 mM NaOH, 1.18 mM pyruvate and 0.18 mM NADH. Lipid peroxidation (LPO) was determined by quantifying the levels of malondialdehyde (MDA), according to Esterbauer et al.. The MDA assay was based on the reaction of the chromogenic reagent 1-methyl- 2-phenylindole with MDA at 45˚C, giving rise to a chromophore with absorbance at 586nm. Samples were incubated with 5mM 1-methyl- 2-phenylindole in acetonitrile:methanol (3:1 v/v), 5.55% of HCl and 0.01% BHT at 45˚C, for 40 minutes. Absorbance was read at 560nm, and MDA content in each sample was extrapolated from the standard curve of 1,1,3,3-tetramethoxypropane (TMP) treated under similar conditions as samples. The final results were normalized by tissue wet weight (g ww). Total protein concentrations were accessed by the Bradford method using bovine serum albumin (BSA) as a standard. ## Statistical analyses Differences between TAN treatments and sites were analysed by means of a generalized lineal mixed model (GLMM). For swimming activity, “Swimming” and “Not visible” were analysed separately and used as a dependent variable. The variable “Resting” was not analysed as it was a complementary variable to the other two. For feeding behaviour, “Latency”, “Voracity” and “Satiety” were analysed separately and used as dependent variables. In all cases, “site” (2 levels: pre-exposed and non pre-exposed fish) and “TAN treatments” (4 levels: “0”, “1”, “5”, and “8” mg/L TAN) were used as factors together with their interaction. The gamma distribution was assumed in the analysis of swimming activity, and the Poisson distribution was assumed in the analysis of feeding behaviour. The variable “Individual” was added to the model as a random factor. Biomarker responses across fish from the two sites (pre-exposed and non pre- exposed fish) and TAN treatments were analysed through a lineal model (LM) with the same factors (“site” and “TAN treatments”). Differences between TAN treatments against control ones were further compared using Dunnett’s post-hoc test. All analyses were conducted with R 3.4.3. GLMM assuming a Poisson or a gamma distribution was performed using glmer (package “lme4”:). Non-significant interactions were removed from the final models. Homogeneity and normality of residuals were visually checked for all models. All significant differences are *P* ≤ 0.05. # Results ## Behavioural variables The “Swimming” and “Not visible” GLMM models showed no significant effect of TAN treatments within and across sites (interaction) (*P* \> 0.05). Only a significant effect of the site (pre-exposed and non pre-exposed fish) (*P* \< 0.001) was shown for these two variables. The swimming activity of fish that had been pre-exposed to ammonia pollution in the field was lower than that of non pre-exposed fish. Non pre-exposed fish swam for a longer time (67% of the time; mean = 643.89 s; 95% confidence interval = 504.10–891.34) than pre-exposed fish (57.3% of the time; mean = 548.92 s; 95% confidence interval = 444.92–716.43) regardless of the TAN treatments they were in. Similarly, non pre-exposed fish spent significantly less time hidden inside the shelter (mean = 206.76 s; 95% confidence interval = 150.01–332.90) than pre-exposed fish (mean = 232.99 s; 95% confidence interval = 165.76–392.43). The analysis of feeding behaviour showed no significant effects of the interaction between TAN treatment and site for none of the variables (*P* \> 0.05). For “Latency” GLMM model showed a significant effect between sites (*P* \< 0.002) but not for TAN treatments (*P* \> 0.05). Non pre-exposed fish had a higher latency than pre-exposed fish. In contrast, GLMM models for “Voracity” and “Satiety” variables showed a significant effect between sites and TAN treatments within each site (*P* \< 0.001). Non pre-exposed fish had a higher voracity and were satiated later than pre-exposed ones. In both cases (sites), significant differences were found between the control TAN concentration (0 mg/L) and the three TAN treatments (1, 5, 8 mg/L) for “Voracity” and “Satiety” variables. ## Biochemical determination The results of the analysis of biomarkers show that there were significant (*P* \< 0.05) differences between sites (pre-exposed and non pre-exposed fish) in three out of the five studied biomarkers. TAN treatment within and across sites (interaction) also affected the activities of CAT, GST, and levels of LPO. Pre- exposed fish had lower CAT activities and lower levels of GSH and the activities of CAT and levels of LPO increased across TAN treatments. In fish from both sites, the activities of GST were enhanced at 1 mg/L of TAN. # Discussion Chronic exposure to pollution by inorganic nitrogen compounds (NH₄⁺, NH₃, NO₂^-, HNO_2, and NO₃^-) has effects on the reproduction, growth and survival of freshwater fish. Specifically, exposure to NH₄⁺ and NH₃ (TAN) pollution can cause gill damage, anoxia, disruption of blood vessels and osmoregulatory activity (damage to the liver and kidneys), and a decrease in the effectiveness of the immune system. In addition, NH₄⁺ ions contribute to an internal reduction of Na⁺ which, in turn, increases the toxicity by NH₃. All these effects can result in a reduction in fish feeding activity, fecundity and survival, leading to a reduction in the size of populations. In the present study, wild-caught fish pre-exposed for a long-term period to ammonia pollution in a contaminated river near a WWTP showed an altered behaviour and suffered from increased physiological stress compared to non pre- exposed fish from a pristine stream. Analysis of fish swimming activity showed that, regardless of the TAN treatments, pre-exposed fish were less active and spent more time hiding in the refuge than non pre-exposed fish. The only studies on the effects of ammonia on fish swimming activity have been conducted on salmonids in laboratories or farms. According to Tudorache et al. and Wicks et al., the swimming activity of salmonids is reduced at concentrations between 0.2–1 mg/L of TAN (that is, at 0.009–0.04 mg/L NH₃). Pre-exposed fish spending more time inside the PVC shelters (“not visible” time) might indicate that these fish had their exploratory activity altered. A decrease in the exploratory activity has been reported in several fish species exposed to crude- oil pollution, pesticides, and pharmaceutical products. The feeding behaviour of *B*. *meridionalis* was also altered. Pre-exposed fish had lower voracity than non pre-exposed fish regardless of the TAN treatments (0, 1, 5, and 8 mg/L TAN). Within each site (pre-exposed and non pre-exposed fish), lower voracity was observed from the lowest TAN concentration (1 mg/L). A reduction in voracity has been reported for salmonids under TAN concentrations from 1 to 3 mg/L. According to Schram et al., in a non-salmonid fish (*Clarias gariepinus*), food consumption was also drastically reduced at TAN concentrations higher than 1 mg/L. In the present study, latency (the time that the fish took to start touching the food) was lower in pre-exposed fish, regardless of the TAN treatment. Low latency has been related to a low capacity to find food and capture prey. Furthermore, several studies relate lower latency with lower efficiency to flee from predators. The present study was conducted under a concentration of ammonia within the range of LC₅₀ (the tested range was from 0.007 to 0.645 mg/L NH₃). However, the tolerance to NH3 in cyprinids could be higher. The LC₅₀ for cyprinids ranked between 0.685 and 1.720 mg/L NH₃. For cyprinids, the sublethal concentrations in which negative physiological effects begin to be observed has been described in a range of 0.105–0.247 mg/L NH₃. The limits of tolerance to this and other compounds are variable depending on each fish species, so that it would be necessary to investigate their effects under natural conditions. Antioxidant enzyme activities such as those of CAT and reduced glutathione has been reported to be important antioxidant mechanisms against oxidative stress- mediated effects of ammonia in fish. Pre-exposed fish (from the polluted site) had lower constitutive levels of the above-mentioned antioxidant defences and consequently were unable to detoxify the excess of reactive oxygen species (ROS) generated by ammonia, leading to enhanced tissue levels of oxidative damage measured as LPO. Interestingly, only in fish from the polluted site (pre-exposed fish), the activities of CAT increased in individuals exposed to ammonia, thus indicating that the exposure to this compound increased ROS and, hence, triggered the antioxidant defences of these fish. In fish from the unpolluted site (non pre-exposed fish), the high constitutive levels of antioxidant defences protected them from ROS generated by ammonia. Sinha et al. (2014) reported that fish species intolerant to ammonia, such as trout, rely mainly on glutathione-dependent defensive mechanisms, while more tolerant species, such as carps, utilize antioxidant enzymes such as CAT and ascorbate. High tolerant species, such as goldfish, use both of these protective systems and show more effective anti-oxidative compensatory responses towards oxidative stress induced by ammonia. Thus, our results are in line with previous studies, as *B*. *meridionalis* considered a tolerant species to ammonia. Results in this study indicated that the exposition of fish to high ammonia concentrations did not guarantee, at least short term, the recovery of good health status and/or greater tolerance to a high concentration of this compound. Fish pre-exposed to ammonia pollution in the wild showed an altered behaviour at the control concentration (0 mg/L TAN). This could be a consequence of pre- existing physiological problems due to exposure to ammonia and other pollutants in nature. The feeding behaviour and the response to the oxidative stress of *B*. *meridionalis* (both pre-exposed and no pre-exposed fish) follow the same pattern, reacting equally to the first 1 mg/L TAN treatment. However, pre- exposed fish had a more marked response in feeding behaviour and biomarkers under the different treatments of TAN. A reduction in food intake is directly related to both lower growth and a low rate of protein synthesis. Reported studies have shown that low protein synthesis rates represent a large proportion of energy costs in fish, and this has a direct impact on the growth efficiency of individuals. Alteration of the behaviour parameters analysed in this study can be extrapolated to other traits such as exploration activity, boldness, and ability to avoid predators. Ammonia can affect social interactions as well by altering dominance relationships, hierarchical dynamics, and predator-prey relationships. Ammonia pollution is a common problem in freshwater ecosystems. Despite the efforts of implementing the European Water Framework Directive (2000/60/RC) (2000), there are still many WWTPs that do not have tertiary purification systems of urban wastewater, which leads to an increase in nitrogen compounds in aquatic ecosystems. Improving water quality is an important key to enhance the conservation of river ecosystems. However, our results indicated that fish that previously survived in a polluted environment did not recover their health in more purified waters. In summary, although habitats are improving their environmental quality, the survival of fish populations that have been pre- exposed to contamination could be compromised. In freshwater ecosystems, which have suffered an 83% decline in vertebrate populations from 1970 to 2014, all factors affecting the survival of individuals are of great relevance. # Conclusions *B*. *meridionalis* pre-exposed in the wild to pollution by ammonia presented a swimming activity, feeding activity and the response to oxidative stress altered when placed in non-contaminated water under experimental conditions. Feeding behaviour and the biomarker response of *B*. *meridionalis* was affected by ammonia and pollution history. Pre-exposed fish (from a polluted site) had less voracity and satiated before fish from an unpolluted site (non pre-exposed fish). In addition, pre-exposed fish were more affected by the different TAN treatments, and these alterations appeared from the lowest concentration of TAN (1 mg/L). The results of the swimming activity showed that pre-exposed fish spent less time swimming and more time is hidden. However, this behavioural response was not related to the different TAN treatments and could be related to the damage caused by pre-exposure to ammonia and other pollutants present in the river. Fish pre-exposed to ammonia in the wild also had the antioxidant defences depressed and consequently were less tolerant to high concentrations of ammonia. Therefore, our results indicated that the recovery of water quality is not necessarily related to the restoration of fish health. There is a physiological cost of being adapt to pollution present in rivers. # Supporting information We thank I. Ramirez and F. López of the Department of Biochemistry and Molecular Biomedicine (University of Barcelona), for giving us access to their laboratory. We thank P. Fortuño (University of Barcelona) and X. Romero (biologist and superior technician of environment and natural environment of the Granollers Town Council) for provided data. We also thank J. Guinea and P. Manning for assistance in the laboratory tasks. The authors are grateful to V. Bonet for the English review. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. [^1]: The authors have declared that no competing interests exist.

# Introduction The Ebola Virus Disease (EVD) outbreaks in West Africa infected more than 28,000 people and took more than 11,000 lives in Guinea, Sierra Leone, and Liberia. The greatest tools for breaking the chain of transmission in these outbreaks have been (1) rapid isolation of EVD patients and (2) secure and dignified burials for victims of EVD. The success of these strategies depends on the timely and reliable identification of live EVD patients and suspect deaths through laboratory testing. The gold standard for EVD diagnosis is PCR, which can be done in less than 4 hours, but requires human resources, equipment, facilities, and infrastructure unavailable in remote areas. The time to transport patients or samples to PCR capable facilities can delay results for days. Delays in test results or the need for travel can increase community resistance to public health interventions. Rapid Diagnostic Tests (RDTs) for EVD can alleviate some of the challenges presented by PCR testing by quickly providing test results at the point of care. To expand testing services for EVD, the National Coordination for the Fight against Ebola in Guinea started a pilot program for the implementation of RDTs in the most affected regions in October 2015. The program was led by the Centers for Disease Control and Prevention (CDC) in partnership with the National coordination for the Fight against Ebola, the National Institute for Public Health, the Guinean Red Cross and International Federation of Red Cross and Red Crescent Societies, and the World Health Organization (WHO). The initial roll- out took place in the prefecture of Forécariah where active EVD transmission was present. Red Cross volunteers, who are responsible for the Secure and Dignified Burials of EVD victims, were trained to perform RDTs on the deceased and laboratory technicians at sentinel sites were trained to perform RDTs on suspect patients and patients who die in medical facilities. Previous articles have described an initial evaluation of the pilot implementation of the EVD RDTs at 15 sites in Forécariah and a baseline assessment of the use of EVD RDTs. Based on lessons learned from the initial RDT implementation in Forécariah, the RDT program for patient testing rolled out more broadly in December 2015 by training (or re-training) laboratory technicians at 16 sites in Forécariah, 11 sites in Conakry, and 5 sites in the prefectures in the Forest Region that are contiguous with Liberia (Macenta, Guéckédou, N’Zérékoré, Lola, and Yomou). Two additional sites were added in the prefecture of Guéckédou in February 2016. The trainings were followed up with weekly lab visits, initial visits to reinforce the RDT training and subsequent visits by CDC or WHO epidemiologists for data collection and supportive supervision. This paper presents the results of an evaluation of three operational measures (acceptability, feasibility, and quality assurance) conducted during the implementation of the RDT program in Guinea, as well as concordance data between EVD RDTs and PCR testing in the field. # Methods ## Context and population The capital city of Conakry and the prefectures of Forécariah and the Forest Region were some of the areas most affected by Ebola outbreaks. The decision to implement the RDT program in each region was based on existing EVD transmission, a high risk of transmission from a neighboring region’s outbreak, and/or a large EVD survivor population with potential viral persistence in body fluids. The 34 sentinel labs participating in the EVD RDT program were located in Health Posts (4), Health Centers (17), Communal Medical Centers (4), and Prefectural (6) and National (3) Hospitals. Therefore their level of services, staffing, and population they served varied considerably. ## EVD Rapid Diagnostic Test The OraQuick^® Ebola Rapid Antigen Test (OraSure Technologies, Inc., Bethlehem, PA) was chosen for the RDT program in Guinea based on its high manufacturer reported sensitivity (84% (95% Confidence Interval (CI): 63.92–95.46)) and specificity (98.0% (95% CI: 89.35–99.95)) for whole blood and its broad temperature tolerance for both storage (2–30°C) and testing (15–40°C) conditions. The sensitivity of the test is related to the viral load in the sample, with 100% sensitivity (95% CI: 86.77–100.0%) for samples with PCR Ct range of 15 to 24 (high viral load), but 84.0% sensitivity (95% CI: 63.92–95.46%) at the full PCR Ct range of 15 to 34 (high to low viral load). OraQuick^® has an Emergency Use Authorization issued by the U.S. Food and Drug Administration in July 2015 and was authorized for use by the Government of Guinea. The OraQuick^® Ebola Rapid Antigen Test is a lateral flow, single-use immunoassay which allows qualitative detection of Ebola antigens from the whole blood of patients or saliva of corpses in 30 minutes. The sample is added to the OraQuick^® RDT device, then the device is inserted into a vial of developer solution to facilitate the capillary flow of the specimen into the device and onto an assay strip with a Test Zone and Control Zone. As the specimen flows through the device, Ebola antigens from the specimen are bound by Ebola antibody labeled gold colorimetric reagent. If Ebola antigens are present the labeled complexes bind to the Test Zone resulting in a purple line, and if they are not present the Test Zone will remain colorless. The remaining colloidal gold continues to migrate and binds to the Control Zone resulting in a purple line to demonstrate there was adequate flow and the test was valid, regardless if the sample was positive or negative for Ebola virus. Positive results may be interpreted as soon as lines are visible at the Test and Control Zones, however negative results must be read 30 minutes after inserting the device in the developer vial to allow adequate time for migration of the sample. The intensity of the line color is not directly proportional to the amount of virus in the specimen; the test is interpreted as reactive or non-reactive. ## EVD RDT program training More than 200 healthcare workers, laboratory technicians and laboratory trainees in Forécariah, Conakry, and the Forest Region received one-day trainings by the EVD RDT program partners. The trainings included lectures on the eligibility criteria and algorithms for the use of the EVD RDT, how to use the OraQuick^® EVD RDT, quality assurance, data collection and supervision tools, waste management, and communication strategies for patients and family members of the deceased as well as practice sessions with the OraQuick^® EVD RDTs and putting on and removing Personal Protective Equipment. Evaluations were conducted after each training to improve the content, and different instructors were involved in different regions, therefore there were some minor differences across trainings. The initial training in Forécariah in September taught that the OraQuick^® EVD RDT should be read at 20–30 minutes, but this guidance was updated to exactly 30 minutes in all subsequent trainings. ## Algorithms for the use of EVD RDTs Several algorithms were used to screen for potentially unknown EVD contacts with the EVD RDTs during different time periods and in different locations. While the eligibility criteria for live patients to receive an RDT varied over place in time (i.e. matching the suspect case definition for EVD, having a febrile illness, having fever plus three other EVD symptoms), the procedures after receiving the test remained nearly the same. Patients with reactive RDTs were immediately isolated pending PCR confirmation. In the vast majority of cases, patients with non-reactive RDTs were investigated for alternative diagnoses and did not require PCR confirmation. Therefore patients with non-reactive RDTs rarely received PCR confirmation. All deaths alerted to the surveillance system and all corpses in hospital morgues were eligible for EVD RDTs. The EVD RDT was always performed at the same time as a swab was taken for PCR confirmation. A presidential edict in place for most of the 2014–2015 Ebola outbreak required all deaths in prefectures with active transmission to receive a Secure and Dignified Burial from the Red Cross. When the edict was lifted in December 2015, a non-reactive RDT result allowed the family to proceed with a traditional burial and a reactive RDT required a Secure and Dignified Burial in communities without active transmission. All screening algorithms required only one EVD RDT to be performed per patient or corpse, unless the first result was invalid, in which case a second RDT would be performed. If both RDTs were invalid then the algorithm proceeded to PCR testing. ## Concordance data Concordance data for EVD RDTs and PCR were collected and monitored throughout the use of the RDTs to identify potential adverse events due to false positive or false negative results. On some occasions healthcare workers deviated from the algorithm and performed multiple EVD RDTs on one patient with discordant results; therefore number of tests was reported rather than number of persons. Given the limited number of new EVD cases in Guinea during the implementation of this pilot project, and that PCR tests were not required for all EVD RDT tests on living patients, paired EVD RDT and PCR tests were rare for living patients. PCR testing events were reported through the National Coordination for the Fight against Ebola’s surveillance system. The case histories of these PCR testing events, investigated by field epidemiologists from CDC and WHO, would indicate if the patient had visited a sentinel lab site and the result of their RDT, if any. Concordance data for the deceased were recorded by the Red Cross, who kept a database of RDT and PCR results, and by field epidemiologists in the Forest Region who kept line lists of RDTs performed in sentinel labs. The line lists from the Forest Region were matched with the National Coordination’s PCR testing database based on name, age, date, and location. ## Operational measures After three months of operation, three operational measures were assessed in the laboratory-based RDT program at all 34 sentinel sites in Guinea: (1) acceptability of the test, (2) feasibility, and (3) quality assurance. The indicators and data sources used for each measure are displayed in. Acceptability of the test includes both acceptability to patients and their families as well as acceptability to the technicians who perform the test. Feasibility involves how well technicians have retained knowledge and skills from their RDT training and their subjective opinions on how easily RDTs are integrated into their routine work. It is also related to whether or not the basic infrastructure, equipment, and supplies needed to implement the RDT program are available at the sentinel site. Quality Assurance encompasses many general quality management skills for laboratories, including temperature logging, record keeping, stock management, and quality control testing. ## Questionnaire & Site visit checklist A technician questionnaire was designed to collect their opinions on the acceptability and feasibility of the tests as well as test their current knowledge levels. The questionnaire included qualitative and quantitative components. The questionnaire, which was written in French, was piloted at two labs in Conakry and small adjustments in wording were made based on feedback from the technicians. The final version is two pages of questions which can be completed in 15 minutes or less. In addition to demographic questions, the first page has two quantitative questions and five qualitative questions to capture the acceptability and feasibility of the EVD RDTs. The second page assesses knowledge related to quality assurance through eight quantitative questions testing current knowledge of the EVD RDTs. For example, images of four possible EVD RDT test results are given for the technicians to interpret. Questions 2, 3, and 4 were multiple choice, while questions 1, 5, 6, 7, and 8 were short answer. Only those technicians who were present during the site visit were asked to complete the survey. The responses were anonymous. Following the survey, the evaluator shared all the correct answers to the knowledge questions. A site visit checklist, completed by the evaluator, provided additional information on feasibility by examining the presence or absence of necessary materials for the RDT program, such as thermometers, timers, and log books, and whether they were used appropriately. One checklist was filled out for each sentinel site. The evaluator had a small supply of thermometers, timers, and forms to provide labs that were found lacking. The site visits and questionnaires were conducted in late March 2016 for the 11 sentinel sites in Conakry and 16 sentinel sites in Forécariah. The March 2016 outbreak of EVD in N’Zérékoré and Macenta delayed site visits and questionnaires for the 7 sites in the Forest Region until late April to mid May 2016. A total of 87 lab technicians were surveyed, 25 from Conakry, 33 from Forécariah, and 29 from the Forest Region. ## Practical exam Following the site visits and questionnaires, a member of the CDC Lab Team in Guinea (JL) visited 20 labs (4 in Conakry, all 7 in the Forest Region, and 9 in Forécariah) to observe lab technicians in the performance of quality control tests on EVD RDTs and provide constructive feedback. Reports from these site visits were coded for the key issues observed during test performance. ## Analysis Quantitative data were analyzed with descriptive statistics to get the overall and regional pictures of the operational measures. Differences in knowledge scores between trainees and civil servants, those who received primary and secondary training (those who were trained by their colleague and not program partners), and across regions were assessed using the Mann-Whitney-Wilcoxon test (for two groups) and the Kruskal-Wallis test (for two-plus groups) in R statistical software version 3.2.0. Qualitative data from the survey (one to three sentence written responses) were coded for major themes by two analysts and then summarized by the percent of respondents who addressed that theme in their responses. Only codes elicited by 5% or more of respondents are presented. ## Ethics statement The protocol for the use of EVD RDTs was approved as a non-research, program evaluation activity at CDC and authorized by the Guinean National Coordination for the Ebola Response. These data were collected as part of ongoing public health program monitoring and evaluation. Verbal consent was obtained from site supervisors and lab technicians before collecting observational data and responses to the questionnaire. # Results ## Concordance data The concordance between the EVD RDTs and PCR testing for both living patients and the deceased is shown in. shows tests, rather than persons, as sometimes more than one EVD RDT was performed on the same individual with discordant results. In false negative RDTs are highlighted in **bold** and false positive RDTs are highlighted in *italics*. No false negative RDTs were encountered in the deceased, but two occurred in living patients. Both false negative RDTs occurred when the RDT was used outside of the recommended protocol during the March 2016 outbreak, and both the case histories and PCR cycle threshold values suggest that the patients were at a stage of EVD with low viral load (i.e. early phase of illness or recovery phase). Both false negative RDTs were performed on known EVD contacts who should have been referred directly to PCR testing and skipped the RDT; one false negative RDT was performed by an untrained technician at a non-sentinel laboratory and the second was performed in the community (outside the laboratory) under poor conditions. None of the non-reactive RDTs that did not have a corresponding PCR test led to an outbreak event. There were eight false positive RDTs in living patients and five false positive RDTs in the deceased. The false positive rate for EVD RDTs in the deceased is 0.16% (5 of 3099 paired tests). The false negative rate for EVD RDTs in the deceased is 0% (0 of 3099 paired tests). As few of the RDTs on the living had paired tests, rates of false positives or negatives are not calculated. ## Characteristics of the sentinel laboratorians performing EVD RDTs At the 34 sites that were surveyed, 255 laboratory workers were reported to staff the laboratories. Of these, 131 (51.4%) performed EVD RDTs. The majority (n = 76, 58.0%) of the EVD RDT users were laboratory technician trainees, and the remainders (n: 55, 41.9%) were laboratory technician civil servants. The ratio of trainees to civil servants varied by area (0.5:1 in Conakry, 1.4:1 in Forécariah and 3:1 in the Forest Region). The questionnaire was administered to 87of the 131 laboratory workers who perform EVD RDTs; 48 (55.2%) of the respondents were trainees. Of those surveyed, 52 (59.8%) attended the primary training sessions organized by partners at the beginning of the roll-out in their region. Others received secondary training from those who had already been trained. ## Acceptability Survey respondents described eight major positive or good features of the EVD RDT, including: (1) the ability to have a diagnosis and properly orient the patient (51.2%), (2) the rapidity of the test (30.2%), (3) the reduction of EVD spread and protection of community health (18.6%), (4) immediate medical care for patients (14.0%), (5) the facility of the test (11.6%), (6) the ability to continue surveillance and give alerts (11.6%), (7) access to PPE and disinfectants for security during the test (8.1%), and (8) the reliability of the test (8.1%). These positive features were described by respondents from all three regions, except none of the responders from Conakry listed immediate medical care for patients. Twenty-two percent (22.0%) of respondents said there were no challenges or bad features of the EVD RDT. Other respondents described nine major challenges or bad features of the EVD RDT, including: (1) technical aspects of the test (sensitivity, specificity, reliability, etc.) (20.9%), (2) reticence or fear of violence from the patient or their entourage (19.8%), (3) the lack of financial motivation for the lab staff (12.8%), (4) difficulty in performing the test correctly (10.5%), (5) the lack of or difficulty with Personal Protective Equipment (10.5%), (6) ruptures of RDT stock (8.1%), (7) the lack of other materials or infrastructure needed to conduct the test (7.0%), (8) the lack of a designated location to don PPE and perform the test (5.8%), and (9) difficulties in the transmission of test results (5.8%). These challenges were elicited by responders from all three regions, however difficulty in performing the test correctly was not listed as a challenge in Conakry and in Forécariah neither a lack of other materials needed for the test nor difficulties transmitting test results were listed as challenges. About half of respondents (51.2%) noted no major changes in the EVD RDT program since it began at their site. The main changes observed included less reticence from the general population (12.8%), improved performance of the test (10.5%), no more outages of EVD RDTs (5.8%), and new procedures for orienting patients (5.8%). These last two observations were only reported from for Forécariah and Conakry, the locations of the longest running sites. The majority of respondents reported that they were concerned for their safety when performing the EVD RDT (69.8% worried, 23.3% not worried, 7.0% no response). Of those who gave reasons for their concern (n = 60), the majority stated they were concerned because despite the precautions they take there is never zero percent risk (60.0%). Others were concerned because EVD is a highly contagious and deadly disease (15.0%), because their PPE may not be correctly worn or complete (13.3%), because of reticence from the population (6.7%), or the lack of an isolation space in their facility (5.0%). Patient refusals of an EVD RDT were encountered at least once by 14.9% of respondents, however some of these incidents may be multiple lab technicians present with the same patient. Only one lab technician reported refusing to perform one EVD RDT, but has since started to perform the test. ## Feasibility On the ease of performing the EVD RDT: 11.6% described the RDT as easy to perform, while 10.5% listed correctly performing the test as a challenge. Furthermore, 10.5% of respondents stated that their performance of the test has improved over time. Overall, respondents performed well on the knowledge retention questions with an average score of 6.6 points of a possible 8. The average knowledge scores and the percent of respondents who answered correctly for different groups are listed in. There was no significant difference between the knowledge scores of trainees and civil servants (W = 738.5, p-value = 0.263) or those who received primary training from partners and those who received secondary training (W = 857.5, p-value = 0.098). There was a significant difference in the knowledge scores across regions (Kruskal-Wallis chi-squared = 28.05, df = 2, p-value \< 0.001), and each region was significantly different than the other (Pairwise comparisons using the Wilcoxon test, all p-values \< 0.05). Technicians in Forécariah (average score 5.8) performed the least well. Forécariah performed very poorly on Question \#1 about the timing to read the RDT result, with only 39.4% of respondents answering correctly. An equal number of respondents from Forécariah (13 of 33, or 39.4%) answered 20 to 30 minutes, rather than 30 minutes, as the read time for the RDT, which agrees with the original training given in September 2015. Staff from Conakry were also challenged by this question, with only 64.0% of technicians answering correctly; while 96.6% of the Forest Region technicians answered correctly. Respondents had little difficulty reading the reactive (93.1% correct) and non- reactive (98.9% correct) test results correctly, but had some difficulty reading the invalid test result with the incomplete “test” line (Question \#6, 63.2% correct). The majority of those who read the invalid test incorrectly described it as “reactive” (32.2%), which would be the more conservative response. Again, the Forest Region performed best (82.6% correct) and Forécariah (39.4% correct) the worst. lists by region the percentage of sites with the different elements of basic infrastructure, equipment, and supplies needed for EVD RDT quality assurance. Overall the sites were poorly equipped with the basic items needed, particularly for thermometers and up-to-date job aids. It should be noted that some surveyors counted personal cell phones as timers and others accepted only dedicated laboratory timers. A fridge, which is present at only 58.8% of the sites, is necessary for the storage of the external quality controls but not necessary for the storage of the RDTs. The sites were better equipped for infection prevention and control with biohazardous waste disposal, PPE available, and handwashing stations in the labs. Though the Forest Region reported only 14.3% of labs having a handwashing station in the lab, there are handwashing stations available outside the lab that are shared with the rest of the facility or patients at all sites. Labs that were lacking were provided with updated job aids, thermometers, timers, and RDT and PPE stock management forms during the site visit or soon afterwards. Problems with biohazardous waste management and handwashing stations were drawn to the attention of the lab staff. ## Quality assurance Recordkeeping was very poor, with only 32.4% of sites having fully complete results log books and 41.2% having a quality stock management system. While the logbooks were not fully complete, most sites had nearly complete records. Given that only 26.5% of sites had a thermometer, temperature logging was nonexistent at most sites. Only two of the sites with thermometers kept temperature logs. In the execution of the quality control tests with the RDTs, 70% of the sites observed–all sites in Forécariah, and about half the sites in Conakry and the Forest Region–did not use the calibrated capillary provided with the RDT kits or used the capillary incorrectly. Thirty percent (30%) of the sites observed did not have a laboratory timer or it was in poor condition and staff used personal cell phones to time the reading of the RDT. Seven sites in Forécariah (35% of sites observed) used an incorrect read time for the test. Three sites in Forécariah (15% of sites observed) were noted to have generally poor technique in the performance of the RDT. A correction and demonstration of proper technique was given to the staff on site as issues were identified. # Discussion Implementation of new diagnostics must take into consideration the existing level of infrastructure, equipment, and supplies in the laboratory system. The laboratory network in Guinea has very weak quality assurance and biosafety systems in place which must be built up in tandem with diagnostic capabilities. The need for quality assurance was taken into consideration during training, but these quality assurance systems will need to be reinforced and monitored over the long term. reports the major concerns highlighted by this evaluation and the recommendations or actions taken to improve the EVD RDT program. ## Acceptability Overall the EVD RDT has a high rate of acceptability among laboratory technicians and the general public, and the acceptability has increased over time. Most of the challenges or negative aspects of the EVD RDT described by respondents can be addressed through continued health promotion and education about RDTs and support for quality lab management systems. The mention of the technical aspects of the test (sensitivity, specificity, reliability, etc.) by 20.9% of respondents as a challenge was mostly in response to incidents of false positive or false negative tests and the need for PCR as a confirmation test. From a technical perspective, the number of false positive tests has been very low and the two false negative tests arose only when the EVD RDT was used outside of the defined algorithm. These points should be highlighted in subsequent training materials. Request for additional financial motivation by 12.8% of laboratory workers respondents is perhaps due to the fact that incentives have been provided in the past by organizations for projects and programs related to EVD response activities and other vertical public health programs. However, the use EVD RDTs was approved as an integral function of the public health laboratories in Guinea, rather than an accessory research program. Therefore it should not require additional payment; rather it should be integrated into regular tasks. Efforts should be made to ensure the transfer of responsibility for program supervision to the Guinean government and continued technical assistance from partners. ## Feasibility Respondents were mixed in their opinions about the ease of performing the EVD RDT but some also noted that their performance improved with practice. Objectively speaking, the EVD RDT is around the same level of difficulty as the malaria RDTs in wide use in Guinea, but the PPE requirements are more rigorous for the EVD RDT. Overall, the lab technicians had acceptable knowledge scores, however the lack of consistency on the 30 minute read time is very concerning, especially in Forécariah. The weaker performance of the Forécariah technicians may be a result of the original training in September 2015, which recommended a 20 to 30 minute read time. The protocol has since been updated to a read time at exactly 30 minutes, but re-trainings may not have reached all the original participants. Subsequent updates to the protocol need a dissemination strategy that reaches all technicians performing the test. The non-significant difference in test scores between the technicians that received primary and secondary training suggests that re-training one focal point in each lab who would then re-train their colleagues would be a viable strategy. Some sites that had received new equipment at the start of the EVD RDT program had lost or damaged the equipment by the time of evaluation. The lack of thermometers raises concerns about the proper storage of the RDTs. The lack of timers is particularly concerning given that an RDT non-reactive at the 30 minute read time could be read as a false positive in as little as four minutes post-read time (unpublished observation of an RDT performed on a healthy individual, CDC Guinea)–the exactness of the read time is very important. The lack of timers also becomes a biosafety concern when technicians are using their personal cell phones to time tests in the lab, especially in sites with few staff members where a colleague cannot start and stop the timer for the technician performing the RDT. Sites without refrigerators would have to receive external quality controls on ice and use them immediately upon receipt. Some sites had persistent issues with biohazardous waste disposal and the availability of PPE and handwashing stations. While many of these issues were addressed during the site visit, inventories of and improvements to the basic infrastructure, equipment, and supplies should continue on a regular basis. ## Quality assurance The laboratory network in Guinea has little experience with quality assurance systems, thus temperature monitoring, new record keeping registers, and other lab management tasks complementary to the performance of diagnostic tests may be perceived as additional work superfluous to the use of a new diagnostic. More work is needed to integrate quality lab management systems into daily lab practice, and it should be approached on a broader basis than quality management for one diagnostic test. The difficulties identified in the use of the calibrated capillary must be addressed in future training sessions with hands-on practice. Supervision should continue on a regular basis and should pay particular attention to any changes in protocol, such as the change in recommended read time, and adherence to the testing algorithm. Incidences of false negatives should be minimized if only trained laboratory technicians are performing the RDTs and they follow the defined screening algorithms and protocols. The algorithms used took contact history and date of symptom onset into consideration for the interpretation of the RDT result and next steps. Regular quality control testing with OraQuick^®’s positive and negative controls allows for a check on the performance of the RDTs as well as the performance of the technicians. And continued monitoring of concordance between the EVD RDTs and RT-PCR testing allows quick identification, investigation, and response to any issues. ## Concordance data We cannot draw conclusions about the OraQuick^® EVD RDT’s performance in comparison with PCR given the low prevalence of EVD during the program period. Furthermore, the use of RDT results as eligibility criteria for PCR testing among live patients prevents us from drawing conclusions about the performance of the RDT in live patients. To determine RDT performance, all RDT results should be independently compared to those of quantitative reverse transcription PCR testing, the gold-standard diagnostic assay for detecting and quantifying Ebola virus. ## Overall conclusions The EVD RDT laboratory program is both acceptable and feasible in Guinea, but room for improvement remains, especially in quality assurance. The low percentage of false reactive OraQuick^® EVD RDTs among the deceased in Guinea is promising, but more data are needed on RDT performance. The cost of the RDTs and official validation of the test were not considered as part of this evaluation but might influence the long term feasibility of EVD-RDT use. The lessons learned during the evolution of this program may benefit others who plan to implement rapid diagnostic testing during public health emergencies. The implementation of new diagnostics in weak laboratory systems requires general training in quality assurance, biosafety, and communication with patients in addition to specific training for the new test. Corresponding capacity building in terms of basic equipment and a long-term commitment to transfer supervision and quality improvement to national public health staff are necessary for successful implementation. The future impact of EVD RDTs in Guinea rests not only on strengthening these capacities but also more generally on the strengthening of communication between the laboratory and surveillance systems and the long term sustainability of the program within the Ministry of Health. # Supporting information Thank you to the healthcare workers and laboratory technicians who participated in the pilot program and our partners at the World Health Organization (especially Koumpingnin Yacouba Nebie), the International Federation of Red Cross and Red Crescent Societies, and the Guinean Red Cross. Data collection was supported by Jean-Paul Moke Six, Ramatoulie Jallow, Patrick Mavungu, Giulia Earle-Richardson, and François Mbuyu Kata. [^1]: The authors have declared that no competing interests exist.

# Introduction Species identification, characterization, and naming remain fundamental and critical steps in biological science. Since the establishment of the Linnean system, species have been described mainly on the basis of morphological and phenotypic characteristics. Through time, however, morphology alone has been shown to be limited in its ability to delineate species boundaries, and led to a proliferation of names and nomenclatural instability. In addition, cryptic diversity (reviewed in) remained hidden from traditional morphological approaches. Modern technological developments, including DNA sequencing, provide new tools allowing the detection of hidden diversity. In particular, the DNA barcoding approach, quickly appeared to be an efficient methodology for detecting cryptic biodiversity, even though in certain cases, such as recent divergence or mitochondrial introgression, barcoding may fail to discriminate between species. Morphological data (used in systematics) and molecular data such as DNA barcodes (used in biodiversity studies) are not mutually exlusive, and often are complementary means of delineating species. Indeed, combining multiple data sources is the most efficient way to support robust species hypotheses. Formalized under the designation “*integrative taxonomy*” (reviewed in), this approach tries to use the complementarity of the different fields of study (e.g. morphology, genetics, biogeography, ecology, ethology, etc.) to delineate, describe and name species. Various protocols have been proposed to integrate these different data sources, but most of them correspond to guidelines that explore the different datasets successively to corroborate taxonomic hypotheses, or only focus on a given type of data. The way that results of the different analyses are interpreted, i.e. in a cumulative or a congruent way, also has an impact on the results, leading to an over- estimation of the number of species and lower confidence in species identity in the former case, and to an underestimation of the number of species and higher confidendence in species identity in the latter. Moreover, comparing results of different analyses, which can be based on qualitatively different data (e.g. linear measurements for morphometric analyses, sequence alignments for phylogenetic trees or distances matrices, GPS coordinates for distributional data, etc.), in the same descriptive framework remains a challenge. Community ecologists, confronted by the same issue of combined analysis of various data types, developed multi-table methods. Based on the co-inertia criterion, multi-table analyses look for common structures present in different data sets, and include them in a common analysis. The link between all tables is defined by row, since all different observations (e.g. the abundances, the distributions, the life traits, etc.) rely on the same statistical units (e.g. the specimens, the stations, etc.). These analyses are particularly flexible and allow the inclusion of multiple data types, and have already been used in different fields including e.g. ecology, medical research, agronomy, evolutionary biology, and genomics. In addition, these methods allow evaluation of the statistical significance of the congruence between data types and the amount of common information present in the different tables. We consider the integrative approach of multi-table methods highly appropriate for the resolution of species delineation in a group of poorly-differentiated catfishes from the Guianese Region. The northeastern part of the Guiana Shield, including Suriname and French Guiana, overlaps the Eastern Guianas Neotropical freshwater ecoregion. This region ranges between the Demerara River in the west and the Oyapock River in the east and probably supports more than 500 described species, of which 169 are considered endemic (freshwater ecoregions of the world: <http://www.feow.org/ecoregions/details/311>, accessed 31th Jan. 2017), making the Eastern Guianas a region of high biodiversity importance. The Eastern Guianas’ river system comprises about a dozen important catchments, including, from west to east, the Corantijn, Nickerie, Coppename, Saramacca, Suriname, Maroni (= Marowijne in Suriname), Mana, Sinnamary, Comté-Orapu, Approuague, and Oyapock rivers. All these catchments are independant and flow from south to north into the Atlantic, making the Eastern Guianas rather isolated from the rest of the Guiana Shield, out of the direct influence of the Amazon and Orinoco basins. Le Bail et al. listed 416 species of freshwater and estuarine fishes in French Guiana and Mol et al. 481 freshwater fish species in Suriname. Among this tremendous diversity, the Characiformes was the most important group, representing about 40% of all species, followed by the Silurifomes at around 35%. Among the latter, the Loricariidae is the most diversified catfish family with more than 80 species distributed in French Guiana and Suriname. The Loricariidae is a strictly Neotropical catfish family comprising 937 valid species and an estimated 300 undescribed species distributed in more than 100 genera, making it the most species rich family of the Siluriformes. Loricariids are primarily characterized by a depressed body covered by bony plates, and by an important modification of the mouth into a sucker disk. Among Loricariidae, the subfamily Hypostominae represents half of the familial diversity, comprising 465 valid species distributed in more than 40 valid genera. In French Guiana and Suriname nine genera are recorded, including hyperendemic and monotypic representatives such as *Hemiancistrus medians* or *Pseudoqolus koko*, both restricted to the Maroni Basin. The other genera are more widely distributed in South America, with the exception of *Guyanancistrus*, restricted to the northeastern part of the Guiana Shield. Isbrücker described the genus *Guyanancistrus*, designating *Lasiancistrus brevispinis* Heitmans, Nijssen & Isbrücker 1983, a species present in Suriname and French Guiana, as the type species. *Guyanancistrus* was originally diagnosed on the basis of its similarity to *Lasiancistrus* Regan 1904 while differing from the latter in the absence of the characteristic bristles, or whisker-like odontodes, that are found among the hypertrophied odontodes on their evertible cheek plates. *Guyanancistrus* was placed in the synonymy of *Pseudancistrus* Bleecker 1962 by Armbruster, based on a phylogenetic analysis of morphological characters that included most of the genera then placed in the subfamilies Hypostominae and Ancistrinae. However, a molecular phylogenetic analysis of the group using mitochondrial and nuclear sequence data revealed *Pseudancistrus sensu lato* to be a paraphyletic assemblage of five unrelated lineages. One of the lineages uncovered corresponded to the genus *Guyanancistrus*. As well as *G*. *brevispinis*, two other species were included in the genus: *G*. *niger* (Norman 1926) and *G*. *longispinis* (Heitmans, Nijssen & Isbrücker 1983), both described from French Guiana and restricted to the Oyapock River Basin. Additionally, a possibly new dwarf species collected in mountain streams flowing to the Marowijne River in Suriname was placed as a member of *Guyanancistrus*, and constituted the sister species of *G*. *brevispinis*. This small species (\<6 cm) nicknamed Bigmouth due to its particular morphology was already suspected to be new by Mol who collected it during a Rapid Assessment Program (RAP) survey to the Nassau Mountains. This revealed a highly endemic fauna now threatened with extinction by a bauxite mining project and illegal gold mining. Unlike its congeners, *Guyanancistrus brevispinis* is known to be widespread, common and abundant, its area of distribution covering all the main Guianese river systems of Suriname and French Guiana, from the Corantijn in the west to the Oyapock in the east. Cardoso and Montoya-Burgos analysed the species based on several of its populations, including Amazonian ones (from northern tributaries of the Paru de Oeste and Jari rivers), in order to decipher its historical biogeography, and found that it was genetically highly diversified, with six distinct allopatric lineages (five Guianese and one Amazonian). It was thus considered as a species complex, with the true *G*. *brevispinis* possibly restricted to the Nickerie River system (see). However, additional sources of information from genetic markers were deemed necessary to confirm their taxonomic status. The five Atlantic coastal *G*. *brevispinis* lineages of the Guianas were found to form a monophyletic group that originated from an ancestral colonization event from the Amazonian Basin, hypothesized to have been through river capture between northern Amazon tributaries and the upper Maroni River Basin. In the Guianas, subsequent dispersal would mainly have resulted from temporary connections between adjacent rivers when sea levels were low, and subsequent diversification of isolated populations during periods with high sea levels. Considering the recent genus revalidation and questions about the type species, and the potential existence of new and/or endangered species, the present work uses an integrative approach combining morphology, genetics and spatial data to reappraise *Guyanancistrus*, focusing on the enigmatic *Guyanancistrus brevispinis* species complex. Most known populations were included in this analysis, principally based on material collected by the authors and their collaborators in the past 15 years. After a comparative diagnosis of the genus, the type species is redefined and its morphological and genetic variation delineated. Several new species revealed by the study are also described. Detailed descriptions and morphological comparisons of *Guyanancistrus niger* and *G*. *longispinis* are already available and will not be repeated, but a practical key to all *Guyanancistrus* species is provided. After this taxonomic clarification, the biogeography of all *Guyanancistrus* members is re-evaluated to investigate dispersal processes, putative local extinctions, and speciation events. # Materials and methods ## Ethics statement No protected species (local restrictions, IUCN or CITES listed species) were examined in the study. Most specimens and tissue samples were obtained from museum collections and/or by local populations or fishermen. No experimentation was conducted on live specimens. For specimens and associated tissue samples obtained from the field, specimens were collected and exported with appropriate permits: Préfecture de la Région Guyane, Arrété 03/17/PN/EN to collect in the Réserve Naturelle des Nouragues in 2003; Ministry of Agriculture, Animal Husbandry and Fisheries to export fishes from Suriname in 2005, 2007, 2008, 2012, and 2014. Material obtained from the Parc Amazonien de Guyane in 2014 was collected under the direct supervision of PAG authorities. When collecting occurred in non protected areas of French Guiana, sampled specimens were declared to the French DEAL (French environmental protection ministry) before export. Immediately after collection, fish were anesthetized and sacrificed using water containing a lethal dose of eugenol (clove oil). Fin clips were taken after death and specimens fixed for long term preservation in museum collections. All work was conducted in accordance with relevant national and international guidelines, and conforms to the legal requirements (Directive 2010/63/EU of the European Parliament and of the Council on the protection of animals used for scientific purposes, the Swiss ordinance OPAn 455.1 of OSAV, and recommendations and regulations of DETA-DGNP permit number 20160422/01 AS). ## Materials Materials examined for this study are deposited in the following institutions and collections: Auburn University Museum of Natural History (AUM); The Natural History Museum, London (BMNH); Museum of Comparative Zoology, Cambridge (MCZ); Muséum d’histoire naturelle, Genève (MHNG); Muséum national d’histoire naturelle, Paris (MNHN); Museu de Zoologia da Universidade de Sao Paulo (MZUSP); National Zoological Collection of Suriname (NZCS); National Museum of Natural History-Naturalis, Leiden (RMNH), presently holding the former Zoological Museum Amsterdam (ZMA) collection; Museum für Naturkunde, Berlin (ZMB); Zoologisches Staatssammlung, Munich (ZSM). Specimens included in morphometric analyses are indicated by an asterisk in specific lists of materials in main text and supplementary file, followed by number when needed. ## Morphology Measurements and counts were obtained from a total of 269 specimens, and were only carried out on one side in cases of paired characters. Specimens were measured with a digital calliper to the nearest 0.01 mm following Fisch-Muller et al.. Measurements are presented in tabular form as percentages of standard length (SL) except for subunits of the head, which are expressed as percentages of head length (HL). Counts are the following: premaxillary and mandibular teeth were counted for the emergent row, adding obviously missing teeth shown by gaps in tooth rows. Dermal plate counts included: 1) lateral plates in the median series of rows, according to Schaefer, 2) plates bordering the supraoccipital, 3) predorsal plates, counted dorsally along a median line between supraoccipital and nuchal plate, 4) lateral plates of dorsal series along the dorsal-fin base, 5) lateral plates in dorsal series between end of dorsal-fin base and adipose- fin insertion, 6) lateral plates in dorsal series between adipose-fin insertion and caudal fin, 7) lateral plates in ventral series between the anal and the caudal fins, 8) lateral plates in dorsal series between end of dorsal fin when adpressed and adipose-fin spine insertion, and 9) lateral plates in dorsal series along unpaired median plate(s) preceeding adipose fin. All plate counts are whole numbers except (8) and (9) that were counted to the nearest half- plate. Dorsal-fin and anal-fin branched rays were counted; other fin-ray counts do not vary among Hypostominae species and were only obtained for type specimens and part of the non-type material. ## Morphometry Morphometric and meristic data were subjected to multivariate analyses to reveal the morphological structure of the different species and populations of *Guyanancistrus* under study, with the addition of the possibly congeneric *Pseudancistrus megacephalus* to resolve taxonomic uncertainties. Prior to the analyses, all specimens smaller than 20 mm were excluded to minimize the bias introduced by allometric growth. Missing data, due to broken fin rays, were estimated for specimens belonging to a given population using the least squares method with the standard length (SL) used as the explanatory variable. Then, all morphometric data were standardized by SL and log transformed to control for size effect, to preserve and linearize allometric growth, and to prevent spurious correlations in the use of simple ratios. Meristic data were used raw. The final table included data from 269 specimens belonging to 27 different morphs and populations, and contained 38 variables (24 morphometric and 14 meristic). This table was centered and reduced to allow comparison of variables expressed in different units, and submitted to a principal component analysis (PCA) using the correlation matrix to reveal its structuring. Because *Guyanancistrus* members appeared morphologically very close, and given the large number of variables in relation to the number of groups, a between group analysis (BGA) was secondarily performed on PCA results. To prevent artificial groupings, the different populations and morphs collected in different places for a given species were considered independently and used as a grouping factor (n = 27 groups). Prior to the BGA, a Monte Carlo permutation test on the value of between-group inertia was conducted using 9,999 random permutations to test against the absence of group effect. Multivariate analyses were performed using the ade4 1.7–4 and ade4TkGUI 0.2–9 packages in R 3.3.2. ## Genetics To estimate the genetic diversity and species boundaries of *Guyanancistrus* members, the 5’ region of the cytochrome *c* oxidase I (COI) mitochondrial gene was amplified for a DNA barcode analysis. In addition, a molecular phylogeny was reconstructed for 77 putative *Guyanancistrus* members and 12 outgroup species based on the analysis of mitochondrial and nuclear gene fragments. Outgroup representatives were chosen from other genera, clades, and subfamilies of the Loricariidae following results of Covain & Fisch-Muller and Lujan et al.. The samples analyzed came from the tissue collection of MHNG. Markers were selected for their ability to resolve between- and within-species relationships, as well as deeper relationships at the intra-familial rank. For this we selected fast evolving markers such as the mtCOI, and the intronic regions of the nuclear *Fish Reticulon-4 receptor* (*f-rtn4r*) gene, whereas more conserved exonic regions of *f-rtn4r* and *recombination activating gene 1* (*rag1*) provided information for deeper relationships. Total genomic DNA was extracted with the DNeasy Tissue Kit (Qiagen) following the instructions of the manufacturer. The PCR amplifications were carried out using the Taq PCR Core Kit (Qiagen). To amplify and sequence in a single run the standard 650 bp barcode region with high quality, fragment size was increased to 900 bp using two newly designed primers: 5COI-F (`5’-CTC GGC CAT CCT ACC TGT G-3’`) and 5COI-R2 (`5’-CGG GTG TCT ACG TCC ATT CCA ACT G-3’`). The amplifications were performed in a total volume of 50 μl, containing 5 μl of 10x reaction buffer, 1 μl of dNTP mix at 10mM each, 1 μl of each primer at 10 μM, 0.2 μl of *Taq* DNA Polymerase equivalent to 1 unit of Polymerase per tube, and 1 μl of DNA. Cycles of amplification were programmed with the following profile: (1) 3 min. at 94°C (initial denaturing), (2) 35 sec. at 94°C, (3) 30 sec. at 53°C, (4) 55 sec. at 72°C, and (5) 5 min. at 72°C (final elongation). Steps 2 to 4 were repeated 39 times. Amplifications of the nuclear *rag1* and *f-rtn4r* genes followed Sullivan et al. and Covain et al. respectively. PCR products were purified with the High Pure PCR Product Purification Kit (Roche). Sequencing reactions were performed with the Big Dye Terminator Cycle Sequencing Ready Reaction 3.1 Kit (Applied Biosystems) following instructions of the manufacturer, and were loaded on an automatic sequencer 3100-Avant Genetic Analyzer (Applied Biosystems, Perkin-Elmer). Newly generated sequences were deposited in GenBank and BOLD with accession numbers provided in, while complementary sequences from previously published studies were obtained from GenBank (accession numbers and corresponding references). The DNA sequences were edited and assembled using BioEdit 7.0.1, aligned using ClustalW, and final alignment was optimized by eye. For the barcode analysis, the aligned COI sequences were converted into a distance matrix to evaluate sequence divergences using the Kimura 2 Parameter (K2P) metrics with pairwise deletion for missing data as implemented in spider 1.3.0 in R. This K2P matrix was used to compute between- and within-species distances to allow threshold optimization and evaluate the existence of a barcoding gap for correct species identification. A levelplot graph allowing a graphical representation of the distance matrix was also computed using the lattice 0.20–34 and colorspace 1.2.7 packages in R. For the phylogenetic reconstruction, four partitions were created corresponding to the COI, *rag1*, exonic regions of *f-rtn4r*, and intronic regions of *f-rtn4r* genes. Two phylogenetic reconstruction methods allowing the analysis of partitioned data were used. First, a maximum likelihood (ML) reconstruction was performed with RAxML 7.2.6 and raxmlGUI 1.0 using the GTRGAMMA model with each partition assigned its own parameters. Robustness of the results was estimated by rapid bootstrapping with 1,000 pseudoreplicates. Second, a Bayesian inference analysis was conducted in MrBayes 3.2.6. Two runs of four chains (one cold, three heated) were conducted simultaneously for 2 x 10⁷ generations using the same model as the ML analysis (nst = 6, rates = gamma, and each partition assigned its own parameters), with the tree space sampled each 1000th generation. After a visual representation of the evolution of the likelihood scores, and checking for the stationarity of all model parameters using Tracer 1.5 (i.e. potential scale reduction factor (PSRF), uncorrected roughly approached 1 as runs converged, and Effective Sample Size (ESS) of all parameters above 200), the 2 x 10⁶ first generations were discarded as burn-in. The remaining trees were used to compute the consensus tree. Bayesian inference was performed using the CIPRES Science Gateway 3.3. ## Distribution To explore the distributional patterns of the different species of *Guyanancistrus*, georeferenced data were recorded for the localities of specimens deposited in official collections. These collections were selected for housing a large sampling of Guianese species, i.e. MHNG, MNHN, RMNH, and ZMA (the two latter now grouped in the Naturalis Biodiversity Center, Leiden). In addition, relative abundances were computed according to the number of specimens in each batch for each species and locality. Relative abundances per species and per locality were then ploted as pie charts onto a geographic map. Because of the heterogeneity of samples (i.e. comprising a collection of small and large numbers of specimens), a constant was added for each abundance estimate for readability (i.e. occurrence + abundance). The map was reconstructed using raster images and shapefiles obtained from the HydroSHEDS project website (<http://www.worldwildlife.org/pages/hydrosheds>) in conjunction with the shapefiles 0.7, mapplots 1.5, raster 2.5.8, and colorRamps 2.3 packages in R. Pie charts were computed with the mapplots package. ## Multi-table analysis Because *Guyanancistrus brevispinis* has been claimed to contain a species complex comprising at least five species, and that morphological characteristics were ambiguous, an integrative approach appeared necessary to clarify species boundaries. Because preliminary analyses provided information about the morphometric, phylogenetic, and spatial structures of *G*. *brevispinis*, the three types of information were united in a multiple co-inertia analysis (MCOA) to identify the possible common structures of all datasets. For this, the three datasets were restricted to the subset of individuals and populations (n = 51 individuals distributed in 15 populations) for which all three types of information were available. Each of the three reduced tables was reanalysed separately. The morphometric data table was reanalysed by a PCA but with within- population variability eliminated by the computation of average values of the morphometric variables for each population. For the phylogenetic data table, a patristic distances matrix was computed from the branch lengths of the phylogenetic tree using ape 3.5 package in R. Then, a principal coordinate analysis (PCoA) using Cailliez correction for non-Euclidian distance matrices was performed, to reveal its structuring. This analysis provided a tree-free representation of the distance matrix, where the pairwise distances between individuals on the axes are equal to the genetic pairwise distances of the matrix. The spatial structure was revealed by a PCoA performed on pairwise geographic distances computed from GPS coordinates using the geosphere 1.5.5 package in R. A first assessment of a possible link between the three tables was obtained by performing pairwise Monte-Carlo permutation tests on the value of the RV coefficient using 9,999 random permutations. Preliminary analyses, permutation tests, and multi-tables analyses were performed using the ade4 package in R. ## Biogeography Finally, after clarification of all systematic issues, a reappraisal of the phylogeography of *Guyanancistrus* members was performed to elucidate the different routes of dispersal, and extinction and speciation events which occurred in the Eastern Guianas (*sensu*). For this a Dispersal Extinction Cladogenesis (DEC) analysis was performed using the BioGeoBEARS 0.2.1 package in R. The DEC model possesses two free parameters (*d* = dispersal and *e* = extinction) and allows maximum likelihood estimates of ancestral areas along branches and nodes of a phylogenetic tree. The DEC model implemented in BioGeoBEARS is equivalent to the one implemented in LAGRANGE but with the possibility of adding an additional parameter *j* for founder effect events as an additional explanation for cladogenesis. For ancestral area reconstructions, the phylogenetic tree was reduced to the ingroup, polytomies resolved using ape in R, and 11 areas were defined that corresponded to the present catchment areas of the coastal rivers of the Guianas where *Guyanancistrus* members were collected, i.e. from west to east: the Corantijn, Nickerie, Saramacca, Suriname, Maroni, Mana, Sinnamary, Comté-Orapu, Approuague, and Oyapock rivers, with the addition of the Amazon Basin (including Paru de Oeste and Jari rivers) for species distributed outside of the Guianas. For the calculation, the maximum number of ancestral areas allowed to be reconstructed to a given node was set to four. The best model for the maximum likelihood reconstruction was evaluated by likelihood ratio test (LRT) since DEC and DEC + *j* were nested models differing in a single parameter (*j*). ## Nomenclatural acts The electronic edition of this article conforms to the requirements of the amended International Code of Zoological Nomenclature, and hence the new names contained herein are available under that Code from the electronic edition of this article. This published work and the nomenclatural acts it contains have been registered in ZooBank, the online registration system for the ICZN. The ZooBank LSIDs (Life Science Identifiers) can be resolved and the associated information viewed through any standard web browser by appending the LSID to the prefix “<http://zoobank.org/>”. The LSID for this publication is: urn:lsid:zoobank.org:pub:72F36C36-69D7-4E59-AF6A-8A88C78CFD86. The electronic edition of this work was published in a journal with an ISSN, and has been archived and is available from the following digital repositories: PubMed Central, LOCKSS. # Results ## Morphometry The between-group inertia recorded by the BGA represented 48.13% of the total inertia of the preliminary PCA (sum of eigenvalues of the BGA / sum of eigenvalues of the PCA: 19.73155 / 41 = 0.4812573). The permutation test was highly significant, with none of the null hypothesis sampling distribution of randomized values greater than the observed value of between-group inertia (X_obs = 0.4812573; p_Xrand ≥ p_Xobs = 0.0001). A significant group effect was thus present in the data, and observed between- group differences were not due to chance. Morphometric data were mainly structured on the first two axes of BGA which explained 57.52% of the total between-class inertia (30.78% for axis 1 and 26.74% for axis 2). The first morphometric plane (axes 1 and 2) of individuals split the different populations and species of *Guyanancistrus* into four main groups. On the negative side of the first axis, the first group corresponded to a population originating from the northern slope of the Mitaraka Mountains (Gsp3; Jari River slope), followed by a mix of mostly *G*. *brevispinis* populations, thereafter called *brevispinis* group. In decreasing order of negative scores of variables, species and population were rather characterized by higher values for number of branched anal-fin rays, interdorsal distance, number of plates bordering the supraoccipital, number of predorsal plates, caudal peduncle length, and upper caudal spine length. On the positive side, different species were aligned along the axis including *G*. *niger*, *G*. *longispinis*, the holotype of *Pseudancistrus megacephalus*, a population of *Guyanancistrus* from the Potaro River in Guyana also identified as *P*. *megacephalus* by Eigenmann in 1908 (Gpot), and a population from the Nassau Mountains (GBM). These species and populations were rather characterized by higher values of cleithral width, supracleithral width, caudal peduncle depth, and head depth at supraoccipital. Along the second axis, on the positive side, *G*. *niger*, *G*. *longispinis*, *P*. *megacephalus*, and the population from Potaro River were split from all other populations and species and constituted the *longispinis* group. Higher values of pectoral spine length, dorsal spine length, and dorsal-fin base length split the *longispinis* group from other *Guyanancistrus* members. On the negative side of Axis 2, the population from the Nassau Mountains was separated by higher values of premaxillary tooth cup length, interbranchial distance, and dentary tooth cup length. The *brevispinis* group remained poorly characterized morphologically, with all of its members grouped at the center of principal axes. The morphometric approach was thus insufficient to delimit species boundaries of the *brevispinis* group members, and only six species could be characterized: *G*. *niger*, *G*. *longispinis*, *P*. *megacephalus*, a putatively three new species Gpot from Potaro River (= *P*. *megacephalus* sensu Eigenmann, 1908), Gsp3 from Mitaraka Mountains, and GBM from Nassau Mountains. ## DNA barcodes The sequence alignment of 77 COI barcodes reached a total length of 889 positions. No insertions, deletions, or stop codons were observed in any sequence. Five lineages and three levels of variations were highlighted by the K2P distances heatmap. The first lineage comprised all populations of *G*. *brevispinis* and the one from the Nassau Mountains. Within-group variations ranged between 0 and 0.023 (mean = 0.011) whereas between-group ones ranged between 0.174 and 0.003 (mean = 0.08). The second group was constituted by populations from the Mitaraka Mountains (Jari and Maroni sides), a population from Paru de Oeste River, and one from the Brownsberg Mountains. In this group, within-group distances ranged between 0 and 0.027 (mean = 0.012), and between- group distances between 0.055 and 0.156 (mean = 0.075). The third group was constituted by *G*. *niger* members (within-group K2P distances 0 \< d \< 0.003, mean = 0.002; between-group K2P distances 0.120 \< d \< 0.193, mean = 0.136), and the fourth one by *G*. *longispinis* members (within-group K2P distances 0 \< d \< 0.0016, mean = 0.0008; between-group K2P distances 0.120 \< d \< 0.193, mean = 0.136). The last group comprised a single representative from the Paru de Oeste River. This specimen displayed high between-group variations ranging between 0.13 and 0.17 (mean = 0.14) K2P distances. One level of variation was revealed in global within-group distances, whereas three levels were present in between-group distances. Both global between- and within-group variations showed strong overlap with global within-group distances ranging between 0 and 0.023 (mean = 0.011), and global between-group distances ranging between 0.004 and 0.193 (mean = 0.088). The barcoding gap analysis revealed the absence of positive differences between the furthest intra-group distances and the closest non-specific for 51 individuals. These individuals consisted mainly of representatives of *G*. *brevispinis*, the individuals from the Nassau Mountains, those from both sides of the Mitaraka Mountains, two specimens from Paru de Oeste and one from the Brownsberg Mountains, indicating that the barcoding approach was insufficient to discriminate all of the species of *Guyanancistrus* without ambiguity. The threshold optimization delivered a minimum of false positive matches and minimum cumulative error with a K2P distance of 0.004, much below the usually accepted threshold (around 1–2%). The barcoding approach was thus unable to distinguish between intra and inter specific variations for several species, including members of the putative *G*. *brevispinis* complex, and only seven mitochondrial lineages could be identified (1: *G*. *brevispinis* including GBM from Nassau Mountains; 2: a mix of Gsp3 and Gsp4 from Mitaraka Mountains; 3: the weakly diferenciated Gsp1 from Paru de Oeste and 4: Gkum from Brownsberg Mountains; 5: *G*. *niger*, 6: *G*. *longispinis*, and 7: Gsp2 from Paru de Oeste). ## Molecular phylogeny Given the poor results of the barcoding approach, nuclear marker sequences were added to the data set, and a molecular phylogeny reconstructed. In addition to the 77 mt COI gene fragments of *Guyanancistrus* members, 2 COI sequences of outgroup species, 56 sequences of the partial nuclear gene *f-rtn4r*, and 69 sequences of *rag1* were sequenced. Forty three complementary sequences (6 of COI, 29 of *f-rtn4r*, and 8 of *rag1*) were obtained from GenBank using the accession numbers provided in Cardoso and Montoya-Burgos, Collins et al., Covain and Fisch-Muller, Covain et al., Covain et al., Fisch-Muller et al., and Lujan et al.. Twenty one gene fragments did not amplify (4 of COI, 4 of *f-rtn4r*, and 13 of *rag1*) and were treated as missing data. The final sequence alignment included 3789 positions of which 889 corresponded to the mt COI gene, 1027 to the intronic and 864 to the exonic regions of *f-rtn4r*, and 1009 to the *rag1* gene. Maximum Likelihood and Bayesian phylogenetic reconstructions lead to identical tree topologies. The ML tree (Ln*L* = -15182.6) and Bayesian tree, both placed the specimen from the Paru de Oeste River, already distinct from all other *Guyanancistrus* in the barcoding approach, as a representative of a distinct genus, member of the outgroup, and sister genus of *Corymbophanes* with high statistical support \[99 Bootstrap Probability (BP) and 1 Posterior Probability (PP)\]. Both genera were nested in a clade comprising *Hopliancistrus tricornis* as sister group, and all three genera constituted the sister group of all *Guyanancistrus* members with high statistical support (76 BP, 1 PP). The *Guyanancistrus* clade was highly supported (99 BP, 1 PP) and split into two groups: one comprising *G*. *niger* with *G*. *longispinis* (100 BP, 1 PP), and a second comprising all other *Guyanancistrus* (100 BP, 1 PP). Within the latter, two new groups emerged: one comprising the population of the Nassau Mountains as the sister group of all *Guyanancistrus brevispinis* members (100 BP, 1PP) thus constituting a distinct species, and a second comprising all remaining species of *Guyanancistrus* (100 BP, 1 PP). Within the latter, two groups were highlighted, one comprising the population of the Jari side of the Mitaraka Mountains along with the one of the Maroni side (98 BP, 1 PP), and a second consisted in a population of the Paru de Oeste side of the Four Brothers Mountains, along with a population from the Brownsberg Mountains (99 BP, 1 PP). These four lineages constituted distinct species of *Guyanancistrus*. Within *G*. *brevispinis*, a first lineage originating from the Suriname River split from all other populations (50 BP, 0.97 PP). Then, with the exception of the polytomized population from the Tapanahony River (Marowijne Basin), three groups emerged: one comprising all populations from Corantijn, Nickerie, and Saramacca basins in Suriname (56 BP, 0.89 PP) sister to two sister groups, one constituted of all populations from Maroni (French Guiana), Mana, and Sinnamary rivers (55 BP, 0.90 PP), and the second comprising all populations of Comté-Orapu, Approuague, and Oyapock rivers in French Guiana (99 BP, 1 PP). The phylogenetic analysis revealed the presence of seven species of *Guyanancistrus* and of a new genus and species in the data. No species complex was present within *G*. *brevispinis* but three lineages of infraspecific rank emerged. ## Distribution Different distribution patterns were present in the data. (1) Species could be widespread. This pattern characterized *G*. *brevispinis* which dominated the other species in terms of both occurrences and abundances. The species was distributed in all important drainages including the Corantijn, Nickerie, Saramacca, Suriname, Maroni, Mana, Comté-Orapu, Approuague, and Oyapock rivers, and represented 86.5% of all specimens collected. When *G*. *brevispinis* was co-distributed with other *Guyanancistrus* members, such as in the Oyapock Basin, it appeared less frequent and abundant. (2) Species could be restricted to a region including a few basins or a single basin. This pattern was observed for *G*. *niger* and *G*. *longispinis*, both endemic to the Oyapock River, but distributed throughout the Oyapock drainage. (3) Species could be hyperendemic and restricted to a single place. This pattern concerned species restricted to mountainous areas of the Nassau, Brownsberg, Four Brothers, and Mitaraka mountains. Three basins had different distribution patterns for different species; patterns 1 and 2 were present in the Oyapock, whereas patterns 1 and 3 were present in the Maroni and Saramacca rivers. ## Multi-tables analysis Because only phylogenetic information was able to discriminate *G*. *brevispinis* among all other *Guyanancistrus* members and none of the five lineages claimed as putative new species could be clearly delineated by the different analyses, the multi-table approach was applied. Prior to the analysis, the DNA barcode table was removed to minimize redundancy and avoid over weighting this information since the COI gene had been used to reconstruct the phylogeny. A first assessment of the relationships between genetics (i.e. the phylogeny), morphology, and geography was performed using pairwise RV tests between preliminary analyses of the reduced datasets (PCoA for genetics and geography, and PCA for morphometric data). All pairwise tests showed strong and significant vector correlations between tables (p-value = 0.0001), with the pairwise correlations indicating that the genetic data were slightly more correlated to the geography (RV = 0.566, p-value = 0.0001) and morphological data (RV = 0.532, p-value = 0.0001) than the latter were to the geography (RV = 0.491, p-value = 0.0001). However, all RV coefficients were globally comparable among pairwise comparisons with around 50% of common signal between tables. The first plane of MCOA accounted for 69.31% of the total co-structure (52.98% for axis 1 and 16.33% for axis 2). The amount of variation explained by MCOA axes was similar to those obtained in the separate analyses, but with a lesser contribution from morphology. Indeed, 99.86% ((0.396 + 0.338)/(0.457 + 0.278) = 0.734/0.735) of the genetic data structure, 69.2% of the morphological data structure, and 100% of the geographic data structure were represented on the first two axes of the MCOA (Co-Inertia/Inertia). The contribution of each table to the quantity maximized by MCOA (i.e. sum of squared covariance between the linear combinations of the variables of each table and the compromise = Cov²) highlighted the relative importance of geography for the first axis, and of genetics for the second. Morphology contributed least to the compromise for both axes. The associated correlations (Cos²) showed that the first two axes of the compromise were strongly linked to each separated table except for the second axis derived from geographic data (0.890 and 0.962 for the genetic data, 0.865 and 0.758 for the morphometric data, and 0.965 and 0.157 for the geographic data). The first axis of the individuals’ plane of the MCOA aligned three groups of *G*. *brevispinis*. In negative scores, a first group comprised populations distributed in the Oyapock, Approuague and Comté- Orapu rivers in eastern French Guiana followed by a second group comprising populations in the Sinnamary, Mana and Maroni basins in central and western French Guiana but excluding those from the Tapanahony River, a western tributary of Maroni River in Suriname. In positive scores, a third group comprised populations from the Corantijn, Nickerie, Saramacca, Suriname, and Tapanahony rivers in Suriname. The second axis split the representatives of *G*. *brevispinis* from central and western French Guiana in negative scores from those from Suriname and eastern French Guiana. Correlations of variables of the preliminary analyses with MCOA axes showed high scores, in decreasing order of positive scores on axis 1, for the first principal coordinate of the PCoA of the geographic data table (i.e. the longitude), first and second principal coordinates of the PCoA of the phylogenetic data table (i.e. deeper structures of the phylogenetic tree restricted to *G*. *brevispinis* corresponding to the splitting of the different populations from Suriname and French Guiana, see), opercle length, interorbital width, cleithral width, supracleithral width and interbranchial distance. For negative scores (in decreasing values of absolute values of negative scores) these variables corresponded to the: number of lateral plates, number of predorsal plates, dorsal–fin base length, number of plates between adpressed dorsal fin and adipose fin, caudal peduncle length and interdorsal distance. On the second axis the variables with greater scores corresponded, in decreasing order of positive scores, to the second principal coordinate of the PCoA of the phylogenetic table, second principal coordinate of the PCoA of the geographic table (i.e. the latitude), and caudal peduncle depth. In negative values, these variables corresponded, in decreasing order of absolute values, to the: first principal coordinate of the PCoA of the phylogenetic table, number of plates between dorsal-fin base and adipose-fin spine, number of plates along adipose-fin base (median platelets), snout length, and number of anal to caudal plates. Three groups of infra-specific rank showing significantly structured data concerning their distribution, genetics, and morphology were consequently recognized. ## Taxonomic account and descriptions ### *Guyanancistrus* Isbrücker in Isbrücker et al., 2001 *Guyanancistrus* Isbrücker in Isbrücker et al., 2001: 19 (type species: *Lasiancistrus brevispinis* Heitmans, Nijssen & Isbrücker, 1983; type by original designation; masculine); Fisch-Muller, 2003: 384 (checklist, valid); Armbruster, 2004 (synonymization with *Pseudancistrus* based on morphological phylogeny); Ferraris, 2007: 287 (checklist, synonym of *Pseudancistrus*); Covain & Fisch-Muller, 2012: 235 (generic revalidation based on molecular phylogeny); Silva et al., 2014: 12 (validity confirmed based on same data); Lujan et al., 2015: 276 (phylogenetic placement in redefined tribe Ancistrini). **Diagnosis.** *Guyanancistrus* was shown to be monophyletic based on mitochondrial and nuclear DNA sequences. No unique morphological character was found to diagnose the genus which belongs to the Ancistrini tribe of the Hypostominae subfamily. The following combination of characters distinguishes *Guyanancistrus* from all other Hypostominae genera: head and body dorsoventrally depressed; head and body plates not forming prominent ridge or crest; snout rounded and flattened; snout covered with plates except tip region and, sometimes, a small area on each side of tip of snout; plates on all parts of snout forming a rigid armor covered with numerous short odontodes; presence of odontodes over a broad area on the opercle; presence of enlarged cheek odontodes supported by evertible plates; these odontodes straight with tips slightly curved, not strongly hook-shaped; absence of whisker-like cheek odontodes; absence of enlarged odontodes along snout margin; presence of a dorsal iris operculum; lips forming an oval disk; dentary and premaxillary with numerous viliform and bicuspid teeth; presence of a small buccal papilla, no enlarged dentary papilla; seven branched dorsal-fin rays; presence of an adipose fin; no membranous extension between end of dorsal fin and adipose fin; five series of lateral plates extending to caudal fin; lateral plates not keeled and not bearing enlarged odontodes; lateral plates of ventral series on caudal peduncle angular but not keeled; abdominal region entirely naked. *Guyanancistrus* is mostly similar to *Cryptancistrus* n. gen in external appearance. It is distinguished from *Cryptancistrus* primarily by the uniformity of its snout plates and odontodes (in *Cryptancistrus*. posterior part of lateral margin of snout do not form a rigid armor but rather a soft fleshy border, and bears slightly enlarged odontodes with small tentacules sensu Sabaj et al.). It can additionally be distinguished from *Cryptancistrus* by the presence of a skin region bordering the exposed portion of opercle that is distinctly narrower than the latter (*vs* roughly as large as the latter). **Etymology.** The name *Guyanancistrus* was originally explained as a contraction of “Guyana” and the generic name *Ancistrus* Kner, 1854. Gender: masculine. **Distribution.** Endemic to the Atlantic coastal rivers and upper Amazonian tributaries of the north-eastern Guiana Shield. ### *Guyanancistrus brevispinis* (Heitmans, Nijssen & Isbrücker, 1983) (Figs, and ;) *Lasiancistrus brevispinis* Heitmans, Nijssen & Isbrücker, 1983: 38, Figs – (type locality: Surinam, district Nickerie \[Sipaliwini\], Fallawatra River, rapid 5 km S.W. of Stondansie Fall, Nickerie River system; holotype: ZMA 107.740); Ouboter & Mol, 1993: 149 (distribution in Suriname); Boujard et al., 1997: 183; Le Bail et al., 2000: 236–237 (complementary description, distribution in French Guiana, illustration); *Guyanancistrus brevispinis* (Heitmans, Nijssen & Isbrücker, 1983): Isbrücker in Isbrücker et al., 2001: 19 (original designation as type species of *Guyanancistrus*); Fisch-Muller, 2003: 385 (checklist, valid); Mol et al. 2007: 112 (Lely Mountains); Covain & Fisch-Muller, 2012: 244 (in molecular phylogeny of *Pseudancistrus sensu lato*, generic reassignation); Mol et al., 2012: 274 (distribution in Suriname); Le Bail et al., 2012: 303 (distribution in French Guiana); Mol, 2012: 448–449 (complementary description and illustration); Lujan et al., 2015: 278 (in a molecular phylogeny of Loricariidae); Melo et al., 2016: 134 (collected in Amapá, Brazil). *Lasiancistrus niger* (not of Heitmans, Nijssen & Isbrücker, 1983): Montoya- Burgos et al., 1998: 367 (in a molecular phylogeny of Loricariidae) *Pseudancistrus brevispinis* (Heitmans, Nijssen & Isbrücker, 1983): Armbruster, 2004a (in a morphological phylogeny), 2004b (illustration), 2008; Ferraris, 2007: 287 (checklist); Cardoso & Montoya-Burgos, 2009: 947 (diversity and historical biogeography); Willink et al., 2010: 41 (comparison to *P*. *kwinti*). **Diagnosis.** *Guyanancistrus brevispinis* is discriminated from all congeners except *G*. *nassauensis* n. sp. by specific barcode sequences (see BOLD numbers) and by much shorter evertible cheek odontodes, longest ones usually only reaching the first half of the opercle, or, in some large specimens measuring more than 70 mm SL (probably adult males), surpassing the middle of the opercle, but not reaching its last quarter (*vs* reaching between last quarter up to far beyond posterior end of opercle except in very small specimens). Evertible cheek odontodes are shorter in *G*. *brevispinis* than in *G*. *nassauensis* with regard to size of specimens; in the latter, large specimens (likely adult males) that have odontodes reaching beyond the middle of the opercle measure only 40 mm. *Guyanancistrus brevispinis* is a larger species than *G*. *nassauensis* (maximum known SL: 152 mm *vs* 61 mm). It is further discriminated from *G*. *nassauensis* by smaller dentary and premaxillary tooth cusps (in % of head length, respectively: 15.8–23.6, mean 19.5, *vs* 24.2–31.9, mean 27.6; 16.3–24.5, mean 20.1, *vs* 25.4–31.4, mean 28.1) and by an anal fin with 5 branched rays (*vs* 4), apart exceptional individuals. *Guyanancistrus brevispinis* is readily distinguished from *G*. *longispinis* and *G*. *niger* by color pattern (body and fins medium brown with paler yellow to orange medium-sized spots to transverse bands, *vs* brown-black with either small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). *Guyanancistrus brevispinis* can also be distinguished from *G*. *brownsbergensis* and *G*. *tenuis* by a smaller number of plates between adpressed dorsal fin and adipose-fin spine (0.5–3, mean 2, *vs* respectively 3–3.5, mean 3, and 3–4, mean 3.5), from *G*. *brownsbergensis* by a lower peduncle depth (8.6–11.3, mean 10.0% of SL, *vs* 11.4–11.6, mean 11.5), and from *G*. *megastictus* by the lower number of plates bordering the supraoccipital (2–3, mean 3 *vs* 4) and by a color pattern with smaller roundish spots (not covering four plates) or straighter bands on posterior part of body and fins. **Description.** Morphometric and meristic data in. Relatively large-sized species for *Guyanancistrus* (up to 152 mm SL). Head and body dorsoventrally depressed. Dorsal profile gently convex from snout tip to dorsal-fin origin, usually more flattened posterior to orbit, slightly convex and sloped ventrally from dorsal-fin origin to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Dorsal contour of head smooth, no ridge or keel, inconspicuous rounded elevations on the midline of the snout and anterior to orbits, supraoccipital nearly flat. Dorsal margin gently flattened from base of first branched dorsal- fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout rounded anteriorly. Eye moderately large. Lips forming an oval disk, covered with short papillae. Presence of a single narrow buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Very short maxillary barbel. Teeth bicuspid, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformly distributed odontodes. Tip of snout always naked, and, except in some large specimens, small naked area meeting the dorsolateral edge of upper lip on each side of tip of snout; in large specimens, dorsolateral margin of the upper lip supporting from few odontodes up to several plates covered with small odontodes. Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes. A narrow unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with enlarged odontodes in highly variable number, from less than ten up to approximately 40 in some large specimens. These cheek odontodes straight with tips curved, longest reaching from the first quarter (in small specimens) up to the third quarter (in large specimens) of the opercle. Usually three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine only slightly enlarged, except in large specimens (presumably males). Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; presence of a large smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin relatively short; when adpressed, never reaching adipose fin, often very distant from it. Adipose fin roughly triangular, preceded by one (or two fused into one) median unpaired raised plate. Adipose spine straight or slightly convex dorsally, membrane posteriorly convex. Pectoral-spine tip reaching to approximately one- third of pelvic spine in most specimens, exceptionally extending over its middle in large specimens (presumably males). Anal fin with weak spine, its margin convex. Caudal fin concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5 (except 3 specimens with I,4); caudal i,14, i. **Coloration.** Dorsal coloration pattern highly variable according to size of specimen and collection locality (see and discussion). In life, base color light grey-brown or orange-brown to dark brown, except whitish ventral region. Dorsal spotting pattern varies from presence of indistinct paler spots, to presence of numerous small to medium-sized, roundish to elongated, faded to brilliant yellow-orange spots on the whole body, or on its anterior part only; in that case spots form similarly colored transverse bands on posterior part of body. Such bands are observed on juveniles of all populations, and remain on larger specimens in some populations. Spots of similar color to those of body are usually present on dorsal and caudal fins, centred on fin rays and often combined to form regular or irregular transverse bands, more generally on caudal fin. Pectoral and pelvic fins with less-distinct spots, sometimes with large pale areas. Fin margins can be of same color as spots, especially in small specimens. Fin membranes usually grayish- brown. **Distribution and habitat.** *Guyanancistrus brevispinis* was found in the main Guianese river systems of Suriname and French Guiana, including the Corantijn at Guyanese-Surinamese border, the Nickerie, Coppename, Saramacca, Suriname, Maroni, Mana, Sinnamary, Mahury (Comté-Orapu), Kaw, Approuague and the Oyapock at the French-Brazilian border. These rivers all have a south-north orientation and flow into the Atlantic Ocean. It inhabits most clear-water rivers, streams and creeks, with running waters and rocky or sandy bottoms, over which specimens blend remarkably well. The species may be locally abundant. Its habitat is usually shadowed by primary rainforest, with measured water temperature varying from 23.7 to 28.4°C, conductivity from 13 to 42 us and pH from 5.96 to 7.06. **Etymology.** The specific name *brevispinis*, derived from Latin and meaning short thorns, was originally given in reference to the short evertible cheek odontodes. ### *Guyanancistrus brevispinis brevispinis* (Heitmans, Nijssen & Isbrücker, 1983) **Holotype.** Same as nominal species: ZMA 107.740, 126 mm SL\*; Surinam, Nickerie River system, district Sipaliwini \[not Nickerie\], Fallawatra River, rapid 5 km SW of Stondansie Fall; Nijssen, 6 April 1967. **Non type material.** See. **Diagnosis.** The nominal subspecies of *G*. *b*. *brevispinis* is differentiated from the subspecies *bifax* and *orientalis* by characteristic barcode sequences (see BOLD numbers). No morphometric variable strictly distinguishes *brevispinis* from the two other subspecies, but most mean values are significantly different from the other subspecies. These variables show that in comparison to them, on average, the body of *brevispinis* is significantly wider (cleithral width in % of SL: 31.98 ± 0.89% *vs* 30.75 ± 0.69 for *bifax* and 30.85 ± 0.75 for *orientalis*; p-value 3.498e^-16) (width at dorsal-fin origin in % of SL: 27.11 ± 1.66% *vs* 26.26 ± 1.28 for *bifax* and 26.13 ± 1.00 for *orientalis*; p-value 0.0004472); on average, the interbranchial distance is larger (in % of SL: 22.73 ± 0.97% *vs* 21.91 ± 0.65 for *bifax* and 22.22 ± 0.95 for *orientalis*; p-value 1.624e^-7), the interorbital wider (in % of SL: 11.41 ± 0.40 *vs* 11.09 ± 0.44 for *bifax* and 10.98 ± 0.35 for *orientalis*; p-value 3.15e^-7), the exposed part of opercle longer (in% of SL: 6.00 ± 0.64 *vs* 5.39 ± 0.53 for *bifax* and 5.34 ± 0.50 for *orientalis*; p-value 3.482e^-10), the thoracic length greater (in % of SL: 23.63 ± 1.01 *vs* 22.70 ± 1.15 for *bifax* and 22.59 ± 1.07 for *orientalis*; p-value 8.626e^-8), and the caudal peduncle deeper (in % of SL: 10.50 ± 0.40 *vs* 9.45 ± 0.47 for *bifax* and 9.97 ± 0.40 for *orientalis*; p-value 2.2e^-16). Also, *brevispinis* generally has fewer plates between adpressed dorsal-fin tip and adipose spine (1.5 ± 0.5 *vs* 2 ± 0.5 for *bifax* and 2 ± 0.5 for *orientalis*; p-value 3.484e^-6). **Distribution.** The nominate subspecies occurs from the Corantijn River basin in western Suriname to the upper Tapanahony River basin, a Marowijne River tributary in eastern Suriname. It was not found in French Maroni River tributaries. **Etymology.** As for nominal species. ### *Guyanancistrus brevispinis bifax* new subspecies urn:lsid:zoobank.org:act: 42111DB4-A0EA-4AC8-88DE-3A0BCE878091 (Figs, and ;) **Holotype.** MNHN 2017–0448 (ex MHNG 2734.090; GFSU12-140), 102.8 mm SL; French Guiana: Crique Petit Laussat, right tributary of Mana River (05°24'28.6"N 53°34'53.6"W); Covain & Fisch-Muller, 24 Oct. 2012. **Paratypes.** MHNG 2734.090 (GFSU12-141), 21; MNHN 2017–0449, 4; (11 measured, 67.0–106.5 mm SL), same data as holotype. MNHN 2015–227, 3, Mana River, Saut Fracas; Le Bail et al., 21 Sep. 1999. MHNG 2593.087, 5; MNHN 2015–221, 3; (4 measured, 46.8–96.3 mm SL) Grand Inini River, Saut “S”, right tributary of Maroni River; Le Bail et al., 1 Oct. 1997. **Non type material.** See. **Diagnosis.** *Guyanancistrus brevispinis bifax* is differentiated from other subspecies and species of *Guyanancistrus* by its barcode sequences (see BOLD numbers), which are however not unique as they are partly shared by *Guyanancistrus nassauensis*. It is distinguished from the latter by the diagnostic characters listed under species diagnosis. No morphometric variable unambiguously distinguishes *bifax* from other subspecies, but several have significantly different mean values. Compared to both *brevispinis* and *orientalis*, on average, *bifax* has a smaller interbranchial distance (in % of SL: 21.91 ± 0.65 *vs* 22.73 ± 0.97 for *brevispinis* and 22.22 ± 0.95 for *orientalis*; p-value 1.624e^-8), a more depressed caudal peduncle (in % of SL: 9.45 ± 0.47*vs* 10.50 ± 0.40 for *brevispinis* and 9.97 ± 0.40 for *orientalis*; p-value 2.2e^-8), and a longer caudal fin (lower caudal- fin spine in % of SL: 33.63 ± 2.38 *vs* 32.79 ± 2.36 for *brevispinis* and 32.53 ± 2.11 for *orientalis*; p-value 0.00684). Additional differences of morphometric variables from either *brevispinis* or *orientalis* are statistically significant, some being listed under the respective diagnoses of these subspecies. On average, the number of plates between dorsal base and adipose fin is smaller for *bifax* (6. ± 0.5 *vs* 6 ± 1 for *brevispinis* and 6 ± 0.5 for *orientalis*; p-value 0.01807). **Remark**: In several populations of the subspecies throughout its area of distribution, the head of some specimens (MHNG 2683.043; MHNG 2699.060; MHNG 2723.008) has a distinctive appearance, with an enlarged forehead part, usually coupled with a slightly longer snout, head and/or predorsal length, as well as an enlarged mouth. This difference between individuals appears not linked to sex, and apparently independant of the genetic data examined. **Distribution.** *Guyanancistrus b*. *bifax* occurs from the Maroni River and its eastern tributaries up to the Mana and Sinnamary rivers basins in French Guiana. **Etymology.** Named *bifax*, a noun in apposition, meaning two faces, for the different appearances of the head observed within the subspecies (see Remark). ### *Guyanancistrus brevispinis orientalis* new subspecies urn:lsid:zoobank.org:act: 116ABD70-36B3-4AB1-B35D-DDFC9ADC8CFE (Figs and ;) **Holotype.** MNHN 2017–0450 (ex MHNG 2681.098; GF06 183), 113.5 mm SL; French Guiana: forest creek, left tributary of Upper Oyapock River, in front of Roche Mon Père (03°16'56.3"N 52°12'36.6"W); Fisch-Muller et al., 6 Nov. 2006. **Paratypes.** MHNG 2681.098, 1 juvenile (measured, 33.4 mm SL), same data as holotype; MHNG 2723.012, 5; MHNG 2723.013, 5; MNHN 2015–225, 5; (8 measured, 32.2–122.1mm SL) first rapids of Crique Gabaret, left tributary of Lower Oyapock River; Fisch-Muller et al. 21 Oct. 1999. **Non type material.** See. **Diagnosis.** The subspecies *orientalis* is differentiated from other subspecies of *G*. *brevispinis* by characteristic barcode sequences (see BOLD numbers). No morphometric variable unambiguously distinguishes *orientalis* from the two other subspecies, but several have significantly different mean values. On average, *Guyanancistrus brevispinis orientalis* has a smaller internostril distance (in % of SL: 3.47 ± 0.48 *vs* 3.82 ± 0.54 for *brevispinis* and 3.85 ± 0.46 for *bifax*; p-value = 8.527e^-5), and a longer caudal peduncle (in % of SL: 29.42 ± 0.93 *vs* 28.87 ± 1.06 for *brevispinis* and 28.68 ± 1.09 for *bifax*; p-value = 0.001369). It also has fewer plates along adipose-fin base (mean 1.5 ± 0.5 *vs* 2 ± 0.5 for *brevispinis* and 2 ± 0.5 for *bifax*; p-value = 0.03897). Its interbranchial distance is larger than for *bifax* but smaller than for *brevispinis* (mean in % of SL: 22.22 ± 0.95 *vs* respectively 21.81 ± 0.65 and 22.73 ± 0.97; p-value = 1.624e^-7), and its caudal peduncle is deeper than for *bifax* but lower than for the nominal subspecies (mean in % of SL: 9.97 ± 0.40 *vs* respectively 9.45 ± 0.47 and 10.50 ± 0.40; p-value = 2.2e^-16). Additional differences of morphometric variables from either *brevispinis* or *bifax* are statistically significant, some being listed under the respective diagnoses of these subspecies. **Distribution.** *Guyanancistrus brevispinis orientalis* is distributed in eastern French Guiana, from the Mahury (Comté-Orapu) River Basin to the Approuague and Oyapock river basins. **Etymology.** The name *orientalis*, from the Latin name *oriens*, is given because of the eastern distribution of the subspecies. ### *Guyanancistrus nassauensis* Mol, Fisch-Muller & Covain, new species urn:lsid:zoobank.org:act: 523943AF-5A39-4A33-B783-0A7DA9A3AD0D (Figs and ;) *Guyanancistrus* sp. « big mouth »: Mol et al., 2007: 112 (potentially new species; collection localities); Wan Tong You, 2007: 249 (behavior in aquarium) *Guyanancistrus* sp. « Bigmouth »: Mol, 2012: 450–451 (short description, distribution and illustration) *Pseudancistrus* sp. Nassau: Covain & Fisch-Muller, 2012: 233 (molecular phylogeny of *Pseudancistrus sensu lato*) *Guyanancistrus* sp. Nassau: Covain & Fisch-Muller, 2012: 244 (generic placement); Mol et al., 2012: 274 (distribution in Suriname), 286 (threatened species) **Holotype.** MHNG 2679.100, 42.0 mm SL; Suriname: Sipaliwini: Paramaka Creek (Na3 site), Marowijne River Drainage, Nassau Mountains (4°51’36” N 54°35’ 30” W); J. H. Mol et al., RAP expedition, 3 April 2006. **Paratypes.** All from Suriname: Sipaliwini, Nassau Mountains, Paramaka Creek Basin, Marowijne River Drainage. MHNG 2745.064 (ex MHNG 2679.100), 2, 28.9–42.8 mm SL; collected with the holotype. MHNG 2679.099 (MUS 299–302), 4, 20.3–34.7 mm SL (1 measured, 34.7 mm SL); MHNG 2690.022, 1 postlarve; Paramaka Creek; J. H. Mol et al., RAP expedition, 29 March– 4 April 2006. AUM 50388, 14 (8 measured, 22.9–49.7 mm SL); NZCS F 7095 (ex AUM 50388), 1; NZCS F 7096 (ex AUM 50388), 1, IJs Creek, headwater tributary of Paramaka Creek (4°49‘14” N 54°36’19” W); J. W. Armbruster et al., 9 Sept. 2009. AUM 50396, 16 (5 measured, 38.8–49.6 mm SL); NZCS F 7097 (ex AUM 50396), 1; NZCS F 7098 (ex AUM 50396), 1, unnamed tributary of IJs Creek (4°51’04” N 54°35’24” W); J. W. Armbruster et al., 12 Sept. 2009. AUM 50740, 6 (3 measured, 37.2–47.0 mm SL); Creek entering Paramaka Creek below the mouth of IJs Creek; J. W. Armbruster & J. L. Wiley, 18 March 2010. AUM 50737, 4 (2 measured, 35.7–50.1 mm SL); Paramaka Creek (4°51’22” N 54°35’01” W); J. W. Armbruster & J. L. Wiley, 18 March 2010. AUM 50763, 2 (1 measured, 61.0 mm SL); Paramaka Creek just downstream of mouth of IJs Creek (4°51’39” N 54°353’59” W); J. W. Armbruster & J. L. Wiley, 19 March 2010. **Diagnosis.** *Guyanancistrus nassauensis* is distinguished from all congeners except *G*. *brevispinis* by its specific barcode sequences (GBOL093-13 and GBOL732-14). It is morphologically discriminated from all congeners by a small adult size (largest specimen observed 61 mm SL; adult size likely reached around 40 mm SL), by a reduced number of anal-fin rays (4 branched rays *vs* 5, apart from exceptional specimens), and by a wide oval mouth with both large dentary and premaxillary tooth cups (in % of head length, respectively: 24.2–31.9, mean 27.6, *vs* 23.6 or less except in *G*. *niger*, and 25.4–31.4, mean 28.1, *vs* 24.5 or less). Only *Guyanancistrus niger* has dentaries nearly as large (22.5–26.3, mean 25.0% of HL) but its premaxillaries are shorter (21.7–23.6, mean 22.6% of HL). *Guyanancistrus nassauensis* is distinguished from *G*. *longispinis* and from *G*. *niger* by a much shorter pectoral-fin spine (in % of SL: 22.2–26.3, mean 24.4, *vs* 31.9–45.5, mean 40.2, and 33.3–48.0, mean 42.8, respectively), and by color pattern (body and fins uniformly brown or with indistinct medium sized paler spots, *vs* brown-black with either small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). It is further separated from all *G*. *brevispinis* group species by having, on average, the widest body, the deepest and longest head, the largest interbranchial distance, the shortest fins, and the highest number of teeth. **Description.** Morphometric and meristic data in. Small-sized species (largest specimen observed 61.0 mm SL; holotype, 42.0 mm SL, likely a breeding male). Head and body dorsoventrally depressed and wide. Dorsal profile gently convex from snout tip to dorsal-fin origin, usually more flattened posterior to orbit, slightly convex and sloped ventrally from dorsal-fin origin to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. A low median ridge from tip of snout to nostrils, sometimes bordered by lateral depression, a slight elevation anterior to orbits, supraoccipital slightly convex to flat. Dorsal margin gently flattened from base of first branched dorsal-fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Large, rounded and laterally flattened snout. Eye relatively small. Large oval mouth, lower lip wide, not or just reaching pectoral girdle, upper lip narrower. Lips forming an oval disk, covered with short papillae. Presence of a single narrow buccal papilla. Very short maxillary barbel. Teeth short and strong with a relatively long bicuspid crown, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformely distributed odontodes.Tip of snout largely naked. Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes. A narrow unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with enlarged odontodes in highly variable number, from fewer than ten up to approximately 35 in some large specimens. These cheek odontodes straight with tips curved, the longest usually reaching middle of the opercle, or beyond in large specimens. Two to four rows of plates between supraoccipital plate and dorsal-fin spinelet, nuchal plate often covered by skin. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine enlarged, only slightly in small specimens, much more significantly in large specimens (presumably males). Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle covered with plates showing a highly reduced number of odontodes. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin short; when adpressed, far from reaching preadipose unpaired plate. Adipose fin roughly triangular, preceded by one, or two fused into one, median unpaired raised plate. Adipose spine straight or slightly convex dorsally, membrane posteriorly convex. Pectoral-spine short, tip usually reaching the first quarter of pelvic spine, exceptionally extending up to the first third in large specimens (presumably males). Anal fin short with weak spine, its margin convex. Caudal fin slightly concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,4 (except 1 specimen, 49.6 mm SL, with i,5); caudal i,14, i. **Coloration.** In alcohol, dorsal part of body uniformly grey-brown, ventral part yellowish except usually patches of melanophores on lateral parts and in the anal region, and abdomen whitish. Fin-rays brownish, with medium-sized spots by some specimens, these spots forming or not forming bands; margin of caudal fin often orange- or red-brown; fin-membranes usually not pigmented, or pigment restricted to areas bordering rays. In life (based on a photograph of one specimen), dorsal coloration of body brown with some lighter ill-defined orange- brown spots; fins orange-brown, fin-membranes hardly pigmented. **Distribution and habitat.** *Guyanancistrus nassauensis* is known solely from Paramaka Creek and some of its tributaries, Marowijne River Basin, in the Surinamese Nassau Mountains (an area of approximately 20x20 km2). At an elevation of 277 m, the type locality is located in a northern branch of Paramaka Creek, a medium-sized and shallow stream (3–7 m. width; less that 50 cm depth) with pools and some riffle habitat, a rocky substrate, and bordered by terra firme rainforest. Water was transparent, with a mean pH of 6.26, conductivity 24.2 μS/cm and temperature 23.2°C. Specimens were collected there by electrofishing with set seine, along with several *Harttiella crassicauda*, another species endemic to streams in the Nassau Mountains. **Etymology.** The name *nassauensis* is a reference to the distribution of the new species which is only known in streams in the Nassau Mountains, an area now under threat of a proposed bauxite mine and illegal gold mining. ### *Guyanancistrus brownsbergensis* Mol, Fisch-Muller & Covain, new species urn:lsid:zoobank.org:act: 02C600BB-5C91-4E0E-B25C-86738918BE28 (Figs and ;) **Holotype.** MHNG 2745.065 (JM14 01), 63.8 mm SL; Suriname: Brokopondo: Kumbu Creek above Kumbu Falls, Saramacca River Basin, Brownsberg Nature Park, Brownsberg Mountains (4°56’57” N 55°11’07” W); K. Wan Tong You, 15–16 Feb. 2014. **Paratypes.** NZCS F 7093 (JM14 02), 55.4 mm SL; MHNG 2745.066 (JM14 03), 49.2 mm SL; collected with the holotype. NZCS F 7094 (SU01-291), 1; MHNG 2723.037 (SU01-280), 1; MHNG 2724–008 (SU01-285), 1; MHNG 2724.009 (SU01-286), 1; MHNG 2724.011 (SU01-296): K. Wan Tong You, 01 July 2011 (all juveniles, 16.6–23.8 mm SL; not measured). **Diagnosis.** *Guyanancistrus brownsbergensis* is differentiated from all congeners by its specific barcode sequences (GBOL697-14, GBOL689-14, GBOL690-14, GBOL691-14, GBOL726-14, GBOL725-14, and GBOL724-14). It is morphologically distinguished from *G*. *longispinis* and *G*. *niger* by a much shorter pectoral-fin spine (in % of SL: 27.7–29.7, mean 28.7, in SL *vs* 31.9–45.5, mean 40.2, and 33.3–48.0, mean 42.8, respectively), having shorter odontodes, and by color pattern (body and fins medium grey-brown with yellowish beige to light brown medium to large-sized spots, *vs* brown-black with either small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). *Guyanancistrus brownsbergensis* is distinguished from all species of the *Guyanancistrus brevispinis* group except *G*. *nassauensis* by a deeper caudal peduncle (11.4–11.6, mean 11.5, *vs* 11.3 or less % of SL). It is differentiated from *G*. *nassauensis* by smaller dentary and premaxillary tooth cusps (18.5 and 17.6% of head length, *vs* 24.2–31.9, mean 27.6, and 25.4–31.4, mean 28.1) by an anal fin with 5 branched rays (*vs* 4). The caudal peduncle of *G*. *brownsbergensis* is not only deep, but it is also short compared to *G*. *tenuis* and *G*. *megastictus* (28.0–29.1, mean 28.5 in % of SL *vs* respectively 29.4–32.1, mean 31.1, and 31.0–31.6, mean 31.3). *Guyanancistrus brownsbergensis* can further be distinguished from *G*. *brevispinis* by longer evertible cheek odontodes (reaching beyond posterior end of opercle, *vs* not reaching its last quarter), and by a larger number of plates between adpressed dorsal fin and adipose fin (3–3.5, mean 3, *vs* 0.5–3, mean 2). It is separated from *G*. *teretirostris* and *G*. *megastictus* by a longer pelvic-fin spine (reaching beyond end of anal-fin base *vs*, respectively, not reaching origin of anal fin and not reaching beyond end of anal-fin base). **Description.** Morphometric and meristic data in. Head and body strongly dorsoventrally depressed. Dorsal profile gently convex from snout tip to dorsal- fin origin, flattened posterior to orbit, slightly convex and sloped ventrally from dorsal-fin origin to end of adipose fin, then slightly concave and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Very low median ridge from tip of snout to nostrils present, parallel to this, similar inconspicuous ridges from snout border to nostrils, then somewhat more elevated to orbits, supraoccipital nearly flat. Dorsal margin gently flattened along dorsal-fin base between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle high, roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout rounded anteriorly. Eye relatively small. Lips forming an oval disk, covered with short papillae. Presence of a single small and narrow buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Very short maxillary barbel. Teeth slender, bicuspid, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformely distributed odontodes. Tip of snout naked; small (small specimens) to minute (holotype, which is the largest specimen) naked area on each side of the latter; dorsolateral margin of the upper lip supporting patches of odontodes (very small in the smallest specimen). Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes. A narrow unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with approximately 25 to 50 enlarged odontodes. These cheek odontodes straight with tips curved, the longest reaching beyond end of opercle. Three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine moderately enlarged. Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; presence of a large smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin relatively short; when adpressed, distant by one (holotype) or two plates from median unpaired plate preceding adipose fin. Adipose fin roughly triangular; spine slightly convex dorsally, membrane straight or slightly concave posteriorly. Pectoral-spine tip reaching first quarter of pelvic spine. Anal fin with weak spine, its margin convex. Caudal fin concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5; caudal i,14, i. **Coloration.** Dorsal coloration pattern grey-brown in life, lighter in alcohol. Medium to large-sized spots irregularly distributed on dorsal part of body. These spots are yellowish-beige to light brown. Spots on fins, at least on caudal, forming bands. Border of dorsal and caudal fins orangish colored in life. Ventral coloration pale yellow and unspotted. **Distribution and habitat.** Collected only in the Upper Kumbu Creek in the Brownsberg Nature Park, Brownsberg Mountains, at altitude 200–430 m above mean sea level. The Upper Kumbu Creek at Kumbu Falls (430 m asl) is a small mountain stream (2.5–3.7 m wide, 28–50 cm water depth) with cool (23.1–23.2°C) water, high dissolved oxygen content (93–96% saturation; 7.08–7.72 mg/L), a pH of 7.0–7.5, conductivity 30.8–31.6 μS/cm, and a current strength of 0.29–0.56 m/s (12 July 2014). The bottom substrate consists of sand, gravel, pebbles, boulders and bedrock. The water is mostly clear. Overhanging vegetation, leaf litter and some woody debris offer shelter. **Etymology.** Species named for the Brownsberg Nature Park in Brownsberg Mountains, in which it was found, and which is presently under threat from illegal gold mining. ### *Guyanancistrus teretirostris*, new species urn:lsid:zoobank.org:act:3D9F8677-4505-47EE-8111-7636ABF48A25 **Holotype.** MZUSP 117149 (ex MHNG 2723.004; SU07-654), 97.6 mm SL; Brazil: Sipaliwini-Parú Savannah in Trio Amerindian territory at the Suriname-Brazil border, Vier Gebroeders (Four Brothers) Mountains in a tributary of the Parú de Oeste River, gift of the Trio tribe in Sipaliwini, 20–21 Oct. 2007. **Paratypes.** MHNG 2723.004 (SU07-652, 653), 2, 87.0 and 97.2 mm SL; same data as holotype. **Diagnosis.** *Guyanancistrus teretirostris* is distinguished from all congeners by its specific barcode sequences (GBOL735-14, GBOL734-14, and GBOL733-14). It is morphologically distinguished from *G*. *longispinis* and from *G*. *niger* by a much shorter pectoral-fin spine (in % of SL: 27.4–29.7, mean 28.5, *vs* respectively 31.9–45.5, mean 40.2, and 33.3–48.0, mean 42.8, mean 44.5) supporting shorter odontodes, and by color pattern (yellow-beige to light brown small to medium-sized spots on body and fins, *vs* either small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). In the *brevispinis* group, *Guyancistrus teretirostris* is distinguished from *G*. *nassauensis*, *G*. *brownsbergensis*, and *G*. *megastictus* by a narrower body (cleithral width in % of SL 29.7–31.1, mean 30.5, *vs* respectively: 32.2–36.6, mean 34.3; 31.5–31.7, mean 31.6; and 31.8–32.7, mean 32.2), and, from the latter three species, by shorter dorsal- and pelvic-fin spines (dorsal spine in % of SL: 23.0–23.5, mean 23.3, *vs* respectively: 24.3–25.6, mean 25.0; 26.0; and 24.5–26.1, mean 25.3; pelvic spine in % of SL: 21.5–23.5, mean 22.5, *vs* respectively: 25.0–26.7, mean 26.0; 25.3; 24.6–26.3, mean 25.5). Pelvic-fin length discriminates *teretirostris* from *G*. *tenuis* (23.5–26.1, mean 24.8), from which it can futher be distinguished by caudal peduncle depth (10.5–10.8, mean 10.6% of SL, *vs* 8.9–9.6, mean 9.3), and by mean number of plates bordering supraoccipital (3–4, mean 3.5, *vs* 3–5, mean 4.5), and separating adpressed dorsal fin and adipose fin (2–3, mean 2.5, *vs* 3–4, mean 3). *Guyanancistrus teretirostris* is distinguished from *G*. *brevispinis* by longer evertible cheek odontodes (reaching end of opercle or beyond *vs* reaching first to third quarter). A particularly short but also depressed head further discriminates G. *teretirostris* from G. *nassauensis* (in % of SL, head length: 31.7–32.3, mean 31.9, *vs* 32.2–40.7, mean 36.4; head depth: 13.8–15.3, mean 14.7, *vs* 16.0–18.1, mean 17.1). On average, head length distinguishes it from all other species of the *brevispinis* group, including *G*. *brevispinis*. **Description.** Morphometric and meristic data in. Head and body dorsoventrally depressed. Dorsal profile gently convex from snout tip to orbit level, then nearly flat, slightly convex and sloped ventrally from dorsal-fin origin to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Dorsal contour of head smooth, no ridge or keel, inconspicuous rounded elevations on the midline of the snout and anterior to orbits, supraoccipital nearly flat. Dorsal margin gently flattened from base of first branched dorsal- fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout fully rounded anteriorly. Eye moderately large. Lips forming an oval disk, covered with short papillae. Presence of a single narrow buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Short maxillary barbel. Teeth bicuspid, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformely distributed odontodes. Tip of snout naked; very small area on each side of the latter is also naked in the smaller paratypes; dorsolateral margin of the upper lip supporting several platelets with short odontodes, or naked (smallest paratype). Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes. A narrow unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with approximaterly 25–35 enlarged odontodes. These cheek odontodes straight with tips curved, longest nearly reaching posterior end of opercle (smallest paratype) or beyond (holotype and second paratype). Three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine only slightly enlarged. Abdomen totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; presence of a moderately large smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin relatively short; when adpressed, distant by at least one plate from median unpaired plate preceding adipose fin. Adipose fin roughly triangular; spine slightly convex dorsally, membrane posteriorly convex. Pectoral-spine tip nearly reaching third of pelvic spine (holotype), or less. Anal fin with weak spine, its margin convex. Caudal fin concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5 (i,4 in paratype 87 mm SL); caudal i,14,I. **Coloration.** In alcohol, dorsal ground color of body medium grey-brown, somewhat darker on head and lighter on lower part of caudal peduncle. Body dorsally covered with yellow-beige to light-brown small to medium-sized spots, usually rounded anteriorly and more irregular in shape posteriorly. Body ventrally yellow-beige, with abdomen whitish, unspotted. All fins of similar color to body, and spotted, anal fin excepted. Spots of dorsal-fin medium sized and rounded. Spots of paired fins similar but less distinct; anterior part of these fins clearly darker that posterior part. Spots of caudal fin forming three to four large light and irregular transverse bands; margin of fin apparently orangish colored. **Distribution.** Known from the Upper Parú de Oeste River. **Etymology.** The species name *teretirostris*, is derived from the Latin words *teres*, meaning rounded and smooth, and *rostris*, meaning snout; an allusion to the snout shape of this species. ### *Guyanancistrus tenuis*, new species urn:lsid:zoobank.org:act:225451B6-1C40-4DC0-BC5A-553A4B2530D4 (Figs and ;) **Holotype.** MZUSP 117148 (ex MNHN 2002–3537; GF Mit06), 90.9 mm SL; Brazil: Para: small tributary of Rio Mapaoni, upper Jari River Basin, Massif du Mitaraka (2°16’45”N 54°32’39”W); P. Keith & P. Gaucher, 25 Oct. 2002. **Paratypes.** MNHN 2002–3537, 19, 23.7–81.1 mm SL; MHNG 2745.067, 12, 25.0–89.8 mm SL (GF Mit01-05); same data as holotype. **Diagnosis.** *Guyanancistrus tenuis* is distinguished from all congeners by its specific barcode sequences (GBOL739-14, GBOL738-14, and GBOL737-14). Morphologically, it is distinguished from *G*. *longispinis* and *G*. *niger* by a much shorter pectoral-fin spine (in % of SL: 26.0–28.1, mean 27.0, *vs* respectively 31.9–45.5, mean 40.2, and 33.3–48.0, mean 42.8) supporting shorter odontodes, and by color pattern (yellow-beige medium to large-sized spots and bands on body and fins, *vs* small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). *Guyanancistrus tenuis* is a particularly slender species, and, with *G*. *megastictus*, it is the most depressed of all *brevispinis* group species. *Guyanancistrus tenuis* can be separated from *G*. *nassauensis* and *G*. *brownsbergensis* by lower head depth values (13.5–14.8 mean 14.0% of SL *vs* 15.2 or more), and from all species but *G*. *brevispinis* by a lower caudal peduncle depth (8.9–9.6 mean 9.3% of SL *vs* 10.4 or more), but only mean head depth discriminates *G*. *tenuis* from *G*. *brevispinis* (14.1–19.6, mean 16.2% of SL in the latter). In the *brevispinis* species group, *G*. *tenuis* additionally shows the narrowest body, distinguishing it from *G*. *nassauensis* and *G*. *megastictus* (cleithral width in % of SL: 27.9–31.7, mean 30.1, *vs* respectively: 32.2–36.6, mean 34.3; and 31.8–32.7, mean 32.2). *Guyanancistrus tenuis* is distinguished from *G*. *brevispinis* by longer evertible cheek odontodes (reaching last quarter of opercle up to largely beyond end of opercle in specimens of approximately 60 mm SL *vs* reaching first to third quarter of it). *Guyanancistrus tenuis* can further be distinguished from *G*. *brevispinis* by a higher number of plates bordering the supraoccipital (3–5, mean 4.5, *vs* 2–3, mean 3) and between the adpressed dorsal fin and the adipose fin (3–4, mean 3, *vs* respectively 0.5–3, mean 2), from *G*. *nassauensis* by smaller dentary and premaxillary tooth cusps (in % of head length, respectively 15.8–20.6, mean 17.4, *vs* 24.2–31.9, mean 27.6, and 15.1–21.1, mean 18.0, *vs* 25.4–31.4, mean 28.1) and by an anal fin with 5 branched rays (*vs* 4), and from *teretirostris* by a longer pelvic-fin spine (23.5–26.1, mean 24.8% of SL *vs* 21.5–23.5, mean 22.5). **Description.** Morphometric and meristic data in. Head and body up to caudal peduncle very dorsoventrally depressed and narrow, resulting in a slender aspect. Dorsal profile gently convex from snout tip to orbit level, then nearly flat to dorsal-fin origin, slightly convex and sloped ventrally from that point to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Dorsal contour of head smooth, usually a very low median ridge from tip of snout to nostrils, slight elevation anterior to orbits, sometimes (including holotype) bordered by a shallow lateral depression, supraoccipital nearly flat. Dorsal margin gently flattened from base of first branched dorsal-fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout rounded anteriorly. Eye moderately large. Lips forming an oval disk, covered with short papillae. Presence of a single narrow buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Short maxillary barbel. Teeth slender, bicuspid, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformely distributed odontodes. Tip of snout naked, and often (particularly in small specimens) also a very small naked area on each side of the latter, separated by a plated area which continues for a short distance on dorsolateral margin of upper lip. Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes; on its ventral margin, odontodes usually slightly enlarged. A relatively large unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with enlarged odontodes in highly variable number, from approximately 10 up to approximately 35 in large specimens. These cheek odontodes straight with tips curved, longest not reaching middle of opercle in smallest specimens, but reaching it or beyond in specimens of approximately 50 mm SL, and reaching last quarter to well beyond end of opercle in specimens of approximately 60 mm SL (beyond in holotype). Usually three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine very slightly enlarged. Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; a moderately large smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin short; when adpressed, tip of fin very distant from adipose fin, and even far from reaching preadipose unpaired plate. Adipose fin roughly triangular, preceded by one, or two fused into one, median unpaired raised plate. Adipose spine straight or slightly convex dorsally, membrane posteriorly straight or slightly convex. Pectoral-spine tip reaching slightly over pelvic-fin origin. Anal fin with weak spine, its margin convex. Caudal fin concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5; caudal i,14, i. **Coloration.** In alcohol, dorsal ground color of body brown, covered with yellow-beige medium-sized spots on head, becoming gradually much larger spots up to end of caudal peduncle. In small specimens, some spots are roundish and large (but covering fewer than four plates) but spots usually coalesce to form large and highly contrasted stripes on posterior part of body. Ventrally, color of body more or less uniformly light brown apart from the abdomen, which is mainly whitish, sometimes with diffuse brown pigmentation. All fins colored similarly to dorsum, and spotted. Spots of dorsal fin medium sized, forming transverse bars or not; usually a dark spot on membrane between origin of spine and first branched ray. Spots of other fins less distinct. Spots of caudal fin forming two to three highly constrasted, large and irregular light-colored transverse bands; tips of fin light-colored. **Distribution.** *Guyanancistrus tenuis* is known solely from a small forest tributary of the Mapaoni River, Upper Jari River Basin, in the Massif du Mitakara, a mountain range in the far southwest of French Guiana. This north- south oriented mountain creek was essentially rocky, shallow (20–60 cm depth), with medium to strong currents, and some pools. **Etymology.** The name *tenuis* is a Latin word meaning thin, in reference to the slender body of the species. ### *Guyanancistrus megastictus*, new species urn:lsid:zoobank.org:act:AAB3CE9E-055D-4DD6-8B9C-682FC8F5E50B (Figs and ;) **Holotype.** MNHN 2002–3508, 62.7 mm SL; French Guiana: Crique Alama, tributary of Crique Saranou, Maroni River Basin, Massif du Mitaraka (2°18’08”N 54°32’00”W); P. Keith & P. Gaucher, 25 Oct. 2002. **Paratype.** MHNG 2745.068 (ex MNHN 2002–3508), 1, 57.1 mm SL; same data as holotype. **Diagnosis.** *Guyanancistrus megastictus* is distinguished from all congeners by specific barcode sequences (GBOL897-15 and GBOL898-15). Morphologically, it is distinguished from *G*. *longispinis* and *G*. *niger* by a shorter pectoral- fin spine (in % of SL: 28.1–28.3, mean 28.2, *vs* respectively 31.9–45.5, mean 40.2, and 33.3–48.0, mean 42.8), supporting shorter odontodes, and by color pattern (pale yellowish medium to very large sized spots or bars on body and fins, *vs* small roundish yellow spots for *G*. *longispinis*, or white dots for *G*. *niger*). Color pattern, with particularly large spots on body posterior to dorsal fin, and a caudal fin mainly light colored by the presence of a single very large yellowish bar (*vs* several light spots or bands), also distinguish *G*. *megastictus* from all *brevispinis* group species. *Guyanancistrus megastictus* is distinguished: from *G*. *brevispinis* by longer evertible cheek odontodes (reaching last quarter of opercle or beyond its posterior end, *vs* not reaching last quarter of opercle); from *G*. *nassauensis* by smaller dentary and premaxillary tooth cusps (in % of head length, respectively: 16.8, *vs* 24.2–31.9, mean 27.6, and 17.4–17.8, mean 17.6, *vs* 25.4–31.4, mean 28.1) and by an anal fin with 5 branched rays (*vs* 4); from *G*. *brownsbergensis* by less deep head (13.9–15.1, mean 14.5% of SL *vs* 15.2–15.7, mean 15.5) and lower caudal peduncle (10.4–10.9, mean 10.7% of SL *vs* 11.4–11.6, mean 11.5); and from *G*. *teretirostris* and *G*. *tenuis* by a larger body (in % of SL, 31.8–32.7, mean 32.2 *vs* respectively 29.7–31.1, mean 30.5, and 27.9–31.7, mean 30.1). **Description.** Morphometric and meristic data in. Dorsal profile gently convex from snout tip to orbit level, then nearly flat to dorsal-fin origin, slightly convex and sloped ventrally from that point to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Dorsal contour of head smooth, a very low median ridge from tip of snout to nostrils, slight elevation anterior to orbits, bordered (paratype) or not (holotype) by a shallow lateral depression, supraoccipital nearly flat. Dorsal margin gently flattened from base of first branched dorsal-fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout rounded anteriorly. Eye relatively small. Lips forming an oval disk, covered with short papillae. Presence of a single triangular buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Short maxillary barbel. Teeth slender, bicuspid, lateral lobe about half the size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformely distributed odontodes. Tip of snout naked, and a minute naked area on each side of the latter, separated by a plated area which continues for a short distance on dorsolateral margin of upper lip. Lateral margin of snout covered with plates forming a rigid armor with short odontodes. Opercle supporting odontodes. A straight unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with approximately 25–30 enlarged odontodes, straight with tips curved, longest nearly reaching posterior end of opercle or beyond. Three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine very slightly enlarged. Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; a moderately large smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin short; when adpressed, tip of fin very distant from adipose fin, and even far from reaching preadipose unpaired plate. Adipose fin roughly triangular, preceded by one, or two fused into one, median unpaired raised plate. Adipose spine relatively long compared to other *Guyanancistrus* species, slightly convex dorsally, membrane posteriorly straight. Pectoral-spine tip reaching slightly beyond pelvic-fin origin or nearly one fifth of fin spine (holotype). Anal fin with weak spine, its margin convex. Caudal fin concave, ventral lobe longer than dorsal lobe. Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5; caudal i,14, i. **Coloration.** In alcohol, dorsal ground color of body a bleached brown, covered with yellow-beige medium-sized spots on head, then large spots, and very large roundish spots (covering at least six lateral plates) and bars posterior to dorsal-fin origin level and up to end of caudal peduncle. Ventrally, color of body more or less uniformly yellowish apart from mainly whitish abdomen. All fins except anal similarly colored to dorsum, and lightly spotted. Spots of dorsal fin large, forming one or two large transverse bars. Spots on paired fins less distinct. Anal fin yellowish. Caudal fin mainly light colored: narrow brownish base, followed by a very large lightly colored transverse bar, then a narrower brownish transverse bar, and tips of fin light. In life, background color greenish-brown, with darker areas surounding the light spots, and caudal-fin base also darker. **Distribution.** *Guyanancistrus megastictus* is known from a small forest tributary of the Upper Maroni River Basin in the Massif du Mitaraka, a mountain range in the far south-west of French Guiana. The only two known specimens were caught poison fishing in a shallow (20–60 cm depth) and mainly sandy portion of this river named Crique Alama. **Etymology.** The Latin word *megastictus* is derived from the Ancient Greek *mega*, meaning large, and *stictos*, meaning spotted, in reference to the presence of very large size spots on body and fins. ### Key to species of *Guyanancistrus* 1. 1 - Presence of distinct spots on body and fins, all spots roundish and smaller than size of a lateral dermal plate; pectoral-fin spine length 31.9–45.5% of SL...........................2     - Absence of distinct spots on body and fins, or presence of spots at least as large as a dermal plate, or coalescing, or forming bands on posterior part of body and fins; pectoral-fin spine length 22.2–34.4% of SL...............................................3 2. 2 - Body and fins covered with small roundish yellow spots; odontodes on dorsolateral margin of the upper lip minute; dorsal-fin base length 29.0–32.0% of SL............ *G*. *longispinis*     - Body and fins covered with minute white dots; odontodes on dorsolateral margin of the upper lip elongated; dorsal-fin base length 24.8–28.8% of SL.........................*G*. *niger* 3. 3 - Anal fin with 4 branched rays; dentary tooth cup 24.2–31.9% of head length *G*. *nassauensis*     - Anal fin with 5 branched rays; dentary tooth cup 23.6% or less of head length...................................4 4. 4 - Longest evertible cheek odontodes reaching the first half of the opercle (except in some large specimens surpassing 70 mm SL reaching the third quarter but not reaching its last quarter)....................................*G*. *brevispinis*     - Longest evertible cheek odontodesreaching the last quarter of opercle or beyond its posterior end (except in very small specimens)...........................5 5. 5 - Pelvic-fin spine not reaching origin of anal fin........................*G*. *teretirostris*     - Pelvic-fin spine reaching beyond origin of anal fin....................................6 6. 6 - Orbital diameter 1.7–2.1 times in interorbital width; depth of caudal peduncle 3.1–3.6 times in its length...............................*G*. *tenuis*     - Orbital diameter 2.2–2.4 times in interorbital width; depth of caudal peduncle 2.5–3.0 times in its length..............................................7 7. 7 - Pelvic-fin spine reaching beyond end of anal fin base; depth of caudal peduncle 2.5 times in its length................................... *G*. *brownsbergensis*     - Pelvic-fin spine not reaching beyond end of anal fin base; depth of caudal peduncle 2.9–3.0 times in its length...................................... *G*. *megastictus* ### *Cryptancistrus* new genus urn:lsid:zoobank.org:act:F5ADD7D1-7A87-41A2-9DFE-78F04612C4BA **Type-species**. *Cryptancistrus similis*, new species urn:lsid:zoobank.org:act:6FE8FA29-E129-4CCA-BA43-27A51DB314E2 **Diagnosis.** *Cryptancistrus* is characterized by its unique barcode sequence (GBOL736-14). No unique morphological character was found to diagnose the genus which belongs to the Ancistrini tribe of the Hypstominae subfamily. The following combination of characters distinguishes *Cryptancistrus* from all other Hypostominae genera: head and body dorsoventrally depressed; head and body plates not forming prominent ridge or crest; snout rounded, and covered with contiguous plates except tip region, and posterior part of lateral margin of snout; latter area forming a soft fleshy border, and bearing slightly enlarged odontodes associated with small fleshy tentacules sensu Sabaj et al.); presence of odontodes over a broad area on the opercle; presence of numerous enlarged cheek odontodes supported by evertible plates; these odontodes straight with tips slightly curved, as opposed to strongly hook-shaped; absence of whisker- like cheek odontodes; absence of enlarged odontodes along snout margin; presence of a dorsal iris operculum; lips forming an oval disk; dentary and premaxillary with numerous viliform and bicuspid teeth; presence of a small buccal papilla, no enlarged dentary papilla; seven branched dorsal-fin rays; presence of an adipose fin; no membranous extension between end of dorsal fin and adipose fin; five series of lateral plates extending to caudal fin; lateral plates not keeled and not bearing enlarged odontodes; lateral plates of ventral series on caudal peduncle angular but not keeled; abdominal region entirely naked. *Cryptancistrus* is externally mostly similar to *Guyanancistrus*. It is distinguished from *Guyanancistrus* primarily by the fleshy posterior part of lateral margin of snout bearing slightly enlarged odontodes associated with small fleshy tentacules (*vs* plates along margin of snout forming a rigid armour covered with minute odontodes, absence of tentacules). It can additionally be distinguished from *Gruyanancistrus* by a skin region bordering the exposed portion of opercle roughly as large as the latter (*vs* distinctly narrower than the latter). **Etymology**. The name *Cryptancistrus* is derived from the Greek names *kryptos*, meaning hidden, and *ankistron*, meaning hook, in reference to the genera *Ancistrus*, type genus of the tribe Ancistrini to which it and *Guyanancistrus* Isbrücker, 2001, to which it is externally the most similar, belong. **Distribution**. Known only from type species locality, Upper Parú de Oeste River basin, Brazil. ### *Cryptancistrus similis*, new species (Figs and ;) **Holotype.** MZUSP 117150 (ex MHNG 2723.005; SU07-672), 61.7 mm SL; Brazil: Sipaliwini-Parú Savannah in Trio Amerindian territory at the Suriname-Brazil border, Vier Gebroeders (Four Brothers) Mountains in a tributary of the Parú de Oeste River, gift of the Trio tribe in Sipaliwini, 20–21 Oct. 2007. **Diagnosis.** As given for genus. **Description.** Morphometric and meristic data of the holotype (only known specimen) in. Head and body dorsoventrally depressed. Dorsal profile gently convex from snout tip to orbit level, then nearly flat, slightly convex and sloped ventrally from dorsal-fin origin to adipose fin, then slightly concave to procurrent caudal-fin rays, and rising to caudal fin. Ventral profile flat from snout to base of caudal fin. Dorsal contour of head smooth, no ridge or keel, inconspicuous rounded elevations on the midline of the snout and anterior to orbits, supraoccipital nearly flat. Dorsal margin gently flattened from base of first branched dorsal- fin ray to base of adipose fin between very slight ridges formed with lateral plates of dorsal series. First lateral plates of mid-ventral series forming low lateral ridge. Caudal peduncle roughly ovoid in cross section, flattened ventrally, and more compressed posteriorly. Snout rounded anteriorly. Eye relatively large. Lips forming an oval disk, covered with short papillae. Presence of a single narrow buccal papilla. Lower lip wide, not reaching pectoral girdle, upper lip narrower. Very short maxillary barbel. Teeth bicuspid, lateral lobe about half size of medial lobe. Head and body plated dorsally, plates generally covered by short and uniformly distributed odontodes. Snout plated except tip naked. Anterior margin of snout carrying slightly enlarged odontodes; meeting the latter, dorsolateral margin of upper lip supporting plates and short odontodes. Posterior part of lateral margin of snout forming a soft fleshy border bearing slightly enlarged odontodes with small tentacules sensu Sabaj et al., cutaneous sheath surrounding base of odontodes being enlarged and partially detached from odontodes. Opercle supporting odontodes, those on inferior margin slightly enlarged. A large unplated area bordering posterodorsal margin of opercle. Evertible cheek plates with approximately 40 enlarged odontodes, straight with tips curved, longest reaching beyond the end of opercle (on right side of the holotype; longest odontodes missing on left side). Usually three rows of plates and a curved nuchal plate between supraoccipital plate and dorsal-fin spinelet. Five series of lateral plates extending to caudal fin. Odontodes on lateral series of plates not forming keels. Odontodes on posterior part of pectoral-fin spine enlarged. Abdominal region totally naked. No platelike structure before the anal fin. Ventral part of caudal peduncle plated; presence of a small smooth area devoid of odontodes around anal fin. Dorsal-fin origin slightly anterior to pelvic-fin origin. Dorsal fin relatively large; when adpressed, nearly reaching adipose fin. Adipose fin roughly triangular, preceded by a median unpaired raised plate. Adipose spine nearly straight, membrane posteriorly convex. Pectoral-spine tip reaching approximately one-fifth of pelvic spine. Anal fin with weak spine, its margin convex. Caudal fin obliquely truncate, very slightly concave, inferior part longer (part of upper lobe damaged by holotype). Fin-ray formulae: dorsal II,7; pectoral I,6; pelvic i,5; anal i,5; caudal i,14, i. **Coloration.** In alcohol, dorsal ground color of body medium brown, covered with yellowish-beige spots. Spots roundish, medium to large-sized, larger on posterior part than on anterior part, and non-coalescent. Ventrally, plated parts of body yellow-beige, abdomen whitish with some yellow-beige areas. All fins of slightly darker color than body, and similarly spotted apart from anal. Spots of dorsal fin and paired fins medium sized; a dark spot on membrane between origin of spine and first branched ray. Spots of caudal fin larger, some of them coalescent, more or less forming two large light and irregular transverse bands; margin of fin light. **Distribution.** Known from a single specimen from the Upper Parú de Oeste River, and collected along with *Guyanancistrus teretirostris* n. sp. and an unidentified *Hypostomus* species, and with two recently described Loricariinae, *Cteniloricaria napova* and *Harttia tuna*. **Etymology.** The Latin name *similis*, meaning similar, refers to the strong morphological resemblance between the new species of *Cryptancistrus* and the type species of *Guyanancistrus*, *G*. *brevispinis*. ## Biogeography of *Guyanancistrus* members Comparison between likelihoods of ancestral area reconstructions along the phylogenetic tree using DEC and DEC + *j* models showed that the latter had significantly better fit. Resolutions of the different polytomies of the phylogenetic tree placed the population of *Guyanancistrus brevispinis brevispinis* of Tapanahony River in sister position to *G*. *b*. *bifax* and *G*. *b*. *orientalis* members, and split the population of *G*. *b*. *brevispinis* of Saramacca River in two subpopulations. Even though apparently contradictory to previous results, these resolutions did not impact ancestral area reconstructions, and reinforced the power of the analysis. The biogeographic analysis of *Guyancistrus* members under DEC + *j* model reconstructed a broad ancestral area comprising Amazonian headwaters including Upper Jari and Paru de Oeste rivers, the Oyapock, and Maroni rivers at the root of the phylogenetic tree, even though this reconstruction was ambiguous (see pie charts of states probabilities in). From this ancestral area, the *G*. *niger* and *G*. *longispinis* lineages split from all other ancestral *Guyanancistrus* by vicariance of the Oyapock Basin. Then a second vicariant event occurred between Amazonian headwaters and the Maroni Basin, splitting the *G*. *megastictus*, *G*. *tenuis*, *G*. *teretirostris* and *G*. *brownsbergensis* lineages from that of *G*. *nassauensis* and *G*. *brevispinis*. In the Amazonian group, two dispersals followed by speciation occurred. From an Amazonian ancestor (likely from Jari River), the ancestors of *G*. *megastictus* dispersed toward the Maroni Basin, whereas ancestors of *G*. *brownsbergensis* likely dispersed from the Paru de Oeste River toward the Saramacca River. In the Maroni group, two speciation events occurred, leading on one side to *G*. *nassauensis* and on the other to *G*. *brevispinis*. Dispersal patterns of *G*. *brevispinis* members appeared more complex, with multiple dispersals among Guianese rivers. From the central Maroni River, a first dispersal occurred to the west toward the Suriname River. Then, a second dispersal from headwaters of the Surinamese Maroni (i.e. Marowijne River) took place toward the headwaters of the Corantijn River, whereas part of the ancestral population of the Maroni River stayed in this basin (now present in the Tapanahony River). From the Upper Corantijn River, ancestral populations spread toward the Lower Corantijn, and then dispersed again to the East toward the Nickerie River, and from the latter to the Saramacca River. All these movements lead to the differentiation of the western form *G*. *brevispinis brevispinis*. A second dispersal, again from the Maroni, occurred to the east toward the Oyapock River. From there, ancestral populations successively dispersed to the west toward the Approuague and Comté- Orapu rivers, leading to the establishment of the present eastern form *G*. *brevispinis orientalis*. Finally, once more from the Maroni, ancestral populations dispersed toward the Upper Sinnamary River, and then toward the Mana River, and from the latter toward the Sinnamary leading to the establishment of the present Central form *G*. *brevispinis bifax*. # Discussion ## The genus *Guyanancistrus* A general similarity of head and body shape, hardly elevated and describing smooth contours, seems to unite all *Guyanancistrus* species. However no shared characteristic was found to be unique for the group. *Guyanancistrus longispinis* and *G*. *niger* appear morphologically quite distinct compared with species of the *brevispinis* group. Even so, mitochondrial and nuclear sequences of the various species unambiguously showed *Guyanancistrus* to be monophyletic (, present results), confirming the validity of the genus as it was originally separated from *Lasiancistrus* by its author (Isbrücker in). Nevertheless, while the latter is well diagnosed, it is not very closely related to *Guyanancistrus* (see also Armbruster: 12, based on *G*. *brevispinis*). The sister clade of *Guyanancistrus* contained three genera: *Hopliancistrus* confirming results of Covain & Fisch-Muller, *Corymbophanes* as in Lujan et al., and the new genus *Cryptancistrus*. *Hopliancistrus* is globally similar to *Guyanancistrus* but clearly distinguished by the diagnostic presence of up to three strongly hook-shaped cheek odontodes and of enlarged odontodes on the sides of the snout. *Corymbophanes* is also easily distinguished from *Guyanancistrus* by the absence of evertible cheek odontodes, the absence of an adipose fin, and by the presence of an elongate postdorsal ridge of 13–17 raised unpaired platelets. In contrast, *Cryptancistrus* is only distinguished from *Guyanancistrus* by the posterior part of lateral margin of its snout forming a soft fleshy border and bearing slightly enlarged odontodes with small tentacules (; see). It is interesting to note that this organisation of odontodes on sides of snout is reminiscent of the condition observed in *Hopliancistrus*, one of its sister genera. These odontodes grow larger in large specimens of *Hopliancistrus* (even becoming stout at the corner of the snout in males), a condition that might be hypothesized for *Cryptancistrus* in the absence of material apart from the single holotype. Despite the absence of obvious morphological characteristics for *Guyanancistrus*, a unique combination of external characters allows genus recognition with respect to its sister genera, and species assignation within the genus, was found for the diagnosis presented. Armbruster (: 12) suggested that *Guyanancistrus* (which he placed in synonymy of *Pseudancistrus*) was not likely to be a monophyletic entity because of divergence of external characters between species. He particularly cited the development of « at least small hypertrophied odontodes on the snout » of *G*. *niger* (*vs* lack in the others), however these enlarged odontodes are present solely in two lateral tufts on the dorsolateral edges of the upper lip, not on the plates outlining the snout contour as in the case of *Pseudancistrus*. *Hopliancistrus tricornis* also has two lateral regions of the snout with enlarged odontodes, that suggest similarity with *G*. *niger*; again, however, they are not on the lip, but on the dorsolateral margin of the snout. Odontodes are present to a variable extent on the dorsal margin of the upper lip in several Hypostomines, including other *Guyanancistrus* species and *Hopliancistrus tricornis*, but they are usually small. Tufts of enlarged odontodes on the dorsolateral edge of the upper lip are thus characteristic of *G*. *niger*. The degree of evertibility of the cheek plates in *G*. *brevispinis* and *G*. *niger* was also found to be divergent by Armbruster (loc. cit.), but we saw no significant difference in this character between *Guyanancistrus* species. *Chaetostomus megacephalus*, described by Günther in 1868, is an additional but currently insufficiently known species that might also belong to *Guyanancistrus*. Long considered as an *Hemiancistrus* species, it was moved to *Pseudancistrus* by Armbruster, who recently stated that this taxon needs further work, but it is evident that it does not correspond to *Pseudancistrus* as defined by Covain & Fisch-Muller (*= Pseudancistrus barbatus* group of de Chambrier & Montoya-Burgos). Indeed, the holotype has evertible cheek odontodes and no enlarged odontodes along the snout margin despite its large size (122.8 mm SL). Morphologically it is similar to *G*. *longispinis* group members, but is quite distinct from the nominal species. The type locality of *C*. *megacephalus* was indicated as “Surinam” in the original description, but Günther later added that it was obtained from the collection of Dr. van Lidth de Jeude specifying that it was “probably from Surinam”. Unfortunately we can only refer to the holotype. Specimens collected in the Essequibo River Basin in British Guiana described by Eigenmann (: 231) as *Hemiancistrus megacephalus* appear more likely a distinct and probably new species morphologically very close to *G*. *longispinis*. Despite extensive field collecting in Suriname, we have been unable to find any additional specimens (as well as other teams in Guyana; J. W. Armbruster, pers. com.). The case is similar for *Chaetostomus macrops* Lütken, 1874, also known from a single specimen from “*aquis Surinamensibus*”, and considered a synonym of *megacephalus* from the early 20^th century until recently. It has a particularly wide and elevated orbital rim reminiscent of that observed in *Hemiancistrus medians*, from which, however, it is easily distinguished by the absence of keeled and rough-toothed trunk plates and by the presence of odontodes over a broad area on the opercle. Morphologically, *C*. *macrops* is most similar to the species collected in the Potaro River by Eigenmann than to *P*. *megacephalus* and could correspond to a distinct species. The collection of fresh material is essential before making taxonomic decisions concerning these species. ## The *G*. *brevispinis* complex The integrative taxonomy methodology reviewed in Padial et al. was sufficient to congruently discriminate eight species of *Guyanancistrus*, including five new species, and a new genus of the Loricariidae. However, *G*. *brevispinis* could not be significantly distinguished from all other species using the different approaches except phylogeny alone, and its different subspecies could not be clearly delineated regardless of the method employed (morphometry, DNA barcodes, phylogeny, or distribution). In this context, use of the multi-table approach integrating all available information was particularly suitable. This method allowed the evaluation of the amount of common information present in the different datasets and its significance through RV tests, and demonstrated that half of the variation recorded in the different tables was significantly linked. Indeed, the unifying structure provided by the MCOA simultaneously revealed significant covariations between morphometric characteristics, phylogenetic structure, and distributional patterns in all available populations of *G*. *brevispinis*, and clearly highlighted three groups of infraspecific rank. Surprisingly, the presence of two lineages of *G*. *brevispinis* within the Maroni Basin was revealed. One lineage included all populations of the eastern Maroni grouped with other populations of *G*. *b*. *bifax*, and the second was located in the Upper Tapanohony River, a western tributary of the Maroni River, grouped with the populations of *G*. *b*. *brevispinis* close to populations from Upper Suriname and Corantijn rivers. This unexpected result highlights the central role played by the Maroni Basin in the distributional pattern of *G*. *brevispinis* members, with an east-west partition of this drainage. This central role was also confirmed by the biogeographic reconstruction, which resolved no fewer than five successive dispersal events originating from the Maroni Basin toward other drainages of the Eastern Guianas. Two of them concerned ancestral populations of *G*. *brevispinis* which dispersed to the west toward the Upper Suriname and Corantijn rivers respectively (leading to establishment of the future *G*. *b*. *brevispinis*), a third eastward toward the Upper Oyapock (future *G*.*b*. *orientalis*), and a fourth and fifth toward the Sinnamary and Mana respectively (future *G*.*b*. *bifax*), but with the persistence of two distinct lineages within the Maroni Basin. These results only partially corroborate the findings of Cardoso and Montoya- Burgos who recovered five lineages (instead of three) among *G*. *brevispinis* including distinct lineages from: (1) Oyapock-Comté-Approuague basins, (2) Maroni-Mana-Sinnamary basins, (3) Suriname River, (4) Corantijn River, and (5) Nickerie River. However, genetic distances between the Nickerie and Suriname rivers’ representatives in their phylogenetic tree were not markedly greater than those within their Maroni-Mana-Sinnamary lineage (our *G*. *b*. *bifax*), such as the representatives of the Sinnamary and Mana basins. These authors also highlighted the central role played by the Maroni Basin as the gateway of ancestors of *G*. *brevispinis* from the Amazon Basin. Using phylogenetic topological tests, they hypothesized a single entrance from headwaters of the Maroni River followed by a first westward dispersal, assuming a stepping-stone pattern of dispersal. Then dispersal strategies were evaluated based on haplotypic diversity and genetic-geographic structure comparisons between populations here described as *G*. *b*. *bifax*, revealing favoured dispersal routes through coalescing river mouths during low sea level periods. If the entrance of ancestral forms of *Guyanancistrus* originating from the Amazon Basin in the Maroni River is confirmed by the present study, the reconstructed dispersal pattern of *G*. *brevispinis* members is much more complex than a simple stepping-stone process, probably related to river capture events (direct dispersal from the Maroni to the Corantijn and Oyapock rivers). In addition, Cardoso and Montoya-Burgos reported a single Amazonas lineage which has been shown in the present study to contain three distinct species: *G*. *teretirostris* from Paru de Oeste River (Pb.BR652, Pb.BR653, and Pb.BR654 in Cardoso and Montoya-Burgos, 2009), *G*. *tenuis* from Jari River (Pb.MIT03, and Pb.MIT04 in Cardoso and Montoya-Burgos, 2009), and *G*. *megastictus* from Maroni River (Pb.MIT02 in Cardoso and Montoya-Burgos, 2009), leading to two dispersal events from the Amazonian tributaries toward the Maroni Basin for their study. The present study also revealed a third dispersal between Paru de Oeste and Saramacca rivers. All these dispersals resulted in speciation within the Eastern Guianas, with the particularity of the Maroni River hosting three newly formed species; two hyperendemics restricted to montaneous areas (*G*. *nassauensis* and *G*. *megastictus*), and one widely distributed (*G*. *brevispinis*) and comprising two subspecies (*G*. *b*. *brevispinis* and *G*. *b*. *bifax*). The Maroni Basin was thus a center of speciation for *Guyanancistrus* members resulting in increased local endemicity, as well as a source of dispersal to other drainages of the Eastern Guianas. Cardoso and Montoya-Burgos tentavely provided diagnostic characters to distinguish their different lineages, but most of them relyed on global estimates of shape and color patterns. Conversely, the MCOA used here, by unifying different variables contained in different data sets within the same analysis, allowed the extraction of diagnostic characteristics for each group. Moreover, the ability to include phylogeny with the other data sets allowed the interpretation of covariations between morphometric, phylogenetic, and distributional variables in an evolutionary perspective. Evolution of shape of *G*. *brevispinis* members was thereby linked to genetic and geographic divergences. *Guyanancistrus brevispinis bifax* evolved a mean shape intermediary between *G*. *b*. *brevispinis* to the west and *G*. *b*. *orientalis* to the east, characterized by more numerous plates on the caudal peduncle, which was also less deep in this subspecies. This result contrasted with that of Cardoso and Montoya-Burgos, who characterized the same group as having the highest body shape. *Guyanancistrus b*. *brevispinis* evolved a broader head and anterior body in the west whereas *G*. *b*. *orientalis* evolved a slender appearance with a longer caudal peduncle and more numerous plates along the body in the east, a result in general agreement with Cardoso and Montoya-Burgos. It is interesting to note that *G*. *b*. *bifax* possesses two morphs, both present in the whole area of distribution of the subspecies. Even though not statistically supported, morphotypes with broad mouth (identified by the letters BM) were all placed closer to *G*. *nassauensis* in the morphometric analysis, and appeared clearly distinct from other morphotypes with a normal mouth. Given that *G*. *nassauensis* was introgressed by *G*. *brevispinis* , implying hybridization between these two co-occuring (at least in the Nassau Mountains) sister species, this morphological characteristic may result from retention of genes from *G*. *nassauensis* in the genome of *G*. *b*. *bifax*. Color patterns also appeared highly variable in *G*. *brevispinis*. Polychromatism is not rare in fish and can be related to sexual selection driving the appearance of strong sexual dimorphism (\[–\], reviewed in). Numerous genera of the Loricariidae exibit strong sexual dimorphism through the development of hypertrophied odontodes (e.g. in *Peckoltia*, *Panaque*, *Neblinichthys*, *Sturisoma*, *Farlowella*, *Spatuloricaria*, *Rineloricaria*), development of fleshy tentacles on snout (e.g. in *Ancistrus*), lip enlargement (e.g. in *Loricariichthys*, *Hemiodontichthys*), or teeth characteristics (e.g. in *Loricaria*), but dichromatism has not previously been reported, even though several genera display colorful patterns (e.g. *Pseudacanthicus*, *Leporacanthicus*, *Hypancistrus*, *Scobinancistrus*, *Peckoltia*, *Panaqolus*). Color variations in Loricariidae appear sex-independent, and rather related to natural selection (e.g. for camouflage over the substrate) or to random drift. However, such variations in *G*. *brevispinis*, with the appearance of very diverse patterns ranging from spots to marbling and reticulations imply rather relaxed selective constraints acting on phenotypes since different patterns can be observed within the same basin. Alternatively similar patterns can be observed between distant basins implying multiple convergent evolution and/or retention of ancestral patterns among populations. Cardoso and Montoya-Burgos tentatively classified their different lineages using pattern characteristics, but if this criterion applied for a given population, it often failed to characterize other populations of the same basin or equally applied to other populations of a distinct lineage. For example, Cardoso and Montoya-Burgos distinguished their lineage from the Corantijn River (Sipaliwini River;), from all other populations by the head having small light vermiform marks and the body faint parallel light bands becoming highly visible on the caudal peduncle. However, corresponding to another population from the Corantijn Basin (Kabalebo River), shows that this population was particularly dark, without such obvious markings. The same situation occurred with the Comté-Approuague-Oyapock lineage (*G*. *b*. *orientalis*), which was supposed to be distinguished from all other lineages by the head having small light dots and the body faint, parallel, light bands becoming highly visible on the caudal peduncle; the specimen from the Comté-Orapu Basin and the one from the Oyapock River, are clearly distinct from each other. Moreover, the characteristics of the Oyapock population better reflected the definition provided for the Maroni-Mana-Sinnamary lineage (*G*. *b*. *bifax*), theoretically distinguished by large light dots, irregular in shape. Given the high variability of the species, color patterns do not appear to be relevant for identification purposes. ## Notes on the ecology of *Guyanancistrus* species *Guyanancistrus brevispinis* occurs in the lowland rivers and large tributaries in the interior of Suriname and French Guiana (i.e. upstream of the most downstream rapids), mainly in strong currents in or immediately downstream of rapids. During the day adult *G*. *brevispinis* were observed foraging on the algal biofilm on boulders and bedrock in the Middle Suriname River together with *Cteniloricaria platystoma* and *Harttia surinamensis*. Adult *G*. *brevispinis* are well camouflaged when feeding on these substrates. Juveniles of *G*. *brevispinis* were collected in a mountain stream (400 m above mean sea level) in Lely Mountains with cool (23.3°C), clear (Secchi disc visibility 150 cm) water with low conductivity (24 μS cm^-1) and neutral pH of 7.5. Postlarvae of *G*. *brevispinis* (15 mm TL) were collected in a headwater tributary of the Upper Palumeu River in a deep (\> 1 m) pool under a 60-m high waterfall (Fig 8.1. in Mol & Wan Tong You); the water was cool (23.5°C), clear (turbidity 5 NTU), slightly acidic (pH 5.9) with low conductivity (20 μS cm^-1), low alkalinity (4.75 mg CaCO₃ L^-1) and some tannins (2.6 mg L^-1). *G*. *brevispinis* has the largest distribution within Eastern Guianas, with an area of distribution ranging from the Corantijn River in western Suriname to the Oyapock River in eastern French Guiana. Alternatively, most of the other Guianese species (the Amazonian species are insufficiently known) appear highly restricted to mountains, having a similar distributional pattern to *Harttiella*, a group of hyperendemic dwarf loricariids restricted to mountainous forest creeks. At least two species of *Guyanancistrus* (*G*. *nassauensis* and *G*. *brownsbergensis*) have developed adaptations to this kind of biotope (small streams, cool water temperature, low productivity…) including dwarfism. *Guyanancistrus nassauensis* and *G*. *brownsbergensis* are each known from a single mountain stream, in the Nassau Mountains (Paramaka Creek) and Brownsberg Mountains (Kumbu Creek), respectively. With this very restricted distribution (\< 20x20 km²) both species can be considered hyperendemics and currently the two species are threatened with extinction by proposed and ongoing mining activities. In Paramaka Creek, *Guyanancistrus nassauensis* occurs syntopically with juvenile *Guyanancistrus brevispinis* and with *Harttiella crassicauda*, a second endemic species from the Nassau Mountains. However, *G*. *nassauensis* occurs both on the plateau in perennial flowing headwaters and in the upper mainstem of Paramaka Creek (lower slopes of the plateau; altitude range 120–530 m amsl), whereas *H*. *crassicauda* only occurs on the plateau proper (230–530 m amsl;). In the IJs Creek tributary of Paramaka Creek on the Nassau plateau (467 m amsl) both *G*. *nassauensis* and *H*. *crassicauda* occur in cool (22.6°C), shallow (40 cm water depth), clear (Secchi transparency \> 40 cm) water with low conductivity (28 μS cm^-1), neutral pH of 7, low inorganic N (0.067–0.120 mg L^-1), relatively high organic N (0.307–0.592 mg L^-1), low total P (0.002–0.010 mg L^-1) and high organic C (2.916–4.972 mg L^-1). The bottom substrate is gravel with boulders and bedrock (with the red filamentous algae *Batrachospermum* sp. attached to it) and near the edge of the plateau in slightly deeper water (approximately 50 cm) stands of the emergent macrophyte *Thurnia sphaerocephala* occur. In the upper mainstem of Paramaka Creek, as well as in some upstream branches on the plateau, *G*. *nassauensis* occurs syntopically with *G*. *brevispinis*. *Guyanancistrus brownsbergensis* was collected only in the upper reaches of Kumbu Creek on the Brownsberg bauxite plateau at an altitude of 200–430 m amsl. The habitat of *G*. *brownsbergensis* seems very similar to that of *G*. *nassauensis* in the Upper Paramaka Creek on the Nassau bauxite plateau. The upper Kumbu Creek is a small (2.5–3.7 m wide, 26–52 cm deep) mountain stream with moderate to strong flow (0.3–0.56 m s^-1) and cool (23.1–23.2°C), mostly clear water with high dissolved oxygen content (7.1–7.7. mg L^-1), neutral pH of 7–7.5 and low electrolyte content (30.8–31.6 μS cm^-1). The bottom substrate is mainly gravel, boulders and bedrock. We observed no aquatic vegetation in the stream, but overhanging terrestrial vegetation, submersed root masses, woody debris, leaf litter and rock crevices offered ample hiding places for *G*. *brownsbergensis*. During the day, adult *G*. *brownsbergensis* were observed on several occasions throughout the year resting in moderate current in front of a rock crevice in a relatively deep (50 cm) pool upstream of the 50-m high Kumbu Falls. We have no information on the ecology of *G*. *teretirostris*, *C*. *similis*, *G*. *tenuis* and *G*. *megastictus*, although the latter two species apparently also occur in small mountain streams, perhaps comparable to the habitat of *G*. *nassauensis* and *G*. *brownsbergensis*. Other species are larger, particularly when they inhabit the main stream of rivers as does *G*. *niger*, the largest species of *Guyanancistrus*, which lives in the rapids of the Oyapock River along with *Pseudancistrus barbatus*. *Guyanancistrus niger* is much less abundant than other species, and we collected only adult specimens, suggesting that adults and juveniles may not be syntopic. *Guyanancistrus brevispinis* and *G*. *longispinis* were collected together in the Oyapock drainage, but while *G*. *brevispinis* seems to prefer small forest streams, *G*. *longispinis* was more often found in the main channel, on the rocky bottom of riffles, where it can be relatively abundant (; personal observations). # Supporting information We are grateful to Kenneth Wan Tong You (NZCS), Regis Vigouroux and Philippe Cerdan (Hydreco), Pierre-Yves Le Bail (INRA), Michel Jégu (IRD), Philippe Gaucher (CNRS), Raphaelle Rinaldo and Guillaume Longin (PAG), Sophie de Chambrier, Alexandre Lemopoulos, Pedro Hollanda Carvalho, and Claude Weber (MHNG), Philippe Keith and François Meunier (MNHN), Juan Montoya-Burgos (University of Geneva), and Gregory Quartarollo (Guyane Wild Fish), for logistic assistance, support, and help during different field collects. Philippe Keith, Pierre-Yves Le Bail, Philippe Gaucher, Gregory Quartarollo, and Regis Vigouroux are also acknowledged for gift of material they collected, as well as Frédéric Melki (Biotope), Dominique Ponton (IRD), Antoine Baglan (Guyane Wild Fish), and Trio Amerindians of the Sipaliwini village. Frédéric Melki also provided a picture of a live *G*. *megastictus* to illustrate the species. The French Guiana DEAL, PAG, and Préfecture; and the Surinamese Ministry of Agriculture, Animal Husbandry and Fisheries provided the necessary authorizations and collecting permits. We greatly thank for his welcome in collection Dirk Neumann (ZSM), and for loans of specimens Jonathan Armbruster (AUM), James Maclaine (BMNH), David Catania (MCZ), Romain Causse and Patrice Pruvost (MNHN), Ronald de Ruiter (RMNH), Hielke Praagman (ZMA), Peter Bartsch (ZMB), Rainer Stawikowski and Uwe Werner. In MHNG, Alain Merguin provided significant logistical help, Philippe Wagneur photographed the type specimens, and John Hollier improved English usage and style of the manuscript. Nathan K. Lujan is acknowledged for useful comments on the manuscript as well as Jon W. Armbruster who also generously provided a picture of a live *G*. *nassauensis* to illustrate the species in the present contribution. [^1]: The authors have declared that no competing interests exist.

# Introduction The early stages of a new, romantic relationship can be a powerful and absorbing experience. Individuals in new romantic relationships report feeling euphoric and energetic. They also become emotionally dependent on, desire closeness with, and have highly focused attention on their partner. Human neuroimaging studies have shown that feelings experienced during the early stages of a romantic relationship are associated with neural activations in several reward-system and affect-processing regions of the brain. Those studies displayed pictures of participants' own romantic partners to reliably evoke acute positive affect and self-reported feelings of love. In one such functional magnetic resonance imaging (fMRI) study, Aron and colleagues instructed participants in new, romantic relationships to view pictures of their partner, and pictures of a familiar acquaintance who was the same age and sex as the participant's partner. Neural activations specific to viewing pictures of the romantic partner were observed in several reward-processing regions, such as the bilateral caudate nucleus and right ventral tegmental area. An earlier fMRI study using a similar protocol reported neural activations specific to the romantic partner pictures in reward regions such as the bilateral caudate nucleus and bilateral hippocampus. The activation of reward structures caused by viewing pictures of a romantic partner has also been confirmed in a Chinese sample, suggesting the phenomenon may be culturally universal. Collectively, these neuroimaging studies demonstrate that reward-system activation is a central component of self- reported feelings of love in new romantic relationships. The engagement of reward systems by viewing pictures of a romantic partner is pertinent to the study of pain because several basic animal studies have shown reward-processing regions to be critically involved in analgesia. For example, the nucleus accumbens and ventral tegmental area (two key reward processing structures) both play an important role in analgesic processes – – perhaps explaining why pleasurable and appetitive states such as sucrose consumption and anticipation of food reward reduce pain. The results of those studies suggest that the activation of reward systems (perhaps even non-pharmacologically) could reduce pain in humans. Indeed, a recent behavioral study demonstrated that the presentation of romantic partner pictures is sufficient to reduce experimentally-induced pain. The partner pictures reduced pain significantly more than when participants viewed pictures of a stranger or affect-neutral object, and the analgesic benefit was as strong as holding the partner's hand. The study was important in that it showed a mere representation of a romantic partner can reduce pain; however, the study was not designed to characterize the central mechanisms related to reward-induced analgesia. Given recent behavioral results suggesting an analgesic benefit of rewarding experiences, and neuroimaging data showing those experiences to produce reward system activation, we hypothesized that analgesia during evoked feelings of love would be associated with reward system activation. Using fMRI of the human brain, we investigated the neurophysiologic substrates of analgesia produced by viewing pictures of a romantic partner. # Materials and Methods ## Participants Participants were 15 right-handed students (8 women and 7 men, age range 19–21 years, M = 20 years) in their first 9 months of a romantic relationship. All participants described themselves as intensely in love, and scored a minimum sum of 90 on the 9-point scale, 15-item short form of the Passionate Love Scale (PLS). ## Experimental paradigm All study procedures were approved by the Institutional Review Board at the Stanford University School of Medicine, and all participants provided written informed consent. Before arriving for the scan session, each participant provided three digital pictures of his or her romantic partner, and three pictures of an acquaintance of the same gender and attractiveness as the romantic partner. We sought to balance partner and acquaintance attractiveness because previous research has shown attractiveness to be associated with neural activations in reward areas. Acquaintances were also individuals who the participants had known for approximately the same length of time as their partner and for whom the participant reported no romantic feelings. The attractiveness of both the romantic partners and acquaintances were rated independently (on a 0–10 numerical scale) by eight individuals who were blinded to the relationship type and who were not otherwise involved with the study. There was no difference in attractiveness of the acquaintances versus partners (*t*(28) = −0.35, *p* = 0.73). All pictures were cropped to display only the face. At each participant's scan session, we first determined what temperature would produce moderate and high levels of pain. To determine temperatures used for moderate-pain and high-pain heat, participants were exposed to 15-second heat blocks, starting at 40 degrees Celsius (a non-painful temperature). Following each heat block, participants were asked to report the degree of pain on an 11-point visual scale (0 = no pain at all, 10 = worst pain imaginable). Each successive heat block was increased by 1 degree Celsius, until the participant reached their 10/10 (maximum) pain score. Then, the lower temperatures were retested to verify what temperatures elicited 4/10 (moderate) and 7/10 (high) pain. We presented thermal stimuli with a Medoc (Durham, NC) Advanced Thermal Stimulator 3×3 cm Peltier contact thermode. When in the scanner, the thermode was attached to the thenar eminence of the left hand, so that the right hand could be free for inputting pain ratings via a button box. The ramp rate for all trials (in and out of the scanner) was 10°C per second, requiring a maximum of 1.5 seconds to reach the target temperature. While in the scanner, participants performed three distinct tasks: an acquaintance baseline condition, a romantic partner active condition, and a distraction control condition. In the *acquaintance* baseline condition, each participant was shown the pictures of his or her acquaintance via a projector and mirror display mounted on the head coil. Following the protocol of Aron and colleagues, participants were asked to focus on the picture and think about the displayed person. The use of the active baseline condition allowed us to separate neural activity specific to viewing pictures of a romantic partner from those of simply looking at an equally attractive and familiar face. In the *romantic partner* condition, participants viewed pictures of their partner, and were asked to focus on the picture and think about the person. Participants also underwent a *distraction* control condition. During the distraction trials, participants were asked to complete a word-association task that had been shown to effectively reduce pain in previous fMRI studies. The distraction control condition allowed us to determine whether or not the partner pictures were simply serving as a salient distractor from pain. In the distraction trials, a seed phrase was displayed (e.g., “Sports that do not use a ball”), and participants were instructed to silently think of as many responses as possible. The distraction task demands a high degree of attention, requires no movement, and is relatively free of an emotional component. Each of the three conditions described above was performed under periods of no pain, moderate pain, and high pain. Participants were told that a range of temperatures would be presented, and were not told that only three discreet temperatures would be administered. All of the “no pain” presentations were given at a baseline temperature of 32 degrees Celsius. Moderate- and high-pain presentations used each individual's 4/10 and 7/10 temperature levels that were determined before the scan. Each condition (partner, acquaintance, and distraction) by pain (none, moderate, and high) combination was repeated 6 times, for a total of 54 randomly ordered trials. Following each trial, the participant rated his or her evoked pain, using the equipped button box and a projected visual analog scale. Pain ratings were collected immediately following (rather than during) the pain stimulus, so that task performance would be minimally affected by sensorimotor processing associated with rating pain on the response box. Following the pain rating, participants completed a mental arithmetic count-back task for 13 seconds. This task (adapted from Aron and colleagues) was designed to minimize emotional and sensory carryover between trials. In the task, participants were visually presented a 4-digit number, and were asked to count backwards by 7's as quickly and accurately as possible. The task was also part of the study manipulation, as participants were told their performance on the task was a central component of the experiment. The time course of each trial was as follows: trial ready cue (2 sec), acquaintance, partner, or distraction task with thermal stimulus (16 sec), pain rating (10 sec), and count-back (13 sec). ## Behavioral analysis and statistics A single univariate ANOVA was performed to determine the effects of the partner and distraction tasks on self-reported pain. Two independent variables were entered into the model: heat pain level (no pain, moderate pain, and high pain), and condition (acquaintance, partner, and distraction). Pain ratings were entered as the dependent variable. Pain ratings were averaged over trial repetitions to yield a single pain score per subject, for each temperature by condition combination – with subject treated as a random effect. After scanning was finished, participants completed a brief outtake form. To assess for possible demand characteristics, participants were asked, “What do you think was the purpose of this experiment?” Responses were evaluated to determine the number of participants who correctly determined the purpose of the study. ## fMRI data acquisition We conducted scans at the Stanford University Lucas Center, using a 3T GE Signa system and 8-channel head coil. Functional blood oxygen level-dependent (BOLD) data were acquired using a T2\*-sensitive spiral in/out pulse sequence. Functional volumes consisted of 28 oblique (parallel to the AC-PC plane) slices covering the brain and brainstem (4 mm thickness, 0.5 mm gap, in-plane resolution 3.125×3.125 mm, repetition time = 2 sec, TE = 30 ms, flip angle = 90°, field of view = 20×20 cm). High-order shimming was performed before the functional scans. A T1-weighted 3D-IR-FSPGR scan was acquired for anatomical reference (TE = 1.7ms, 124 slices, 1.2mm isotropic resolution). ## fMRI processing and analysis Functional images were first corrected for cardiac and respiratory noise , and then realigned, resliced, and smoothed by 6mm, using SPM8 (Wellcome Department of Imaging Neuroscience, London). First-level statistics were performed on an individual level in native space. Condition-specific effects were estimated using a general linear model (GLM) approach. Conditions were described with a boxcar design and then convolved with the canonical hemodynamic response function. All phases of the scan protocol (cue, task, pain rating, and countback) were modeled, though only the task periods were used in contrasts. Statistical results maps for all planned contrasts were coregistered with the high-resolution structural images, normalized to MNI space, and resampled at a 1 mm isotropic voxel size using the DARTEL toolbox in SPM8. The spatially- normalized contrast maps were then used to conduct second-level (group) statistics with participant as a random effect. A grey matter voxel mask was applied to all second-level contrast maps. Two major contrast analyses were performed. The first contrast identified neural activations and deactivations associated with viewing pictures of a romantic partner, while in pain, and controlling for both the effects of viewing pictures of an equally attractive acquaintance, and performing a distraction task. A conjunction analysis approach was used, which requires all identified voxels to demonstrate greater increase or decrease in activity compared to both the acquaintance and distraction tasks. The conservative approach (using the “conjunction” option in SPM8) ensured that all identified clusters were significantly different from both control conditions. The second contrast identified neural regions associated with analgesia resulting from viewing pictures of a romantic partner, distinct from distraction analgesia. Analgesia (pain reduction) during the partner task was determined by subtracting pain ratings in the partner trials from pain ratings in the baseline acquaintance trials. Degree of analgesia was also calculated for the distraction task, by subtracting pain in distraction trials from pain in the acquaintance trials. To determine the neural responses specific to analgesia caused by viewing pictures of a romantic partner, analgesia was entered as a covariate in the second-level analysis, yielding a contrast map of all BOLD increases and decreases significantly associated with pain relief. The contrast map also masked out any significant BOLD responses associated with distraction analgesia, to identify only those analgesia responses specific to viewing pictures of a romantic partner. By reversing the distraction-analgesia mask, a separate map was also created to show analgesia-associated BOLD responses occurring in both the partner and distraction tasks. All statistics were performed on the whole brain, with no region-of-interest or small-volume corrections. The group-level (mixed effects) contrast images were thresholded with a voxel-height significance threshold of *p*\<.005 (uncorrected), requiring a *t*-value of 3.01. An FDR-corrected spatial extent threshold was not employed, because several reward and pain-modulatory nuclei have total structural volumes that are below FDR-corrected thresholds for the whole brain. For example, the ventral tegmental area has a volume below the FDR- corrected, *p*\<.05 spatial threshold of approximately 271 mm³. Instead of an FDR correction, a spatial extent threshold of 64 contiguous voxels (a 64 mm³ region) was used to allow smaller regions to emerge in a whole-brain analysis. The use of combined height and spatial-extent thresholds in this manner has been demonstrated to provide a good balance between risk of Type I and Type II error. However, even with the extent threshold we used, it is still possible that smaller reward structures (or specific regions of reward structures) would be too small to be identified. # Results ## Behavioral All participants scored highly on the self-reported scale of passionate love (mean = 109.8, SD = 11.2, range = 91.5–132.0), meeting the minimum required sum score of 90. presents pain ratings for the three tasks (acquaintance picture baseline, distraction control, and partner picture task), and three temperature levels (no pain, moderate-pain, and high-pain). The thresholding procedure was effective at determining temperatures to elicit 4/10 (moderate) and 7/10 (high) pain in the baseline condition. Averaging across all acquaintance baseline trials, pain during the moderate-intensity trials was rated at 3.7, and pain during the high- intensity trials was rated at 7.0. A univariate ANOVA of pain ratings (with task and temperature as predictors) was performed. The two-predictor model strongly predicted pain ratings (adjusted R squared = 0.81). Temperature was a significant predictor of pain (*F*(2,14) = 276.95, *p*\<0.0001). Bonferroni-adjusted post-hoc tests revealed that reported pain was significantly different between all levels of heat intensity (all pairwise contrast *p*'s\<0.0001). Task was also a significant predictor of pain (*F*(2,14) = 4.72, *p* = 0.011). Bonferroni-adjusted post-hoc tests showed that pain was significantly reduced in both the distraction and partner conditions, contrasted with the acquaintance baseline condition (*p*'s = 0.026). There was no difference in pain between the distraction and partner conditions (*p* = 1.0). The “temperature by condition” interaction was not significant (*F*(4,14) = 1.07, *p* = 0.372). Because we were interested in neural mechanisms that generalize across pain levels, moderate- and high- intensity trials were aggregated for all neuroimaging analyses. To determine the possible role of demand characteristics on self-reported pain, responses to the manipulation check were examined. A response was counted as a correct guess if the participant identified pain as the dependent variable (e.g., “to see how emotions affect pain”). Six out of the fifteen participants correctly guessed the purpose of the study. ## fMRI – BOLD responses during presentation of romantic partner pictures while in pain The first group contrast identified the main effects of the partner task on neural responses during periods of moderate- and high-intensity pain. Identified clusters represent BOLD signal changes associated with the partner task that occurred over and above both the acquaintance and distraction tasks. The conjunction analysis revealed several regional areas of BOLD activity increase and decrease associated with viewing pictures of a romantic partner during pain. The largest cluster of activation was found in the bilateral frontal cortex, projecting out from the pregenual anterior cingulate cortex, and extending into the medial orbitofrontal cortex. Separate clusters of activation were observed in the subgenual anterior cingulate (BA 25) and mid-cingulate (BA 23) cortices, as well as the left precuneus (BA 31;), left amygdala, and right hypothalamus. Several BOLD activity decreases were also associated with viewing pictures of a romantic partner. Signal decreases were seen in the bilateral posterior insula, left thalamus (ventral lateral nucleus;), left inferior frontal cortex, right frontopolar area (BA 10), left supplementary motor area, and right precentral gyrus (BA 6). ## BOLD responses associated with pain relief while viewing pictures of a romantic partner The second group contrast assessed neural activity changes associated with analgesia produced by the romantic partner picture task. BOLD activity increases and decreases significantly correlated with pain relief during the partner trials were identified, masking out any regions also associated with pain relief in the distraction task. BOLD activity in a number of regions was positively associated with pain relief in the partner task, including the bilateral caudate head, bilateral nucleus accumbens, right dorsolateral prefrontal cortex, right superior temporal gyrus, bilateral lateral orbitofrontal cortex , left amygdala, and right thalamus (ventral anterior nucleus). Greater pain relief was associated with decreased BOLD activity in the right superior frontal gyrus, left dorsal anterior cingulate cortex , right brainstem (located in the substantia nigra and red nucleus region;), left anterior insula, right putamen, left supplementary motor area, and left parahippocampal area. To determine if any BOLD responses were associated with analgesia in *both* the partner and distraction conditions, the distraction mask was reversed. Only one area of overlap was identified: the right lateral orbitofrontal cortex (green cluster). Distraction analgesia was instead associated with activations in the left rostral anterior cingulate, left medial frontal gyrus, left middle frontal gyrus, bilateral putamen, left superior parietal cortex, right anterior cingulate cortex, left dorsolateral prefrontal cortex, left Broca's area, left orbitofrontal cortex (BA 47), and bilateral orbitofrontal cortex (BA 11). # Discussion In this study, we demonstrate that pain relief experienced while viewing pictures of a romantic partner is associated with reward system activation. We further show that the neural processes associated with reward-induced analgesia are distinct from those associated with distraction-induced analgesia. Our first goal was to determine if viewing pictures of a romantic partner activates reward and limbic regions of the brain, even during periods of moderate- and high-intensity pain. We found a main effect for the romantic partner task on BOLD response in several reward and limbic regions. Viewing pictures of a romantic partner activated reward processing areas such as the amygdala (stimulus reward value learning), hypothalamus (reward-stimulus associations), pregenual anterior cingulate cortex (reward-related cognition), and medial orbitofrontal cortex (hedonic experience processing). Several additional limbic regions, such as the precuneus, mid-cingulate, and subgenual anterior cingulate cortex were also associated with the viewing of romantic pictures. Because the contrast map controlled for both: 1) viewing pictures of an equally attractive and familiar acquaintance, and 2) distraction, the observed activations were likely specific to feelings evoked by the pictures of a romantic partner. BOLD activity decreases were also observed with the romantic pictures task, and were localized mostly in pain-processing regions. Activity was suppressed in both the left and right posterior insula, areas responsible for sensory processing of pain. BOLD activity was suppressed in the left ventral lateral nucleus of the thalamus, suggesting early-stage suppression of nociceptive signals. Activity in the stimulus-contralateral premotor cortex was also suppressed, suggesting further reduction of pain processing. Not all regions showing depressed BOLD activity were associated with classic pain-processing regions. Regions showing an activity decrease, but not typically connected to the pain experience, included the left inferior frontal cortex, and right frontopolar area. As suggested by previous behavioral research, viewing pictures of a romantic partner effectively reduced self-reported pain. We found that BOLD increases in several reward system regions were associated with greater pain relief during the partner task. Associations between pain relief and activations in the bilateral caudate head and nucleus accumbens were seen, identifying aspects of both the classic mesolimbic and nigrostriatal reward pathways. Furthermore, regions known to modulate reward processing, such as the amygdala, lateral orbitofrontal cortex, and dorsolateral cortex, were associated with pain relief. Many of those regions are commonly identified in reward-related neuroimaging tasks and make up a corticostriatal network of reward processing. The corticostriatal reward network provides one route by which strongly valenced external cues may reduce the experience of pain. Cue-evoked reward prediction is known to involve orbitofrontal and dorsolateral prefrontal cortex activity in humans, and both of those regions are strongly connected to reward-based evaluation and decision-making. In the case where pictures of a beloved may serve as a reward-related cue, activity in the orbitofrontal cortex or dorsolateral prefrontal cortex would modulate activity in mesolimbic and nigrostriatal reward systems via pronounced projections to the nucleus accumbens, and caudate head –. Analgesia may then result from reward system projections to descending pain modulatory systems, which can inhibit ascending nociceptive messages at the spinal level. The hypothesis that reward activation may suppress nociceptive processing at an early supraspinal or spinal stage is further supported by the wide range of pain processing regions that exhibited analgesia-correlated activity decrease. Pain relief produced by viewing pictures of a romantic partner was associated with suppressed activity in sensory (anterior insula and brainstem), affective (putamen, hippocampus, and anterior cingulate), and cognitive (supplementary motor area, superior frontal gyrus) aspects of the neural pain response network. The reward-system activity we observed to be associated with pain relief during viewing of romantic partners was unlikely to be a general effect of analgesia, as an equal amount of pain relief evoked by a distraction task showed little engagement of reward systems. Distraction analgesia was associated with mainly cortical activations, and in many regions previously associated with the distraction task used. The results suggest that there are multiple routes by which cognitive tasks can reduce pain, with emotion-based and distraction-based analgesia being two such possibilities. However, it is also true that the experience of pain relief itself can also serve as a rewarding experience. The nucleus accumbens and amygdala can be activated during various experiences of pain relief. Placebo analgesia in particular has been associated with increased opioid transmission in a range of reward systems, including the orbitofrontal and dorsolateral prefrontal cortices, nucleus accumbens, and amygdala. In the present study, pain relief resulting from both the partner pictures and distraction task activated an overlapping region of the right lateral orbitofrontal cortex. Therefore, the two types of analgesia were found to have largely separate, but somewhat overlapping, neural substrates. Reward system activation may be one way in which analgesia systems are engaged, and neural overlap between the two systems is strong. The relationship between reward processing and pain relief is supported by prior research showing analgesic benefits from pharmacologic manipulation of key reward systems. It is possible that the analgesic benefit of a rewarding state confers a particular evolutionary advantage on organisms, including humans. The reduction of physical pain during the pursuit of a rewarding stimulus may allow individuals to pursue important goals even in the face of noxious and punishing stimuli, allowing an appetitive state to reduce the influence of aversive states on behaviors. Such inhibitory interactions between appetitive and aversive stimuli have long been reported in the psychological literature. The engagement of reward systems provides one neurobiological route potentially underlying a number of recent findings, such as the presence or likeness of a partner reducing threat response and pain. While we have focused our discussion on reward systems, it is certainly true that the experience of viewing pictures of a beloved involves complex motivational, evaluative, memory, and other processes. Viewing pictures of a romantic partner is likely to be a more active process than the simple, passive experience of a rewarding state. While we attempted to control for active processes with the distraction task, it may not be possible to delineate and experimentally control all the aspects of the partner task. Several of the regions activated by the love task have been associated with other processes in fMRI tasks. The precuneus, for example, has been linked to episodic memory, perhaps indicating activation of memories linked to the picture. The region of the mid-cingulate we identified has been associated with visuospatial attention. And, several of the regions identified (orbitofrontal cortex, hypothalamus, and amygdala) have also been observed during sexual arousal, especially in males. Some methodological issues limit interpretation of the results. First, while participants retrospectively reported high attention to the tasks, there was no objective measure of task adherence or performance. Individuals paying close attention to the tasks would likely experience greater analgesia. It has been previously demonstrated that the suppression of neural pain processing activity is dependent on behaviorally-measured task performance. Future studies may include a measure of task-attention (e.g., eye tracking). A second limitation is that the small sample size precluded the analysis of gender differences in the romantic partner analgesia effect. Third, demands characteristics could play a role in the observed analgesic responses. Six out of the fifteen participants correctly guessed the purpose of the experiment. If those individuals determined the purpose of the experiment early in the session, their self-reported pain may have been affected. Fourth, despite our attempts to control for attractiveness, it is likely that participants found their partner more attractive than their acquaintance. Some of the reward-processing regions we identified were also reported in a previous study examining BOLD response to attractive faces. Surprisingly, we found no regions that showed *both* a main effect for the love task *and* a correlation with degree of analgesia. It is therefore not possible for us to determine what structure or system is critical for love-induced analgesia. The results also demonstrate that there is considerable individual variability in the analgesia experienced when looking at pictures of a beloved. The observed variability could be due to attention to the task, or could be a feature of the relationship (e.g., degree of obsession with the partner, or strength of the relationship). Considerable advances in our understanding of pain and analgesia have been made in recent years, fueled to a great extent by emerging neuroimaging technologies. We show here that the activation of reward systems by viewing pictures of one's romantic partner is associated with reduced pain. A better understanding of these analgesic pathways may allow us to identify new targets and methods for producing effective pain relief. # Supporting Information We thank Vanisha Gandhi and Xu Cui for their invaluable assistance with the study. [^1]: Conceived and designed the experiments: JWY AA SM. Performed the experiments: JWY SP. Analyzed the data: JWY NC. Wrote the paper: JWY AA SP NC SM. [^2]: The authors have declared that no competing interests exist.

# Introduction Working memory is a key cognitive ability that helps us navigate the world around us. We use working memory to temporarily hold information in an active state and to manipulate it in aid of our current goals. Because of its importance for intelligent behaviors, there has great interest in improving working memory through “training” manipulations where working memory demands become progressively more difficult over time. Even without an adaptive training design, working memory performance greatly improves over time with simple practice. Here, we tested whether behavioral feedback aimed at reducing failures of visual working memory performance could augment simple practice benefits, and whether practice-related working memory benefits might lead to improvement on other cognitive tasks. What are working memory failures, and why target them with feedback? Using a whole report working memory task, we can measure trial-by-trial fluctuations in working memory performance and identify failures. In this task, participants briefly view an array of colored squares, remember them for a short duration, and then are required to report all items from the array. Accuracy for each trial is scored as the number of correctly reported items. By holding memory load constant at a difficult set-size (e.g. 6 items), we can measure endogenous fluctuations in performance. In this task, most participants have a mode of 3 items correct (and model fits are consistent with a mostly common capacity of 3 items), but they differ in how frequently they fail to reach this typical capacity limit. Surprisingly frequently (\~12% of trials), participants perform no better than chance (0 or 1 correct). Other studies using change detection similarly have found that participants had lapses for a similar proportion of trials. Mind wandering and attentional lapses occur frequently in everyday life, and lapses of attention influence all sorts of cognitive tasks, including working memory tasks. As such, reducing failures represents a potentially fruitful way to improve cognitive performance across a wide variety of domains. This is in part because of the simplicity of the goal. Rather than trying to boost maximal performance, we can simply try to eliminate abject failures. The current study is related to, but distinct from, the broader “working memory training” literature. First, we focus specifically on practice benefits to a visual working memory task (which involves storage of color-space pairings). Some work has characterized training of visual working memory, but the vast majority of the working memory training literature has employed either the dual n-back task (which contains auditory and visual information and involves updating over time) or complex span tasks (which may contain verbal, numerical, or spatial information, and involve both “storage” and “processing” demands). Second, our experimental design compares working memory gains with practice alone versus practice with performance feedback. Here, we held task difficulty (set size) constant throughout practice. In contrast, the majority of the literature focuses on how well “adaptive” training designs may improve working memory performance (and whether these performance gains transfer to other domains). Previous work has shown that visual working memory performance can be improved by giving participants trial-by-trial feedback about accuracy. The most effective working memory feedback focused participants on reducing failures rather than attempting to store more items. In this feedback design, points were awarded for trials with at least 3 items correct, and points were subtracted for trials that did not reach this goal. Thus, this feedback structure used a weighted combination of both positive and negative feedback, and this weighted combination was more effective than positive feedback alone (e.g. +1 point per item). With weighted feedback, participants’ performance dramatically improved in blocks with feedback compared to without feedback. The improvement to working memory with weighted feedback is likely due to a combination of improved awareness of failures, motivation to reduce these failures, and changes to task strategy (for further discussion see). Unfortunately, this improvement disappeared shortly after feedback was taken away. In a within-subjects design, participants who received feedback in the first half of the experiment experienced more failures immediately after the trial-by-trial feedback was taken away. Brief exposure to feedback may have been insufficient to produce lasting benefits on working memory performance. Despite the relatively high frequency of working memory failures, participants are relatively unaware of them. Previously, we found that participants had poor meta-awareness of working memory failures, catching them only around 30% of the time. Thus, we hypothesized that feedback may be effective in improving working memory performance because it helps to counter-act deficiencies in metaknowledge. Extensive exposure to external feedback may be necessary to bring about improvements in metaknowledge. To test the potential for extensive feedback to provide lasting benefits to working memory performance, we had two groups of participants complete 6 practice sessions of a working memory task. One group completed the practice sessions with trial-by-trial feedback, the other group received no feedback. We also included an active control group (crossword puzzle practice) and a passive control group (no-contact) to measure baseline changes in working memory performance over time. We predicted that practicing with feedback would lead to stronger improvements in working memory performance, and that it would also lead to improvements in participant’s meta-awareness of working memory failures. We found robust effects of practice on working memory performance that were modestly augmented by the presence of feedback. However, we did not find evidence that feedback improves metaknowledge and we also found no “transfer” of working memory practice benefits to other cognitive tasks (visual search, antisaccade, Raven’s matrices). Finally, we report the results of measures included to assess the validity of our design and participants’ expectations for improvement. First, we conducted post-hoc power analyses on current and previously published data, and we found that exposure to the task (i.e. practice effects) lowered power to detect a between-subjects effect. This decrease in power with practice may be important for planning new multi-session training studies with a between-groups effect. In addition, we included measures of participants’ effort, perceived improvement, and attitudes toward the malleability of intelligence. Subjective measures of effort and perceived improvement allowed us to test whether participants’ expectations were matched across training and control groups. We also included measures of subjects’ beliefs about intelligence (e.g. Theories of Intelligence Scale) to test whether individuals’ beliefs predicted the degree of motivation- or placebo-related improvement from pre- to post-test (regardless of training group). For example, prior work by Jaeggi et al. examining theories of intelligence and working memory (n-back) training found that participants reporting more “fixed” theories of intelligence (e.g. “You can’t really change how intelligent you are”) showed less improvement from pre- to post-test, regardless of their training group (i.e. n-back versus active control). # Materials and methods ## Participants Procedures were approved by the University of Oregon Institutional Review Board. Participants were recruited from the University of Oregon and the surrounding community and provided written, informed consent. They were between the ages of 18 and 35 (*M* = 20.5, *SD* = 2.66), and they self-reported normal or corrected- to-normal visual acuity and normal color vision. Participants were paid a total of \$200 for completing all sessions (2-hour pre-test, 2-hour post-test, and six 1-hour training sessions). They received \$30 after the pre-test session. Upon completing all 6 training sessions and the post-test, they received an additional \$170. If participants chose to withdraw from the study early, they were compensated at a pro-rated rate for their participation (\$3.75 per 15 min). We initially recruited 79 subjects. A total of 6 participants withdrew after the pre-test, and 1 additional participant withdrew after the first training session. This left a total of 72 subjects (48 female) for analysis (23, 25, and 24 per group). At a later date, we recruited an additional 35 subjects to serve as a passive control group. A total of 2 participants no-showed for their first scheduled appointment, and an additional 4 participants did not return for their second appointment. This left 29 participants (22 female) for final analyses. These participants were paid \$20 total for a 2-hour pre-test and \$30 total for a 2-hour post-test. ## Procedures Participants completed a 2-hour pre-test session, six 1-hour training sessions, and a 2-hour post-test session. After the pre-test, participants were pseudo- randomly assigned to one of three training groups. Pseudo-random assignment matched groups for average pre-test performance across all pre-test tasks (color whole report, crossword puzzles, color change detection, orientation whole report, visual search, antisaccade, Raven’s), as differences in pre-test performance often prevent sensible interpretation of results. To do so, we randomly shuffled group assignment until all tasks yielded no main effect of Group (*p* \>.05). Pseudo-randomization was performed once after the first 3 groups (“Working Memory—Feedback”, “Working Memory—No Feedback”, “Active Control”) completed the pre-test session but before any practice sessions occurred. The fourth group (“Passive Control”) was added at a later date and was not included in the pseudo-randomization procedure. The critical training group (“Working Memory, Feedback”, n = 25) practiced the discrete whole report task and received feedback during their training sessions. A second group (“Working Memory, No Feedback”, n = 23) practiced the discrete whole report task but did not receive any feedback about their performance. A third group (“Active Control”, n = 24) practiced doing crossword puzzles. This active control group served as a baseline comparison to working memory practice while attempting to control for researcher contact and expectations for improvement. Finally, a fourth group of subjects (“Passive Control, n = 29) completed only the pre-test and post-test sessions. Given concerns about the placebo effect causing differences in performance for passive control groups relative to active control groups, we wanted to include both. During the pre-test and post-test sessions, participants completed six cognitive tasks in the same order (Color Change Detection, Color Whole Report, Orientation Whole Report, Visual Search, Antisaccade, Raven’s Advanced Progressive Matrices). During the post-test session, participants filled out some questionnaires after completing all the cognitive tasks. Cognitive tasks and questionnaires are described below. During training sessions, participants were required to practice their assigned task for the full hour. All sessions took place in the lab in individual testing rooms. Crossword puzzles and questionnaires were completed with pencil and paper; all other cognitive tasks were administered on a PC running Windows XP with a 17-inch CRT monitor (refresh rate = 60 Hz). Stimuli were presented using MATLAB (The Mathworks, Natick, MA) with Psychtoolbox. Participants were seated approximately 60 cm from the monitor. ## Training tasks ### Color whole report The color whole report task is a working memory task that measures trial-by- trial fluctuations in performance. This task was used both for the Working Memory–Feedback group and for the Working Memory–No Feedback group. During each practice session, participants completed whole report trials for one hour in total; trials were divided into blocks of 30 trials each. The exact number of blocks completed varied across participants (*M* = 8.60, *SD* =.41). Critically, those in the Feedback group did not complete a different number of blocks per session compared to those in the No Feedback group (*p* =.41). On each trial, participants briefly viewed (250 ms) an array of 6 colored squares and remembered the items across a blank delay (1,000 ms). Colors were chosen from a set of nine easily discriminable colors (red, orange, yellow, green, blue, magenta, cyan, black, white). At test, place-holders containing a 3x3 grid of all possible colors appeared at the locations of the remembered items. The arrangement of colors in the response grid was fixed. Participants clicked the color in the place-holder corresponding to the color that was presented at each of the 6 locations. Accuracy for each trial was calculated as the number of correctly reported colors (out of 6 possible). After reporting colors for all items, the place-holders disappeared. The No Feedback group saw a blank gray screen after completing each trial. The Feedback group saw performance feedback after completing each trial. After presentation of feedback or a blank gray screen, the next trial began after a mouse click (followed by a 1,000 ms inter-trial interval). Performance feedback awarded participants points based on their performance. If participants performed well, they accrued points (+1 for 3 correct, +2 for 4 correct, +3 for 5 or 6 correct). If they performed poorly, they lost points or earned no additional points (-2 for 0 correct, -1 for 1 correct, 0 for 2 correct). In addition, participants earned a “streak bonus” for consistently performing well; the streak bonus was equal to the number of trials in a row that participants got at least 3 correct (e.g. 5 trials in a row = +5). Points accrued across each block of trials and reset to 0 at the beginning of each new block. Each trial’s feedback screen showed the number of correct items, the number of points gained or lost, total score for the current block, and an overall block high score. This points manipulation was previously shown to be very effective at reducing the frequency of working memory failures relative to no feedback or to simple feedback. While making responses, participants were asked to indicate their confidence by using a mouse-button press. If they were confident about an item, they were instructed to report the color of that item by clicking with the left mouse button. If they felt like they were guessing, they were instructed to instead use the right mouse button to respond. These confidence ratings can be used to calculate subjects’ metaknowledge of fluctuations in working memory performance. ### Crossword puzzles Participants were given a new packet of 3 crossword puzzles to work on during each 1-hour practice session. Puzzles were acquired from Boatload of Crosswords (Boatload Puzzles, Yorktown Heights, NY, USA). ## Pre- and post-test tasks ### Color change detection Change detection is a measure of working memory capacity. On each trial, participants remembered an array of 4, 6, or 8 briefly presented (150 ms) colored squares across a blank delay (1,000 ms). Colors were chosen from a set of nine easily discriminable colors (red, orange, yellow, green, blue, magenta, cyan, orange, black, white). At test, a single test square was presented at the location of one of the remembered items. The color of the test square was the same or different from the remembered color (50% probability). If the color was the same, participants pressed the “z” key; if it was different they pressed the “/” key. Participants completed 5 blocks of 30 trials each. ### Color whole report Stimuli and trial procedures were the same as described for the No Feedback version of the training task. For the pre- and post-test assessments, participants completed 4 blocks of 25 trials each. Unless participants in the Feedback group can “transfer” feedback-related performance improvements to later task contexts (without feedback), then we would expect no main effect of feedback group during the post-test. ### Orientation whole report This task was very similar to the color whole report task. Instead of remembering color, participants instead remembered orientations of circles with wedges cut out from them (“wrench-head” stimuli). Orientations were chosen from one of 4 values (up, down, left, or right). On each trial, participants briefly viewed (200 ms) an array of 6 orientation stimuli and remembered the array across a blank delay (1,150 ms). At test, place-holders appeared at each of the remembered locations. These placeholders contained “cross-hairs” (intersected vertical and horizontal lines). Participants clicked the arm of the cross-hair that matched the remembered orientation. After participants responded to all items, the place-holders disappeared. The next trial began after a spacebar press (followed by a 500 ms inter-trial interval). Participants completed 2 blocks of 30 trials each. ### Visual search Visual search measures how quickly participants can find a target among distractors. Participants searched for an upright “L” among homogenous distractors (the letter “T” rotated 0, 90, 180, or 270 degrees). The vertical part of the target “L” was slightly offset so that it more closely matched the upside-down “T” distractors. Participants pressed the left arrow key if the slightly longer side of the “L” target was facing left, and they pressed the right arrow key if it was facing right. The number of distractors ranged from 1 to 8. There were 5 blocks of 48 trials. ### Antisaccade The antisaccade task is a measure of attention and cognitive control. Participant fixated a cross in the center of the screen. After an unpredictable duration (.2–2.2 seconds), a cue (“=“) quickly flashed to the left or right of fixation. The cue flashed twice (100 ms on, 50 ms off, 100 ms on). Following a 50 ms delay, a target appeared in the opposite hemifield for 100 ms. The target was the letter “P”, “B”, or “R”. Following a 50 ms delay, the target letter was masked twice (letter “H” for 100 ms, blank 50 ms, number “8” for 100 ms). To detect the target, participants needed resist capture by the cue and quickly move their attention and eyes to the opposite hemifield. Participants reported the target by pressing the letter “P”, “B”, or “R” on the keyboard. For correct trials, the entire screen flashed green for 500 ms. For incorrect trials, the entire screen flashed red. There were 4 blocks of 36 trials total. ### Raven’s Advanced Progressive Matrices The Advanced Progressive Matrices task is a measure of abstract reasoning ability and fluid intelligence. For each question, participants viewed a 3 by 3 grid of abstract geometric shapes. These shapes are related to one another (e.g. an abstract rule dictates similarity across columns / rows of the grid). The bottom right corner of the grid is missing, and participants must choose the item that best belongs from one of 8 choices. The full test is 36 questions that are presented in ascending order of difficulty. To measure change from pre- to post-test, questions were divided into two sets of 18 (even and odd questions). Most participants received even questions at pre-test and odd questions post- test (due to clerical error, 4 participants received odd questions at pre-test and even questions at post-test). Participants were given 10 minutes to work on the set of 18 questions; scores were calculated as the total number of correct questions. ### Timed crossword puzzle Participants were given 1 crossword puzzle, and they were given 10 minutes work on the puzzle. Participants were instructed that they should try to get as many words correct as possible in the allotted time. Crossword puzzle accuracy was scored as the total number of complete, correct words per minute of task time. ## Questionnaires Questionnaires were administered at the end of the post-test session. All questions and scales administered to participants are available on our Open Science Framework page (<https://osf.io/839dz/>). We collected information about subject demographics (e.g. age, handedness). In addition, we were interested in whether participants thought that they improved from the pre-test to the post- test. We also used a few different questionnaires from Dweck (2000) to characterize the attitude of participants in our sample towards learning and intelligence. We reasoned that participants who thought they improved more or who view intelligence as more malleable might show greater practice benefits and/or placebo effects. ### Demographic information and perceived performance We collected participants’ age, gender, native language(s), handedness, and parental education level. We also asked participants to estimate their typical amount of sleep per night, and to give ratings of their average levels of alertness and motivation across the training sessions. In addition, we asked participants 6 questions about their perceived level of effort and improvement across their completed sessions. An example of subjects’ perceived effort is the statement, “I found it difficult to care very much about how I was doing on the tasks.” An example of subjects’ perceived performance is the statement, “I feel that my performance on the tasks improved from the first session to the last session.” Participants rated their endorsement of each statement from 1 (strongly agree) to 6 (strongly disagree). For ease of interpretation, these measures are plotted such that 6 represents the strongest agreement with each construct (e.g. more effort) rather than disagreement. ### Theories of intelligence scale for adults This 8-question scale quantifies participants’ views on the malleability of intelligence. Participants rated their endorsement of each statement from 1 (strongly agree) to 6 (strongly disagree). People who view intelligence as more fixed (“entity theorists”) are more likely to endorse the statement, “To be honest, you can’t really change how intelligent you are.” Those who view intelligence as malleable (“incremental theorists”) would endorse a statement such as, “You can change even your basic intelligence level considerably.” ### Goal choice items questionnaire This 4-item questionnaire places learning goals against performance goals. Learning goals represent engaging in an activity in which the participant would learn a lot, but would not necessarily perform well. Performance goals represent engaging in an activity where the participant would excel, but not necessarily be challenged. For example, someone who values performance goals over learning goals would be more likely to endorse the statement, “Although I hate to admit it, I sometimes would rather do well in a class than learn a lot.” Previous work has shown that people who view intelligence as malleable are more likely to endorse learning goals over performance goals. ### Confidence in one’s intelligence This 3-item measure asks participants to rate their confidence in their intelligence. This measure is most often used to show that those who view intelligence as fixed versus malleable do not differ in their confidence or optimism toward their own intelligence. # Results ## Improvement on practiced task To test whether feedback differentially changed the rate of improvement across all sessions, we ran a mixed ANOVA with between-subjects factor Feedback and within-subjects factor Session (Pre-test, 6 training sessions, and Post-test). We quantified behavioral performance in a couple of ways. First, we looked at mean performance (the average number of correct items on each trial). Second, we looked at the change of “poor performance” trials (less than 3 correct). The feedback manipulation used here incentivized participants to get 3 correct by awarding points only when this goal was achieved. Consistent with this incentive structure, we found previously that this feedback manipulation had the greatest impact on the proportion of poor performance trials. As such, we expected *a priori* that this measure of performance should be most sensitive to any changes in performance. We found a large main effect of session on mean performance, *F*(2.52,115.71) = 30.0, *p* \<.001, η_p² =.40, and proportion of poor performance trials, *F*(3.23,148.42) = 28.9, *p* \<.001, η_p² =.39, indicating that practice led to significant improvements in working memory performance. Note, the Greenhouse-Geisser correction is applied where the assumption of sphericity is violated. There was no main effect of feedback for either measure (*p* \>.11). However, there was a significant interaction between feedback and session for the poor-performance measure, *F*(3.23,148.42) = 3.68, *p* =.01, η_p² =.07, and a trending interaction for the mean performance measure, *F*(2.52,115.71) = 2.33, *p* =.09, η_p² =.05. Post-hoc comparisons revealed that the likely cause of the interaction was because of a larger difference between the feedback and no-feedback groups for earlier practice sessions. Two- tailed t-tests revealed a significant difference in the proportion of poor performance trials for only the first (*p* =.038) and third (*p* =.045) practice sessions. Indeed, ANOVAs testing for an interaction between pre-test scores and scores on the first session were significant (mean number correct, *p* =.004, proportion poor performance trials, *p* =.002). By design, there was no difference in pre-test scores between groups. Thus, this significant interaction supports the conclusion that there was an initial difference between feedback and no feedback groups in the first practice session that did not persist across all practice sessions. We also tested the hypothesis that participants in the feedback group would develop superior meta-knowledge relative to the feedback group. We found no supporting evidence for this hypothesis. To quantify metaknowledge accuracy that is relatively independent of overall bias in confidence reports, we calculated a “metaknowledge correlation” for each individual participant. To do so, we plotted the number of correctly reported items for each trial against the number of confident items for each trial then calculated the correlation coefficient (Pearson’s) for each participant. Note, some subjects occasionally used the same button (“no guess”) for all responses within the entire session. For these subjects, the correlation coefficient is undefined so they are excluded from the analysis (total remaining: 18 in no-feedback group, 20 in feedback group). We found that metaknowledge performance increased over time, as shown by increased correlation strength, *F*(3.91,140.86) = 6.58, *p* \<.001, η_p² =.15. There was also a main effect of feedback, but not in the predicted direction. Those in the no-feedback group had higher metaknowledge than those in the predicted group, *F*(1,36) = 11.0, *p* =.002, η_p² =.24, and there was no interaction between feedback and session, *F*(3.91,140.86) = 1.4, *p* =.24, η_p² =.04. Post-hoc paired t-tests revealed that there were no significant differences between groups in either the pre-test (*p* =.27) or post-test sessions (*p* \>.07). However, during all 6 training sessions the no-feedback group had higher metaknowledge relative to the feedback group (*p* \<.001 to *p* =.015). This indicates that the main effect of group was dependent on the presence of feedback; the between-groups differences went away when no feedback was administered in the post-test session. As a second test of metaknowledge that would not exclude any subjects, we instead looked at overall level of confidence relative to actual performance (measured as the absolute value of the difference between confidence and accuracy for each trial). Unlike the first measure, this difference measure is most similar to measures of “bias” in ratings, rather than accuracy. For this measure, values close to 0 would indicate better metaknowledge. This difference measure revealed the same significant effect of session; participants’ metaknowledge improved across sessions, *F*(3.07, 141.25) = 3.99, *p* =.009, η_p² =.08. There was again a main effect of feedback group, *F*(1,46) = 6.614, *p* =.013, η_p² =.13, and no feedback by session interaction (*p* =.32). Once again, the main effect of group was limited to sessions where feedback was actually presented. There was no significant difference between feedback groups for either the pre-test (*p* =.90) or post- test sessions (*p* =.42). However, there was a significant difference between groups for 4 of 6 practice sessions (*p* \<.05). Unlike the change in metaknowledge correlation coefficient, the observed improvement over time was limited to the change from the pre-test to the first practice session. If only the 6 practice sessions were included in the ANOVA, there was no main effect of session (*p* =.29). Across our two measures of metaknowledge, we found that practicing the working memory task led to improved metaknowledge (with or without feedback). However, this improvement was not necessarily acquired gradually over time; we found inconsistent time-courses of improvement for metaknowledge accuracy (correlation measure) versus bias (difference measure). Surprisingly, we also found that the presence of feedback altered metaknowledge in a counter-intuitive way. Specifically, those who received trial-by-trial feedback had *worse* metaknowledge relative to those who did not receive any feedback. However, these differences did not carry through to a post-test session with no feedback. This surprising finding may indicate that those who received feedback paid less attention to internal judgments. Finally, we checked whether the active control group (crossword puzzles) improved on their practiced task over time. A repeated-measures ANOVA with factor Session revealed a significant difference over time, *F*(3.57,78.56) = 24.52, *p* \<.001, η_p² =.53. This difference was not entirely driven by the large rise from the last practice session to the final post-test. There was still a significant improvement across only the 6 practice sessions, *F*(5,110) = 4.67, *p* =.001, η_p² =.18, with a significant linear trend (*p* =.01). Thus, it seems that participants in the active control group were effortfully engaged throughout the practice sessions. ## Improvement from pre-test to post-test First, we compared improvement from pre-test to post-test for each of the practiced tasks. We found that those who were in either of the two color whole report groups improved significantly more on the color whole report post-test measure relative to controls. Data for all four groups are plotted. For this measure, there were no significant differences or between the two working memory groups (feedback vs. no feedback, *p* =.97) or the two control groups (passive vs. active, *p* =.67). As such, we collapsed the data into two groups: “working memory practice” and “control” groups. This analysis revealed a significant interaction between group and session, F(1,99) = 51.68, *p* \<.001, η_p² =.34. Those who practiced the working memory task improved from pre- to post-test *t*(47) = 7.66, *p* \<.001, whereas those in the control groups did not improve, *t*(52) =.13, *p* =.90. On the other hand, we did not find a significant interaction between these groups for improvement in metaknowledge performance (absolute value of number confident–number correct). Both groups improved their metaknowledge (*p* =.003), but there was no interaction between group and metaknowledge improvement (*p* =.21). However, this measure is difficult to interpret, as there were overall group differences at baseline (*p* =.02). Likewise, participants in the crossword puzzle group improved more on the crossword puzzle post-test relative to the other three groups. Data for all four groups are plotted in. Accuracy for the crossword test was quantified as words per minute. One subject was excluded from the crossword puzzle analysis because of a clerical error (their test was not timed to precisely 10 minutes, and they were instead allowed an unknown extra amount of time to work). There was no difference in performance for any of the 3 groups who did not practice the crossword puzzle task (feedback, no feedback, and passive control), *p* =.68. As such, they were collapsed into one group and compared to the crossword puzzle group. Both the crossword puzzle practice group (*p* \<.001) and the non- crossword group (*p* \<.001) improved from pre- to post-test. However, the crossword group improved more than the non-crossword group as indicated by a significant interaction between group and session, F(1,98) = 8.78, *p* =.004, η_p² =.08. Next, we examined whether practice on the color whole report working memory task led to improvements on the other two working memory tasks. These two tasks each differed in one aspect from the practiced task. The color change detection task used the same stimulus set (9 colors) but required a different response (same or different judgment). Conversely, the orientation whole report task used a different stimulus set (4 orientations), but used the same response mode (clicking each item). We found evidence for stimulus-specific benefits of practice. Those in the working memory practice groups improved more on color change detection than did those in the control groups. There was again no difference between the two working memory groups (*p* =.86) or between the two control groups (*p* =.997) so they were collapsed. We found a significant interaction between group and session, F(1, 99) = 14.91, *p* \<.001, η_p² =.13, indicating that those who received working memory practice improved more than did those in the control group. In fact, there was no significant change in the control group’s performance from pre- to post-test, *t*(52) = 1.09, *p* =.28. Unlike the color task, we did not find any evidence of a systematic benefit of color working memory practice on the orientation task. There were some baseline differences in performance across groups, so we did not collapse into two groups. We found a small benefit of practice on performance, F(1,51) = 6.03, *p* =.02, η_p² =.06, but no interaction between group and session, *p* =.24. We looked separately at each group for evidence of improvement from pre- to post- test, and we found that only the no-contact control group improved from pre- to post-test (*p* =.001). Thus, we found no evidence that practice with a color working memory task led to improvement on an orientation task. Instead, we conclude that the practice benefits obtained were highly feature-specific, similar to prior work examining visual working memory training (but also see). Finally, we found no differences in groups’ performance for any of the other cognitive tasks (Raven’s, Antisaccade, Visual Search), as shown in. However, we did find overall improvement on these tasks from pre-test to post-test. Raven’s accuracy was scored as the total number of correct items completed in 10 minutes. There was an increase in accuracy on the Raven’s from pre- to post- test, F(1,97) = 48.47, *p* \<.001, η_p² =.33, but no effect of group, F(3,97) =.29, *p* =.83, η_p² =.009, and no interaction between group and performance, F(3, 97) = 2.10, *p* =.11, η_p² =.06. Antisaccade performance was scored as percent error (wrong letter reported) and as average reaction time. There was an increase in antisaccade accuracy from pre- to post-test, F(1,97) = 23.84, *p* \<.001, η_p² =.20, but no effect of group, F(3,97) = 1.27, *p* =.29, η_p² =.04, and no interaction between group and performance, F(3, 97) =.29, *p* =.84, η_p² =.01. The same pattern of results was seen for reaction time. Finally, visual search performance was quantified as the average response time. We found improvement in visual search reaction times from pre- to post-test, F(1,97) = 17.61, *p* \< .001, η_p² =.15, but no effect of group, F(3,97) = 1.56, *p* =.20, η_p² =.05, and no interaction between group and performance, F(3, 97) = 1.41, *p* =.24, η_p² =.04. The same pattern of results was observed for search slopes. Thus, practicing a color working memory task did not lead to any marked improvement in other cognitive tasks. Correlations between measures and the reliability of measures across pre- and post-test sessions are shown in – Tables in the Supporting Information. Reliabilities for visual search reaction times and for Raven’s matrices were particularly poor (*r* \<.5), so these tasks should be interpreted with caution. Reliability values for the working memory tasks were quite a bit higher for the control group (*r \~*.8) than for the working memory practice group (*r* \~.6) presumably because there was some variability in how much participants’ benefited from practice (thus reducing the correlation between pre- and post- test). ## Growth mindset did not predict performance improvement Next, we examined our hypothesis that participants with a growth mindset may show greater improvement relative to those with a fixed mindset. This effect might even bear out in tasks where we would expect no improvement due to practice alone. That is, those who think they are capable of growth might be more susceptible to expectancy-based placebo effects. We found no support for this hypothesis. To create a single “growth mindset” score, we re-ordered reverse-scored items and then averaged across all 8 questions from the Theories of Intelligence scale (Cronbach’s alpha of the 8 items =.95). A score of 1 indicates the greatest possible degree of having a growth or “incremental mindset” whereas a score of 6 indicates the maximum degree of having a “fixed mindset”. The average growth mindset score was 2.72 (SD = 1.01, skew =.24), indicating that our sample leaned toward having a growth mindset (t-test compared to expected middle score of 3.5, *t*(100) = 7.82, *p* \<.001, 95% CI \[2.52, 2.92\]). For each of our cognitive measures, we computed a difference score (post- test–pre-test) as a measure of overall task improvement. We found no significant correlation between task improvement and growth mind-set for any of the tasks: Color Whole Report (*r* =.01, *p* =.96), Crossword Puzzles, (*r* =.17, *p* =.09), Color Change Detection (*r* =.07, p =.47), Orientation Whole Report, (*r* =.05, *p* =.65), Raven’s (*r* =.02, p =.82), Antisaccade accuracy (*r* =.02, *p* =.84), or Visual Search speed (*r* =.04, p =.67). Likewise, there were no significant correlations between task performance and growth mindset when correlations were examined separately for control groups and working memory practice groups. We also computed composite scores for the “goal choice” questionnaire and for the “confidence in intelligence” questionnaire. We found no relationship between the growth mind-set score and goal choice, *r* =.09, *p* =.39. We also found no relationship between growth mind-set and confidence in intelligence, *r* =.18, p =.08. ## Differences in perceived improvement and effort across groups We were interested whether those in different groups reported subjective differences in perceived improvement or effort. First, we looked at differences across all 4 groups using a one-way ANOVA. We found that there was no difference in self-reported effort between groups, F(3,97) = 1.67, *p* =.18. However, we did find a significant difference in perceived improvement, F(3, 97) = 3.12, *p* =.03. By eye, this effect of group appeared to be mostly driven by the difference between the two working memory groups relative to the controls. A post hoc test collapsing into two groups (“working memory task” or “control”), we found a difference in perceived improvement, F(1,99) = 9.37, *p* =.003. Although we found no large effect of subjective effort across the four groups, we nevertheless found differences in expectation (as indexed by subjective perceived improvement). This result emphasizes how difficult it is to find control conditions that can truly eradicate differences in expectation and control for placebo effects. ## Post-hoc power analyses based on prior work We think it is important to qualify these results with a power analysis and a brief discussion of within- versus between-subjects manipulations in studies of working memory. In earlier work, we found robust within-subject effects of performance feedback on working memory performance. However, because individual differences in working memory are both large and stable, the expected size of an intervention’s effect is typically much smaller than the range of individual differences. Unfortunately, we did not have the foresight to conduct both within- and between-subjects power analyses before collecting this experiment’s data. Post hoc, however, we can illustrate the extent of change to expected power for the same effect run between versus within subjects. Power estimates were calculated using the G\*power 3.1 application. First, we looked at the effect size and power for the feedback effect observed in Experiment 3 of Adam and Vogel (2016), as this was the same feedback manipulation used in the current study. In Experiment 3, participants received feedback for half of the experiment and no feedback for the other half of the experiment. The order of these two conditions was blocked and counterbalanced across participants. The within-subjects effect of feedback reported in Adam and Vogel (2016) was very strong, with a calculated effect size of *d* =.93 and power (1 − β) \>.99 (n = 52). Next, we instead calculated effect size as between-subjects effects. We calculated between-subjects power separately for the first half and the second half of the experiment so that we could see if the computed effect size changed after subjects had exposure to the task (i.e. after experiencing one of the 2 conditions during the first half of the experiment). For the first half of the experiment, the between-subjects effect size was *d* =.78. With group sizes of 24 and 28, we found that power was sufficient, (1 − β) =.79. Note, these sample sizes are similar to the ones we chose for the present experiment. Unfortunately, however, the calculated effect size decreased in the second half of the Adam and Vogel (2016) experiment. The effect size for the second half of the experiment decreased to *d* =.51, (1 − β) =.44. Thus, had we run this full set of analyses before conducting the current experiment, we may have suspected that we might need larger group sizes to counteract a decrease in effect size potentially caused by practice effects or some other aspect of previous exposure to the task. For the first practice session in the present study (after exposure to the task in the pre-test), the difference between the feedback and no feedback groups was closer to the second, smaller effect found in Adam and Vogel (2016). We calculated an effect of size of *d* =.39, meaning that around 100 subjects per group would be needed to achieve power (1 − β) =.80 to detect a between-subjects effect within this single session. Given the robust effects of feedback for our earlier within-subjects manipulation and our relatively low power here, we speculate that the small effects of feedback that we saw would hold up with a larger between-groups sample. However, these results should nevertheless be interpreted conservatively; we see hints that feedback boosts practice benefits, but these effects are relatively short-lived. The relative effects of feedback appear to dissipate as practice effects increase. # Discussion Feedback can improve working memory performance when it points participants toward an optimal goal. Here, we asked whether practicing with feedback can help participants more quickly and efficaciously reduce the frequency of working memory failures. Our goal of reducing failures with feedback is similar to, but distinct from, the goal to increase capacity throughout the very large “working memory training” literature. Both goals would have the same consequence of improving overall working memory performance but would rely upon distinct mechanisms and strategies. We hypothesized that practicing with feedback relative to without feedback would lead to faster, more robust improvements in working memory performance and would also improve subjects’ awareness of working memory failures. We found only partial support for these hypotheses. Consistent with our predictions, practicing with performance feedback increased working memory performance relative to practicing without feedback. However, the size of this effect was rather small and did not persist over time. Relative to the no- feedback group, participants in the feedback group more successfully reduced the frequency of poor performance trials only for some sessions. In addition, training with feedback did not improve subjects’ metaknowledge performance. If anything, subjects’ metaknowledge performance actually *declined* during feedback. ## The benefits of visual working memory practice are highly stimulus specific Practicing on one visual working memory task led to improved performance on another visual working memory task with the same stimuli and a different response mode (color change detection) but led to no improvement on a task with different stimuli but the same response mode (orientation whole report). This finding is consistent with work by Gaspar and colleagues showing that visual working memory training benefits were highly stimulus specific (extensive adaptive training on an object change detection task did not confer benefits to an orientation change detection task). Likewise, Buschkuehl and colleagues recently found that adaptive visual working memory training yielded highly task- specific improvements. Future work is needed to establish the mechanisms underlying these stimulus-specific improvements to working memory performance (e.g. familiarity affecting encoding or storage, “chunking” strategies, or improved retrieval). Unfortunately, the stimulus-specificity of visual working memory practice greatly limits the scope of the expected benefits of practice. Practice may be beneficial if the desired outcome is to more effectively remember or mentally manipulate the same stimulus set (e.g. the same set of icons in a complex software display), but may be of limited utility if transfer to novel stimulus sets is desired. ## Robust practice benefits with fixed task difficulty Frequently, the use of an adaptive task is described as critical for yielding improvements to working memory performance, and direct comparisons have found that adaptive tasks yield larger improvements to performance than non-adaptive tasks. However, when comparing adaptive and non-adaptive training protocols, it is important to control for difficulty, otherwise the manipulation of adaptiveness can be confounded with overall task difficulty during training. For example, work by von Bastian and Eschen found that there was no difference between adaptive and non-adaptive working memory practice when overall difficulty was matched. Thus, mere exposure to a variety of difficulty levels has been shown to yield large increases in working memory performance. In the current study, we found that exposure to a difficult condition alone was sufficient to yield robust increases in visual working memory performance. Participants never encountered any easy trials (all trials were set size 6), yet we observed robust improvements in performance across sessions. The effects of adaptive versus non-adaptive procedures have yet to be directly compared with a similar visual working memory task, so we cannot make strong claims about the efficacy of adaptive versus non-adaptive practice in this context. However, our finding of robust practice effects is consistent with previous observations of improvements to visual working memory performance across sessions, both for adaptive and non-adaptive tasks. An interesting open question for future work is to what extent participants may use “self-set” adaptive strategies when they are only given difficult trials, and whether such self-set goals underlie the practice benefits that we observed. Since a whole report task requires reporting each item individually, participants’ improvement may be aided by self-selected goals that change across practice, for example, “To start out, I’ll just try to get 2 correct and then guess on the rest.” ## Spacing may influence visual working memory practice benefits Here, we reported robust improvements to visual working memory performance with practice, consistent with previous work. However, inconsistent with this core result, a study by Olson and Jiang reported that visual working memory is impervious to training. We think that the spacing of practice may explain the discrepancy between these two results. In Olson and Jiang (2004), all of the training was “massed” within a single practice session. Conversely, studies reporting robust practice effects on visual working memory performance have spaced practice across multiple days. A “spacing effect” is defined by a larger performance benefit when an equivalent amount of practice is separated in time as opposed to being massed within a single practice session. Spacing effects have been extremely well-characterized in the domains of episodic memory and motor learning, but have not been formally quantified in the working memory literature. However, a qualitative assessment of the working memory literature suggests that spacing effects in working memory may respect a non-monotonic function, like that reported for episodic memory. With short “block breaks” within a single session (e.g., minutes), working memory practice effects are absent. With intermediate spacing (e.g., 1 day– 1 week), a robust practice benefit is observed from the first to the second session. However, with more distant spacing (e.g., many weeks to months), no robust practice benefit is observed (as for the Passive and Active Control Groups, ; also see). We think this is an intriguing qualitative pattern of results, but future work is needed to systematically quantify the effects of spaced practice on visual working memory performance. ## Previous exposure to tasks may reduce power We found that practice decreased a between-groups effect size, which has profound implications for other studies attempting to calculate power for multi- session data based on single-session pilot studies. We reanalyzed an earlier dataset in which we could make the same comparison (feedback versus no-feedback) in both a within-subjects and a between-subjects manner. As expected, the overall between-subjects effect size was smaller than the within-subjects effect size. Critically, the same between-subjects effect size was reduced after all participants had experience with the task. This suggests that power estimates made from single-session task performance may not generalize well to multi- session studies. Thus, practice effects should be considered when planning sample sizes for large training studies. ## Feedback about accuracy did not improve metacognition Consistent with previous work, we observed a reliable increase in metaknowledge performance with practice. Surprisingly, however, we found that subjects’ metaknowledge performance was *worse* during sessions with feedback. This finding suggests that feedback may actually undermine the goal of improving metaknowledge if the feedback leads participants to spend less effort on monitoring their own performance. For example, if feedback emphasizes improving accuracy, then participants may neglect the secondary task of accurately rating their metaknowledge to better maximize available resources for the working memory task. Because our feedback focused on memory accuracy and did not reward metacognitive accuracy, we think that participants engaged in such a tradeoff. Note, we have used the term “with feedback” to discuss behavioral effects related to our specific weighted feedback intervention relative to no feedback. However, we only tested a single feedback manipulation in this study, so our conclusions should not be interpreted to mean that all feedback manipulations would yield similar effects. Changes to the timing, frequency, content, or modality of feedback would be expected to modulate the effect of feedback on behavioral performance. Manipulating the precise nature of performance feedback would be useful for investigating the limits of feedback-related improvements to working memory and metaknowledge performance. For example, a new weighted feedback design which combines both working memory accuracy and metacognitive accuracy may prove effective at boosting both working memory performance and metacognition. Finally, future work employing near real-time feedback about behavioral, neural, and physiological markers of attentional state could be used to provide participants precise, theoretically-driven feedback and to test the specific mechanisms underlying feedback-related improvements. ## Mixed evidence that crossword puzzles are an adequate active control One critical aspect of any intervention is the choice of a control group. The problem of placebo effects is particularly pernicious in multi-session behavioral interventions. Spurious placebo-like effects might be generated from differential amounts of experimenter contact, task engagement, or expectations for improvement. If these issues cannot be avoided altogether, they can at least be measured. As an example, Boot and colleagues performed an online study where they measured the expectations of participants. Participants viewed a training task (e.g. action video game) and then rated whether they thought by practicing that task they might improve on some other tasks and abilities. Unfortunately, Boot and colleagues found that expectations were not well matched by commonly- used control tasks (e.g. Tetris). Here, we used crossword puzzle practice as a control for working memory practice. Because people commonly believe that doing crossword puzzles reflects intelligence and can allow one to stay mentally “fit”, we thought this control task would have a good chance of leading subjects to believe that practice benefits were expected. Others, however, have speculated that crossword puzzles do not adequately control for demand and expectations relative to working memory tasks. In our experiment, subjective measures of overall effort and perceived improvement (rated as average improvement for all pre- and post-test tasks) revealed mixed results for the efficacy of crossword puzzle control groups. On the one hand, those in the crossword puzzle group received the same amount of experimenter contact and did not report lower levels of effort throughout the experiment, indicating that they stayed engaged throughout the training sessions. On the other hand, they reported less perceived improvement than those in the working memory groups. Like the working memory groups, the crossword puzzle group did actually improve more on their trained task than the other groups. So, it would be accurate for this group to report that they improved *slightly* more (for the average across all of the tasks) than the passive control group. However, this was not the case. Instead, the active control group reported equivalent perceived improvement to the passive control group and lower perceived improvement than the two working memory practice groups. Thus, it is unclear whether or not crossword puzzle training adequately controlled for participant’s expectations relative to working memory practice groups, and future work is needed to more precisely characterize participants’ beliefs and expectations about a wide array of potential control tasks. ## Conclusions We found robust practice-related improvements to visual working memory performance, both with and without performance feedback. Performance feedback somewhat augmented practice-related improvements, but the effects of feedback were somewhat weak and transient. Once subjects were well-practiced on the working memory task, the benefits of feedback dissipated. In addition, the benefits of feedback did not persist after feedback was taken away in the post- test session. Participants got better at monitoring their own performance with practice, but feedback did not play a role in this improvement. If anything, trial-by-trial feedback actually reduced participants’ self-monitoring and metacognitive accuracy. Finally, despite robust, stable improvements in working memory ability with practice, we found no evidence that practice benefits led to concomitant improvements in other cognitive abilities. # Supporting information We thank Irida Mance for assistance with task preparation and data collection, and we thank Richard Matullo for organization of subject payments and data collection. Thanks also to Will McGuirk, David Grady, and Emily Taylor for assistance with data collection. [^1]: The authors have declared that no competing interests exist.

# Introduction Among those with depression, it is older adults who are most likely to attempt and complete suicide. The role of concomitant physical disease, and the etiology of depression among seniors are poorly understood and confusing. Older adults are more likely than the young to have chronic illnesses such as cardiovascular disease, diabetes, or respiratory compromise, all recognized both as risk factors for and complications of depression.\[–\] Yet, paradoxically, most research demonstrates a decrease in prevalence of depression with increasing age.\[–\] There is evidence that some individuals, including seniors, with physical illness may experience a response shift of recalibrating their physical health expectations, reprioritizing their values, or reconceptualizing constructs in order to maintain their quality of life while experiencing a decline in health.\[–\] The differential impact on women and men, that is, of sex/gender, adds to the confusion. In German and Canadian studies of older adults, increased chronic disease aligned with increased depressive symptoms in women. However, a similar study in Italy found the reverse, with worsened physical health being more highly associated with depression in men. Others have found that the association between chronic disease and depression diminishes with age but is unrelated to sex/gender. These conflicting findings are further muddied by variations in susceptibility to depression among older adults depending on country income and rural residency. An international cross-sectional study found that the prevalence of depression decreased with age in high-income, but not low- and middle-income countries. Some have found that older adults living in metropolitan areas have decreased levels of depression relative to those in rural areas, whereas others find no association between rural setting and depression. The role of setting is important in understanding the etiology of mood changes in seniors. If older adults in high-income and metropolitan settings, but not those in middle-income or rural settings, experience decreases in depression, this hints at the role of cultural or social contexts. Conversely, a decline in depression across all settings would suggest that seniors experience improved mood as they adapt to the aging process itself, regardless of other social factors. Understanding how physical decline contributes to changes in mood will enable clinicians to better identify and treat older patients at risk. To our knowledge, no studies have previously used a longitudinal, population-based design to measure changes in mood and physical health across middle- and high- income countries. Our aim was to address this knowledge gap using data collected over four years from four study sites in Canada, Colombia, and Brazil to assess the relationships between physical decline and depression in older populations. # Methods ## Data sources/study populations The International Mobility in Aging Study (IMIAS), is a population-based longitudinal study of community-dwelling seniors, aged 65–74 on enrolment in 2012, and living in Kingston (Canada), St. Hyacinthe (Canada), Natal (Brazil), Manizales (Colombia), or Tirana (Albania). Each study setting reflects a different way of life, culture, and language, but inhabitants within each setting are relatively homogenous ethnically and culturally. Kingston is a university city in Ontario, St. Hyacinthe is an agricultural centre in Quebec, Manizales has a large coffee-growing region in the Andes, and Natal is a coastal city in a relatively low-income area of Brazil. Because of uncertainty regarding the accuracy of the depression measure for Tirana, it was excluded from this study although preliminary analyses (not reported here) found that its inclusion did not change outcomes. Respondents were recruited randomly from primary care medical centers in each setting. However, to satisfy local ethics committees' stipulations in Canada, researchers were not allowed to invite potential participants directly. Instead, patients of the target age in one large family practice received a letter signed by their primary care physician inviting those willing to participate to contact the study coordinator. In Manizales, the Public Health Insurance registry, which includes most adults age 65–74, was used to identify those to be invited to participate. In Natal, invitees were identified using neighbourhood primary care registers. Participation rates varied by study setting. In Kingston and St. Hyacinthe, 30% of those mailed the invitation contacted IMIAS, and 95% of that group chose to participate. Participation rates in Latin America were over 90%, which may reflect the appreciation respondents expressed at receiving free medical exams and blood tests. IMIAS recruitment and study design are described in greater detail in Zunzunegui et al., 2015. ## Study design In 2012 approximately 200 community-dwelling men and 200 women were enrolled per site for a total of 768 men and 840 women (1608 total). A small number of trained personnel conducted face-to-face interviews in offices (Kingston) or respondents’ homes (St. Hyacinthe, Manizales, Natal), and performed a battery of physical tests and measurements. Our analysis is limited to those who had a valid depression score at all three time points, as measured by the Center for Epidemiologic Studies—Depression (CES-D) in 2012, 2014, and 2016, resulting in 1161 respondents. Of the respondents who were lost to follow up, 123 moved, 85 died, 24 failed the cognitive screen, and 180 refused or were too sick to participate. Our inclusion criteria and the percentage of respondents who were excluded or lost to follow-up due to declining capacity or death are consistent with other longitudinal studies on aging including the Baltimore Longitudinal Study on Aging and the Italian Longitudinal Study on Aging.. Earlier analyses showed that the proportion of all 2012 participants (ie including those subsequently lost to follow-up) with depression was 244/1601 or 15.2%. This is nearly identical to the depression outcomes for 2012 of those who remained in the cohort, suggesting there was no substantive difference in mood at baseline across the two groups. ## Exclusion criteria Those living in hospital or an assisted living facility were excluded from the initial cohort as our overall aim was to examine changes in mobility over time and therefore we enrolled participants who were able to live independently. Respondents remained enrolled in the study as long as they remained in the community, unless their cognitive ability declined such that they had four or more errors on the Leganes Cognitive Test (LCT). ## Outcome measure: Change in depression The primary outcome was change in depression (Δ depression), measured by the difference in respondents’ depression score in 2016 compared to 2012. We assessed depression using the Center for Epidemiological Studies Depression Scale (CES-D), a screen for depression which has been validated among older, community-dwelling adults and in English, French, Spanish, and Portuguese.\[–\] The CES-D consists of 20 questions about mood over the preceding 2 weeks, each assessed from 0 (rarely/never) to 3 (most/all of the time). Potential scores range of 0–60, and scores of 16 or more suggest a significant symptom burden indicative of depression. Change in depression score was defined as the difference between CES-D 2016 and CES-D 2012, yielding a continuous variable, ranging from -40 (less depressed) to 32 (more depressed). ## Predictors of change in depression The primary predictors examined were change in chronic health conditions (Δ CHC), change in grip strength (Δ grip) reflecting objective physical function, and change in self-rated health (Δ SRH), reflecting subjective perceptions of health. CHC was assessed via a checklist of 8 self-reported diseases (hypertension, diabetes, malignancy, chronic lung disease, heart disease, cerebral aneurysm/stroke, osteoarthritis, and osteoporosis). We assessed baseline number of CHC in 2012 by asking respondents “Has a doctor or nurse ever told you that you have \[condition\]?” and in 2014 and 2016, we asked respondents, “Since we last saw you, has a doctor or nurse told you that you have \[condition\]?”. In 2014 and 2016, we used the total number of diseases endorsed at any point, accounting for duplication. CHC was a continuous variable with a possible range from 0–8, and an actual range of 0–7. Δ CHC was the increase in CHC from 2012 to 2016, with greater values indicating increases in CHC. We measured hand grip strength with a grip dynamometer (Jamar Hydraulic Hand Dynamometer). The highest of three sequential measures at each time point with the dominant hand was utilized. Grip strength is negatively correlated with all-cause mortality and poor health outcomes in healthy older adults, and thus reflects objective physical health.\[–\] We calculated Δ grip by comparing values for 2016 and 2012, with negative scores indicating decreases in strength on this continuous scale. Self-rated health (SRH) was reported as 1 (very poor) to 5 (very good), assessed in 2012, 2014, and 2016. Poor self-rated health has been associated with increased depression in elderly populations, irrespective of objective physical health. We calculated Δ SRH by comparing SRH scores in 2016 and 2012. Positive values reflected increased subjective physical health on this continuous scale. ## Covariates Also included as potential explanations for change in depression were site, sex, age, income sufficiency, and smoking. Study site and sex were included because of discrepancies in prevalence of depression across countries, and between women and men. Smoking was included because of its association with depression and many of the comorbidities assessed within CHC. Income sufficiency was selected as the measure of socioeconomic status (SES) because it overcomes international discrepancies in income, cost of living, education, and standard of living across study sites. It has been used by others as an SES control variable in international studies of mental illness with diverse study sites. Income sufficiency was defined as the extent to which income met the respondent’s needs from 1 (insufficient) to 3 (very sufficient). ## Statistical analysis All analyses were performed using SPSS V. 24. Normality was assessed with visual inspection, Q-Q plots, box plots, and measuring skewness, and kurtosis. We used paired sample t-tests of changes in depression, CHC, grip strength, and self- reported health between 2012 to 2016 with a 5% threshold for statistical significance, two-tailed. We reported on findings for the whole cohort and then for women and men separately. Data from 2014 are included to demonstrate time trends only. Using multiple linear regression, we developed a model to predict Δ depression from 2012 to 2016 for men, women, and overall. Bivariate correlations identified those covariates significantly associated with Δ depression, with Spearman correlations being applied to analysis with one or more non-continuous variables, and Pearson correlations being used for the remainder. These, along with our primary predictor variables, were used as independent variables in Model 1. Only significant variables from Model 1 were then included in Model 2. ## Ethics statement The study was approved by all relevant university or health sciences ethics committees: Queen’s University (Kingston), Centre de Recherche du Centre Hospitalier de l’Université de Montreal (St. Hyacinthe), the Universidad de Caldas (Colombia), and the Universidade Federal de Rio Grande do Norte (Brazil). Written informed consent was obtained from all participants at recruitment (2012). # Results Visual inspection with histograms, normal Q-Q plot, and box plots showed that the majority of data were approximately normally distributed. Change in depression was within the normal range for males (skewness -0.260, SE 0.104, kurtosis 2.958, SE 0.209) and females (skewness -0.556, SE 0.099, kurtosis 2.689, SE 0.197). In 2012, the mean age of respondents was 68.97 and 52.9% were female. Income insufficiency varied by site, with Kingston and St. Hyacinthe having much lower rates of income insufficiency (2.8% and 6.1%) than Manizales and Natal (66.5% and 71.1%). CES-D scores in Latin America were higher (ie more depressed) than those in Canada at all stages. Overall CES-D scores decreased at all sites, although these site-specific findings were only significant for St. Hyacinthe and Manizales (*p*s \< 0.05). In bivariate analyses, decreases in depression were significantly correlated with higher 2012 CES-D (i.e. greater depression initially) (*p* \< 0.001). Respondents with insufficient income at baseline and those with increased SRH were also more likely to experience decreases in CES-D (*p*s \< 0.001). Age, sex, and changes in objective physical health (i.e. Δ CHC, Δ grip, and smoking) were not significantly associated with Δ depression (*p*s \> 0.05). To refine understanding of predictors of Δ depression, we next developed a linear regression model using as independent variables those that were significantly correlated with Δ depression in bivariate analyses: CES-D 2012, Δ SRH, and income sufficiency (Model 1). Only CES-D 2012 was highly correlated with Δ depression, with Δ SRH and income sufficiency having a small correlation with Δ depression. We also included Δ CHC and Δ grip as objective measures of physical health as these were of primary interest despite their lack of significance in bivariate analyses. Data were analyzed for the whole sample, and then separately for men and women. Greater depression in 2012 remained highly predictive of a decrease in depression over time in men, women, and overall (*p*s \< 0.001). Improved SRH also aligned with a decrease in depression overall and among women (*p*s \< 0.001) but not in men. Increases in CHC were associated with increases in depressive symptoms for men (*p* = 0.007) but not for women or among the cohort, overall (*p*s \> 0.05). With the inclusion of multiple predictors, income insufficiency ceased to be significant and Δ grip remained insignificant (*p*s \> 0.05). In our final model (Model 2), greater baseline CES-D and increases in SRH were associated with decreases in depression overall and in women, R² = 0.310 and R² = 0.325, respectively, whereas for men, higher 2012 CES-D scores and absence of increased chronic disease were the significant predictors, R² = 0.308. # Discussion As respondents aged, they reported fewer depressive symptoms despite declining physical health. This improvement in mood was consistent for men and women, and across high- and middle-income settings. The most consistent predictor of a decrease in depression was a higher initial CES-D. This could demonstrate nothing more than regression to the mean, however the consistency of the finding across settings and groups suggests something more. While objectively-measured physical health declined and number of chronic diseases increased, SRH increased or remained stable, implying that subjective assessments of one's health sometimes extend beyond a count of chronic diagnoses to include elements of mental well-being. Sex/gender differences emerged when men and women were examined separately. Improved SRH was associated with decreased depression among women, despite women having poorer objective physical health, overall. It is possible, therefore, that what we measure when we measure SRH may be gendered. Women, in particular, may take an inclusive view of subjective health, incorporating self-rated mental wellness along with their physical health. In men, increase in CHCs aligned with decreases in mood, which supports literature findings that depression in men is associated with mobility and functional impairment. In keeping with others, we found that women reported more depressive symptoms than men at any given time, but improvements in mood with aging were similar for all. Our findings suggest that mental health is fundamentally different from and unrelated to physical health in older adults although among women, SRH may include both. Our research is consistent with existing evidence in showing decreased depression and increased SRH over time among older adults, despite declines in objective physical health. However, unlike others who found a decline in depression in high-income and metropolitan, but not middle-income or rural settings, we found that depressive symptoms decreased in all settings, although middle-income sites had higher CES-D scores at any given time. Prior research was cross-sectional and may not have fully accounted for higher baseline rates of depression in middle-income countries. Surprisingly, while IMIAS respondents with insufficient income reported more depressive symptoms at baseline, they experienced greater improvements in mood over time than their economically better off counterparts. Our study’s longitudinal and international design provides new insight into the role of context and income sufficiency on mood. Because men and women across settings experienced fewer depressive symptoms over time, our results support the concept that older adults experience a fundamental shift in outlook as part of their aging experience despite increasing physical illness or socioeconomic position. Socioemotional selectivity theory (SST) postulates that the elderly prune memories to favour positive recollections, and this may explain why aging brings improved mood. According to SST, older adults may screen out negative thoughts (e.g. about declining physical health) in favour of positive ones (improved mood and self-rated health). Others speculate that elders may develop adaptive coping techniques including response shifts in order to maintain their well-being while facing physical, economic, and social stressors. This may help explain why those with insufficient income reported more depressive symptoms at baseline, but experienced greater improvements in mood with time. Our findings reinforce the concept that older adults shift their conceptualization of health risks, detaching these from a sense of successful aging and, in the process, experiencing an improvement in mood. Those with higher baseline depression scores seemed to experience the greatest shift, as they had larger decreases in depression over time. There is clinical importance to these findings. Physicians should not dismiss increasing depression in the elderly as an inevitable outcome of declining physical health. Rather, doctors should recognize that lower mood in older adults is a red flag for a serious yet treatable condition. In addition, physicians might put less emphasis on enumerating risks and numbers of chronic diseases, and instead help older patients build a sense of successful aging despite declining objective physical health. Our longitudinal study design with an international sample allows us to shed light on the pre-existing, conflicting research findings about depression and physical illness in older adults across different settings. While factors such as sex/gender, chronic illness, income insufficiency can contribute to a higher baseline prevalence of depression, older adults across diverse settings appear to have decreases in depressive symptoms as they age. ## Limitations Our study results must be interpreted within limitations: only community- dwelling respondents who passed a cognitive screening test were included. Older adults with dementia, and those in assisted living facilities have higher levels of depression, but as fewer than 1% of those 65–74 live in collective dwellings, their exclusion is unlikely to have dramatically altered outcomes. Our research is limited to those who participated in IMIAS across the four years of data collection and does not account for those who died or withdrew from IMIAS. We do not attempt to generalize to those with dementia or to those with the most severe illnesses. We did not control for medication use, and it is possible that antidepressants started between 2012 and 2016 may have affected associations between mood, CHC, and SRH, particularly for those with a higher burden of depressive symptoms at baseline. We used a single, albeit evidence-based indicator of physical function, that is, grip strength. The variation across sites in depression and other variables examined almost certainly speaks to beneficial or harmful social norms and values in each setting. However, we could not measure these and therefore can only speculate as to what they might be. Finally, IMIAS enrolled those aged 65–74 in 2012, and we cannot confidently generalize findings to the oldest old. # Conclusion Our longitudinal, international study demonstrates that decreased mood is not an inevitable consequence of physical illness in older adults. While risk factors such as sex/gender, chronic illness, and income insufficiency can contribute to a higher baseline prevalence of depressive symptoms, adults across middle and high income settings appear to have improvements in mood as they age. Health care providers treating older adults should not dismiss depression as an expected consequence of aging. Instead they might help patients develop a sense of successful aging despite increasing physical comorbidities. The authors would like to acknowledge the International Mobility in Aging Study participants and interviewers. [^1]: The authors have declared that no competing interests exist.

# Introduction Cytomegalovirus (CMV) infection is a frequent infectious complication in an immunocompromised host, as is the case in an organ transplant recipient who faces life-long immunosuppression. CMV infection can occur either from reactivation of recipient or donor CMV or *de novo* infection in the recipient after transplantation. The clinical spectrum ranges from asymptomatic to systemic infection and end-organ disease with pneumonitis, meningitis, esophagitis, and colitis. Furthermore, CMV infection also results in an increased host susceptibility to secondary infections and increased alloimmunity, leading to acute and chronic allograft rejection. Evaluation of peripheral mRNA changes from clinical and sub-clinical CMV primary infection and reactivation has provided insight into causal factors for allograft dysfunction secondary to CMV, host biological pathways actively altered by the virus, and the dynamic temporal changes of cell cycle, DNA damage molecules that occur during the evolution of CMV viral replication. Mass spectrometry-based proteomics of human serum and plasma has been used for identifying markers for disease diagnosis, and for the study of viral infections such as influenza, COVID-19, and HIV. Characterizing the host response to CMV infection will allow for an increased understanding biological events associated with CMV infection severity and will also provide insights into the negative impact of CMV infection on the transplanted organ. CMV infection drives alloimmune injury of the engrafted organ through heterologous immunity in an organ transplant recipient and is an important confounder for higher rates of graft failure after kidney transplantation. To address these key questions, we have utilized a highly characterized, propensity-matched cohort of kidney transplant patients with and without post-transplant CMV DNAemia, with serially collected samples at 3 months (pre-CMV infection) and 12 months post-transplant (post-CMV infection), as well as immediate post-CMV time points at 1 week and 1 month post CMV DNAemia and also utilized preexisting transcriptomic data on time-matching PBMCs from the same cohort. # Materials and methods ## Patient enrollment and patient characteristics Plasma was isolated from the peripheral blood that was collected from kidney transplant recipients enrolled and consented in an IRB-approved study, “New ways to test patients for Kidney and/or Pancreas transplant rejection (IRB#11–001387) at the University of California Los Angeles(UCLA). Patients provided informed written consent for the use of their samples and medical records used for this research study. The study only included adults and did not include any minors. Written informed consent was obtained by the physician enrolling the patient in the study. All study personnel completed CITI and HIPAA training. Patient samples and medical record data were anonymized before study personnel accessed them. Patients received induction with either anti-thymocyte globulin (ATG) or basiliximab depending on pre-transplant levels of sensitization and donor kidney quality, followed by protocolized immunosuppression with tacrolimus, mycophenolate mofetil, and prednisone, as previously described. CMV prevention was performed according to center protocols summarized as follows: 6 months of valganciclovir for high-risk donor-positive (D+) and recipient negative (R-) patients and 3 months of valganciclovir for intermediate risk recipient positive (R+) patients who received ATG induction. R+ patients who received basiliximab or low-risk (D-/R-) patients received acyclovir prophylaxis to prevent HSV and VZV infection. All study subjects underwent regular CMV PCR screening at 3-, 6- and 12-months post-transplantation to detect CMV DNA in peripheral blood, and further at 1 week and 1 month after detection of CMV viremia. ## Patient selection A matched cohort of patients with (n = 31) and without (n = 31) CMV DNAemia post-transplant (\>137 IU/ml), were identified by review of clinical records of 7,331 PNMCs collected from 536 kidney transplant recipients. 31 cases (i.e., patients with DNA-viremia 3–7 months posttransplant) were carefully selected based on having a ‘complete set’ of samples relative to the onset of viremia (specimens at baseline, 1-week, 1-month, and long-term). An additional 71 control recipients were identified without evidence of DNAemia at any time post- transplant. Next, 31 of the controls were successfully matched 1:1 by age, race, sex, and year of transplant using nearest neighbor propensity scores generated from logit regression models fit separately across 3 blocks (or strata) of patients generated according to induction-therapy practice and CMV-IgG status pre-transplant: (1) CMV+ patients who received Anti-Thymocyte Globulin (ATG), (2) CMV+ patients who did not receive ATG; and (3) pre-transplantation CMV- patients who did not receive ATG induction. The following blood samples were processed for proteomics: (1) Baseline (3-month post-transplant sample, after discontinuation of CMV prophylaxis, prior to DNAemia start) (2) One-week post- DNAemia (week 1, early post-CMV) (3) One-month post-DNAemia (month 1, intermediate post-CMV) (4) 12 month after transplantation (long-term post CMV). Four of the 31 patients that developed CMV DNAemia after transplantation were donor IgG seropositive, and recipient seronegative; the others were recipient IgG seropositive. For patients with multiple episodes of CMV DNAemia over 137 IU/ml, the first episode closest to transplantation was studied. Patients with a history of CMV DNAemia were matched on a 1:1 basis to a cohort of kidney transplant recipients without a history of CMV DNAemia via propensity scores estimated from a logit model with variables donor type (deceased versus living donor status), recipient, sex, race, and induction type in kidney transplant recipients who were either D+/R- or R+. Samples were selected for each control patient that corresponded in terms of time post-transplant with the baseline and long-term CMV DNAemia patient samples. Of these CMV DNAemia-negative patients, 24 were CMV seropositive and 7 were seronegative with CMV-positive donors. ## Targeted LC/MS-based proteomics The plasma samples were processed using a proteomics blotting methodology in a 96-well format. In brief, 1 μL of plasma (\~50 μg of protein) was added to 100 μL of urea buffer (8 M urea in 50 mM ammonium bicarbonate buffer (ABC). Following reduction using dithiothreitol (DTT, 50 mM in urea buffer) and alkylation of the cysteine side chains using iodoacetamide (IAA, 250 mM in urea buffer), 10–15 μg of proteins were loaded onto a 96-well plate with an activated and primed polyvinylidene fluoride (PVDF) membrane at the bottom (Millipore- Sigma). Proteins adsorbed to the membrane were trypsinized by incubation with the protease for 2 hours at 37°C. The resulting tryptic peptides were eluted off the membrane with 40% acetonitrile (ACN)/0.1% formic acid (FA). The peptides were subsequently desalted using a 96-well MACROSPIN C18 plate (TARGA, The NestGroup Inc.). Approximately 1 μg of tryptic peptides was analyzed using a Mikros Liquid Chromatography connected to LCMS 8060 triple quadrupole mass spectrometer (both: Shimadzu Corp.). The mass spectrometer was operated in MRM mode. The proteotypic peptides were selected and validated using pooled plasma samples. A total 629 unique peptides corresponding to 315 proteins were monitored in each LC/MS run with a total run time of 15 minutes per sample. Each peptide was monitored using 3 transitions. The raw data were exported into Skyline software (v20.2.1.315) for peak area and retention time refinement. The means of the peptide intensities were used for the different protein abundances, which were exported for further analysis. ## Data analysis Data preprocessing was conducted with Python (3.8.8), Pandas (1.2.0), Numpy (1.19.5), and Re (2.2.1). The data was set at a 90% limit for missing data. Any aberrated or missing values were imputed by the minimum value observed in a similar cohort member. Moreover, all raw abundance data was log2 normalized prior to any further analysis. All exploratory data analysis and descriptive statistical plots were generated by using Seaborn (0.11.1). Differential Expression Analysis was conducted by linear models by using R (4.0.4), limma (3.46.0) and ggplot2 (3.3.5). Three major comparisons were created to discern a CMV + signature: CMV+ Baseline samples vs all CMV- samples, CMV+ samples vs all CMV- samples, and CMV+ longitudinal samples vs CMV- longitudinal samples. Statistical significance was accepted with a p-value less than or equal to 0.05. Corresponding results were illustrated as volcano plots using ggplot2. Additionally, statistically significant proteins per each time point were summarized in a Venn diagram by using Python (3.8.8) & Matplotlib_venn (0.11.6). Multivariable analysis for clinical confounders revealed no bias for gender, race, ethnicity, age, donor Type, recipient age, induction immunosuppression, or CMV risk group based on recipient and donor CMV IgG status, and time post- transplant for the time of CMV reactivation or infection. A two-component linear discriminant analysis was used to distinguish the clustering of patient CMV status over time (Python 3.8.8, Sklearn: 0.24.1). Furthermore, a XGBoost machine learning model (Python 3.8.8, XGBoost: 1.4.1) was utilized to identify significant proteins that are associated with CMV DNAemia positivity. The data was normalized by standard scaling. The mean and standard deviation was computed by fit transformation (Sklearn: 0.24.1). Data was then split to preserve 30% data for testing. A param grid of learning_rate, min_child_weight, gamma, subsample, n_estimators, colsample_bytree, and max_depth was automated by using a randomized search cv. The random search was composed of 50 iterations, scoring based on roc_auc, 4 jobs, and a cross- validation composed of a stratified k fold search of 10 splits. Root Mean Square Error (RMSE, 0.345033), Prediction Accuracy (88.1%), CV score (85.21%), and AUC (96%) were used as metrics to observe the accuracy and validity of the model. Moreover, the top contributing features were extracted from the machine learning model and evaluated by the F score; resulting in a set of 37 unique proteins that was run through STRING DB (functional protein association networks) to determine the GO Biological Processes, Molecular Functions, and proteomic interactions. ## Proteome & transcriptome correlation We utilized a previously reported⁶ matching transcriptome data on PBMCs collected from the same cohort with matching time points to perform an integrative multiomic analysis. Data was downloaded from Gene Expression Omnibus (GSE168598), in which there were 153 overlapping time-matching samples present with an overlap of 186 gene/protein present in both transcriptome and proteome datasets. Transcriptome data was converted into log(cpm) analyzed using R (4.0.4), edgeR (3.32.1) and limma-voom transformation (3.46.0). Differential expression was then calculated for a baseline signature, CMV signature and longitudinal signature. The results were then analyzed for a Pearson correlation between all matching gene transcripts and proteins by using R’s corrplot (0.92). # Results and discussion ## Results ### Patient characteristics A summarization of patient demographics is provided in. As expected, due to the propensity matching design for patient selection, age, sex, race, and induction type were comparable between CMV DNAemia positive (≥ 137 IU/mL of CMV detected in a PCR test: abbreviated PCR+) and CMV DNAemia negative (PCR-) patients. A similar number of patients were high risk for CMV (D+/R-) as intermediate risk (R+) in both groups. Patients experiencing CMV DNAemia were PCR positive at a median of 80 days after transplantation (range: 10 to 561 days). Median peak viral load was 757 IU/ml (range 146 to 13900 IU/ml). Two CMV DNAemia-positive patients were diagnosed with clinical CMV disease, based on standard definitions. ### Plasma proteome provides CMV-DNAemia-specific protein profiles To compare protein changes associated with CMV DNAemia, we analyzed the identified plasma protein data from CMV DNAemia-positive patients and compared it with protein data from CMV DNAemia-negative patients. A total of 241 proteins were used for this analysis. To analyze distinct proteomic features among patients CMV DNAemia status using a two-component linear discriminant analysis was used to distinguish clustering of patient CMV DNAemia status over time. LDA separated samples based on CMV DNAemia status among CMV DNAemia positive and CMV DNAemia negative cohort. Within CMV DNAemia positive cohort, the samples were separated based on post-CMV DNAemia timepoint. Among samples collected from CMV DNAemia individuals, samples collected 1 week and 1-month post-DNAemia are interspersed, suggesting the signal of CMV infection persists until 1-month post-infection. ### Pre- CMV DNAemia perturbation in plasma proteome To investigate if there are changes in protein profiles pre-CMV infection, we analyzed plasma protein profiles generated from samples collected at pre-CMV infection time from CMV DNAemia positive and matching CMV DNAemia negative cohort as described in. The analysis resulted in 17 proteins whose levels were either increased (n = 6) or decreased (n = 11). The significance of the changes is demonstrated by the Volcano plot in. Relative fold increase and decrease and statistical significance are presented in a bubble plot (**).** The most significantly increased protein was Lysine Methyltransferase 2C (KMT2C) with a 3.38-fold increase (p = 0.05) in CMV DNAemia positive samples and the most significantly decreased protein was Immunoglobulin Lambda Variable 7–43 (IGLV7-43) with 2.17 fold decrease (p = 0.01). A complete list of proteins with fold change and P value is provided in. The top 3 biological processes enriched by the significant proteins were platelet degranulation (FDR, 4.83E-06), acute inflammatory response (FDR, 0.0018), blood coagulation (FDR, 0.0018). A complete list is provided in **.** ### CMV DNAemia-specific protein signature **T**he LDA analysis demonstrates that the plasma proteomic profile of patients at 1-week and 1-month post-CMV infection are similar and interspersed. Next, we compared plasma protein profiles for samples that were collected at 1-week and 1-month post-CMV DNAemia from CMV DNAemia positive cohort and compared them against proteins profiles generated from CMV DNAemia negative cohort that included Pre-DNAemia and 1-year timepoint. The analysis resulted in significant changes in 16 proteins, with an increase in their level with CMV infection (n = 11) and a decrease in their level with CMV infection (n = 5). Among the increased proteins mostly were immunoglobin complex components, IGLV9-49, IGLV3-10, IGLV3-19, IGKV1-5, IGHV3-49, IGHG1. Other were plasma proteins complement factor H related 1 (CFHR1), apolipoprotein F (APOF), Lipopolysaccharide Binding Protein (LBP), inter-alpha-trypsin inhibitor heavy chain 3 (ITIH3), and ceruloplasmin (CP), decreased proteins were immune complex components, IGHG2, IGLV1-36, complement factor P (CFP), kallikrein B1 (KLKB1), CFP (complement factor properdin), and transaldolase (TALDO1). Changes in individual proteins are presented as a Volcano Plot. This demonstrated a rise in the level of a set of plasma proteins immediately after CMV DNAemia. We also analyzed proteomics data to see long-term proteomic changes 1 year after DNAemia by comparing protein profiles of 12-month post-DNAemia time points in between positive and negative DNAemia cohorts. As expected, the signature of CMV at the time of DNAemia was lost. We identified only 4 proteins, namely, IGLV3-19 (1.7-fold increase, p = 0.003), IGHG3 (1.5 fold increase, p = 0.02), PSMB4 (1.3-fold increase, p = 0.03), CFHR5 (1.8 fold increase, p = 0.03) were increased in CMV DNAemia positive cohort indicating a long-term residual impact of CMV DNAemia in kidney transplant recipients. ### Temporal proteome changes due to CMV DNAemia We performed an analysis of the proteomic dataset for quantitative alteration of plasma proteins that correspond to CMV DNAemia. Protein profiles of baseline samples of CMV DNAemia-positive patients were compared with protein profiles of 1-week post-infection samples which resulted in ten significantly changed proteins. The significant proteins expression direction and significance is presented with a volcano plot. Increased proteins at the time of CMV DNAemia positive infection (1-week post-DNAemia) included Serpin Family A Member 12 (SERPINA12) with P value 0.01 and 2.47 fold increase and Immunoglobulin Heavy Variable 3–72 (IGHV3-72) with P value 0.02 and 1.65 fold increase. Transthyretin (TTR) and Lysine Methyltransferase 2C (KMT2C). The proteins were involved in the enrichment of biological processes such as complement activation (FDR = 0.03), Humoral immune response (FDR = 0.01), and innate immune response (FDR = 0.01). ### CMV DNAemia induced proteomic signatures A summary of protein changes due to CMV DNAemia is presented in **,** listing plasma proteomic differences in the CMV DNAemic cohort pre-, immediately post- and 1 year post -DNAemia. Pre-CMV DNAemia samples showed an increase in serine protease inhibitors SERPINA3 and SERPING1 and an enrichment of proteins involved in acute inflammatory responses, such as Serum amyloid protein A (SAA1). Protein profiles immediately after CMV DNAemia showed higher amounts of immunoglobins and immunoglobulin complex components; ceruloplasmin (CP), vaspin (SERPINA12), Lipopolysaccharide-binding protein (LBP), C4b-binding protein alpha chain (C4BPA), Complement Factor H (CFH) and immune complex protein (IGHV3-72) were significantly increased in 1-week post-CMV DNAemia. The increased levels of these proteins were significantly lower in both 1-mo- and 1-yr-post-DNAemia plasma samples. Using XGBoost, a decision-tree-based ensemble machine learning algorithm that uses a gradient boosting framework, we developed a classifier that provided an accuracy of 88%, the sensitivity of 85%, specificity of 91%, and a ROC AUC of 96% in identifying patients with CMV DNAemia. This classifier also identified a subset of 37 proteins **** that contributed to distinguishing CMV DNAemia ****. Furthermore, we used feature selection with the Boruta algorithm. The algorithm ran at 800,000 iterations was integrated to detect the top essential protein features. This process reduced the 241-protein list down to 5 proteins and increased AUC by 1%. This set of proteins included HBG1, KMT2C, IGLV7-43, IGKV3D-15, and CP **.** When tested for its performance on 41 samples with 21 CMV DNAemia positive and 20 CMV DNAemia negative status. The validation of the panel of proteins to identify CMV DNAmeia-positive status irrespective of post CMV DNAemia time distinguished CMV positives with sensitivity at 80%, and specificity at 95%, **** along with a ROC AUC of 0.97. ### Comparing plasma proteome with the transcriptome of PBMCs in matching blood samples We performed a comparative multiome analysis of proteomics data from the plasma samples and transcriptome data available from the matching PBMCs from the same cohort, which has been analyzed separately in a previously published paper⁶. Given two different milieus interrogated by transcriptomics (PBMCs) and proteomics (plasma), a significant overlap of target genes/proteins was not expected. Across all gene expression data and protein data, the average correlation for the most significant targets at the time of CMV DNAemia was 0.11, with the highest observed correlation between proteome and transcriptome seen in immunoglobulin heavy constant gamma 1 (IGHG1) (Pearson, r = 0.44). There are six unique readouts when observing the differentially expressed CMV DNAemia positive signature between transcriptome and proteome. Among the six genes and proteins existing in the CMV signature, Orosomucoid 1 (ORM1) holds a positive correlation (Pearson, r = 0.17). ORM1 gene encodes a key acute phase plasma protein which is also known to be decreased in blood samples infected with SARS- CoV-2. # Discussion CMV modulates host immunity through several sophisticated strategies to achieve persistence in infected individuals. One of the mechanisms of modulation is reported to be the deployment of proteins to target host factors for degradation and suppression on host immunity against the virus in an *in vitro* model of CMV infection. The virus has an impact on host gene and protein expression. We recently reported the establishment of subset of prememory-like NK cells expressing NKG2C and lacking Fc epsilon RIgamma increased during viremia in CMV viremic patients. An interactome study of CMV-host and virus-virus protein interactions identifies multiple degradation hubs. These virus-host protein interactions enable the virus to evade host-immunity. These cited works use either peripheral blood cells or in vitro cell system to study CMV’s role in CMV disease in humans. We performed proteomic analysis of plasma samples using blotting methodology and made significant observations about plasma protein markers and pathways associated with CMV DNAemia. This is first of its kind of studies that has been carried out in plasma samples from kidney transplant patients with and without DNAemia. Dimensionality reduction with LDA based on 241 proteins identified in plasma of transplant patients successfully differentiated patients with and without CMV DNAemia. Furthermore, LDA separated patients into were groups based on post-CMV DNAemia timepoint. This highlights the fact that CMV induces changes in plasma protein composition that is unique to CMV DNAemia when compared to CMV DNAemia negative cohort. In addition, we also observed that plasma samples collected 1 week and 1-month post-DNAemia share proteomic composition in plasma suggesting the signal of CMV infection persists until 1-month post-infection. Next, we identified CMV-specific proteins and biological pathways in pre-DNAemia plasma. We identified a set of 17 proteins (6 increased and 11 decreased) in pre-CMV DNAemia plasma samples which were enriched in biological processes such as platelet degranulation, acute inflammatory response, and blood coagulation. The most significantly increased protein was Lysine Methyltransferase 2C (KMT2C) is a potential tumor suppressor but has not been reported to be associated with CMV and is a novel finding. Interestingly, expression of serum amyloid A1 (SAA1) protein was significantly elevated pre-DNAemia in patients that progressed to CMV DNAemia. SAA is an acute phase protein with multiple immunological functions involved lipid metabolism and inflammation. In a cohort of lung transplant recipients, SAA concentrations were found to be significantly elevated in acute rejection and infection but not in stable transplants suggesting that SAA1 elevation is not an intrinsic response to graft. Further evaluation of SAA1 levels in kidney transplant patients will provide validation of these results. These observations are intriguing and have the potential to lead to early markers for susceptibility for CMV infection among kidney transplant cohort. Earlier study profiling the transcriptional changes associated with CMV DNAemia did not detect any differences in pre-viremia signature of patients that went on to develop CMV DNAemia versus patients that did not. This further underscore the importance of circulating proteins in the blood can provide crucial information about the immune status. Further validation of these proteins as early markers of CMV risk will provide better strategies in managing CMV infection in kidney transplantation. It is known that CMV modulates host cells, downregulating and degrading hundreds of proteins. Differential expression of proteins has been reported to control innate immunity linked to host response to viral infection. Using in vitro system, a previously published report listed 71 human proteins and 12 proteins encoded by known viral open reading frames (ORFs). For the first time, we identified plasma proteins whose level is significantly altered with CMV DNAemia. Increased immunoglobin complex components, IGLV9-49, IGLV3-10, IGLV3-19, IGKV1-5, IGHV3-49, IGHG1, plasma proteins such as complement factor H related 1 (CFHR1), apolipoprotein F (APOF), Lipopolysaccharide Binding Protein (LBP), inter-alpha-trypsin inhibitor heavy chain 3 (ITIH3), and ceruloplasmin (CP) provides information on perturbation on plasma proteomic landscape due to CMV DNAemia. Given, the different outcomes of CMV infection such as latent infection, subclinical infection, active infection, and CMV disease, these plasma proteins offer potential biomarkers to stratify CMV infection types. Especially, subclinical infection of CMV is of concern, and it would be interesting to explore that aspect of the impact of latent infection of other viruses. Despite, the study’s first-of-its-kind nature, we acknowledge certain limitations. This study was done on samples collected from a single center. We do not know the impact of antiviral therapy on post-DNAemia protein profiles and the majority of patients have a history of CMV infection which may have contributed to the plasma protein profile as reactivation of CMV instead of primary infection. In summary, we believe the data presented provides insight into CMV DNAemia-induced biological perturbations and potential protein markers for CMV infection. Further larger-scale follow-up studies will validate mechanisms and identify biomarkers for CMV disease and infection using assays such as ELISA. # Supporting information This study was supported by a National Institutes of Health grant U19 AI128913 (PIs: E.R. and M.S.). CMV Systems Immunobiology Group (in alphabetical order): Richard Ahn, Janice Arakawa-Hoyt, Patrick Boada, Jenny Brook, Suphamai Bunnapradist, Jim Cimino, Izabella Damm, Nakul Datta, Mario Deng, Don Diamond, Tin Doung, David Elashoff, Janette Gadzhyan, David Gjertson, Alexander Hoffman, Kenichi Ishiyama, Maggie Kerwin, Lewis Lanier, Megan Llamas, Erik Lum, Dane Munar, Rajesh Parmar, Harry Pickering, Zach Qian, Priyanka Rashmi, Elaine F. Reed, Maura Rossetti, Dmitry Rychkov, Minnie Sarwal, Joanna Schaenman, Subha Sen, Tara Sigdel, Danielle Sim, Marina Sirota, Jun Shoji, Angela Sun, Swastika Sur, Parhom Towfighi, Flavio Vicenti, Otto Yang. [^1]: The authors have declared that no competing interests exist. [^2]: ¶ the complete membership of the author group can be found in the Acknowledgments.

# Introduction Encoding of viewed third party interactions appears to be automatic and can influence cognitive processes such as attention, working memory and longer-term memory. Of particular note, the encoding of social interactions appears to be fast and automatic in that during a visual search task detection of two people facing towards each other is faster than for two people who ignore each other (see also): a social priority effect. It was argued that the detection of which people in a scene are interacting in this way is a critical early stage to quickly interpret the social information. According to this view, such social binding processes provide a basic visual input for more sophisticated processes, such as detection of deception, social affiliation, social dominance, and more generally a forward modelling ability to predict potential future actions. Indeed, in our initial research programme investigating the form of representation mediating these social binding effects, we provided evidence for the role of representations of social interactions rather than low-level perceptual features driving the behaviour. This evidence was harvested from a series of studies examining visual search, working memory and longer-term memory. In that paper we determined that “*…results are consistent with the social binding hypothesis*, *and alternative explanations based on low level perceptual features and attentional effects are ruled out*. *We conclude that automatic midlevel grouping processes bind individuals into groups on the basis of their perceived interaction*”, pp 1251). More broadly, it is now well established that visual search processes are influenced by prior learning of the emotional properties of a stimulus, and such attention capture effects cannot be explained by low-level physical properties of a stimulus. Such studies have examined a wide range of situations from electric shock to negative social feedback (e.g., and for review). A central feature of such social interactions is of course joint attention via gaze direction (see for review). That is, when two people are interacting they are typically attending to one another. Hence it is possible that attention processes evoked by the direction of gaze are contributing to the assessment of whether or not they are interacting, and consequently to the social binding priority effect. However, a critical issue concerns whether general attention orienting mechanisms are necessary and sufficient to account for the social binding priority effect (, see also for review) or whether processes involving the representation of social agents is critical. One approach to this issue is to examine effects with stimuli that do not have the mental states that mediate social interactions but nevertheless orient attention. Arrows have been shown to possess these properties. For example, arrows produce attention orienting effects that are very similar to the social attention cues of a person’s gaze direction. Clearly, arrows are not biological stimuli with social intent and do not act in an interactive manner. Rather, arrow-like stimuli have intrinsic low-level physical properties that imply direction. The basic visual properties of direction can be seen, for example, in the shape of the arrow launched from a bow, the stream-lined shape of the fighter jet or sports car, or the body shape of the diving gannet. In each case the pointing shape exists to facilitate movement in a particular direction. Hence simple visual features, such as those possessed by arrows, imply direction. If attention orienting by simple physical cues (independent from implied social interactions) is central to the previously observed effects then a clear prediction is that towards facing arrows will be detected faster than away facing arrows. Indeed, recent work by Vestner et al. has demonstrated that this is the case. That is, just as towards facing people are detected faster than those facing away from one another (see), so also towards facing arrows are detected faster than away facing arrows. That such simple stimuli with non- social properties can produce the same effects challenges prior claims that they play no role in the effects observed by Vestner et al.. However, though similar effects are obtained with low-level visual stimuli not containing social information, this does <u>not</u> demonstrate that the latter higher-level processes play no role in social binding processes. This is especially the case if effects are at ceiling. The brain represents multiple properties of the visual world, from lower-level features such as simple shape, colour, proximity and motion to higher-level representations of object identity, emotion and social properties. Indeed, Ristic et al. found that gaze and arrow cues are managed by separate systems to the same functional outcome. The parallel co-existence of multiple forms of representation across cortical and subcortical networks leads to the possibility that the effects of some internal representations, not observable in some tasks, might be detected in other situations. Therefore further work employing converging techniques is necessary. In the current series of experiments we further investigate the low-level visual features that influence visual search, but also simultaneously manipulate higher-level representations of third-party social interactions that can be congruent or incongruent with low-level properties. Such an approach avoids the interpretational issues caused by potential floor or ceiling effects. ## Our approach In our new studies we manipulate these two forms of processing within the same experimental stimuli. Consider, which shows examples of the displays employed in Experiments 1 (A&B) and 2 (C&D). These teardrop stimuli have low-level basic features that imply direction. That is, in a basic Posner cueing task targets presented to the pointed side of the teardrop stimulus are detected significantly faster than targets presented to the round side of the stimulus. Hence in the visual search task featuring such stimuli, target detection will be faster when stimuli point towards rather than away due to attention being jointly focused to one location, replicating Vestner et al. with these new stimuli. However, prior to the search task we employ a social learning stage where participants are presented with Heider & Simmel type video displays, adapted from those of Over & Carpenter. Hereafter these videos are described as ‘social priming videos’. In such displays, stimuli move in particular spatial patterns that evoke a powerful experience of complex social interactions, such as inclusion/exclusion, teasing, and cooperation/competition, and even characters with particular personalities. This is a potent technique which appears to evoke universals in social perception based on object interactions that are similar across cultures, develop early and can reveal individual differences in social perception (e.g., in autism see). Furthermore, observation of such displays showing positive or negative social interactions can influence basic perceptual processes such as judgments of distance and they can prime particular states that can influence behaviour at a later time. In these displays, we manipulate the orientation and direction of motion of the socially interacting objects. For example, in one condition during the pro- and anti-social interactions of the objects, the pointing end is equivalent to the face/front end. However, in a second condition the stimuli are reversed such that the round end is equivalent to the face/front end during the social interactions. A critical features of our displays is the direction of motion. Hernik et al have shown that motion direction can reliably disambiguate the front acting part of an object from it’s back. Furthermore, classic studies from ethology show that direction of motion can influence the identity of an ambiguous stimulus. For example, the famous hawk/goose stimulus is perceived as a hawk when the short end is the movement direction, and as a goose when the long end is the motion direction. This direction of motion of the goose/hawk silhouette can determine whether predator avoidance responses are evoked in a number of bird species. A second important feature in our video displays beyond motion direction is the interaction with other objects. The side of a stimulus that interacts with and has an influence on other objects is perceived to be the active action end of the object. A final aspect of our motion stimuli is that they contain no intrinsic visual features that could be construed as a face. Previous work has shown that features to one side of an object can introduce directionality to an object representation, possibly evoked by attention capture by the features. Hence to avoid this potential confound we have employed very simple objects with no salient intrinsic face-like features. Via this approach we attempt to manipulate the interpretation of exactly the same stimuli via prior social priming. provides a schematic example of the motion displays where the round end is perceived as the active “face” end, but we recommend viewing the actual videos at osf.io/qxk8z. Because observation of such third-party interactive displays evokes a powerful sense of agency, personality, and the achievement of social goals they should create internal representations of individuals that socially interact. Certainly, looking ahead, our displays are sufficiently potent to evoke a face/front end of targets in the overwhelming majority of participants. With this in mind, the search performance of the participants who have observed the round end of the interacting objects as the “face” might be influenced in a subsequent visual search task. There are three data patterns that identify the roles of the two processes of low-level visual direction features and higher- level social interaction: First, if only low-level stimulus feature-based attention processes are at play, then independent of what participants observe in the prior social priming videos, detection of targets in (point inwards) will always be faster than the targets of (round inwards). Second, internal representations of socially interacting individuals might dominate low-level visual features. If so, search performance will be driven by the participant’s experience of the prior social priming videos. Detection of targets in (point inwards) would be faster than (round inwards) when the pointed end of the objects are represented as “face”; whereas the opposite pattern will be observed in participants who observe the round end as “face” in the social priming video. Finally, it is possible that both the low-level visual feature processes and higher-level social interaction processes will influence search simultaneously. In this case, there will be an interaction between the social priming condition and the point towards vs point away search effect where the largest effect will be observed after priming of social interactions with the pointed end as “face” due to the combined facilitation effects of low-level and higher-level stimulus properties. In contrast, after social priming where the round end is “face” there will be competition between low-level properties of pointedness and higher-level social representations, reducing the point towards vs away effect. Because this is a somewhat complex article containing 7 experiments utilising different techniques to examine the roles of both low-level visual features and higher-level social representations, we felt it would facilitate comprehension to preview our findings of each experiment at this point. We found that attention orienting visual features (Experiment 1 and 4) dominated visual search for target pairs even following social priming (Experiment 2) and social priming with semantic labelling and biological animacy (Experiments 3 and 5). Further, even when using targets which lacked low-level attention orienting shape features (Experiment 6) social priming with semantic labelling and biological animacy was insufficient to orient attention in a visual search task (Experiment 7). With regard to rapid visual search for interacting pairs, this consistent pattern of findings provides evidence for attention orienting by visual features, and no evidence for attention orienting by social representation. # Experiment 1 (‘Teardrop’ attention pre-test) Because the teardrop stimuli to be presented in the social priming video are novel and have not been used before, we have to first demonstrate their ability to automatically orient attention. That is, unlike gaze and arrows, which have been extensively studied and produce robust automatic shifts of attention, it is possible that the teardrop stimuli we employ will not have these basic low-level orienting properties. Therefore in Experiment 1, we present the teardrop in a Posner cueing design. The teardrop stimulus is presented in the centre of the screen and a target is presented either to its left and right. Because this cue is irrelevant to the participant’s task, and it does not predict target location, we can examine whether it evokes automatic attention orienting responses. We predict that the basic visual property of the pointed end of the object will shift attention to that side of space, just as gaze and arrow cues do. ## Method ### Apparatus The experiment was built and hosted in Gorilla Experiment Builder ([www.gorilla.sc](http://www.gorilla.sc/),). Browsers were restricted to Chrome, Firefox, Safari, Edge and Internet Explorer. Devices were restricted to desktop and laptop computers. ### Design Participants completed a practice block and a task block. Before each block, participants were shown instructions on the screen. Verbatim copies of the instructions given to participants are available at osf.io/qxk8z. At the start of a trial, two identical boxes appeared to the left and right of screen centre. After 500ms the cue would appear in the centre and then either 200 or 600 ms later a target (a cross) or a distractor (a circle) would appear in either the left or right box. If the target cross appeared then participants were to press ‘F’ if it was on the left or ‘J’ if it was on the right. However, if the distractor circle appeared then participants were to not press anything and should simply wait for the end of the trial. Participants were not instructed on which fingers to use for the task though in pilot testing all participants used their left index finger for ‘F’ (left side of the keyboard) and right index finger for ‘J’ (right side of the keyboard) on a QWERTY keyboard. If an incorrect response was made (wrong side reported for a cross or any response to a circle) then a ‘thumbs down’ appeared over the cue for 500 ms before the trial ended. Target trials ended when a response was made. Distractor trials ended when either a response was made or 2000 after distractor appearance. At the end of a trial the screen was blank. See. Cues in the practice block were black dots (lacking directional attention cue) whereas cues in the task block were teardrop shapes (i.e. possessing directional attentional cue, see) in purple or red. Cued trials are those in which the cue pointed to the same side that the target/distractor would appear, and uncued trials are those in which the cue pointed to the opposite side that the target/distractor would appear. The point direction is defined by the acute end of the cue (i.e. as a standard arrow). The target circle had the same height and width dimensions as the target cross. All stimuli and stimulus size details are available at osf.io/qxk8z. The practice block had 12 trials. Half had a stimulus onset asynchrony (SOA) of 200 ms and half had a SOA of 600 ms. Of each set of 6 there were 2 leftwards and 2 rightwards target trials, and 1 leftwards and 1 rightwards distractor trial. The task block contained 112 trials. Half had a SOA of 200 ms and half had a SOA of 600 ms. Of each set of 56 there were 12 cued leftwards target trials, 12 uncued leftwards target trials, 12 cued rightwards target trials, 12 uncued rightwards target trials, 2 cued leftwards distractor trials, 2 uncued leftwards distractor trials, 2 cued rightwards distractor trials, and 2 uncued rightwards distractor trials. ### Participants & analysis Protocols were approved by the University of York’s Psychology Departmental Ethics Committee and were in accord with the tenets of the Declaration of Helsinki. Participants were recruited through Prolific with filters of age from 18 to 50 and vision as normal or corrected to normal. Thirty-two participants were tested but 26 remained following exclusions (see Data exclusion and analysis). No participant completed more than one experiment in this programme of experiments. Informed consent was obtained prior to participation. Data was collected on the 12^th and 13^th of December 2019. Bayesian analysis was planned for all experiments in this manuscript using JASP v0.13. Participant N thresholds are not necessary to interpret Bayes Factors (BFs) which indicate the weight of evidence in favour of, or against, the null/alternative hypothesis (specified in each model). As such, power analyses were not performed and our sample size of 26 was based on typical simple attention cueing and visual search designs. We report evidence categories after each BF to aid interpretation of values e.g. (BF₁₀ = 3.005e+7 \[*extreme evidence for H1*\],…). ### Data exclusion Full exclusion details can be found at osf.io/qxk8z. Briefly: two participants were excluded due to error rate (\>25% errors in either the target (i.e. \>24) or distractor (i.e. \>3) trials); and 4 participants were excluded due to too few remaining trials following RT exclusion (\<75% in any SOA × cue condition). The mean ± SD percentage of trials remaining in each SOA × cue for each participant 94.4 ± 6.2%. ## Results & discussion Reaction times are shown in. Bayesian repeated measures ANOVA on RTs with within subject factors of SOA (200/600 ms) and Cue (cued/uncued) support a model including both main terms (BF₁₀ = 3.005e+7 \[*extreme evidence for H1*\], *p*(H1\|Data) =.603; SOA BF_incl. = 1.376e+7; Cue BF_incl. = 4.493). Reaction times were shorter for the 600 ms SOA than for the 200 ms SOA, and were shorter when the target was cued (point towards the target) than when the target was uncued (point away from the target). Frequentist modelling supports these findings. Models at osf.io/qxk8z. The results are clear and demonstrate that these teardrop stimuli do indeed produce automatic attention orienting. That is, even though the cue is non- predictive of target location and to be ignored, it shifts attention rapidly (within 200ms) and the attention shift remains stable until at least 600ms. Therefore, after featuring in social priming videos, these teardrop stimuli will have the basic low-level visual properties that evoke attention shifts and higher-level priming of third-party interactions. # Experiment 2 (‘Teardrop’ social priming) This experiment attempts to manipulate the properties of the teardrop objects from Experiment 1 by featuring those objects in social priming videos (adapted from Over & Carpenter) in which they are observed interacting in dynamic and complex spatial movement patterns. These patterns of motion evoke a salient and powerful impression of third-party social interactions, which display complex social states such as cooperation, in- vs out-group structure, and finally rejection and sadness. ## Method ### Apparatus Participants sat at a table in a dimmed room facing a 23" touch screen monitor (Iiyama (Tokyo, Japan) ProLite T2735MSC-B2, 1920×1080 pixels) at approximately 50 cm distance. A keyboard was positioned on the table between the participant and the screen. Participants and the keyboard spacebar were positioned at the screen’s horizontal centre. Stimulus presentation (60Hz) and response recording were achieved using custom scripts and Psychtoolbox 3.0.11 operating within Matlab 2018a (The MathWorks Inc., Natick, USA) on a PC (Dell (Round Rock, USA) XPS, Intel (R) Core (TM) i5-4430, 3 GHz CPU, 12 GB RAM, 64 bit Windows 10 Enterprise). ### Experiment design Participants completed a practice block, then four task blocks, and finally a target orientation question. In the practice and task blocks participants searched for and then touch the target pair in a set of four simultaneously presented pairs. Participants were shown the target pair on an instruction screen before each of those blocks. The target pair changed between blocks and was not present on every trial. The target orientation question was to check that our priming videos had effectively demonstrated the face (agency/attention) direction of targets. ### Practice & task trial composition At the start of a trial a cross divided the screen into four sections indicating that the participant should press and hold the space bar. Pressing the spacebar caused a pair of characters to appear in each section. Participants searched for the target pair and, upon discovery, released the space bar and reached out to tap that pair with the same finger that was pressing the spacebar. Releasing the space bar caused all four pairs to disappear so the target had to be identified before beginning a reaching action. If the target pair was not present (i.e. all four pairs were distractors) then the participant should keep holding the space bar down until the end of the trial. A trial ended either when the screen was tapped or 5s after the space bar was first pressed. Reaction times were measured from the moment the four simultaneously presented pairs appeared to the moment of space bar release. Movement time is measured from the moment of space bar release to the moment of screen contact. ### Target and distractor pairs Each of the four pairs presented on a trial was made-up of two shapes that possessed directional attentional cues. The shapes in the target pair either cued inwards or outwards whereas the shapes in the other pairs (the distractor pairs) always cued both leftwards or both rights. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Practice block The practice block was to familiarise participants with the task. Participants were instructed to “find and tap the target pair as quickly as possible” but if the target pair was not present (i.e. only distractor pairs were present) then they were to hold down the space bar until the end of the trial. Practice block instructions were presented on-screen and verbally by the experimenter. Participants had the opportunity to ask questions before starting the task blocks which they would complete in isolation. Verbatim copies of the instructions given to participants are available at osf.io/qxk8z. ### Task blocks and social priming Participants were shown a video prime before each task block. This 55s Heider & Simmel (1944) type video was adapted from those of Over & Carpenter (2009). It featured the purple and red teardrop shapes from Experiment 1 as well as a yellow teardrop shape. The purple and red shapes were seen playing and acting jointly to exclude the yellow shape, which was trying to play with them (see, and osf.io/qxk8z for full video). The intent was to prime participants to understand that either the pointed or rounded end (see Conditions) of the shapes was their “face”. The pointed and rounded videos were identical apart from the orientation of the teardrop which was mirrored to create a pointed or rounded prime. The prime for pointed or round alternated between participants. ### Conditions Every block contained standard trials (in which one of the four pairs was the target pair) and catch trials (in which all four pairs were distractor pairs). The orientation of target pair (pointing inwards or outwards) for the practice block was determined randomly between participants. The orientation of the target pair for the 4 task blocks was alternated in an A-B-A-B or B-A-B-A pattern as participants completed the experiment. For example, participant 1 would receive the target pairs \[inwards-outwards-inwards-outwards\] across blocks then participants 2 would receive the target pairs \[outwards-inwards- outwards-inwards\] across blocks. On a catch trial, pairs would all point leftwards or all point rightwards. On a standard trial, one pair would be in the target orientation and the other pairs would all point leftwards or all point rightwards. Within each pair the targets would be in the colours they had seen in the social priming video. Two left hand targets from different pairs would be one colour and two would be the other colour. All colours and distractor pair orientations were counterbalanced within task blocks. The section of the screen in which the target would appear was randomised between trials. The practice block contained 6 trials. There were two catch trials. In one catch trial all targets pointed leftwards and in the other all targets pointed rightwards. For the remaining four trials the orientation of the targets (shown in the instructions) was determined randomly between participants. The task blocks each contained 24 trials. There were four catch trials. In two catch trials all targets pointed leftwards and in the others all targets pointed rightwards. For the remaining twenty trials the orientation of the targets was determined by block number. Half of these trials had distractors pointing leftwards and half pointing rightwards. ### Target orientation question After completing the final task block participants were presented with a single question about target orientation. Participants were presented with two pairs of targets in the centre of the screen and asked to tap on the pair that was facing each other. The targets were the same size and shape as those in the practice block but were white rather than red or purple. ### Participants Protocols were approved by the University of York’s Psychology Departmental Ethics Committee and were in accord with the tenets of the Declaration of Helsinki. Participants were recruited through the University of York’s Psychology Department participant recruitment system. Informed consent was obtained prior to participation. For the pointy social prime 33 participants were tested and 26 (age mean±SD = 21.7±9.1, 6 male, 1 undisclosed) remained following exclusions. For the round social prime 31 participants were tested and 26 (age mean±SD = 19.8±2.4, 4 male) remained following exclusions. Note—the data in this experiment (and the other social/identity priming experiments: Experiments 2, 3, 5, 7) are from participants recruited and tested ‘in-lab’ whereas the data from pre-test experiments (Experiments 1, 4, and 6) are from participants recruited and tested ‘on-line’. ### Data exclusion & analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: 10 participants were excluded due to error rate (errors on \>20% of trials); 1 participant was excluded for using two hands; and 1 participant was excluded due to too few remaining trials following reaction time and movement time exclusion (\<75% trials). The mean ± SD percentage of trials remaining in each condition for each participant was 98.4±3.2 in the pointy face prime condition and 97.4±4.0 in the round face prime condition. Following all exclusion there were an equal number (n = 13) of participants in the A-B-A-B and B-A-B-A designs of each prime condition. Movement time is not considered a principle indicator though analysis of movement time is provided for all visual search experiments (Experiments 2, 3, 5 and 7) at osf.io/qxk8z for completeness. ## Results & discussion Bayesian repeated measures ANOVA on RTs with a within-subjects factor of target orientation (point inwards/point outwards) and a between-subjects factor of social prime type (pointed face/rounded face) support a model including only the target orientation term (BF₁₀ = 391375.418 \[*extreme evidence for H1*\], *p*(H1\|Data) =.697; BF_incl. = 293804.334). Reaction times were faster when detecting inwards pointing targets regardless of the prime. In the target orientation task (at the end of the experiment where participants were required to report which way the objects faced) Bayesian binomial tests (test value = 0.5) indicated that the majority of participants were able to identify the facing pair in both the rounded (24/26 participants, BF₊₀ = 15295.424 \[*extreme evidence*\]) and pointed prime conditions (19/26 participants, BF₊₀ = 7.485 \[*moderate evidence*\]). Frequentist modelling supports all findings. All models at osf.io/qxk8z. This experiment has demonstrated quite clearly that 1) our priming technique were effective, and 2) that low-level properties of attentional orienting are computed and can guide visual search. That is, just as with towards facing faces or arrows, we find the attention orienting direction of teardrop stimuli facilitates target detection. Importantly, we find the prior exposure to the video displays did not influence the subsequent visual search performance. That is, whether participants observed social interactions where the pointed end of the objects were perceived as the “face” or the round end as the “face”, had no effect on target search performance. However, clearly, to propose that high-level third-party social interaction relationships play no role in the visual search performance would be premature based on one experiment. It might be the case that although the video priming technique appears to have influenced object perception when tested at the end of the experiment, it is possible that during visual search such representations are less salient. Therefore Experiment 3 is a replication, but with a more compelling approach to boost the priming effects of the social interaction videos. Previous research has clearly demonstrated attention cueing effects with cartoon characters, avatars and animals. Hence we decided to increase the potency of object identities by first introducing the teardrops as seals and show some initial interactions before showing the social priming video and completing the search task of the present experiment. Providing semantic labels to the objects might further increase the sense of biological animacy when observing the third-party interactions. # Experiment 3 (‘Seal’ social priming) ## Method ### Apparatus The apparatus was identical to that of Experiment 2. ### Design The design was identical to that of Experiment 2 with two exceptions. First, just prior to viewing the social priming video a further priming block provided the stimuli with the semantic identity–the teardrop shapes were described as seals. Second, participants were never cued to view the pointed end as the “face” of the teardrop shape. The new priming block was intended to give a much stronger cue to view the rounded end of the teardrop as the “face”. At the start of the priming block, participants were told that this experiment was about seals. They were shown a modified target shape as a cue and then shown what the seals would look like in this experiment. Next they were shown three videos in which a lone seal moved, two seals greeted each other, and three seals were dyed from white to red, yellow and purple. This last video was to colour the white target shapes to match those of Experiment 2 thus allowing an identical task block presentation. Immediately after this further semantic priming, participants observed the same social priming video of Experiment 2, and these latter videos were then observed before every visual search block. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Conditions Conditions were identical to Experiment 2 but participants never saw the pointed front prime version of the social priming video illustrated in i.e. only the round end (congruent with the seal head) was ever primed as the “face”. ### Participants Protocols were approved by the University of York’s Psychology Departmental Ethics Committee and were in accord with the tenets of the Declaration of Helsinki. Participants were recruited through the University of York’s Psychology Department participant recruitment system. Informed consent was obtained prior to participation. Twenty-nine participants were tested but 26 (age mean±SD = 19.6±3.5, 2 male) remained following exclusions (see Data exclusion and analysis). ### Data exclusion and analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: two participants were excluded due to error rate (errors on \>20% of trials); and 1 participant was excluded due to few remaining trials following RT and MT exclusion (\<75% trials). The mean ± SD percentage of trials remaining in each condition for each participant was 94.8±11.5. Following all exclusion there were an equal number (n = 13) of participants in the A-B-A-B and B-A-B-A designs. ## Results & discussion Bayesian repeated measures ANOVA on RTs with a within-subjects factor of target orientation (point inwards/point outwards) support a model including the target orientation term (BF₁₀ = 10.866 \[*strong evidence for H1*\], *p*(H1\|Data) =.916). To confirm the consistency of these effects, a further analysis combining the results of Experiment 3 with those of Experiment 2 did not favour a model featuring social prime type (pointed face/rounded face/rounded (seal) face). Instead, it supported a model featuring only the target orientation term (BF₁₀ = 1.162e+7 \[*extreme evidence for H1*\], *p*(H1\|Data) =.631). Reaction times were faster when detecting inwards pointing. In the target orientation task, Bayesian binomial tests (test value = 0.5) indicated that participants were able to identify the facing pair (which were primed by the video interaction) (26/26 participants, BF₊₀ = 4.971e+6, \[*extreme evidence*\]). Frequentist modelling supports all findings. All models at osf.io/qxk8z. Again, the results are clear. The priming was effective at communicating target front and, despite semantic labelling and multiple observations of the round end as the “face” during social priming, the low-level physical property of the pointedness automatically triggered attention orienting which resulted in faster detection of targets pointing inwards. # Experiment 4 (‘Seagull’ attention pre-test) When arguing for the absence of an effect, it is important to employ alternative approaches. Such converging techniques increase our confidence in the conclusions drawn from the observed data. Therefore, in the next two experiments we use an ambiguous figure for which two sides can be viewed as its “face”. Such ambiguous figures have intrigued students of perception for many years and here we use a version of Fisher’s well-known ambiguous rabbit/duck figure. Prior to running Experiment 4, we explored whether this version of the duck/rabbit was in fact seen as those two animals. Anecdotal reporting from naïve viewers indicated the rabbit was always perceived but that the alternative interpretation was more strongly reminiscent of a seagull than a duck. Consequently, our images were described as *rabbit* or *seagull* to participants and for the remainder of this manuscript. Our intention in this experiment is to bias the interpretation of this stimulus’ “face” using priming techniques and then show the stimuli in a visual search task in which the stimuli are “facing” towards or away from one another, as in the previous visual search tasks. Our logic is the same as Tinbergen’s goose/hawk ambiguous figure where the same physical object can have quite different identities. In the present experiment we explore the attention cueing nature of our rabbit/seagull stimulus using the Posner design of Experiment 1. Because the potential attention cueing properties of these stimuli have never been examined, it was necessary to examine them in a Posner cueing procedure similar to that of Experiment 1. Our initial speculation is that the ears/beak side of the object (the left side in the orientation shown), with its protruding distinctive features might be more salient and draw attention to that side of the object, and hence to targets on that side of space, as previously demonstrated by Leek & Johnston. ## Method ### Apparatus & design The apparatus and design were identical to that of Experiment 1 with the exception of the cue stimulus. Rather than using a teardrop shape, this experiment used a modified version of Fisher’s ambiguous figure. The object appeared an equal number of times with the ‘beak’ to the left and to the right in a fully counterbalanced design of 112 trials. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Participants Protocols approval, recruitment technique and recruitment criteria were identical to Experiment 1. Twenty-nine participants were tested but 26 remained following exclusions (see **Data exclusion and analysis)**. Data was collected on the 20^th December 2019. ### Data exclusion and analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: 1 participant was excluded due to error rate (\>25% errors in either the target (i.e. \>24) or distractor (i.e. \>3) trials); and 2 participants were excluded due to few remaining trials following RT exclusion (\<75% in any SOA × cue condition). The mean ± SD percentage of trials remaining in each SOA × cue for each participant 95.8 ± 5.5%. ## Results & discussion Reaction times are shown in. Bayesian repeated measures ANOVA on RTs with within subject factors of SOA (200/600 ms) and Cue (cued/uncued) support a model including both main terms (BF₁₀ = 5.730e+14 \[*extreme evidence for H1*\], *p*(H1\|Data) =.761; SOA BF_incl. = 1.001e+14; Cue BF_incl. = 241.422). Reaction times were shorter for the 600 ms SOA than for the 200 ms SOA, and were shorter for the when the target was cued (seagull faced toward the target/rabbit faced away from the target) than when the target was uncued (seagull faced away from the target/rabbit faced toward the target). Frequentist modelling supports these findings. All models at osf.io/qxk8z. The attention cueing experiment provides clear information concerning the properties of the rabbit/seagull stimulus. Targets are detected more rapidly when presented to the side of the ears/beak. This attention orienting is fast, within 200ms; stable as it is maintained up to 600ms; and automatic in that the rabbit/seagull stimulus does not predict the location of the up-coming target. As noted above, this asymmetric physical bias with distinctive features to one side of an object confirms that they can introduce directionality to an object representation, possibly evoked by attention capture by the features (e.g. Leek & Johnston). Therefore, as in the previous experiments, we have clear low-level properties that rapidly orient attention. The question in Experiment 5 is therefore whether priming a particular semantic interpretation can influence the third-party grouping in the visual search task. In that experiment one group of participants is presented with stimuli semantically associated with “rabbit” and the other with “seagull”. For example, the rabbit group are told from the start they are involved in a rabbit experiment, they are shown a variety of drawings of rabbits in different postures, and in the search task they are to search for rabbits either facing towards or away from each other. # Experiment 5 (‘Seagull’ identity priming) ## Method ### Apparatus The apparatus was identical to that of Experiment 2. ### Design The design was identical to that of Experiment 2 with two exceptions. Firstly, no social priming videos were shown. Secondly, the instruction screen for the main task block included the semantic prime for either rabbit or seagull by showing line drawings of the appropriate animal next to the instructions. For each participant the target shape in the search task was thereafter referred to as a ‘rabbit’ or ‘seagull’ as appropriate and the appropriate rabbit or seagull images where consistently observed before every block of search trials. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Conditions Conditions were identical to Experiment 2. ### Participants Protocols were approved by the University of York’s Psychology Departmental Ethics Committee and were in accord with the tenets of the Declaration of Helsinki. Participants were recruited through the University of York’s Psychology Department participant recruitment system. Informed consent was obtained prior to participation. For the rabbit prime condition 30 participants were tested and 26 (age mean±SD = 19.6±1.4, 4 male) remained following exclusions (see Data exclusion and analysis). For the seagull prime condition 30 participants were tested and 26 (age mean±SD = 21.1±7.8, 2 male) remained following exclusions (see Data exclusion and analysis). ### Data exclusion and analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: 6 participants were excluded due to error rate (errors on \>20% of trials); and 2 participants was excluded due to few remaining trials following RT and MT exclusion (\<75% trials). The mean ± SD percentage of trials remaining in each condition for each participant was 95.6±8.4 in the rabbit prime condition and 97.4±4.3 in the seagull prime condition. Following all exclusion there were an equal number (n = 13) of participants in the A-B-A-B and B-A-B-A designs. ## Results & discussion Bayesian repeated measures ANOVA on RTs with a within subjects factor of target orientation (seagull inwards/seagull outwards) and a between subjects factor of prime type (seagull face/rabbit face) support a model including only the target orientation term (BF₁₀ = 48.955 \[v*ery strong evidence for H1*\], *p*(H1\|Data) =.448; BF_incl. = 38.352). Reaction times were faster when detecting inwards pointing seagulls regardless of the prime. In the target orientation task, Bayesian binomial tests (test value = 0.5) indicated that the majority of participants were able to identify the facing pair in both the seagull (22/26 participants, BF₊₀ = 332.459 \[*extreme evidence*\]) and rabbit prime conditions (24/26 participants, BF₊₀ = 15292.424, *extreme evidence*). Frequentist modelling supports all findings. All models at osf.io/qxk8z. Confirming our previous findings, this experiment demonstrates in the final object orientation question that our priming techniques are effective in creating an internal representation of object identity and facing direction, and we assume such representations are active during visual search. Nevertheless, low-level visual features that orient attention dominate and drive the grouping effects when searching for targets amongst distractors. That is, independently of whether participants were primed to perceive rabbits or seagulls, the search data was equivalent: targets were detected significantly faster when the beaks/ears were oriented towards each other, which is exactly the result predicted from the prior attention cueing study and the results of Experiment 2 and 3. # Experiment 6 (‘Symmetrical shape’ attention pre-test) The data thus far clearly show that the low-level visual features of an object that evoke attention orienting determine the visual search performance. So far we have no evidence for higher-level social interaction experience influencing how the objects are represented and searched for. However, the results thus far do not unequivocally demonstrate that higher-level semantic representations of object identity and exposure to third-party social interactions are not represented. For this issue, consider, which represents Gestalt grouping processes. For most participants the image is initially grouped in terms of vertical columns, based on low-level grey-scale features. The higher-level subsequently computed horizontal grouping by shape (circle vs square) has little effect on how the display is grouped during the initial processing of the display. However, clearly the high-level shape information is internally represented, and indeed with further focussed processing, this structure can be extracted. The low-level earlier computed grey scale dominates initial attention capture and perceptual processing. Hence it is possible to make a similar claim in the experiments described thus far. It is possible that the third-party social interaction relationships have been activated and are represented, but the low-level visual features, which are computed at earlier stages, dominate rapid search performance. Indeed, this idea that encoding of social interactions might be a slow process that loses the race to more basic visual processes, has recently been supported by Isik et al.. They report that the encoding of social interactions is a relatively late process taking around 300 to 500ms. This is in contrast to other complex visual processes such as face, object and scene processes that can be computed within 100 to 200ms. In our current experiments, the low-level features such as pointed ends that orient attention are even simpler than face, object and scene analysis, and hence can be encoded rapidly, further increasing the temporal contrast with the later third-party interaction processing. Therefore, in our final experiments, we utilize stimuli that are symmetrical and do not have low-level visual features that orient attention to one side of space. This provides a further test of whether priming higher-level third party representations which bias which part of an object is forward facing, can influence grouping and performance in a visual search task. However, before the visual search study, we have to investigate the attentional orienting properties of these new stimuli. That is, an initial study with these new symmetrical stimuli will ensure that there are no low-level visual properties that could bias attention to one side of space. For example, the stimuli differ in colour, and it is possible that one colour is more salient than the other. Hence we again employ the attention cueing task of Experiments 1 and 4, and now expect to see no attention orienting cueing effects. The lack of orienting effect will enable a further test of whether the higher-level third- party representations of the objects can be primed and influence search performance. ## Method ### Apparatus & design The apparatus and design were identical to that of Experiment 1 with the exception of the cue stimulus. Rather than using a teardrop shape, this experiment used a lemon shaped object that was two colours separated along its midline. The object appeared an equal number of times with yellow on its left and on its right in a fully counterbalanced design of 112 trials. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Participants Protocols approval, recruitment technique and recruitment criteria were identical to Experiment 1. Thirty-one participants were tested but 26 remained following exclusions (see Data exclusion and analysis). Data was collected on the 20^th December 2019. ### Data exclusion and analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: 3 participants were excluded due to error rate (\>25% errors in either the target (i.e. \>24) or distractor (i.e. \>3) trials); and 2 participants were excluded due to few remaining trials following RT exclusion (\<75% in any SOA × cue condition). The mean ± SD percentage of trials remaining in each SOA × congruency for each participant 96.3 ± 5.4%. ## Results & discussion Reaction times are shown in. Bayesian repeated measures ANOVA on RTs with within subject factors of SOA (200/600 ms) and Cue (yellow face/ purple face) support a model including only the SOA term (BF₁₀ = 2.343e+11 \[*extreme evidence for H1*\], *p*(H1\|Data) =.790; SOA BF_incl. = 1.634e+11). Reaction times were shorter for the 600 ms SOA than for the 200 ms SOA. Frequentist modelling supports these findings. All models at osf.io/qxk8z. In contrast to Experiments 1 and 4, we find no evidence for attention cueing effects for either side of the stimulus. The lack of low-level visual features orienting attention ensures that this new stimulus provides an appropriate vehicle for testing whether higher-level representations of third-party interactions can be activated and influence visual search performance. # Experiment 7 (‘Symmetrical shape’ social priming) ## Method ### Apparatus The apparatus was identical to that of Experiments 2 and 3. ### Design The design was identical to that of Experiment 3 with three exceptions. Firstly, the target changed from a teardrop shape to a symmetrical two-colour lemon shape from Experiment 6. Second, both coloured ends of the lemon shape could be primed as the “face” (as was the case in Experiment 2). Third, though the priming block remained, the video with targets changing colour was removed since the targets were already coloured. Rather than introducing a white seal in the priming block, instead a purple/yellow seal was presented. The between-subjects condition was whether the yellow or purple front was indicated as the face in this prime and in the two subsequent videos. The motions and size of these targets were identical to the targets in Experiment 2 and 3. The colour change video from Experiment 2 was not used in the present experiment since the shapes would remain purple and yellow. After this initial semantic identity priming procedure, the social interaction videos containing 3 objects were presented before each search block as in Experiments 2 and 3. All stimuli and stimulus size details are available at [osf.io/qxk8z](http://osf.io/qxk8z). ### Conditions Conditions were identical to those of Experiment 2 (i.e. they differed from Experiment 3 in that either end (purple or yellow) of the target could be primed as the front. ### Participants Protocols were approved by the University of York’s Psychology Departmental Ethics Committee and were in accord with the tenets of the Declaration of Helsinki. Participants were recruited through the University of York’s Psychology Department participant recruitment system. Informed consent was obtained prior to participation. For the yellow face prime condition 37 participants were tested and 26 (age mean±SD = 19.6±1.4, 4 male) remained following exclusions (see Data exclusion and analysis). For the purple face prime condition 40 participants were tested and 26 (age mean±SD = 21.7±6.3, 4 male) remained following exclusions (see Data exclusion and analysis). ### Data exclusion and analysis Full exclusion details can be found at osf.io/qxk8z. Briefly: 21 participants were excluded due to error rate (errors on \>20% of trials); and 4 participants were excluded due to few remaining trials following RT and MT exclusion (\<75% trials). The mean ± SD percentage of trials remaining in each condition for each participant was 95.8±8.9 in the yellow face prime condition and 97.7±6.6 in the purple face prime condition. Following all exclusion there were an equal number (n = 13) of participants in the A-B-A-B and B-A-B-A designs of each prime condition. ## Results & discussion Bayesian repeated measures ANOVA on RTs (see with a within-subjects factor of target orientation (yellow inwards/yellow outwards) and a between-subjects factor of prime type (yellow front/purple front) does not support any model over the null (BF₁₀ \< =.757). Reaction times did not differ between conditions. In the task where participants judged which target pairs were facing each other, Bayesian binomial tests (test value = 0.5) indicated that the majority of participants were able to identify the facing pair in both the yellow (21/26 participants, BF₊₀ = 75.513 \[*very strong evidence*\]) and purple prime conditions (25/26 participants, BF₊₀ = 191193.305, \[*extreme evidence*\]). Frequentist modelling supports all findings. All models at osf.io/qxk8z. The results confirm our previous observations. First, priming techniques were effective at communicating target front even with symmetrical targets. Second, even after extensive repeated experience of object identities and observing object social interactions, the representations do not influence performance on a visual search task. For example, if a participant had been exposed to a series of events showing that the yellow end of an object has agency and is equivalent to the “face” during a series of video displays, detection of towards facing yellow stimuli is not facilitated during search. This result confirms the findings of Experiments 2, 3, and 5. # General discussion Previous research has shown that in complex and cluttered environments participants are able to detect interacting people faster than non-interacting people. This process would appear to be a valuable way to parse complex social scenes into interacting individuals where important social processes might be taking place, which would be worthy of further analysis to interpret the scene and predict potential future behaviours. For example, when two people are perceived to be interacting, the emotion of one influences the perceived emotion of the other and subsequent short- and longer-term memory of the interaction is influenced by the initial computation of the social interaction. The issue we have examined here concerns what the specific mechanisms might be that mediate this initial processing that enables the structuring of social scenes. On the one hand, there may be rapid encoding of the third-party interaction between social beings. Such encoding of high-level interpersonal processes might predict that the visual search effects would only be detected when observing social animals such as humans. Indeed, Vestner et al. explicitly argued that effects were caused by higher-level representations of social interactions, rather than lower level perceptual processes. However, subsequent work has challenged this conclusion, at least in visual search tasks, by demonstrating that in fact the same social priority effects in visual search can be detected when participants search for towards vs away facing arrows. As these arrow stimuli do not possess the social properties of biological systems, this would suggest that the effect is driven by a non-social low-level attention orienting process (e.g.). The interpretation in terms of general attention mechanisms is that inwards faces or arrows orient the beam of attention to one central location between the two critical target objects facilitating search. In contrast, away facing people or arrows evoke attention shifts in opposite and hence competing directions. Such splitting of attention would impair the judgment of the relationship between the two objects. However, although the effects demonstrated with arrows are equivalent to those of faces, suggesting that higher-level social representations of interacting individuals are not necessary to produce the effects, they do <u>not</u> unequivocally demonstrate that higher-level social processes are not involved due to potential ceiling effects preventing the detection of additive effects. And indeed, a range of studies have in fact argued for the role of mentalizing processes during social shifts of attention, such as learning of trust and action intention (e.g., and for review). Therefore, in this series of studies we have investigated this issue further. A series of experiments using a range of converging methods has examined whether creating higher-level representations of objects as interacting individuals could influence the initial structural encoding of visual scenes, independently of low-level visual properties. A wide range of previous research has demonstrated the potency of the video priming techniques we have used, where effects can be observed in pre-language 6-month olds for example. And we confirmed that such techniques appear to influence the representation of social agency in the direction of the object’s attention. The evidence is clear within this research programme that such higher-level representations of social interactions created by our social priming techniques are not playing a significant role. Clearly, as this is a null finding, we have to be cautious with our interpretations. Nevertheless, further analysis combining Experiments 2, 3 and 5 provides substantial power to support our conclusions. Within this analysis, there is substantial evidence that our visual search effects are dominated by low-level visual features that automatically trigger attention orienting (BF₁₀ = 2.435e+9 \[*extreme evidence for H1\]*, *p*(H1\|Data) =.687; BF_incl. = 1.784e+9), and any inclusion of social priming considerably worsens model fitting by a factor of 15 or greater (see model at osf.io/qxk8z). As discussed, we are not discounting higher-level social processes in all situations, and indeed in the Vestner et al. studies, it is possible that such effects are taking place at the later processing stages of working memory maintenance and retrieval from longer-term memory. Rather, we argue it is likely that they play a limited role in the earliest stages of processing, where there is rapid structuring of the visual environment to create internal representations for further processing. Thus, we suggest that the higher-level mentalizing processes may be at play at later stages as suggested by evidence from neuroscience and behavioural studies showing that the effects of mentalizing on gaze following are slower non-automatic processes (e.g., though also note recent work potentially indicating mentalising as a fast automatic process e.g.). As an example of two potential processes, one rapid and another slower, consider the attention cueing study of. In that study the rapid attentional cueing effects evoked by gazing faces, as measured by reaction time to detect peripheral targets, were unaffected by the emotion of the face (smile vs disgust). However, in the same task a slower social learning process was simultaneously at play, where subsequent decisions concerning object liking were influenced by the interaction between gaze and emotion. In a similar vane, tasks requiring deeper encoding where participants actively switch perception from their own first-person to another’s third-person perspective before judging a visual scene, also provide evidence for slower mentalizing processes. In conclusion, rapid parsing of social scenes into interacting vs non- interacting individuals is an important process that provides initial representations for further more sophisticated processing. However, our data suggest that this initial process structuring the social world is not based on complex and sophisticated representations of socially interacting individuals. Rather, the highly efficient attention systems utilizing basic perceptual features that have evolved for the interaction between vision and action to enable selective goal-directed action can also serve these social computations. The employment of basic attention processes would appear to be the most parsimonious and efficient way of rapidly structuring visual inputs to reflect social interactions. We would like to thank Bryony G. McKean and Edward Hindmarsh for assistance with data collection. 10.1371/journal.pone.0258832.r001 Decision Letter 0 Steinborn Michael B. Academic Editor 2022 Michael B. Steinborn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 26 Apr 2021 PONE-D-21-06244 Rapid detection of social interactions is the result of domain general attentional processes PLOS ONE Dear Dr. Flavell, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Three independent experts commented on your manuscript. One of them disclosed as Hauke Meyerhoff. As you can see from the reviews, all referees found the general topic addressed in your manuscript interesting and they have many nice things to say about the study. At the same time, they have a whole number of remarkably constructive and excellently detailed suggestions how to further improve the paper. The comments speak for themselves, but it is obvious that one reoccurring theme is the need for more specificity regarding the theoretical concepts., and some minor methodological issues. While this will call for some extra efforts, I consider it worthwhile. Therefore, we invite you to submit a revision of the manuscript that addresses the remaining points together with a cover letter that contains point-by-point replies. Please submit your revised manuscript by Jun 10 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Michael B. Steinborn, PhD Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <u><https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf></u> and [https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf](about:blank) \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Partly \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: This is interesting research. The authors conducted seven experiments that systematically assess the influence of basic perceptual features vs. that of higher-level social representation on visual search. The authors make data, results and stimuli publicly available. They report Bayesian analyses in the main text but also attach frequentist statistics in the online supplemental materials. I really enjoyed reading this paper and many of my remarks are intended to improve readability. However, I also have some suggestions for extensions of this manuscript General remarks: \- Experiment 1/4/6 were conducted online while experiments 2, 3, 5 and 7 were conducted in the lab. Most likely, the participants of experiments 1/4/6 vs. 2/3/5/7 stem from different cohorts. I don’t think that this is a big issue since the authors don’t directly compare results from these experiments, but I still think that this should be mentioned briefly in the discussion \- In the document containing the models at OSF, it is not readily apparent what the factor levels refer to. E.g. in experiment 2, there is the factor “TA” with the levels “T” and “A”. Another example: In experiment 6, there is a factor labeled “congruency”. What does this name refer to when there were no directional cues in this experiment? Could the authors please rename each factor (level) in each experiment so that their meaning is obvious? Please also go through all headings and captions and check whether everything is self- explanatory. For example, it is not clear to me what “MTs” refers to (e.g. in “ANOVA on MTs”). Please clarify this, e.g. in the caption. \- Also, in the model document in the osf respository, it is now the case that the model solution containing the interaction term is printed in bold (e.g., in the first table, the solution with two main effects plus the interaction term (“SOA + Congruency + SOA \* Congruency”) is printed in bold although this does not provide the best solution). This is a bit misleading as I would expect the best solution to stick out. My suggestion is to either highlight the best model or none Introduction: I found the introduction well understandable and straight-forward. The hypotheses are clear and plausible. I have two small suggestions: \- In the introduction, the authors make a point about parallel processing streams for social and directional cues. Perhaps this could be strengthened by citing Ristic, Friesen & Kingstone (2002), Psychonomic Bulletin & Review (e.g. on page 5, lines 81 ff.) \- I found the experiment overview figure in the OSF database very helpful to keep track/get an overview of the setup of the seven experiments. Perhaps this figure could also be included in the manuscript Generally, I found the descriptions of the experiments clear and easy to follow. The results are convincing and I agree with the author’s interpretations. I have some remarks regarding some of the experiments: Experiment 1: \- Maybe the authors could add on page 11 para 1 whether the participants were explicitly instructed how to press the respond keys (e.g. both index fingers or two fingers of the same hand) Experiment 2: \- Could the authors please report the visual angle of stimulus presentation? SR research provides a nice tool to calculate it: <https://www.sr- research.com/visual-angle-calculator/> \- Is there a specific reason why, in Figure 7, the upper two left-hand boxes say ‘with round as “face”’ while the lower two say ‘(round as “face”)’? If not, I suggest to chose either one and keep it consistent \- I presume that this was just a technical error in the uploading process, however, the axis labels of Figure 9 are very hard to read. Also, I suggest no to divide e.g. 1200 ms with a comma (1,200 ms) \- Could the authors please add short descriptives to the main text that indicate the direction of an effect (e.g. lines 379-380: were inwards or outwards facing targets responded to faster?) Experiment 3 \- I think, Figures 10 and 11 were mixed up in the submission. In my PDF, Figure 10 depicts the seagull/rabbit Figure and Figure 11 shows the seal Experiment 4 \- No remarks Experiment 5 \- Since the participants were Psychology students, I wonder whether they (or some of them) might have been aware of the rabbit/duck figure. Did the authors assess this (e.g. by asking participants about it, or how they perceived the figure) and exclude participants who were familiar with it? If yes, this should be reported as well. If not, I don’t think it is much of a problem because binomial tests confirm that the cues were perceived as intended Experiment 6 \- No remarks Experiment 7 \- The number of participants excluded due to error rates is much higher in this experiment compared to experiment 1-6 (11 and 14 compared to 3-6 participants). Do the authors have any idea about why this is the case? General Discussion As the rest of the manuscript, the general discussion is neat and well-readable. I have two points to add: \- I miss a little bit of a summary of the main findings of the experiments and how they are interrelated / what big picture they form. This would be particularly helpful for readers who don’t have time to read the whole paper in detail \- On page 38, the authors discuss the interpretation of null results and introduce new results in this section. I feel that this paragraph is worth extension and more results could be included. Perhaps this could be done (in part) in a separate section prior to the general discussion Minor points: Line 769: in a similar vein Reviewer \#2: Review for Flavell et al. (submitted to PlosOne). Rapid detection of social interactions is the result of domain general attentional processes Summary The authors present a big set of experiment investigating whether low-level perceptual features of high-level social features guide attentional processing in visual search tasks. They test shapes with orientation information, schematic animals (ambiguous drawings), as well as neutral shapes with differently colored ends. The pattern of results is consistent with the interpretation that low- level properties guide attention. Evaluation This is a good set of experiments. I only have a couple of points and references which I think can improve the manuscript. I want to explicitly emphasize that I appreciated the way how the authors managed to keep the manuscript short despite reporting so many experiments. Major Comments 1\. One shortcoming in the experiments is that the authors do not compare the different stimuli within the same experiment. Such a design would allow to test for interactions. Instead, they report the presence of an effect in one experiment and the absence of an effect in another experiment (which does not allow the conclusion that both effects differ from each other). I think this should be mentioned and discussed. Given the rather strong results, however, I do not think that this needs to be run as an additional experiment. 2\. There is a related field of research which came up with matching findings which the authors might find interesting to broaden their discussion (which is rather narrow). Visual search and attentional guidance research on perceptual animacy has shown that rather low-level properties rather than the social properties guide the detection of such interactions (Meyerhoff, Schwan, & Huff, 2014, JEP:HPP; Meyerhoff, Schwan, & Huff, 2014; PB&R). In the same way, the authors might find the literature on the wolfpack effect interesting (Gao, MacCarthy, & Scholl, 2010, Psych. Sci.) 3\. The authors report the BF10 for the final models (i.e. for the factors which explain the results). As a part of the story is to say that the other factors (e.g. the priming manipulations) are irrelevant for explaining the results. It would be nice to see the BF01 or BF10 for these factors as well (maybe more general the BFs for factors rather than models?). Minor Stuff \- There seem to be some residuals from previous submissions (e.g., there always will be color in an online journal) which should be removed from the ms. \- Figure 11 is missing \- Stimuli sizes are probably important here, so they should be in the Methods. Signed, Hauke Meyerhoff Reviewer \#3: This paper seemed great in several ways: the question motivating the study is interesting and important, the results are clear and robust, and it's a well-written and well-organized manuscript. I also find this series of experiments to be a nice addition to the authors’ past work, and admire the depth of the research program as a whole. Along with my enthusiasm, I have a few concerns about the interpretation and discussion of the results, as well as some minor comments. Major comments 1\. A crucial assumption underlying the conclusions is that the videos primed higher-level social interpretations; this was assessed via a final forced-choice question asking participants to select which stimuli were facing each other. Based on responses largely congruent with the primes, the authors conclude they have successfully manipulated “the representation of social agency in the direction of the object’s attention” (p.38). But the DV (i.e., participants’ ability to select the front as instructed) seems different from the conclusions (i.e., participants’ representations of social properties/joint attention). First of all, this forced-choice question is extremely susceptible to demand characteristics. For example, in Expts. 3/7 it would have been very surprising if participants had indicated the pointy/purple end of the seal as its front after seeing the speech bubble coming out of its round/yellow end. This is even more apparent in Expt. 5, where the instructions prompted participants to “pick which pair of seagulls are facing each other”; the fact the intended interpretation of the stimuli is re-iterated in the question makes me wonder whether participants were simply selecting what they had been taught was the correct option – regardless of their actual percepts. So I am on board when the authors label these primes as “semantic labels” (p. 26, 34), but not when they describe them as directly manipulating “the representation of social agency in the direction of the object’s attention” (p.38) or “high-level representations of social interactions” (p.38), or when equating these manipulations to the perception of a face or social interactions – which seems fundamentally different in nature. 2\. Even assuming the primes worked as intended, I wonder if the relevant interpretations were active during the search task itself. Perceptual interpretations of ambiguous figures have been shown to rely on fixation patterns and covert attention (e.g. Peterson & Gibson, 1991, JEP:HPP; Toppino, 2003, P&P) – but both fixations and covert attention are key for successful visual search, and so top-down interpretations of ambiguous stimuli are directly in competition with search performance (which is especially high here). This is not a problem in itself, but it highlights another problem with generalizing from ambiguous stimuli to faces and people, since face perception doesn't require top-down control (or not even awareness, e.g. Stein et al., 2012, Cognition). 3\. Another assumption here is that default percepts of fronts reflect lower- level visual features; but I don't think this can be taken for granted. For example, the fact that seagulls facing each other are prioritized in Expt. 5 could be an effect of the interpretation of seagulls as seagulls (in both conditions, if the prime is inactive as per points \#1 and \#2) and it is in fact their perceived facing direction that is driving attention. (Relatedly, the claim on p.7 that the stimuli “contain no intrinsic visual features that could be construed as a face” and have “no salient intrinsic face-like features” doesn’t seem right to me, since the seagulls/rabbits have a beak, ears, and eye.) This also applies to Expt. 2, as some authors have suggested that arrows can in fact be socially meaningful (e.g. Kingstone et al., 2003) or suggest an agentic presence, especially when multiple stimuli point towards the same region of space (e.g. Gao et al., 2010, Psych Sci; Takahashi et al., 2013, Front Hum Neurosci; Colombatto et al., 2019, Perception). Since these results as described have potential implications for that literature, it would be great to see this discussed. 4\. One of the reasons social binding was initially thought to be a social effect was that it vanished with inverted faces that are equivalent to upright faces in lower-level properties, but differ in the higher-level social properties (although related to my point \#1, participants would still be able to indicate whether people are facing each other, even when they are inverted!). Those findings contradict the conclusion that lower-level properties only drive social binding – and since this work directly follows up on those experiments, it would be interesting to see the authors discuss this. Minor comments 1\. The main text for both Expts. 6 and 7 describes a ‘yellow’ vs. ‘purple’ manipulation (also depicted in Figs. 5, 14-15). But then Fig. 9 depicts ‘orange’ vs. ‘purple’ conditions, and the datafiles report ‘orange’ and ‘purple’ for Expt. 6, and ‘rabbit’ and ‘orange’ for Expt. 7 (in both the RT and MT files); could these be made consistent? 2\. How were subjects assigned to the prime conditions – were they alternating? (This would be helpful to clarify as potentially related to block order assignment.) 3\. How were the prime videos generated for the ‘round-face’ conditions? Were the stimuli simply mirrored/rotated? I ask because this seems to have produced some artifacts, e.g. the ‘round-faces’ overlap at the end of the videos and during the ball tossing game, which would be weird for agents, or the ‘round- face’ stimuli either toss the ball at a distance (rightmost), or overlap with the ball (leftmost). 4\. The target orientation question for Exp. 2 is phrased in the instructions as “You need to pick which pair of seals are facing each other”, but I don’t think ‘seals’ was ever used elsewhere in Exp. 2. How could participants answer this question? 5\. This statement seems inaccurate/controversial: “higher-level mentalising processes are at play at later stages as suggested by \[…\] behavioural studies showing that the effects of mentalising on gaze following are slower non- automatic”: the matter is still debated, but recent evidence showing the contrary should be cited, e.g. that mentalising such as perspective taking can occur rapidly and automatically (e.g. Ward et al., 2019, Curr Bio), and that percepts of mental states can influence lower-level processes such as gaze cueing, even quickly and automatically (e.g. Colombatto et al., 2020, PNAS). 6\. For additional evidence that the pointy end of the teardrop is seen as its front, the authors may find helpful Chen & Scholl, 2018, PB&R – who used teardrop stimuli to elicit the same effects (on aesthetic preferences) as other types of fronts. (I wouldn't normally mention too many papers from our group in a review, but each one here seems warranted as potentially helpful or directly relevant to the discussion.) \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Christina Breil Reviewer \#2: No Reviewer \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0258832.r002 Author response to Decision Letter 0 16 Aug 2021 RESPONSE TO REVIEWERS NOTE - copied from the uploaded document 'Response to Reviewers_v2.docx. Formatting may vary. Dear Dr Steinborn, Thank you for your letter concerning our paper “Rapid detection of social interactions is the result of domain general attentional processes”. We are grateful for the opportunity to resubmit our revised manuscript. As always, we appreciate the time and effort you and the reviewers have committed to our work, and we were very pleased by the positive and helpful feedback. In the following we describe our response to the reviewer feedback. Reviewer comments are shown in black and our responses are shown in blue. Yours sincerely, Jonathan Flavell, Harriet Over, Tim Vestner, Richard Cook, Steven Tipper -o0o- REVIEWER \#1 This is interesting research. The authors conducted seven experiments that systematically assess the influence of basic perceptual features vs. that of higher-level social representation on visual search. The authors make data, results and stimuli publicly available. They report Bayesian analyses in the main text but also attach frequentist statistics in the online supplemental materials. I really enjoyed reading this paper and many of my remarks are intended to improve readability. However, I also have some suggestions for extensions of this manuscript Thank you for these kind comments. General remarks: Experiment 1/4/6 were conducted online while experiments 2, 3, 5 and 7 were conducted in the lab. Most likely, the participants of experiments 1/4/6 vs. 2/3/5/7 stem from different cohorts. I don’t think that this is a big issue since the authors don’t directly compare results from these experiments, but I still think that this should be mentioned briefly in the discussion We have now raised this in ‘Participants’ of Experiment 2. In the document containing the models at OSF, it is not readily apparent what the factor levels refer to. E.g. in experiment 2, there is the factor “TA” with the levels “T” and “A”. Another example: In experiment 6, there is a factor labeled “congruency”. What does this name refer to when there were no directional cues in this experiment? Could the authors please rename each factor (level) in each experiment so that their meaning is obvious? Please also go through all headings and captions and check whether everything is self- explanatory. For example, it is not clear to me what “MTs” refers to (e.g. in “ANOVA on MTs”). Please clarify this, e.g. in the caption. Also, in the model document in the osf respository, it is now the case that the model solution containing the interaction term is printed in bold (e.g., in the first table, the solution with two main effects plus the interaction term (“SOA + Congruency + SOA \* Congruency”) is printed in bold although this does not provide the best solution). This is a bit misleading as I would expect the best solution to stick out. My suggestion is to either highlight the best model or none Thank you for taking the time to check the ‘Models’ document. We apologise for the unclear labelling and the misleading bold type. A new ‘Models’ document has been uploaded to the OSF with un-bolded models (this was a formatting error in the first upload) and labels changed to match the terminology used in the manuscript. See below for label changes (from → to): • Congruency → Cue o Con → Cued o Incon → Uncued; • TA → Orientation o T → Inwards o A → Outwards • RT → Reaction Time • MT → Movement Time Introduction: I found the introduction well understandable and straight-forward. The hypotheses are clear and plausible. I have two small suggestions: \- In the introduction, the authors make a point about parallel processing streams for social and directional cues. Perhaps this could be strengthened by citing Ristic, Friesen & Kingstone (2002), Psychonomic Bulletin & Review (e.g. on page 5, lines 81 ff.) Thank you for the helpful reference. We’ve added it to the introduction as suggested. \- I found the experiment overview figure in the OSF database very helpful to keep track/get an overview of the setup of the seven experiments. Perhaps this figure could also be included in the manuscript We have also added the experiment overview figure to the introduction. Generally, I found the descriptions of the experiments clear and easy to follow. The results are convincing and I agree with the author’s interpretations. I have some remarks regarding some of the experiments: Experiment 1: \- Maybe the authors could add on page 11 para 1 whether the participants were explicitly instructed how to press the respond keys (e.g. both index fingers or two fingers of the same hand) Participants were not instructed on which fingers to use for the task (we have now indicated this in the manuscript) but in piloting participants always used their left index finger for ‘F’ and their right index for ‘J’ on the lab QWERTY keyboard. Experiment 2: \- Could the authors please report the visual angle of stimulus presentation? SR research provides a nice tool to calculate it: <https://www.sr- research.com/visual-angle-calculator/> Participants were not head rested and head position was not so horizontal and vertical visual angles are not available. The reported viewing distance of \~50cm from the screen is intended only as a guide to set-up. \- Is there a specific reason why, in Figure 7, the upper two left-hand boxes say ‘with round as “face”’ while the lower two say ‘(round as “face”)’? If not, I suggest to chose either one and keep it consistent Thank you for pointing this out. All items were supposed to be identical. This has been corrected. Note that this is now Figure 8. \- I presume that this was just a technical error in the uploading process, however, the axis labels of Figure 9 are very hard to read. Also, I suggest no to divide e.g. 1200 ms with a comma (1,200 ms). Axis labels in the uploaded.tiff file appear correct so we agree that it must be an upload error. We have removed the comma from y axis labels. A png version of the image is below in case a re-submitted figure has the same issue. Note that this is now Figure 10. \- Could the authors please add short descriptives to the main text that indicate the direction of an effect (e.g. lines 379-380: were inwards or outwards facing targets responded to faster?) Thank you for suggesting this. We have added a single sentence to the end of each model description of the main text to describe the outcome in terms of difference in RT by condition. Experiment 3 \- I think, Figures 10 and 11 were mixed up in the submission. In my PDF, Figure 10 depicts the seagull/rabbit Figure and Figure 11 shows the seal Thank you for spotting this. It was an error in submission. We will double check figure file labels on re-submission. Experiment 4 \- No remarks Experiment 5 \- Since the participants were Psychology students, I wonder whether they (or some of them) might have been aware of the rabbit/duck figure. Did the authors assess this (e.g. by asking participants about it, or how they perceived the figure) and exclude participants who were familiar with it? If yes, this should be reported as well. If not, I don’t think it is much of a problem because binomial tests confirm that the cues were perceived as intended One of the authors demonstrates the rabbit/duck figure in a lecture series every year so this had been considered. The experiment was conducted early in the term before this lecture took place. Participants were not asked about familiarity with the figure. Experiment 6 \- No remarks Experiment 7 \- The number of participants excluded due to error rates is much higher in this experiment compared to experiment 1-6 (11 and 14 compared to 3-6 participants). Do the authors have any idea about why this is the case? The number of exclusions due to error rate was indeed higher in Experiment 7 than it was using the same design with different stimuli (Experiment 2, 3 and 5). The majority of all participants excluded for error rate have exceptionally poor performance generally due to, anecdotally, not understanding the task or lack of effort. Experiment 7 did not differ in that respect but 8 of the excluded participants had error rates at \~30% or less so under less stringent conditions their data may have been included. The remaining participants typically had error rates \>75%. We believe that the relatively higher error in Experiment 7 is likely due to stringent performance requirements to ensure data quality coupled with a few less diligent participants. General Discussion As the rest of the manuscript, the general discussion is neat and well-readable. I have two points to add: \- I miss a little bit of a summary of the main findings of the experiments and how they are interrelated / what big picture they form. This would be particularly helpful for readers who don’t have time to read the whole paper in detail. Thank you for this excellent suggestion. We appreciate that the article is quite complex with a range of different approaches used in 7 experiments. In light of this we have now included a preview of our findings and a new figure 4 at the end of the Introduction. \- On page 38, the authors discuss the interpretation of null results and introduce new results in this section. I feel that this paragraph is worth extension and more results could be included. Perhaps this could be done (in part) in a separate section prior to the general discussion Features were found to dominate attention in all cases and this new combined analysis is intended only to lend credence to those findings. As a result, we have sought to keep the discussion relatively brief. Minor points: Line 769: in a similar vein We are not sure what the reviewer is referring to but we have briefly expanded on the paper by Furlanetto et al. -o0o- REVIEWER \#2 Review for Flavell et al. (submitted to PlosOne). Rapid detection of social interactions is the result of domain general attentional processes Summary The authors present a big set of experiment investigating whether low-level perceptual features of high-level social features guide attentional processing in visual search tasks. They test shapes with orientation information, schematic animals (ambiguous drawings), as well as neutral shapes with differently colored ends. The pattern of results is consistent with the interpretation that low- level properties guide attention. Evaluation This is a good set of experiments. I only have a couple of points and references which I think can improve the manuscript. I want to explicitly emphasize that I appreciated the way how the authors managed to keep the manuscript short despite reporting so many experiments. Thank you for this kind and generous evaluation of our work. Major Comments 1\. One shortcoming in the experiments is that the authors do not compare the different stimuli within the same experiment. Such a design would allow to test for interactions. Instead, they report the presence of an effect in one experiment and the absence of an effect in another experiment (which does not allow the conclusion that both effects differ from each other). I think this should be mentioned and discussed. Given the rather strong results, however, I do not think that this needs to be run as an additional experiment. Re. Posner cueing experiments. We believe that including the teardrop, seagull, symmetrical shape stimuli in a single Posner cueing task is likely to reveal similar results to those found in the three separate tasks – principally that shape direction strongly orients attention to one side or another. Re. Visual search experiments. We assume the the reviewer is considering a 2x2 grid visual search task featuring symmetrical shape pairs and non-symmetrical shape pairs with either the shape type being primed as a social agent beforehand. Again, we believe that similar results to the current experiment would likely be found – faster responses when detecting inwards pointing pairs with no prime effect. However, we agree that it could be interesting to pursue a line work comparing the influence of asocial directional cues (teardrop) with social directional cue (seal and seagull) and social non-direction cues (symmetrical shape) using primes for one or more types. 2\. There is a related field of research which came up with matching findings which the authors might find interesting to broaden their discussion (which is rather narrow). Visual search and attentional guidance research on perceptual animacy has shown that rather low-level properties rather than the social properties guide the detection of such interactions (Meyerhoff, Schwan, & Huff, 2014, JEP:HPP; Meyerhoff, Schwan, & Huff, 2014; PB&R). In the same way, the authors might find the literature on the wolfpack effect interesting (Gao, MacCarthy, & Scholl, 2010, Psych. Sci.) Thank you for these interesting articles. We have now mentioned this work in the Introduction, in set the scene for our new experiments. 3\. The authors report the BF10 for the final models (i.e. for the factors which explain the results). As a part of the story is to say that the other factors (e.g. the priming manipulations) are irrelevant for explaining the results. It would be nice to see the BF01 or BF10 for these factors as well (maybe more general the BFs for factors rather than models?). From each Bayesian ANOVAs we report only the ‘best’ (the model with the largest BF10) model in the manuscript. This does not mean that all other models (with different factors or different combinations of factors) are null. For example, in the table below (copied from the supplementary file ‘Models.docx’ at www.osf.io/qxk8z) all the models fit the data better than the null model though the extent of this varies between models. A model with the just the Cue factor is rather poor whereas the model including SOA + Cue factors and the model including the SOA + Cue + SOA\*Cue factors predict the data extremely well. In the manuscript we report the SOA + Cue model because it fits the data \~2.25 times better than the SOA + Cue + SOA\*Cue model. As described in the Results sections, all models are available open access at www.osf.io/qxk8z. To keep the manuscript succinct and convey the relevant points clearly, we would prefer to keep the ‘rejected’ models and that supplementary material. Note that BF01 can be found as 1/BF10. So a BF10=3.005e+7 (an extremely large number) is equal to BF01=4.519e-4 (an extremely small number). Model Comparison Models P(M) P(M\|data) BF M BF 10 error % Null model (incl. subject) 0.200 2.005e -8 8.021e -8 1.000 SOA + Cue 0.200 0.603 6.065 3.005e +7 2.489 SOA + Cue + SOA  ✻  Cue 0.200 0.268 1.466 1.338e +7 2.449 SOA 0.200 0.129 0.594 6.444e +6 4.574 Cue 0.200 2.840e -8 1.136e -7 1.417 1.153 Note. All models include subject Analysis of Effects Effects P(incl) P(incl\|data) BF incl SOA 0.600 1.000 1.376e +7 Cue 0.600 0.871 4.493 SOA  ✻  Cue 0.200 0.268 1.466 Minor Stuff \- There seem to be some residuals from previous submissions (e.g., there always will be color in an online journal) which should be removed from the ms. Thank you for spotting this error, we have corrected it. \- Figure 11 is missing In our PLOSONE generated pdf version of the manuscript all of the figures are presented at the end of the manuscript but the order is slightly off… figure 8, figure 9, figure 11, figure 10, figure 12. \- Stimuli sizes are probably important here, so they should be in the Methods. We have added details of stimulus sizes and relative positions to a separate document on the OSF names StimulusSizeAndPosition.docx. -o0o- REVIEWER \#3 This paper seemed great in several ways: the question motivating the study is interesting and important, the results are clear and robust, and it's a well- written and well-organized manuscript. I also find this series of experiments to be a nice addition to the authors’ past work, and admire the depth of the research program as a whole. Along with my enthusiasm, I have a few concerns about the interpretation and discussion of the results, as well as some minor comments. Thank you for the kind evaluation of our work. Major comments 1\. A crucial assumption underlying the conclusions is that the videos primed higher-level social interpretations; this was assessed via a final forced-choice question asking participants to select which stimuli were facing each other. Based on responses largely congruent with the primes, the authors conclude they have successfully manipulated “the representation of social agency in the direction of the object’s attention” (p.38). But the DV (i.e., participants’ ability to select the front as instructed) seems different from the conclusions (i.e., participants’ representations of social properties/joint attention). First of all, this forced-choice question is extremely susceptible to demand characteristics. For example, in Expts. 3/7 it would have been very surprising if participants had indicated the pointy/purple end of the seal as its front after seeing the speech bubble coming out of its round/yellow end. This is even more apparent in Expt. 5, where the instructions prompted participants to “pick which pair of seagulls are facing each other”; the fact the intended interpretation of the stimuli is re-iterated in the question makes me wonder whether participants were simply selecting what they had been taught was the correct option – regardless of their actual percepts. So I am on board when the authors label these primes as “semantic labels” (p. 26, 34), but not when they describe them as directly manipulating “the representation of social agency in the direction of the object’s attention” (p.38) or “high-level representations of social interactions” (p.38), or when equating these manipulations to the perception of a face or social interactions – which seems fundamentally different in nature. We appreciate the reviewers concerns. Our initial hypothesis, and “hoped for” result, was to detect the effects of higher-level social representations in the visual search task. Unfortunately we did not detect these effects. Our approach has been to employ converging operations, where we attempted to detect the effects via a variety of techniques. These approaches have been established as effective. For example, the video action sequences have been used in many studies ranging from young children to clinical populations, and they always evoke potent experiences of animate interactive objects (see the video demos at osf.io/qxk8z). On the specific point concerning the measure of object identity biased by our techniques, first we would like to state that the orientation question results from Experiment 2 (no speech bubble and question does not refer to either side) are comparable to the results of Experiments 3 and 7 (target with speech bubble) and Experiment 5 (question with animal reference). This suggests that the question of front/face may not be as susceptible to demands as the reviewer suggests. In considerations of the reviewer’s strong feelings on the text on page 38, we have toned down and qualified the specified statements. And indeed, we do not rule out the possibility that as yet undiscovered techniques that are better than our approaches might demonstrate the role of high-level social representations in the search tasks. 2\. Even assuming the primes worked as intended, I wonder if the relevant interpretations were active during the search task itself. Perceptual interpretations of ambiguous figures have been shown to rely on fixation patterns and covert attention (e.g. Peterson & Gibson, 1991, JEP:HPP; Toppino, 2003, P&P) – but both fixations and covert attention are key for successful visual search, and so top-down interpretations of ambiguous stimuli are directly in competition with search performance (which is especially high here). This is not a problem in itself, but it highlights another problem with generalizing from ambiguous stimuli to faces and people, since face perception doesn't require top-down control (or not even awareness, e.g. Stein et al., 2012, Cognition). Indeed, we agree with the reviewer’s point. We actually suggest that the higher- level object identity processes (perhaps further eye-movements) may require some further processing, which is slower than the basic simple feature detection that triggers Posner-like orienting. Furthermore, we did not intend to generalise from our simple geometric and ambiguous figures directly to human faces and people generally. Rather we are investigating the influence of mechanisms processing lower-level features and implied ‘face/front’ by social priming. Note that we have been quite conservative in our General Discussion and suggest simply that higher-level social processes may have less influence at the earliest stages of processing than at later ones 3\. Another assumption here is that default percepts of fronts reflect lower- level visual features; but I don't think this can be taken for granted. For example, the fact that seagulls facing each other are prioritized in Expt. 5 could be an effect of the interpretation of seagulls as seagulls (in both conditions, if the prime is inactive as per points \#1 and \#2) and it is in fact their perceived facing direction that is driving attention. Participants in Experiment 5 were assigned to either a seagull or a rabbit condition and all stimuli were referred to accordingly for each participant. There was no priority for seagulls. The experiment was counterbalanced as described in Methods. The target was indeed driving attention with faster responses to seagull-inwards facing pairs regardless of prime condition being seagull or rabbit front. (Relatedly, the claim on p.7 that the stimuli “contain no intrinsic visual features that could be construed as a face” and have “no salient intrinsic face- like features” doesn’t seem right to me, since the seagulls/rabbits have a beak, ears, and eye.) We apologise. It was not clear this was actually referring to the stimuli in a different experiment, where the stimuli were presented in motion (teardrops and symmetrical shape). This has been amended. This also applies to Expt. 2, as some authors have suggested that arrows can in fact be socially meaningful (e.g. Kingstone et al., 2003) or suggest an agentic presence, especially when multiple stimuli point towards the same region of space (e.g. Gao et al., 2010, Psych Sci; Takahashi et al., 2013, Front Hum Neurosci; Colombatto et al., 2019, Perception). Since these results as described have potential implications for that literature, it would be great to see this discussed. It remains debatable whether such stimuli have social properties. As we note, low-level shape properties imply/afford directional action based representations, such as arrows and the shape of diving gannets, and object motion patterns also afford directional cues even without social content (a falling leaf or moving cloud). Furthermore, Kingstone focusses more on arrows being socially meaningful rather than being social agents i.e. directing attention by feature rather than social interpretation. Gao’s work uses multiple moving arrows whose attention is towards a target square whereas Takahashi’s and Colombatto’s work uses multiple moving cones whose ‘attention’ is towards the participant observer. We feel these publications are quite different in focus than our current work which is exploring ascribed sociality to objects which can have feature cues in the opposite direction to the ‘social direction’. Further, it is difficult to project our findings onto less time-pressured situations as we point out in our discussion and as you indicate elsewhere in review. 4\. One of the reasons social binding was initially thought to be a social effect was that it vanished with inverted faces that are equivalent to upright faces in lower-level properties, but differ in the higher-level social properties (although related to my point \#1, participants would still be able to indicate whether people are facing each other, even when they are inverted!). Those findings contradict the conclusion that lower-level properties only drive social binding – and since this work directly follows up on those experiments, it would be interesting to see the authors discuss this. The reviewer is correct that related findings have previously been interpreted as social effects due to the sensitivity of the effects to inversion. This was thought to account for low-level features, such as symmetry, distance, and centre of mass, and therefore rule out lower-level mechanisms based on such features. However, more recently an explanation for such apparent social effects has been proposed based on directional cueing (Vestner, Gray, & Cook, 2020). This explanation suggests gaze cueing effects elicited by faces to be the cause for "social" interaction effects. These have been shown to be sensitive to stimulus inversion (Vestner, Gray, & Cook, 2021) while not requiring the stimulus itself to be perceived as social. It was later confirmed that these directional cueing effects are domain general and not specifically social, and can be extended to any number of non-social but directionally cueing objects (Vestner, Over, Gray, & Cook, 2021; Vestner, Over, Gray, Tipper, & Cook, 2021). Minor comments 1\. The main text for both Expts. 6 and 7 describes a ‘yellow’ vs. ‘purple’ manipulation (also depicted in Figs. 5, 14-15). But then Fig. 9 depicts ‘orange’ vs. ‘purple’ conditions, and the datafiles report ‘orange’ and ‘purple’ for Expt. 6, and ‘rabbit’ and ‘orange’ for Expt. 7 (in both the RT and MT files); could these be made consistent? Thank you for pointing out this inconsistency. We have changed any ‘orange’ to ‘yellow’ in all documents. 2\. How were subjects assigned to the prime conditions – were they alternating? (This would be helpful to clarify as potentially related to block order assignment.) Thank you for identifying this omission. Participants were indeed alternated (now described in Experiment 2 \> Task blocks and social priming). 3\. How were the prime videos generated for the ‘round-face’ conditions? Were the stimuli simply mirrored/rotated? I ask because this seems to have produced some artifacts, e.g. the ‘round-faces’ overlap at the end of the videos and during the ball tossing game, which would be weird for agents, or the ‘round- face’ stimuli either toss the ball at a distance (rightmost), or overlap with the ball (leftmost). Yes the teardrop stimuli were rotated 180° about their centre. We recognise that there are some gaps and overlaps of ball to characters during ‘throwing’ and between characters during close interaction but we don’t believe that this would be sufficient to break the illusion that the characters have intentions, opinions and feelings. Indeed, during initial testing of our stimuli observers always reported the animate nature of the objects in all versions of the videos (see video demos in OSF). 4\. The target orientation question for Exp. 2 is phrased in the instructions as “You need to pick which pair of seals are facing each other”, but I don’t think ‘seals’ was ever used elsewhere in Exp. 2. How could participants answer this question? Well spotted! This was a copy/paste error. We have checked the experiment code and the script for the orientation question in Experiment 2 reads “You need to pick which pair of characters are facing each other”. We have the instructions document. 5\. This statement seems inaccurate/controversial: “higher-level mentalising processes are at play at later stages as suggested by \[…\] behavioural studies showing that the effects of mentalising on gaze following are slower non- automatic”: the matter is still debated, but recent evidence showing the contrary should be cited, e.g. that mentalising such as perspective taking can occur rapidly and automatically (e.g. Ward et al., 2019, Curr Bio), and that percepts of mental states can influence lower-level processes such as gaze cueing, even quickly and automatically (e.g. Colombatto et al., 2020, PNAS). Thank you for making us aware of this new work which we now cite in the manuscript. 6\. For additional evidence that the pointy end of the teardrop is seen as its front, the authors may find helpful Chen & Scholl, 2018, PB&R – who used teardrop stimuli to elicit the same effects (on aesthetic preferences) as other types of fronts. (I wouldn't normally mention too many papers from our group in a review, but each one here seems warranted as potentially helpful or directly relevant to the discussion.) Thank you for this suggestion. We assume the reviewer is referring to Experiment 3 of the 2014 paper ‘Seeing and liking: biased perception of ambiguous figures consistent with the “inward bias” in aesthetic preferences’. We have added this reference to our introduction. 10.1371/journal.pone.0258832.r003 Decision Letter 1 Steinborn Michael B. Section Editor 2022 Michael B. Steinborn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 1 Sep 2021 PONE-D-21-06244R1 Rapid detection of social interactions is the result of domain general attentional processes PLOS ONE Dear Dr. Flavell, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Editor comment. Two of the previous referees commented again on your manuscript. As you can see from the reviews, they are positive about the new version and there are only some minor issues remaining. I would ask you to consider these points in a final revision of your manuscript, and I think further reviewer rounds will not be necessary. Therefore, we invite you to submit a revision of the manuscript that addresses the remaining points together with a cover letter that contains point-by-point replies. Please submit your revised manuscript by Oct 16 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Michael B. Steinborn, PhD Section Editor, Cognitive Psychology PLOS ONE Journal Requirements: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. Additional Editor Comments (if provided): \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed Reviewer \#3: (No Response) \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: I feel that the authors have addressed all of my comments adequately and I strongly endorse publication. Reviewer \#3: The authors have addressed most of my earlier comments; I have a couple additional suggestions, but in general I believe with their revisions (along with their responses to other reviewers' comments), the manuscript is now in a better shape. Regarding my point about labeling vs. perceiving social interactions, I see that the authors edited the example I provided, and I appreciated that the new introductory paragraph was worded accordingly. I disagree however that Experiment 2 rules out possible contributions of demand characteristics, since it looks like the instructions and question still mentioned “characters” that were “facing towards each other” – which is of course less apparent as a manipulation compared to speech bubbles, but could still plausibly be perceived as such. In the previous round of review, I had discussed the results of Expt. 5, namely that seagulls facing each other are prioritized in visual search (and how it is consistent with a ‘social’ interpretation: if during the search task participants automatically perceive the stimuli as seagulls – regardless of which prime they had been exposed to – then the advantage for inward seagulls might reflect a prioritization of ‘facing’ vs. ‘non-facing’ dyads). The authors rule out this interpretation mentioning that (1) condition assignment was randomized between subjects, and (2) “there was no effect of seagulls”. For (1), I am not sure how random assignment might speak to the possibility that all participants might perceive the stimuli as seagulls during the search task, regardless of previous priming; and (2) also confused me, since my understanding is there was no effect of prime, but there was an effect of stimulus orientation for seagulls (faster for facing vs. non-facing seagulls, i.e. faster for non- facing vs. facing rabbits). I think this might just be a misunderstanding, but if not (i.e. if there was instead no effect of facing seagulls), the results section for Expt. 5 should be clarified. I might have missed this, but I don’t think the stimulus generation method for the ‘round-face’ videos has been clarified in the manuscript? I still see that the conditions of Expt. 7 labeled as ‘rabbit’ and ‘orange’ in the RT file, the MT file, and the data exclusion file. The averages for Expt. 2 in the raw data file seem to be inconsistent with the figure (i.e. mean RT for inward point = 1014ms, inward round = 1003ms, but the depicted mean in Fig. 10 is higher for round). (As an aside, one would naturally associate pointed vs. rounded shapes with the point vs. round condition, but pointed vs. rounded shapes actually signal SOA in the figure, while point vs. round condition is signaled by color.) The OSF stimuli folder for Exp. 6 contains priming videos and search targets – but wasn’t Exp. 6 simply a cueing task? Apologies if I missed this in the manuscript, but it should be clarified that RT was computed as the time of spacebar release (i.e. without the reaching movement). Both the main text and supplement state that in Expt. 1, 33 participants were tested and 26 remained following exclusions; but of the (presumably 7) excluded subjects, only 6 are mentioned (2 based on accuracy, and 4 based on RT). \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Christina Breil Reviewer \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0258832.r004 Author response to Decision Letter 1 21 Sep 2021 Please see 'Response to Reviewers.docx' for formmated version. RESPONSE TO REVIEWERS Dear Dr Steinborn, Once more we’d like to thank you and the reviewers for the time and further comments concerning our paper “Rapid detection of social interactions is the result of domain general attentional processes”. We appreciate the opportunity to clarify and re-submit the work after. In the following we describe our response to the reviewer feedback. Reviewer comments are shown in black and our responses are shown in blue. Yours sincerely, Jonathan Flavell, Harriet Over, Tim Vestner, Richard Cook, Steven Tipper -o0o- REVIEWER \#1 I feel that the authors have addressed all of my comments adequately and I strongly endorse publication. Thank you for your time and for your recommendation. REVIEWER \#3 The authors have addressed most of my earlier comments; I have a couple additional suggestions, but in general I believe with their revisions (along with their responses to other reviewers' comments), the manuscript is now in a better shape. Thank you for another thorough and considered review once again. Regarding my point about labeling vs. perceiving social interactions, I see that the authors edited the example I provided, and I appreciated that the new introductory paragraph was worded accordingly. I disagree however that Experiment 2 rules out possible contributions of demand characteristics, since it looks like the instructions and question still mentioned “characters” that were “facing towards each other” – which is of course less apparent as a manipulation compared to speech bubbles, but could still plausibly be perceived as such. To clarify, in our previous Response to Reviewers we “…the question of front/face may not be as susceptible to demands…” rather than stating ruling them out entirely (which the present set of studies cannot do conclusively). Note that we have added some additional text to Experiment 2 \> Results and Discussion. In the previous round of review, I had discussed the results of Expt. 5, namely that seagulls facing each other are prioritized in visual search (and how it is consistent with a ‘social’ interpretation: if during the search task participants automatically perceive the stimuli as seagulls – regardless of which prime they had been exposed to – then the advantage for inward seagulls might reflect a prioritization of ‘facing’ vs. ‘non-facing’ dyads). The authors rule out this interpretation mentioning that (1) condition assignment was randomized between subjects, and (2) “there was no effect of seagulls”. For (1), I am not sure how random assignment might speak to the possibility that all participants might perceive the stimuli as seagulls during the search task, regardless of previous priming; and (2) also confused me, since my understanding is there was no effect of prime, but there was an effect of stimulus orientation for seagulls (faster for facing vs. non-facing seagulls, i.e. faster for non- facing vs. facing rabbits). I think this might just be a misunderstanding, but if not (i.e. if there was instead no effect of facing seagulls), the results section for Expt. 5 should be clarified. Re. (1). Perhaps we had crossed wires. We thought that your description of seagulls as “prioritized in Expt. 5” was implying a methodological prioritisation. Re. (2). You are correct in that RTs were affected by seagull orientation but not by prime type. We cannot find the quote “there was no effect of seagulls” in our previous Manuscript or Response to Reviewers. Assuming you meant our words “There was no priority for seagulls” in Response to Reviewers, this is explained by the above point. We have edited the text in Experiment 5 \> Results and Discussion and hope that this clarifies our own view on these data. I might have missed this, but I don’t think the stimulus generation method for the ‘round-face’ videos has been clarified in the manuscript? We did not realise that you were requesting an edit to the manuscript in your previous comment. We have now added the following to Experiment 2 \> Method \> Task blocks and social priming: “The pointed and rounded videos were identical apart from the orientation of the teardrop which was mirrored to create a pointed or rounded prime.” I still see that the conditions of Expt. 7 labeled as ‘rabbit’ and ‘orange’ in the RT file, the MT file, and the data exclusion file. Thankyou for identifying this error. These documents have now been corrected. The averages for Expt. 2 in the raw data file seem to be inconsistent with the figure (i.e. mean RT for inward point = 1014ms, inward round = 1003ms, but the depicted mean in Fig. 10 is higher for round). (As an aside, one would naturally associate pointed vs. rounded shapes with the point vs. round condition, but pointed vs. rounded shapes actually signal SOA in the figure, while point vs. round condition is signaled by color.) The figure has been corrected. The data are corrected but the figure labels were switched accidentally. We have corrected this and taken the opportunity to switch the icons in Figure 10 so that circles are for the rounded front condition and the triangles are for the pointed front condition. The OSF stimuli folder for Exp. 6 contains priming videos and search targets – but wasn’t Exp. 6 simply a cueing task? You are correct. There were no priming videos in Experiment 6 and these were uploaded in error. The six files have been removed from the OSF. Apologies if I missed this in the manuscript, but it should be clarified that RT was computed as the time of spacebar release (i.e. without the reaching movement). Thankyou for identifying this omission. We have added the following text to Experiment 2 \> Method \> Practice & trial composition: “Reaction times were measured from the moment the four simultaneously presented pairs appeared to the moment of space bar release. Movement time is measured from the moment of space bar release to the moment of screen contact.” Both the main text and supplement state that in Expt. 1, 33 participants were tested and 26 remained following exclusions; but of the (presumably 7) excluded subjects, only 6 are mentioned (2 based on accuracy, and 4 based on RT). We have double checked and it was the recruited size that was in error. It has been corrected to “Thirty-three two participants were tested…”. Thank you for identifying this mistake. 10.1371/journal.pone.0258832.r005 Decision Letter 2 Steinborn Michael B. Section Editor 2022 Michael B. Steinborn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 7 Oct 2021 Rapid detection of social interactions is the result of domain general attentional processes PONE-D-21-06244R2 Dear Dr. Flavell, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. Kind regards, Michael B. Steinborn, PhD Section Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0258832.r006 Acceptance letter Steinborn Michael B. Section Editor 2022 Michael B. Steinborn This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 7 Jan 2022 PONE-D-21-06244R2 Rapid detection of social interactions is the result of domain general attentional processes Dear Dr. Flavell: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. If we can help with anything else, please email us at <plosone@plos.org>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Michael B. Steinborn Section Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction The impact of research is fundamentally dependent on how well it is disseminated to the end-user. Conventional routes of dissemination involve journal publications, conference presentations and, ultimately although often years later, textbooks. This model of dissemination requires the end-user to search for, or ‘pull’, the relevant knowledge from the literature base. With regard to health and medical research, this approach might be ineffective because the end- users are often clinicians who do not subscribe to journals, nor attend conferences. The rise of open access publication reduces one barrier to effective dissemination by making literature freely available for all who wish to consult it, but it still relies on the end-user pulling out the relevant knowledge. The rapid rise in popularity of web logs (blogs) and social media sites such as Facebook and Twitter, has positioned them as critical tools with which to aid dissemination. Health and medical research is no exception - high profile journals such as the New England Journal of Medicine (NEJM) and the British Medical Journal (BMJ) have established cohesive digital strategies that incorporate both blogs and social media sites, presumably in the hope of improving the dissemination of knowledge. This approach contrasts with the pull approach insofar as it ‘pushes’ the knowledge to the end-user. By having different blog and social media sites, journals allow the end-user to self- select the genre of knowledge they wish to receive. RSS (Really Simple Syndication) is another example of how users can self-select information. Although not a pure social media tool, RSS feeds enable the pushing of individualised information and blog contents. RSS permits some user interaction and information sharing. The fundamental importance of a digital strategy is emphatically stressed by social media advocates. Markers such as the number of ‘likes’, or the number of Facebook or Twitter followers are cited as measures of research impact, collectively captured by concepts such as ‘altmetrics’. We contend, however, that the most common altmetrics are not measuring impact, insofar as impact relates to the effect of research on clinical practice or thinking. Moreover, the definitions of various terms are not clear and they mean different things to different people. For the purposes of this experiment, we define the key concepts as set out in. We took ‘reach’ to be the number of people who have been alerted to the presence of a web page and have the opportunity to view it. Reach reflects the number of people who could potentially see the blog, either directly because they subscribe to the blog through RSS feed or email alerts, or through following the blog on various social media sites for example Facebook, Twitter, LinkedIn, Google+ or ResearchBlogging. One step closer to impact is engagement, defined here as the number of people who view the web page and then do something in response to viewing it – for example they ‘like’ it, re-tweet it, or they share it with their friends. The concept of ‘virality’ attempts to capture a stronger level of engagement and a reflection of the propensity of the message to ‘go viral’. Here we use the percentage of engagers who then write a story on the post on Facebook or begin a new tweet. This distinction between terms is important because as few as 16% of Facebook followers actually read a new post and about 1% of people who see and ‘like’ a Facebook page actually comment on it or start a new story on it. A substantial gap in our understanding of the link between social media and impact, is the effect that a social medial release about a research article has on dissemination of the article itself. Remarkably, despite the apparent acceptance of social media reach, engagement and virality being evidence of impact, there seems to be no empirical evidence to support this claim. This observation was recently noted by Priem et al - ‘Researchers must ask if altmetrics really reflect impact, or just empty buzz’. We undertook a blinded, randomised repeated measures experiment to test the hypothesis that social media release of an original research article in the clinical pain sciences increases viewing and downloads of the article, thereby demonstrating increased dissemination of the research and end-user behavioural change. # Methods Sixteen original research articles were selected from the PLOS ONE group of journals. Inclusion criteria were: (i) relevance to the clinical pain sciences; (ii) of interest to the readership of our research group’s blog (bodyinmind.org), a readership that consists primarily of clinicians who work in a pain-related field; (iii) first published on-line between 01/01/2006 and 31/12/2011; (iv) not previously mentioned in a bodyinmind.org blog post. Research articles were randomly allocated to four researchers in our group, each of whom wrote a blog post of around 500 words based on the original article, and which included a tag line directing the reader to the on-line version of the article for more information. All posts were released on a Tuesday (between 6 and 7 am) or between 11 pm Thursday and 2 am Friday, Australian Eastern Summer time. Other posts, not part of the current experiment, were also released during the experimental period (14/08/2011–02/02/2012). For each blog on a research article, two dates were randomly selected from all possible post-dates during the experimental period. Of the two dates, one was randomly selected as the release date and one as the control date. Each blog post was broadcast via ResearchBlogging.org, Facebook, Twitter and LinkedIn on the day of the blog post. The experiment was undertaken covertly, so there was no risk that end- users who knew the experiment was being conducted would visit the original article as a result of that knowledge. The primary outcome variables were rate of HTML views and PDF downloads over a seven day period. The former reflects some engagement with the target article by visiting it on the PLOS website. The latter reflects a higher level of engagement with the target article by adding it to a user library, presumably for future reference. Both outcome variables represent a behavioural change associated with the target article. Each primary outcome variable was assessed during four seven-day periods: the seven days before and after the release date, and the seven days before and after the control date. Facebook statistics were provided by ‘Facebook page insights’ for 28 days post publication of the blog post. They are not broken down into individual days. Twitter comments and re-tweets were searched for manually and relied on the Twitter search engine to identify all mentions. ## Statistical Analysis We undertook a 2×2 repeated-measures ANOVA on each primary outcome variable. The first factor was ‘Date’ (two levels: release date or control date). The second factor was ‘Week’ (two levels: before date or after date). In order to maximise the likelihood of detecting an effect on each primary outcome variable, which we took to reflect different levels of dissemination, we did not correct for multiple measures and set α = 0.05. ## Secondary Analyses - Relating Citations, HTML Views and PDF Downloads to Social Media Reach and Engagement We calculated the relationship between the primary outcome variables and recognised measures of social media reach and social media engagement. We undertook two linear regressions with the increase in HTML views or PDF downloads as the dependent variable, and the following measures of social media reach and engagement as the independent variables: ### Reach The number of unique visitors who were alerted to the blog post and had the opportunity to view it. ### Engagement The number of unique people who liked, commented on, or shared the blog post on [www.bodyinmind.org](http://www.bodyinmind.org), Facebook, Twitter or LinkedIn. ### Virality The percentage of unique viewers who then created a story from the blog post on Facebook, twitter, or blogged about it separately. We investigated whether social media reach or engagement related to a conventional measure of impact - citation count, as provided by Scopus. We did this using a third linear regression, with reach and engagement as regressors, and citation count at 03/09/2012 as the dependent variable. We also investigated whether HTML views or PDF downloads related to citations by correlating citation count at 03/09/2012 with total HTML views and total PDF downloads at the end of the week after the social medial release. We tested whether there was a ‘blogger effect’ (ie, do some blogger’s posts have a greater impact than others?) by first calculating the difference in the change or rate of increase in the primary outcome variables between the social media release date and the control date. We called this the blog effect. We then compared the blog effect between reviewers using a Kruskal-Wallis test. Finally, we tested whether there was an ‘age effect’ (ie, is there an effect of the age of the article on our outcome variables?) by relating the blog effect to the days between publication of the article and the social media release. No correction was applied for multiple measures because these were secondary and therefore exploratory, hypothesis-generating analyses. # Results Over the 18-week study period, the blog (bodyinmind.org) had an average of 2585 unique views per week. Each post was viewed a mean (SD) of 507 (160) times in the week following publication. In the 28 days after publication, a mean (SD) of 693 (135) unique visitors saw the post in their Facebook newsfeed; 35 (16) unique visitors clicked on each post; 6 (4) unique visitors created a like, comment or share from the post. Of the total number of unique visitors who saw the post on Facebook, 0.93% (0. 66%) created a story from it. ## HTML Views The rate of HTML views was higher during the second week than during the first, regardless of the date. That is, there was a main effect of Week on HTML views (F(1,15) = 6.27, p = 0.024). The rate of HTML views was also higher either side of the social media release than it was either side of the control date (main effect of Date on HTML views – F(1,15) = 7.39, p = 0.016). However, visual inspection of the data show that these main effects were driven to a large extent by an interaction, such that the social media release was associated with a larger increase in the rate of HTML views than the control date was (Week x Date interaction: F(1,15) = 7.39, p 0.016). The mean ± SD rate of HTML views in the week after the social media release was 18±18 per day, whereas the rate during the other three weeks was no more than 6±3 per day, which equates to an effect size (Cohen’s d) of 0.9. That is, in the week after the social media release, about 12 people per day viewed the research article as a result of the social medial release. ## PDF Downloads The results for PDF downloads reflected those for HTML views: The rate of PDF downloads was higher during the second week than during the first (main effect of Week – F(1,15) = 10.83, p = 0.005) and higher either side of the social media release than it was either side of the control Date (main effect of date – F(1,15) = 6.57, p = 0.022). Again, these effects were driven by an interaction, such that the social media release was associated with a larger increase in the rate of PDF downloads than the control date was (Week x Date interaction: F(1,15) = 14.74, p = 0.002). The mean ± SD rate of PDF downloads in the week after the social media release was 4±4 per day, whereas the rate during the other three weeks was less than 1±1 per day, which equates to an effect size (Cohen’s d) of 1. That is about 3 people per day downloaded the research article as a result of the social media release. ## How well do reach, engagement and virality of the social media release relate to HTML views and PDF downloads of the research article? Engagement was 5.3% of reach and virality was 0.9% of engagement. None of the social media metrics related to the increase in rate of HTML views of the research article (p = 0.947 for reach; p = 0.809 for engagement; p = 0.544 for virality), nor to the increase in PDF downloads of the research article (p = 0.323 for reach; p = 0.864 for engagement; p = 0.934 for virality). The only relationship that approached significance was that between the number of HTML views of the blog post and PDF downloads (p = 0.09). ## Relationship between Reach, Virality and Citations There was no relationship between citations on Scopus about one year after publication and any of the social media metrics (p\>0.68 for all). Total PDF downloads at the end of the week after social media release related to total HTML views at the same time (Pearson r = 0.72; p = 0.002). Interestingly, citations at 03/09/2012 related to total PDF downloads (Pearson r = 0.51; p = 0.045) but not to total HTML views (Pearson r = 0.06; p = 0.826). ## Was there a ‘blogger’ Effect? One blogger wrote seven posts, two wrote four posts and one wrote one post. There was no difference between bloggers for either HTML views or PDF downloads (p\>0.88 for both). ## Was there an Article Age Effect? The age of the article at the time of blogging was not related to the rate of HTML views, or the rate of PDF downloads, during the any of the four one-week periods (p\>0.71 for all). The blog effect was not affected by the age of the article at the time of blogging (p = 0.28). # Discussion We hypothesised that social media release of an original research article in the clinical pain sciences increases viewing and downloads of the article. The results support our hypothesis. In the week after the social media release, there were about 12 extra views of the HTML of the research article per day, and 3 extra downloads of the article itself per day, that we can attribute to the social media release. The effects were variable between articles, showing that multiple factors mediate the effect of a social media release on our chosen outcome variables. Although the absolute magnitude of the effect might be considered small (about 0.01% of people we reached were sufficiently interested to download the PDF), the effect size of the intervention was large (Cohen’s d \>0.9 for both outcomes). The effect of social media release was probably smaller for our site, which is small, young and specialised, than it would be for sites with greater gravitas, for example NEJM or BMJ or indeed, PLOS. ## Relationship between Reach and Impact The idea of social media reach is fairly straightforward - it can be considered as the number of people in a network, for example the number of Facebook friends or Twitter followers. A blog may have 2,000 Facebook ‘likes’, 700 Twitter followers and 300 subscribers - a reach of three thousand people. Impact is less straightforward. As depicted in, the various definitions of social media each reflects a substantially larger population than our most proximal measure of impact – HTML views and PDF downloads of the original article. One might suggest that impact should reflect some sense of engagement with the material, for example the number of people within a network who make a comment on a post. From a clinical pain sciences perspective, change in clinical practice or clinician knowledge would be clear signs of impact, but such metrics are very difficult to obtain. Perhaps this is part of the reason that researchers are using, we believe erroneously, social media reach as a measure of social media impact. There are now several social media options that researchers integrate into their overall ‘impact strategy’, for example listing their research on open non- subscription sites such as Mendeley, and joining discussions about research on social media sites such as Twitter and on blogs. Certainly, current measures of dissemination, most notably citations of articles or the impact factor of the journals in which they are published, do not take into account the social media impact of the article. New measurements, such as altmetrics and article-level metrics such as those provided by PLOS, aim to take into account the views, citations, social network conversations, blog posts and media coverage in an attempt to analyse the influence of research across a global community. There is merit in this pursuit, but, although our study relates to clinical pain sciences research, our results strongly suggest that we need to be careful in equating such measures with impact or influence, or using them as a surrogate for dissemination. Indeed, not even virality, which estimates the propensity of an item to ‘go viral’, was related with HTML views or PDF downloads. This is very important because our results actually suggest that we may be measuring the wrong thing when it comes to determining the social media impact of research. That is, we showed a very clear effect of the social media release on both HTML views and PDF downloads of the target article. However, we did not detect any relationship between either outcome and the social media metrics we used. The only variable that related to either outcome was the number of HTML views, of the original blog post, in the week after social media release. It seems clear then, that it is not the total number of people you tell about your study, nor the number of people they tell, nor the number of people who follow you or who re-tweet your tweets. In fact, it appears that we are missing more of how the release improves dissemination than we are capturing. The final result, that citation count did not relate to any social media measures, casts doubt over the intuitively sensible idea that social media impact reflects future citation-related impact. We used the Scopus citation count, taken almost 9 months after the completion of the experimental period, and 1–2 years after the publication date of the target articles, as a conventional measure of impact. There was no relationship between citation count and our measures of social media reach or virality. One must be cautious when interpreting this result because citation count so soon (1–2 years) after publication might be unlikely to capture new research that was triggered by the original article – although, importantly, journal impact factors are calculated on the basis of citations in the two years after publication. Suffice here to observe that the apparent popularity of an article on social media does not necessarily predict its short-term citation count. Although this is the first empirical evaluation of social media impact in the clinical pain sciences and we have employed a conservative and robust design, we acknowledge several limitations. Social media dissemination in the clinical sciences relies on clinicians having access to, and using, social media. It will have no effect for those who do not use the web and who rely on more traditional means of dissemination - ‘pulling’ the evidence. Although there was an increase in HTML views and PDF downloads as a result of social media dissemination, we do not know if people read the article or whether it changed their practice. We presumed that a portion of those who viewed the HTML version of the article would then go onto download it, however our data suggest that a different pattern of access is occurring. Unfortunately, our data do not allow us to determine whether the same people both viewed the HTML and downloaded the article PDF or whether different people viewed the HTML and downloaded the article PDF. Downloading a PDF version of a paper does not necessarily imply that they would later read it, but it does increase the probability of such. Citations and impact factors measure the impact within the scientific community whereas views by social media will also include interested clinicians and laypeople and, as such, measure uptake by different audiences. Although we used a variety of different social media platforms to disseminate to as wide an audience as possible, we do not know who the audience is - we can only surmise that they are a mixture of researchers, clinicians, people in pain and interested laypeople. Further, each social media strategy comes with inherent limitations in regards to data collection of usage statistics related to a blog post. Gathering Facebook and Twitter statistics for each article is still cumbersome and is probably not always accurate. The risk in using search engines to gather data is that there is no way of knowing whether all the data have been identified. For Twitter there is no way to retrospectively calculate the number of re-tweets accurately over a longer period retrospectively for each post. As a result, our Twitter data is a best estimate and my have underestimated the true values but, critically, we would expect this effect to be unrelated to our blog post and therefore not impact on our findings. Regarding Facebook, shares, likes and comments are grouped as one statistic but in reality only shares and comments show engagement with the post and indicate that people are more likely to have read it. Regarding LinkedIn, the only available data was the number of members of the BodyInMind group and as such, we have no way of knowing how many viewed the actual blog post. The blog, BodyInMind.org, through which the original blog posts of PLoS ONE articles were released, experienced a technical interruption half-way through the experiment. In spite of an attempt by PLOS to retrieve the statistics, approximately five days of data were lost on several of the blog posts. This also meant that additional data on traffic, such as percentage of traffic for each blog post from external sources such as Facebook, Twitter, LinkedIn and ResearchBlogging could not be measured during this period. Critically and fortuitously, this period did not coincide with data collection weeks. PLOS indicated that this technical problem has now been fixed, but similar problems may arise in the future and present an ongoing risk to studies such as ours. Although disconcerting for those keenly following social media data, this problem would be very unlikely to have affected our primary outcomes because none of our dates fell within the period that was affected. Social influence can produce an effect whereby something that is popular becomes more popular and something that is unpopular becomes even less popular. It seems possible that articles on BodyInMind.org were shared because the site is popular among a discrete community and not because the article itself merited circulation. This possibility does not confound our main result but it adds a possible argument to the common objective of making a blog more popular as a device to boost social media impact of individual posts. Finally, our study relied on the target articles being freely available to the public. Many journals are not open access, particularly those in the clinical pain sciences. Therefore, we must be cautious extrapolating our results to subscription only access journals. In conclusion, our results clearly support the hypothesis that social media can increase the number of people who view or download an original research article in the clinical pain sciences. However, the size of the effect is not related to conventional social media metrics, such as reach, engagement and virality. Our results highlight the difference between social media reach and social media impact and suggest that the latter is not a simple function of the former. # Supporting Information [^1]: This experiment was conducted using [www.bodyinmind.org](http://www.bodyinmind.org). Web metrics of this website are used as evidence of the authors’ research social media reach. The authors therefore stand to benefit indirectly should this manuscript increase the reach of that website. This does not alter the authors' adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: GLM HGA TRS FDP. Performed the experiments: GLM HGA TRS FDP. Analyzed the data: GLM. Wrote the paper: GLM HGA TRS FDP.

# Introduction *Mycobacterium tuberculosis* (*M. tuberculosis*), firstly discovered by Robert Koch, is a pathogenic species and the causative agent of most tuberculosis. The World Health Organization (WHO) has recognized the global threat imposed by *M. tuberculosis*, and statistics show that about one-third of the world's population has been infected. It was reported by the WHO that the increasing rate of new clinical cases was 8 million each year, with at least 3 million people deaths. *M. tuberculosis* has an unusual, waxy coating on the cell surface (primarily mycolic acid), which highlights that there must be a large number of enzymes involved in lipid metabolism. In 1998, the whole genome of *M. tuberculosis* H37Rv strain was sequenced by the Sanger Center and the Institut Pasteur, showing at least 250 enzymes related to lipid metabolism including extracellular secreted enzymes, integrated cell wall enzymes and intracellular esterases/lipases, compared with about 50 enzymes in *E. coli*. The genomic organization and gene functionality of *M. tuberculosis* are invaluable for understanding the slowly growing pathogen. Mycobacterial genes that are involved in lipid metabolism, cell division chromosomal partitioning, and secretion are more likely to be required for survival in mice. Lamichhane and colleagues detected 31 *M. tuberculosis* genes that were found to be required for *in vivo* survival in mouse lungs. Mutation of six of the Mycobacterial membrane protein (mmpL) family genes severely compromised the ability of the respective mutants to multiply in mouse lungs. In 2007, a *M. tuberculosis* CDC1551 (or Rv2224c of H37Rv) gene, *MT2282*, was identified as a virulence gene belonging to the microbial esterase/lipase family with an active site consensus sequence of G-X-S-X-G. In fact, the esterase was a cell wall-associated carboxyl esterase rather than a protease as initially annotated. Further research found that the MT2282 esterase was required for bacterial survival in mice and full virulence of *M. tuberculosis*. The Rv0045c protein is a putative hydrolase, probably involved in ester/lipid metabolism of *M. tuberculosis*. Alignment among amino acid sequences showed that the Rv0045c protein shares little amino acid sequence similarity with members of the esterase/lipase family identified in *Bacteria, Archaea, Eukaryotes* and some viruses, such as Aes acetyl-esterase from *E. coli*, and *mosquito* carboxylesterase Estα2¹ (A₂). Here, we experimentally characterized the Rv0045c protein via protein expression, purification, biochemical characterization and enzyme activity analysis, and finally demonstrated that Rv0045c is a novel esterase in *M. tuberculosis*. # Results ## Expression and purification of the Rv0045c protein In order to allow easy purification and to attenuate the effect of a large tag on the biological activity of the Rv0045c protein, a 6×His-tag was chosen and added to its N-terminal. The fusion protein was overexpressed at 37°C and induced with 1 mM IPTG. SDS-PAGE analysis showed a major protein band with the expected 35.5 kDa size, but the recombinant Rv0045c protein was in form of inclusion bodies (data not shown). To make purification easy and to maintain the biological activity of the recombinant protein, the expression condition was optimized by raising the major fraction as a soluble protein under a feasible condition with 0.3 mM IPTG at 16°C (lane 3). First, we purified the soluble protein from supernatant using Ni²⁺-affinity chromatography (lane 5 to lane 8). Subsequently, the eluted protein was concentrated, and loaded onto an anion exchange chromatography column and a cation exchange chromatography column. Finally, the protein was further purified through gel filtration chromatography to \>98% purity. ## MALDI-TOF mass spectrometry analysis of the Rv0045c protein We analyzed the purified Rv0045c protein by mass spectrometry. The MALDI-TOF MS spectrometry of the digested protein is shown in. The peptide mass fingerprinting (PMF) of the protein was observed and submitted to Mascot. Consequently, only NP_214559 protein from *M. tuberculosis* was obtained as a result with a score of 112. The results provided convincing evidence that the purified Rv0045c protein is the NP_214559 protein from *M. tuberculosis.* ## Circular dichroism spectroscopy analysis of the Rv0045c protein To gain insight into the secondary structural elements in the Rv0045c protein, a circular dichroism (CD) spectroscopy was collected in the wavelength range from 240 to 190 nm at room temperature (25°C) and the pH range 2.0–12.0 (with an interval of 1.0, except pH 5.0 because the protein precipitates and may be too close to its pI).. The curves converged together in the range pH 6.0–10.0, but were nevertheless distorted and disordered at extreme pH (≤pH 4.0 and ≥pH 11.0). Near physiological conditions (at pH 7.0 and pH 8.0), the protein was much more stable and the negative trough at 216 nm with crossover at 195 nm is the characteristic feature of β-sheet secondary structure. The native state of the protein was estimated to entail 11∼14% α-helix, 54∼60% β-sheet, 4∼8% turn and 24∼26% random region, measured according to Yang and colleagues. The high β-sheet content suggested that the Rv0045c protein possessed abundant β-sheet secondary structures, which is in accordance with the α/β-hydrolase fold and implied that the Rv0045c protein may fall into the description of the α/β-hydrolase fold by Nardini and colleagues. In the ranges pH 2.0–4.0 and pH 11.0–12.0, the structure of the protein had been denatured, showing that the conformations were quite different from those at pH 7.0 (as shown). In order to assess the thermal stability of the protein, CD spectra was also collected at various temperatures (ranging from 10°C to 70°C, with an interval of 10°C) with the pH fixed at 7.5. The conformation of the Rv0045c protein was stable at temperatures ≤40°C and the curves converged together. The proportions of α-helix and β-sheet secondary structures at 30°C and 40°C (at 30°C: α  = 10.0%, β  = 61.3%, turn  = 4.0%; at 40°C: α  = 11.0%, β  = 58.1%, turn  = 7.6%) were similar to those for pH 7.0 and pH 8.0 at room temperature (25°C). When the temperature went down to ≤20°C, folding of the protein is consistent with inactivity (data not shown), although the percentages of α-helix and turn (at 20°C: α  = 16.1%, turn  = 14.8%; at 10°C: α  = 20.5%, turn  = 21.1%) notably increased. It was reported that the active site was fully available for substrate binding only when the protein was in the active and open conformation, and hence the Rv0045c protein adopts an inactive closed conformation at low temperatures, causing the enzyme activity to be extremely low. In contrast, when the temperature was increased to ≥50°C, the α-helical secondary structure was lost (e.g. α  = 4.7% at 60°C and α  = 4.8% at 70°C) and curves began to deviate from those for temeratures ≤40°C (as shown), which showed that the structure of the protein had been partially or largely denatured. ## Enzyme activity analysis of the Rv0045c protein Based on the above results, and in order to test whether the Rv0045c protein has esterase activity, we experimentally analyzed the enzyme activity of the Rv0045c protein using *p*-nitrophenyl derivatives (*p*-nitrophenyl acetate (C₂), butyrate (C₄), caproate (C₆), caprylate (C₈), laurate (C₁₂), myristate (C₁₄) and palmitate (C₁₆)) as substrates according to previously described methods. As shown in, at pH 7.0 and 37°C, the Rv0045c protein could hydrolyze a wide range of *p*-nitrophenyl derivative (C₂–C₁₄) substrates, of which *p*-nitrophenyl caproate (C₆) was effectively hydrolyzed. The substrates *p*-nitrophenyl acetate (C₂) and *p*-nitrophenyl myristate (C₁₄) were also visibly hydrolyzed with more than 50% maximal activity. In contrast, no enzyme activity towards longer *p*-nitrophenyl esters (C₁₆) was detected. *M. tuberculosis* is known to presents a certain degree of resistance to aberrant potential of hydrogen. Activity of the Rv0045c protein was examined over a broad pH range from pH 2.0 to pH 12.0. No or poor activity was detected at ≤pH 4.0 and ≥11.0 (data not shown). Based on the CD spectroscopy data, the enzyme displays a conformation-dependent esterase activity, with activity declining dramatically or almost lost at ≤ pH 4.0 and ≥11.0 as a result of the enzyme becoming denatured. Activity was also too low to be detected at pH 9.0 and pH 10.0, for the reason that substrates spontaneously decomposed causing a deep background (data not shown). To determine the dynamic activity of the enzyme, we tested the activity using *p*-nitrophenyl caprylate (C₆) as substrate at certain pH conditions (pH 6.0–8.0) in the temperature range around body temperature (from 36°C to 40°C), respectively. As shown in, the highest enzyme activity at pH 6.0 occurred at 37°C. At both pH 7.0 and pH 8.0, however, the optimal temperature for the enzyme activity is shown to be 39°C, In addition, the activity as a whole and also the highest activity at the optimal temperature exhibited a rapid and dramatic increase along with pH, suggesting that the Rv0045c protein adopted a pH-dependent activity in the pH range from 6.0 to 8.0, and can be described by the electrostatic potential distribution on the enzyme surface at alkaline pH making the substrate-binding and/or hydrolysis more effective. # Discussion Esterases or lipases are types of hydrolases which are widely distributed from prokaryotes to eukaryotes, and which are involved in lipid metabolism. As previously reported, *M. tuberculosis* is understood to contain more than 250 enzymes related to ester/lipid metabolism. In this study, we confirmed that the *M. tuberculosis* Rv0045c protein is a novel esterase. Compared with other esterases in the α/β-hydrolase fold family, two esterases Rv3487c and Rv1399c from the *M. tuberculosis*, both of which have been functionally characterized as esterases, shared no obvious sequence identity to our Rv0045c protein, in a multiple sequence alignment calculated using ClustalW software (data not shown). All esterases in the α/β-hydrolase fold family have a nucleophile-histidine-acid catalytic triad evolved to efficiently operate on substrates with diverse chemical compositions or physicochemical properties. Alignment among amino acid sequences showed that the active site G-X-S-X-G sequence motif within esterases is highly conserved (data not shown), and that the main catalytic residues (Ser89, Asp113, Ser206, His234) in the esterase ybfF are also well conserved in our Rv0045c protein sequence. However, the Rv0045c protein shares as low as 23% amino acid sequence identity with ybfF. Additionally, residues around the active site in ybfF are quite divergent from those in the Rv0045c protein, suggesting that the Rv0045c protein has distinct substrate specificity and catalytic properties that set it apart from other esterases. As with the proteins Rv3487c and Rv1399c, the Rv0045c protein can efficiently catalyze short-chain synthetic substrates (C₂–C₈), but can also hydrolyze *p*-nitrophenyl myristate (C₁₄) with more than 50% of the maximal relative activity. Being the causative agent of most cases of tuberculosis, *M. tuberculosis* infects the lungs of the mammalian respiratory system and can persist in the human body at normal temperature (36°C–37°C) and pH (pH 7.3–pH 7.4) conditions for many decades. Thus, *p*-nitrophenyl acetate (C₆) was used to determine the dynamic activity of the enzyme at mild pH conditions (pH 6.0–8.0) over the temperature range from 36°C to 40°C, which was around body temperature. Compared with the optimum reaction temperature of 30°C for Vlip509, a new esterase from a strict marine bacterium, *Vibrio* sp. GMD509, the optimal temperature for the Rv0045c protein activity turned out to be 37°C at pH 6.0 and 39°C at both pH 7.0 and pH 8.0. This is probably the result of *M. tuberculosis* commonly living in the bodies of humans or animals whereas the *Vibrio* sp. GMD509 marine bacterium parasitizes in the eggs of the sea hare, a cold-blooded animal living at relatively low temperatures. It has also been observed that the average and the highest activity of the enzyme increased rapidly and dramatically following increased pH, indicating that the metabolism of esters/lipids in this pathogen was more active when the circumstances become less favorable, especially more basic, and further suggests that *M. tuberculosis* becomes more pathogenic at alkaline pH. *M. tuberculosis* is pathogenic to humans, and to some extent shows resistance to aberrant hydrogen potential. In our research, enzyme activities were determined over a broad pH spectrum (pH 2.0–12.0), yet little or no activity was detectable at extreme hydrogen potential (≤ pH 4.0 and ≥pH 11.0, data not shown). Results from CD spectroscopy analysis also indicated that, at extreme hydrogen potential (≤ pH 4.0 and ≥pH 11.0), the enzyme is partially or almost completely denatured, especially the α-helical secondary structure. These data suggest that the enzyme activity of the Rv0045c protein is conformation- dependent. Data from CD spectroscopy analysis showed that the Rv0045c protein is rich in β-sheet secondary structure, indicating that the enzyme should possess a very stable and substantial β-sheet core which helps to stabilize the architecture of the enzyme, thus ensuring that the pathogen can survive in strong environments. However, at extreme hydrogen potential (≤ pH 4.0 and ≥pH 11.0), the α-helical secondary structure of the enzyme was mostly denatured, and simultaneously the activity of the enzyme was not detectable. Based on the above evidence, it can be deduced that the β-sheet comprises the skeleton and backbone of the enzyme, while the α-helices or other secondary structure elements, e.g. turns, are required for the catalytic reaction. In addition, the conformation of the enzyme is very stable at temperatures ≤ 40°C, and the thermal denaturing temperature of the Rv0045c protein was determined to be 50°C, which can be utilized for dry heat sterilization to deactivate the enzyme and possibly even the pathogen. The Rv0045c protein is just one of hydrolases involved in ester/lipid metabolism in *M. tuberculosis*, of which many members haven't been identified or haven't been studied. Biochemical characterization and functional analysis of those undefined esterase/lipase members should help to reveal the mechanism of ester/lipid metabolism of *M. tuberculosis*. In order to illustrate the relationship between the tertiary structure and function of the Rv0045c esterase, and to explain the molecular mechanism and principles of the Rv0045c protein participating in hydrolyzing esters, crystallography of the protein is under progress. # Materials and Methods ## Protein expression Based on the template of the whole genome of *M. tuberculosis*, the *Rv0045c* gene was amplified using a standard PCR procedure with primers R1 (5′-CGCGGATCCCTATCTGACGACGAACTGACC-3′, contained a *BamH I* digestible site) and R2 (5′-TCCGCTCGAGTCAGCGTGTGTCGAGCACCCC-3′, attached a *Xho I* site), and subcloned into the *BamH I* and *Xho I* sites of the pET28a vector (Novagen) with *6*×*his-tag* gene on N-terminal. The Rv0045c protein was overexpressed in *E. coli* BL21 (DE3) strain (Novagen) as a fusion protein with the 6×His-tag. Briefly, *E. coli* BL21 (DE3) carrying the *Rv0045c* gene was grown in LB medium at 37°C with 50 µg/mL kanamycin until the OD₆₀₀ reached 0.6–0.8, and then induced with 0.3 mM IPTG at 16°C for 20 hrs. Protein expression was verified by SDS-PAGE analysis. ## Protein purification For 1L culture, the cells harvested by centrifugation were homogenized in 80 mL buffer A (20 mM Tris, 150 mM NaCl, 10 mM Imidazole, pH 7.5) and disrupted by ultrasonication (400 W, 3 s/3 s, 4°C). Cell debris was removed by centrifugation at 15,000 rpm for 30 min at 4°C. The supernatant collected was loaded onto Ni Sepharose^TM 6 Fast Flow resin (GE Healthcare), which was pre- equilibrated with buffer A. The resin was washed with buffer B (20 mM Tris, 150 mM NaCl, 20 mM Imidazole, pH 7.5), and the objective protein was eluted with buffer C (20 mM Tris, 150 mM NaCl, 200 mM Imidazole, pH 7.5) and buffer D (20 mM Tris, 150 mM NaCl, 500 mM Imidazole, pH 7.5), sequentially. Collections were verified by SDS-PAGE analysis. The target protein was dialyzed against buffer E (20 mM Tris, pH 7.5) at 4°C to remove the imidazole and salt, and then concentrated using a 10 kDa Centricon concentrator (Millipore). Concentrated protein was successively applied to Resource Q and Resource S 1 mL columns (GE Healthcare), and the protein was eluted from the column using buffer E with a gradient of NaCl from 0 M to 2 M. Finally, the protein was loaded onto a Superdex 75 10/300 GL column (GE Healthcare) in buffer F (10 mM Tris, 150 mM NaCl, 2 mM DTT, pH 7.5). All peak fractions were collected, and the protein purity was analyzed by SDS-PAGE. ## Mass spectrometry analysis The gel strip was removed from the SDS-PAGE gel, cut into small pieces, washed with 100 µL 25 mM ammonium bicarbonate (pH 8.0) containing 50% acetonitrile for 15 min twice with vortexing. Gel pieces were dehydrated with 100 µL acetonitrile and completely dried with a Speed-Vac before tryptic digestion. The volume of the dried gel was evaluated and three volumes of 12.5 ng/mL trypsin (Promega) in 25 mM NH₄HCO₃ (freshly diluted) were added. The digestion was performed at 30°C overnight, and then the mixture was sonicated for 10 min and centrifuged. The supernatant was removed for matrix-assisted laser desorption/ionization time-of flight mass spectrometry (MALDI-TOF MS) analysis. For MALDI-TOF MS analysis, 1 µL of the digested sample was spotted onto the MALDI target plate, and coated with 1 µL of matrix solution (5 mg/mL α-cyano-4-hydroxycinnamic-acid in 50% (v/v) acetonitrile and 0.1% (w/v) trifluoroacetic acid), then left to air-dry. Mass data were analysed with a prOTOFTM 2000 mass spectrometer interfaced with TOFWorksTM software (PerkinElmer/SCIEX). In this study, a 2-point external calibration of the prOTOF instrument was performed before acquiring the spectra from samples. Protein identification was performed by searching for bacteria in the NCBI non- redundant database using the Mascot search engine (Matrix Science), using the following parameters: monoisotopic; mass accuracy, 0.1 Da; missed cleavages, 1. ## Circular dichroism spectroscopy analysis During circular dichroism (CD) spectroscopy analysis, purified 6×His N-terminally tagged Rv0045c protein (0.35 mg/mL) was solubilized in 20 mM Tris (pH 7.5) and measured in the presence of room temperature with different pH (pH 2.0–pH 12.0) and pH 7.5 with different temperatures (10°C–70°C), respectively. UV CD spectra between 190 and 250 nm were collected on a JASCO 715 spectropolarimeter (JASCO) using 1 mm quartz cuvettes containing 200 µL of the protein solutions, with a data pitch of 0.1 nm, bandwidth of 2.0 nm and scanning speed of 50 nm/min. Every sample was measured in triplicate, and data were analyzed using the Jasco Jwsse 32 secondary structure estimation software. ## Enzyme activity analysis Enzyme activity of the Rv0045c protein was measured as previously described, using seven substrates: *p*-nitrophenyl acetate (C₂), butyrate (C₄), caprylate (C₆), caproate (C₈), myristate (C₁₂), laurate (C₁₄) and palmitate (C₁₆). The activities were determined by applying 10 mM *p*-nitrophenyl esters (C₂–C₁₆) as substrates at different pH (pH 6.0–pH 9.0) and different temperature (36°C–40°C). The substrate of *p*-nitrophenyl caprylate (C₆) was also used to estimate the dynamic activity of the enzyme at pH from 6.0 to 8.0 in the presence of mild temperatures (36°C–40°C)..For each standard assay, 50 µL 10 mM sodium taurocholate, 20 µL 10 mM substrate (dissolved in chloroform), and 420 µL Britton-Robinson buffer solution with different pH (pH 6.0–pH 9.0) were mixed in 1.5 mL Eppendorf tube separately, and then 10 uL protein (0.2 mg/mL) was added into each tube. After incubating at different temperatures for 15 min, the reaction was terminated by adding 700 µL 5∶2 (v/v) acetone/hexane solution. The mixture was then centrifuged at 4,600 g for 2.5 min at room temperature and the OD₄₀₅ of the lower phase was measured. Simultaneously, three controls were made: one prepared by adding the Rv0045c protein after adding acetone/hexane solution to observe instant hydrolysis; another prepared by substituting substrate solution with chloroform; and the other prepared by substituting the Rv0045c protein with 20 mM Tris (pH 7.5). Five parallel tests were repeated for every substrate at different pH and temperatures. [^1]: Conceived and designed the experiments: SL HP. Performed the experiments: JG XZ LX ZL KX SL TW HP. Analyzed the data: JG XZ HP. Contributed reagents/materials/analysis tools: HP. Wrote the paper: JG XZ HP. [^2]: The authors have declared that no competing interests exist.

# Introduction As of September 2020, infectious disease caused by the novel coronavirus (coronavirus disease 2019, or COVID-19) continues to spread on a global scale. This most recently discovered coronavirus emerged in the city of Wuhan in Hubei Province of China in December 2019. As of September 9, 2020, approximately 27 million people (confirmed cases) have been infected worldwide; the number of deaths has risen to approximately 900,000. The first case of infection in Japan was confirmed on January 16, 2020: a Chinese man residing in Kanagawa Prefecture who had lived for some time in Wuhan. Subsequently, the spread of infection among local residents who had never visited Wuhan and cluster infections occurred in Japan, which led to a rapid rise in the number of infected people, particularly in Tokyo and Osaka. On April 16, 2020, the Japanese government declared a national state of emergency. By September 9, 2020, the number of infected persons nationwide had risen to 72,726, leading to 1,393 deaths. Such a widespread social effect created confusion among the general public. More specifically, reactions such as bulk purchases (stockpiling) of face masks and sanitizers soared, creating secondary disruptions. Because fear of the COVID-19 epidemic can adversely affect disease management, fear of COVID-19 should be assessed appropriately. Some researchers have already conducted studies of people’s fear of infection with COVID-19 that is currently spreading worldwide. The Fear of COVID-19 Scale (FCV-19S) was developed with the aim of quelling fear of COVID-19 and for other goals. The FCV-19S is positively associated with anxiety and depression, and with perceptions of vulnerability to infection. The FCV-19S, based on the Protection Motivation Theory, has a unidimensional factor structure. The FCV-19S has been confirmed to have reliability and validity in various countries such as Bangladesh, Iran, Israel, Italy, New Zealand, Russia and Belarus, Saudi Arabia, Turkey, and Vietnam. It has come to have more widespread use than other corona-related measures. Results obtained using FCV-19S have been found to be associated with various factors including socio-demographic and residential environments. Being female, older, smoking, using health care services for COVID-19-related stress, and worries related to lockdown are factors associated with higher FCV-19S. In addition, fear and anxiety can affect social behavior. Among the psychological responses and coping behaviors of people following the 2009 influenza A (H1N1) outbreak, those with less resistance to uncertainty were more likely to use coping behaviors aimed at releasing their emotions. In the current context, fear of COVID-19 is not only positively associated with prevention behaviors and health-related behaviors such as increasing alcohol and tobacco use; it is also associated with bulk buying behaviors. However, these studies have not examined reasons underlying those behaviors. In Japan, where peer pressure is high, some people wear masks not only because it is necessary as a preventive behavior, but also because they are worried about how others will think of them if they do not wear a mask. Therefore, reasons of two types are assumed for actions: self-determining reasons, by which a person decides to act independently out of a sense of need, and conformity reasons, by which a person attunes self-behavior to the actions of surrounding people out of fear or anxiety. This study therefore examines a model of the relation between fear of COVID-19 and coping behavior with reasons of two types mediating fear and coping behavior. The purposes of this study were twofold. First, this study translates FCV-19S, established by Ahorsu et al., into Japanese and assesses its reliability and validity in Japan based on a procedure equivalent to that used by Ahorsu et al.. Secondly, this study develops a model of COVID-19 fear effects on coping behaviors, including reasons for the behaviors. # Materials and methods ## Procedure Data were collected in Japan. The participants were 450 residents of Japan (291 men and 159 women) aged 18 and older with average age of 48.13 years (SD = 14.04). They were recruited through an online service provided by Yahoo (<https://crowdsourcing.yahoo.co.jp/>), a major crowdsourcing service in Japan managed by Yahoo Japan Corp. The online service coordinates requests from clients with crowdsourcing workers. The current research was posted by the authors as a psychology research project in the "questionnaire" category on the Yahoo information board. The number of participants in the survey was set to be terminated when 450 participants had been recruited. The number of participants was determined based on the number of participants in another FCV-19S study to exceed the minimum sample size required by CFA and SEM. The participants read the study descriptions and agreed to take part in it by opting into the study themselves. Participants who completed the survey were assigned 105 points (about one dollar) to be used at a specific store. The survey questionnaire was administered on April 18 and 19, 2020. ## Survey description Socio-demographic variable: The survey questionnaire asked for free answers from each respondent for sex, age, nationality, and residential area (city and prefecture). This item was followed by a multiple-choice question asking about the participants’ health condition at the time, which allowed for many responses such as 1 = “in normal condition,” 2 = “having a fever of 37.5°C or higher,” 3 = “having a sore throat,” 4 = “having deep fatigue,” 5 = “having a cough,” and 6 = “having other symptoms”. The subsequent analysis combined answers that included at least one from choices 3 through 6 into one group, with three other groups including 1 = “in normal condition,” 2 = “having a fever of 37.5°C or higher,” and 3 = “having other symptoms.” The next question categorized the diseases being treated at the time into respiratory diseases, mental disorders (anxiety disorder, depression, and other mental disorders), and other diseases and asked the participants to indicate whether they had a disease, and if they did, to give the specific name(s) of the disease(s). The analysis combined the types of diseases into one group and divided the answers based on whether or not the participants had a disease being treated at the time. Subsequent questions asked all participants whether they smoked and asked female participants whether they were pregnant. All participants in this survey were Japanese citizens. For this study, the variables of sex, age, health condition at the time, diseases being treated, and smoking habits, which were expected to be associated with FCV in earlier studies, were used for the analysis. Working status: The questionnaire asked whether the participants worked at a job at the time. If they did, the questionnaire asked them to describe their job. Participants were also asked whether they were able to work from home, how they commuted, and at what hierarchical level in the company their job was positioned. After combining responses indicating whether the participants worked at a job and whether they were able to work from home, three categories were created for analyses, including a group being able to work from home, one unable to work from home, and a group of participants not working at a job. Family composition: Participants were asked about the number of family members living with them, their relation to the participant, their age, whether they had respiratory or other diseases, their smoking history, and their pregnancy status (only female family members residing with the participant). In addition, participants were asked about changes in the amount of conversations and conflict of opinions in the family living together in the last month. Only people living with family members who were expected to be associated with FCV in earlier studies were included in the analysis. Sources of information about COVID-19: Participants were asked a multiple-choice question about media that they regarded as a valuable source of information about COVID-19. They were also asked to rank such information sources from first to the third based on their importance. Specific sources of information indicated in the answers were the following: 1 = newspaper; 2, news on TV; 3, talk shows on television; 4, websites of public organizations; 5, news on the internet; 6, Twitter; 7, Facebook; 8, Instagram; 9, other social networking services (“SNS”); and 10, other media. The analysis used only the most prioritized information sources and combined choices 6–9 into a single group called “SNS.” Presence of persons infected with COVID-19 around the participant: Participants were asked whether anyone with whom they were acquainted had contracted COVID-19. Any respondent acquainted with an infected person was asked to describe their relation. To assess the status of COVID-19 infection for individuals around the participant, the survey asked about an infected person’s residence: 1, “in the same prefecture”; 2, “in the same municipality”; or 3, “in the same district” as the participant; or 4, no one was infected around the participant. The analysis combined responses for 1 through 3 into one group, designated as 1, “there is an infected person nearby” and the rest 2 “there is no infected person nearby.” Measuring fear of COVID-19: The study used a Japanese translation of FCV-19S developed by Ahorsu et al.. The following describes the procedure for translating the scale. First, permission to produce a Japanese translation of FCV-19S was obtained from Dr. Amir H. Pakpour, one author of the original article on FCV-19S. A translation agency performed the translation. The author of this report and the translation agency then modified Japanese expressions used for the items of the scale. A Korean psychologist fluent in Korean, English, and Japanese subsequently performed a reverse translation of the Japanese translation without seeing the original version. Finally, a native speaker of English compared this reverse translation and the original version of the scale and confirmed that they were fundamentally the same. The Japanese version was therefore finalized. The Japanese version of the scale that was completed consisted of seven items in the same manner as the original. These seven items were made a provisional Japanese version of the Fear of COVID-19 Scale (“FCV-19S-J”). FCV-19S-J asks questions to be answered on a scale of 1, “I am not afraid of COVID-19 at all” to 5, “I am most afraid of COVID-19.” A higher score reflects a greater fear of COVID-19. Measuring anxiety and depression: The Japanese version of the Hospital Anxiety and Depression Scale (HADS) prepared by Hatta et al. was used. Although the HADS was designed originally for patients in nonpsychiatric hospital clinics, it has been found to be reliable and valid in general samples. HADS was positively correlated with FCV-19S. It comprises a total of 14 items, including seven that measure recent anxiety and another seven that measure recent depression. It seeks responses from four choices each. Higher scores denote severer anxiety or depression experienced recently. Both the measures of anxiety and depression are expected to show positive correlation with FCV-19S-J. Measuring perception of vulnerability to infection: The study used the Japanese version of the perceived vulnerability to disease (PVD) scale developed by Fukukawa, Oda, Usami, and Kawahito. The PVD scale, which has been demonstrated as positively correlated with FCV-19S, consists of 15 items comprising 7 items of susceptibility to infection and 8 items of germ aversion. The response to each item is selected from seven choices from 1 = “strongly disagree” to 7 = “strongly agree.” A high score represents high susceptibility to infection or high germ aversion. Both the measures of susceptibility to infection and germ aversion are expected to represent positive correlation with FCV-19S-J. Behavior to cope with COVID-19: Participants were asked about their actions taken to cope with COVID-19. Items for coping behaviors were produced based on a report issued by the Ministry of Health, Labour and Welfare: "What we are doing to prevent new corona infection" of the "1st National Survey for Countermeasures against COVID-19". In addition, based on reports on social issues, items related to social behavior were added. Each of 19 items (e.g., “avoided places with large crowds”) was rated on a six-point Likert scale ranging from 1 (not at all) to 6 (very much). Reason for behavior: We asked about reasons for the behaviors above to cope with COVID-19. We developed three items as proactive reasons (e.g., “I did it because I felt it was necessary for myself”) and four items as passive reasons (e.g., “I did it because other people told me to”). Each of 7 items was rated on a six- point scale ranging from 1 (not applicable at all) to 6 (highly applicable). ## Ethical consideration After the first section of the survey presented a statement of the survey purpose, the form stated that participation was voluntary and that the survey was anonymous: personal information would not be disclosed to third parties. Only those who agreed to cooperate in the survey would be able to proceed to the questionnaire. Additionally, the Tohoku University Graduate School of Education’s ethics committee granted ethical approval for this study (ID: 20-1-003). ## Data analysis Statistical operations were conducted using software (Mplus 8.1) and a computer program (R 3.6.3). Analyses examined the reliability and validity of the Japanese version of the FCV-19S and the effect of fear of COVID-19 on coping behavior. To investigate the reliability and validity of the Japanese version of FCV-19S, confirmatory factor analysis, correlation analysis, and calculation of reliability coefficients were conducted. We also conducted t-tests and analysis of variance with the socio-demographic variable as the independent variable and FCV-19S as the dependent variable. For confirmatory factor analysis (CFA), a robust maximum likelihood estimator (MLR) was applied in this study. To test goodness of fit, we conducted the following analyses: comparative fit index (CFI), root mean square error of approximation (RMSEA), standardized root mean square residual (SRMR), and Bayesian information criterion (BIC). The cut-off values for acceptable model fit used for this study were: RMSEA \<.10 for acceptable fit and \<.06 for good fit; CFI \>.90 for acceptable fit and \>.95 for good fit; and SRMR \<.10 for acceptable fit and \<.08 for good fit. Error correlation was assumed to be related to the modified index (MI) if the goodness of fit showed an inadequate value. An MI indicates the expected parameter change if a particular specification were included in the model. Reliability was calculated for Cronbach’s alpha coefficients (α) and McDonald’s omega coefficients (ω). Correlations between the FCV-19S-J and other measures were established by calculating Pearson’s correlation coefficients. A t-test was performed when examining the difference in means between the two groups. Analysis of variance was conducted when examining the difference in means among the three or more groups. Reported effect sizes are interpreted using Cohen’s d and η², respectively including 95% confidence intervals. To elucidate the effect of FCV-19S on coping behavior, an exploratory factor analysis of coping behavior and reasons for behavior was conducted. A structural equation model was tested. In exploratory factor analyses (EFA), MLR and goemin rotation was applied for coping behavior and reason for behavior. We removed items with factor loadings lower than.35. In the structural equation model (SEM), a robust maximum likelihood estimator (MLR) was applied in this study. The same indices were used for testing goodness of fit as for confirmatory factor analysis. We assumed a direct path from FCV-19S to coping behavior and a path from FCV-19S to coping behavior by mediating the reason for behavior. Therefore, we examine not only the direct but also indirect effects of FCV-19S on coping behavior. All statistical analyses used two-tailed tests. For all statistical evaluations, p values less than.05 were inferred as significant. Missing values were visible only for age. Therefore, pairwise deletion was used for missing data. # Results ## Factor structure, reliability and validity of FCV-19S-J presents participants’ basic characteristics. The present study participants were 450 participants, mostly men (65%). Confirmatory factor analyses, as described by Ahorsu et al., were used to examine the goodness of fit. Results show that the Japanese FCV-19S did not fit well (Model 1). To improve the model fit, MI were used. The MI between items 1 and 4 (MI = 47.72) and between items 2 and 5 (MI = 35.30) were higher values. Therefore, a within factor error-covariance between items 1 and 4 (Model 2) and between items 2 and 5 (Model 3) was included. The model was modified. The model that includes error correlations of items 1 and 4 / items 2 and 5 is Model 4. Results indicate that the modified model (Model 4) was more acceptable (CFI =.943, RMSEA =.105, SRMR =.052) than Model 2 or 3. This was the final model. The descriptive statistics of the Japanese FCV-19S and other scales are presented in. Internal consistency reliability was acceptable for FCV-19S scores (α =.87/ ω =.92). Correlations between the Japanese FCV-19S and the HADS and PVD were significant, ranging from r =.29 to r =.56, p \<.01. ## Differences based on relevant variables Means and standard deviations of all compared groups are presented in. No significant difference was found among the variables other than the important source of information. Groups of the important source of information differed significantly with regard to the FCV-19S-J, F(6, 443) = 3.469, p\<.01, η² =.044, 95% CI \[.006;.076\]. Post-hoc tests revealed that the News on TV group scored higher on the FCV-19S-J than the Other groups, p =.04, d =.97, 95% CI \[.38; 1.56\]. ## Effects of fear of COVID-19 on coping behavior Factor analysis was conducted of coping behaviors and the reasons for these behaviors. Results show that 13 items were extracted from three factors for coping behavior; six items from two factors were extracted for reasons for the behavior. The descriptive statistics of these scales are presented in. Structural equation modeling (SEM) was used to investigate the effect of fear of COVID-19 on coping behavior. The model had acceptable fit to the data, X²(42) = 179.934, p\<.05, CFI =.905, RMSEA =.085, (90% CI.073.098), SRMR =.057. The path from FCV-19S-J to each coping behavior and conformity reason had a small and moderate effect (β =.206.358, p\<.001). The path from conformity reason to stockpiling and from self-determining reason to daily attention had a small and moderate effect (β =.113 –.119, p\<.05). An indirect effect was examined to ascertain whether the conformity reason significantly mediated the relation between FCV-19S-J and stockpiling. Results show that FCV-19S-J had a significant indirect effect on stockpiling, through the effect of conformity reason (β =.035, p\<.05) for a modest total effect (β =.311, *p*\<.01). # Discussion The first purpose of this study was development of the Japanese version of FCV-19S. As described by Ahorsu et al., the results of factor analysis indicated a single factor structure. The α coefficient and omega coefficient in FCV-19S-J returned sufficient values. HADS indicated a significant correlation between “depression” and “anxiety”; PVD revealed significant correlation between “perceived infectability” and “germ aversion,” which suggests adequate reliability and validity. In FCV-19S-J, the goodness of fit was made an acceptable value by assuming an error correlation between item 1 (“I am most afraid of coronavirus-19.”) and item 4 (“I am afraid of losing my life because of coronavirus-19.”) and between item 2 (“It makes me uncomfortable to think about coronavirus-19.”) and item 5 (“When watching news and stories about coronavirus-19 on social media, I become nervous or anxious.”). For the error correlations, the Saudi Arabian version was on items 1 and 2, items 3 and 6, items 3 and 7, and items 6 and 7, whereas the Turkish version was on items 3 and 6, items 3 and 7, and items 6 and 7. We found no study with reported error correlations between items 1 and 4, items 2 and 5 as in the results of this study. FCV-19S has a unidimensional structure. However items 3, 6, and 7 are regarded as somatic responses to COVID-19 fear; items 1, 2, 4, and 5 are regarded as representing the general level of fear. Therefore, the error correlations in this study are error correlations between items that indicate a general level of fear. In Japan, a state of emergency declaration was expanded to include all of Japan on April 16. This event was a couple of days before the participants in this study participated in the study. This social context might have influenced the association between items of general fear rather than somatic responses. Results of t-tests and analysis of variance suggested that participants who prioritized news on television as an information source tended to have greater anxiety related to COVID-19 than others who prioritized other information sources. The FCV-19S scores, however, were not found to be significantly different based on other factors such as the sex, age, living with family, and presence of persons infected. Earlier studies indicated that risk factors which increase the fear of COVID-19 include being female, older, smoking, using health care services for COVID-19-related stress, worries related to lockdown, and not living with a family member. Differences between the results of this study and earlier studies are likely to be attributable to social conditions. Because a state of emergency was declared in Japan, it is likely that everyone living in Japan is almost equally fearful, irrespective of the demographic characteristics of the study participants. Therefore, we believe that these study participants did not differ significantly in terms of their fear of COVID-19 depending on their attributes. The second purpose of this study was development of a model of COVID-19 fear effects on coping behaviors, including reasons for the behaviors. Results revealed that fear of COVID-19 encouraged measures taken to prevent infection such as care in daily life and health conditions. This result is consistent with results indicating that FCV-19S predicted positive behavioral change (e.g. hand washing, changed travel) and that it was positively associated with adherence to New Zealand's lockdown rules (e.g. maintaining the two-meter rule when out in public). Particularly, Harper examined behavioral change. Therefore, various behaviors such as hand-washing and stockpiling were combined into a single variable. However, this study not only categorized coping behaviors as being careful in daily life, stockpiling, and health monitoring; it also asked about reasons for the behaviors. Results of this study demonstrated that fear of COVID is associated with preventive behavior, and also that it is related directly to nuisance behaviors such as stockpiling. Furthermore, we have demonstrated that fear of COVID is associated with stockpiling via conformity. The fear of COVID-19 did not affect self-determination of reasons. However, self-determination of reasons contributed to an increase in care in daily life. At least in Japan, the self-determination of reasons and fear of COVID-19 might be useful for encouraging caution in daily life. ## Limitations This study has some limitations. Study participants were solicited using the internet. The study was able to collect people in wide-ranging age groups, but those who were willing to participate in the survey were likely to have been mentally stable, with sufficient psychological capacity to contemplate COVID-19 effects. The survey must be expanded to include participants such as medical professionals and people who have been adversely affected financially through events such as a loss of work because of COVID-19. Moreover, the analysis used for this study used cross-sectional data, which were inadequate to verify causal relations between anxiety about COVID-19 and coping behavior. Particularly, the cross-sectional survey does not enable us to ascertain whether anxiety and fear arouses preventive behavior or whether it is aroused by performance of preventive behavior. Subsequent studies must identify relations between coping behavior and a fear of COVID-19 using a longitudinal study. # Conclusions Despite the limitations described above, this study has explained the factor structure of FCV-19S-J. This report is the first in Japan to describe a study identifying the relations between fear of COVID-19 and coping behavior. The environment surrounding COVID-19 changes day by day. Appropriately measuring people’s anxiety and fear of COVID-19 likely to contribute to an understanding of increasing anxiety experienced by many people and to prevention of difficulties associated with COVID-19. # Supporting information The authors would like to thank the members of our study team and the participants who took part in our study. 10.1371/journal.pone.0241958.r001 Decision Letter 0 Montazeri Ali Academic Editor 2020 Ali Montazeri This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 24 Aug 2020 PONE-D-20-15563 Variables related to fear of novel coronavirus infection (COVID-19) and coping behavior PLOS ONE Dear Dr. Wakashima, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Oct 08 2020 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols> We look forward to receiving your revised manuscript. Kind regards, Ali Montazeri Academic Editor PLOS ONE Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main _body.pdf> and <https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf> 2\. Since one of the key outcomes of your study was the development of a translated version of the FCV-19S, you may wish to include this in the title of your manuscript. 3\. PLOS ONE has specific requirements for studies that are presenting a new method or tool as the primary focus, including a newly developed or modified questionnaire or scale (<https://journals.plos.org/plosone/s/submission- guidelines#loc-methods-software-databases-and-tools>.) One requirement is that the questionnaire or scale must be openly available under a license no more restrictive than CC BY. In light of this, before we proceed, please include a copy of your questionnaire or scale as a Supporting Information file (in the original language) or provide a link if it is available through an online repository. 4\. Thank you for stating the following financial disclosure: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." At this time, please address the following queries: Please clarify the sources of funding (financial or material support) for your study. List the grants or organizations that supported your study, including funding received from your institution.State what role the funders took in the study. If the funders had no role in your study, please state: “The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.”If any authors received a salary from any of your funders, please state which authors and which funders.If you did not receive any funding for this study, please state: “The authors received no specific funding for this work.” Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 5\. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager. Please see the following video for instructions on linking an ORCID iD to your Editorial Manager account: <https://www.youtube.com/watch?v=_xcclfuvtxQ> 6\. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: [http://journals.plos.org/plosone/s/supporting-information](about:blank). \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: N/A \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: This is an interesting and timely study. I have some concerns that will help to improve the manuscript. 1- I would add Japanese or japan in the title 2- Abstract: sampling procedure is missing. Real results and statistics are not reported the abstract. I cannot understand what you really mean on “were relatively comparable to those of the original FCV-19S”? how did you assess coping behavior? Please clearly report mediation results rather than writing a vague interpretation. 3- Introduction is good but needs to improving by adding some information of the latest statistics on covid-19 in japan. Please also add the following references: Pakpour, A. H., Griffiths, M. D., Chang, K. C., Chen, Y. P., Kuo, Y. J., & Lin, C. Y. (2020). Assessing the fear of COVID-19 among different populations: A response to Ransing et al.(2020). Brain, Behavior, and Immunity. Pakpour, A. H., & Griffiths, M. D. (2020). The fear of COVID-19 and its role in preventive behaviors. Journal of Concurrent Disorders. Lin, C. Y. (2020). Social reaction toward the 2019 novel coronavirus (COVID-19). Social Health and Behavior, 3(1), 1. 4- I think that the authors need to compare their results with previous validation version. Please see the following: Nguyen, H.T.; Do, B.N.; Pham, K.M.; Kim, G.B.; Dam, H.T.; Nguyen, T.T.; Nguyen, T.T.; Nguyen, Y.H.; Sørensen, K.; Pleasant, A.; Duong, T.V. Fear of COVID-19 Scale—Associations of Its Scores with Health Literacy and Health-Related Behaviors among Medical Students. Int. J. Environ. Res. Public Health 2020, 17, 4164. Soraci, P., Ferrari, A., Abbiati, F. A., Del Fante, E., De Pace, R., Urso, A., & Griffiths, M. D. (2020). Validation and psychometric evaluation of the Italian version of the Fear of COVID-19 Scale. International Journal of Mental Health and Addiction, 1-10. Haktanir, A., Seki, T., & Dilmaç, B. (2020). Adaptation and evaluation of Turkish version of the fear of COVID-19 scale. Death Studies, 1-9. Harper, C. A., Satchell, L. P., Fido, D., & Latzman, R. D. (2020). Functional fear predicts public health compliance in the COVID-19 pandemic. International journal of mental health and addiction. Reznik, A., Gritsenko, V., Konstantinov, V., Khamenka, N., & Isralowitz, R. (2020). COVID-19 fear in Eastern Europe: Validation of the Fear of COVID-19 Scale. International journal of mental health and addiction, 1. Alyami, Mohsen, Marcus Henning, Christian U. Krägeloh, and Hussain Alyami. "Psychometric evaluation of the Arabic version of the Fear of COVID-19 Scale." International journal of mental health and addiction (2020): 1. Bitan, D. T., Grossman-Giron, A., Bloch, Y., Mayer, Y., Shiffman, N., & Mendlovic, S. (2020). Fear of COVID-19 scale: Psychometric characteristics, reliability and validity in the Israeli population. Psychiatry Research, 113100. Winter, Taylor, Benjamin Riordan, Amir Pakpour, Mark Griffiths, Andre Mason, John Poulgrain, and Scarf Damian. "Evaluation of the English version of the Fear of COVID-19 Scale and its relationship with behavior change and political beliefs." (2020). Sakib, N., Bhuiyan, A. I., Hossain, S., Al Mamun, F., Hosen, I., Abdullah, A. H.... & Sikder, M. T. (2020). Psychometric validation of the Bangla Fear of COVID-19 Scale: Confirmatory factor analysis and Rasch analysis. International Journal of Mental Health and Addiction. Reviewer \#2: The introduction is not convincing. It fails to show the current literature gap. The aim of the study is not clear. The study should clearly specify the mentioned issues; whether you are validating FCV-19S or developing a model of COVID-19 fear effects…? Method: how did you invite the participants to the study? How many were approached? What were the main reasons for non-participation? The rational for study sample size should be stated. The rationale behind selecting the variables included in the questionnaire should be stated. The statistical analysis should be clearly explained. The conceptual model that guided the authors for analysis should be explained in details. Table 1. please indicate which test was used for analysis. Table 2: writing error, heart races! Why did you select model 4 as the final model? Did you change any item in the scale? What modification/s was/were made in to the scale? Why did you use HAD for structural validity among a community population? What do you mean by (α=.93/ ω=.93) in table 3? These should be explained. The main concern is that, could we use the same sample for both validation study and SEM? What are the differences between table 4 and 5? The discussion is highly poor. In this part you should compare the results obtained from your study with similar studies either confirming or rejecting the current results. Then, express your explanations or justifications. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Amir Pakpour Reviewer \#2: **Yes: **Marzieh Aaraban \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0241958.r002 Author response to Decision Letter 0 1 Oct 2020 Reviewer \#1 : Comment 1 : Title I would add Japanese or japan in the title. Response: We agree with your opinion. We have changed the title from “Variables related to fear of novel coronavirus infection (COVID-19) and coping behavior” to “The Japanese version of the Fear of COVID-19 scale: Reliability, validity, and relation to coping behavior”. Comment 2 : Abstract Sampling procedure is missing. Real results and statistics are not reported the abstract. I cannot understand what you really mean on “were relatively comparable to those of the original FCV-19S”? how did you assess coping behavior? Please clearly report mediation results rather than writing a vague interpretation. Response: We agree with the reviewer’s point. We have made the following modifications. \(1\) We have added “450 Japanese participants were recruited from a crowdsourcing platform” to the Abstract as the sampling procedure (p.2 lines 21-22). \(2\) We have added real results and statistics to the Abstract (p.2 lines 28–33). \(3\) We have changed the sentence from “were relatively comparable to those of the original FCV-19S” to “These results suggest that the Japanese FCV-19S is a psychometric scale with the same reliability and validity as the original FCV-19S.” Comment 3 : Introduction Introduction is good but needs to improving by adding some information of the latest statistics on covid-19 in japan. Please also add the following references: Response: To page 3 of the revised manuscript, we have included the latest statistics on COVID-19 from the entire world and from Japan. We have added references to lines 53–54 (Lin, 2020), lines 56–57 (Pakpour and Griffiths, 2020), and lines 62–63 (Pakpour et al., 2020). Comment 4 : I think that the authors need to compare their results with previous validation version. Response: Thank you for introducing us to the many references. We have cited nine references to support the validation lines 60–73. These are the following: Sakib et al. (2020), Bitan et al. (2020), Soraci et al. (2020), Winter et al. (2020), Reznik et al. (2020), Alyami et al. (2020), Haktanir et al. (2020), Nguyen et al. (2020), and Harper et al. (2020). Reviewer \#2 : Comment 1 : Introduction The introduction is not convincing. It fails to show the current literature gap. The aim of the study is not clear. The study should clearly specify the mentioned issues; whether you are validating FCV-19S or developing a model of COVID-19 fear effects…? Response: The reviewer has commented on the lack of persuasiveness of the Introduction and the lack of clarity of purpose. We agree with the reviewer’s points. We have particularly cited the references introduced by Reviewer 1 (lines 60–73) and have described the purpose as follows: The purpose of this study was twofold. First, this study translates FCV-19S established by Ahorsu et al. \[10\] into Japanese and assesses reliability and validity in Japan based on a procedure equivalent to that used by Ahorsu et al. \[10\]. Secondly, this study develops a model of the effects of COVID-19 fear on coping behaviors, including the reasons for the behavior (p. 5 lines 82–85). Comment 2 : Method How did you invite the participants to the study? How many were approached? What were the main reasons for non-participation? The rational for study sample size should be stated. Response: The reviewer has commented on the lack of clarity related to data collection in the Materials and Methods section. We have specifically added information related to how we recruit participants and the reasonableness of our sample size (Page 5, lines 89–101). Because we used crowdsourcing and decided to end the study when the number of participants reached 450, we cannot state the reasons for non-participants. Comment 3 : The rationale behind selecting the variables included in the questionnaire should be stated. Response: We agree with the reviewer’s point. We have stated the rationale for our choice of variables by citing earlier studies (Page 6, lines 118–120; Page 7, lines 132–134). Comment 4 : The statistical analysis should be clearly explained. The conceptual model that guided the authors for analysis should be explained in details. Response: Thank you for your comment. We realized that our original explanation was unclear and revised as follows: \(1\) The Data analysis section described all the analyses used for this study (lines 201–230). \(2\) Regarding the conceptual model, we discussed the need to measure the reasons for behavior in the Introduction section and the relation between FCV-19S, coping behavior, and reasons for behavior, as assumed in this study (lines 74–81). In addition, the conceptual model was explained in the Data analysis section (lines 222–230). Comment 5 : Table 1. please indicate which test was used for analysis. Response: We have provided a brief description of the analysis of the Table 1 footnote of revised manuscript. Note: \*\* p\<0.01, sex, smoking habit, diseases being treated, living with family and presence of persons infected were subjected to t-tests, and analysis of variance was performed for age, health condition, work status, and sources of information. Comment 6 : Table 2: writing error, heart races! Response: Thank you for pointing out this error. An earlier manuscript did not present all sentences in Item 7. We have therefore added all sentences to item 7. The original FCV-19S (Ahorsu et al., 2020) also uses the word "heart races". Comment 7 : Why did you select model 4 as the final model? Did you change any item in the scale? What modification/s was/were made in to the scale? Response: The reviewer has commented on the lack of clarity related to the final model decision. We decided to use Model 4 as the final model because it was an acceptable fit over the other two improved models (Model 2 and Model 3) (lines 242–243). We have added a description of MI to the Data analysis section to show that we have assumed error correlation related to MI (lines 213–215). Some other studies of FCV-19S have also assumed error correlations (Alyami et al., 2020; Haktanir et al., 2020) (lines 335–337). We did not change the items in the scale. Comment 8 : Why did you use HAD for structural validity among a community population? Response: HAD has been used in the validation of FCV-19S and has been found to be reliable and valid in studies of community samples. We have added the following explanation to lines 166–168 of the revised manuscript. Although the HADS \[35\] was designed originally for patients in nonpsychiatric hospital clinics, it has been found to be reliable and valid in general samples \[34\]. HADS was positively correlated with FCV-19S \[10\]. Comment 9 : What do you mean by (α=.93/ ω=.93) in table 3? These should be explained. Response: We have provided a brief description of � and � in the Table 2 and Table 3 footnotes. We have also given the use of � and � as indicators of reliability in Data analysis section (lines 215–216). Comment 10 : The main concern is that, could we use the same sample for both validation study and SEM? Response: The reviewer has asked about the propriety of using the same sample for validation and SEM. Many other studies have applied factor analysis and other analyses (including SEM) to the same samples. Therefore we believe there is no particular problem with their application here. Comment 11 : What are the differences between table 4 and 5? Response: Table 4 presents results of exploratory factor analysis for “coping behavior.” Table 5 shows the results of exploratory factor analysis for “reason for behavior.” For “reason for behavior,” we changed the factor names to clarify it that it is a reason (Previous manuscript: self-determination and conformity behavior. Revised manuscript: self-determining reason and conformity reason). Comment 12 : The discussion is highly poor. In this part, you should compare the results obtained from your study with similar studies either confirming or rejecting the current results. Then, express your explanations or justifications. Response: The reviewer is concerned about the lack of sufficient discussion. The reviewer is correct. We appreciate the chance to clarify our exposition. We have revised the paper as explained below. \(1\) Results of the confirmatory factor analysis are discussed in comparison to results obtained for other countries (lines 335–346). \(2\) The relation between sociodemographic variables and FCV-19S was discussed in comparison to the results of earlier studies (lines 349–358). \(3\) Effects of FCV-19S on coping behaviors were discussed in comparison to Winter et al., 2020 and to Harper et al., 2020 (lines 359–375). Other points to change ERROR CORRECTED p.6 line 103: “Socio-demographic variable” (previous manuscript: Attributes of the respondents) p.6 line 108: “6 = having other symptoms” (previous manuscript: 9 = having other symptoms) p.11 line 236: “FCV-19S-J” (previous manuscript: Japanese FCV-19S-J) p.23 line 389: “longitudinal study” (previous manuscript: cross-sectional surveys) 10.1371/journal.pone.0241958.r003 Decision Letter 1 Montazeri Ali Academic Editor 2020 Ali Montazeri This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 26 Oct 2020 The Japanese version of the Fear of COVID-19 scale: Reliability, validity, and relation to coping behavior PONE-D-20-15563R1 Dear Dr. Wakashima, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. Kind regards, Ali Montazeri Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The authors have addressed my comments. The paper can be published. Thank you so much for sending the paper for review in PLOS One. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No 10.1371/journal.pone.0241958.r004 Acceptance letter Montazeri Ali Academic Editor 2020 Ali Montazeri This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 28 Oct 2020 PONE-D-20-15563R1 The Japanese version of the Fear of COVID-19 scale: Reliability, validity, and relation to coping behavior Dear Dr. Wakashima: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. If we can help with anything else, please email us at <plosone@plos.org>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Ali Montazeri Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Network-based modeling and characterization of brain architectures has provided both a framework for integrating imaging data as well as for understanding the function and dynamics of the brain. Brain networks are traditionally constructed either from structural or functional imaging data. Functional brain networks represent the associations between regions estimated by statistical similarities in regional time series, as measured by correlation or coherence. In the case of fMRI data, regional gray matter activity is measured by the *blood oxygenation level dependent (BOLD)* signal. Brain networks are commonly studied using techniques drawn from graph theory and machine learning. These techniques provide fundamental and generalizable mathematical representations of complex neuroimaging data: nodes represent brain regions and edges represent structural or functional connectivity. This simplified graphical representation enables the principled examination of patterns of brain connectivity across cognitive and disease states. While the majority of network-based studies have focused on the brain’s resting state, more recent efforts have turned to understanding brain connectivity elicited by task demands, including visual processing and learning. Global network analysis of both functional and structural connectivity has demonstrated that brain networks have characteristic topological properties, including dense modular structures and efficient long-distance paths. Traditional network analysis tools are not necessarily sensitive to small perturbations in functional or structural connectivity as they rely on network- wide statistics. Recent efforts have focused on developing new algorithms to identify specific subgraphs that are discriminative between brain states (cognitive or disease) and therefore critical for an understanding of local neurophysiological processes. Zalesky and colleagues describe a set of methods to identify groups of edges that are significantly different between two groups of networks. Kim et al. recently proposed powerful statistical tests for identifying functional edges that differ between groups. Motifs, defined patterns of local connectivity that occur frequently across sessions and subjects, are groups of edges with particular topological properties that may play specific functional roles. Hyperedges, considered by Bassett and colleagues, were defined as groups of edges that vary significantly in weight over time, for example during adaptive functions like learning or during higher- order cognitive processes like memory and attention. In general, all of these tools seek to associate local network features (or subgraphs) with cognitive function, offering fundamental understanding and the opportunity to inform therapeutic interventions. Here, we take an approach that is complementary to previous approaches focused on individual brain regions or single connections. Our focus is on the network architecture of the brain, and how that architecture relates to behavior. Specifically, in this study, we develop and apply a novel analysis framework for identifying subgraphs that discriminate between individuals with differing behavioral variables. Our approach is: (i) data-driven in that it does not make assumptions about network sparsity, edge independence, etc.; (ii) enforces subnetwork connectivity in both the discovery and statistical scoring of discriminative subnetworks; and (iii) draws on novel machine learning methods for network state classification. Particularly, we extend recent techniques from labeled network mining, whose goal is to classify network instances based on individual node/edge states. In our setting, the goal is uncover a network- specific type of *biomarker*—a connected subgraph of functional edges—whose coherence predicts whether individuals are learning a motor sequencing task at a *high* or *low* rate. We learn a low-dimensional subspace of connected brain regions that discriminates among the categories representing high and low learning rates. To ensure generalization in the presence of few training instances, we build multiple models by performing k-fold validation and repeating the fold partitioning multiple times. We then mine conserved discriminative subnetworks across runs using frequent subgraph mining and employ randomized statistical tests to establish the significance (in terms of q-value) of the discovered subgraphs while implementing a *false discovery rate* correction for multiple comparisons. We employ our framework to uncover the subgraphs that maximize the discriminative potential in explaining the differences in the rate of motor learning between individuals. The data for our analysis comes from a motor learning task in which subjects’ neural activity was measured using fMRI in repeated sessions as they learned a set of 12-note finger sequences. Subjects learned three different sequences, each of which was presented as a string of spatial notes on a 4-line tablature. Each line corresponded to one of four fingers. The performance measure was movement time, which is the time required to perform a given 12-note sequence. We assign individual sessions to two categories—*high* and *low* rate learner sessions based on the measured slopes of learning rate. Intuitively, if a subject’s time in a given session decreases significantly, we classify the subject in this session as high-rate learner. The fMRI data was aligned to the Harvard-Oxford Brain Atlas (part of the FSL tool) involving 112 cortical and subcortical regions. A functional edge strength linking two cortical areas was estimated as the wavelet-based coherence of the corresponding regional time series. Our work is the first to propose a general methodology for identifying network- based biomarkers for learning rate based on connected subgraph mining. Amongst tens of thousands of possible edges between 112 brain regions, we find connected functional subgraphs comprised of 1-5 edges that predict whether a subject is a high or low rate learner. These learning biomarkers agree with observations from previous studies and further suggest new brain region relations which are essential to learning. Our proposed methodology is general in that it can be applied for studying the differences in other kinds of cognitive states or functional connectivity differences between disease and controls. # Materials and methods All human participants provided written informed consent after the study was approved by the University of California Santa Barbara Institutional Review Board. Our goal is to detect a set of functional edges interconnecting cortical regions whose coherence can predict differences among subjects in fMRI cognitive or disease-related studies. In the context of our data, the goal is to predict individual learning rates. We expect that learning-related changes in functional connectivity will be located in coordinated neural circuits involving co- activated regions forming a connected component, and we therefore restrict our attention to predictive edges that form a connected subgraph. As the rate of learning increases, some functional edges within a subgraph of interest will fall into a low coherence state (i.e., coherence between their adjacent regions’ activation will approach 0), while others will move into a high coherence state. We seek to understand these dynamics and extract structures (in the form of functional subgraphs) that predict the global behavioral state of the individual: high/low rates of learning that are statistically significant. *Definition:* A subnetwork *biomarker* is a statistically significant connected subgraph of functional edges whose coherence states can collectively *differentiate* between cognitive scores (e.g., learning rate) of subjects. The potential of a biomarker to differentiate between functional networks corresponding to low or high learning rate subject sessions is called *discriminative power* (and the biomarker is called *discriminative*). An overview of our approach is presented in. We employ a discriminative biomarker mining approach to analyze functional networks constructed from fMRI scans of 18 subjects performing a motor learning task over 3 learning sessions that occurred on 3 different days. The sessions are divided into low and high learning rate sessions based on the average reduction in movement time to complete the motor task. The goal of our analysis is to identify biomarkers (subgraphs) that are *discriminative* of the session type (high *vs.* low rate learners), while at the same time statistically significant. We describe the steps in our approach in more detail in what follows. ## Data acquisition and preparation The data for our analysis was collected during a motor learning experiment in which subjects’ neural activity was measured using fMRI. The data was originally used to analyze the brain’s functional flexibility during learning. Here, we follow the same protocol for data preparation, but focus on subgraph biomarkers associated with learning. Next, we provide a description of the original experimental procedure, apparatus, imaging protocol and data processing. **Experimental procedure.** While in supine position, subjects performed a cued sequence production (CSP) task using the four fingers of their left hand, thumb excluded. To maximize comfort and to provide an angled surface to position the response box, subjects were positioned inside the scanner with foam padding under their knees. Padding was also placed under the left arm to provide extra support when responding during the task. Subjects performed the CSP task by responding to visually cued sequences on the response box using their left hand. The sequences were presented in static form, as a series of 12 music notes on a 4-line music staff. Subjects were instructed to type the sequences, reading from left to right, so that the top line of the staff mapped to the leftmost finger and the bottom line mapped the rightmost finger. Each 12-element sequence contained 3 notes per line. Each trial began with a fixation ‘+’, displayed for 2 seconds. The complete 12-element sequence was presented immediately following the offset of the fixation ‘+’, and participants were instructed to initiate responding quickly and accurately. The sequence remained on screen for the responding duration, or a time limit of 8 seconds, whichever came first. After completion of a correct sequence, the notes were replaced with a fixation signal until the trial duration was reached. With any incorrect press, a verbal cue “INCORRECT” appeared and the participant waited for the next trial. Subjects trained on 16 different sequences, with training divided into 3 levels of intensity. Of these, 3 sequences were trained frequently (189 trials/sequence), with training distributed for these “frequent” sequences evenly across the training sessions. In addition, a second set of three sequences were presented moderately for 30 trials, and a third set of ten additional sequences, were presented rarely between 4 and 8 trials during training. Frequent sequences were practiced in blocks of 10 trials, with 9 out of 10 trials in a block belonging to the same frequent sequence, and the other trial belonging to one of the ten rarely trained sequences. Trials were presented using an event-related structure, with sequence trials separated using an interstimulus interval that ranged between 0 and 20 seconds, along with additional time remaining following the completion of the previous trial. To provide some motivation for good performance, after each block of trials, subjects received feedback that detailed the number of correct trials and the average movement time (defined as the time to complete a sequence) needed to complete a correct sequence in that block. Scan epochs lasted 40 trials (4 blocks, 345 scan TRs), and each training session contained 6 scan epochs (2070 scan TRs). **Apparatus.** Task presentation and online behavioral data acquisition was handled using a Dell Latitude D620 laptop computer running MATLAB 7.1 (Mathworks, Natick, MA) and the Cogent 2000 toolbox. Key-press responses and response times were collected using a custom response box with fiber optic signal transduction connected to a response card (DAQCard-6024e, National Instruments, Austin, TX). **Imaging protocol.** Functional MR images were collected using a 3T Siemens Trio with a 12-channel phased-array head coil. For each scan epoch, a single- shot echo planar imaging sequence sensitive to BOLD contrast was used to acquire 33 slices per scan (repetition time \[TR\]/echo time\[TE\] 2000/30 ms, 3 mm axial slice, 0.5 mm gap, flip angle of 90°, field of view 192 mm, 64 × 64 inplane acquisition matrix). Prior to the initial functional scan epoch, a high- resolution T1-weighted scan (TR/echo time 15.0/4.2 ms, flip angle of 9°, 3D acquisition, 0.89 axial slice, field of view 256 mm, 256 × 256 inplane acquisition matrix) was collected. Image preprocessing was performed using the FMRIB (Oxford Centre for Functional Magnetic Resonance Imaging of the Brain) included in the Software Library (FSL). Functional imaging time series realignment was performed using the fully automated program MCFLIRT (Motion Correction using FMRIB’s Linear Image Registration Tool) with realignment to the middle time series image. Images were high-pass filtered (50*s* cutoff), and spatially smoothed 8*mm* with a Gaussian kernel. No temporal smoothing was used. Further, signal intensity was normalized across all functional volumes in order to control for possible fluctuations in across sessions. Functional volumes were normalized to the Montreal Neurological Institute (MNI)-152 template with affine transformation (12 DOF) using FLIRT (FMRIB’s Linear Image Registration Tool). We parcellated the brain into 112 cortical and subcortical regions using the Harvard-Oxford structural atlas as supplied with FSL in standard space (MNI-152). For each individual participant and for each of the 112 regions, we calculated the regional activation level time series by finding the mean across all voxels in a region. **Subjects.** The study involved 18 paid young adult participants without formal training in playing a musical instrument, with normal vision, and without neurological or psychiatric disorders. The number and order of sequence trials was identical for all participants. All participants completed three training sessions in a five-day period. **Functional networks and learning rate labels.** Edges between nodes represented the pairwise coherence of the average fMRI time series for a pair of brain regions. More specifically, we estimated a magnitude squared wavelet coherence, which identified areas in time-frequency space where two time series co-varied in the frequency band 0.06-0.12 Hz (we used fixed binning of this interval). This measure of functional connectivity was estimated using the minimum-variance distortionless response method, and provides a measure of nonlinear functional association between any two time series. In using the coherence, which has been demonstrated to be useful in the context of fMRI neuroimaging data, we were able to measure frequency-specific linear relationships between time series. The coherence matrix of every session corresponds to a fully connected graph (a *clique*) involving all brain regions as nodes and coherence values associated with edges. Apart from the coherence values, each session is also characterized by movement time—the time to complete the sequence. Motor learning is well- characterized by an exponential drop-off in movement time; early learning shows a fast rate of drop-off, and is well-fit by one exponential curve, and later learning shows a slower rate of drop-off, and is well-fit by a second exponential curve. Here we study early learning, taking place over 3 days of moderate practice, and therefore examine a single exponential fit of the movement time versus trial bin. The magnitude of the exponential drop-off parameter indicates the gradient of the learning slope, where a sharper drop-off in movement time corresponds to individuals who are faster learners in the session, and a less-sharp drop-off in movement time corresponds to individuals who are slower learners respectively. We derive a learning rate class label (high and low) for each session-specific network based on the drop-off in learning slope. To estimate the threshold between the two classes, we cluster the drop-off values in two groups multiple times and adopt the consistent pivot between clusters over multiple runs as a threshold. Details on the threshold estimation are provided in the Supporting Information. It is worth mentioning that multiple levels of learning rate can also be accommodated within our framework, requiring only minor modifications, however we restrict the analysis to two classes due to the limited number of instances. Further discussion of multiple-class analysis is available in the following section. ## Mining discriminative subgraphs As a result of the data preparation we obtain a set of *global-state networks*: graphs with local coherence values on edges and a global network state indicating whether high or low learning rate was observed in the corresponding session. The setting is similar to that in a classification task in machine learning in that we have a set of instances (functional coherence networks) characterized by features (coherence values) and associated with labels (learning rates). The distinctive aspect in global network state classification is that there is an inherent structure imposed on the features: the shared network topology. The aim in global state network classification and feature selection is to find small connected subgraphs involving the most discriminative nodes, whose labels predict the global state of network instances. In our case, these subgraphs will correspond to interconnected brain regions, whose patterns of pairwise coherence discriminates between high and low learning rates. We employ our recently proposed method for global state classification to mine candidate substructures. Prior to describing our method, let us briefly discuss existing techniques that deal with global network state classifications. They typically work with node labels as opposed to edge labels. Hence, in order to employ those techniques, we represent our data in this common framework by transforming the original functional network into its edge-dual graph. Original functional edges become vertices in the edge-dual graph. Two vertices have a link between them if their corresponding edges in the original network shared a common end-node (region in the brain). To avoid confusion, we will use edges and nodes when discussing the original functional graph among brain regions (112 nodes and 6216 edges); and vertices and links to refer to the elements of the edge-dual graph (6216 vertices and 344988 links). Similar consideration of original edges as vertices in a line graph has been also adopted by others in the network analysis literature. The transformation is demonstrated in for a small example network of 4 nodes and 6 edges. If we start with a complete graph *G*(*N*, *E*) of \|*N*\| nodes and \|*E*\| = \|*N*\|(\|*N*\| − 1)/2 edges, the corresponding edge-dual graph *G*_*ed*(*E*, *L*) will have \|*E*\| vertices and \|*L*\| = \|*E*\|(\|*N*\| − 2) links. In our example graph, we have 4 nodes and 6 edges and in the corresponding edge-dual graph we obtain 6 vertices and 12 links. In this transformation, originally adjacent edges become vertices in the dual graph and two vertices are connected by a link. Note that this transformation ensures that a subgraph of connected vertices in the dual graph corresponds to a connected subgraph of edges in the original functional network. Discriminative subgraphs in a global network state setting can accurately differentiate between the global states. The space of potential discriminative subgraphs encompasses the set of all possible connected substructures of the functional brain network, which is exponential in the number of brain regions. Hence, the fair consideration of all subgraphs is computationally intractable even at the spatial resolution of 112 cortical and subcortical regions. Early methods in this area resort to sampling. While computationally efficient and ensuring the connectivity of discovered subgraphs, such approaches produce subgraphs that vary significantly across runs. To avoid such variation, we employ our recently introduced spectral learning method called *Sub-Network Learning (SNL)* to produce candidate biomarkers. The key idea behind SNL is to project the original network instances to a low-dimensional subspace in which instances of different global states are well separated. Simultaneously, the method regularizes the learning process by enforcing locality in the edge-dual network topology within the projection. Consequently, dimensions of the learned low-dimensional subspace correspond to combinations of connected subnetworks with high global-state classification accuracy. SNL constructs two meta-networks that capture the similarity relationships among network samples corresponding to individual sessions. It is important to note that vertices in these meta-networks correspond to edge-dual graphs, while a value associated with an edge represents the similarity between two edge-dual graphs. We denote with *G*⁺ the first meta-network, and with **A**⁺ its associated affinity matrix where an entry $\mathbf{A}_{ij}^{+}$ captures the similarity between two session-specific graphs *i* and *j* that have *the same* global state value. Likewise, we use *G*⁻ to denote the second meta-network and **A**⁻ its associated affinity matrix where entry $\mathbf{A}_{pq}^{-}$ captures the similarity between graphs *p* and *q* that have *different* global states. In we discuss several ways to compute the similarity among graph samples and for this study we adopt the cosine distance. Under this measure, all edge coherence values within a session are modeled as a vector and the pairwise cosine distances between session-specific vectors are used as weights in *G*⁺ and *G*⁻. Modeling more than two classes, e.g., different levels of learning rate, is possible within the framework where the semantics of *G*⁺ and *G*⁻ remain unchanged. For ordinal classes, i.e. learning rates with an order imposed on them, it might be desirable to apply an appropriate scaling for elements of *G*⁻ such that pairs of instances of “closer” labels incur smaller penalty. Prediction of continuous network labels (regression) would require a significant redesign of our learning framework, a direction we plan to explore in the future. It is important to note that higher number of classes or regression analysis would require much bigger training sets, hence we focus on two-class analysis for the sensorimotor experiment at hand. Given the dissimilarity information encoded in *G*⁺ and *G*⁻, we learn a transformation function that maps graph samples from the original space to a *d*-dimensional latent subspace, of which graphs with the same global labels are mapped close to each other, while graphs with different global labels are rendered far apart. This objective ensures that graphs from different classes are well-discriminated. Learning this transformation function is further regularized by the graph topology, which promotes the inclusion of well-connected subgraphs that are related to the prediction of the global low and high learning rates. SNL can learn an optimal subspace of arbitrary dimensionality *d* as long as *d* is smaller than the total number of vertices in the edge-dual graphs. However, similar to the approach of spectral clustering or Fisher’s linear discriminant analysis, one often chooses *d* equal to one less than the number of expected clusters (or classes). Hence, to discriminate between low and high learning states, we project network samples onto a 1-dimensional subspace (i.e., *d* = 1). The dual-optimization function can be formulated as follows: $$\begin{matrix} & \left\{ \begin{array}{l} {\operatorname{minimize}\sum\limits_{i = 1}^{m}\sum\limits_{j = 1}^{m}{\parallel \mathbf{u}^{T}\mathbf{v}_{i} - \mathbf{u}^{T}\mathbf{v}_{j} \parallel}^{2}\mathbf{A}_{ij}^{+}} \\ {\operatorname{maximize}\sum\limits_{p = 1}^{m}\sum\limits_{q = 1}^{m}{\parallel \mathbf{u}^{T}\mathbf{v}_{p} - \mathbf{u}^{T}\mathbf{v}_{q} \parallel}^{2}\mathbf{A}_{pq}^{-}} \\ \end{array}\operatorname{} \right. \\ & {\text{subject}\text{to}\mathbf{u}^{T}\mathbf{C}\mathbf{u} \leq t} \\ & {\text{and}\mathbf{u}^{T}\mathbf{V}\mathbf{D}\mathbf{V}^{T}\mathbf{u} = 1,} \\ \end{matrix}$$ where *m* denotes the number of edge-dual graphs, and vector **v**_*i* stores vertex values (i.e., coherence) of the *i*-th graph sample. Session-specific vectors **v**_*i* comprise the columns of matrix **V**. Matrix **D** is diagonal with entries $\mathbf{D}_{ii} = \sum_{j}\mathbf{A}_{ij}^{+}$. Additionally, **C** is the combinatorial Laplacian matrix encoding the topology of the edge-dual graphs, and *t* is a parameter that captures the impact of this topology on the coefficients of the mapping vector **u**. Unlike Lasso where a solution path can be found via varying a parameter controlling the L1-norm sparsity, here we impose a quadratic form penalty similar to L2-norm ridge regression. Thus, the solution defines a rank order (based on **u**’s elements) of network edges. The optimal value for *t* is selected via cross-validation which is typical for such settings. More specifically, we tune not only *t*, but also the number of selected edges based on grid search using an inner cross-validation on the training data. The first constraint in is similar to ridge-shrinkage in linear regression based on the graph topology, and thus its purpose is to “shrink” values of irrelevant vertices to zero, while the second constraint removes the scale freedom of vector **u** to ensure uniqueness of the solution. Detailed derivations and algorithmic solutions for optimizing this objective function can be found in. Compared to recent alternative techniques, SNL always produces a unique solution and more importantly, its solution is globally optimal with respect to the optimization function. Specifically, one alternative method, termed MINDS, is based on Markov Chain Monte Carlo (MCMC) while while another, called NGF, relies on a heuristic sampling procedure from the exponential space of all possible subgraphs. Both produce unstable results across runs due to their sampling from a large (exponential) space of possible subgraphs. These alternatives are, thus, less suitable for selecting consistent subgraph biomarkers within cohorts. For our task of *learning rate* classification, we learn the projection vector **u** using SNL and further threshold it to obtain a subset of the features. A linear classification model based on the retained features produces on average cross-validation accuracy of 81%. In contrast, a widely adopted support vector machine (SVM) classifier presented with the full set of 6216 features/vertices achieves 76% classification accuracy. Note that both results are based on 9-fold cross validation of the session instances. The reason for this difference is that SVM does not take advantage of the inherent graph structure among features which in our case model coherence values. Importantly, SNL outperforms SVM while using only a small subset of the features (i.e., those included in the discriminative subgraphs), while SVM uses all features. While feature selection can be performed as a preliminary step to improve SVM’s accuracy, off-the-shelf feature selection methods cannot enforce graph connectivity of the selected features. On the contrary, SNL optimizes classification accuracy and connectivity of employed features simultaneously. Hence, beyond classification accuracy, we employ SNL as a feature selector in order to identify candidate subnetworks that we then subject to significance testing, retaining only conserved and statistically significant subnetwork biomarkers. SNL’s source code is available at <http://www.cs.ucsb.edu/~dbl/publications.php>. ## Significant biomarkers Our substructure mining approach extracts a set of connected discriminative subgraphs based on the classes of global network state instances. For our application, one can view SNL as a feature selection method with regularization based on the graph structure captured using edge coherence values. When the number of training instances is small and the instances are high-dimensional, feature selection approaches may suffer from overfitting to the training data. Our goal, however, is to discover biomarker subgraphs whose predictive power is expected to generalize to novel unseen instances of functional networks and is also statistically significant. To achieve these desirable properties, we perform subgraph selection multiple times for random subsets of the training instances and focus on conserved subgraphs that are consistently selected. This approach is conceptually similar to Bootstrap aggregation with the distinction that our goal here is to detect a stable subset of discriminative features as opposed to combine the predictions of multiple classifiers. Therefore, since there is no available ground truth for biomarkers (i.e. functional edges that are guaranteed to be associated with learning), conventional quality measures like sensitivity, specificity and ROC curves are not applicable for quantifying the quality of the outcome. Instead, we evaluate the statistical significance of the predictive power of conserved subgraphs using a *q*-value statistic that implements a strict false discovery rate correction for multiple comparisons. **Conserved subgraphs.** A common machine learning approach used to improve the generality and stability of a learned model is Bootstrap aggregation, where multiple versions of a training set are generated and the individually trained models are aggregated to produce a single model. Our method generally follows this strategy. However, unlike Bootstrap aggregation which samples the data with replacement and actually is an ensemble method, we train our models based on cross validation and perform such cross validation multiple times in order to evaluate the consistency of the uncovered subnetworks. Specifically, we perform 9-fold cross validation 5 times (45 training sets in total) and select candidate biomarker subgraphs based on their consistency in cross validation, i.e., subgraphs conserved over multiple training runs. We also ensure connectivity in the resulting candidate biomarker subgraphs, as our goal is to capture differences in coordination among communicating brain regions involved in a common cognitive function or neurophysiological process (further details available). The problem of obtaining conserved subgraphs using SNL is similar to frequent connected subgraph mining (FSM). Given a database of subgraphs and a frequency threshold, the goal of FSM is to compute connected subgraphs that appear more frequently in the database than a specified frequency threshold. For our setting, such subgraphs will appear in at least a pre-specified number of trained models (i.e., subgraphs obtained by SNL). The rich literature on the general problem of frequent subgraph mining includes applications to computational chemistry, program analysis, and others with multiple proposed variations of the problem and corresponding methods. For our analysis, we employ gSpan, a commonly used and computationally efficient approach for general subgraphs. The input to the algorithm is the set of possibly disconnected subgraphs obtained over multiple runs of SNL and a frequency threshold. For our analysis, we require that subgraphs are selected in at least a third of the runs of SNL. Less conservative frequency thresholds, significantly increase the number of candidate subgraphs but do not increase the number of subgraphs that pass the significance test (described next). Further details on this step of the analysis are provided in the Supporting Information. **Testing the statistical significance of conserved subgraphs.** To evaluate the significance of the individual discriminative power of the obtained conserved subgraphs, we compute their *q*-values with respect to a random population of connected subgraphs of matching size. We choose the *q*-value as our significance measure as it reflects the false discovery rate (FDR), as opposed to the false positive rate (FPR) captured by *p*-values. The *q*-value measure of statistical significance has been employed for genome-wide studies and has significant advantages over alternative corrections for FDR. One challenge in computing the *q*-value for our subgraphs is that we need a background model of expected discriminative power of random graphs. To estimate this background model for subgraphs of varying number of edges, we sample connected graphs of fixed size uniformly at random and compute their accuracy in classifying all training instances based on an SVM classifier with polynomial kernel, involving the corresponding subset of features. To ensure uniform sampling of connected subgraphs, we employ a random-walk based sampling technique with degree-based rejection (details available). To estimate the *p*-values and subsequently *q*-values for our conserved subgraph candidates, we use the corresponding background accuracy distributions. We retain subgraphs of *q*-value ≤0.015, thus the FDR for our selected subgraphs is 0.015. # Results We apply our biomarker mining technique to the fMRI data acquired during the sensorimotor learning task described above consisting of session-specific functional networks coupled with global states: high and low learning rates. We discover 21 subgraph biomarkers that are both conserved in the sets of candidates produced by SNL and whose individual accuracy is significant (*q*-value ≤ 0.015). These candidate subgraphs are selected based on their consistent detection (at least in a third of all runs) in 9-fold cross validation performed 5 times where the optimal parameters are chosen based on the training accuracy and tested on the left-out fold according to. The sizes of the subgraphs range between 1 and 5 edges which connect 2 to 6 brain regions. Their individual training accuracy ranges between 74% and 85% (based on an SVM with a polynomial kernel). The list of all conserved and significant subgraphs as well as their individual accuracy and significance is provided in the Supporting Information. These subgraphs share edges and larger substructures, i.e., there is redundancy in the brain regions and connections they cover. We, thus, focus our analysis on the union of their edges which corresponds to two disjoint connected regions. shows the two biomarker regions (mapping of all brain region ids to anatomical names is provided). The first biomarker is the bilateral superior temporal- parietal (BSTP) biomarker, and it involves bilateral planum temporale and parietal operculum as well as the right superior temporal gyrus (the posterior portion). We represent the correlation between edge coherence and learning rate in. The arrows associated with edges capture both the correlation sign (positive correlation corresponding to upward green arrow) and the magnitude (arrow length). In the case of the BSTP biomarker, the dominant correlation is positive and in particular higher coherence in the circuit composed of the left and right planum temporale and the superior temporal gyrus corresponds to sessions of high learning rate. Particularly interesting in the bilateral superior temporal-parietal biomarker is the involvement of the parietal operculum whose coherence with planum temporale has a negative correlation with learning rate. Notably, the parietal opercular cortex is involved in manipulation and macroscopic tactile sensation, processes that are critical to participants learning this task which requires a growing familiarity with the keyboard and the mapping of stimuli to movements. Parietal opercular cortex is also involved in predicting sensory consequences of motor commands. In light of these roles for the parietal operculum in motor tasks, we can now interpret the BSTP biomarker. The learning rate is high when the earlier required dependence on simple sensory motor mapping has been completed, and therefore parietal operculum is no longer needed. The second biomarker is predominantly in the occipital lobe and we will refer to it as the bilateral occipito-temporal (BOT) biomarker. It interconnects the bilateral occipital fusiform gyrus, lingual gyrus, occipital pole, and intracalcarine cortex, as well as the right supercalcarine cortex. All of the above regions are involved in visual processing. Overall, the coherence between the cortical regions in the left and right hemisphere is negatively correlated with learning rate. This means that there is communication among visual cortex regions in the early sessions when subjects are still getting familiar with the visual cues and do not register high rates of task completion. This visual cortex coordination is reduced in high learning rate sessions. An exception to that trend is the circuit involving the left and right lingual gyrus and its coherence with the occipital poles and occipital fusiform gyri. This circuit is more coherent in high learning rate sessions and less coherent in low learning rate sessions. Unlike the calcarine cortices, the lingual and fusiform gyrus are higher order visual areas. We speculate that the nature of the task—visually decoding tablature with 12 different colored notes—is highly dependent on these higher order visual processing areas. Stronger coherence among these areas would lead to faster learning. The CSP task requires subjects to plan sequential movements from relatively complex visual stimuli. Interestingly, we found that subjects with more shallow learning curves had greater connectivity of primary visual cortical areas, typically involved in processing of low level visual features. On the other hand, we found that steeper learning curves were correlated with increased connectivity in higher-order visual regions, regions which are involved in visual object recognition, including words. These results indicate that greater synchrony in higher-level visual regions supports quicker learning in particular, when learning a motor skill involves the parsing of complex visual stimuli. This suggests that greater connectivity of higher-level visual regions involved in recognition might signify the perception of sequential note patterns as unique motor sequence identifiers and less as collections of individual notes. It is important to note that the available training data in our learning task do not deem any significant and predictive bridging edges/paths that connect the two biomarker regions. Under more observations or different learning tasks the two biomarker regions may merge in one or may change in terms of edge and region memberships. All distinct biomarker edges are also listed in together with corresponding statistics such as average coherence in low and high learning rate sessions and the actual value of correlation with learning rate. Additional analysis and discussion of the two biomarker regions follows in the subsequent sections. ## Differences in high- and low-rate learning sessions Biomarkers are discriminative between learning rates based on their individual edge coherence values. However, to better understand the potential differences in their coordination of distributed neural circuits, we next examine the biomarkers’ average coherence profile patterns. We visualize both biomarkers in their anatomical locations in the brain, and we vary the thickness of the edges to represent the average coherence observed in high and low learning rate sessions. The actual average values and their difference are also reported in. The most salient features of high-rate learning sessions associated with substantially *high* coherence are edges 40L/48L (left occipital fusiform gyrus—left occipital pole), 88R/96R (right occipital fusiform gyrus and right occipital pole) and 46L/94R (left and right planum temporale). These areas are key players in visual processing and color recognition, critical cognitive processes required by the task. The most salient features of high-rate learning sessions associated with substantially *lower* coherence are edges 47L/72R (left supercalcarine cortex and right intracalcarine cortex), 47L/96R (left supercalcarine cotex and right occipital pole) and 91R/94R (right parietal operculum and right planum temporale). In, we display edges with reduced coherence in high learners in red and edges with increased coherence in high learners in blue; the thickness of each edge corresponds to the magnitude of the difference. # Discussion Identifying the neural-circuit level drivers of higher order cognitive processes in humans is a critical frontier in human neuroscience. Reaching this goal will require concerted efforts across many fields of science, leading to the development of technologies from novel imaging techniques to new powerful computational tools that can extract both descriptive statistics and predictive features. We seek to address the latter challenge by offering a methodological approach to identify discriminative network biomarkers in the form of connected subgraphs that can be used to separate non-invasive neuroimaging measurements according to cognitive performance. Our approach capitalizes on recent advances in computer science and graph mining, and is accompanied by strict statistical testing and validation. We exercise our approach in the context of visuo-motor skill learning, and uncover two predictive biomarkers that distinguish high from low learning rate training sessions. Together, the method and application offer an important perspective on higher-order cognition in humans, and provide a generalizable toolset for use in other studies. ## Visuo-motor learning as a network process Traditionally, visuo-motor learning in both humans and animals has been studied at a relatively local level from the general perspective of brain mapping. In this view, regions of the brain, or ensembles of neurons, are identified as physical volumes whose activity may change as the animal learns. This approach has led to the extensive insights that build our intuitions about motor learning today. However, in recent years, this regionally-focused view has begun to be complemented by other perspectives highlighting the fact that changes in time series properties, functional connections, or even large-scale connectivity patterns may each play a role in neural computation and its relationship to behavior. Focusing on this latter approach, previous network-based studies of visuo-motor learning in humans in the specific context of finger sequences have delineated several features of the network dynamics or reconfiguration properties associated with learning. These studies together demonstrated that temporal changes in the community structure and centrality of brain regions (network nodes) can be detected in networks estimated from functional magnetic resonance imaging, complementing prior work in electrophysiological measurement modalities such as EEG and MEG. Here we take a complementary approach and instead ask the question of whether we can identify local subnetwork patterns that can predict high from low learning, irrespective of the time at which that learning occurred (e.g., early *versus* late in training). We address this question by comparing and contrasting the brain network structures observed in groups characterized by low *versus* high learning rate, and thereby identifying the most discriminative subnetworks. The uncovered patterns are thus in the form of local interactions among functionally correlated brain regions, and thus provide a deeper understanding of the local network structures facilitating visuo-motor learning in healthy adult humans. Our approach—deeply steeped in recent advances in computer science and graph mining—complements other recently developed statistical approaches. For example, Kim and colleagues recently proposed a set of statistical tests for comparing the functional connectivity between brain (or mental) states. The authors employ sparse matrix estimation based on the *graphical lasso* by imposing a fixed level of sparsity within a set of observations of one kind (e.g., disease or control), followed by regression and normalization of individual edges. Using this preprocessed data, the authors then apply the spatial pairwise clustering approach, or the network based statistic approach to test for network differences between classes. Importnatly, this analysis of Kim et al. enables a robust characterization of individual essential edges. Here, by comparison, we focus on the discovery of connected subgraphs that are discriminative using edges scores of both classes simultaneously. Our framework is, in a sense, orthogonal to the statistical tests for individual edges, in that the scores estimated by Kim et al.’s methodology can serve as features in our methods instead of raw coherence values. The novelty of our approach comes from enforcing connectivity of subgraphs (as opposed to disjoint edges) in both the discovery process and in the subsequent significance tests. That is, we test subgraph significance as opposed to individual-edge or all-edges significance. Our study also differs from prior work building on this same data, which has predominantly focused on extracting network patterns that distinguish good from poor learners in a purely descriptive manner that did not incorporate any sophisticated tools from machine learning. Indeed, the aforementioned studies focused on describing global network changes as a new skill was acquired over an extended training period of 6 weeks. For example, the results described in reveal that the core-periphery organization of the brain can predict individual differences in extended learning estimated from out-of-scanner behavior. More specifically, good learners were more likely to have a greater separation between the network core and the network periphery than poor learners. In a more recent study, the group sought to identify the fundamental functional modules present during visuo-motor skill learning, and reported a growing autonomy of these systems as participants acquired the new skill. In contrast to these descriptive approaches, the subgraph biomarker analysis that we present here offers an alternative lens in which to understand the differences in learners in a long-term training setting. Indeed, rather than identifying meso-scale structures such as communities, or macro-scale features such as centrality, this approach identifies sparse, local network motifs or subgraphs whose pattern of coherence can predict the behavioral outcome. In other words, this approach offers a much more parsimonious account of network characteristics supporting human learning. ## Biomarker regions, learning and further interpretation The human parietal operculum (OP) is a heterogeneous cortical area overlapping with multiple Broadmann regions. Evidence shows that it contains at least 2 sensory representation maps of the body and is involved in the processing of somatosensory information, responding to both non-noxious and noxious stimulation. Moreover, this region is involved in proprioceptive feedback during active movements, with recent work suggesting that this region is involved in the coordination of finger movements. The planum temporale (PT) is considered secondary auditory cortex and is functionally involved in higher order auditory and language processing, including speech, reading, and auditory-motor integration. Together, converging evidence suggests that this region acts as a specialized hub for spectrotemporal processing of stimuli. We found that reduced connectivity between the PT and OP was related to faster learning. This suggests that a greater independence of specialized hubs, one that is involved in hand related sensorimotor feedback (OP) and the other, in spectrotemporal processing (PT), served to promote learning. Moreover, greater cross-hemispheric connectivity of the PT with other neighbors, might serve to strengthen learning as well. In this regard, we presented participants with stimuli that represented a music-like notion, which required participants to “read” the notes, and map vertical and horizontal position to the appropriate finger. So, it is possible that greater connectivity with PT leads to swift learning, which could reflect a greater proficiency in translation of the music staff to motor output. It is unlikely that this effect is due to individual differences in music training because we selected participants with minimal music experience (less than 4 years total). More generally, the relationship between individual differences of learning rate and local functional connectivity is consistent with an emerging literature that considers the evolution of brain activity across networks rather than local regions. Historically, both increases and decreases of regional activity have correlated with amount of training, but not individual differences of learning rate. With the development of functional connectivity metrics, it was evident that local, pairwise connectivity could also change with learning and correlate with depth of knowledge. In recent studies that consider the evolution of functional connectivity across larger brain networks, it has become apparent that the strength of connectivity, both positive and negative, between separable network communities can evolve with training. Allegiance of a network node to a community, as well as allegiance between networks are strong predictors of learning as well as individual differences in the rate of learning. Indeed, it has been possible to show that over-involvement of prefrontal areas associated with executive control are associated with slower rates of learning, presumably by delaying the emergence of autonomous motor behavior. In this context, the current results provide additional complementary evidence by demonstrating that increased connectivity between secondary somatosensory cortex and the temporal cortex associated with abstract sequential information can also result in slower learning. Whether this is due to a delay in the development of autonomous motor behavior, a competition between brain systems that represent sequential information differently or some other process remains to be determined. ## Methodological considerations There are several important methodological considerations pertinent to this work. First, it is important to note that the methodological approach that we develop and apply in this work—based on subgraph biomarker mining—does not consider dynamic or time-evolving aspects of the functional interactions. However, extensions of the approach that we develop here to time-evolving networks could be particularly useful in studying the ability of a brain region to broadcast or receive information in learning tasks. Such extensions are likely possible by building on the recent methods for extracting significant dynamic subgraphs from temporal networks of various genres. Indeed in future, it will be particularly interesting to ask how to mine neuroimaging data for significant subgraph biomarkers as the functional (or structural) networks evolve in time. Answering this question could offer an important view into the dynamics of circuit function essential for human learning. A second important consideration lies in the empirical challenges inherent in collecting long-term training data. In this study, we used data acquired in 3 sessions spaced over 5 days from 18 healthy adult individuals. Such a longitudinal study is extremely difficult to complete in terms of recruiting, cost, and personnel resources, and is therefore a particularly valuable resource. Nevertheless, it would be important in future to validate our results on similar longitudinal data sets acquired in a separate set of healthy adult human subjects. A third important factor that deserves consideration is the length of training time. We kept each practice session below 1.5*hr* to decrease potential for fatigue, thus extending our study over 3 days to adequately sample of the first, fast rate of improvement characteristic of early motor skill learning. We anticipate that the extracted biomarker is relevant for the initial stages of learning, but we cannot claim that this same biomarker would be identified if we studied longer term learning, where other cognitive processes are thought to be involved. Another consideration of interest is that the discriminative subgraphs that we identify in this work are likely to be highly-specific to the particular visuo- motor task performed. Other types of learning tasks, and even other types of visuo-motor tasks (such as visuo-motor tracking) may display different discriminative subgraphs. It will be particularly interesting in future to catalogue the discriminative subgraphs that distinguish high *versus* low cognitive performance or task performance across a range of motor and non-motor learning tasks. ## Limitations There are several limitations of the current imaging and analysis, that will be important to address in the future via additional experimentation and generalized computational analysis. One important limitation of the neuroimaging component is that the acquisition was performed without multiband technology, thus decreasing the temporal resolution of the data. It would be useful in future studies to use recently developed higher-resolution temporal sampling techniques to increase the statistical power to detect individual differences in neural markers of learning. The complexity of the experimental task resulting in a limited number of subjects and respectively learning sessions is another limitation. While we have taken extensive measures to alleviate this drawback: (i) repeated cross validation with fold re-sampling, (ii) L2 regularization to increase the stability of selected subgraphs and (iii) subsequent subgraph statistical significance testing based on q-values, we expect that larger datasets will enable even more stable and statistically significant biomarkers. Beyond more data, further computational extensions can be considered to further alleviate the relatively small number of instances compared to features. In future studies, we will seek to incorporate L1-norm constraints which can shrink coefficients of irrelevant edges to zero (sparsity). Another improvement may be enabled by adopting structural connectivity maps as priors for coordination among regions as opposed to solely the observed functional coherence from fMRI. Such analysis will benefit from low level of spurious interactions due to noise, imaging artifacts or concurrent processes in the brain, but will require diffusion imaging scans for the participating subjects. # Conclusions We developed a general approach for the discovery of brain subgraph biomarkers from fMRI data associated with global labels. Our approach is based on discriminative subspace learning in network space coupled with significant conserved subgraph mining. We applied our method to data acquired during the performance of a sensorimotor learning task. We obtained two significant biomarkers involving circuits related to visual processing, motor performance, and learning, which together suggest novel interactions among regions that may play a critical role in visuo-motor skill learning. While we focused on data from a learning experiment as a case study for our method, our framework can be applied to a variety of settings. Beyond analysis of other cognitive tasks, one can also adopt our method to detect biomarkers specific to neurological and psychiatric diseases, by applying the method to fMRI data acquired in patients and controls. # Supporting information AKS would like to acknowledge support from the National Science Foundation grant IIS-1219254. This data is based on support to STG by the US National Institutes of Health (grant: P01 NS044393) and the Institute for Collaborative Biotechnologies through contract W911NF-09-0001 from the U.S. Army Research Office. DSB would also like to acknowledge support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research Laboratory and the Army Research Office through contract numbers W911NF-10-2-0022 and W911NF-14-1-0679, the National Institute of Health (2-R01-DC-009209-11, 1R01HD086888-01, R01-MH107235, R01-MH107703, R01MH109520, and R21-M MH-106799), the Office of Naval Research, and the National Science Foundation (BCS-1441502, CAREER PHY-1554488, and BCS-1631550). The content is solely the responsibility of the authors and does not necessarily represent the official views of any of the funding agencies. [^1]: ND’s commercial affiliation with her current employer Ticketmaster does not alter our adherence to PLOS ONE policies on sharing data and materials since ND contributed to the manuscript while pursuing her MS degree at UCSB and before joining Ticketmaster.

# Introduction Ischemic stroke is characterized by overstimulation of glutamate receptors of the N-methyl-D-aspartate type (NMDARs) and increased inflow of intracellular Ca²⁺, \[Ca²⁺\](i). NMDAR overactivation disrupts antioxidant defenses and critical survival pathways, which not only increase the susceptibility of neurons and glia to ischemic damage but also trigger numerous ischemic cascades, leading to further neuronal degeneration, swelling, or even deaths. Nonetheless, efforts to inhibit NMDARs have generally failed, mainly due to critical roles of these receptors in neuronal survival and synaptic plasticity. Other strategies to improve neuronal defense against the glutamate- induced excitotoxicity and/or to decrease the \[Ca²⁺\](i) inflow into the ischemic neurons have, therefore, been suggested. Several different classes of antioxidants and/or neuroprotectants such as calpain inhibitors have been shown to protect against ischemia-induced excitotoxicity and, thus, decrease brain damage caused by experimental stroke. Magnolol, a blood-brain barrier permeable phenolic constituent (5, 5′-dially-2, 2′-dihydroxydiphenyl) of magnolia bark, is known to be a central nervous system depressant agent and potent antioxidant. Magnolol has been shown to protect against brain damage in an experimental heatstroke model. We have previously shown that magnolol protects against hind limb ischemic-reperfusion injury in rats by reducing post-ischemic rises in the levels of nitrite/nitrate (NOX), malondialdehyde (MDA) and myeloperoxidase (MPO). Alternatively, magnolol is an inhibitor of voltage-dependent Ca²⁺ channel and can reduce necrotic cell deaths in mixed neuron-astrocyte cultures exposed to chemical hypoxia. Accordingly, we suspected that magnolol might protect the brain against ischemic stroke. In the study, we evaluated the protective effects of magnolol against cell damage and swelling as well as increased inflow of \[Ca²⁺\](i) in cultured neurons exposed to glutamate. Additionally, we investigated neuroprotective efficacy of magnolol in rats subjected to permanent focal cerebral ischemia. # Results Neurotoxicity of magnolol was observed with a concentration beyond 100 µM. The LD₅₀ value was 129.3±0.1 µM (*P\<*0.05;). Alternatively, the glutamate- and NMDA-induced neurotoxicity was significantly attenuated by magnolol at 0.1–1.0 µM (*P\<*0.05;, respectively), and their ED₅₀ values were 0.3±0.1 and 0.2±0.1 µM, respectively. At 10 *DIV* cultured neurons, the addition of glutamate (300 µM) induced abrupt rises of \[Ca²⁺\](i) levels up to ≈1000 nM. Magnolol, however, exhibited a hormetic inhibitory response. Only magnolol at 0.1 µM, but not at 0.01 and 1 µM, effectively inhibited the increased \[Ca²⁺\](i) inflow over time (*P\<*0.05;). On contrast, treatment with magnolol at 0.1–1 µM invariably attenuated the glutamate-induced neuronal cell swelling over time (*P\<*0.05;). Nine animals (7.8%) died prior to completing the protocol following pMCAO and were excluded: 4 (8.2%) were in the vehicle-treated groups and 5 (5.4%) were in the magnolol-treated groups. Following the ischemic onset, the ipsilateral LCBP declined to 14–22% and 32–38% of the baseline data in the ischemic core and penumbral areas, respectively. The LCBP was not significantly different among experiment groups, and was independent of magnolol treatments (*P*\>0.05; data not shown). The other physiological parameters were kept within normal limits and did not differ significantly among experiment groups, except that high-dose (200 mg/kg) magnolol-treated animals had arterial pCO₂ retention along with reduced heart rate and arterial pH. Ischemic animals invariably experienced spontaneous hyperthermia throughout the recovery period. Animals treated with magnolol at 200 mg/kg, but not at lower doses, had modest temperature reductions by ≈3°C, and this temperature-lowering effect remained effective up to 2 hrs post-insult. Animals which were pre-treated with magnolol, at 50 mg/kg (n = 11), 100 mg/kg (n = 8), 150 mg/kg (n = 9), or 200 mg/kg (n = 12), but not at 25 mg/kg (n = 13), 1 hr before the ischemic onset, showed significant infarct size reductions (*P\<*0.05) when compared with controls (n = 15). Infarction lesions were reduced by 30.9, 33.8, 37.8 and 35.3%, in animals treated with magnolol at 50 mg/kg, 100 mg/kg, 150 mg/kg, and 200 mg/kg, respectively. Animals treated with magnolol (100–200 mg/kg) also showed significantly improved sensory neurologic scores taken 22–24 hrs post-insult than did controls (*P\<*0.05;). Additionally, significantly less body weight loss was observed in animals treated with magnolol (25–200 mg/kg), compared with controls (*P\<*0.05;). In the delayed treatment paradigm, our results indicated that magnolol (100 mg/kg) resulted in significant infarct volume reductions when administrated within 4 hrs after the ischemic onset (*P*\<0.05;). Relative to controls (n = 30), infarction were reduced by 42.5, 28.5 and 20.6%, respectively, when magnolol was given at 1 (n = 7), 2 (n = 7), and 4 hrs (n = 9) post-insult (*P*\<0.05;). Magnolol treated at 6 hrs post-insult (n = 9) did not significantly reduced brain infarction. However, delayed treatment with magnolol significantly improved sensory neurologic scores, even when administered up to 6 hrs post-insult (*P*\<0.05), and effectively reduced post-ischemic body weight loss, when administered up to 2 hrs post-insult (*P*\<0.05), but did not affect post-ischemic motor scores (*P\>*0.05;). The physiological parameters were kept within normal limits and did not differ statistically between study and control animals (data not shown). # Discussion Our results indicated that magnolol (50–200 mg/kg) reduced infarct volumes and improved neurobehavioral outcomes in rats subjected to permanent focal cerebral ischemia. Additionally, we found that magnolol (100 mg/kg) was effective in reducing brain infarction and improving neurobehavioral outcomes even when administrated up to 4 hrs post-insult. Moreover, we demonstrated that magnolol not only effectively attenuated both glutamate- and NMDA-induced neurotoxicity, but also reduced the glutamate-induced increases in the \[Ca²⁺\](i) inflow and neuronal swelling. This neuroprotection cannot be accounted for by changes in glucose, hemodilution (as measured by blood hematocrit), or differences in mean arterial blood pressure, since these were not significantly different when compared between vehicle-injected and magnolol-treated animals. The changed physiologic parameters were decreases in arterial pH and heart rate, associated with a rise in pCO₂, seen in the animals treated with magnolol at 200 mg/kg. These findings suggested that magnolol (200 mg/kg) might have induced a cardiopulmonary suppression, probably due to its centrally-acting muscular relaxant effect. Exactly by which mechanisms in the glutamate-stimulated cultured neurons the dose-responsive regimen seen with magnolol for cell swelling inhibition was inconsistent with the “U-shaped” hormetic response observed for inhibiting the rises of \[Ca²⁺\](i) remains to be elucidated. Curiously, hormetic neuroprotective responses were also observed in the magnolol-treated stroke animals in which a low-dosing regimen was ineffective whereas high dosage (200 mg/kg) induced adverse effects along with a temperature-lowering action. Thus, the *in vitro* dosing response might not represent the trend of dosing response observed *in vivo*. It was very likely that magnolol actually had multiple mechanisms acted, independently or in combined, to exhibit neuroprotection observed here. A therapeutic window of 4 hrs seen with magnolol in reducing brain infarction compares favorably with those of glutamate receptor antagonist and other anti- oxidant and radical-scavenging agents, but not as well as that reported with a calpain inhibitor. Perhaps using multiple effective, smaller doses of magnolol, combined with an intravenous administration route, the therapeutic window may be extended and/or the degree of neuroprotection improved. Further studies are needed to determine whether magnolol can protect against reperfusion damage and late-onset ischemic insults following cerebral ischemia/reperfusion after a prolonged reperfusion period. In additional, more mechanisms underlying neuroprotection observed here need to be elucidated. In conclusion, magnolol protects against permanent focal cerebral ischemia with a therapeutic window up to 4 hrs post-insult. This neuroprotection may be partly mediated by its ability to attenuate the glutamate and NMDA-induced neurotoxicity. # Materials and Methods All procedures performed were approved by the Subcommittee on Research Animal Care of the University. All chemicals were purchased from Sigma-Aldrich Co. (St Louis, MO) unless otherwise indicated. Hank’s balanced salt solution (HBSS 10×, GIBCO, Grand Island, NY) was composed of (mM): glucose 55.56, KCl 53.33, NaCl 1379.31, KH₂PO₄ 0.44 and Na₂PO₄ 3.36; pH 7.1. Magnolol (Wako Pure Chemical Industries, Ltd., Osaka, Japan) was dissolved in PEG 400 or dimethylsulfoxide (DMSO). ## Neuronal Cultures and Cytotoxicity Assay According the method described previously, cultured neurons were obtained from cerebral cortices of 1-day-old Sprague-Dawley rats. Cytotoxicity was determined at 24 hrs after treatment by using a LDH assay kit (Promega, Madison, WI). Experiments were undertaken on cultured neurons between 10 and 14 days *in vitro* (DIV). Neurons were incubated magnolol (0–300 µM) or vehicle (0.1% DMSO). The LD₅₀ value was defined as the concentration of compound required to induce 50% of cell deaths in 24 hrs at 37°C. ## Glutamate- and N-methyl-D-aspartate (NMDA)-induced Cell Cytotoxicity Cultured neurons were pre-treated with magnolol (0.1–1 µM) or vehicle (0.1% DMSO) for 30 min and, then, were exposed to glutamate (300 µM) or NMDA (100 µM) for 24 hrs. The ED₅₀ value was defined as the concentration of compound required to reduce 50% of cell deaths of controls in 24 hrs at 37°C. ## Intracellular Ca²⁺ Measurement The level of \[Ca²⁺\](i) were measured on a single cell fluorimeter. Briefly, neuronal cultures were incubated with 3 µM fura 2-acetoxymethylester (Fura-2 AM) and 10 µM ionomycin in a standard buffer (composition in mM: NaCl, 140; KCl, 3.5; KH₂PO₄, 0.4; Na₂HPO₄, 1.25; CaCl₂, 2.2; MgSO₄, 2; glucose, 10; HEPES, 10, pH 7.3) for 30 min, followed by incubation in dye-free standard buffer for 30 min and, then, the addition of vehicle or magnolol (0.01, 0.1, or 1 µM) for 20 min and the exposure of glutamate (300 µM). During experiments, standard buffer was replaced by low Mg²⁺ saline (composition in mM: NaCl, 140; KCl, 3.5; KH₂PO₄, 0.4; Na₂HPO₄, 1.25; CaCl₂, 2.2; MgSO₄, 0.03; glucose, 10; HEPES, 10, pH 7.3). The glass coverslip was placed into the stage chamber of an Olympus IX71 inverted microscope, equipped with a 75 W xenon illumination system, a cooled charge-couple device (CCD) camera (300T-RC; Dage-MTI, Michigan City, IN) coupled to an image intensifier (Gen II S-25 image intensifier; Dage-MTI), a Lambda 10-2 filter-wheel and shutter (Sutter Instruments, Novato, CA) and a computerized image analyzer (MCID Elite, Imaging Research Inc., St. Catherines, Ontario, Canada). The cells were alternatively illuminated with the light of 340 and 380 nm wavelengths and the emitted light was passed through a 510 nm barrier filter. The 340 and 380 nm images were captured at 6 second intervals and the ratio signals (340 nm excited image/380 nm excited image) were processed and examined for real changes in \[Ca²⁺\](i)_. Approximately 10 neurons in each microscopic field were individually measured. The \[Ca²⁺\](i) level was calculated by using the equation: \[Ca²⁺\](i)  =  Kd×(Fo/Fs)×\[(R−Rmin)/(Rmax−R)\] where Kd is the dissociation constant for fura -2 in the cytosol (225 nM), and Fo/Fs is the fluorescence emitted at 380 nm excitation at minimum Ca²⁺ level divided by the same emission fluorescence at the fura-saturated concentration. R is the ratio fluorescence intensity recorded at 340 and 380 nm, and Rmin and Rmax are the rations of 340/380 nm fluorescence intensity recorded at minimum Ca²⁺ and the fura-saturated Ca²⁺ concentrations, respectively. We used the Fura-2 Calcium Imaging Calibration Kit (F-6774; Invitrogen Molecular Probes, Eugene, OR) to detect the Kd level under conditions. Measurements of Fo and Rmin were performed in nominally Ca²⁺-free isotonic solution containing 10 mM EGTA. Cells were then superfused with isotonic solution containing 1 µM thapsigargin, 10 µM ionomycin and 10 mM Ca²⁺ to evaluate Fs and Rmax. ## Cell Swelling Measurements The glutamate (300 µM)-induced neuronal morphologic changes were measured by time-lapse imaging techniques in a microscope equipped with a thermo- controllable heating stage, differential interference contrast (DIC) lens and an image analyzer (MCID Elite) by the method described previously. DIC images of pyramid-shaped neurons were measured and compared over time. Three randomly selected fields were counted and averaged per culture (approximately 12 to 15 neurons per culture). Data are expressed as a percentage relative to the baseline values. ## Animal Preparation, Anesthesia, and Monitoring Male Sprague-Dawley rats, weighting 220–270 g, were supplied by the University Laboratory Animal Center, and were allowed free access to food and water before and after surgery. Animals were anesthetized with 1–2% halothane in 70% N₂O/30% O₂. During surgery, body temperature was maintained at 37±0.5°C using a thermostatically controlled heating blanket and rectal probe (Harvard Apparatus, South Natick, MA). The right femoral artery was cannulated for measuring arterial blood gases, glucose, hematocrit and blood pressure. ## Experimental Model Focal cerebral ischemia was employed by permanent occlusion of the proximal right middle cerebral artery (pMCAO) with a 4-0 nylon suture occluder, as described previously. Successful MCA occlusion was ensured by a sharp decrease of local cortical blood perfusion (LCBP) to about 20% of baseline as determined by Laser-Doppler flowmetry (LDF, Laserflo BMP², Vasamedics, St. Paul, MN). ## Drug Administration and Grouping of Animals In the first series of experiments, animals were received either magnolol (25 mg/kg, 50 mg/kg, 100 mg/kg, 150 mg/kg, or 200 mg/kg, i.p.; n = 58) or vehicle (the same volume of PEG 400, i.p.; n = 16), 1 hr pre-insult, to test the neuroprotective dose response. An additional set of rats, received magnolol (100 mg/kg, i.p.; n = 34) or vehicle (PEG 400, i.p.; n = 33) at 1, 2, 4 or 6 hrs post-insult, was used to evaluate the therapeutic window of opportunity. ## Neurobehavioral Testing Neurologic and body weight measurements were conducted by an investigator unaware of treatment protocol at 24 hrs post-insult,. Five categories of motor neurologic findings were scored: 0, no observable deficit; 1, forelimb flexion; 2, forelimb flexion and decreased resistance to lateral push; 3, forelimb flexion, decreased resistance to lateral push and unilateral circling; 4, forelimb flexion, unable or difficult to ambulate. The affected forelimb also received forward and sideways visual placing tests which were scored as follows: 0, complete immediate placing; 1, incomplete and/or delayed placing (\<2 seconds); 2, absence of placing. ## Animal Sacrifice and Quantification of Ischemic Damage Sacrifice was performed at 24 hrs post-insult by decapitation under anesthesia. The brain was cut into 2 mm coronal sections using a rat brain matrix (RBM 4000 C, ASI Instrument, Inc., Warren, MI) and stained according to standard 2, 3, 5-triphenyltetrazolium chloride (TTC) method. Briefly, the brain was cut into 2 mm coronal sections using a rat brain matrix (RBM 4000 C, ASI Instrument, Inc., Warren, MI) and stained according to the standard 2, 3, 5-triphenyltetrazolium chloride (TTC) method. A total of 7 brain sections were traced and measured using a computerized image analyzer (MCID Elite). The calculated infarction areas were then compiled to obtain the infarct volumes per brain (in mm³). Brain Infarct volumes were expressed as a percentage of the contralateral hemisphere volume. ## Statistical Analysis All data were expressed as the mean±standard deviation (S.D.). Paired Students’ *t* test was used to evaluate the response to a change in conditions, and unpaired Students’ *t* test/one-way analysis of variance (one-way ANOVA) with Fisher’s protected least significant difference (LSD) *posthoc* comparison was used to evaluate differences between groups. Neurobehavioral scores were analyzed by the Kruskal-Wallis/Mann-Whitney *U* test. *P*\<0.05 was selected for statistical significance. [^1]: Conceived and designed the experiments: WTL MHL EJL TSW. Performed the experiments: WTL MHL EJL YCH SHT HYC. Analyzed the data: MHL TYC. Contributed reagents/materials/analysis tools: MHL TYC TSW. Wrote the paper: WTL MHL EJL. [^2]: The authors have declared that no competing interests exist.

# Introduction *Trypanosoma cruzi* is the etiological agent of Chagas’ disease, which affects between 16 and 18 million people, primarily in Central and South America. This single-cell protozoan has a complex life cycle, alternating between an insect vector (triatomine) and mammalian hosts. As a consequence, *T. cruzi* is continuously exposed to drastic environmental changes. Therefore, in order to achieve a rapid adaptation to such conditions, *T. cruzi* requires a flexible modulation of its gene expression profile. Gene expression in trypanosomes presents several differences compared to most part of the eukaryotic lineage. Among some of these differences, there is a general absence of defined promoters for protein-encoding genes. In addition, transcription of mRNAs is polycistronic and constitutively generates pre-mRNAs, which are finally processed into individual mRNAs by trans-splicing and polyadenylation. Trans-splicing consists in the addition of a 39-nucleotide capped RNA (known as mini-exon or Spliced Leader RNA) at the 5′end of all mRNAs. Hence, it is considered that gene expression is primarily modulated at the post- transcriptional level, mainly by mRNA stability and translation control. In recent years, Stress Granules (SG) and P-bodies have also emerged as another post-transcriptional layer of gene expression regulation. These cytoplasmic structures have been postulated to modulate translation, operating either as protective or degradative mRNA reservoirs, especially when cells are subjected to stress. In this regard, it has been shown that trypanosomes exposed to certain stress conditions, such as starvation and severe heat shock, are also able to display those cytoplasmic structures. Recently, the resolution of the nucleolar proteomes of yeast, *Arabidopsis,* and humans has surprisingly shown the presence of several RNA Binding Proteins (RBPs) involved in different steps of mRNA metabolism. In addition, it has been reported that the nucleolus could also be involved in the regulation of other mechanisms related to mRNA metabolism, such as production of small interfering RNAs, maturation of microRNAs, accumulation of aberrantly spliced mRNAs, and export of viral mRNAs to the cytoplasm. These observations support the notion that, besides its traditional function in ribosome biogenesis, the nucleolus could also be involved in the regulation of other cellular processes including mRNA metabolism. There is also a growing body of evidence arguing in favour of its possible role as a sensor and coordinator of the stress response, for instance, by sequestering key factors essential for the regulation of gene expression. In this regard, previous works in yeast have shown that poly(A)+ RNA could be accumulated into the nucleolus in response to severe heat shock, supporting the idea that the nucleolus is involved in mRNA transport. Furthermore, it has been shown that whereas mRNAs transcribed from intron- containing genes are exported to the cytoplasm through a nucleolar phase, mRNAs generated from intron-less genes are not, suggesting that spliced and unspliced transcripts are exported by different pathways in fission yeast. More recently, we have reported that the *T. cruzi* nucleolus is probably involved in the stress response since several RBPs and the poly(A)+ RNA are relocalized into this structure in response to transcription inhibition by Actinomycin D (ActD). In this work, we found that most mRNA population is partially and reversibly accumulated into the *T. cruzi* nucleolus under an environmental stress, such as is a severe heat shock. Interestingly, the Heat shock protein70 (Hsp70) mRNA was able to bypass such nucleolar accumulation. These data reinforce the notion about the potential role of the nucleolus in the mRNA metabolism in *T. cruzi*. # Results ## Severe Heat Shock Induces Nucleolar Accumulation of mRNAs in *T. cruzi* We have recently reported that poly(A)+ RNA is accumulated into the nucleolus under ActD treatment in *T. cruzi*. Taken into account our previous results showing that severe heat shock stress induces the relocalization of some RBPs to the nucleolus, we wondered whether this stress could also induce a similar poly(A)+ RNA behaviour. To test this, we performed RNA-FISH using a Cy3-oligo(dT) probe against the poly(A) tail of mRNAs. Under normal conditions (28°C), the bulk of the poly(A)+ RNA population was distributed mainly throughout the cytoplasm, and also in the nucleus but to a much lesser extent level (panel 1,). However, when an epimastigote culture was subjected to heat shock at 40°C for 2 h, in addition to the cytoplasmic signal, a significant nucleolar accumulation of poly(A)+ RNA was observed in 57% of parasites ( panel 2,). To further support this result, we repeated the RNA-FISH assays using a probe against the mini-exon sequence, which is present at the 5′end of all mature mRNAs in trypanosomes. With this probe we obtained similar results, with 62% of the parasites displaying mRNA nucleolar accumulation in response to heat shock (panels 1 and 2, and). As controls, we performed the following assays: i) RNase A digestion before the hybridization step (panels 4), and ii) RNA-FISH using a Cy3-oligo(dA) probe. In both control experiments, the fluorescence remaining after the hybridizations was negligible, thus confirming the probe specificity and the RNA nature of the nucleolar signals. To find out whether the mRNA nucleolar accumulation observed was a reversible process, parasites were heat-shocked for 2 h and then re-incubated at the normal growth temperature. After 6 h at 28°C, the parasites showed an mRNA distribution similar to that of control parasites in almost the whole population (panels 3; see for a quantitative analysis). From these results, we conclude that severe heat shock in *T. cruzi* epimastigotes induces a partial relocalization of the bulk of mRNA to the nucleolus in a reversible way. ## Nucleolar Accumulation of mRNAs could be Bypassed by Hsp70 mRNAs In yeast cells, the poly(A)+ RNA is accumulated into the nucleolus in response to severe heat shock. Interestingly, under this condition, the export of mRNAs, such as that of Hsp70, required to mount the stress response is not inhibited. To test whether a selective nucleolar accumulation of mRNAs could also take place when subjecting *T. cruzi* to severe heat shock, we evaluated the behaviour of two well-characterized mRNAs in trypanosomes: α-Tubulin (Tc00.1047053411235.9) and Hsp70 (Tc00.1047053511211.160, Tc00.1047053511211.170). It is worth mentioning that the Hsp70 mRNAs, in clear contrast with most mRNAs, are continuously translated even when parasites are subjected to severe heat shock. Under normal conditions, both transcripts showed mainly a cytoplasmic localization (, top panels, respectively; S2 and S3). However, after incubating parasites at 40°C for 2 h, both mRNAs showed a different sub-cellular distribution. As expected, a significant fraction of the α-Tubulin mRNA pool showed a nucleolar accumulation in 50% of the parasites (middle panels, S2 and S3). On the other hand, the Hsp70 mRNA did not show a significant alteration in its localization pattern, being mainly localized throughout the cytoplasm (middle panels, S2 and S3). These results suggest that in *T. cruzi*, mRNAs of proteins required to activate cellular defence mechanisms under stressed conditions are refractory to nucleolar accumulation, as it has been previously shown in yeast. To further support that non-heat-shock response mRNAs are accumulated into the nucleolus under severe heat shock, we evaluated the distribution of mRNAs belonging to the Small Mucin-like family (Smug),. We performed the assay using a probe that could potentially hybridize with all its family members. Under normal conditions (top panels, S2 and S3) Smug mRNAs were distributed throughout the cytoplasm similarly to both Hsp70 and α-Tubulin mRNAs. When parasites were subjected to heat shock, the Smug mRNAs were partially accumulated into the nucleolus in 52% of the parasites (middle panels, S2 and S3). To validate that the *in situ* hybridization assays were specifically detecting the mRNAs evaluated, we performed several control experiments. The addition of an excess of the unlabelled oligonucleotide probe to the hybridization buffer significantly decreased the signal intensity for each mRNA analysed, whereas the addition of the same amount of an unlabelled oligonucleotide generated by randomizing the corresponding probe sequence did not significantly affect the intensity of the mRNA signal. In addition, we performed RT-PCR using each probe as a reverse primer in combination with a forward primer against the mini-exon sequence, which hybridizes with all mRNAs. For each mRNA evaluated, we observed only one band of the expected size, confirming the specificity of each probe by an independent approach. Finally, pre-treatment with RNase A before the hybridization steps almost completely abolished the nucleolar signal in heat- shocked parasites, thus demonstrating the RNA nature of the signals observed (bottom panels). Together, these results suggest that severe heat shock induces a selective nucleolar accumulation of mRNAs, where those mRNAs that codify for proteins not directly involved in the heat shock response are susceptible to be partially relocalized into the nucleolus; in contrast, mRNAs codifying proteins required for the heat shock response, for instance Hsp70, could bypass such nucleolar retention. ## Nucleolar Accumulation of Poly(A)+ RNA is Absent in *T. brucei* Recently, the analysis of several Stress Granules (SG) protein markers (for instance TbPABP2) has shown that severe heat shock leads to SG formation in *T. brucei*. Therefore, we wondered whether a fraction of the poly(A)+ RNA population could also be accumulated into the nucleolus in response to such stress. Under normal conditions, the poly(A)+ RNA population was distributed throughout the cytoplasm, similarly to that observed in *T. cruzi*. Nevertheless, after subjecting *T. brucei* procyclic forms to heat shock at 40°C for 2 h, the poly(A)+ RNA distribution was notably modified, being the formation of cytoplasmic stress granules the most remarkable structures observed in almost the whole population of parasites (8). However, we did not observe nucleolar accumulation of poly(A)+ RNA under this stress condition. As previously reported by us, poly(A)+ RNA nucleolar relocalization can be induced by ActD treatment in *T. cruzi*. So, we also tested this treatment in *T. brucei*. Under such condition we did not observe obvious nucleolar accumulation (compare); instead the poly(A)+ RNA population remained distributed throughout the cytoplasm, but in lower amounts than those observed in untreated parasites. We repeated the experiment evaluating longer times (24 h) and found only a higher cytoplasmic signal decay (not shown). These results are in agreement with the absence of nucleolar relocalization of RBPs in response to ActD treatment reported by our group in *T. brucei*. Taken together, these results indicate that in *T. brucei* neither poly(A)+ RNA nor RBPs are mobilized to the nucleolus in response to stress conditions. # Discussion We have recently reported that several RBPs involved in the mRNA metabolism of *T. cruzi,* as well as the poly(A)+ RNA, are accumulated into the nucleolus in response to ActD treatment. In this frame, we aimed to further characterize the potential role of the nucleolus in mRNA metabolism from a “physiological” perspective. Therefore, we focused our attention on severe heat shock, since this stress has been reported to induce a reduction in transcription levels in trypanosomes, , and, more importantly, because it is an environmental stress to which *T. cruzi* could be exposed during its life cycle. In this work, we found that the bulk of the mRNA population is partially relocalized to the nucleolus when *T. cruzi* epimastigotes are subjected to heat shock at 40°C. In addition, we showed that this process was fully reversed when parasites were shifted to normal growth conditions (see, panels 3). It should be mentioned that, in addition to transcription, trans-splicing was significantly reduced when exposing parasites at 40°C since we detected an accumulation of the SL RNA. Along this line, we have also observed a nucleolar accumulation of SL RNA. In this regard, it is possible that partially processed mRNAs could also have accumulated in the nucleolus. However, we believe that the observed signals likely correspond to mature mRNAs since, as reported, transcription inhibition prevents the accumulation of incompletely processed mRNAs. In addition, it has been shown that such incompletely processed mRNAs could not be detected by RNA FISH, even under conditions where these mRNAs are significantly accumulated. Taking this into consideration, our results showing that poly(A)+ RNA is accumulated in the nucleolus even when inducing transcription inhibition before exposing parasites to heat shock suggest that the signals observed correspond to processed mRNAs. We also showed that these mRNAs were present before subjecting parasites to heat shock. It has been previously reported that mRNA export in yeast is also affected by heat shock. In this regard, we could not initially discard, the notion that the mRNA nucleo-cytoplasmic transport might also be affected, thus resulting in the observed mRNA nucleolar accumulation. However, previously reported data from experiments performed with *T. cruzi* showed that mRNA export inhibition in response either to Leptomycin B or sodium arsenite led to a nuclear, but no nucleolar, accumulation of mRNAs. This evidence favours the argument that the observed nucleolar accumulation of mRNAs would not be a “secondary effect” generated as a result of an mRNA export arrest. Interestingly, the Hsp70 mRNA bypassed the nucleolar accumulation in heat- shocked parasites, suggesting that retention of mRNAs into the nucleolus might occur in a selective fashion. This observation is in agreement with previous works performed in yeast exposed to severe heat shock, showing that the bulk population of poly(A)+ RNA was accumulated into the nucleolus, ; but, the SSA4 mRNA (which codifies for Hsp70) was distributed throughout the cytoplasm under the same condition. As proposed by Saavedra et al., this difference in localization is likely based on the presence of multiple signals at the sequence and/or structure level in those mRNAs that avoid such retention leading to their export, even when cells are subjected to severe heat shock conditions. We next aimed to determine whether the RBPs that are mobilized to the nucleolus during stress conditions could directly participate in the nucleolar accumulation of the poly(A)+ RNA population. Since *T. cruzi* does not have a full RNAi machinery, we attempted an antisense knockdown strategy to inhibit the expression of different RBPs. This strategy did not work, and we were thus not able to dissect the role of the RBPs in the nucleolar accumulation of poly(A)+ RNA. We have previously reported that the mechanism driving the nucleolar relocalization of RBPs, induced by transcription inhibition, is not functional in *T. brucei*. Here, we extended these findings by showing that the nucleolar accumulation of poly(A)+ RNA is also absent in this parasite. So, both results strongly indicate that, unlike that observed in *T. cruzi*, the *T. brucei* nucleolus does not play a role in mRNA metabolism during stress conditions, and also suggest that this role has been differently conserved in the trypanosomatid lineage. Strikingly, it seems that trypanosomatids have developed different responses to deal with the same stress condition. For instance, confronted to severe heat shock, *T. brucei* responds with the formation of cytoplasmic SG , while *T. cruzi* accumulates poly(A)+ and certain RBPs in its nucleolus, without apparent SG formation. It should be highlighted that, to our knowledge, the nucleolar accumulation of mRNAs in response to severe heat shock has been observed only in yeast and *T. cruzi* (this work). Taking this information into account, an interesting question is why mRNAs need to be stored in the nucleolus in response to heat shock in *T. cruzi*. It has been previously shown that cytoplasmic poly(A)+ RNA granule formation induced by starvation persisted after ActD treatment in trypanosomes, thus indicating that the mRNA is stabilized in such structures. On the other hand, it has been reported in mammals that ActD treatment leads to the loss of P bodies favouring the argument that transcripts in those structures are degraded if there is not enough supply of mRNAs to support them. Based on these observations, if the *T. cruzi* nucleolus were a focus of mRNA protection, then we would expect that the nucleolar mRNA in heat- shocked parasites might remain even after treating with ActD. Preliminary experiments showed that the nucleolar poly(A)+ RNA signal still persisted under such conditions. Moreover, we have previously demonstrated that long-term incubation with ActD results in cytoplasmic poly(A)+ RNA decay, but also nucleolar mRNA accumulation, even after 24 h of ActD treatment. Together, these results suggest a potential role of the nucleolus as a protective structure for mRNAs. Together, our results raise the interesting possibility that the *T. cruzi* nucleolus plays an important role in the response to environmental stress conditions, such as severe heat shock (outlined in). Under this condition, the nucleolus may behave as a storage/protection structure for factors involved in gene expression, such as RBPs and most mRNAs, until favourable conditions are restored. However, factors (for instance Hsp70) and mRNAs codifying proteins required to overcome the heat shock stress bypass the nucleolar retention. These results also support the idea that the additional functions of the nucleolus might be already present in ancestral organisms such as trypanosomes, arguing in favour of the fact that this potential alternative way of post- transcriptional gene regulation could have been acquired early in the evolution of the eukaryotic cell. # Materials and Methods ## Parasites *T. cruzi* epimastigotes (CL strain) were cultured in BHT 10% medium (brain heart infusion, 0.3% tryptose, 0.002% bovine hemin) supplemented with 10% heat- inactivated fetal calf serum, streptomycin 0.1 mg/ml and penicillin 100 U/ml at 28°C. *T. brucei* procyclic (Lister 427 strain) form cells were cultured in SDM79 medium. Parasite cultures were taken in a late logarithmic growth phase at a cell density of 2.5−3.5×10⁷/ml parasites for *T. cruzi* and of 0.5×10⁷/ml parasites for *T. brucei*. ## Treatments For the heat shock experiments, parasites were incubated in a water bath at 40°C. Transcription inhibition was induced incubating parasites with ActD 50 µg/ml. ## RNA Preparation Total RNA was isolated using TRIzol reagent (Invitrogen) and residual genomic DNA was removed by DNase treatment using RNase-free DNase I (Ambion). Both procedures were performed according to the manufacturer’s instructions. ## Immunofluorescence Trypanosomes were centrifuged from log phase cultures for 2 minutes at 2000 g, washed in PBS twice, allowed to settle on poly-L-lysine-coated slides and fixed in paraformaldehyde (PFA) 4% in PBS at room temperature (RT) for 10 minutes. After two brief washes in PBS at RT, fixed cells were incubated at RT with 25 mM NH₄Cl in PBS for 10 minutes. Cells were washed twice with PBS, permeabilized and blocked with 0.5% saponin, 1% Bovine serum albumin, 2% goat normal serum in PBS for 60 minutes at RT. After blocking, cells were first incubated with the primary antibody (diluted in 0.1% saponin and 1% BSA in PBS) for 60 minutes and then washed 3 times with PBS. Afterwards, slides were incubated with secondary antibodies (diluted in 0.1% saponin and 1% BSA in PBS) for 60 minutes, washed three times with PBS and once in Milli-Q water. The primary antibody was polyclonal anti-TcPABP2 (1∶1000). The secondary goat anti- rabbit antibody AlexaFluor 488 (Molecular Probes) was used at 1∶1000 dilutions. ## Fluorescence *in situ* Hybridization (FISH) For detection of total poly(A)+ RNA by FISH, parasites were harvested, allowed to adhere to poly-L-lysine-coated microscope slides, fixed with 4% paraformaldehyde in PBS at RT for 10 minutes, followed by a 10-min incubation with 25 mM NH₄Cl. Fixed parasites were permeabilized and blocked for 1 h in 0.5% saponin (Sigma), 2% BSA (Blocking Buffer), followed by 2-h prehybridization at RT in 2% BSA, 5X Denhardt, 4X SSC, 5% dextran sulphate, 35% deionized formamide (Sigma), 0.5 µg/µl yeast tRNA (Sigma) and 10 U/ml RNasin (Promega) (Hybridization Solution). Hybridizations were performed overnight at 28°C in a humid chamber either in the presence of 1 ng/µl Cy3-conjugated oligo-(dT)30 or Cy3-conjugated oligo-(dA)30 in Hybridization Solution. Slides were washed twice in 4X SSC at RT. Slides were mounted in 1 µg/ml DAPI prepared in Fluorsave (Calbiochem). RNase A pretreatment was performed at 37°C for 30 minutes before hybridization. The following probes were also used: - Mini-exon: \*caatatagtacagaaactgtatcaataatagGgtt - α-Tub (Tc00.1047053411235.9): - ○ \*CGACGAGTTAAATCATAAATTGCTT - ○Random α-Tub: CCGATTGATGATCATAATAAATGTC - Smug (Tc00.1047053504539.30, Tc00.1047053504539.20, Tc00.1047053504539.10): - ○ \*CTCAAACACAGCAGCATCGT - ○ Random Smug: CATCCACATAAGGGTCCAAC - Hsp70 (Tc00.1047053511211.160, Tc00.1047053511211.170): - ○ \*CAATCTCCTTCATCTTTGACAGGAC - ○Random Hsp70: CATCGCCCAGTGTTATTTCCAATCA The asterisks indicate the position of Cy3 in each probe. # Supporting Information We thank Agustina Chidichimo, Berta Franke de Cazzulo and Liliana Sferco for parasite cultures. D.O.S. is a career investigator from Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) and E.N. is a fellow from Universidad Nacional de San Martín (UNSAM). [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: EN DOS. Performed the experiments: EN. Analyzed the data: EN REV DOS. Contributed reagents/materials/analysis tools: REV DOS. Wrote the paper: EN DOS.

# 1. Introduction We previously used serologic analysis of a recombinant cDNA expression library (SEREX) in U937 acute monocytic leukemia cells to identify leukemia-associated antigens. Using this approach, we identified a novel leukemia-associated gene, *MLAA-34*, which is a novel splice variant of *CAB39L* that is associated with acute monocytic leukemia. At present, little is known about the function of the *CAB39L* gene. A previous study identified *MLAA-34* as a novel anti-apoptotic gene in U937 cells. Clinical research has shown that *MLAA-34* is highly expressed in primary and recurrent acute mononuclear cell leukemia (M5) patients, while it shows low expression in M5 patients with complete remission. Examination of *MLAA-34* in 30 kinds of tumor cell lines, 11 kinds of non-tumor cell lines and 5 kinds of hybrid cells revealed high expression of *MLAA-34* in U937 and MHCC97-H cells and peripheral blood mononuclear cells in patients with M5. The expression was relatively weak in other leukemia and lymphoma cell lines, as well as in some solid tumor cell lines; almost no expression was detected in non-tumor cell lines. *MLAA-34* gene was not expressed in the HeLa cervical cancer cell line. Further study found that the anti-apoptotic effect of *MLAA-34* in U937 cells may be related to activation of the Wnt/β-catenin signaling pathway. So far, research on the function of *MLAA-34* gene has been limited to the acute monocytic leukemia cell line U937. However, whether *MLAA-34* has an anti-apoptotic role in other tumor cells has not yet been reported. Arsenic trioxide (ATO) has been successfully used to treat acute promyelocytic leukemia and has also been researched as a possible treatment for other hematological and solid cancers. ATO has shown therapeutic efficacy in the treatment of cervical cancer and has been demonstrated to effectively induce apoptosis of cervical cancer cells at low concentrations *in vitro*. In this study, we explored the effect of *MLAA-34* expression on apoptosis and the cell cycle of HeLa cells treated with ATO and investigated whether the mechanism of action is related to the Wnt/β-catenin signaling pathway. # 2. Materials and methods ## 2.1. Cell culture HeLa cells were obtained from the Institute for Cancer Research, the School of Life Science and Technology, Xi’an Jiaotong University (China). Cells were cultured in DMEM medium with 10% fetal calf serum. In all experiments, cells were used in the log-growth phase. ATO was diluted in medium and the concentration of ATO in experiments was 1 μmol/L. ## 2.2. Construction of the *MLAA-34* recombinant lentiviral vector and virus production The recombinant vector containing *MLAA-34* cDNA was constructed using the lentiviral expression vector pGC-LV (GeneChem Limited Company. Shanghai, China); this vector was named pGC-FU-MLAA-34. The vector that does not contain the *MLAA-34* cDNA was termed as pGC-FU. The construction of the recombinant lentiviral vector, virus packaging and virus titer determination method were performed as previously described. ## 2.3. Establishment of *MLAA-34* overexpressing stable cell lines HeLa cells were cultured in a 6-well tray in DMEM supplemented with 10% FBS. The recombinant *MLAA-34* lentivirus was added to HeLa cells at multiplicity of infections (MOIs) of 30, 50 and 100 with ENi.S and 5 μg/ml polybrene when the fusion degree is about 70%. The expression of green fluorescence was observed by fluorescence microscope after 72 h, and the optimal infection rate was determined. The infection rate was determined by flow cytometry. Recombinant *MLAA-34* lentivirus was used to infect HeLa cells at the optimal MOI, and the culture medium were replaced with fresh medium after 24 h. Stably infected clones were established by selection with G418 (Sigma-Aldrich, St. Louis, MO) at a concentration of 600 μg/ml for 3 weeks. Resistant clones were classified and expanded by culture; the expression of GFP was confirmed by fluorescence microscopy to determine the efficiency of infection. The expression level of *MLAA-34* gene and protein was detected by RT-PCR and western blot assay. After 5 weeks of extended culture post-G418 selection, we obtained stable highly expressing *MLAA-34* HeLa cell lines. ## 2.4. RNA analysis Total RNA was isolated from cells using an acid guanidinium-phenol-chloroform method (Trizol, Invitrogen). RT-PCR was carried out using the HighFidelity PrimeScriptTM RT-PCR Kit (Takara, DaLian, China). Primers for analysis were designed using Primer Premier Software (PREMIER Biosoft International, Palo Alto, USA) based on known sequences. Primer sequences used for amplification are listed in. Cycling parameters for the reactions were as follows: 95°C for 3 min, 30 cycles of 98°C for 10s, 55°C for 30 s, 72°C for 1 min, and a final extension at 72°C for 10 min. PCR products were analyzed by 2% agarose gel electrophoresis. Quantity One 4.6.2 software was used to analyze the gray value of electrophoresis results. ## 2.5. Western blot analysis Cells were collected in 80 μl lysis buffer containing protease inhibitors (Beyotime, China). Nuclear proteins were extracted using the CelLytic™ NuCLEAR™ Extraction Kit (Sigma, USA). Protein concentration was determined using the BCA kit (Sigma, USA). *MLAA-34* is a novel splice variant of the *CAB39L* gene, but with the same coding frame, encoding the same protein, therefore the CAB39L monoclonal antibody (Santa Cruz Biotechnology, sc-100390), which can recognize MLAA-34 protein, served as MLAA-34 antibody. Other antibodies were used as follows: anti-β-actin (Sigma-Aldrich), anti-β-catenin (Santa Cruz Biotechnology), anti-c-Myc (Abcam, UK), anti-cyclin B1 (Santa Cruz Biotechnology), anti-cyclin D1 (Santa Cruz Biotechnology), anti-Histone H3 (Sigma-Aldrich), HRP-labeled Goat Anti-Mouse IgG (Beyotime, China), and HRP- labeled Goat Anti-Rabbit IgG (Beyotime, China). ## 2.6. Cell viability assay Cells (3×10³) were plated on a 96-well plate and allowed to adhere overnight. The treatment group was treated with ATO to a final concentration of 1 μmol/L, and then cultured for 1, 2, 3, 5 and 7 d. MTT assays were then carried out. Absorbance was measured at 540 nm. All assays were performed in triplicate and were repeated three times under independent conditions. Data are presented as means±SD. Cell viability was calculated as % of control cells, and the viability of HeLa cells (control group) was considered 100%. ## 2.7. Flat plate clone formation assay Cells in logarithmic growth phase were digested into a single cell suspension with trypsin-EDTA solution and then 5 ml of the suspension was seeded into 25 ml cell culture flasks at a density of 40 cells/ml. The treatment group was treated with ATO at a final concentration of 1 μmol/L, and the culture medium was replaced with fresh medium after 24 h. Cells were cultured for about 3 weeks and colony formation was monitored under a magnification of 100 x. The assay was done in triplicate. ## 2.8. Analysis of cell cycle and apoptosis Cells were seeded into a six-well tissue culture plate. The treatment group was treated with ATO at final concentration of 1 μmol/L. After 48 h, cells were harvested. Cell cycle analysis was performed as previously described. Annexin V (Ann-V) and propidium iodide (PI) staining were performed using the Annexin V-FITC Apoptosis Detection Kit (BD Biosciences), followed by flow cytometric analysis. All assays were performed in triplicate and were repeated three times under independent conditions. Relative G₂/M cell growth rate = (G₂/M cells in ATO treatment group—G₂/M cells in control group)/G₂/M cells in control group×100%. ## 2.9. Cignal TCF/LEF reporter assay The Cignal TCF/LEF reporter assay (Qiagen, Dusseldorf, Germany) was performed according to the manufacturer’s instructions. After preparation of complex formation according to the instructions, suspended cells were obtained from exponential phase of growth cells (4×10⁵/ml, 100 μl/well). The 96-well plates were cultured at 37°C with 5% CO₂ and16 hours later, the medium was replaced with assay medium (Opti-MEM® containing 0.5% fetal bovine serum, 1% NEAA, 100 U/ml penicillin and 100 μg/ml streptomycin).The next day, ATO was added and the cells were cultured for another 48 h. Luciferase assays were then carried out using the Dual-Luciferase Reporter Assay System (Promega, WI, USA). The TCF/LEF reporter activity is presented as the relative ratio of Firefly luciferase activity to Renilla luciferase activity. All experiments were performed three times with triplicate replicates. ## 2.10. Statistical analysis Statistical significance was assessed by comparing mean (±SD) values with Student’s t-test for independent groups. *P* \< 0.05 was considered statistically significant. All analyses were performed using SPSS software. # 3. Results ## 3.1. Establishment of stable HeLa cell lines with *MLAA-34* overexpression To examine the function of *MLAA-34* in HeLa cells, we generated HeLa cell lines stably overexpressing *MLAA-34* using our previously published lentivirus vector (pGC-FU-MLAA-34). The positive clones were selected by G418 for 3 weeks, and then after 5 weeks of extended culture, we confirmed highly expressing *MLAA-34* HeLa cell lines. As shown in (– Figs), the *MLAA-34* stable HeLa cell lines showed increased *MLAA-34* mRNA and protein levels as compared to control and empty vector-transfected cells, confirming that the exogenous *MLAA-34* gene was stably expressed in HeLa cells. ## 3.2. *MLAA-34* overexpressing stable HeLa cells showed increase cell viability under ATO treatment We next examined the cell viability in response to ATO treatment using MTT assays. In the absence of ATO treatment, the cell viability of the *MLAA-34* stable cell lines compared with the control cell lines was not statistically different, indicating that overexpression of *MLAA-34* had no impact on HeLa cell viability. Upon treatment with ATO (1 μmol/L), the cell viability markedly decreased in all groups over time. Notably, the cell viability of *MLAA-34* overexpressing stable cells treated with ATO was significantly higher than that of the control cells treated with ATO. This showed that overexpression of *MLAA-34* could block, in part, the inhibitory effects of ATO on cell viability in HeLa cells. ## 3.3. *MLAA-34* overexpressing stable HeLa cells showed increased colony formation ability under ATO treatment We next examined colony formation ability of the HeLa cell lines with ATO treatment. In the untreated cell groups, the colony formation ability of control cells and *MLAA-34* stable cells was 98.5%±4.2% and 103.4%±5.3% respectively, with no statistical significance between the two groups. Upon treatment with ATO (1 μmol/L), the colony formation ability of all cell lines markedly decreased. The colony formation abilities of both ATO-treated control groups were 62.5%±3.9% and 66.7%±2.8%. However, the colony formation ability of *MLAA-34* stable cells treated with ATO was 84.3%±3.3%, and this was significantly higher compared with the ATO treated control group. These findings showed that *MLAA-34* overexpression blocked, in part, the inhibition of colony formation ability by ATO treatment. ## 3.4. Effects of *MLAA-34* overexpression on cell cycle distribution and apoptosis of HeLa cells under ATO treatment We next examined the apoptosis rate and cell cycle distribution of the HeLa cell lines. The *MLAA-34* stable cells showed no significant difference in apoptosis or cell cycle distribution compared to control HeLa cells, suggesting that overexpression of the *MLAA-34* gene had no significant effects on apoptosis and cell proliferation. Upon treatment with ATO, the apoptosis rate of the control group was 21.8%, while the apoptosis rate of the *MLAA-34* overexpression stable cell line was 10.8%, indicating that *MLAA-34* overexpression significantly reduced the apoptosis induced by ATO. In line with this observation, the relative G₂/M cell growth rate of the control group treated with ATO was 194.9%, while that of the *MLAA-34* overexpression stable cell line treated with ATO was only 56.5%, indicating that *MLAA-34* overexpression significantly reduced the increased numbers of cells in G₂/M phase induced by ATO. These results show that upon induction of apoptosis by ATO in HeLa cells, *MLAA-34* overexpression had an anti-apoptotic effect and reduced the G₂/M phase arrest. ## 3.5. Effects of ATO on the Wnt/β-catenin signaling pathway in U937 cells Because the Wnt/β-catenin signaling contributes to the development of U937 cells, and U937 cells could be induced apoptosis by ATO, we investigated the Wnt/β-catenin signaling pathway in U937 cells following exposure to ATO. U937 Cell lines were treated with ATO (0, 1, 2, 4 μmol/L) and then cultured for 48 h, the Cignal LCF/TCF reporter assay was used to evaluate the activity of Wnt/β-catenin signaling. As shown in, Different concentrations of ATO could significantly inhibit the transcriptional activity of the TCF/LEF reporter vector compared with untreated U937 cells (*P*\<0.05), and in a dose-dependent manner. Moreover, western blot results showed that the expression of nuclear β-catenin in ATO-treated cells was significantly decreased compared with control U937 cells (and Figs). We further assessed the expression of MLAA-34 protein, as well as the expression of proteins of Wnt signaling pathway downstream gene c-Myc, cyclin B1 and cyclin D1. As shown in (– Figs), ATO decreased the protein expression of MLAA-34, c-Myc, cyclin B1 and cyclin D1 in a dose-dependent manner (P\<0.05). These results indicate that ATO reduces the expression of MLAA-34 and inhibits the activity of the Wnt/β-catenin signaling pathway in U937 cells. ## 3.6. Effects of *MLAA-34* overexpression on the Wnt/β-catenin signaling pathway in Hela cells We next investigated whether the anti-apoptotic effect of the *MLAA-34* gene was related to the Wnt/β-catenin signaling pathway in HeLa cells. Cell lines were treated with ATO (1 μmol/L) and then cultured for 48 h. We assessed the expression of *β-catenin* (the Wnt signaling pathway key gene) mRNA and protein, as well as the expression of proteins of Wnt signaling pathway downstream genes c-Myc, cyclin B1 and cyclin D1. As shown in (– Figs), in the absence of ATO treatment, *MLAA-34* overexpression had no impact on the mRNA and protein expression of *β-catenin* or the protein expression of c-Myc, cyclin B1 and cyclin D1. These results were consistent with the fact that *MLAA-34* overexpression did not affect the growth curve, apoptosis rate and cell cycle of HeLa cells without ATO treatment. Upon treatment with ATO, the mRNA and protein expression of *β-catenin* and c-Myc, cyclin B1 and cyclin D1 protein expressions were significantly decreased compared with untreated cells, and the difference was statistically significant (*P* \< 0.05). This result suggests that the induction of apoptosis by ATO in HeLa cells may be partly achieved by inhibiting the Wnt/β-catenin signaling pathway. Furthermore, in the *MLAA-34* stable cells treated with ATO, the mRNA and protein expression of *β-catenin* and c-Myc, cyclin B1 and cyclin D1 protein expressions were significantly higher than those of the ATO-treated control group, and the difference was statistically significant (*P* \< 0.05); however, these levels were still lower than those of untreated HeLa cells. These results suggest the inhibitory effect of MLAA-34 on ATO induced HeLa cell apoptosis may be partially achieved through its blocking of ATO-mediated disruption of the Wnt/β-catenin signaling pathway. To validate the involvement of the Wnt/β-catenin signaling cascade, the Cignal LCF/TCF reporter assay was used to evaluate the activity of Wnt/β-catenin signaling. As shown in, ATO (1 μmol/L) treatment could significantly inhibit the transcriptional activity of the TCF/LEF reporter vector compared with untreated HeLa cells (0.89±0.25, *P* \< 0.05). ATO treatment in *MLAA-34* overexpressing stable cells significantly increased the TCF/LEF reporter activity compared with control HeLa cells treated with ATO (2.01±0.19, *P* \< 0.05). Furthermore, western blot results showed that the expression of nuclear β-catenin in ATO- treated cells was significantly decreased compared with untreated HeLa cells. In contrast, the expression of nuclear β-catenin in *MLAA-34* overexpression cells was significantly increased compared with control HeLa cells treated with ATO (and Figs). These results suggest that *MLAA-34* overexpression increases the nuclear localization of β-catenin protein in HeLa cells treated with ATO. # 4. Discussion In our previous studies, we found that *MLAA-34* gene (GenBank no: AY288977.2) is a new anti-apoptotic gene related to acute monocytic leukemia and is a novel splice variant of *CAB39L*. At present, little is known about the function of *CAB39L* gene. Lo et al. found significantly higher expression of mRNA and protein of *CAB39L* in the tumor tissues of 37 patients with oral cancer compared with surrounding normal tissues. A genome-wide association analysis and gene expression analysis found a considerable association between *CAB39L* and coffee addiction, and indicated that the protein encoded by *CAB39* may be involved in the metabolism of caffeine. Rahmioglu et al. reported that *CAB39L* is a susceptibility gene for endometriosis and obesity. These studies suggest that the function of *CAB39L* gene may be related to the metabolism of the substance, and it may be involved in the occurrence and development of tumors. Our previous study found that *MLAA-34* exhibits anti-apoptotic effects in U937 cells and clinical research has shown high expression of *MLAA-34* in primary and recurrent M5 patients and low expression in M5 patients with complete remission. In our previous study, we generated U937 cells stably expressing *MLAA-34* and found that *MLAA-34* overexpression can significantly reduce the spontaneous apoptosis of U937 cells, increase the percentage of G₂/M phase cells, and promote cell proliferation. Co-immunoprecipitation, shotgun and bioinformatic analysis showed that the Wnt/β-catenin signaling pathway may be involved in the anti-apoptotic effect of *MLAA-34* in U937 cells. The present study found that when ATO induced apoptosis of U937 cells, the Wnt/β-catenin signaling pathway was inhibited and the expression of MLAA-34 protein was downregulated, suggesting that *MLAA-34* is involved in regulating the Wnt/β-catenin signaling pathway. To further confirm the anti-apoptotic effect of the *MLAA-34* gene, we investigated its function in another tumor cell line of non-hematologic tumors that does not express the *MLAA-34* gene. Several studies have demonstrated that ATO significantly decreased HeLa cell survival and induced apoptosis and cell cycle arrest in G₂/M phase. HeLa cells are very sensitive to ATO, with 1 μmol/L treatment producing apoptosis. This dose has been very close to ATO in patients with acute myeloid leukemia in the treatment of dose, so as to induce HeLa cells apoptosis, the treatment concentration of ATO was set to 1μmol/L in our research. Although many studies have examined the mechanism of ATO inducing apoptosis in HeLa cells, the mechanism is still not clear and under discussion. The inhibition of the expression of the HPV oncogenic gene *E6* by ATO may be one of the mechanisms. In addition, *c-myc*, *Bc1-2* gene, a cell differentiation gene (*NDRG1*), intracellular ROS level, and the change of telomerase activity may be involved in the induction of apoptosis by ATO in HeLa cells. Our results showed that inhibition of the Wnt/β-catenin signaling pathway may be one of the contributory mechanisms to ATO-induced apoptosis in HeLa cells. Our study showed that overexpression of the *MLAA-34* gene had no effect on the growth, apoptosis and cell cycle of HeLa cells. Upon ATO treatment, the cell viability and colony formation ability of HeLa cells with *MLAA-34* overexpression were significantly higher than that of the control group, and the apoptosis rate and proportion of G₂/M cells decreased. These results showed that the *MLAA-34* gene not only has anti-apoptotic effects in the acute monocytic leukemia cell line U937 but also in the HeLa cervical cancer cell line. Our results showed that ATO treatment resulted in decreased expression of *β-catenin* mRNA and protein and the downstream target proteins c-Myc, cyclin B1, and cyclin D1 in HeLa cells. However, in *MLAA-34* stable cells, the ATO- mediated reduction of these factors was compromised, and the levels were significantly higher than that in the ATO-treated HeLa cells. Furthermore, we found that ATO-treated *MLAA-34* stable cells showed higher expression of nuclear *β-catenin* levels compared with HeLa cells treated with ATO. This suggested that the inhibitory effect of MLAA-34 protein on ATO-induced HeLa cell apoptosis may be partially achieved through activation of the Wnt/β-catenin signaling pathway. Previous overexpression and RNA interference experiments have suggested that *MLAA-34* gene plays an antiapoptotic role in U937 cells with high expression of *MLAA-34*, and may play a role by participating in the Wnt/β-catenin signaling pathway. Recently, we have successfully expressed and purified MLAA-34 protein and isolated a fully human ScFv antibody (MA1) against MLAA-34 from a large human ScFv library. MA1 can not only specifically bind with U937 cells, but also inhibit the proliferation of U937 cells. The results of this study showed that in the absence of *MLAA-34* expressing Hela cells, the expression of *MLAA-34* after exogenous vectors could also partially reduce ATO-induced apoptosis of HeLa cells, and the inhibitory effect of *MLAA-34* may be partially achieved through activation of the Wnt/β-catenin signaling pathway. In previous study, we found that the *MLAA-34* gene was highly expressed in U937 cells and expressed strongly in K562 and HL-60 cells, so we selected U937, K562 and HL-60 cells to do the knockdown experiments using RNA interference technique. The results showed that there was significant correlation between *MLAA-34* and the proliferation of K562 and HL-60 cells, the apoptosis rates of cells with siRNA infection were higher than that of control group. We found that MLAA-34-siRNA could induce the apoptosis of K562 and HL60 cells, it also significantly decreases the levels of β-catenin and TCF4, and so *MLAA-34* anti-apoptotic effect may be through the Wnt/β-catenin signaling pathway. It is still not clear how *MLAA-34* activates and functions in the Wnt/β-catenin signaling pathway and whether the *MLAA-34* gene affects the Ras signaling pathway simultaneously. # 5. Conclusions Our results show that the Wnt/β-catenin signaling pathway is related to ATO- induced apoptosis in HeLa cells. The *MLAA-34* gene reduced ATO-induced apoptosis and G₂/M arrest, and the anti-apoptotic effect may be achieved by activating the Wnt/β-catenin signaling pathway in HeLa cells. # Supporting information We express sincere thanks to Associate Prof. Yingli Xue for assistance with the manuscript. We thank Gabrielle White Wolf, PhD, from Edanz Group ([www.edanzediting.com/ac](http://www.edanzediting.com/ac)) for editing a draft of this manuscript. [^1]: The authors have declared that no competing interests exist.

# Introduction Intra-arterial radioembolization with yttrium-90 microspheres (⁹⁰Y-RE) is an increasingly applied treatment option for patients with unresectable primary or secondary hepatic malignancies, refractory to systemic therapies. The treatment consists of intra-arterial administration of microspheres tagged with or containing yttrium-90 (⁹⁰Y), a radioisotope that emits high-energy beta radiation. In contrast to the normal liver parenchyma, which mainly relies on the portal vein, intrahepatic malignancies mainly depend on the hepatic artery for their blood supply. As a consequence, these tumors can be selectively targeted by instillation of ⁹⁰Y-microspheres in the hepatic artery. There is growing evidence for an overall beneficial effect of ⁹⁰Y-RE regarding time to progression, overall survival and quality of life in salvage patients with either primary or metastatic hepatic malignancies.– The effect of ⁹⁰Y-RE in terms of tumor response varies widely, with disease control rates (complete response+partial response+stable disease) ranging from 56% –100%. Given the wide variety in tumor response rates, great effort is put into optimal patient selection through the identification of prognostic factors for a favorable outcome after ⁹⁰Y-RE.– Improved selection may increase the efficacy of this therapy and prevent patients from futile treatment and unnecessary toxicity. Although minimally invasive, ⁹⁰Y-RE is not without adverse effects. Common adverse effects related to ⁹⁰Y-RE are symptoms of the post- embolization syndrome, comprising fatigue, nausea, vomiting, abdominal pain, loss of appetite and fever.– In general, these symptoms appear on the day of treatment and last up to three days after treatment. More serious complications can occur when an excessive radiation dose is applied to non-target tissue. An excessive dose to the healthy liver parenchyma, which can be due to either a high overall administered activity or an unfavorable tumor to non-tumor activity distribution ratio, can cause radiation induced liver disease (RILD). Alternatively, distribution of microspheres in organs other than the liver could cause serious morbidity and even mortality (e.g. radiation pneumonitis or gastric ulceration). These severe complications occur in less than 10% of patients.–. Laboratory toxicity in terms of elevated liver function tests and liver enzymes can be expected after ⁹⁰Y-RE. It is important to monitor laboratory toxicity, because this may be an early indicator for RILD. Relatively little is known, however, about the normal range of laboratory toxicities following ⁹⁰Y-RE in patients who do not develop RILD. The primary objective of this study was to investigate clinical and laboratory toxicity in patients with liver metastases, treated with ⁹⁰Y-RE. Secondary objectives were assessment of tumor response and overall survival. # Materials and Methods ## Patient Selection Records of all liver metastases patients who were not participating in a clinical trial and had received a pre-treatment angiographic procedure for treatment with ⁹⁰Y-RE at our institute between February 1^st 2009 and March 31^st 2012 were retrospectively analyzed. Patients that were eligible for ⁹⁰Y-RE had unresectable liver dominant metastases and had progressive disease under systemic treatment, or were no longer treated systemically due to contraindications. The Medical Ethics Committee of the University Medical Center Utrecht waived the need for informed-consent and approved this study. ## Procedure ⁹⁰Y-RE was carried out over two sessions: a pre-treatment diagnostic angiography and a treatment angiography. Patients were admitted to the hospital on the evening before angiography. They received 1.5 L per 24 h NaCl 0.9% intravenously for pre- and post-hydration. Pre-treatment diagnostic angiography started with selective visceral catheterization (celiac axis and superior mesenteric artery) in order to obtain an angiographic map of the patients’ vascular anatomy. Specific extrahepatic vessels were coil-embolized to prevent ⁹⁰Y-microspheres that were injected into one of the hepatic arteries, to be distributed to visceral organs other than the liver. Arteries that were actively searched for and embolized using coils included the gastroduodenal artery, the right gastric artery, and pancreaticoduodenal vessels and any other relevant arteries depending on the patient’s specific anatomy. Subsequently, 150 MBq technetium-99m-labelled macro-albumin aggregates (^99mTc-MAA) were injected into the hepatic artery to simulate the ⁹⁰Y-microspheres distribution. Next, single photon emission computed tomography (SPECT) and planar nuclear imaging were performed. In order to assess whether part of the dose was deposited in abdominal organs other than the liver, the SPECT images were analyzed after fusion with computed tomography (CT). Planar nuclear imaging was used to calculate the lung shunt fraction; patients with a lung shunt \<10% received the full dose of ⁹⁰Y-microspheres, when lung shunt fraction was between 10%–15% or 15%–20% the dose of ⁹⁰Y-microspheres was reduced with 20% and 40%, respectively. Lung shunt fractions of \>20% implied that no treatment could be given. If radioactivity was detected in non-target organs, such as pancreas, duodenum or stomach, further angiographic investigation was performed with additional coiling and/or a more distal injection position of ^99mTc-MAA. Patients stayed one night in the hospital for observation. Treatment angiography was performed within two weeks after the pre-treatment angiography. Patients were readmitted to hospital the day before angiography, where they again received pre- and post-hydration. One hour before angiography, patients received a single intravenous dose of dexamethason (10 mg) and ondansetron (8 mg). The dose of radioactive resin microspheres (SIR-Spheres®, SIRTeX, Lane Cove, Australia) for each individual patient was calculated according to the body surface area method provided by the manufacturer. The tumor volume and total liver volume were calculated by volumetric assessment of CT imaging. Subsequently, the dose of ⁹⁰Y-microspheres was administered with the catheter tip in the hepatic artery or one of its branches, at the same position as used for the injection of ^99mTc-MAA. The total liver weight (*m_liver*) was derived from CT-volumetric measurements assuming a density of 1 kg/l. The net amount of administered radioactivity (*A_net*) (prepared activity minus residual activity in administration system and catheter) was calculated. The whole liver absorbed dose (*D_liver*), assuming a homogeneous distribution and full absorption of activity in the liver, was then estimated using the following Medical Internal Radiation Dose (MIRD) committee-based formula : Patients received ⁹⁰Y-RE as a whole liver treatment in a single angiographic procedure (i.e. whole liver delivery), whole liver treatment in two sessions (i.e. sequential delivery) or as treatment of a single lobe (i.e. lobar treatment). In cases of sequential delivery, the aim was to perform both treatment sessions within a commonly accepted interval of 30–45 days. The distribution of ⁹⁰Y-microspheres was assessed with either bremsstrahlung SPECT or ⁹⁰Y-positron emission tomography computed tomography (PET-CT). Our institution’s radiation safety committee required all patients to stay in the hospital for a minimum of 12 hours after treatment. ## Toxicity Assessment Post-treatment, patients reported to the outpatient clinics at intervals of approximately four weeks. At these visits, physical examination and laboratory tests were performed. The following laboratory investigations were included in our analysis in order to assess laboratory toxicity: total bilirubin, alkaline phosphatase (ALP), gamma-glutamyl transferase (GGT), aspartate aminotransferase (AST), alanine aminotransferase (ALT), albumin, hemoglobin (Hb) and white blood cell count (leukocytes). Blood samples, taken up to four weeks prior to ⁹⁰Y-administration and during a four months follow-up were used for toxicity analysis. Laboratory toxicity was graded according to the Common Terminology Criteria for Adverse Events (CTCAE) v4.0. GGT, AST, ALT and Hb reference values were gender dependent. For each patient, baseline CTCAE grades and maximal CTCAE grades during follow-up were determined. In addition, new toxicity or progression of baseline toxicity to a higher CTCAE grade was grouped separately and will be referred to as “new toxicity“. Patients, in whom data on baseline and/or follow-up laboratory investigations were not available in our center, were excluded from the laboratory toxicity assessment. The clinical toxicity assessment was based on the reporting of periprocedural complications, treatment-related symptoms (CTCAE grade 1–2) and serious adverse events (CTCAE grade 3–4), in the patient’s charts. ## Response Assessment Baseline imaging was performed with CT or magnetic resonance imaging (MRI) of the liver. In addition, patients with (suspected) ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG)-avid tumors received ¹⁸F-FDG-PET to assess the presence of extrahepatic metastases. Follow-up imaging was performed with CT or MRI of the liver (depending on the modality used for baseline imaging) at approximately 1, 3 and 6 months post- treatment. Response assessment was performed in accordance with the Response Evaluation Criteria in Solid Tumors (RECIST 1.1) on the level of target lesions (TL), whole liver (including non-target lesions) and overall response (including non-target lesions and extrahepatic disease) at 3 months (range 2.0–4.5 months) and 6 months (range 4.5–7.5 months) after the first ⁹⁰Y-RE procedure. Up to five target lesions per patient were identified by an observer (either MS or CR) and the maximal cross-sectional diameter of each target lesion was subsequently measured by the other observer. Observers were blinded for the identity and characteristics of the patient; date of imaging and whether it was a baseline or follow-up scan. Data on progression of non-target lesions, new liver lesions and progression of extrahepatic disease were extracted from radiologic reports. Patients who were lost to follow-up were regarded as having progressive disease (PD) on the ‘overall response level’ at the time of death. Median time to progression (TTP) was calculated for all response levels per Kaplan-Meier analysis. ## Survival Analysis Overall survival was defined as the interval between the date of (first) ⁹⁰Y-RE treatment and the date of death or most recent contact (alive). Median overall survival (including corresponding 95% CI) was calculated through Kaplan-Meier survival-analysis. Statistical analyses were performed with SPSS Statistics 20.0 for windows (IBM SPSS, Chicago, IL). All percentages were rounded to the nearest whole number. # Results ## Patients Between February 1^st 2009 and March 31^st 2012, a total of 73 consecutive patients (excluding patients participating in a prospective clinical trial) with liver metastases were considered eligible for ⁹⁰Y-RE treatment at our institute and received a pre-treatment angiographic procedure with ^99mTc MAA. A flowchart of the study design and patient treatment is presented in. Fourteen patients (19%) could not be treated with ⁹⁰Y-RE, due to persistent extrahepatic deposition (PED) of ^99mTc-MAA (n = 11), rapidly progressive disease (n = 2) and a lung shunt fraction exceeding twenty percent (26%, n = 1). Fifty-nine patients received ⁹⁰Y-RE treatment. Baseline characteristics of these patients are presented in. The majority of the patients (30/59, 51%) had colorectal cancer liver metastases, six patients (10%) had neuroendocrine tumor (NET) liver metastases, and 23 patients (39%) suffered from liver metastases from various other primary tumors. Treatment details are presented in. The majority of the patients received a whole liver treatment in one session (n = 38, 64%), with a selective administration of ⁹⁰Y-microspheres in the left and right hepatic artery (n = 28) or administration in the proper (n = 9) or common hepatic artery (n = 1). In ten patients, whole liver treatment was performed selectively in sequential sessions (n = 10, 17%), with a median interval of 14 days (range 12–77 days) between both treatment sessions. Eleven patients received unilobar treatment (n = 11, 19%). The mean net administered activity was 1473 MBq (standard deviation 447) with an estimated mean liver-absorbed dose of 42.0 Gy (standard deviation 14.3). Post-treatment bremsstrahlung scintigraphy or ⁹⁰Y-PET, revealed no extrahepatic deposition of radioactivity in any of the patients. Four patients were retreated with ⁹⁰Y-RE after disease progression had occurred, with a median interval of 9 months (range 5–25 months) between the first and second treatment. Median time of hospital admission was 2 days (range 1–4 days). Fifty-four patients (92%) were discharged the day after treatment. The other five patients required longer hospitalization (one or two days extra), due to comorbidities such as renal insufficiency, diabetes mellitus or heart failure. ## Toxicity Eleven patients (19%) were excluded from laboratory toxicity analysis, because data on laboratory investigations at baseline or during follow-up, within our defined intervals, were not available in our center. In the remaining 48 patients, there were values missing for some laboratory parameters, therefore the denominator was adjusted accordingly when calculating incidences. CTCAE grades at baseline, maximum CTCAE grades during follow-up and corresponding new toxicity are presented in. Grade 3–4 toxicity at baseline was observed for GGT (16/47, 34%) and ALP (1/47, 1%). Grade 3–4 new toxicity was observed in 18 patients (38%), including following parameters: GGT (13/47, 27%), ALP (10/27, 21%), bilirubin (1/41, 2%), AST (1/47, 2%), ALT (1/47, 2%), and albumin (1/42, 2%). In addition, the incidence of grade 3–4 new toxicity was stratified according to treatment strategy. Ten out of 28 evaluable patients (36%) who received whole liver treatment in one session had grade 3–4 new toxicity, compared to five out of ten patients (50%) who received whole liver treatment in sequential sessions, and three out of ten patients (30%) who received unilobar treatment. The following periprocedural complications were reported: allergic reaction to contrast agent (n = 6), arterial dissection (n = 2), nausea/vomitus during angiography (n = 1), delayed hemostasis at the access site requiring prolonged clamping (n = 1), inguinal hematoma at the access site (n = 1). Complications did not prevent any patients from receiving therapy. Back pain or abdominal pain during angiography was managed with fentanyl (37% of patients, range 50–200 mcg i.v.) and/or diclofenac (35% of patients, range 50–125 mg i.v.). Clinical symptoms associated with the postembolization syndrome (CTCAE grade 1–2) were observed in the majority of the treated patients. This syndrome comprised the following symptoms (in order of frequency): fatigue and loss of appetite, pain/discomfort in the right upper abdominal quadrant requiring analgesics (paracetamol and/or diclofenac and/or morphine), nausea and vomitus, fever and general discomfort. In general, these symptoms started on the day of treatment and lasted up to two weeks after treatment. No grade 3–4 clinical toxicity was observed after ⁹⁰Y-RE treatment and no serious treatment-related complications such as duodenal or gastric ulceration, radiation pneumonitis or RILD, were observed. ## Response Target lesions-, whole liver- and overall response rates and TTP (for all patients and per tumor type) at 3- and 6-months are displayed in. Target lesion, whole liver and overall disease control rates (complete response+partial response+stable disease) at 3-months post-treatment were 35%, 21% and 19% respectively. Corresponding disease control rates at 6-months were 25%, 13% and 12%. Median TTP for all patients was 6.2 months (95% CI 2.2–10.0) for target lesions, 3.3 months (95% CI 2.8–3.8) for the whole liver and 3.0 months (95% CI 2.4–3.5) overall. ## Survival At the time of analysis, 49 patients had died and 10 patients were still alive. Median overall survival for the entire group of patients (n = 59) was 8.9 months (95% CI 7.2–10.6). The Kaplan-Meier survival curve is displayed in. Median overall survival was 8.9 months (95% CI 6.9–10.9) for colorectal cancer liver metastases (n = 30), 40.3 months (0–107.9) for NET metastases (n = 6) and 7.8 months (95% CI 5.0–10.6) for other metastases (n = 23). # Discussion The primary objective of this study was to investigate treatment-related clinical and laboratory toxicity in patients with unresectable liver metastases, treated with ⁹⁰Y-RE. Secondary objectives were to assess tumor response and overall survival. Clinical toxicity was confined to grade 1–2 symptoms of the post-embolization syndrome. No RILD or other grade 3–4 clinical toxicity was observed, whereas laboratory toxicity grade 3–4 was observed in 38% of patients. In this cohort, a disease control rate of up to 35% was obtained at 3-months post-treatment, and median overall survival was 8.9 months. Tumor response rates vary widely in the ⁹⁰Y-RE literature. This may be explained in part by differences in methodology for response assessment. Various studies do not specify whether RECIST criteria have been followed. According to these criteria, tumor response should be differentiated in target lesion, liver and overall response. In order to improve interpretability of overall response rates, studies should indicate whether patients had evidence of extrahepatic disease at baseline. Response rates are commonly divided into 3- and 6-months rates post-treatment. However, it should be clearly stated which imaging intervals are chosen to represent this 3- and 6-months measurements. In addition, it would be preferable to score target lesion response blindly, to assure objective measurements. In a comprehensive review of the ⁹⁰Y-RE literature, twelve studies were identified that reported a 3-month disease control rate, ranging from 63–100%. In most of these studies, the level on which response assessment had been performed was not specified. Assuming these are whole-liver disease control rates, our 3-month disease control rate was much lower: 21%. This difference could be attributable to differences in methodology of response assessment, as mentioned above. However, less stringent patient selection criteria and the heterogeneity of our cohort, including hyper- and hypovascular liver metastases from various primary tumors, could also have attributed to lower response rates. Toxicity due to radiation to the liver has first been described after external radiation therapy., It was found that the liver is very sensitive to radiation and patients may develop radiation induced liver disease (RILD), months after an overdose of radiation. Histopathologically, RILD is characterized by veno- occlusive disease with congestion of the central veins and sinusoids.– The symptoms of RILD comprise fatigue, anicteric ascites, hepatomegaly, and elevated liver function tests (especially alkaline phosphatase). High dose corticosteroids can be given to mitigate the course of this disease. It is however, hard to recognize RILD since it has a long latency time and many of its symptoms can also occur after non-complicated treatment with ⁹⁰Y-RE. A better understanding of the physiological variation of treatment-related laboratory toxicity after ⁹⁰Y-RE would be very helpful in discriminating early signs of RILD from transient laboratory abnormalities after treatment. Mild toxicity (grade 1–2) of liver function tests is common after ⁹⁰Y-RE, occurring in up to 70% of the patients.– Reported incidences of grade 3–4 toxicity are much lower and vary widely across studies. Van Hazel *et al.* observed no grade 3–4 toxicity in their study, Piana *et al.* found an overall incidence of 7% and Kennedy *et al.* reported an incidence of up to 20.5% for ALP. In the study of Piana *et al.*, one patient died of RILD. In our study we found higher incidences of laboratory toxicity, with new laboratory toxicity grade 3–4 occurring in up to 38% of the patients. However, we did not observe any serious treatment-related complications, nor did we observe any RILD. This indicates that serious laboratory toxicity regarding transaminases and liver function tests can occur as part of the physiological reaction of the liver to ⁹⁰Y-RE treatment. One of the factors complicating the interpretation of toxicity results is that abnormalities in liver function tests and transaminases could be the result of tumor progression instead of treatment-related toxicity. Moreover, results of toxicity are often incompletely reported in the ⁹⁰Y-RE literature. Many studies do not specify how CTCAE scores for laboratory toxicity have been determined. This could inadvertently lead to an underestimation of treatment toxicity and it limits the comparability of studies. Therefore, we aimed to report our methods and results in an unambiguous and transparent fashion. The most important limitations of this study were its retrospective design and the lack of standardization of laboratory investigations and reporting of clinical symptoms during physical examination. Therefore, our results in terms of the incidence of laboratory or clinical toxicity are likely to be underestimations of the real incidence of toxicity. Another limitation was the heterogeneity of our study population. However, this heterogenic group does reflect the typical population of patients referred for ⁹⁰Y-RE treatment. Fourteen of the 73 patients (19%) who received work-up angiography did not receive ⁹⁰Y-RE. The majority of these patients (n = 11) were not eligible because of persisting extrahepatic deposition (PED) of ^99mTc-MAA. This PED rate of 11/73 (15%) is much higher than the rates reported in the literature (ranging from 0% to 10%)., A likely cause of the high PED rate in this study is the relative large number of proximal injection positions (i.e. proper or common hepatic artery). Several studies have demonstrated that extrahepatic deposition can be solved/prevented by more distal injection positions (left/middle/right hepatic artery or even more selective)., We have changed our current practice accordingly and we rarely perform whole liver treatments from the proper hepatic artery anymore. In addition, our center and many others increasingly use c-arm cone beam computed tomography during the pre-treatment angiography to help prevent extrahepatic distribution and identify culprit vessels.. The whole liver approach has also been associated with increased toxicity. Seidensticker *et al.* have reported that a whole liver approach, in non- cirrhotic liver metastases patients, resulted in a higher number of liver- related CTCAE grade 3–4 events as compared to a sequential lobar approach. We could not confirm this finding in our patients. In fact, the number of patients with CTCAE grade 3–4 laboratory toxicity was even lower in the whole liver approach group (36%) than in the sequential lobar group (50%). Selection bias, and confounding due to differences in baseline characteristics, may play a significant role in this matter. However, we do recognize the clinical importance, and we think that the question whether treating the whole liver at once increases toxicity, should be determined using a randomized controlled trial. The majority of the patients (70%) treated in our cohort, received radio- embolization as salvage therapy. This illustrates that ⁹⁰Y-RE is still regarded as a treatment option of last resort, for patients who have unresectable and chemorefractory liver tumors. The costs of radioembolization treatment (approximately **€**11.000 for one dose of SIR-spheres plus the costs of the procedure, the involved imaging, hospitalization and follow-up) need to be weighed against the potential benefit to the patient. For this purpose, prospective comparative studies evaluating survival, tumor response, and quality of life after ⁹⁰Y-RE are strongly warranted. In addition, it will become increasingly important to select those patients that will benefit most from this therapy. Performing radioembolization at an earlier stage in patients with liver metastases might for instance translate into improved tumor response rates and overall survival. Two large randomized controlled trials are currently ongoing, investigating the effect on overall survival (SIRFLOX study) and progression free survival (FOXFIRE study) of the addition of ⁹⁰Y-RE to FOLFOX (fluorouracil, leucovorin, oxaliplatin) with or without bevacizumab as first-line treatment for patients with unresectable colorectal liver metastases.. ## Conclusion The risk of severe complications or grade 3–4 clinical toxicity in patients with liver metastases of various primary tumors undergoing ⁹⁰Y-RE is low. In contrast, laboratory toxicity grade 3–4 was observed in more than one-third of the patients without any signs of RILD. This physiological reaction of liver enzymes to ⁹⁰Y-RE therapy may mask early signs of toxicity due to RILD. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MS AvdH CR MK MvdB. Performed the experiments: MS AvdH CR BZ ML JN MK MvdB. Analyzed the data: MS AvdH CR. Contributed reagents/materials/analysis tools: MS AvdH CR. Wrote the paper: MS AvdH CR BZ ML JN MK MvdB.

# Introduction Widespread uptake of safe and effective vaccines is critical to controlling the COVID-19 pandemic. Three COVID-19 vaccines have been authorized or approved by the U.S. Food and Drug Administration (FDA) for use in the United States, including the 2-dose Pfizer-BioNTech BNT162b2 mRNA and Moderna mRNA-1273 vaccines and the single-dose adenovirus-based Johnson & Johnson/Janssen Ad.26.COV2.S vaccine. In large, randomized controlled trials, the Pfizer- BioNTech and Moderna mRNA vaccines each had an efficacy of ≥94% in preventing symptomatic, laboratory-confirmed COVID-19 following the 2-dose series. Among the over 32,000 people who received either the Pfizer-BioNTech or Moderna vaccine during those clinical trials, 20 developed COVID-19 after vaccination. Among the 19,514 people randomized to receive Janssen vaccine during those trials, 116 developed COVID-19 following vaccination, resulting in an efficacy of 67% against moderate-to-severe COVID-19. The Pfizer-BioNTech and Moderna vaccines were authorized by FDA and recommended by the Advisory Committee of Immunization Practices for use in the United States in December 2020, while the Janssen vaccine was authorized and recommended for use at the end of February 2021. Vaccine administration in the United States began within a few days of authorization for each vaccine. As no vaccine is 100% effective at preventing illness, COVID-19 occurring among vaccinated people, often referred to as vaccine breakthrough infections, are expected. Amid increases in vaccination coverage in the setting of widespread SARS-CoV-2 transmission, the numbers of COVID-19 cases among vaccinated people could be substantial. We estimated the number of symptomatic vaccine breakthrough infections expected in the United States based on published vaccine efficacy (VE) data, percent of the population that had been fully vaccinated, and reported COVID-19 case counts. # Methods We developed a tool in Microsoft Excel^© to estimate the expected number of symptomatic COVID-19 cases among vaccinated persons per day in the United States using publicly available data. Inputs are the 7-day moving average for daily numbers of COVID-19 cases in the United States as reported to the Centers for Disease Control and Prevention (CDC), the cumulative number of persons fully vaccinated with each vaccine as reported to CDC as of 14 days prior to each 7-day average case count, and VE data from phase 3 trials of the three vaccines authorized in the United States. The number of symptomatic vaccine breakthrough infections, rather than all symptomatic and asymptomatic vaccine breakthrough infections, were calculated because prevention of symptomatic disease was the primary phase 3 clinical trial endpoint reported for the COVID-19 vaccines authorized in the United States and corresponds to published vaccine efficacy figures. As this project incorporates secondary use of publicly available data, human subjects research review was deemed unnecessary. Breakthrough cases were defined as those occurring in persons ≥14 days after completion of vaccination with an authorized COVID-19 vaccine, a delay to reflect when maximum immunity conveyed by vaccination is reached. Given the similar reported VE from clinical trials for the Pfizer-BioNTech and Moderna vaccines, the average VE of 94.6% was used for both mRNA vaccines in the calculator, while 66.9% VE was used for the Janssen vaccine. Calculations were restricted to persons aged ≥18 years for January through May, the primary population that received vaccines during that time; the population reflected in calculations was expanded to approximate persons aged ≥12 years beginning at the end of May. Available data suggest that 87.4% of reported U.S. cases to date as of the end of July had occurred among persons aged ≥18 years, and 93.7% of total cases had occurred among persons aged ≥12 years. We approximated the average number of COVID-19 cases occurring per day among the vaccine eligible population accordingly. We also proportionally restricted the denominator data used to approximate the proportion of the vaccine eligible population that was fully vaccinated per 2019 U.S. census estimates. The number of persons fully vaccinated with Janssen vaccine registered as of 14 days prior to each date for case count ascertainment was subtracted from the total number of persons vaccinated as of that date to approximate number of persons vaccinated with Pfizer-BioNTech or Moderna vaccines at that time. The number of symptomatic vaccine breakthrough infections expected per day is a function of VE and vaccination coverage in the population. For these calculations, *C* denotes the approximated 7-day moving average for daily number of reported COVID-19 cases among the vaccine eligible population and *V* represents different vaccination “groups” according to numeric subscripts: 0 for unvaccinated or not fully vaccinated, 1 for Janssen vaccine, and 2 for Moderna and Pfizer-BioNTech vaccines; *%V* is the percent of the population fully vaccinated in each vaccine group. VE is calculated as (1 –the risk ratio \[RR\]), where RR is the ratio of confirmed symptomatic SARS-CoV-2 infections per 1000 person-years among those receiving vaccine in phase 3 trials divided by those receiving placebo. The Janssen RR₁ was 0.331 (VE₁ = 0.669), the Pfizer-BioNTech and Moderna RR₂ was 0.054 (VE₂ = 0.946), while RR₀ was defined as 1 (VE₀ = 0) for people who are unvaccinated or not fully vaccinated. The expected number of symptomatic vaccine breakthrough infections per day is calculated as: $$\frac{C\left( {\hat{{RR}_{1}}*\% V_{1}} \right) + \mspace{360mu}\left( {\hat{{RR}_{2}}*\% V_{2}} \right)}{{\% V}_{0} + \mspace{360mu}\left( {\hat{{RR}_{1}}*\% V_{1}} \right) + \mspace{360mu}\left( {\hat{{RR}_{2}}*\% V_{2}} \right)}$$ Variance was calculated based on available phase 3 clinical trial data for the Pfizer-BioNTech and Moderna vaccine trials using Poisson regression models. The pooled variance of the expected symptomatic vaccine breakthrough infections was estimated to be $\text{Var}({\hat{\beta}}_{1}) = 0.01149$, $\text{Var}({\hat{\beta}}_{2}) = 0.05551$, with $\text{Cov}({\hat{\beta}}_{1},{\hat{\beta}}_{2}) = 0$ and calculated as $$C^{2}\frac{\left( {\hat{RR_{1}}*\text{\%}V_{0}*\text{\%}V_{1}} \right)}{\left\{ {\left\lbrack {\text{\%}V_{0} + \mspace{360mu}\left( {\hat{RR_{1}}*\text{\%}V_{1}} \right) + \mspace{360mu}\left( {\hat{RR_{2}}*\text{\%}V_{2}} \right)} \right\rbrack*2} \right\}^{2}}\mspace{360mu}\text{Var}{\hat{\beta}}_{1} + \mspace{360mu} C^{2}\frac{\left( {\hat{RR_{2}}*\text{\%}V_{0}*\text{\%}V_{2}} \right)}{\left\{ {\left\lbrack {\text{\%}V_{0} + \mspace{360mu}\left( {\hat{RR_{1}}*\text{\%}V_{1}} \right) + \mspace{360mu}\left( {\hat{RR_{2}}*\text{\%}V_{2}} \right)} \right\rbrack*2} \right\}^{2}}\mspace{360mu}\text{Var}{\hat{\beta}}_{2}$$ The first persons in the United States to be vaccinated against SARS-CoV-2 completed their 2-dose series during the week of January 4, 2021. Therefore, we began calculating the expected number of symptomatic vaccine breakthrough infections 14 days later, the week beginning January 17. We calculated weekly estimates using approximated average case counts among the vaccine eligible population and vaccination coverage as of the Sunday beginning each week and then multiplied the daily estimate by seven. We estimated per-week symptomatic breakthrough infections through the last week of July. Incorporation of Janssen vaccine into the estimates began in mid-March. Case counts and eligible population included persons aged 12yearss during the week beginning May 30. We calculated cumulative expected counts to date by summing weekly expected vaccine breakthrough case counts. We derived 95% confidence intervals (CI) around cumulative counts by summing weekly variances as described above into standard CI calculations. To understand the relative influence of community transmission and VE in determining the number of expected vaccine breakthrough infections, we calculated expected cumulative counts during the same time period under two hypothetical scenarios: 1) doubling the average daily COVID-19 case counts each week; and 2) modifying population vaccination coverage during January–July such that it entirely reflected VE of 67%, VE associated with the Janssen vaccine. # Results Nearly 12 million COVID-19 cases were reported in the United States during January–July 2021. The number of COVID-19 cases reported per day during this period ranged from a high of approximately 210,000 cases in mid-January to a low of approximately 12,000 cases in late June. The estimated number of symptomatic vaccine breakthrough infections in the United States ranged from a low of almost two per day (11 per week) in January to nearly 5,000 per day (34,000 per week) during the last week of July. As of the end of July, we estimate that a total of 198,840 symptomatic vaccine breakthrough infections (95% CI: 183,346–214,333 cases) occurred in the United States among \>156 million fully vaccinated people ( and). On average, starting in February, the number of expected vaccine breakthrough infections increased by 37% each week, but slowed beginning the last week of April amid falling numbers of COVID-19 cases in the United States. This trajectory in the number of expected symptomatic vaccine breakthrough cases each week shifted rapidly with the increasing SARS-CoV-2 transmission driven by the spread of the Delta variant beginning in late June. The number of expected vaccine breakthrough infections during that time translates to a cumulative incidence of approximately 127 vaccine breakthrough infections per 100,000 fully vaccinated people. The expected number of vaccine breakthrough infections varied substantially under different hypothetical scenarios reflective of 1) doubling daily average case counts, and 2) all vaccination occurring at VE of 67%. The relationship between COVID-19 cases and the expected number of vaccine breakthrough infections was proportional (i.e., when case counts doubled, so did symptomatic vaccine breakthrough infections), whereas VE had far more influence on the expected number of symptomatic vaccine breakthrough infections. Compared to vaccination with an average VE of nearly 95%, as occurred during January through July in the United States, a hypothetical scenario in which all vaccination occurred at VE of about 67% nearly quadrupled the number of expected symptomatic vaccine breakthrough infections without modifying other parameters. # Discussion We created a spreadsheet-based calculator to estimate the number of symptomatic vaccine breakthrough infections in the United States based on the average number of COVID-19 cases, percent of the population fully vaccinated, and published efficacies of the three FDA-authorized vaccines. Using this tool, we estimate that approximately 200,000 symptomatic SARS-CoV-2 vaccine breakthrough infections occurred by the end of July among the over 156 million people fully vaccinated in the United States by the middle of July. Vaccine breakthrough infections occur in only a small fraction of all vaccinated persons. Understanding the number of expected vaccine breakthrough infections is important for accurate public health messaging to help ensure that the occurrence of such cases does not negatively affect vaccine perceptions, confidence, and uptake. We developed this tool incorporating ideal VE scenarios. Real-world vaccine effectiveness may be lower, particularly for people who are older or have underlying health conditions. Lower effectiveness also may result from decreased protection against certain SARS-CoV-2 variants, including the Delta variant. Early vaccine effectiveness studies in the United States and elsewhere demonstrated high effectiveness of mRNA vaccines against symptomatic infection and severe disease in various real-world situations, including approximately 95% effectiveness among large cohorts of healthcare workers and \>85% among residents of skilled nursing facilities. However, more recent estimates of vaccine effectiveness against infection with the Delta variant have decreased, while effectiveness remains high against hospitalization and death. Although we used efficacy data from clinical trials as inputs to estimate expected vaccine breakthrough infections, these inputs could be updated using additional vaccine effectiveness data as those become increasingly available. COVID-19 vaccines have demonstrated the ability to mitigate risk of severe disease, hospitalization, and death among persons infected following vaccination. Clinical trial endpoints utilized here were for prevention of symptomatic infection; the effectiveness of authorized vaccines at preventing asymptomatic SARS-CoV-2 infection is still unclear but preliminary reports suggest the mRNA vaccines were \>90% effective at preventing infection prior to circulation of the Delta variant. We did not incorporate asymptomatic infections into these calculations, nor did we update efficacy inputs to reflect decreased vaccine effectiveness against the Delta variant, which was just beginning to circulate at the end of this study period. With decreased vaccine effectiveness and increasing numbers of vaccinated persons, the expected numbers of symptomatic vaccine breakthrough cases will represent a larger proportion of total COVID-19 cases. The analytic approach described here is based on several assumptions and limitations that affect how our results should be interpreted. First, this approach is based on reported case counts and does not account for the population at-risk, susceptibility of persons previously infected, or duration of immunity following vaccination. Second, these calculations assume that vaccinated and unvaccinated people have the same risk of exposure to SARS-CoV-2, which may not be true at the population level. Third, we define vaccine breakthrough infections as those occurring more than two weeks after completion of vaccination; these figures do not include people who may become infected following partial vaccination or prior to 14 days following completion of vaccination. Fourth, reported COVID-19 case counts stratified by patient age are not available from all states. Our assumption that adults comprise 87.4% of reported cases and persons aged ≥12 years comprise 93.7% of total cases reflects cumulative trends since the beginning of the pandemic; these data may not reflect the age distribution during the weeks included here and only approximate the number of COVID-19 cases occurring among vaccine eligible people. If the proportion of total cases occurring among the vaccine eligible population decreased over time, our assumptions would yield an overestimate of vaccine breakthrough cases. Lastly, reporting delays among both COVID-19 case and vaccine administration data vary, and data are often updated retrospectively. Therefore, the figures used for these calculations should be viewed as approximations. Collectively, because of these limitations, the specific estimated counts should be interpreted with caution. Nevertheless, they provide useful context for guiding expectations. Risk reduction provided by any vaccination is inherently relative, and the number of cases among vaccinated persons assuming equal exposure as unvaccinated persons is directly linked to vaccination coverage and disease incidence. Even with highly effective vaccines, given the large number of people being vaccinated in the United States and high levels of SARS-CoV-2 transmission in many parts of the country, hundreds of thousands of symptomatic vaccine breakthrough infections are expected, and will continue to accumulate amid high- levels of SARS-CoV-2 circulation. The methods described here can be used by public health officials to determine if the frequency of vaccine breakthrough infections reported in their jurisdictions are consistent with expectations based on vaccine efficacy from clinical trials. Furthermore, public health messaging regarding expected vaccine breakthrough infections is important to assure the public that this is expected, is not cause for alarm, and does not indicate that vaccines are not preventing severe COVID-19. Vaccine breakthrough infections are expected to continue to accumulate amid ongoing widespread community transmission of SARS-CoV-2 and high vaccination coverage. However, the number of COVID-19 cases, hospitalizations, and deaths prevented among vaccinated persons will far exceed the numbers of vaccine breakthrough infections. CDC continues to collaborate with public health officials nationwide to monitor COVID-19 trends among vaccine persons and to identify unexpected trends in characteristics of people with vaccine breakthrough infections or patterns associated with infecting strains. # Supporting information CDC’s COVID-19 Vaccine Breakthrough Investigation Team; Lindsey Duca, Jaymin Patel, Perrine Marcenac, CDC, for helpful review of the methodologic approach, and Paul Mead for figure suggestions. **Disclaimers**: The findings and conclusions in this report are those of the author(s) and do not necessarily represent the official position of the Centers for Disease Control and Prevention. Names of specific vendors, manufacturers, or products are included for informational purposes and does not imply endorsement of the vendors, manufacturers, or products by the Centers for Disease Control and Prevention or the U.S Department of Health and Human Services. [^1]: The authors have declared that no competing interests exist.

# Introduction Aquaculture production is increasing worldwide as a source of fish for human consumption. Aquaculture introduces land-derived microbes, nutrients, metals and other chemicals such as antibiotics to the water environment. The prophylactic and therapeutic use of antibiotics results in the occurrence of antibiotic- resistant bacteria and antibiotic resistance genes (ARGs) in the aquaculture environment. This may lead to seawater and the sediment becoming reservoirs for ARGs. The ARGs in aquaculture environments can be transferred horizontally among microbes and ultimately be transferred to fish pathogens. Thus, the presence of ARGs in aquaculture environments may lead to inefficiency in treating fish diseases using antibiotics. To avoid production losses in the fish-farming industry, it is important to control the occurrence and spread of ARGs in aquaculture facilities. The spreading of ARGs in the environment is mediated by horizontal gene transfer (HGT). Therefore, genes associated with HGT should also be examined when determining the prevalence of ARGs in the environment. Integrons can contribute to the occurrence of HGT of ARGs in bacterial populations, as a consequence of possessing a site-specific recombination system capable of capturing gene cassettes containing ARGs. Integrons can be carried by mobile genetic elements such as transposons and plasmids that promote their wide distribution within bacterial communities. Class 1 integrons, which contain an *intI1* gene encoding an integrase of the tyrosine recombinase family, are known to carry gene cassettes containing ARGs. Class 1 integrons have been found in cultured fish pathogens and from cultured bacteria in the aquaculture environments. Sulphonamides potentiated with trimethoprim or ormethoprim and florfenicol are some of the antibiotics commonly used in aquaculture. Consequently, the presence of several antibiotic-resistant bacteria in aquaculture environments has been reported previously using culture-dependent methods. Bacteria of the *Actinobacter* and *Bacillus* genera have been observed to carry the resistance genes to sulphonamides and *Aeromonas* and *Pseudomonas* to florfenicol. Also aquatic bacteria of the genera *Proteus* and *Pseudomonas* can carry the resistance genes to both sulphonamides and trimethoprim. However, culture- dependent methods may introduce bias when determining the prevalence of ARGs due to the inability to cultivate the majority of bacteria from environmental samples. Culture-independent methods, including the measurement of gene copy numbers by quantitative polymerase chain reaction (qPCR), give a less biased estimation of the ARG amounts in the environment. qPCR has been widely used to study ARGs in environmental samples – and aquaculture environments. However, little is known about the long-term persistence of ARGs at aquaculture sites and their dispersal to the surrounding environment. To investigate these aspects, we collected sediment samples below two open-cage fish farms located in the Turku Archipelago, Finland in the northern Baltic Sea during the summers of 2006 to 2012. Sediment samples were also collected 200-m and 1000-m from fish farms, as well as from transect sites at 200-m intervals up to 1000 m from one farm to observe the dispersal of ARGs to the surrounding sediment environment. We quantified ARGs for sulphonamides, trimethoprim, and florfenicol and an integrase gene of class 1 integrons using qPCR. The antibiotic concentrations (sulphamethoxazole, sulphadiazine and trimethoprim) were measured using liquid chromatography-mass spectrometry (LC-MS). Our findings show that the ARGs were abundant and persistent in the sediments below the fish farm cages during the 6-year observation period, but were not detected in the sediments even at the closest 200-m distance from the cages. Moreover, a correlation was found between the amount of class 1 integrons and the *sul1* gene. # Results ## Antibiotic resistance genes and class 1 integrons in sediments The detection of trimethoprim resistance genes (*dfrA1, dfrA2, dfrA5, dfrA12, dfrA15, dfrA16, dfrA17,* and *dfrA19),* sulphonamide resistance genes (*sul1, sul2* and *sul3*), florfenicol resistance gene (*floR*) and an integrase gene (*intl1*) of class 1 integrons was done for six sediment samples chosen from northern Baltic Sea fish farm sites, using standard PCR. From the targeted ARGs, the *dfrA1, sul1, sul2* and *intI1* genes were found at two fish farms (FIN1 and FIN2). The amounts of the genes detected were measured, using qPCR in all 51 sediment samples. The copy numbers of the four genes were under the detection limit in every sediment sample taken 200 m to 1000 m from the farms as shown in. The dispersal of the ARGs was not considerable, since the genes were not detected even at the closest 200-m distance from each farm. The *sul1, sul2* and *intI1* genes were present in every sample and the *dfrA1* gene in most samples from the FIN1 and FIN2 farms throughout the 10 sampling times from 2006 to 2012. The *sul1*, *sul2* and *intI1* gene copy numbers at the two farms were similar, with about 10⁻³–10⁻² copies in proportion to the 16S ribosomal RNA (rRNA) gene copies. The *dfrA1* gene copy numbers varied approximately 10⁻⁵–10⁻² copies in proportion to the 16S rRNA gene copies. The abundance of the *dfrA1* gene was lower than and significantly different from that of the *sul1* (*P*\<0.001), *sul2* (*P*\<0.001) and *intI1* (*P*\<0.001) genes. ## Correlation between ARGs and class 1 integrons in the sediments Linear regression analysis was performed to test whether the copy number of the class 1 integron was correlated with any of the three ARGs detected (*dfrA1*, *sul1* and *sul2*) and thus could have played a role in the prevalence of the ARGs. Significant correlation (*F*_1,22  = 19.39; *P* = 0.000225; *R²* = 0.47) was found only between the average copy numbers of the *intI1* and *sul1* genes. The prevalence of the *sul1* gene in the Baltic Sea farm sediment may therefore be associated with class 1 integrons. ## Antibiotic concentrations in the sediments The sulphamethoxazole, sulphadiazine, and trimethoprim concentrations were measured, using LC-MS analysis to estimate the presence of selection pressure in the sediments. The antibiotic concentrations in the sediments are shown in. All the antibiotic concentrations measured in the farm sediments were very low (1.5–101 ng g⁻¹ of dry sediment) and some even below the detection limit (\<1 ng g⁻¹ of dry sediment). The antibiotic concentrations were also below the detection limit in the sediments taken 200 m to 1000 m from the farms. Hence, there was no clear selection pressure in the sediments. # Discussion Our results showed that the sulphonamide resistance genes (*sul1* and *sul2*) and trimethoprim resistance gene (*dfrA1*) were persistent in the Baltic Sea farm sediments during the 6-year observation period. We assume that the Baltic Sea farms have relatively small impact from human habitats and agriculture and therefore municipal and agricultural ARG sources can be excluded. The antibiotics and organic matter from uneaten fish feeds and fish excrement can enter the sediments directly from the water since there is no purification process in the open-cage farming system. Thus, the selection for the resistant bacteria and ARGs may have occurred in the medicated fish feeds or inside the fish intestines and fish faeces that entered the sediments, or selection through the antibiotics present in the sediments. However, the LC-MS results indicated low concentrations of sulphonamides and trimethoprim, suggesting that there was no clear selection pressure in the farm sediments. Furthermore, sulphonamides are decomposed by chemical and biological factors with half-lives of 7–85 days. The persistence of ARGs in the farm sediments may, therefore, have been due to a constant introduction of ARGs from external sources, such as uneaten medicated fish feeds and fish faeces. On the other hand, very low concentrations of the antibiotics may also play a role in maintaining the ARGs in the farm sediments by selection and enrichment of resistant bacteria by a subinhibitory antibiotic concentration, which was shown in a previous study. Moreover, the presence of antibiotics in a subinhibitory concentration may induce the HGT system in bacterial communities, which further increases the prevalence of ARGs. Therefore, the potential for low antibiotic concentrations in maintaining resistant bacteria also needs to be examined to further understand the persistence of ARGs in northern Baltic Sea farm sediments. In this study, the dispersal of ARGs (*sul1, sul2 and dfrA1*) from the Baltic Sea aquaculture farms to the surrounding sediments was not detected. Previous work from the same fish farms locations shows similar results; the four tetracycline resistance genes studied were not detected in the sediments even at the closest 200-m distance from the farm cages. These results suggest that the resistance genes are potentially a problem for the fish-farming industry, but impact less the surrounding sediment environment in the northern Baltic Sea. The copy number of the class 1 integron *intl1* was significantly correlated with the *sul1* gene copies in the farm sediments. This was expected, since the *sul1* gene is one of the backbone genes of the 3′-conserved segments in class 1 integrons. Class 1 integrons are, therefore, likely involved in the prevalence of the *sul1* gene in Baltic Sea fish farm sediments. Similar correlations between the copy numbers of the *intI1* and *sul1* genes have been observed in riverine sediments in Haihe, China and in Colorado, USA. Only *dfrA1* of the eight trimethoprim resistance genes analysed (*dfrA1, dfrA2, dfrA5, dfrA12, dfrA15, dfrA16, dfrA17 and dfrA19*), which are commonly associated with class 1 integrons, was detected in the Baltic Sea farm sediments. There was no significant correlation between the amounts of the *intI1* gene of class 1 integrons and the *dfrA1* gene copies, suggesting that the *dfrA1* gene was not associated with class 1 integrons in the farm sediments. The prevalence of the *dfrA1* gene may have been mediated by other mobile elements or even independently of them. The copy number of *dfrA1* was significantly lower than that of *sul.* Although the amount of *dfrA1* genes was lower than the amount of *sul* genes, the prevalence of both genes in the aquaculture environment deserves equal focus, because sulphonamides and trimethoprim are used in combination as so-called potentiated sulphonamides. While many studies have demonstrated sulphonamide resistance in aquaculture environments, to our knowledge, this is the first study reporting qPCR measurements of the trimethoprim resistance gene (*dfrA1*) in aquaculture-impacted sediment samples. In conclusion, the persistence of ARGs in farm sediments may lead to problems in the efficiency of antibiotics used to treat fish diseases and eventually to production losses at the fish farms. It is important for the fish-farming management to control the use of antibiotics to avoid the emergence of ARGs at the farms. Since the ARGs were not detected 200 m and 1000 m from the farms, their presence at the farms is unlikely to cause serious effects in the aquatic environment surrounding fish farms in the northern Baltic Sea in the current environmental conditions. However, a change in environmental conditions or an extended exposure to nearby fish farming activity could conceivably lead to an emergence of antibiotic resistance genes in the future. The sources and the spread of ARGs in aquaculture process chains should also be studied, as well as their potential risk to human health. # Materials and Methods ## Study site and sampling The sediment samples were collected from two fish farms (FIN1 and FIN2) and nearby areas located in the Turku Archipelago, Finland in the northern Baltic Sea from 2006 to 2012. The northern Baltic Sea is a unique brackish water marine environment (mean salinity: 6.7 parts per thousand) and no tide. The sampling locations are described in. Both the FIN1 and FIN2 farms use open-cage systems in which the fish are kept in net cages that allow free transfer of uneaten fish feeds and fish excrement from the cages to the surrounding waters and eventually to the sediments. The farms raise European whitefish (*Coregonus lavaretus* (L.)) and rainbow trout (*Oncorhynchus mykiss* (Walbaum)). Each farm produces approximately 50 tons of fish annually. The record amount of antibiotics used at the fish farms was not available. Northern Baltic Sea aquaculture farms operate only in summer. Sampling was done 10 times during the 6-year observation: June, July, August and September 2006, June 2007, June and September 2008, June 2009, September 2011 and September 2012. In addition, transect interval samples were collected at sites 200-m up to 1000-m distance from the FIN1 farm on September 2008. Three replicate samples were collected in each year from 2006 to 2009 and the replicates from each year were pooled. In 2011 and 2012, three biological replicates were individually collected. In all, 51 samples from the FIN1 farm and FIN2 farm and surrounding areas were collected using a Limnos sediment probe (Limnos Ltd., Turku, Finland). Each sample was homogenized manually inside a zipper storage plastic bag and immediately frozen on dry ice. The sediments were stored at −80°C until DNA extraction. ## DNA extraction The environmental total DNA was extracted from 0.5 g wet weight sediment, using the FastDNA® SPIN kit for soil (MP Biomedicals, Illkrich, France). The standard protocol was modified by adding an extra washing step with 5.5 M of guanidine thiocyanate (Sigma Life Science, Steinheim, Germany), according to the manufacturer's instructions for removing humic acids. The DNA quality and concentration were analysed with a Nanodrop 1000 spectrophotometer (Thermo Scientific, Wilmington, DE, USA). The extracted DNA was stored at −20°C. ## Standard PCR The PCR primers and conditions for detecting the presence of targeted genes in the sediments were described previously: sulphonamide resistance genes (*sul1, sul2* and *sul3*), trimethoprim resistance genes associated with mobile elements (*dhfr1, dfrII, dfrV, dhfrIX, dfrXII, dhfrXV, dfr16, dfr17* and *dfrA19*), currently known as *dfrA1, dfrA2, dfrA5, dfrA9, dfrA12, dfrA15, dfrA16, dfrA17* and *dfrA19*, florfenicol resistance gene (*floR*) as well as the integrase gene of the class 1 integron (*intI1*). Six sediment samples from the FIN1 farm, FIN2 farm and 1000-m distance from the FIN1 farm, taken during years 2007 and 2011, were chosen for gene detection in PCR. The 25-μl PCR reactions consisted of 1× Taq buffer with (NH₄)2SO₄, 2 mM MgCl₂, 0.2 mM of each deoxyribonucleotide triphosphate (dNTP), 50 U Taq DNA polymerase, recombinant (Finnzymes, ThermoFisher Scientific, Espoo, Finland), 0.2 μM of each primer (Oligomer Oy, Helsinki, Finland) and the DNA template. The negative controls had nuclease-free water and were done for every PCR reaction. All the PCR reactions were done in triplicate using a PTC-200 thermal cycler (MJ Research, Watertown, MA, USA). The PCR products were purified, using QIAquick PCR purification kit (Qiagen, Hilden, Germany) and sequenced by the DNA sequencing service at the Institute of Biotechnology (University of Helsinki, Finland) to confirm the sequences of the PCR product. ## Quantitative PCR measurement The primers for the *dfrA1* gene and the *intI1* gene used in standard PCR were not optimal for the qPCR assays in this study. Thus, new primer sets were designed. The *dfrA1* and *intI1* gene sequences (accession numbers indicated) were submitted to Primer3 v.2.3.4 to produce primer sets with melting temperatures above 60°C and amplifying 150–250 base-pair (bp)-long fragments. The primers designed were compared with known gene sequences that were retrieved from GenBank and aligned with mafft to choose primer sets that are located in conserved regions. Basic Local Alignment Search Tool (BLAST) analysis of the primer sequences against the National Center for Biotechnology Information (NCBI) database was performed to avoid nonspecific amplification. Plasmids R388, RSF1010 and pUV441 were used as the *sul1, sul2,* and *sul3* gene standards. Presynthesized pUC57 vectors containing either the entire 459-bp sequence of the *dfrA1* gene or a 250-bp fragment of the *floR* gene were ordered from GenScript (Piscataway, NJ, USA). The chromosomal DNA of *Escherichia coli* K12 was used as the 16S rRNA gene standard. The plasmid standards were linearized and purified. The standard copy numbers per μl were calculated using the estimated molar mass for the DNA bp (650 Da/1 bp). Determination of the qPCR efficiency was done, using a five-point 10-fold dilution series. In addition, inhibition tests were performed as previously described to observe whether the sediment samples had the same amplification efficiency as the standard. The average inhibition was 2.5% (+/− SD 1.9%). The qPCR was performed using a 7300 real-time PCR system (Applied Biosystems, Foster City, CA, USA). The 20-μl qPCR reactions contained 1× DyNAmo Flash SYBR Green Master Mix (Thermo Scientific) with 1× of ROX passive reference dye, 0.625-μM primers for the targeted gene listed in and freshly diluted DNA sample. Nuclease-free water was added instead of DNA samples for the no template control (NTC) sample. The qPCR programs consisted of 7 min at 95°C, 40 cycles of 10 s at 95°C, 30 s at the annealing temperatures (Ta) listed in and melting curve analysis. Three technical replicates of each sample, NTC and standard dilution were performed in each measurement and the measurements were repeated once to determine the reproducibility of the qPCR assays. Based on previous studies of the validation of qPCR assays, each assay was performed until the qPCR assay characteristics were achieved as described in. The limit of quantification (LOQ) of the qPCR assay was 5.92×10⁻⁵ copies in proportion to the average number of 16S rRNA gene copies for *sul1*, *sul2*, *sul3* and *floR*, 1.18×10⁻⁵ copies in proportion to the average of 16S rRNA gene copies for *dfrA1* and 2.96×10⁻⁴ copies in proportion to the average number of 16S rRNA gene copies for *intI1*. Data analysis was performed manually, using the 7300 System SDS v.1.2 Software (Applied Biosystems). ## Statistical analysis Student's t-test was performed to determine whether the copy numbers of the ARGs from all farm samples varied significantly. The correlation between the average gene copy numbers of the integrase gene of the class 1 integron and the ARGs detected in the farm sediments was analysed, using linear regression. The t-test and the linear regression with log-transformed variables were performed using RStudio v.0.97.168 (RStudio, Boston, MA, 2012). All models were considered to be significant at p-values less than 0.05. ## Analytical methods for antibiotic quantification The liquid chromatography analyses were performed with a Hewlett-Packard 1100 (Hewlett-Packard Co., Palo Alto, CA, USA). The reagents used and extraction methods are described in. The antibiotics were separated on a reverse-phase column (YMC Pro C18, 3 μm, 150 mm×2 mm; YMC America Inc., Allentown, PA, USA) operated at 30°C at a flow rate of 0.15 ml min⁻¹. The mobile-phase solvents were water-acidified with 1% (v/v) formic acid (eluent A) and methanol- acidified with 1% (v/v) formic acid (eluent B) to a pH of 2.5. The HPLC gradient programmes are contained in Table S1 in. The antibiotics were detected with a Micromass Quattro Ultima triple-quadrupole MS (Micromass, Milford, MA, USA), equipped with electrospray ionization. The analyses were performed in the positive ion mode. The protonated molecular ion (\[M+H\]+) of the compounds was selected as the parent ion. Detection was performed in the multiple reaction- monitoring mode, using the two most intense and specific fragment ions. Table S2 in lists the optimized conditions of the individual analytes. To identify the antibiotics, we compared the retention times and the area ratios of the two product ions in each sample with the average retention time and peak ratios of the standards in all measurements. The criteria difference between the samples and the standard was within 0.3 min for the retention time and 20% for the area ratio of the two product ions. The concentrations of the samples were calculated by an external standard method, based on the peak area of the sum of the two product ions monitored. The calibration lines of five concentration points (10, 20, 30, 40, and 50 μg l⁻¹ in water-methanol \[1∶1\]) of the individual antibiotics were used for quantification. The linearity of the calibration curve in this range was confirmed (*R*²\>0.99). The final concentrations of most of the samples in the vials were within the range of the calibration lines.The LOQs were defined as 10 times the noise level of the baseline in the chromatograms signal to noise (S/N) ratio and were 1 ng g⁻¹. ## Ethics statement No specific permits were required for the field study described. The sampling did not affect any protected or endangered organisms. # Supporting Information We would like to thank Kornelia Smalla (Julius Kühn-Institut Federal Research Centre for Cultivated Crops, Braunschweig, Germany) for providing the plasmids R388, RSF1010 and pUV441. We thank Johanna Muurinen and Leena Pitkänen for technical assistance during sampling. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WIM MT MV. Performed the experiments: WIM SM KP. Analyzed the data: WIM SM AK CL MT SS MV. Contributed reagents/materials/analysis tools: SM MV. Wrote the paper: WIM SM KP.

# Introduction Worldwide, most long-distance terrestrial migrations have been lost or greatly reduced largely due to anthropogenic factors creating barriers to wildlife movements. Restricted access to food and water is a primary threat to migrations worldwide. In particular, fencing has played a large role in ungulate population declines where fencing closes off parks, delineates national boundaries and separates rangelands, thus physically cutting off access to necessary resources, and in grassland bird decline, where birds may not be able to see fencing and die due to collision or entanglement. In Africa, fences were implemented in the 1950's to prohibit transfer of wildlife disease between livestock and wildlife populations. In Kruger National Park, fencing restricted migrations of wildebeest and the population, cut off from seasonal water sources, declined by nearly 88%. In North America, 75% of seasonal migrations, mostly those of bison (*Bison bison*) and pronghorn antelope (*Antilocapra americana*) have been lost largely due to overhunting and disruption of migration routes and it is hypothesized that fence collisions have contributed to the decline of sage grouse populations. Wire fencing was introduced in the North American west by the homesteaders of the 19^th century to avoid importing timber or stone to create barriers around their land. By the late 19^th century, barbed wire had been invented and was being produced commercially at the high demand of homesteaders wishing to claim their land. By 1880–1884, barbed wire production reached a peak with an estimated 643,000–965,000 km being produced annually. Eventually ranchers and farmers in the west began using fencing pervasively. Subsequently, access to roads and water holes were inadvertently restricted for cattle. Today much of the fencing erected in the 1800's still exists and is profuse across northern Montana. It is used by land owners to delineate property, section off agricultural fields, demarcate parcel boundaries of the same ownership, corral cattle, and along roads for safety. Fencing type in northern Montana varies with land ownership and operation type. Three and four strand barbed wire are most common, however, mesh woven fences are also common in the Hi-Line region of northern Montana (A. Jakes, University of Calgary, *unpublished data*) and these wire fences can pose significant impediments to the movements of wildlife, thus modifying habitat availability and migration opportunities of various native species. As such, fences can influence a particular species movement patterns at multiple spatial and temporal scales. Species perceive the surrounding landscape differently from one another to fulfill life-history requirements. Appreciating that wildlife may have multi- scaled responses to anthropogenic factors such as fences requires that we too address anthropogenic factors at various scales. Loss of migrations could result in extirpation of wildlife populations, resulting in overall contraction of species range due to decreased access to forage and safe calving grounds. Thus, to conserve migrations and allow continued movement to optimal seasonal ranges, conservation and maintenance of habitat outside of protected areas is required. Because wildlife generally track high quality forage and water sources through their migrations, erecting fencing along or through migration routes or breeding habitat can restrict access to necessary habitat, resulting in declines in population numbers. For example, in a study in Colorado and Utah, USA, most mule deer (*Odocoileus hemionus*), elk (*Cervus canadensis*) and pronghorn (*Antilocapra americana*) fence-related mortalities were due to becoming entangled in fence wires. Furthermore pronghorn density and fence-related mortality were positively correlated; indicating fencing may negatively impact pronghorn populations. In addition, increased indirect mortalities of fawns being separated from does was observed from woven wire fences as opposed to plain barbed wire fences and additionally, mortalities increased with increasing fence height. In a region where fencing is a common landscape feature such as northern Montana, migratory opportunities could be diminished and as a result, population numbers of wildlife could suffer declines unless landscape permeability is maintained. Although North American migrations are among the best studied, and despite the importance of fencing information on wildlife movements at multiple scales, regional datasets on fence locations and fence densities do not exist. In this study we create a regional fence location and density model for the Hi-Line area of northern Montana. To aid in future studies identifying wildlife movement pathways or identifying factors affecting seasonal habitat selections we determined it was important to have a readily available predicted fence location and density information for this region. We provide methods for modeling fence locations and density at a large scale, using publicly available information in an effort to encourage the creation of fence spatial datasets for wildlife research and management in other regions. # Methods ## Study Area The study area is comprised of 13 counties in the Montana Hi-Line region of northern Montana. These counties are bounded by the Canadian border on the north and the Marias and Missouri Rivers to the south. The study area was chosen initially due to ongoing research on pronghorn migrations and movements at the northern terminus of pronghorn range (completed under wildlife capture and handling permit \#11-2007 from the Montana Fish, Wildlife & Parks Institutional Animal Care and Use Committee). The total area for which fencing was modeled was 103,426 km². Privately owned land is the dominate tenure type, more than double public land ownership. ## Ground Truth Surveys ### Transect Identification A seamless land cover dataset developed by bordering state and provincial wildlife agencies was clipped to the bounds of the study area. From this land cover data, three generalized habitat type regions were delineated by extracting polygons around land cover of similar types. Habitat types included “grassland”, “agriculture” and “shrubland”. To create a fourth generalized habitat type, we used removed areas not previously defined and classified as “mixed” habitat to include the remaining areas. Next, we identified roads, paved or unpaved, from the roads dataset created by the Northern Sagebrush Steppe Initiative and intersected this layer with the generalized habitat type region within the Montana portion of the Northern Sagebrush Steppe Initiative land cover data. New areas were identified, based on both generalized habitat type and road type, leaving a total of eight different classifications: Grass/Unpaved; Grass/Paved; Agriculture/Unpaved; Agriculture/Paved; Shrub/Unpaved; Shrub/Paved; Mix/Unpaved and; Mix/Paved. Within each of these generalized areas, random points along roads were generated, each which served as the middle point of each 3.2 km road transect for field surveying. This distance was used to capture a majority of land tenure changes within the “checkerboard pattern” of land ownership found within this area. Random points were generated at a minimum distance of 3.5 km apart, to avoid transect overlap. Each random point had a unique transect associated with it and unique numbers were generated and manually entered for each random point. Transect identification took place in ArcGIS 9.3 (ESRI 2009) and random points and transect regions coordinates were defined in WGS 1984. ### Field Surveys During June-August 2009, 2,362 km of fence along roadside transects were sampled in Blain, Phillips, Valley and Daniels Counties. In ArcGIS 9.3 (ESRI 2009), we created random midpoints for 3.2 km long survey transects. At each point along the 3.2 km transect that a roadside fencing ended, appeared, or changed in fence structure, a GPS location was recorded. Locations were also recorded at every interior fence and road intersection. Information on GPS locations, transect number and heading, fence structure, ground cover and road type was recorded. Locations and structure of internal fencing and twinning and tripled roadside fences were also recorded. Fencing had to be within 200 m of the road to be considered along the roadside and therefore within the transect. Additionally, changes in fence structure had to be longer than 100 m to be recorded. GPS waypoints were downloaded after each field day to create an overall layer of fence changes along roadsides or where interior fences converged with roadsides across the area. Complete sampling protocols were created to standardize methodologies for unique fencing schematics found across the landscape and are applicable to all survey areas across the study area. ## Fencing Location Modeling Fence location and density were modeled using private land ownership data provided by the Montana Public Land Ownership dataset in ArcGIS 10.0 (ESRI 2012). We predicted fence presence based on parcel ownership, size and ownership adjacency. We used publicly available free datasets including land tenure data from the Montana Cadastral Database (Montana Department of Administration/Information Technology Services, 2011), pasture data provided by the Bureau of Land Management (Bureau of Land Management, 2011), the 2010 TIGER roads dataset (US Census Bureau, 2010) and land cover from the National Gap Analysis Program Land Cover Dataset (United States Geologic Service 2000) and the 2000 Land Cover for Agricultural Regions of Canada (Canada Agri-Geomatics Service 2000) national land cover dataset. Historically, most ranches were based on ‘sections’ (2.59 km²) in this region of Montana, and we use this unit of measurement as our base unit of area. We began the modeling process with the land ownership polygon dataset. We edited this dataset based on a series of assumptions about where fencing exists around parcels, along roads and where it defines crop fields. The remaining outlines of polygons after applying these assumptions and their associated GIS functions represented potential fencing. For example, we assumed private lands with the same mailing address \<½ section in size are not fenced if adjacent to each other, so parcels \<½ section that are privately owned are merged together in the GIS polygon dataset and the boundary of the two merged parcels is a fence. To create the assumptions, we consulted a variety of local experts, including BLM, Montana Fish, Wildlife and Parks, World Wildlife Fund, Montana Department of Natural Resources & Conservation, Rocky Boys Indian Reservation, Fort Belknap Indian Reservation and, Fort Peck Indian Reservation local personnel. Fence data for the Charles M. Russell National Wildlife Refuge (CMR) was provided, so this area was not included in the modeling efforts, but was added in for density estimates. We first modeled fencing by dissolving and merging the land tenure parcels based on the land tenure-based assumptions and then combined this dataset with the fencing modeled using the land cover-based assumptions. Within areas of large cropland, all land tenure-based fencing was removed, assuming there would be no parcel fencing within these areas, and fencing in these areas follows different rules (See below). Finally, roads-based fencing was overlaid on the two previously combined datasets. ### Land Tenure Fence Modeling To model fencing based on land ownership, we removed all BLM land from the land tenure layer and replaced it with pastures data we received from the BLM. We assumed the pasture polygon outline represented fences and that no additional fencing would exist on the BLM. Next, we eliminated the boundaries of neighboring state lands, assuming that if more than one state-owned parcel was adjacent, they would be fenced together. State-owned lands that were surrounded by BLM lands were dissolved into the BLM lands, assuming that the state lands would be leased to the BLM. Bureau of Reclamation lands were then treated the same as state lands, in that if two or more parcels were adjacent, they would be fenced together. This process was repeated to remove boundaries separating more than one Fish and Wildlife Service parcel. For private and tribal lands, we first merged parcels based on ownership and adjacency (parcels owned by the same party were combined as one if they were adjacent) and then selected resulting parcels that were greater than 2 sections (5.2 km²). These parcels were combined with the overall fence layer. The smaller parcels were dissolved into a neighboring parcel of similar size (5.2 km²) and then combined with the fence dataset. Next, we removed fencing on National Park Service and Forest Service land. This resulted in the final layer representing land tenure fencing. ### Road Fence Modeling We identified primary and secondary roads and buffered them using the estimates of road width as buffer width. These buffer outlines, 19 m wide on primary roads, and 11 m wide on secondary roads then approximated fence lines along roads. These square-ended buffers were dissolved, merged and converted to lines. So fences would not bisect roads, we removed the buffer ends by removing line segments that were exactly 38 m long, in the case of primary roads, and 22 m long in the case of secondary roads. This process left fence lines on either side of these roads. To model fencing along local roads, we first removed roads from the CMR. Next, length was calculated and long local roads (≥1,200 m) were merged with primary and secondary roads. Because local roads were assumed to have fencing only on one side, we used these local road polylines to represent fencing along them. To include local fenced roads in the model without including the misclassified local roads which were two-tracks or driveways we then selected short roads (\<1200 m) that intersected long local roads. This began an iterative process where increasingly smaller roads which intersected larger roads were selected and added to the fenced roads classification. This process ensured small line fragments (two tracks, trails, and possible data errors) from the roads dataset were not included in the fence dataset. We merged the fenced short local roads identified through this process with longer local roads. We buffered this combined dataset by 11 m on each side, and converted the buffer to a line. One side of this double line was erased, and the remaining line represented fencing. Again, the remaining buffer ends were removed, and the resulting dataset of one-sided fencing along local roads was combined with the double-sided fencing along primary and secondary roads, defined by the US Census TIGER roads classification. ### Land Cover Fence Modeling The land cover layer was first converted to a polygon shapefile. To remove inconsistencies and small sections of non-cropland within large areas of crop (usually wheat and corn in this area) we identified crops, dissolve these polygons together and then calculated their area. We did the same for non-crop land cover classes. Next, we identified non-crop polygons ≤½ section (1.3 km²) that intersected large crop (≥3 sections or 7.8 km²). These non-crop sections were dissolved into the large areas of crop. Based on advice we received from experts, we assumed there would always be a fence line between large areas of cropland (≥3 sections or 7.8 km²) and larger areas of native prairie (half of a section or \>1.3 km²). These areas of prairie were identified and erased from the large sections of crop, which created a polygon boundary. This layer of large croplands was then converted to a raster, reclassified, and expanded from within to remove any remaining holes. This was then again converted to a polygon and represented fences around large croplands, and between large croplands and native prairie. We erased all land tenure fencing from the areas of large croplands, assuming the only other fencing on large croplands would be those along roads. ### Fence Dataset Synthesis To combine the three polygon fence layers, roads, land tenure and land cover, we first removed additional fence lines from the land cover layer from the BLM layer, again assuming the polygon outlines of the BLM pastures dataset would represent all of the fencing on this land type. We erased boundaries of large lakes and rivers polygons and lines that intersected study area boundary lines. The polygon layers of parcel fencing and land cover fencing were then combined and this layer was converted to a line dataset. Assuming that there would not be parcel boundary fence parallel to nearby roads fencing, we next removed roads fencing that were completely within a 20 m buffer of other types of fencing. Finally, the remaining roads fencing was merged with the land cover and land tenure fencing. ## Model Accuracy Assessment Because fencing along roads was modeled separately from internal parcel fencing, we completed individual accuracy assessments for these portions of the fence model, as well as a combined accuracy assessment. For the roads accuracy assessment, all fenced transect lines (a fence on either or both sides) were merged and clipped to the accuracy assessment study area to identify the true positives and all non-fenced transects were merged, clipped, and used to identify the true negatives. These line transects were then converted to points and buffered by 30 m to account for spatial error. Only modeled fencing from the roads fencing model were used in this assessment. We identified areas where buffered GPS points representing true positives and true negatives intersected modeled fences along roads, to result in the true positives and false positives, respectively. The number of true positives was subtracted from the total number of fence transect points to result in the number of false negatives. The number of false positives was subtracted from the total number of non-fenced transect points to identify the number of true negatives. Total number of samples was 1,832 for true positives and 469 for true negatives. For the internal fencing accuracy assessment, all GPS points were merged and clipped to the accuracy assessment study area to identify the true positives. We selected GPS locations that were along fenced transects and these points were then buffered by 30 m. We isolated the modeled internal fencing by removing roads fencing and identifying intersections in the modeled internal fencing. The points of intersection were then buffered by 30 m. To identify true positives, we identified where these buffered modeled intersection points intersected the buffered GPS points. The true positives were then subtracted from the total number of GPS points used to result in the false positives. To identify the true negatives, the buffered areas around the GPS sample points were subtracted from the fenced transects to ensure we used actual areas with fences, but with no internal fences (indicated by GPS locations) for this assessment. On the remaining transect lines, where fencing was not observed, we created random points and buffered them by 30 m. Identifying locations where these buffered points intersected modeled internal fences resulted in the false positives. The false positives were then subtracted from the total number of random points to result in the true negatives. Total number of samples was 1,333 for true positives and 1,290 for true negatives. To calculate the total accuracy of the fence dataset, all GPS locations and the modeled fencing were buffered by 30 m to account for spatial error. Random points were created along the non-fenced transects and similarly buffered by 30 m. GPS locations intersecting modeled fencing resulted in the true positive rate and random points along non-fenced transects intersecting modeled fencing resulted in the false positive rate. The number of true positives were subtracted from the total number of GPS locations to result in the false negative rate and the number of false positives were subtracted from the number of non-fence transect random points to result in the true negatives. There were 1,655 samples for the true positive and the number of samples for true negative as 1,113. We then calculated Cohen's Kappa, an accuracy measure commonly used in remote sensing applications. The Kappa statistic is the chance agreement subtracted from the observed accuracy divided by chance agreement subtracted from 1. The Kappa statistic can range from −1 to 1, where 1 represents 100% accuracy, and 0 represents accuracy no better than that due to chance. Negative values are rare and generally indicate accuracy worse than random; a mismatch between ground truth locations and modeled data. Cohen's Kappa and confidence intervals were calculated in R statistical software (R Core Development Team 2012), using the FMSB package, which tests the null-hypothesis that the agreement between the model is the same as random, with Kappa  = 0. Because the advice we received from our experts (local personnel from various organizations) about fencing on large croplands and BLM lands was variable we completed additional accuracy assessments. To identify causes of inaccuracies, we first removed BLM land from the model and repeated the accuracy assessment, and then removed land cover fencing. ## Fence Density After we created the fence model, the density of fences was calculated using ArcGIS 10.0. The density function search radius for the entire study area was 10,000 m and cell size was 1500 m. # Results ## Fence Location and Density Modeling A total of 263,308 km of fencing was predicted. Maximum fence density was 6.79 kilometers of fencing per km². Mean fence density was 2.37 km of fencing per km². Fence density was highest along the U.S. Highway 2 corridor. Roosevelt County had the highest average density at 3.40 km/km² and Rosebud County had the lowest density at 1.22 km/km². ## Model Accuracy Assessment To compare how well our different decision rules reflected true fence lines, we created three fence datasets: all fencing on the landscape; all fencing except fencing associated with BLM lands; and all fencing except fencing associated with large crop lands. From these three layers, we then calculated accuracy for all fence types together; fencing only along roads and fencing only around land parcels. We chose to examine the rules associated with BLM lands and large croplands because our experts suggested these rules may vary greatly across the landscape. Accuracy for all fencing and all types of fencing within the dataset was more accurate than random (Kappa  = 0), with a Kappa of 0.40. Within this dataset, the accuracy of the roads was lowest with a total accuracy (proportion of ground truth points that matched modeled fencing) of 0.63 and Kappa of 0.12. In the dataset excluding fences associated with BLM land, we found a consistent decrease in accuracy. When assessing the accuracy of the roads fencing only, Kappa was −0.07. The internal (parcel) fencing was the most accurate within this layer, with a Kappa of 0.28. Accuracy was highest when large areas of cropland were excluded, in the third fence dataset. In this assessment, Kappa was 0.56. In analyzing accuracy for roads fencing in this dataset, Kappa was 0.27 and Kappa for land tenure fencing was 0.41 in the absence of fencing associated with large croplands (Table5). # Discussion Although these fence location and density datasets can benefit from improvements in data analysis and sampling methods, this exercise did produce noteworthy results. Higher fence densities appear along Highway 2, where residential areas contributed to the increase in density. Less developed areas, such as the CMR and Glacier National Park contribute to the areas of low fence density. From our accuracy assessment, our model displayed actual fence locations along predicted fence lines at a Kappa of 0.40–0.56, considered moderate agreement between the modeled and ground truth data. Therefore, using this modeling approach offers moderately accurate fence locations and density over a large spatial scale. In addition, regional rules can be created to hone the methodology to specific states or provinces of interest. Our accuracy varied slightly depending on whether croplands and BLM lands were included in the model. Because these decision rules were based on expert opinion and experience, and our experts were not as confident with large areas of cropland and BLM lands, we believe that adding experts from these areas may further improve model accuracy. In some areas of the West, fences are used on BLM lands. Since we assumed the BLM pasture boundaries were fences on BLM land, accuracy may be increased when other fence types are included in the model. We also assumed that fencing on areas of croplands would be different than non- cropland areas, however, the increase in accuracy gained when removing the croplands from the model may suggest otherwise. Accuracy was lowest when fencing associated with large croplands were included in the fence dataset, but we believe that our model overall is an accurate reflection of fencing on the landscape. Improvements in our analysis methods may improve our overall accuracy of the fence dataset. Different fence structure types have different effects on wildlife movement and habitat selection and the fence layer created here did not include fence structure data. Fence structure type is difficult to model over large regions and is more a result of private landowner's preference and type of livestock production per ranch. Wire mesh fencing has been used in this study area to corral sheep and may be particularly hard for wildlife to cross. Additional improvements in accuracy of the fence layer may be made by amending the fence sampling protocol. Fence surveyors at times recorded fence locations at as much as 100 m from the actual fence due to railroad right-of-ways and property rights and interior pasture fences off roadways were not ground- truthed. Future sampling forays are planned to sample more of the study area, which will be imperative for future landscape fence permeability analyses for wildlife. Finally, because this data was created with the continued reliance on assumptions about neighboring parcels of land, there may be a decrease in accuracy around the edges of the study area. We recognize that individual fences may not be accurately modeled. Although improvements may be made, this novel effort uses extractable methodologies we believe will assist in modeling a key variable towards unraveling wildlife movement and habitat selection. Fences can exhibit both indirect and direct effects on native wildlife populations worldwide. In North America, indirect effects of fencing such as animal displacement, reduced habitat availability, and habitat fragmentation may have a higher impact on pronghorn and other wildlife populations than direct effects, by altering behavior and movement rates resulting in eventual population decline. Concerning pronghorn, and in particular during sever winter conditions, snow and ice can accumulate over the bottom-most fence wire during winter, thus preventing pronghorn from crawling underneath fences. Because of this, fencing may prevent migrating pronghorn from reaching higher quality habitat, during which time they have expended energy without finding better conditions. Fencing therefore can alter behavior at multiple scales and place North American wildlife in perilous situations as it has in other areas of the world. A regional fence layer allows both wildlife and land managers to assess effects to wildlife at various scales, including at home range and within home range level of habitat selection, as well as identifying important population-level movement pathways between seasonal ranges during migration. Certainly, it can aid researchers through inclusion into predictive modeling efforts to assess habitat suitability and connectivity at regional level scales. As a priority, on-the-ground management practices could identify high fence densities (here along the developed Highway 2), along migratory pathways and within breeding grounds; ecological necessities to sustain North America's dwindling grassland wildlife. Federal, state and provincial agencies, along with non-profit organizations and community organizations all can play an important role by undertaking cooperative projects to modify fences in strategic locations. These could include key-linkage areas which are geographic or anthropogenic areas and/or critical stopover locations along the migration pathway and fawning and/or lek areas. Without planning and the proper data, the cumulative anthropogenic changes to landscapes will continue to erode wildlife habitats and seasonal migration opportunities, reducing effective habitat patch size, potentially leading to sustained population declines and contraction of overall species range. # Supporting Information We thank Lee Smith and Carrie Breneman in Alberta, and Samantha Howlett in Montana for their many hours in the field and the numerous people we consulted to develop our fencing assumptions: Kelvin Johnson, Scott Thompson, Mark Sullivan, Adam Messer, MT FWP; John Carlson, Steve Zellmer, Abby Hall BLM; Dennis Jorgensen, WWF; Randy Matchett, USFWS; Paul Jones ACA; Matt Lopez, US BIA; John Doney, Ft. Peck Indian Reservation; Hoyt Richards, MT DNRC. We also thank our two anonymous reviewers at PLoS ONE. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: EEP AJ CL MS. Performed the experiments: EEP AJ. Analyzed the data: EEP AJ. Contributed reagents/materials/analysis tools: EEP AJ CL MS. Wrote the paper: EEP AJ CL MS.

# Introduction It is estimated that one third of the world's population is infected with *Mycobacterium tuberculosis* (*M. tb.*). In most infected individuals the host response contains this infection, or may result in pathogen clearance. However, the global burden of disease remains a significant problem. In 2007, there were estimated to be 9.2 million new cases of tuberculosis (TB) and in 2006, there were 0.5 million cases of multi-drug resistant TB. The only available vaccine against TB, *Mycobacterium bovis* Bacille Calmette-Guerin (BCG), has variable efficacy. When BCG is administered at birth, it confers consistent and reliable protection against disseminated disease in childhood in the developing world. However, BCG fails to protect against pulmonary disease in these regions. Any improved vaccine regime against TB should include BCG, administered in infancy, in order to retain the protective effects against severe childhood disease. Revaccination with BCG in adolescence has been routine practice in many countries throughout the world, with evidence to suggest that it confers improved protection against diseases such as TB meningitis and leprosy. In a large clinical trial of BCG revaccination with over 200,000 children in Brazil, no enhanced protective effect of boosting with BCG against pulmonary disease was observed. In light of this and other studies, revaccination is not recommended by the WHO. The immunological markers of protection against TB have not yet been defined. However, immunity to *M.tb*. is dependent on the generation of a Th1-type cellular immune response. Secretion of IFNγ is central to the activation of *M.tb*-infected macrophages and the measurement of IFNγ release from antigen- specific T cells provides the best available immunological correlate of protection against TB. TB is the commonest HIV-related disease and the most frequent cause of mortality. The annual risk of developing TB disease from reactivation of latent *M.tb* infection in those co-infected with HIV is 5–15% annually, compared to a 5–10% lifetime risk in immunocompetant individuals. In addition to the importance of CD4+ T cells in protective immunity, CD8+ T cells are also important in the control of *M.tb* infection. CD8+ T cells may be of greater importance during the latent stages of *M.tb* infection, by maintaining control of slowly replicating bacteria, rather than in acute infection. Improved induction, activation, or functionality of TB-specific CD8+ T cells may enhance protective immunity against TB. We have developed a new vaccine, Modified Vaccinia Ankara expressing antigen 85A (MVA85A) using a recombinant viral vector system that is designed to induce and amplify the cellular immune response. This vaccine has been developed as a booster vaccine for BCG. Antigen 85A is an enzyme (mycolyl-transferase) involved in cell wall biosynthesis, is highly conserved in all mycobacterial species, and is present in all strains of BCG. Antigen 85A is immunodominant in animal and human studies, and immunization with antigen 85A conferred significant protection against *M.tb* challenge. Boosting BCG with MVA85A induces high levels of antigen specific IFNγ secreting CD4+ and CD8+ cells and enhanced protection against *M.tb* challenge when administered intranasally to mice. In larger animals enhanced protection was observed when MVA85A was administered intradermally to boost BCG in non human primates (Verreck F et al, submitted) and also to cattle where protection against *M. bovis* challenge was enhanced (Vordermeier M et al, submitted). In a series of clinical trials we have shown that MVA85A is effective at inducing high levels of antigen specific CD4+ T cells when given alone or as a boost to BCG induced immune responses. However, detecting TB-specific CD8+ T cells in peripheral blood mononuclear cells (PBMC) samples from these clinical trials proved to be difficult. Using *ex vivo* ELISpot assays and PBMC based flow cytometry, no antigen specific CD8+ T cell responses were detected either before or after MVA85A vaccination in any of the studies to date. In other work, dendritic cells (DC) have been employed as professional antigen presenting cells, to increase the sensitivity of CD8+ T cell detection. This is due to the high levels of MHC I and MHC II, co- stimulatory molecules and relevant cytokine profile expressed by mature DC which facilitates expansion of antigen-specific T cell population that otherwise may not be detected. The aim of the clinical trial presented here was to investigate the safety and immunogenicity of boosting BCG vaccinated subjects with BCG, and to compare these results with data from previous trials where BCG vaccinated subjects had been boosted with MVA85A. In addition, the particular immunological focus was to establish a methodology for detecting antigen specific CD8+ T cell responses in these subjects. # Methods Consort flowcharts for each of the two trials discussed here are presented in. The protocol for this trial and supporting CONSORT checklist are available as supporting information; see and. ## Participants Subjects were recruited for both clinical trials under protocols approved by the Oxfordshire Research Ethics Committee and enrolled only after obtaining written informed consent. Follow up was for 24 weeks; however protocols were amended during the trial so that subjects still living in the area were invited for a week 52 assessment. The trials reported here were open label observational trials. Subjects enrolled in the clinical trials were all healthy and had previously been vaccinated with BCG. This was confirmed visually by the presence of a BCG scar. They were all seronegative for HIV, HBV and HCV at screening, and routine laboratory haematology and biochemistry were performed prior to vaccination. All values were within normal limits. All subjects had a Heaf test and anyone with a Heaf test score of greater than Grade 2 was excluded. All subjects were negative for the *M.tb* -specific antigens ESAT6 and CFP10, as determined by an *ex vivo* ELISpot assay. There were 14 subjects eligible for enrolment into the BCG-BCG trial with a median time interval of 12 years from their first BCG vaccination. The median age of subjects was 28 years (range 19 to 51 years). Subjects in the BCG-MVA85A trial were from a previously reported study (n = 17), with an additional 4 subjects subsequently recruited into the BCG-MVA85A group. All 21 subjects were used in the analysis for this paper. Similar inclusion/exclusion criteria were applied to subjects recruited to the BCG-BCG and BCG-MVA85A studies. The median interval period between BCG vaccination and boosting with MVA85A was 18 years in this group, and the median age of subjects was 31 years (range 22 to 54 years). Demographic information on all the subjects is summarised in. All subjects completed a diary card recording local and systemic adverse events and body temperature for 7 days following vaccination. In addition, solicited adverse events were collected at each clinic visit. ## Interventions In the BCG-BCG study, volunteers were vaccinated with BCG (a single immunisation with BCG (BCG Denmark, Statens Serum Institute, Copenhagen, Denmark), 100 µl administered intradermally over the deltoid region (n = 14)). In the BCG-MVA85A study, subjects were vaccinated intradermally with a single immunisation of 5×10⁷ pfu MVA85A. ## Immunogenicity Analyses ### ELISpot Analysis A standardised *ex vivo* IFNγ ELISpot assay was performed on PBMC taken at the following time points: at screening (prior to the Heaf test), and then at 1, 4, 8, 12, and 24 weeks after vaccination on fresh PBMC from subjects in both trial groups. Blood was also taken at 52 weeks after vaccination in those subjects still residing in the area. Briefly, 300,000 fresh PBMCs per well were plated directly onto the ELISpot plate (MAIP, Millipore) in the presence of either Tuberculin Purified Protein Derivative (PPD) at 20 µg/ml, purified antigen 85 complex (10 µg/ml), or antigen 85A pooled peptides (7 pools of 9–10 15-mer peptides spanning the length of antigen 85A, which overlapped by 10 amino- acids). The final concentration of each peptide in the well was 10 µg/ml. The PBMC were incubated with antigen (PPD, antigen 85, or antigen 85A peptide pools) for 18 hours at 37°C. Streptokinase/Streptodornase and PHA were used in all assays as positive controls and cells and media alone as the negative control. Assays were performed in duplicate and the results were averaged. The ELISpot data were analysed by subtracting the background counts (mean number of spots in the medium and cells alone control wells) from the mean number of spots in wells with antigen and cells, using an AID ELISpot 04 reader (AID Diagnostika GmbH, Strassberg, Germany). Counts of less than 5 spots/well were considered as negative. A well was considered positive if the count was at least twice that in the negative control wells and at least 5 spots more than the negative control wells. For the peptide pool wells, the results were summed across all the peptide pools (SPP) for each volunteer, at each time point. This will count twice a T cell that responds to any of the 10-mer overlap regions that occur in two pools with adjacent peptides, as each pool contains non- overlapping peptides. An area under the curve (AUC) analysis for IFNγ ELISpot responses was performed to compare BCG - BCG and BCG - MVA85A vaccine regimens over time (up to 52 weeks). ## Dendritic Cell Mediated CD8+ T Cell Amplification The flow cytometry assays were performed on cryopreserved PBMC obtained from subjects in both the BCG-BCG and BCG-MVA85A groups at screening (prior to the tuberculin skin test), and then at 1, 4, 8, 12 and 24 weeks after vaccination allowing comparability between groups. In this assay, PBMC were co-cultured with monocyte derived dendritic cells (DC) to enhance the detection of *ex vivo* CD8+ T cell responses. DC were produced using the standard method of culturing with IL-4 and GMCSF. At day 5 they were activated with 10 ng/ml LPS. At 16 hour post-activation, the DC were washed, peptide-pulsed for 45 minutes, and washed again. Preliminary experiments performed with DC pulsed with BCG showed no detectable CD8+ T cell response in the BCG-BCG regime and so this was not investigated further. PBMC were co-cultured in 24-well plates at a ratio of 15∶1 (PBMC∶DC) in the presence of IL-7 at 25 ng/ml on day 1, then IL-2 was added at 0.5 ng/ml on day 3, and subsequently on day 7 and day 10. The culture was washed on day 13 and peptide re-stimulation (10 µg/ml) was performed on day 14, alongside PHA and unstimulated cells. Brefeldin A was added after 4 hours and after a further 12 hours intracellular cytokine staining was performed to detect IFNγ using the Becton Dickinson intracellular cytokine staining kit with GolgiPlug™, according to the manufacturer's protocol. An HLA-A2 specific CD8+ immunodominant antigen 85A peptide pentamer, KLIANNTRV was selected as this has previously been shown to be recognised in BCG vaccinated subjects. A previously defined HLA-A2 restricted CD8+ T cell epitope was selected as HLA-A2 is the most common HLA Class I antigen and this therefore enabled us to analyse the maximum number of subjects using a single epitope, KLIANNTRV. To evaluate and quantify CD8+ T cell responses, 5 subjects in the BCG-BCG trial and 8 subjects in the BCG-MVA85A trial all of whom were HLA-A2 positive were analysed. Tetramers for EBV (EBNA 3A, HLA-B35-restricted YPLHEQHGM) and HIV epitopes (gag; p17 HLA-A24-restricted KYRLKHLVW and p46 HLA-A6802-restricted ETAYFILKL) were used as a positive and negative control (for short term culture conditions), respectively.All antigens were used at 10 µg/ml. Cells were incubated with KLIANNTRV pentamer labelled with APC for 60 minutes at 37°C. Cells were then cooled to 5°C and incubated with PE-conjugated CD8 mAb (Becton Dickinson) for 30 minutes, washed and incubated with FITC-conjugated IFNγ and PerCP-conjugated CD3 mAb (Becton Dickinson) for 30 minutes at 5°C. Cells were acquired using flow cytometry on a FACS Calibur (Becton Dickinson) machine and were analyzed using FlowJo software by gating on the lymphocyte population in forward scatter (FSC) and side scatter (SSC), and gating on live cell populations. Gating was then performed for CD8^HI+ population to determine if they were antigen specific and producing IFNγ. These gated statistics were used and the specific responses (i.e. the percentage of positive stimulated cells minus the percentage of positive unstimulated cells) were displayed in a histogram. The data shown are representative of triplicate experiments. ## Statistical Methods A Mann-Whitney test was used for all comparisons between groups, areas under the curve and comparisons between early and late time points (week 1 and week 52) between groups, and the Binomial method was used to estimate the confidence intervals using STATA 9 # Results ## Recruitment and Demographics Subjects were recruited from February 2004 to November 2005 for the BCG-BCG group, and from March 2003 to June 2006 for the BCG-MVA85A group. All subjects were followed up for safety and immunogenicity for 24 weeks after vaccination, with an additional visit at 52 weeks for 12 subjects from the BCG-BCG regime and 13 subjects from the BCG-MVA85A regime. There was 1 (7%) of 14 subjects in the BCG-BCG group born outside the UK, whereas 7 (33%) of 21 subjects in the BCG- MVA85A group were born outside the UK. Three (21%) of BCG-BCG subjects had significant travel history compared with 2 (10%) of BCG-MVA85A subjects. ## Safety of BCG-BCG Vaccination Compared With BCG-MVA85A Vaccination The BCG-BCG and BCG-MVA85A regimens were well tolerated with no serious or severe vaccine related systemic adverse events. There were no marked differences in local and systemic adverse events between BCG-BCG (n = 14) and BCG-MVA85A (n = 21) vaccination regimens. All subjects in the BCG-BCG and BCG-MVA85A regimens experienced some mild local adverse events related to the vaccination. This included redness, pain and induration that was experienced by all subjects in both regimens, with some subjects also experiencing pruritus. All systemic adverse events reported were mild in severity. The most common systemic adverse events were headache (5 (36%) of 14 subjects receiving BCG-BCG and 8 (38%) of 21 subjects receiving BCG-MVA85A), and myalgia (5 (36%) of 14 subjects receiving BCG-BCG and 5 (24%) of 21 subjects receiving BCG-MVA85A). The adverse event profiles were comparable between the two vaccine regimens. ## BCG-MVA85A induced significantly stronger cellular immune responses to antigen 85 protein and antigen 85A SPP than BCG-BCG vaccination The immune responses to PPD, antigen 85, and antigen 85A SPP were compared between subjects in the BCG-BCG and BCG-MVA85A trials in PBMC at various time points over the 52-week post-vaccination period. In the BCG-BCG subjects, a strong cellular immune response to PPD was elicited at 1-week post-vaccination, resulting in an increase in IFNγ ELISpot responses (median 824 spot forming cells (SFC)/10⁶ peripheral blood mononuclear cells (PBMC)) above levels observed at screening (median 205 SFC/10⁶ PBMC;). The cellular immune response was lower by 12 weeks post-vaccination (median 467 SFC/10⁶ PBMC), and reached a plateau at these levels which persisted for at least 52 weeks after vaccination (median 402 SFC/10⁶ PBMC). Similarly, in BCG-MVA85A subjects, the peak response to PPD was observed 1-week post-vaccination (median 783 SFC/10⁶ PBMC). These levels dropped at 12 weeks post-vaccination (median 223 SFC/10⁶ PBMC), increasing marginally at 52 weeks (median 347 SFC/10⁶ PBMC). The response to antigen 85 protein in the BCG-BCG subjects was lower than the PPD response; peaking at 1 week post-vaccination (median 340 SFC/10⁶ PBMC) from screening (median 105 SFC/10⁶ PBMC). A low level response to this antigen continued to be observed throughout the 52-week post-vaccination period. In contrast, there was a marked increase in response to antigen 85 protein in the BCG-MVA85A subjects at 1-week post-vaccination (median 790 SFC/10⁶ PBMC) over values observed at screening (median 27 SFC/10⁶ PBMC;). The response to antigen 85A declined up to 12-weeks post-vaccination (median 113 SFC/10⁶ PBMC), and response levels were marginally higher at 52 weeks (median 237 SFC/10⁶ PBMC). In BCG-BCG subjects, the response to SPP increased at 1 week post-vaccination (median 236 SFC/10⁶ PBMC) compared with screening (median 102 SFC/10⁶ PBMC). The response was reduced to baseline levels (median 95 SFC/10⁶ PBMC) at 8-weeks post-vaccination, where this plateau continued to 52 weeks (median 34 SFC/10⁶ PBMC). In BCG-MVA85A subjects, the response to SPP increased markedly at 1 week post-vaccination (median 2147 SFC/10⁶ PBMC), compared with values at screening (median 13 SFC/10⁶ PBMC). The response to SPP declined by 12 weeks (median 390 SFC/10⁶ PBMC), and then increased marginally by 52 weeks post- vaccination (median 506 SFC/10⁶ PBMC). To analyse the effect of the vaccine regimens over the 52-week post-vaccination period, an area under the curve (AUC) analysis of the IFNγ ELISpot responses was performed on subjects present at all time points (1, 2, 4, 8, 12, 24, and 52 weeks). The results demonstrated that the boosting effect of MVA85A (n = 13) was significantly greater than the booster effect of BCG (n = 12) in enhancing the cellular response to antigen 85A (p = 0.009) and SPP (p = 0.0001) over the 52-week period ( and). The response to PPD was not significantly different between the BCG–BCG and BCG–MVA85A groups. ## 1 Week Post-Vaccination To examine the significance of the peak effect of the vaccination regimens at 1-week post-vaccination, statistical analysis of the cellular immune responses to PPD, antigen 85A protein and SPP was performed (BCG∶BCG n = 21 BCG∶MVA85A n = 14). A strong cellular immune response was observed for PPD-specific PBMC in both BCG-BCG and BCG-MVA85A subjects at 1 week after vaccination (median 707 SFC/10⁶ PBMC and 824 SFC/10⁶ PBMC, respectively), with no significant differences in response to PPD between the 2 groups (p = 0.95;). In contrast, at 1 week after vaccination, a significantly higher response for PBMC from BCG-MVA85A subjects was observed against antigen 85A protein (median 790 SFC/10⁶ PBMC) and SPP (median 2147 SFC/10⁶ PBMC) than in PBMC from BCG-BCG subjects (median 340 SFC/10⁶ PBMC against antigen 85A protein, p = 0.002, and median 236 SFC/10⁶ PBMC against SPP, p\<0.0001;,). ## 52 Weeks Post-Vaccination To determine the level of cellular immune responses at the end of the follow-up period, 52-week post-vaccination samples were examined, (BCG∶BCG n = 12, BCG∶MVA85A n = 13). Statistical analysis of subjects at 52-weeks post- vaccination demonstrated a moderate response to PPD in PBMC from BCG-BCG subjects (median 402 SFC/10⁶ PBMC) and BCG-MVA85A subjects (median 347 SFC/10⁶ PBMC), with no significant difference. A significantly greater antigen 85A specific cellular immune response was detected in PBMC from BCG-MVA85A subjects (median 237 SFC/10⁶ PBMC to recombinant protein and median 506 SFC/10⁶ PBMC to the SPP) than in PBMC from BCG-BCG subjects (median 83 SFC/10⁶ PBMC against antigen 85A, p = 0.0003, and 34 SFC/10⁶ PBMC against SPP, p = 0.0001; and). ## Dendritic cell mediated amplification of IFNγ secreting, antigen 85A peptide specific, CD8+ T cells from PBMC In the five HLA-A2+ BCG-BCG subjects studied in this analysis, only one subject had very low levels of IFNγ secreting, CD8+ T cells recognising KLIANNTRV (1.5% of total CD8 population)at one time point (4-weeks) post-vaccination. In contrast, in the eight HLA-A2+ BCG-MVA85A subjects, IFNγ secreting, CD8+ T cells recognising KLIANNTRV were detectable at 2- and 4-weeks post-vaccination in 4 (50%) of 8 subjects, at levels up to approximately 10-fold higher than those observed for BCG-BCG subjects (up to 14% total CD8 population). # Discussion In this current study, revaccination with BCG was well tolerated with no unexpected or serious adverse events up to 52 weeks post-vaccination. This was similar to the adverse event profile observed in subjects in a previously reported trial using the BCG-MVA85A vaccine regimen. Vaccination with BCG has variable efficacy, failing to protect against pulmonary disease in adults, whilst affording protection in childhood. The most widely accepted explanation for the variable efficacy of BCG in many global trials is that prior exposure to environmental mycobacteria either inhibits the replication and development of BCG induced protective immune response or masks the effect of BCG vaccination by providing a similar degree of anti- mycobacterial immunity. In this current study, the differences in cellular immune responses between these two vaccine regimens could not be explained by possible differences in the levels of mycobacterial antigen exposure at baseline. To confirm there was no difference between mycobacterial antigen exposure between trials, the trial inclusion criteria, including baseline Tuberculin skin test responses, were the same for both trials. The results from this study provide evidence that there are elevated cellular immune responses to *M.tb* antigens following BCG-BCG and BCG-MVA85A vaccine regimens. The IFNγ ELISpot response to PPD in BCG-BCG subjects was high, and similar to BCG-MVA85A subjects at both early (week 1) and late (week 52) time points, and over the 52 weeks period (as determined by AUC values), with no significant differences observed between the two vaccine regimes. Boosting BCG with BCG may result in boosting of many of the different secreted antigens within PPD, whereas boosting with MVA85A will only result in boosting of the antigen 85A component of BCG. In contrast to the PPD responses, in this study, we have shown that subjects vaccinated with MVA85A have significantly higher cellular immune responses to antigen 85A protein and antigen 85A SPP at peak (week 1) and plateau (week 52) time points, and over the 52 week period (as determined by AUC values) than in BCG-BCG subjects. Studies have demonstrated that reactivity against antigen 85A and antigen 85B protein , invoke a protective immunity to TB. Immunization with a vaccine construct that encodes antigen 85A confers protection against challenge with live TB in small animals. Although significant differences were observed using the *ex vivo* IFNγ ELISpot assays, this assay only detected CD4+ T cell responses. No CD8+ T cell responses have been identified before this current study. Therefore, we used DC to amplify the low numbers of CD8+ T cells that were otherwise not detectable, and compared between boosting regimens. An HLA-A2 restricted immunodominant peptide (KLIANNTRV) from antigen 85A was selected to identify MHC class I-restricted, CD8+ T cells. KLIANNTRV has been shown to be recognised by PBMC stimulated with BCG in 50% of BCG-vaccinated subjects. Following amplification of CD8+ T cells using DC, we observed that revaccination with BCG induced detectable, but low frequency of pentamer (KLIANNTRV)+, IFNγ+, CD8+ T cells at a single time point in only one of 5 HLA-A2+ subjects. The magnitude (up to 10-fold higher) of CD8+ T cell recognition of the immunodominant KLIANNTRV peptide was higher in the BCG-MVA85A subjects, was detectable in 4 of 8 subjects and detectable at 2 different time points (in 50% of subjects). It is clear that a higher frequency of CD8+ T cells can be detected in the BCG-MVA85A regimen than in the BCG-BCG regimen using this highly sensitive assay. The clinical significance of this enhanced CD8+ T cell response can only be determined in large scale efficacy trials. It is possible that the levels of antigen-specific CD8+ T cell responses remain below the limit of detection in BCG-BCG subjects, but given the increased sensitivity of the DC-mediated amplification method used in this study, this would imply they are at very low frequency indeed. The results of this study demonstrate that DC are a useful and powerful tool in providing the stimulation required to amplify the low frequency CD8+ T cells in PBMC samples to detectable levels. CD8+ T cell responses to mycobacteria have been shown to be induced through vaccination with BCG in older subjects, or following latent infection or active disease. However, the widely held view is that BCG vaccination results in suboptimal CD8+ T cell responses. Although CD4+ T cells are the predominant cells controlling TB infection, there is evidence in animal studies that CD8+ T cells are involved in protection against TB. Protection against *M. tb*, an intracellular pathogen, probably requires a finely balanced interplay between cells of the innate and adaptive immune system. CD8+ T cell responses may play role in the initial control of TB infection and this may be elicited by presentation from neutrophils during the initial innate immune response, or through interactions with neutrophils and DC. Neutrophils have been shown to cross-present pathogen-derived and soluble antigens to CD8+ T cells *in vivo*. This may go some way to explain the differences observed in whole blood assays where neutrophils are present and may initiate CD8+ T cell proliferation, compared with PBMC preparations where neutrophils are not present. Indeed, it has been shown that CD8+ T cells can be detected by flow cytometry in whole blood samples from adults, and infants, vaccinated with BCG. Although pentamer (KLIANNTRV)+, IFNγ+, CD8+ T cells could be detected in the BCG-BCG and BCG- MVA85A subjects in our study, it is possible that other, functional, CD8+ T cells exist as we have shown in subjects who received BCG-MVA85A. As these cells occur at low frequency, it may be possible that the use of whole blood samples coupled with DC amplification could be used as a powerful tool for amplifying low frequency CD8+ T cells to asses their functionality. In summary, the results in this study have demonstrated that in the BCG-MVA85A vaccination regimen, significantly stronger cellular immune responses to antigen 85A can be induced compared with revaccination with BCG-BCG. IFNγ is clearly an important cytokine involved in protection against TB and has been used as a measure of vaccine “take” in Phase I BCG-MVA85A clinical trials. As further research is performed, it is highly likely that other correlates of protection will emerge. Proofs of concept efficacy trials are due to start in early 2009, and will enable us to evaluate whether enhanced immune responses following BCG- MVA85A result in improved protection. # Supporting Information We thank all the subjects who took part in the studies reported here. We thank Kris Huygen for providing purified antigen 85 for use in these studies. Oxford University was the sponsor for all the clinical trials reported here. [^1]: Conceived and designed the experiments: KTW AVH HM. Performed the experiments: KTW AAP. Analyzed the data: KTW AAP HF NCA HM. Contributed reagents/materials/analysis tools: CRS HF IP. Wrote the paper: KTW HM. [^2]: AAP, AVSH and HMcS are all co-inventors on a composition of matter patent for MVA85A and are shareholders in a Joint Venture established to develop this vaccine.

# Introduction Schizophrenia is a serious mental illness and a chronic disease for most patients. Patients with schizophrenia often suffer from somatic comorbidities and have a 15–25 year reduced life-expectancy\[–\]. Patients fare best when in continuous, integrated healthcare, which consists of psychiatric and somatic care. This may help to prevent psychotic relapse, crisis treatment, hospitalization, and early mortality\[–\]. The Netherlands has a well-established care system for patients with schizophrenia\[–\]. Patients can choose any healthcare provider. All inhabitants in The Netherlands have compulsory health insurance and are free to choose and have to be accepted by any health insurance company (regardless of their health or income)\[\]. 73% of patients in a cohort received continuity of care in 2009–2011. Rising costs of healthcare in The Netherlands, but also in many other countries, led to re-evaluation of the way healthcare is financed\[,–\]. The Netherlands has a long-standing policy with universal access to a wide range of evidence-based mental health treatments and few financial barriers to such care. Co-payments for patients in The Netherlands were traditionally modest compared to other countries but are rising. In the period between 2012 and 2013 health-related co-payments were raised significantly, for everyone, from €155 to €360 euro. In addition, during 2012 a separate co-payment for psychiatric care was charged. Although there is evidence that the effects of co-payments are smaller among patients with severe health problems, there is also evidence that effects are larger among patients with mental disorders\[–\]. Continuity of elective outpatient care is essential for patients with schizophrenia. Co-payments may disrupt this. Unwanted effects may include a rise in use of acute psychiatric care (inpatient or crisis care). The introduction and rise of co-payments offered a natural experiment to study the effects of co-payments on healthcare use of patients with schizophrenia. Other factors than co-payments which influence healthcare use have to be considered, for instance the national policy to substitute inpatient care with elective planned outpatient treatment. A previous paper has shown that, overall, the use of mental healthcare declined substantially after the introduction of co-payments and that this effect was stronger for patients with lower incomes. Patients with schizophrenia often have very low incomes\[–\]. Patients with schizophrenia may be even more vulnerable to unwanted effects of co-payments because a substantial proportion of schizophrenia patients has decreased insight in their illness and/or is reluctant to accept treatment. We expected that (i) overall there would be a decline in the long-term use of psychiatric care and (ii) tested whether there were significant deviations on top of the long-term declining trend that correlated temporally with the abrupt rise in co-payments in 2012 in the Netherlands. The deviations we expected were a relative decrease in continuous outpatient psychiatric care and a relative increase in acute psychiatric care (inpatient or crisis treatment) among patients with schizophrenia. # Methods ## Study design and patient selection Computerized registry data of the largest Dutch health insurer (Zilveren Kruis) were collected for all insured persons. All patients insured by Zilveren Kruis with a diagnosis of schizophrenia in 2008 under 70 years of age were selected. Mental and somatic healthcare use, including prescription data of antipsychotic medication, of these patients over 2009–2014 were analyzed in a retrospective, longitudinal registry-based cohort study. Zilveren Kruis provided health coverage for about 30% of the 16.4 million residents in the Netherlands in 2008. Those insured by Zilveren Kruis were representative of the Dutch population. In 2008, the first year mental health care was provided under the Dutch Health Insurance Law, there were 15 552 patients who had claims with the diagnosis schizophrenia. Excluded were: 907 (5.8%) patients who were 70 years or older on January 1 2008, 1041 (6.7%) patients who died and 2 693 (17.3%) patients who were not insured the whole period by Zilveren Kruis. 10 911 (70.2%) patients, under 70 years of age and insured during the whole study period by Zilveren Kruis, were included in the analysis. 61% were men (mean age 40.2, SD 11.4) and 39% were women (mean age 46.0, SD 12.3). ## Data source: Dutch computerized health insurance registry data All data are derived from Zilveren Kruis health insurance registry data. Dutch health insurance companies thoroughly process and pay the claims for all healthcare which is covered by the Dutch Health Insurance Law. The claims process is regulated by the National Care Authority. Dutch health insurers have implemented and maintain very strict rules and regulations about privacy of their insured and their healthcare providers according to prevailing law in the Netherlands. The selection and analysis of the necessary data for this study took place according to these rules. The analysis was carried out with data through which individual patients could not be identified. Therefore, no informed consent nor approval of a Medical Ethical Committee was needed. The Zilveren Kruis database contains data concerning all the care, as covered by the Dutch Health Insurance Law, received by these patients from all their healthcare providers. The information concerning diagnoses is limited. The main groups of DSM-IV diagnoses are registered in the registry, not the detailed codes of the schizophrenia spectrum. The diagnoses of schizophrenia in this study were registered by the treating psychiatrist and pertain to the time of inclusion and the period covered by the study. Although registry based diagnoses can be unreliable, recent work shows that diagnoses of schizophrenia are sufficiently reliable for use in in study such as ours. Separate treatments for more than one psychiatric disorder will result in separate claims. For short-term treatment no diagnosis has to be provided. A psychiatric crisis situation has to be registered on the claim. The maximum duration for such a claim is 28 days, for all other claims the maximum duration is one year. The claims for outpatient medication, which were prescribed and picked up at the pharmacy, include the name, price, and the Defined Daily Dose (DDD) of every medication. Information on inpatient medication is not available in the registry dataset. ## Co-payments In the Netherlands the level of co-payment is low compared to other countries, but the level of co-payment has risen from €155 in 2009 to €360 in 2014 (Figs). Individual co-payments per patient are not available in our dataset. There were no co-payments concerning visits to a general practitioner. Only during 2012 extra co-payments for psychiatric care had to be paid. These extra co-payments were €100-€200 per year for psychiatric treatment and €145 per month of inpatient care starting at the second month of treatment. Crisis treatment, involuntary treatment, and treatment of patients under 18 years of age were exempted from these co-payments. ## Measures We evaluated the association between the level of co-payments and healthcare received over the years 2009–2014. Healthcare as covered by the Health Insurance Law was divided in (i) elective outpatient psychiatric care (non-crisis psychiatric outpatient care, antipsychotic medication), (ii) acute psychiatric care (psychiatric crisis treatment and psychiatric hospitalization), and (iii) somatic care. All measures were aggregated per quarter of the years 2009–2014. The quarters per year will be referred as q1-q4, e.g. the third quarter of 2010 as 2010q3. Other measures are the number of patients starting new psychiatric treatments per quarter and the average amount of psychiatric care per patient, reflected by the costs of care. The costs of psychiatric care were calculated using average national prices in euros and allocated to the quarter of the starting date. The costs were adjusted for inflation using price indices for psychiatric care with 2009 as basis. Medication was assigned to a quarter based on the pick-up day of that medication. The average amount of medication is given in Defined Daily Dose (DDD). Somatic care, including all non-antipsychotic medication, was assigned to a quarter based on the starting day of the somatic treatment. The average amount of somatic care, reflected by the costs of care, was calculated and adjusted with cost indices for somatic care derived from all insured by Zilveren Kruis with 2009 as basis. Continuity of elective care was defined as receiving elective psychiatric care during every quarter from 2009–2014. ## Analysis The aim of our analysis was to test whether the abrupt and substantial rise in co-payments from 2012 on was associated with changes in the elective and acute care. Therefore, we analyzed trends in care consumption and deviations from these trends. First, the trends in elective psychiatric care received by the patients in the cohort per quarter of the year were analyzed over the years 2009–2014, followed by the trends in acute psychiatric care, and somatic care. Next, the trends in psychiatric care for the groups with and without continuous elective care were examined. When changes in trends of care usage over time were observed these were tested by Box-Jenkins autoregressive integrated moving average (ARIMA) models. These models describe temporal changes in trends (for an explanation please see). All analyses were performed with SAS Enterprise guide 6.1 \[SAS Institute Cary, NC, USA\]. # Results ## Elective psychiatric care Although the amount of received elective psychiatric care declined over follow- up, the majority of patients with schizophrenia (59%) remained in elective outpatient care over the whole period. However, the percentage of patients in a quarter receiving outpatient psychiatric care plus antipsychotic medication declined from 65% to 56% (-14%) per year over 2009–2014 (Figs). Outpatient psychiatric care without antipsychotic medication declined from 15% to 9% (-42%). The percentage of patients that received only antipsychotic medication increased from 10% to 17% (67% increase). Patients receiving no elective outpatient psychiatric care and no antipsychotic medication increased from 9% to 18% (99% increase). Taken together the percentage of patients with any form of elective care in a quarter declined from 91% to 82%. Outpatient psychiatric care decreased, both with regard to newly initiated treatments (-23%) and in the total amount of care (-30%). The amount of antipsychotic medication used as reflected in DDD remained almost the same over the period. Next, we analyzed if abovementioned trends were constant or whether there were deviations from these trends. First, the increasing trend in percentage of patients treated with only antipsychotic medication shifted to a higher level in 2012q1 and in 2014q3 ( and Tables). Second, the decreasing trend in the number of patients starting new elective outpatient psychiatric care shifted to a higher level from 2012q4-2013q3. Third, the decreasing trend in amount of elective outpatient psychiatric care per patient shifted to a higher level in 2012q4. ## Acute psychiatric care The percentage of patients needing acute psychiatric care declined from 20% to 13% (-37%) over 2009–2014. The relation between elective psychiatric care and acute psychiatric care was examined. The levels of acute psychiatric care among patients with elective outpatient care, with or without antipsychotic medication, per quarter are under 15% average per year over 2009–2014 and declined with 16% from 2009–2014 (Figs). The level of acute psychiatric care in quarters with only antipsychotic medication was very high with an average of 83% in 2009 and declined to 34% in 2014 (-59%). In quarters without elective psychiatric care the average level of acute psychiatric care was also very high in 2009 (49%) declining to 12% (-76%) in 2014. The subgroup of patients with elective psychiatric care during all quarters needed less than half the levels of acute psychiatric care as compared to the other patients over 2009–2014 (Figs). Acute psychiatric care decreased in both newly initiated treatments (-24%) and in the total amount of care (-29%). Several deviations of abovementioned trends occurred. The percentage of patients with acute care showed an increased level shift from 2012q2 ( and Tables). The patients with elective outpatient care plus antipsychotic medication shifted to a higher level of acute psychiatric care in 2012q1 and to a lower level in 2013q4. The patients with elective outpatient care without antipsychotic medication shifted to a higher level of acute psychiatric care in 2012q3. Patients with only antipsychotic medication also shifted to a higher level in 2012q3. The subgroup of patients with quarters without elective psychiatric care shifted to a higher level of acute psychiatric care in 2012q1. The number of patients starting a new episodic treatment shifted to a higher level in 2012q2 and to a lower level in 2013q4. The amount of acute psychiatric care had an increased level shift from 2012q1 and a downward level shift from 2014q1. ## Somatic care The average amount of somatic care, including non-antipsychotic medication, per patient with schizophrenia per year rose with 24%, per quarter from €432-€534 without significant deviations from this trend. # Conclusion and discussion Our goal was to study the healthcare received by a cohort of patients with schizophrenia during a period of substantially rising co-payments. Our most important findings are: over the whole period (2009–2014) consumption of psychiatric healthcare decreased strongly, while consumption of somatic healthcare increased. The percentage of patients receiving only antipsychotic medication increased and the percentage of patients without any elective psychiatric care also increased substantially. Use of acute psychiatric care was highest in quarters when patients received only antipsychotic medication. Continuous elective psychiatric care over the whole period was received by 59% of the patients, who showed less than half the levels of episodic psychiatric care compared to the other patients. On top of this decreasing trend in access to psychiatric healthcare we found that the abrupt rise in co-payments from 2012 onwards coincided with relative increases in patients treated with only antipsychotic medication as well as with relative increases in acute psychiatric care. The overall reduction in consumption of psychiatric healthcare is spectacular and much higher than expected. The overall reduction in acute care may be explained as a result of a strongly pursued national policy to substitute episodic care with elective, planned outpatient treatment, thereby reducing institutionalization and enhancing social participation among patients. However, we would have expected that a reduction in acute care would coincidence with increased elective psychiatric care. This is not what we observed. During the study period both outpatient psychiatric care consumption and episodic psychiatric care consumption decreased. Although causal inferences from a naturalistic study should be interpreted with caution our results support the hypothesis that co-payments have unwanted effects on the appropriate healthcare use among patients with schizophrenia. There are several factors that may have mitigated the potential effects of general co-payments. For instance, health insurance companies and cities have established health insurance contracts for low-income groups with provisions to reduce the co-payments. Patients with schizophrenia have low incomes and may have profited from these arrangements. Furthermore, a number of providers did not collect the specific co-payments for psychiatric care. Had such dampening measures not been taken, the potential effects of co-payments may have been even larger. ## Strengths and limitations We consider the following as strengths of our research. Large registry databases complement the insights provided by trials, especially about the healthcare provided and the effects on large groups at population level. Research on registry data is rare. Most of the available studies analyze one aspect of healthcare for patients with schizophrenia, for instance the use of antipsychotic medication, costs, family services, or continuity of care\[,–,–\]. Our examination of the relationship between continuity of care, co-payments and episodic psychiatric care, with acute psychiatric care as a proxy of quality of care, is new and encompasses all types of care available. The strength of our data is its reliability. Registry data from health insurers in the Netherlands are comprehensive and include both mental and somatic healthcare from all healthcare providers. These data are thought to be reliable because it is important for patients, providers, and insurers that the data are correct and the National Care Authority regulates and controls the claims process thoroughly. A further strength is that, in the Netherlands, there is universal access to (compulsory) health insurance which means that only very few people are uninsured. Throughout the country there are well-developed facilities for the treatment of patients with schizophrenia. This means that there is little room for bias by access to insurance or due the availability of services. The results of our study should be interpreted in the light of several limitations. The associations between co-payments and healthcare use may have been driven by patient characteristics (bias by indication) and the way healthcare providers have adapted to policy changes during the study. During the study period the outcome variables we studied are probably influenced by the nation-wide policy to decrease the amount of clinical care. Second, because detailed information about other DSM-IV diagnoses was not available, associations between co-morbid psychiatric disorders and outcome and costs could not be analyzed. Third, the patients that remained insured by Zilveren Kruis during the whole period may be different from those who left. However, in 2008 only 240 (1.5%) were not insured that whole year by Zilveren Kruis and in the 2009–2014 period, a minority (2 458, 17%) of the patients selected in 2008 were not insured by Zilveren Kruis for the whole period. A fourth limitation is that only two outcomes in a naturalistic study are analyzed: (1) acute care, (2) amount of psychiatric and somatic care. ## Conclusion Although the observational design of our study precludes firm causal inference, we conclude that it is highly likely that the rise in co-payments for mental health care in the Netherlands has substantially contributed to a decrease in patients accessing elective continuous outpatient care and an increase in both their using stand-alone antipsychotic medication and acute crisis or inpatient care. This is clearly and untoward effect of measures taken to control the rising costs of health care. The results are not only relevant for The Netherlands, but also interesting for other countries with intentions to raise or introduce co-payments. We recommend that experiments are started in which co-payments are lifted or where patients are even rewarded when accessing appropriate care. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Predator-prey interactions have played a substantial role in shaping the diversity of life, leading to many adaptations and counter-adaptations for attack and defence. Perhaps the most widespread defence is camouflage, preventing an object from being detected or recognised by an observer. On a basic level, camouflage is intuitively simple, often involving matching the general appearance of the background environment. Yet this ignores a rich complexity to the subject because many different types of camouflage are thought to exist, from background matching and disruptive coloration to countershading and masquerade, and camouflage can also be optimised in a variety of ways. Almost certainly, the most common form of camouflage in nature, and the basis for many other types of concealment, is background matching, where an animal resembles the general colour and pattern of the background. In nature, animals could exhibit background matching through a variety of mechanisms, including genetic adaptation over long time periods, phenotypic plasticity during development, behavioural choice of substrates, using materials and decorations from the environment, and colour change over different time scales. The ability to change colour over a short time frame, physiological colour change, is based on the redistribution of pigment within chromatophore cells. Investigating animals that can change colour is a particularly useful method to study how camouflage works and is tuned to different environments because researchers can manipulate the background on which individuals are found and investigate how the camouflage is changed in response. This approach can also yield insights into how the visual systems of animals work and interpret information in the environment, and how action is mediated via visual pathways. While a great deal of recent research has investigated the different types of camouflage that may exist and how they work, mostly in artificial systems, comparatively little work has studied camouflage optimisation and tuning in real animals. The most extensive work to date has focussed on cephalopods, which show remarkable abilities for rapid colour and pattern change in response to specific features of the environment, such as pattern contrast and edge information. Studies of colour change for camouflage have also been undertaken in, for example, chameleons, flatfish, and crabs. However, outside of studies on cephalopods, relatively little work has directly quantified how effective rapid (here defined as changes occurring in seconds or minutes) colour change for camouflage is, and how quickly this can occur. Little work has also directly investigated the exact form that colour change takes in terms of changes in colour and brightness. A major problem has been that conventional methods using spectrometry to quantify coloration require extensive handling of specimens, leading to stress induced colour change, and are also slow and hence unable to quantify rapid colour change. Fish make an ideal group to study colour change for camouflage because it is widely reported that many species have this ability, and they occur in a wide range of habitats and on many backgrounds. The ability to change colour over short term, often for concealment, is thought to be widespread among teleost fish, both in the marine and freshwater environments, and enables them to occupy a greater range of backgrounds and to cope with heterogeneous habitats. A variety of goby species have been observed to change colour, both to match their backgrounds in order to maintain camouflage, and during breeding phases. However, only one study by Fries in 1942 has specifically conducted experiments investigating camouflage by these fish on different backgrounds, with gobies reported to become paler and less red when on blue colours, darker and more red on red backgrounds, and more yellow on yellow backgrounds, owing to changes in chromatophore cells. However, colour change was monitored by human eye, without quantifying it objectively, how it affected match to the background, or how fast it occurred. In this paper we study the colour change abilities of the abundant and widely distributed rock goby (*Gobius paganellus*) when placed on backgrounds of different colours and brightness, to test whether they can change appearance for camouflage, and how quickly they do this. We use digital image analysis and predator vision modelling to quantify the speed and extent of colour change. Rockpool and intertidal fish such as gobies are excellent candidate species for this type of study for a variety of reasons. First, the environment in which they live is highly changeable, with a wide range of background types existing even over very small areas. In addition, physical disturbance of tides and waves will often push individuals over a range of backgrounds against which they are viewed by predators. Furthermore, owing to the tidal nature of the environment, the fish are under intense predation pressure, at low tide from birds and at high tide from other groups, including larger fish. Therefore, there may be a major advantage in being able to rapidly change colour for camouflage. # Materials and Methods Gobies were collected by dip net in the intertidal zone from Gyllyngvase beach, Falmouth, Cornwall, UK (50° 8′33.4690″N, −005° 04′07.9716″W) between July 2013 and September 2013 for experiment 1 (40 individuals), and between October 2013 and July 2014 (40 individuals) for experiment 2. Once caught, fish were kept in fresh seawater in a grey bucket in order to minimise colour change prior to use. Both experiments were conducted *in* *situ* on Gyllyngvase beach under natural light conditions in shallow trays lined with waterproof paper (see below). All work was conducted under approval from the University of Exeter Biosciences ethics committee (application 2013/149). The field location (specified above) where the experiments were conducted and fish collected is public land and no further licences or permits were needed. Fish were kept no longer than 2 hours, and all individuals were returned unharmed to their original rockpool area after being tested. Rock gobies are not an endangered or protected species. ## Experimental Background Creation Our aim in experiment 1 was to test whether gobies are capable of changes in their luminance when placed on a black or white background. In experiment 2, we aimed to test for changes in colour. Our design here was intended to minimise perceived differences in brightness of the backgrounds by the fish, and to test whether gobies change their actual colours as opposed to just changes in luminance. We used red and blue as two colours at different ends of the visual spectrum that gobies are likely to be able to discriminate. In both experiments we also aimed to test how quickly any changes in appearance occurred. Background colours for both experiments were created by printing colours at 300 dpi on waterproof paper (HP LaserJet Tough paper; Hewlett Packard, Palo Alto, USA), with a Hewlett Packard Colour LaserJet 2605 dn printer. Before the experiments, gobies were placed on an intermediate grey background for at least 15 minutes as a standard background for all individuals. This was to provide the same starting point at the beginning of the experiment, and to remove some of the individual variation that would otherwise exist owing to individuals being found on different rockpool substrates on collection. To produce an intermediate grey midway between white and black we followed past approaches and printed a range of grey squares of different intensity (pixel) values from black through to white made in Photoshop Elements 5.0 (Adobe Systems Inc., San Jose, USA). We measured the reflectance of each square using an Ocean Optics (Dunedin, FL, USA) USB2000+ spectrometer, held at 45° to normal, with illumination by a PX-2 pulsed xenon lamp, and calculated the average reflectance of each square across 400–750 nm (we excluded UV light because the paper and the print toner reflect little UV light), followed by plotting image pixel value against reflectance. We then calculated the mid grey value based on a ratio scale. For experiment 1, we simply printed the darkest black and used the white paper for the two experimental backgrounds. For experiment 2, the red and blue colours were calculated to be the same average brightness across the visible spectrum. This was achieved by photographing a range of different red and blue colours printed on the same paper (as above), followed by measuring their reflectance values in longwave (LW), mediumwave (MW), and shortwave (SW) images. ## Experimental Procedure Fish were placed in a 24 cm wide × 34 cm long × 5 cm deep (internal measurements) white tray that had been covered with a background of midpoint grey paper, calibrated as described above. The tray was filled 2 cm deep with fresh seawater and a spirit level was used to ensure trays were flat and the water level accurate to prevent variation in colour measurements due to water depth, and to ensure all areas of the tray had sufficient water to minimise stress to the fish by ensuring that even the largest individuals were fully submerged, while at the same time keeping water depth low so as to not affect the colour analyses. Fish were given 15 minutes to acclimatize on the grey background then photographed (see below) in the control tray before being transferred individually into a secondary experimental tray of 28.5 cm wide × 39 cm long × 7 cm deep (internal measurements) divided into eight compartments by thin plastic barriers attached with silicone sealant glue. These compartments were either four white and four black for experiment 1, or four red and four blue for experiment 2. Transfer of the fish between trays was done as quickly as possible and with a net in order to minimise any stress associated with capture and handling. Fish were unable to see each other, although barriers were not completely sealed and water was able to flow around the tray. Although this also meant that chemical cues could potentially transfer among individuals, any such effects should not produce directional colour changes in line with responses to background colours and brightness. The experimental trays also ensured that pairs of fish were tested under the same water conditions (e.g. temperature), and the relatively small size of the compartments prevented fish from swimming around too much, which would have made photography difficult. Experimental trials were undertaken in blocks, with a single block consisting of a pair of fish, with one fish placed on each background colour, and with those individuals approximately matched by size to remove bias that may occur due to variation in colour change with individual size. Twenty fish were tested on each background colour for each experiment (40 fish in total per experiment). Fish were subsequently photographed again at 1–2, 10 and 60 minutes while remaining in the tray to establish the extent of colour change over time. Photos were taken using a Nikon D90 SLR camera, which had undergone a quartz conversion to enable ultraviolet sensitivity (Advanced Camera Services, Norfolk, UK) and fitted with a Nikon 105 mm Nikkor lens. In both experiments photographs were taken in human visible (400–700 nm) and ultraviolet (300–400 nm). For the human visible photos a UV/IR blocking filter was used (Baader UV/IR 2″ Cut Filter) and a UV pass filter was used during the ultraviolet photographs (Baader U 2″ Cut Filter). All photographs included a Spectralon 40% grey reflectance standard (Labsphere, Congleton, UK) next to the tray and a ruler. Due to changing light conditions and reflectance from the water surface, a black and silver photographic umbrella (Neewer, Guangdong, China) was used to shade the trays from direct sunlight. ## Image Analysis Images were taken in RAW format with manual white balance and fixed aperture settings. Images were then linearized with regards to light intensity based on camera responses to a set of eight Spectralon grey standards with reflectance values ranging from 2 to 99% (in custom programs written in Image J) in order to correct for the non-linear responses in image values many cameras produce in response to changes in light levels. Image values were then equalised with regards to the 40% grey standard, and each image channel (LW, MW, SW and UV) scaled to reflectance, where 255 on an 8-bit scale is equal to 100% reflectance. We wanted to analyse colour change with regards to one of the likely main predator groups of rockpool fish: shore birds. To obtain data corresponding to avian vision, we transformed the reflectance based image based on spectral sensitivity data from the peafowl (*Pavo cristatus*) using a polynomial mapping technique to convert from camera to avian colour space. The likely predators of rockpool fish include a range of shorebird species found at the intertidal zone. Previous work has shown that these are likely to have a ‘violet’ sensitive system, with the UV cone type shifted in sensitivity to slightly longer wavelengths than species that fall into the ‘ultraviolet’ group (although violet sensitive species can still detect UV light). Although gulls are likely to be predators of rockool fish too, and seem to have a UV visual system, the relatively low levels of UV involved in the backgrounds and fish should mean that differences in the perception between these systems is small. The peafowl is often used as a model species for modelling birds that fall into the violet group. Once calibrated, the outline of each goby was drawn around by hand using Image J and the region of interest (ROI) saved. Each image layer was measured to acquire values for photon catch. We then calculated a series of metrics to analyse the appearance of each goby. Saturation (the amount of a given colour compared to white light) was defined as the distance an object is in a tetrahedral colour space from the achromatic grey point. Larger distances equate to colours that appear more saturated. We next derived a measure of colour type, or hue. Here, we followed past approaches that have defined hue based on a ratio of the relative photoreceptor stimulation in different parts of the light spectrum. Broadly, this approach, whereby colour types are defined in terms of a ratio of the different channels present, is based on the way that opponent colour channels are thought to work in animal vision, and in practical terms is a way of defining a colour type in an intuitive and readily interpretable manner. In experiment 1, we had no *a priori* reason to expect particular changes in colour of fish because all the backgrounds used were achromatic shades of grey. Therefore, we used a standardised ratio that describes colour in terms of differences in the amount of shorter to longer wavelengths of light (an approach commonly used to calculate opponent channels): hue = ((LW+MW)–(SW+UV))/(LW+MW+SW+UV). In experiment 2, whereby we used backgrounds that were either red or blue, we predicted specific changes in coloration with fish moving more towards these two colour types. As such, we defined hue as (LW–SW)/(LW+SW). Higher values mean that an individual is relatively red in colour, whereas smaller values mean an individual is relatively blue. To derive a measure of achromatic change in appearance, we calculated luminance (perceived lightness) based on the double cone values, as in birds achromatic vision is widely thought to be driven by these receptors. Finally, we calculated how changes in the appearance of fish equated to differences in their level of match to the experimental backgrounds. To do so we used a log form of a model of visual discrimination, the Vorobyev-Osorio model, which is based on differences in colour or luminance based on photo catch values, including estimates of neural noise and relative photoreceptor proportions. We used a Weber fraction value of 0.05 for the most abundant cone type \[70, 90\], and relative proportions of cone types in the retina of the peafowl (LW = 0.95, MW = 1.00, SW = 0.86, UV = 0.45;). The model gives values of ‘just noticeable differences’ (JNDs), whereby differences of 1.00–3.00 mean that two stimuli are unlikely to be discriminated by an observer, and larger values above 3.00 are increasingly likely to equate to discriminable differences. ## Statistics We did not specifically expect an overall difference in appearance between fish on each background at all time points. Instead, our key prediction was that there should be no difference at the start of the experiment (time zero) when fish have been on the same intermediate grey background, whereas there should be differences as the experiment progresses. The exact time where differences arise should also depend on the speed of colour change. As such, we conduced a series of planned comparisons between fish on each background type, separately at each time point. Data for all metrics except hue were non-normal and resistant to transformation and so we conduced Wilcoxon Mann-Whitney tests. For hue we conducted two-sample t-tests. Owing to the repeated testing for each experiment (one test per time point), we adjusted the critical p-values needed for significance by using a sequential Bonferroni. For each experiment, p-values are ranked in order of significance and then compared to an adjusted critical value in turn, which becomes more stringent with each additional test. Critical thresholds for significance for each of the four statistical tests per experiment were therefore 0.050, 0.025, 0.016, and 0.012. To test for changes in the level of camouflage over time for both colour and brightness/luminance, we conducted Kruskal-Wallis tests. # Results ## Experiment 1 ### Changes in Colour and Luminance For luminance, there was no significant difference between fish on black or white backgrounds at time 0 (W = 414.0, n = 20, p = 0.925), but there were significant differences at one minute (W = 587.0, n = 20, p\<0.001), 10 minutes (W = 600.0, n = 20, p\<0.001), and at 60 minutes (W = 610.0, n = 20, p\<0.001), with fish on white backgrounds having higher luminance values. Note, however, that the magnitude of differences is generally quite small with the largest difference in luminance values between time 0 and time 60 being 0.09 (with photon catch values on a scale of 0–1), and average differences being 0.03 (across both backgrounds). In terms of colour change, for saturation, there was also no significant difference between fish on black or white backgrounds at time 0 (W = 406.0, n = 20, p = 0.925), but significant differences occurred at one minute (W = 305.0, n = 20, p = 0.005), 10 minutes (W = 268.0, n = 20, p\<0.001), and 60 minutes (W = 308.0, n = 20, p = 0.006), with saturation values being higher on the black backgrounds. The results for hue were very similar, again with no significant difference at time 0 (T = −0.92, df = 35, p = 0.363), but significant differences at one minute (T = −9.15, df = 36, p\<0.001), 10 minutes (T = −12.24, df = 37, p\<0.001), and 60 minutes (T = −10.49, df = 37, p\<0.001), with hue values being higher (more LW and MW and less SW and UV in colour) for fish on white backgrounds. ### Differences in Camouflage Over Time On the white background there was no significant reduction in JNDs over time (better match to the substrate) for colour (H = 0.46, df = 3, p\<0.927), but there was for luminance JNDs (H = 31.77, df = 3, p\<0.001;). On the black background there was a significant difference in colour JNDs with time (H = 12.81, df = 3, p = 0.005). However, note that there was no decline in JNDs with time, but rather lower JND values at times 0 and 60 than at times 1 and 10, indicating that fish did not actually improve camouflage for colour over time intervals. There was no significant difference in luminance JNDs (H = 1.59, df = 3, p = 0.661). Therefore, fish improved in their achromatic match to the white background, but not to the black background. ## Experiment 2 ### Changes in Colour and Luminance There was no significant difference between fish on red or blue backgrounds for luminance at time 0 (W = 410.0, n = 20, p = 1.000), at one minute (W = 373.0, n = 20, p = 0.324), or at 60 minutes (W = 346.0, n = 20, p = 0.086), nor was there a significant difference at 10 minutes when controlling for multiple testing (W = 327.0, n = 20, p = 0.025);. Regarding colour, for saturation, there was no significant difference between fish on red or blue backgrounds at time 0 (W = 378.0, n = 20, p = 0.394), but there were significant differences at one minute (W = 311.0, n = 20, p = 0.008), 10 minutes (W = 307.0, n = 20, p = 0.006), and 60 minutes (W = 289.0, n = 20, p = 0.001), with fish being more saturated on the red background. The results for hue were similar, with no significant difference at time 0 (T = 0.60, df = 31, p = 0.552), but significant differences at one minute (T = −6.35, df = 35, p\<0.001), 10 minutes (T = −6.90, df = 32, p\<0.001), and 60 minutes (T = −7.24, df = 37, p\<0.001). Fish on red backgrounds had higher hue values (more LW and less SW in coloration;). ### Differences in Camouflage Over Time On a red background, there was a significant reduction in JNDs for colour (H = 25.31, df = 3, p\<0.001), but not for luminance JNDs (H = 3.10, df = 3, p = 0.376);. On a blue background, there was also a significant reduction in colour JNDs over time for colour (H = 17.56, df = 3, p = 0.001). Although there was a significant change in luminance JNDs, this was generally a decrease rather than improvement in luminance matching over time 0 (H = 14.64, df = 3, p = 0.002). # Discussion Here, we tested whether rock gobies can change either their luminance (lightness) or colour depending on the background on which they are placed. As predicted, in experiment 1, fish changed in their overall luminance when put onto either a white or a black background, with individuals getting lighter or darker respectively. This led to changes in the level of similarity of fish to each background in terms of luminance, improving camouflage matching over time. In contrast, although there were some statistically significant changes in hue and saturation in this experiment too, these generally did not affect the overall match to the background, indicating that these changes were perceptually small and unlikely to be of significance in terms of camouflage, similar to other work. In contrast, in experiment 2 where fish were placed onto either red or blue backgrounds, individuals underwent marked changes in colour with regards to both hue and saturation. At least some goby species have been shown to have three cone types, sensitive to relatively shorter and medium/longer parts of the spectrum, and so they should be able to distinguish between the blue and red backgrounds. In accordance with this, on a red background fish became more red in colour and more saturated, whereas on a blue background they became less red more grey in colour. These differences led to significant improvements in the level of colour match to the background. In contrast, there was little change in the luminance of fish on these backgrounds, demonstrating that fish can change their overall colour without changing their luminance. The exact mechanism of luminance perception in rock gobies is unknown, but this result suggests that they likely perceived the two background types as being of about the same brightness because in experiment 1, where the backgrounds were very different, fish did change in luminance. Overall, in both experiment 1 and 2 the changes were very rapid, with the majority of colour and brightness change occurring in the first minute. The result that fish changed to become more red in coloration on the red background, yet that changes towards the blue colour were much smaller (they mostly become more grey in colour) is interesting. It suggests that some types of colour are easier for the fish to adopt than others. This fits with the background environment in the habitat where the fish were collected, whereby blue colours are rare, yet red encrusting algae and brown stones and seaweed are common. Past work has shown that different types of chromatophore control different colours, with black melanin being controlled by melanophores, yellow pteridine controlled by xantophores, red carotenoids controlled by erythrophores and more rarely blue cyanophores controlling a yet unknown cyan biochrome. Colour responses may also be elicited by more than one type of chromatophore, and Fries suggested that blue response in common gobies are due to a response of the erythrophores, xanohpores, and iridophores. However, more work is needed to test what cellular mechanisms cause changes in goby coloration, and whether other populations might be capable of greater changes in blue. The levels of change in luminance were relatively small in this study even for the fish that changed the most. Thus the decrease in difference to the background, although significant, was not very large. However, this did equate to a decrease in discrimination thresholds of almost 10 JNDs on average for fish on the white backgrounds. In general across the experiment fish were quite dark, and in nature they are likely to be a better match to the general colour and brightness of the substrate in the rockpools (especially the dark rocks). Therefore, in such cases in the wild when fish are already well matched in appearance to the background even relatively small differences may equate to a valuable benefit in improved camouflage. Furthermore, to our eyes, changes in the brightness of fish are clearly perceptible. One possibility to resolve this apparent discrepancy is that in this study we analysed the appearance of the entire body of each fish. In reality, gobies often have quite strong patterns that to us have key characteristics of disruptive coloration to break up the body shape against the background. We often noted that the prominence of such patterns changed as fish change colour, and we think it quite possible that even when the overall brightness of an individual stays essentially the same that there can be pronounced changes in pattern. For example, on a uniform background fish may reduce the contrast and prominence of their markings and adopt a more uniform appearance, but this may be broadly similar in overall brightness to that of their starting appearance. Otherwise, better brightness match to the background may be brought about through longer-term changes, such as through morphological colour change that occurs over days and months and may be caused by changes in the overall density of chromatophores in the skin. It is interesting that the ability of fish to change colour seems to be better than their ability to change brightness. Until we test fish on more natural coloured backgrounds we can only speculate as to why this may be. In the rockpool environment, the background is highly heterogeneous in terms of brightness, with stones and gravel of a range of shades occurring on a small scale (smaller than the size of the fish). Thus, overall changes in brightness may have a relatively small benefit. In contrast, some rockpools and larger backgrounds seem to have broadly different colours, meaning that colour change may be more valuable. This is likely to be especially the case with changes in shore height too, whereby there are changes in the amount of substrate types, especially greater brown and green algae cover lower down the shore. Here, we have focussed on changes in colour and brightness using relatively artificial background appearances. Next, it will be important to test for colour change and camouflage ability on backgrounds that more closely resemble those in the environment where the individuals live. Moreover, given that gobies often have strongly contrasting patterns that appear disruptive, it would be important to test whether individuals have the capacity to change their markings on backgrounds of different marking sizes and contrasts. Previous studies have shown that flatfish are capable of impressive changes in pattern depending on the substrate appearance. While gobies are unlikely to match the extent of this ability, the potential is there for them to change their patterns for concealment. In addition, the environment in which gobies live is both highly challenging (the intertidal) and heterogeneous, and so being able to adjust individual markings is likely to provide a strong advantage. A number of other species live in the same habitat, including several other common species of goby and blenny that have been suggested to change colour too. Thus, there also exists great potential for comparative studies of colour change within and among habitat types in intertidal fish species. It should also be noted that in this study we have not directly explored the level of individual colour change possible because this would require placing the same individual fish on different backgrounds in a repeated measures design and analysing their colour change abilities. While in the present study we have focussed on colour change for camouflage a number of recent studies on colour change in gobies report change colour in response to breeding, with individuals becoming less camouflaged and more attractive to mates. This change appears to be largely hormonally induced. In gobies, breeding coloration change in other species is also influenced by predation pressure, with courtship colouration less intense under high predation risk. When in aquaria devoid of predators, however, blennies from high predation sites showed full courtship colouration. In other fish, arctic charr (*Salvelinus alpinus*) have shown differences in aggression when paired over white or black backgrounds. Darker skin colour in salmonids may relate to subordination and be used to reduce aggressive interactions between conspecifics. When fish are paired within a tank and acclimatised to light coloured backgrounds, fish generally show increased aggression towards each other, whereas this aggression is not observed if fish have been acclimatised to dark backgrounds. Many species of fish will also change colour under stressful conditions, such as when odour cues suggest a predator is in close proximity. While some species of goby have been found to react to this predator cue, *Gobius paganellus* has not been found to display any defensive reaction to conspecific skin extract. In *Gobius minutus*, physical handling also provoked rapid darkening and reddening in pale fish. Thus, it would be valuable to investigate how capacity for colour change is affected by factors such as season and mating behaviour, dominance, and condition. Overall, colour change in fish and other species presents an excellent system to study camouflage, what makes this effective and how it is controlled, and other life history factors that may affect camouflage responses and tuning. # Supporting Information We thank Zeehan Jaafar and another anonymous referee for a range of helpful comments on the work and manuscript. We also thank Jolyon Troscianko for help by writing the Image J calibration programmes and assistance with camera calibration. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MS AD AEL. Performed the experiments: AEL AD MS. Analyzed the data: MS. Contributed to the writing of the manuscript: MS AEL. Calibrated the images and measured them: AEL AD. Undertook visual modelling of the data: MS.

# Introduction Multiple myeloma (MM) is a malignant plasma cell disorder that accounts for approximately 10% of all hematological cancers. Despite recent advances, long- term survival is rare after autologous stem cell transplantation and/or treatment with recently introduced anti-myeloma agents, and disease recurs in virtually all patients. Therefore, other therapeutic approaches need to be developed to complement the current strategies. Several immune alterations have been described in MM patients. These alterations are caused in part by the replacement of normal bone marrow with malignant plasma cells, suppressing normal hematopoiesis. Moreover, the immune response is directly suppressed by MM cells and through their interactions with the microenvironment. As the immune response impairment contributes to MM progression, cellular immunotherapy appears to be a promising therapeutic approach. Allogeneic stem cell transplantation (allo-SCT) is a form of cellular immunotherapy that is widely used to treat hematological malignancies. Much of the curative potential of allografts is attributed to the “graft-versus-tumor” (GvT) effect. In MM, evidence for a graft-versus-myeloma (GvM) effect was provided by the ability of donor lymphocyte infusions to induce complete responses in patients who initially relapsed after allo-SCT, and by the association between chronic graft-versus-host disease (GvHD) and a decreased incidence of relapse after transplantation. However, despite evidence of GvM effects, allo-SCT has remained a controversial treatment modality in MM. Given the high relapse rate of MM after allo-SCT, some of the current clinical trials focus on combining non-myeloablative allo-SCT with new drugs given for post- transplantation maintenance therapy. However, the introduction of immunomodulating agents that could improve GvT effects may inadvertently induce GvHD. This is well illustrated in a recent study by the HOVON group, where lenalidomide maintenance after non-myeloablative allo-SCT increased acute GvHD, and strongly suggests that new therapies aimed at modulating GvM effects should ideally be tested first in animal models. Mouse models have contributed to the understanding of MM biology and to the introduction of novel agents, and are of great interest in the preclinical evaluation of cellular immunotherapy. Currently, only two immunocompetent murine models have been described in which allo-SCT is associated with a GvM effect, but these models do not resemble human MM disease or do not use allo-SCT as a curative treatment for established disease. So far, an immunocompetent murine GvM model in which allo-SCT is used for the treatment of established MM that resembles human disease, marked by bone marrow tropism and osteolytic lesions, has not been described. In the current study, we investigated the anti-myeloma effects of allo-SCT from B10.D2 mice into myeloma-bearing Balb/cJ mice (H-2^d MHC-identical, but differing at minor histocompatibility loci) which results in sclerodermatous chronic GvHD,. Myeloma-bearing Balb/cJ mice were inoculated with the myeloma cell line MOPC315.BM, originating from Balb/c mice, that presents bone marrow tropism. # Materials and Methods ## Ethical statement All experimental procedures and protocols used in this investigation were reviewed and approved by the Institutional Animal Care and Use Ethics Committee of the University of Liège (Belgium), reference 1016. The “Guide for the Care and Use of Laboratory Animals”, prepared by the Institute of Laboratory Animal Resources, National Research Council, and published by the National Academy Press, was followed carefully as well as European and local legislation. Animal welfare was assessed at least once per day, and all efforts were made to strictly control animal suffering during the experiments (e.g. development of a decisional system to follow the mice, application of humane endpoints with precise response to specific symptoms including use of dietary supplements, analgesic administration and sacrifice). ## Animals Balb/cJ (H-2^d) and B10.D2 (H-2^d) mice were purchased from Jackson Laboratory (Bar Harbor, ME, USA). Strains were kept and bred at the animal facility of our institute. Mice were used when they were between 10- to 14-wk-old. ## Myeloma cell line and model The selection of the MOPC315.BM cell line, which is derived from the mineral oil-induced plasmacytoma cell line MOPC315, was previously described. The parental MOPC315.BM cells and the firefly luciferase transfected cells (MOPC315.BM.Luc) were provided by Prof. Bjarne Bogen. Luciferase-transfected MOPC315.BM cells were used for all experiments. Cells were maintained in culture at 37°C in 5% CO₂ using a RPMI 1640 medium (Sigma-Aldrich, Bornem, Belgium) containing 10% fetal bovine serum (FBS) (Sigma-Aldrich) and 1% Penicillin (100 U/ml)/Streptomycin (0.1 mg/ml) (Sigma-Aldrich). Intravenous (i.v.) injection of MOPC315.BM cells results in tumor development with a restricted localisation in the bone marrow and spleen and is associated with osteolytic lesions, validating the model as a multiple myeloma model. In advanced disease stages, bone marrow infiltration can cause paraplegia in mice through spinal cord compression. Mice injected with MOPC315.BM cells were monitored daily for general condition and locomotion. They were sacrificed when presenting locomotion trouble/paraplegia, deterioration of general condition or apathy. Animals that were not immediately sacrificed when presenting locomotion trouble/paraplegia, e.g. because they were receiving allo-SCT treatment, received analgesic administration (buprenorphine 0.05 mg/kg twice per day) and were very closely monitored for general condition and activity. ## Graft cell suspension Spleens and bone marrows (femurs and tibias) from donor mice were harvested and homogenized in RPMI 1640 medium containing 10% FBS and 1% Penicillin/Streptomycin ( = complete medium). Red blood cells were lysed using sterile filtered RBC lysis buffer (eBioscience, San Diego, USA) and cells were washed, resuspended in phosphate buffered saline (PBS) containing 3% FBS, and filtered through a 70 µM nylon membrane. For CD8 T-cell depletion, the “Mouse CD8α positive selection kit” (Stem Cell, Grenoble, France) was used according to the manufacturer's EASYSEP depletion protocol. Finally, cells were suspended in 200 µl PBS for i.v. injection. ## Bioluminescence measurement Beetle luciferin (Promega, Leiden, Netherlands) solubilized in PBS was injected intra-peritoneally into mice (3 mg/mouse in 100 µl). Bioluminescence was measured within 10 to 20 minutes using VIVOVISION IVIS 200 (Xenogen, Alameda, USA). Results were analysed and quantified using Living Image software (Xenogen). ## *In vivo* experimental design Balb/cJ recipient mice were injected intravenously with 2.5×10⁵ MOPC315.BM.Luc cells. MM development was allowed to proceed for 30 days, and monitored by bioluminescence studies. At day 30 post-inoculation, mice were irradiated with 6 Gy (Total Body Irradiation) from a ¹³⁷Cs source (GammaCell 40, Nordion, Ontario, Canada). After 6 hours, mice were transplanted by i.v. injection of 1×10⁷ bone marrow cells and 7×10⁷ splenocytes from donor mice \[allogeneic: B10.D2 donor; autologous: Balb/cJ donor\]. Myeloma-bearing mice that received allogeneic or autologous transplant are referred to as “<u>Allo-MM</u>” or “<u>Auto-MM</u>” mice, respectively. Mice were sacrificed after appearance of myeloma symptoms (e.g. paraplegia), GvHD symptoms or apathy. Bioluminescence monitoring allowed tracing of luciferase- transfected myeloma cells and assessment of tumor development. This transplantation protocol was adapted from Jaffee and Claman who developed a murine model of chronic GvHD. Experimental and monitoring strategies are summarized in. ## Donor sensitization experiments For GvHD studies, healthy Balb/cJ recipient mice were irradiated and transplanted with B10.D2 grafts as described in “experimental design”. Before transplantation to Balb/cJ mice, B10.D2 donor mice were not sensitized (controls), or sensitized by i.v. injection of 5×10⁵ myeloma cells (MOPC-sensitized) or Balb/cJ splenocytes (Balb/c-sensitized) 21 days before sacrifice for graft harvesting. GvHD symptoms were evaluated with a scoring system adapted from Sakoda et al. The score is based on weight loss (\<10% = 0; 10–20% = 1; \>20% = 2), “hunched-back” position (normal = 0; hunched-back while resting = 1; persistent = 2), activity (normal = 0, reduced activity = 1, apathy = 2), alopecia (normal = 0, \<1 cm² = 1, \>1 cm² = 2) and skin fibrosis (normal = 0, fibrosis = 1; scabs = 2) with a maximum score of 10. The animals' conditions were controlled daily, and the GvHD score was calculated at least 3 times per week. Mice were sacrificed at the latest at a score of 8/10 or when apathic. ## *In vitro* co-cultures Cell suspensions were obtained from spleens as described earlier. Effector cells were obtained from spleens of Allo-MM mice, sacrificed 2 weeks after transplantation. Target cells (MOPC315.BM cells or Balb/cJ splenocytes) were irradiated with 50 Gy using a ¹³⁷Cs source (GammaCell 40) and washed with complete medium after irradiation. The effector:target ratio was 20∶1 and cells were co-cultured for 5 days in 6-well plates using complete medium, containing 0.1% of Heparin (LEO Pharma, Lier, Belgium). Effector cells were harvested and analysed by flow cytometry at the end of the co-cultures. ## Flow cytometry Cell suspensions were obtained from spleen, lymph nodes, bone marrow and blood. Red blood cells were lysed as described earlier. Extracellular staining was performed in PBS containing 3% FBS. Intra-cytoplasmic or intra-nuclear staining was performed using BD Cytofix/Cytoperm (BD Biosciences, San Diego, CA, USA) or Foxp3 Staining Buffer Set (eBioscience), respectively. Antibodies were incubated for 30 min at 4°C. The Streptavidin/PerCPCy5.5 complex and the following antibodies were purchased from eBioscience: anti-CD4/eFluor450 (RM4-5); anti- CD8/PECy7 (53-6.7); anti-CD49b/Biotin (DX5); anti-CD69/APC (H1.2F3); anti-B220/APCeFluor780 (RA3-6B2); anti-Foxp3/PE (FJK-16s); anti-CD44/APC (IM7); anti-CD62L/APCeFluor780 (MEL-14). The following antibodies were purchased from BD Biosciences: anti-CD229.1/FITC (30C7); anti-CD3e/v500 (500A2) or from Invitrogen: anti-IgA/FITC. A specific antibody directed against MOPC315.BM paraprotein (Ab2.1-4/Biotin) was kindly provided by Bjarne Bogen. Quantitation of blood cells was determined using BD Trucount tubes (BD Biosciences). Flow cytometric data were acquired using a BD FACSCanto II flow cytometer (BD Biosciences) and the BD FACS DIVA software, and analysed with the FlowJo software (Tree Star, Ashland, OR, USA). ## Serum paraprotein quantitation Serum paraprotein levels were measured using an ELISA for mouse IgA (Mabtech, Sweden) according to the manufacturer's instructions and analysed with a Multiskan FC Plate reader (Thermo Scientific) with the SkanIt for Multiskan FC 3.1 software. ## TCR Vβ CDR3-size spectratype analysis T cells were isolated from splenocytes of B10.D2 mice, i.e. “non-sensitized” (control mice), “MOPC-sensitized” or “Balb/cJ-sensitized” mice 21 days post- sensitization. CD8 T cells were isolated using the “Mouse CD8α positive selection kit”. The CD8-negative fraction was further depleted to eliminate residual CD8 cells and constituted the “CD4 fraction” (\>80% of T cells were CD4⁺). Cell pellets were suspended in TriPure Isolation Reagent (Roche, Vilvoorde, Belgium). RNA extraction was performed using chloroform and isopropanol following the manufacturer's instructions. Isopropanol phase containing RNA was transferred on “RNeasy Mini kit“ columns (Qiagen, Venlo, Netherlands) and RNA extraction was finalized following the manufacturer's instructions. Genomic DNA was removed using recombinant RNase-free DNaseI (Roche, Vilvoorde, Belgium). cDNA was synthesized from RNA (2 µg) using oligo(dT)₁₈ primers with the “Transcriptor First Strand cDNA synthesis kit” (Roche). Seminested PCR was performed using sense primers for a panel of murine Vβ families and two Cβ anti-sense primers, the second being fluorescently labelled (IDT Technologies, Leuven, Belgium), as previously described. All PCR reagents were purchased from Applied Biosystems (Life Technologies, Gent, Belgium). The fluorescently labelled PCR products were run together with GeneScan ROX 500 Size Standard (Applied Biosystems) on a “DNA Analyzer 3730“ (Applied Biosystems) capillary electrophoresis system at the GIGA-Research Genomics facility of the University of Liège. CDR3-size spectratype analysis was performed with GeneMapper version 4.0 Software (Applied Biosystems). Experiments were repeated three times, and the mean area of the different peaks, representing different CDR3-size lengths, was calculated for each Vβ family. A CDR3-size length was considered skewed when the mean area under the peak was higher than the mean+3SD of the same peak in the control condition (non- sensitized B10.D2 mice), as previously described. ## Statistics Statistical significance between groups was determined using Mann-Whitney tests. Survival curves were compared using the Log-Rank test (Mantel-Cox). These statistical tests were performed with the Prism Software (Graph Pad Software, San Diego, CA). # Results ## Graft-versus-myeloma effect The experimental design is illustrated in. Briefly, Balb/cJ recipient mice were injected intravenously with luciferase-transfected MOPC315.BM cells. After MM development during the first 30 days, mice were irradiated and transplanted by i.v. injection of bone marrow cells and splenocytes from donor mice (allogeneic: B10.D2 donor; autologous: Balb/cJ donor). Myeloma-bearing mice that received allogeneic or autologous transplantation are referred to as “<u>Allo-MM</u>” or “<u>Auto-MM</u>” mice, respectively. MHC-matched allo-SCT in the MOPC315.BM myeloma model resulted in strong anti- myeloma effects. We observed complete bioluminescence disappearance in 17 out of 18 Allo-MM mice from 4 independent experiments (94%), whereas all 13 Auto-MM mice showed an initial decrease in bioluminescence (probably due to the irradiation) followed by increasing bioluminescence signals after transplantation and progressive myeloma disease (p\<0.0001;). Strikingly, two mice in the Allo-MM group already displayed paraplegia before transplantation and recovered completely. Serum paraprotein measurements demonstrated a significant decrease of paraprotein levels in the Allo-MM mice (p\<0.0001). In contrast, paraprotein significantly increased in the Auto-MM group. Moreover, the Allo-MM group showed significantly lower paraprotein levels at sacrifice compared to the Auto-MM group (p\<0.0001;), as well as lower myeloma cell infiltration in the bone marrow (mean±SD: 0.03±0.04 vs. 1.3±1.3%; p\<0.0001) and spleen (0.3±0.3 vs. 3.5±5.3%; p = 0.027). All together, these data demonstrate a potent GvM effect of allogeneic transplantation in this model. A similar anti- tumor effect was also observed when myeloma cells were injected subcutaneously, resulting in formation of solid tumors. After allo-SCT, solid tumors regressed into small residual tumors, whereas tumors in the autologous group continued to grow after transplantation (N = 6/group). Regarding GvHD, 16 out of 18 mice in the Allo-MM group showed symptoms of chronic GvHD (alopecia, skin fibrosis, weight loss, “hunched-back” position, diarrhea) after day 21 post-transplantation, which is the time point for symptom appearance in the B10.D2→Balb/cJ GvHD model. The other 2 mice were sacrificed before day 21, one mouse due to myeloma progression and the other mouse due to a worsening of its general condition, probably due to transplant-related complications. ## Immune cell populations in the graft-versus-myeloma model In order to determine which immune cells might be responsible for the observed GvM effect, we performed flow cytometry analyses on blood samples at different time points. At the time point of sacrifice also lymphoid organs (bone marrow, spleen, lymph nodes) were analysed. At sacrifice, we observed significantly higher percentages of total CD8 T cells in the bone marrow and blood, and activated CD8 T cells in all investigated organs of Allo-MM mice compared to Auto-MM mice or healthy Balb/cJ mice. A large CD8 T-cell expansion (total and activated) was confirmed by absolute cell counts in blood at sacrifice. For activated CD4 T cells, an increase in percentages and absolute counts was also observed in Allo-MM mice at sacrifice, but this increase was much smaller compared to that of CD8 T cells (fold-increase in T-cell counts for Allo-MM vs. Auto-MM: CD8 T cells: 8.2x; activated CD8 T cells: 12x. CD4 T cells: 2x; activated CD4 T cells: 3.5x). In addition, percentages of regulatory T cells were significantly decreased in blood and bone marrow of Allo-MM mice compared to Auto-MM or healthy mice at sacrifice. Prior to sacrifice, kinetics in blood samples already showed higher total and activated T-cell counts (CD4 and CD8) 1 week after SCT in the Allo-MM compared to the Auto-MM group. Three weeks after SCT, activated CD8 were still significantly increased and total CD8 tended to be increased, whereas no increase in total or activated CD4 T-cell counts was present at this moment (data not shown). Furthermore, we observed significantly higher percentages of effector memory subsets (CD44⁺CD62L⁻) within CD4 and CD8 T cells at sacrifice in the blood and spleen of the Allo-MM group, and the same trend was noted in the bone marrow. On the other hand, central memory (CD44⁺CD62L⁺) and naive (CD44⁻CD62L⁺) T-cell subsets were decreased in spleen and bone marrow of this group compared to the Auto-MM group or healthy mice. Percentages and cell counts of NK (DX5⁺ CD3⁻) and NKT (DX5⁺ CD3⁺) cells did not differ between the two groups, whereas percentages of B cells (B220⁺ CD3⁻) in all lymphoid organs and blood were lower in Allo-MM mice compared to Auto-MM mice (data not shown). ## Involvement of CD8 T cells in the graft-versus-myeloma effect Based on the previous results suggesting a possible *in vivo* implication of T cells in the GvM effect, we evaluated T-cell reactivity against myeloma or allogeneic cells *in vitro*. We performed a five-day co-culture of splenocytes from Allo-MM mice, with irradiated target cells (MOPC315.BM or BALB/cJ splenocytes). The results showed an expansion of CD8 T cells, in contrast to CD4 T cells, but the expansion observed in co-cultures with MOPC315.BM cells was not different from co-cultures with Balb/cJ splenocytes, suggesting a close relationship between epitopes recognized in chronic GvHD and GvM processes. Furthermore, increased percentages of activated CD4, and even more so activated CD8 T cells, confirmed T-cell reactivity against myeloma cells and Balb/cJ cells among Allo-MM splenocytes. Our previous results suggested reactivity of CD8 T cells against MM. Moreover, MOPC315.BM cells express MHC I, but do not express MHC II, possibly implicating a direct activation of myeloma-reactive CD8 T cells, whereas CD4 T cells need antigen-presenting cells to become reactive against myeloma cells, as previously described. Thus, we evaluated the contribution of CD8 T cells in the GvM effect *in vivo*, by depleting CD8 T cells in the B10.D2 graft before transplantation to myeloma-bearing Balb/cJ mice. CD8 T-cell and activated CD8 T-cell reconstitution was significantly slower in the depleted group 1 week after transplantation, and a trend for lower CD8 T-cell numbers was still noted 3 weeks after transplantation. In the CD8 T-cell-depleted group, 4 out of 6 mice (66.7%) showed strong bioluminescence signals and myeloma symptoms after transplantation, whereas in the standard Allo-MM group a bioluminescence signal was only observed in one out of 18 mice after transplantation (5.6%, p\<0.01). Higher tumor burden and paraprotein levels in CD8 T-cell-depleted mice confirmed a reduced GvM effect compared to standard allogeneic transplantation, underlining the *in vivo* importance of CD8 T cells in GvM effects. ## Graft-versus-myeloma and graft-versus-host reactivity The separation of GvT effects from unwanted GvHD remains an important challenge in transplantation medicine. In order to determine whether epitopes were shared between immune responses directed against MOPC315.BM myeloma cells and GvH response in our model, we decided to perform sensitization of B10.D2 donor mice by injecting them with myeloma cells prior to cell collection for transplantation. In preliminary experiments (data not shown), Allo-MM recipient mice challenged with myeloma-sensitized donor cells presented with exacerbated chronic GvHD symptoms, indicating possible overlap between myeloma antigens and alloantigens *in vivo.* In order to confirm these results and compare the effects of MM sensitization with allogeneic sensitization, we performed chronic GvHD experiments on larger cohorts, in which B10.D2 donor mice were sensitized with MOPC315.BM myeloma cells (MOPC-sensitized) or Balb/c-splenocytes (Balb/c-sensitized) before transplantation to healthy Balb/c recipient mice (chronic GvHD model). Recipient mice transplanted with grafts from Balb/c-sensitized donors experienced worsened GvHD symptoms and had a shorter survival compared to control animals, with a median survival of 31 vs 42 days, respectively (p = 0.038). Notably, flow cytometric results showed reduced percentages of Treg cells and naive T cells (CD4 and CD8) in these mice compared to control mice in spleen, blood and bone marrow at sacrifice (data not shown). Likewise, recipient mice of MOPC-sensitized donor grafts also showed a trend towards shorter survival (median survival 37 days), suggesting that sensitization with myeloma cells could lead to increased alloreactivity in this model. In order to confirm the possible overlapping responses between GvHD and GvM in our model, we used CDR3-size spectratype analyses to determine which TCR Vβ families were involved in the different B10.D2 T-cell responses. We analysed the TCR Vβ spectratype within both CD4 and CD8 T cells (isolated from MOPC- sensitized or Balb/c-sensitized B10.D2 mice), as our previous results support a role of CD8 T cells in the GvM effect, whereas chronic GvHD in the B10.D2→Balb/c model is mainly dependent on CD4 T cells. Skewed CDR3-size lengths for different Vβ families are summarized in, and indicate clonal or oligoclonal expansions. Within CD4 T cells, the results revealed 6 myeloma-reactive Vβ families (2, 3, 5.1, 5.2, 8.3, 11) and 5 alloreactive Vβ families (5.1, 5.2, 11, 15, 18). Reactivity of Vβ families 2, 3 and 8.3 was unique to the anti-myeloma response, whereas alloreactivity specifically involved Vβ families 15 (with five skewed CDR3-lenghts) and 18. Vβ families 5.1., 5.2 and 11 showed expansion both in the anti-myeloma and allogeneic settings, suggesting overlapping responses of these CD4 T-cell populations. Surprisingly, within CD8 T-cell compartment there were no uniquely expanded Vβ families. We only observed three reactive Vβ families (5.1, 11, 13) skewed in both groups at exactly the same CDR3-size lengths, also suggesting potential overlap between responses to MM and Balb/c antigens in the CD8 T-cell subset. Although we did not observe overall aggravation of GvHD symptoms in the larger cohort of mice transplanted with grafts from MOPC-sensitized donors (data not shown), survival of these mice was shorter compared to control group. This result suggests that sensitization with myeloma cells could lead to increased alloreactivity in this model due to the presence of shared antigens between both allogeneic Balb/cJ and MOPC315.BM cells, as demonstrated by the presence of overlapping CDR-3 size skewed bands in both CD4 and CD8 T-cell populations. # Discussion In the current study, we describe a graft-versus-myeloma effect in the context of MHC-matched allogeneic transplantation in myeloma-bearing mice. So far, only two other immunocompetent mouse models of allogeneic transplantation in MM have been described. In the first model, Balb/c mice were intra-peritoneally injected with plasmacytoma-resembling HOPC-1F cells, which present few characteristics of human MM disease, and transplanted with bone marrow and spleen cells of DBA/2 origin. No long-term disease-free survival could be obtained with unmanipulated SCT alone – since idiotype vaccination of donor mice was needed for the GvM effect - and the transplanted mice developed acute GvHD, which has a distinct pathobiology from that of chronic GvHD. In the other murine MM allo-SCT model, C57Bl/KaLwRij.Hsd (H-2^b) recipient mice first received allo-SCT from MHC-matched C3.SWH2b/SnJ donors. After two months of immune reconstitution, recipients were inoculated with the 5T33MM murine cell line and developed myeloma disease. Donor lymphocyte infusions (DLI) prolonged the median survival of diseased mice. Additional dendritic cell (DC) vaccination of the DLI-recipient mice, using dendritic cells loaded with the H7 minor histocompatibility antigen that differs between donor and recipient strains, further extended survival without inducing GvHD by targeting the H7-presenting MM cells. Percentages of effector memory CD8 T cells were increased in the bone marrow of transplanted MM mice, irrespective of post- transplantation treatment. However, in this GvM model, the observed anti-myeloma effect is entirely due to post-transplantation immunotherapy using DLI, as mice received allo-SCT before the establishment of MM disease, which does not correspond to the clinical scenario. In our study, we established a GvM model using the MOPC315.BM model, which closely resembles human MM disease as tumor cells mainly grow in the bone marrow milieu and induce osteolytic lesions. A GvM effect was obtained using allo-SCT in mice with established MM disease, as a curative treatment, with concomitant chronic GvHD development. The allogeneic graft was composed of bone marrow (source of hematopoietic stem cells) and splenocytes (source of T cells), which is a widely used approach in murine SCT models. In clinical SCT, the most frequently used graft sources are hematopoietic stem cells collected from peripheral blood after a mobilization treatment. These grafts contain 10 to 30-fold higher amounts of T cells and other immune cells (B cells, monocytes, NK and NKT cells) than bone marrow grafts. Thus, murine SCT grafts that contain bone marrow enriched with splenocytes (predominantly consisting in B and T cells, monocytes, NK and NKT cells) display similarities in cellular composition with peripheral blood-derived grafts used in clinics. Even though some dissimilarities may exist between murine SCT protocols and clinical allo-SCT, such murine models are invaluable tools in understanding the immunobiology of SCT. Our data corroborate the role of T cells in the observed GvM effect. In general, both CD4 and CD8 T-cell subsets contribute to graft-versus-leukemia (GvL) reactions. However, the dominant mechanism seems to be strain-specific and varies with the degree of donor-recipient histocompatibility. Mice receiving CD8-depleted donor marrow have a higher leukemic relapse incidence than those receiving CD4-depleted marrow. In experimental mouse transplants, the addition of purified CD8 T cells to the graft had an anti-tumor effect and facilitated engraftment without inducing GvHD. In our model, *in vivo* and *in vitro* data suggest a role for CD8 T cells in the GvM effect, since CD8 T-cell-depletion of the graft reduced GvM effects. We identified three TCR Vβ families within the CD8 T-cell subset, with overlapping reactivity to both myeloma and alloantigens. CD8 T cells can recognize polymorphic peptides derived from non-MHC proteins (i.e. minor histocompatiblity antigens). Thus, we hypothesise that Balb/cJ minor histocompatibility antigens implicated in GvHD pathogenesis are present on MOPC315.BM cells (originating from a Balb/c-derived background). B10.D2 and Balb/cJ mice, both H-2^d, differ at multiple non-MHC loci (including H-1, H-7, H-8, H-9 and H-13) potentially implicated in alloreactivity and possibly expressed by MOPC315.BM cells, which could explain that the same CDR3-size length was found skewed in both the anti-tumor and the alloresponse (band 161 in Vβ family 5.1, band 165 in Vβ family 11 and band 176 in Vβ family 13). CD4 T cells, which are essential for the development of chronic GvHD in the B10.D2→Balb/c model, also probably played a role in the GvM effect. Indeed, we also identified potentially overlapping TCR Vβ families within CD4 T cells (i.e. Vβ families 5.1, 5.2 and 11;), further confirming a link between GvM and GvHD. Interestingly, we identified other Vβ families within the CD4 T-cell subset that are probably implicated specifically in either GvM or GvH effects, suggesting that GvM or GvH reactivity could be separately modulated in this model in future studies. Despite the lack of MHC II expression on MOPC315.BM cells, primary CD4 T-cell responses (mediated by tumor-infiltrating antigen-presenting cells) can be induced, as demonstrated for the MHC II-negative parental MOPC315 cells. CD4 cells probably play a role in the orchestration of the CD8 T-cell response, and in the establishment of GvHD, as previously described. We did not observe complete disappearance of GvHD after CD8 T-cell-depletion (in mice sacrificed after day 21), most likely because of the presence of alloreactive T cells in the CD4 compartment (as suggested by the presence of single and multiple skewed CDR3-size lengths in the CD4 anti-Balb/cJ response of the Vβ 5.1, 5.2, 11, 15 and 18 families, &) confirming the essential role of CD4 T cells for GvHD development in this model. For future studies using this model, it would be interesting to isolate and infuse T cells from Vβ families that are specifically involved in the GvM effect (Vβ 2, 3 and 8.3) in order to determine the effects on GvHD development and the capacity of these T cells to maintain a GvM effect or, in contrast, to deplete from the graft Vβ families specifically involved in the GvH response (Vβ 15 and 18) and determine the persistence of GvM and GvH effects. Similar experiments have already been described in the literature. The Vβ13 family was shown to be highly skewed in the B10.BR CD8 T-cell response against a myeloid leukemia cell line (MMC6), but not in the alloresponse against CBA recipient mice. Transplantation of low doses of CD8⁺Vβ13⁺ T cells, isolated by magnetic cell separation, induced a slight GvT response with no concomitant acute GvHD development. In the B6→Balb.b (MHC-matched) GvHD response, several Vβ families have been found to be skewed within CD4 T cells. Whereas transplantation of the positively selected skewed Vβ families induced lethal GvHD, mice that received skewed Vβ-depleted CD4 T cells all survived with minimal GvHD symptoms. In conclusion, we describe the establishment of a reliable graft-versus-myeloma model using allo-SCT in immunocompetent tumor-bearing mice. Effector memory CD4 and CD8 T cells probably mediated the GvM effect in this model, with CD8 T cells being essential for the observed GvM effect *in vivo*. Within CD4 and CD8 subsets, we identified overlapping Vβ families in the responses against myeloma cells (anti-tumor) and Balb/cJ cells (alloreactive), underlining the relationship between anti-tumor responses and GvHD, whereas some Vβ families within CD4 T cells specifically respond to either myeloma or host alloantigens. The current murine model of GvM should enable future studies of immunomodulatory drugs, acting on the balance between graft-versus-myeloma and graft-versus-host effects. The authors would like to thank Sophie Dubois, Coline Daulne and Amélie Halleux from the GIGA-Research Hematology laboratory (University of Liège) for excellent technical assistance. Thanks also to the laboratory of tumor and development biology (LBTD) for their help in the bioluminescence studies, and to the GIGA- Research (University of Liège) “Genomics facility” and “Imaging and Flow Cytometry platform” for technical help and the use of the BD FACS Canto II (BD Biosciences). [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MB LB FB JC. Performed the experiments: MB EO LB. Analyzed the data: YB BB JZ FB JC. Contributed reagents/materials/analysis tools: BB JZ PD. Wrote the paper: YB MH AB RH FB JC. [^3]: ¶ These authors are joint senior authors on this work.

# Introduction The loss of a baby from stillbirth has detrimental consequences for the family and the community. The causes of many stillbirths are unexplained. Sleep- disordered breathing (SDB), ranging from snoring to obstructive sleep apnoea (OSA), is common during pregnancy. The cardinal symptom, habitual snoring ≥3 nights per week, affects up to 35% of women in the third trimester, and up to 85% of women with pre-eclampsia, while objective measures of OSA are estimated to affect between 8% and 26% of pregnant women. SDB is a risk factor for adverse pregnancy outcomes, including gestational hypertension and pre-eclampsia, hyperglycaemia, impaired fetal growth, and early-term and/or preterm birth. SDB is exacerbated by obesity, advanced gestation, and the supine sleep position, all of which are themselves associated with an increased risk of late stillbirth. Therefore, pregnant women with SDB may have an increased risk of late stillbirth (≥28 weeks’ gestation) and this risk may be magnified if women settle to sleep supine, however the data is lacking. Importantly, the association between SDB and maternal sleep patterns (sleep quality, sleep duration, restless sleep, daytime sleepiness, and daytime naps) with late stillbirth is inconsistent across studies. A meta-analysis, which included the comparison of stillbirth in women with and without SDB as an outcome measure, using subjective (self-reported snoring) and objective (OSA) measurements, reported no association between SDB and stillbirth. The relationship between sleep duration and late stillbirth was reported in several case-control and cross sectional studies, however, the results are not consistent in identifying an association. Subjective sleep quality was also not associated with stillbirth in a cross-sectional and case-control study. Other case-control studies reported that daily naps, compared to no naps, were independently associated with late stillbirth. These inconsistencies may be due to differing measurements of these aspects of maternal sleep between studies, or because some studies did not adjust for potential confounders (such as maternal body mass index \[BMI kg/m2\] and maternal age). Furthermore, as late stillbirth is a relatively rare event, ranging from 1·3 to 8·8/1000 births in high-income countries, individual studies have been underpowered to investigate interactions between supine going-to-sleep position and late stillbirth in women with SDB compared to those without. The triple risk model suggests that late stillbirth may be the culmination of an interplay between stressors (e.g. SDB, supine going-to-sleep position), maternal risk factors (e.g. obesity, age), and fetal-placental vulnerability (e.g. impaired fetal growth, placental dysfunction). Exploration of possible biological pathways of the association of adverse pregnancy outcomes related to SDB suggests that there are multifactorial mechanisms, including sympathetic activation, oxidative stress, inflammation, and endothelial dysfunction, which contribute to maternal cardiovascular dysfunction, metabolic derangement, placental dysfunction, and fetal compromise. Thus, it is plausible that when a mother is in the supine position in late pregnancy and there is reduced maternal-fetal blood flow from aortocaval compression, the addition of partial airway collapse with SDB may exacerbate fetal compromise in a vulnerable fetus. Since SDB and maternal sleep patterns are potentially modifiable during pregnancy (such as lateral position for supine-dependent snoring, continuous positive airway pressure for OSA, and frequency of daytime naps), it is possible that screening and management of these aspects of maternal sleep during pregnancy may support reduction in the rate of late stillbirth. However, there is a need to assess the current evidence from individual studies that have collected data on maternal sleep and stillbirth to determine if they are associated with late stillbirth. We established the Collaborative Individual Participant Data (IPD) Meta-analysis of Sleep and Stillbirth (CRIBSS) group to address if maternal going-to-sleep position was associated with late stillbirth. This included pre-specified secondary questions on symptoms of SDB and maternal sleep patterns, including 1) is SDB associated with late stillbirth, and 2) is supine going-to-sleep position associated with greater risk of late stillbirth in women with SDB compared to those without? # Materials and methods The study population comprised cases with late stillbirth and controls with ongoing pregnancies from the CRIBSS data. This IPD meta-analysis was registered with the PROSPERO register of systematic reviews (CRD42017047703) and followed the IPD meta-analysis protocol, search strategy, risk of bias for non-randomised studies (ROBINS-E) tool, and published results. Five international case-control studies that collected maternal going-to-sleep position and late stillbirth data were included in this pooled IPD meta-analysis. Participant level inclusion criteria were singleton, non-anomalous pregnancy, ≥28 weeks’ gestation. Exclusion criteria were multiple pregnancy, major congenital abnormality, gestation \<28 weeks’ when pregnancy sleep data was collected, termination of pregnancy at ≥28 weeks’, and receiving an intervention that may have affected going-to-sleep position. Maternal sleep data were collected by self-report via face-to-face interview or online survey within six weeks after stillbirth in cases or at a matched gestation in controls. Late stillbirth, using the international definition of stillbirth, “a baby born with no signs of life at or after 28 weeks’ gestation,” was the primary outcome. The analysis included intrapartum stillbirth, with the rationale that the exact time of the stillbirth may be uncertain and that SDB may result in a vulnerable baby that is unable to tolerate labour. ## Data analysis This was a prespecified secondary analysis of an IPD meta-analysis that investigated maternal going-to-sleep position and late stillbirth, with a one- stage approach stratified by study and site. A detailed statistical analysis plan, prior to the analysis, has been published.²⁵ Prespecified potential covariates were: maternal age, earliest pregnancy BMI, ethnicity, parity, education level, marital status, pre-existing hypertension or diabetes, smoking, recreational drug use, supine going-to-sleep position, fetal movements, infant birthweight by customised centiles, and measures of SDB and sleep patterns (‘any’ snoring, habitual snoring, the Berlin Questionnaire \[BQ\], Epworth Sleepiness Scale \[ESS\], sleep quality, sleep restlessness, and sleep duration). Frequency of getting up to use the toilet and daytime naps were also included as these are previously reported independent risk factors for late stillbirth. Where data exists for multiple time frames, only data for the month prior to the stillbirth were used in the analysis. In cases where the last month data were not available, data collected for the ‘last week’ were used. There are currently no validated tools for SDB screening during pregnancy, therefore we investigated habitual snoring, a positive BQ, and daytime sleepiness using the ESS as proxy indicators. The BQ was developed to identify individuals at risk of OSA in non-pregnant primary care populations and has three categories 1) snoring frequency, loudness, and witnessed apnoea, 2) daytime sleepiness, and 3) BMI \>30 and hypertension, with a positive BQ requiring two positive categories. The ESS is a subjective measure of daytime sleepiness with eight questions about the likelihood of dozing off in specified situations, ranging from unlikely (in a car stopped for a few minutes in traffic) to highly likely (lying down to rest in the afternoon). The ESS is coded as 0 = never doze, 1 = slight chance, 2 = moderate chance, and 3 = high chance, with a positive ESS screen indicating clinical levels of daytime sleepiness defined as ≥10. Data on the usual duration of overnight sleep were also collected. The reference for sleep duration was defined as 6 to 9 hours, with duration categorised as \<6, 6–9, or \>9 hours. Restless sleep and sleep quality were each single questions, with ‘average’ restlessness and ‘average’ sleep quality as the reference group. A one-stage approach to meta-analysis was used, so that the data from the participating eligible studies were included in a single model. Logistic regression models were used for the binary outcome. A fixed study effect and study site effect were included in the model specification as strata. Univariable analysis was performed to evaluate the association between the measures of SDB and maternal sleep patterns and the odds of late stillbirth. A multivariable model was developed incorporating prespecified covariates available in all the studies (Appendix 1 in the) and measures of SDB and maternal sleep patterns that were significant in univariable analysis. Some covariates (habitual snoring, the BQ, sleep quality, restless sleep, daytime naps, daytime sleepiness using the ESS, and getting up to use the toilet) were not available in all participating studies. The interaction between supine going-to-sleep position and common measures of SDB (habitual snoring and the BQ) and sleep duration were assessed in bi- variable regression models. Significant interactions were then added to the multivariable model as described above. Estimates of the risk of late stillbirth were reported as odds ratio (OR) with 95% confidence intervals (95% CI). For missing data in each individual study, imputation was not undertaken. Statistical analyses were performed using SAS, version 9.4 (SAS Institute Inc., Cary NC USA). Each individual study obtained ethical approval. Approval for the IPD meta- analysis was obtained from the New Zealand Health and Disability Ethics Committee (NTX/06/05/054/AM06). # Results Participants comprised 851 late stillbirth cases and 2257 controls with ongoing pregnancies from five eligible case-control studies: the Auckland Stillbirth Study, the New Zealand Multicentre Stillbirth Study, the Sydney Stillbirth Study, the UK Midlands and North of England Stillbirth Study, and the International Study of Trends and Associated Risks for Stillbirth Study, comprising women of many ethnicities. Differences in maternal and pregnancy characteristics, infant size, and going- to-sleep position between cases and controls have been previously reported. ‘Any’ snoring (cases n = 473, 56.0%; controls, n = 1182, 54.1%), sleep quality (fairly bad to very bad, cases n = 248, 33.5%; controls, n = 703, 35.3%), daytime sleepiness (positive ESS score ≥10, cases n = 128, 17.5%; controls n = 312, 15.8%), and frequency of getting up to use the toilet (≥1 per night, cases n = 667, 90.0%; controls, n = 1820, 91.5%) last month were not associated with late stillbirth in the univariable analysis. Long sleep duration \>9 hours last month (cases n = 78, 10.5%; controls, n = 129, 6.5%) was independently associated with late stillbirth compared to sleep duration of 6 to 9 hours (adjusted odds ratio \[aOR\] 1.82, 95% CI 1.14–2.90). Reporting a daily daytime nap last month (cases n = 139, 23.7%; controls, n = 216, 12.8%) compared to never reporting a daytime nap was associated with an increase in the odds of late stillbirth (aOR 1.52, 95% CI 1.02–2.28). In addition, a positive BQ (cases n = 176, 30.0%; controls, n = 370, 21.8%) was associated with late stillbirth (aOR 1.44, 95% CI 1.02–2.04), however, when BMI \>30 was removed from the BQ score, a positive BQ showed no significant association with stillbirth (aOR 0.81, 95% CI 0.54–1.21). Restless sleep greater than average last month (cases n = 225, 38.3%; controls, n = 761, 45.2%) was associated with a reduction in the odds of late stillbirth (aOR 0.62, 95% CI 0.44–0.88). Women who had a stillbirth, 689 cases from four participating studies, were asked what time of day they thought their baby had died: 34.8% (n = 240, or 52.3% of 459 cases who could recall a time of day) reported that they thought their baby had died overnight, 19.4% (n = 134) reported afternoon-evening, 11.8% (n = 81) morning, 0.6% (n = 4) during a daytime nap, and 33.4% (n = 230) were unsure. Interactions were assessed between supine going-to-sleep position and habitual snoring, a positive BQ including BMI, sleep duration \>9 hours, and restless sleep greater than average last month. Interactions for a positive BQ (p = 0.56), sleep duration \>9 hours (p = 0.99), and restless sleep greater than average (p = 0.98) were not statistically significant. There was a significant interaction between habitual snoring and supine going-to-sleep position (multivariable interaction p value = 0.001). The combined effect of supine going-to-sleep position and habitual snoring resulted in a reduced odds of late stillbirth in the multivariable model than would be expected.. # Discussion ## Main findings Our study has demonstrated that a positive BQ, long sleep duration \>9 hours, and a daily daytime nap in the last month, were each associated with increased odds of late stillbirth. In contrast, restless sleep greater than average in the last month was protective for late stillbirth. The associations between these aspects of maternal sleep and late stillbirth were adjusted for prespecified covariates available in all the studies (S 1), and measures of SDB and maternal sleep patterns significant in univariable analysis. The \~50% prevalence of ‘any’ snoring and habitual snoring ≥3 nights per week between 17–24% was within the range reported in the pregnancy literature. ‘Any’ snoring, habitual snoring, sleep quality, and daytime sleepiness using the ESS, was not associated with late stillbirth. This is consistent with previous studies: snoring, sleep quality, and daytime sleepiness. A positive BQ was independently associated with late stillbirth, although this association was no longer significant when BMI \>30 was excluded from the BQ (Model 2). This aligns with the suggestion that the BQ used in pregnant women is a proxy for BMI during late pregnancy, due to BMI being a component of the BQ. Indeed, the BQ performs poorly as a screening tool for objective SDB measures during pregnancy, with a 2018 meta-analysis of six studies (n = 604 participants) reporting poor to fair BQ performance during pregnancy with an overall probability of OSA occurrence of 38% if a pregnant woman has a positive BQ. This range may be due to the BQ including risk factors that do not apply to pregnant women (male gender, age \>50 years) and because weight gain is relevant for all pregnancies. Furthermore, symptoms of SDB progress with gestation, and there are differing opinions about the optimal timing of the BQ during pregnancy. Long sleep duration \>9 hours was also associated with late stillbirth , and this association has previously been reported in two case-control studies. While the reason is uncertain, it is plausible that prolonged periods of aortocaval compression during maternal sleep may be a factor. It is also possible that an unmeasured confounder associated with long third trimester sleep (e.g. working night shifts or no paid employment) may lengthen the duration of maternal sleep over the last month and contribute to stillbirth. The definition of long duration in the individual case-control studies is also inconsistent, ranging from \>8 hours to \>9 hours. This range may be due to lack of consensus about what is considered normal sleep duration in healthy pregnancy, although self- reported time to sleep in the third trimester is similar to objectively measured sleep duration and maternal estimates of sleep duration increases in accuracy with increasing duration of sleep. There was no association between short sleep duration during last month and late stillbirth, despite an independent association with short sleep on the night before stillbirth in three case- control studies. This discrepancy may be due to a potentially fatal fetal event (e.g. pre-labour contractions for an acutely compromised fetus) that may shorten sleep on the night before stillbirth. Daily daytime naps were also associated with a 1.5-fold increase in the odds of late stillbirth compared with no daytime naps, and this finding is consistent with individual studies. The physiology behind this is unknown and cannot be explained by overnight sleep duration or daytime sleepiness, as daily naps remained significant when we controlled for these factors. However, we speculate that daily naps in late pregnancy may increase the duration of maternal inactivity, potentially increasing the amount of time that the women spend in the supine position and therefore the duration of aortocaval compression, which when combined with the blood pressure dips that occur during third trimester sleep, may further compromise a vulnerable fetus. Our finding of a 38% reduction in the odds of late stillbirth for women who reported restless sleep more than average during the last month is novel. We speculate that this may be due to maternal body movement facilitating maternal- fetal blood flow, potentially abating adverse fetal effects of aortocaval compression. Furthermore, while maternal hypotension is known to have adverse fetal consequences, such as lower birth weight and stillbirth, increased third trimester arousals related to snoring may assuage prolonged periods of relative hypotension, as deep sleep is commensurate with the lowest overnight blood pressure and arousal with increased blood pressure. Our finding of a protective association between restless sleep more than average and late stillbirth aligns with an international case-control study that reported non-restless sleep in the last month was associated with a 1.7-fold increase in odds of late stillbirth. Similarly, getting up to use the toilet on the night before stillbirth is associated with a 2-fold reduction in late stillbirth, suggesting that maternal body movement on the night before stillbirth may mitigate the effects of a hypoxic event on the fetus. Certainly, pregnant women are susceptible to the development of sleep disturbances, commonly reduced quality and duration of sleep, night waking, daytime sleepiness, and snoring. Causes are most likely to be hormonal and physiological changes of pregnancy, including increased oxygen consumption and metabolic rate, lower overall oxygen reserve, nasopharyngeal oedema, vasomotor rhinitis, and weight gain, which contribute to narrowing of upper airway, reduced functional residual capacity due to diaphragmatic pressure by the growing fetus, and increased arousals during sleep. These physiological changes are exacerbated as pregnancy progresses and when combined with obesity, advanced maternal age, and supine sleep position. Conversely, late pregnancy may provide some protection from SDB, with increased respiratory drive, alteration in the cyclical sleep pattern with decreased rapid eye movement (REM) sleep, and preference for a lateral sleep position. These may be factors contributing to our finding of a significant interaction between habitual snoring during the last month and supine going-to-sleep position, with a lower odds of late stillbirth than expected in women who reported both during the last month. While this may be a chance finding due to low prevalence, with 17 (12 controls and 5 cases) of 92 women reporting habitual snoring and a supine going-to-sleep position, this could also be explained by the women being woken by a sleep companion or experiencing a self-arousal due to snoring, and moving from the supine to a lateral position, which is known to reduce third trimester snoring in obese women and late stillbirth risk. ## Strengths and limitations A limitation of the IPD meta-analysis is that not all participating studies had data for all sleep measures. Minor differences in the design of the individual studies also limited the inclusion of some covariates. Our search had no language restriction and an eligible study from India was identified, however, there was no response from authors or journal editors to repeated invitations to participate. No other eligible randomised trials, prospective cohort studies or studies from low-income countries were identified, thus participating studies were all case-control studies from high-income countries. A limitation of case- control studies include the retrospective data collection which is subject to potential recall bias, although as the relationship between late stillbirth and maternal sleep is not universally well known by pregnant women, systematic bias is unlikely. The longer length of time before interview for cases may have influenced their recall compared to controls, however, case recall is unlikely to be biased towards an association with SDB, with self-reports from a single night of sleep having similar bias and calibration as ‘usual’ sleep. Use of self-reported symptoms of SDB, rather than objective measures using polysomnography may also be considered a limitation. However, self-report of snoring is strongly and reliably associated with the severity of OSA obtained from polysomnography in non-pregnant and pregnant women, therefore self-report is useful for large scale studies where routine access to polysomnography in late pregnancy is costly and impractical. # Conclusion This IPD meta-analysis adds to the evidence on maternal sleep and late stillbirth, using the best available data on the association of SDB and maternal sleep patterns with the risk of late stillbirth. These findings demonstrate that self-reported maternal snoring, a positive BQ screen excluding BMI, daytime sleepiness, sleep quality, and getting up to use the toilet, are not independently associated with late stillbirth last month. Long sleep duration \>9 hours and daily daytime naps are independent risk factors, while sleep more restless than average may reduce the odds of late stillbirth. There is an urgent need to better understand factors associated with long sleep duration and daily daytime naps before recommendations can be made to pregnant women. Meanwhile, pregnant women may be reassured that the commonly reported increased restlessness of sleep during late pregnancy may be physiological and is associated with a reduced risk of late stillbirth. # Supporting information We would like to thank the women who participated in the individual studies and the research midwives who conducted the interviews. 10.1371/journal.pone.0230861.r001 Decision Letter 0 Liguori Claudio Academic Editor 2020 Claudio Liguori This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 5 Jan 2020 PONE-D-19-22011 Associations between symptoms of sleep-disordered breathing and maternal sleep patterns with late stillbirth: findings from an individual participant data meta-analysis. PLOS ONE Dear Mrs Cronin, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. We would appreciate receiving your revised manuscript by Feb 19 2020 11:59PM. When you are ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. To enhance the reproducibility of your results, we recommend that if applicable you deposit your laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission- guidelines#loc-laboratory-protocols> Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). This letter should be uploaded as separate file and labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. This file should be uploaded as separate file and labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. This file should be uploaded as separate file and labeled 'Manuscript'. Please note while forming your response, if your article is accepted, you may have the opportunity to make the peer review history publicly available. The record will include editor decision letters (with reviews) and your responses to reviewer comments. If eligible, we will contact you to opt in or out. We look forward to receiving your revised manuscript. Kind regards, Claudio Liguori Academic Editor PLOS ONE Journal Requirements: 1\. When submitting your revision, we need you to address these additional requirements. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf> and <http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_f ormatting_sample_title_authors_affiliations.pdf> 2.  Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: <http://journals.plos.org/plosone/s/supporting-information>. 3.  Thank you for stating the following in the Financial Disclosure section: Mrs Cronin reports grants from Health Research Council of New Zealand (12/372); Cure Kids (5357); Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., grants from 2016 TransTasman Red Nose/Curekids (6601), grants from Sir John Logan Campbell Medical Trust, during the conduct of the study; Ms. Wilson reports grants from 2016 TransTasman Red Nose/Curekids (6601); Dr. Gordon reports grants from Stillbirth Foundation Australia, during the conduct of the study; other from NHMRC Early Career Fellowship \#1089898,  outside the submitted work; Dr. Li reports grants from Health Research Council of New Zealand (12/372); Cure Kids (5357); Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., grants from 2016 TransTasman Red Nose/Curekids (6601), during the conduct of the study; Dr. Culling has nothing to disclose; Associate Professor Raynes-Greenow reports grants from Stillbirth Foundation Australia, during the conduct of the study; other from NHMRC Career Development Fellowship \#1087062, outside the submitted work; Professor Heazell reports grants from Action Medical Research, during the conduct of the study; grants from Tommy's, grants from NIHR, outside the submitted work; Dr. Stacey reports grants from Cure Kids (3537), Nurture Foundation, Auckland District Health Board Charitable Trust., grants from Health Research Council of New Zealand (12/372); Cure Kids (5357); Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., during the conduct of the study; Professor Askie has nothing to disclose; Professor Mitchell reports grants from Cure Kids (Grant 3537), Nurture Foundation, Auckland District Health Board Trust., grants from Health Research Council of New Zealand (12/372); Cure Kids (5357); Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., grants from 2016 TransTasman Red Nose/Curekids (6601), during the conduct of the study; other from Cure Kids, outside the submitted work; Associate Professor Thompson reports grants from Cure Kids (3537), Nurture Foundation, Auckland District Health Board Charitable Trust., grants from Health Research Council of New Zealand (12/372); Cure Kids (5357); Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., grants from 2016 TransTasman Red Nose/Curekids (6601), during the conduct of the study; Professor McCowan reports grants from Cure Kids (537), Nurture Foundation, Auckland District Health Board Trust., grants from Health Research Council of New Zealand (12/372); Cure Kids (5357);  Mercia Barnes Trust; Nurture Foundation; University of Auckland Faculty Research Development Fund (3700696)., grants from 2016 TransTasman Red Nose/Curekids (6601), during the conduct of the study. Associate Professor O'Brien reports grants from American Sleep Medicine Foundation, grants from ResMed, outside the submitted work; We note that you received funding from a commercial source: ResMed Please provide an amended Competing Interests Statement that explicitly states this commercial funder, along with any other [relevant declarations](http://www.plosone.org/static/competing.action) relating to employment, consultancy, patents, products in development, marketed products, etc. Within this Competing Interests Statement, please confirm that this does not alter your adherence to all PLOS ONE policies on sharing data and materials by including the following statement: "This does not alter our adherence to PLOS ONE policies on sharing data and materials.” (as detailed online in our guide for authors <http://journals.plos.org/plosone/s/competing-interests>).  If there are restrictions on sharing of data and/or materials, please state these. Please note that we cannot proceed with consideration of your article until this information has been declared. Please include your amended Competing Interests Statement within your cover letter. We will change the online submission form on your behalf. Please know it is PLOS ONE policy for corresponding authors to declare, on behalf of all authors, all potential competing interests for the purposes of transparency. PLOS defines a competing interest as anything that interferes with, or could reasonably be perceived as interfering with, the full and objective presentation, peer review, editorial decision-making, or publication of research or non-research articles submitted to one of the journals. Competing interests can be financial or non-financial, professional, or personal. Competing interests can arise in relationship to an organization or another person. Please follow this link to our website for more details on competing interests: <http://journals.plos.org/plosone/s/competing-interests> 4\. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. Additional Editor Comments: Only some minor changes have been requested by the Reviewers; please modify the manuscript as requested. \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: The aim of this work is to find the relationship of symptom of SDB and late stillbirth from previously done case-control studies that is available and pool them together using statistical technique. The point that this paper is a successful collaboration between large teams around the globe combining with a robust individual patient-level data analysis technique is interesting. I congratulate the authors for their efforts. Overall I find the manuscript to be of good quality, however some minor revision points may need to be done: 1\. Study selection process report need to be more transparent, i.e. the authors should describe number of study that were excluded (fig. 1 may need to be revised) how many duplicates were there?, why 124 studies did not meet the eligibility criteria? There should be more detail on the number and reason of studies that were excluded, to provided that the author has established all available evidences to support the meta-analysis. Also, for one study that is not participating, is there any difference in characteristic of that particular study compared to participating studies? (The author may need to describe about this in the result) 2\. Case control studies were prone to recall bias and combining case-control in meta-analysis can still be susceptible to this. I wonder if there any other supporting evidences from cohort study? If there is not the author should specifically mention about the result of the assessment of quality of the evidence in the result, and emphasis on the possibility of bias upon making conclusion, which may need a large pregnancy cohort to be done. Again, I sincerely admire the authors' efforts into the collaborating process. Reviewer \#2: Comments to the Author -How the duration of overnight sleep was evaluated? Did you use a standardized sleep log? -How did you evaluated the sleep position (supine and not supine)? Is this data only self-reported? Or did you use some instruments to evaluate it? If only self-reported, it can probably be inaccurate because you can’t know the real position throughout the night. Can you specify it? -You suggest that a reduction in the odds of late stillbirth for women who reported restless sleep during the last month may be is due to maternal body movement facilitating maternal-fetal blood flow, potentially abating adverse fetal effects of aortocaval compression. Considering that the supine position is associated with an increase in aortocaval compression, have you evaluated interactions between restless sleep and sleep position to understand if the reduced risk of late stillbirth in these cases can be related to the position? -You report that the combined effect of supine going-to-sleep position and habitual snoring resulted in a reduced odds of late stillbirth in the multivariable model and you explain this result assuming that snoring can cause arousal that can cause a change of position from supine to lateral. It would be interesting to test this hypothesis with an objective assessment of the position. -You report that getting up to use the toilet in the night is associated with a reduced risk of late stillbirth, suggesting that maternal body movement on the night may mitigate the effects of a hypoxic event on the fetus. Considering that getting up to use the toilet in the night can be also a clinical manifestation of SBD, have you analyzed interactions between the use of the toilet in the night and Berlin Questionnaire (excluding BMI)? -Non using objective measures (polysomnography) for SBD’s evaluation may be a limitation as you specified. Didn’t you evaluate the possibility to assess the polysomography, if not possibile in all the women, in some of them such as patients with positive Berlin Questionnaire (excluding BMI), in patients with restless sleep and in patients getting up to use the toilet? \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: No Reviewer \#2: Yes: Dott.ssa Francesca Furia \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email us at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0230861.r002 Author response to Decision Letter 0 15 Feb 2020 The authors thank the editorial team and reviewers’ for their comments. The manuscript has been revised according to the comments. EDITORIAL FEEDBACK 1\) When submitting your revision, we need you to address these additional requirements. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <http://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf> and <http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_f ormatting_sample_title_authors_affiliations.pdf> \- AUTHOR RESPONSE: We agree, and have addressed these additional requirements. 2\) Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: <http://journals.plos.org/plosone/s/supporting-information> \- AUTHOR RESPONSE: We agree, and have added detail to the captions for the Supporting Information files at the end of our manuscript. 3\) Thank you for stating the following in the Financial Disclosure section… We note that you received funding from a commercial source: ResMed. Please provide an amended Competing Interests Statement that explicitly states this commercial funder, along with any other relevant declarations relating to employment, consultancy, patents, products in development, marketed products, etc. Within this Competing Interests Statement, please confirm that this does not alter your adherence to all PLOS ONE policies on sharing data and materials by including the following statement: "This does not alter our adherence to PLOS ONE policies on sharing data and materials.” (as detailed online in our guide for authors <http://journals.plos.org/plosone/s/competing-interests>). If there are restrictions on sharing of data and/or materials, please state these. Please note that we cannot proceed with consideration of your article until this information has been declared. Please include your amended Competing Interests Statement within your cover letter. We will change the online submission form on your behalf. \- AUTHOR RESPONSE: We agree, and have amended our Competing Interests Statement as follows: “Associate Professor O'Brien reports grants from ResMed outside the submitted work. This does not alter our adherence to PLOS ONE policies on sharing data and materials.” This amended Competing Interests Statement has been included within our cover letter. Thank you for changing the online submission form on our behalf. 4\) We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see <http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data- access-restrictions>. In your revised cover letter, please address the following prompts: a\) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent. b\) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see <http://www.bmj.com/content/340/bmj.c181.long> for guidelines on how to de- identify and prepare clinical data for publication. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data- availability#loc-recommended-repositories>. We will update your Data Availability statement on your behalf to reflect the information you provide. \- AUTHOR RESPONSE: We have provided additional details regarding the ethical restrictions on sharing our data publicly in our revised cover letter as follows: “Data cannot be shared publicly beyond the Collaborative Individual Participant Data Meta-analysis of Sleep and Stillbirth (CRIBSS) group as no individual participating study obtained consent from participants to make the data publically available. Furthermore, because stillbirth is uncommon there is potential for participants to be identifiable. Contact information for the CRIBSS Data Access Committee is The CRIBSS Data Centre, Department of Obstetrics and Gynaecology, Faculty of Medical and Health Sciences, University of Auckland, Private Bag 92019, Auckland Mail Centre, Auckland 1142.” Thank you for updating our Data Availability statement. REVIEWER FEEDBACK REVIEWER \# 1 1.1) The aim of this work is to find the relationship of symptom of SDB and late stillbirth from previously done case-control studies that is available and pool them together using statistical technique. The point that this paper is a successful collaboration between large teams around the globe combining with a robust individual patient-level data analysis technique is interesting. I congratulate the authors for their efforts. Overall I find the manuscript to be of good quality, however some minor revision points may need to be done: \- AUTHOR RESPONSE: Thank you. 1.2) Study selection process report need to be more transparent, i.e. the authors should describe number of study that were excluded (fig. 1 may need to be revised) how many duplicates were there?, why 124 studies did not meet the eligibility criteria? There should be more detail on the number and reason of studies that were excluded, to provided that the author has established all available evidences to support the meta-analysis. Also, for one study that is not participating, is there any difference in characteristic of that particular study compared to participating studies? (The author may need to describe about this in the result) \- AUTHOR RESPONSE: We agree, and have revised “Figure 1 PRISMA study population flow chart,” which now includes the number of duplicate articles, and participant level inclusion and exclusion criteria. We have also added more detail to the Strengths and Limitations section as follows: “Our search had no language restriction and an eligible study from India was identified, however, there was no response from authors or journal editors to repeated invitations to participate. No other eligible randomised trials, prospective cohort studies or studies from low-income countries were identified, thus participating studies were all case-control studies from high-income countries.” Please note that a December 23, 2019 update for Beall's List of Predatory Journals and Publishers states that the journal that published the non- participating study from India (the International Journal of Reproduction, Contraception, Obstetrics and Gynecology) has been included with “journals that were not originally on the Beall's list but may be predatory.” <https://beallslist.net/standalone-journals/> 1.3) Case control studies were prone to recall bias and combining case-control in meta-analysis can still be susceptible to this. I wonder if there any other supporting evidences from cohort study? If there is not the author should specifically mention about the result of the assessment of quality of the evidence in the result, and emphasis on the possibility of bias upon making conclusion, which may need a large pregnancy cohort to be done. \- AUTHOR RESPONSE: We agree that case-control studies have the limitation of recall bias, and that supporting evidence from cohort studies should be reported where available. This has been explained in the Strengths and Limitations section, and we have clarified that that it is the relationship between late stillbirth and maternal sleep that is not universally well known as follows: “A limitation of case-control studies include the retrospective data collection which is subject to potential recall bias, although as the relationship between late stillbirth and maternal sleep is not universally well known by pregnant women, systematic bias is unlikely. The longer length of time before interview for cases may have influenced their recall compared to controls, however, case recall is unlikely to be biased towards an association with SDB, with self- reports from a single night of sleep having similar bias and calibration as ‘usual’ sleep \[70\]. Use of self-reported symptoms of SDB, rather than objective measures using polysomnography may also be considered a limitation. However, self-report of snoring is strongly and reliably associated with the severity of OSA obtained from polysomnography in non-pregnant \[71\] and pregnant women \[46\]; therefore self-report is useful for large-scale studies where routine access to polysomnography in late pregnancy is costly and impractical.” We also agree that the assessment of the quality of the evidence should be reported. Under the Materials and Methods section, we have already provided a reference to the risk of bias tool and risk of bias report used for this study as follows: “This IPD meta-analysis was registered with the PROSPERO register of systematic reviews (CRD42017047703) and followed the IPD meta-analysis protocol \[38\], search strategy \[24\], risk of bias for non-randomised studies (ROBINS-E) tool \[39\], and published results \[24\].” REVIEWER \#2 2.1) How the duration of overnight sleep was evaluated? Did you use a standardized sleep log? \- AUTHOR RESPONSE: Sleep data, including sleep duration, were collected by self-report in all participating studies. No study collected sleep data via standardised sleep log or actigraphy. To make this clear, we have added the words “self-report” to the Materials and Methods section as follows: “Maternal sleep data were collected by self-report via face-to-face interview \[26, 27, 30, 31\] or online survey \[32\] within six weeks after stillbirth in cases or at a matched gestation in controls” 2.2) a) How did you evaluated the sleep position (supine and not supine)? Is this data only self-reported? Or did you use some instruments to evaluate it? b) If only self-reported, it can probably be inaccurate because you can’t know the real position throughout the night. Can you specify it. \- AUTHOR RESPONSE: a\) Please see answer to 2.1. b\) The participating studies were all questionnaire based, data on maternal sleep position throughout the night was not collected. The supine/non-supine going-to-sleep position variable in our manuscript refers only to going-to-sleep position. The reason that going-to-sleep position was the chosen variable was because this is most easily recalled and modifiable for the majority of women in late pregnancy (Cronin et al., 2017, <https://doi.org/10.1186/s12884-017-1378-5>). We were not able to validate maternal self-report of going-to-sleep position in the individual studies, however, in some of the participating studies the investigators reported that participants often described reference points to remember their going-to-sleep position, such as facing the door. In addition, McIntyre et al’s 2016 overnight sleep study of 30 healthy women in late pregnancy (<https://doi.org/10.1186/s12884-016-0905-0>), reported good agreement for going-to-sleep (sleep onset) position between infared digital video and self- completed questionnaires for participants (who were not given any information about sleep position), with the majority accurately recalling their going-to- sleep position. 2.3) You suggest that a reduction in the odds of late stillbirth for women who reported restless sleep during the last month may be is due to maternal body movement facilitating maternal-fetal blood flow, potentially abating adverse fetal effects of aortocaval compression. Considering that the supine position is associated with an increase in aortocaval compression, have you evaluated interactions between restless sleep and sleep position to understand if the reduced risk of late stillbirth in these cases can be related to the position? \- AUTHOR RESPONSE: We agree that the relationship between restless sleep and supine going-to-sleep position is interesting. Therefore, we have added the analysis of interaction between supine going-to-sleep position and restless sleep greater than average last month to “Table 3: Analysis for interaction between supine going-to-sleep position, and habitual snoring, the Berlin Questionnaire, sleep duration \>9 hours and restless sleep greater than average.” The interaction between supine going-to-sleep position and restless sleep was not statistically significant (p=0.98). 2.4) You report that the combined effect of supine going-to-sleep position and habitual snoring resulted in a reduced odds of late stillbirth in the multivariable model and you explain this result assuming that snoring can cause arousal that can cause a change of position from supine to lateral. It would be interesting to test this hypothesis with an objective assessment of the position. \- AUTHOR RESPONSE: We agree that an objective assessment of sleep position in women with habitual snoring during late pregnancy would be of great interest and could be evaluated in future studies. 2.5) You report that getting up to use the toilet in the night is associated with a reduced risk of late stillbirth, suggesting that maternal body movement on the night may mitigate the effects of a hypoxic event on the fetus. Considering that getting up to use the toilet in the night can be also a clinical manifestation of SBD, have you analyzed interactions between the use of the toilet in the night and Berlin Questionnaire (excluding BMI)? \- AUTHOR RESPONSE: We agree that we reported that getting up to use the toilet on the night before stillbirth was associated with halving of the odds of late stillbirth in three individual studies. However, there was no association found between getting up to use the toilet during the month before stillbirth in the analysis reported in our manuscript. 2.6) Non using objective measures (polysomnography) for SBD’s evaluation may be a limitation as you specified. Didn’t you evaluate the possibility to assess the polysomography, if not possibile in all the women, in some of them such as patients with positive Berlin Questionnaire (excluding BMI), in patients with restless sleep and in patients getting up to use the toilet? \- AUTHOR RESPONSE: The participating studies were all questionnaire based and any referral for polysomnography was not captured. 10.1371/journal.pone.0230861.r003 Decision Letter 1 Liguori Claudio Academic Editor 2020 Claudio Liguori This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 11 Mar 2020 Associations between symptoms of sleep-disordered breathing and maternal sleep patterns with late stillbirth: findings from an individual participant data meta-analysis. PONE-D-19-22011R1 Dear Dr. Cronin, We are pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it complies with all outstanding technical requirements. Within one week, you will receive an e-mail containing information on the amendments required prior to publication. When all required modifications have been addressed, you will receive a formal acceptance letter and your manuscript will proceed to our production department and be scheduled for publication. Shortly after the formal acceptance letter is sent, an invoice for payment will follow. To ensure an efficient production and billing process, please log into Editorial Manager at <https://www.editorialmanager.com/pone/>, click the "Update My Information" link at the top of the page, and update your user information. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, you must inform our press team as soon as possible and no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. With kind regards, Claudio Liguori Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#2: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#2: I congratulate the authors for their work. I find that the manuscript is well written, well analizyzed and of good quality. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#2: Yes: Francesca Furia 10.1371/journal.pone.0230861.r004 Acceptance letter Liguori Claudio Academic Editor 2020 Claudio Liguori This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 13 Mar 2020 PONE-D-19-22011R1 Associations between symptoms of sleep-disordered breathing and maternal sleep patterns with late stillbirth: findings from an individual participant data meta-analysis. Dear Dr. Cronin: I am pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please notify them about your upcoming paper at this point, to enable them to help maximize its impact. If they will be preparing press materials for this manuscript, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. For any other questions or concerns, please email <plosone@plos.org>. Thank you for submitting your work to PLOS ONE. With kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Claudio Liguori Academic Editor PLOS ONE [^1]: Associate Professor O’Brien reports grants from ResMed outside the submitted work. This does not alter our adherence to PLOS ONE policies on sharing data and materials.

# Introduction Flax (*Linum usitatissimum* L.) phloem fibers are a valuable industrial feedstock and are also a convenient model system for studying secondary cell wall formation. The mechanical properties of bast fibers are largely dependent on the composition of their secondary walls. Bast fibers have gelatinous-type walls, which are rich in cellulose (up to 90%) and lack detectable xylan and lignin. Gelatinous fibers are widespread in various land plant taxa, but have been studied primarily in angiosperms. Depending on the species, either phloem or xylem (of either primary or secondary origin) can produce gelatinous fibers in various organs including stems, roots, tendrils, vines, and peduncles,. The mechanisms of gelatinous cell wall development in these fibers remain largely unclear. However, some genes implicated in gelatinous cell wall development have been identified. The list includes fasciclin-like arabinogalactan proteins (FLAs), β-galactosidases, and lipid transfer proteins. A role for β-galactosidases in G-type wall development has been demonstrated functionally. Transcripts of genes encoding chitinase-like proteins are reportedly enriched in fibers, particularly during the cell wall thickening stage of flax phloem cellulosic fiber development. Expression of CTLs during primary or secondary cell wall deposition has also been reported in species other than flax. Plant chitinase-like proteins have been identified in a wide range of organelles and tissues, including the apoplast and vacuole. Chitinase-like proteins belong to a large gene family that includes genuine chitinases (i.e. proteins with proven chitinase activity) and other homologous proteins, which may not have chitinase activity. Here, we refer to both genuine chitinases and their homologs collectively as chitinase-like proteins (CTLs). Chitinases catalyze cleavage of β-1,4-glycoside bonds of chitin and are organized in five classes (Classes I–V), which can be distinguished on the basis of sequence similarity. Classes I, II, and IV belong to glycoside hydrolase family 19 (GH19), found primarily in plants, whereas Classes III and V belong to glycoside hydrolase family 18 (GH18) present in various types of organisms. The Class I chitinases are found in both monocots and dicots, while classes II and IV are found mainly in dicots. Class I and IV chitinases contain a highly- conserved cysteine-rich domain – the chitin binding domain (CBD) – at the N- terminal region, but there are two characteristic deletions in the main catalytic domain in Class IV chitinases. Because chitin is the major component of fungal cell walls, chitinases are classic pathogenesis-related proteins involved in non-host-specific defense. Plants also contain chitinase-like proteins that are not induced by pathogens or stresses. In many cases, these chitinase-like proteins have been shown to lack detectable chitinase activity. Chitinase-like proteins may play an important role during normal plant growth and development. For example, *AtCTL1* is constitutively expressed in many organs of *Arabidopsis*. Mutations of *AtCTL1* lead to ectopic deposition of lignin in the secondary cell wall, reduction of root and hypocotyl lengths, and increased numbers of root hairs. It was suggested that this gene could be involved in root expansion, cellulose biosynthesis, and responses to several environmental stimuli. In particular, co- expression of some CTLs with secondary cell wall cellulose synthases (CESAs) was reported. It has been suggested that these chitinase-like proteins could take part in cellulose biosynthesis and play a key role in establishing interactions between cellulose microfibrils and hemicelluloses. The xylan-type secondary wall is the most common secondary wall in land plants and is characteristically rich in cellulose, xylan, and lignin. Compared to typical xylan-type secondary walls, gelatinous layers are enriched in cellulose, have a higher degree of cellulose crystallinity, larger crystallites, and a distinctive set of matrix polysaccharides (see and references therein). Presumably, cellulose synthase genes have a significant role in gelatinous cell wall formation, but the expression patterns of the complete flax *CESA* family has not been described to date. It is known that at least three isoforms of CESAs comprise the cellulose synthase rosette: CESA1, CESA3, and CESA6 are required for cellulose biosynthesis in primary cell walls, whereas CESA4, CESA7, and CESA8 are required for cellulose biosynthesis during secondary wall deposition. Flax is a useful model for comparative studies of cell wall development: different parts of the flax stem form a primary cell wall, xylan type secondary cell wall, or gelatinous cell wall; these stem parts may be separated and analyzed by diverse approaches, including functional genomics. Furthermore, the publication of a flax whole genome assembly facilitates a thorough study of key gene families. In the present study, we measured expression of all predicted *LusCTL* genes of the GH19 family in various tissues including those that produce gelatinous-type and xylan-type cell walls. We also described the *LusCESA* gene family and measured expression of its transcripts in comparison to *LusCTLs*. Phylogenetic analysis of *LusCTL* and *LusCESA* genes identified distinct groups of *LusCTL* genes that were expressed preferentially at specific stages of bast fiber gelatinous cell wall development. # Materials and Methods ## Plant Growth Flax (*Linum usitatissimum* L.) var. Mogilevsky plants were grown in pots in a growth chamber at 22°C, with a light intensity of approximately 200 µE on a 16 h light/8 h dark cycle. Plants were harvested at the period of rapid growth (4 weeks after sowing). Plant material was sampled with respect to the location of the snap point, which is a mechanically defined stem position in which fibers undergo transition from elongation to secondary cell wall formation. The following seven samples were collected: 1. “Apex” – the apical part of stem (1 cm of length). 2. “TOP” – the following “apex” segment of stem above the snap point with phloem fibers in the process of elongation. 3. “MID” – the stem segment (5 cm of length) below the snap point which contained fibers at early stages of secondary cell wall thickening. 10 cm of the stem downwards from “MID” was divided into Peel (4), which contained epidermis, parenchyma cells, phloem fiber bundles and sieve elements and Xylem (5), which contained parenchyma cells, xylem vessels and xylem fibers. 6. “Fibers” – i.e. isolated phloem fibers were obtained by washing Peels in 80% ethanol in a mortar several times and gently pressing the fiber-bearing tissues with a pestle to release the fibers. 7. Roots. The number of biological replicates was three, with five plants in each replicate. ## Sequence Alignment and Phylogenetic Analysis Predicted amino acid and nucleotide sequences of CTLs (Pfam domain: PF00182) and CESAs (PF03552) were obtained from the Phytozome database v.9.0 (*Linum usitatissimum*, *Populus trichocarpa*, *Arabidopsis thaliana*). CESAs of poplar (PtiCESAs) were renamed according to Kumar et al.. A list of various well- characterized CTLs from different plant species was obtained from previously published works. Sequences were aligned using MUSCLE with default parameters, and a phylogenetic tree was constructed using MEGA5 based on the Maximum Likelihood and Neighbor-Joining methods, bootstrapping 1000 replicates, model WAG+G or JTT+G. Signal peptides for protein sequences were predicted using SignalP (<http://www.cbs.dtu.dk/services/SignalP/>), molecular weights, isoelectric points of the proteins were analyzed by ProtParam (<http://web.expasy.org/protparam/>). ## Reverse Transcription Quantitative Real Time PCR Total RNA from all plant samples was isolated using a Trizol-extraction method combined with an RNeasy Plant Mini Kit (Qiagen) according to the manufacturer’s instructions. RNA quality was evaluated by electrophoresis using a BioAnalyzer (Agilent), and no degradation of RNA was evident. Residual DNA was eliminated by treatment with DNAse I using the DNA-free kit (Ambion). Gene specific primers for CTL and CesA genes were designed using Universal ProbeLibrary Assay Design Center (<http://www.roche-applied-science.com/shop/CategoryDisplay?catalogId=100 01&tab=&identifier=Universal+Probe+Library>). One microgram of total RNA was reverse-transcribed using RevertAid H Minus First Strand cDNA Synthesis Kit (Thermo Scientific). The cDNAs were diluted 1∶32 with nuclease free water. Real- time PCR was performed in a 7900 HT Fast real-time PCR system (Applied Biosystems, USA). Each 10 µL real-time PCR cocktail contained 2.5 µL of 0.4 µM concentrations of both forward and reverse gene-specific primers, and 2.5 µL of cDNA, 5 µL of 2×Dynamite qPCR mastermix (Molecular Biology Service Unit - University of Alberta) which included SYBR green (Molecular Probes) and Platinum Taq (Invitrogen). The thermal cycling conditions were 95°C for 5 minutes, 40 cycles of 95°C for 15 seconds, and 60°C for 1 minute. A 60–95°C melting curve was performed to confirm the specificity of the products. Threshold cycles (CT) were determined using 7900 Fast Software. C_T values were normalized using eukaryotic translation initiation factors 1A, 5A (*LusETIF1, LusETIF5A*) and glyceraldehyde 3-phosphate dehydrogenase (*LusGAPDH*) gene from flax. From each of three biologically independent cDNA samples, two independent technical replications were performed and averaged for further calculations. ΔΔCT values were generated using the apex sample as a reference. Relative transcript abundance calculations were performed using comparative C_T (ΔC_T) method as previously described for flax tissues (TOP/Apex, MID/Apex, Peel/Apex, Xylem/Apex, Fiber/Apex, Root/Apex). Heat maps of expression levels of some genes were then created with MeV v4.8 (Multi Experiment Viewer, <http://www.tm4.org/mev>.) using the means of ΔC_T. # Results ## *LusCTL* Phylogenetic Characterization We searched within the flax genome assembly (version 1.0) for predicted genes with homology to Pfam domain PF00182, which is characteristic of chitinases of the glycosyl hydrolase family 19 (GH 19) family. This search identified 37 predicted chitinase or chitinase-like genes (referred to here collectively as *LusCTL*s). However, only three of the predicted proteins (*LusCTL8, LusCTL10*, and *LusCTL15*) contained a conserved chitin-binding domain (CBD), suggesting that not all of the LusCTLs use chitinase as a substrate. The mean predicted protein size of the 37 *LusCTLs* was 246.5 aa (or 27 kDa), and the majority (30/37) were predicted to be secreted. The labels assigned to the 37 predicted *LusCTLs* are shown in. *LusCTL1* and *LusCTL2* were so named because they encoded proteins that were most similar to *CTL1* and *CTL2*, respectively, which have been characterized in other species (e.g. *A. thaliana* and *G. hirsutum*). The gene *LusCTL37* was predicted to encode only a protein of 69 aa, which is much shorter than the rest of the LusCTLs, and so it was not used in further analyses. The LusCTLs and their inferred phylogenetic relationships are shown in. Based on this dendrogram, the predicted LusCTLs were divided into three groups: Group A included *LusCTL1–6*, Group B included *LusCTL7–14*, and Group C included *LusCTL15–36*. The flax-specific tree shown in was expanded by the addition of representative GH19 CTLs from other species. In this multispecies tree, LusCTLs of Group A, which includes *LusCTL1* and *LusCTL2*, were part of the same clade as the well-characterized *AtCTL2* of *A. thaliana* and *GhCTL1*, *GhCTL2*, *GhCTLVII* of *G. hirsutum,* The Group B *LusCTLs* (*LusCTL7–14*) were in the same clade as the previously defined Classes I, II, III, GH19 chitinases. Most of group B was in the same sub-clade as Class II, although none of the previously defined Classes I–III were monophyletic in our analysis. Finally, our Group C *LusCTLs* (*LusCTL15–36*) formed a monophyletic clade with representatives of the previously defined Class IV GH19 chitinases. ## *LusCTL* Transcript Expression Quantitative real-time reverse-transcription PCR (qRT-PCR) was performed to study *LusCTL* expression patterns of *L. usitatissimum* genes of chitinase-like proteins in various tissues and stages of development. These tissues and their names as used here are equivalent to the names used in previous studies. Only 34 sets of primers were used in this assay, because members of each of two pairs of *LuCTLs* could not be distinguished by unique primers: *LusCTL28* and *LusCTL29* (95.6% aa and 96.3% nt identity), and *LusCTL33* and *LusCTL34* (99.6% aa and 98.8% nt identity). Thus a common set of primers was used for each of these pairs. We observed that transcripts of *LusCTL1* showed enriched levels of expression (compared to the apical part of stem) in tissues in which cell walls were undergoing thickening in xylem and in phloem fibers. Transcripts for this gene were enriched 57-fold in xylem, 28-fold in the MID region, and 20-fold in fiber. Another predicted CTL, *LusCTL2*, showed a similar pattern of enrichment in secondary-wall bearing tissues (8.3, 4.5 and 1.4-fold higher in xylem, MID and fiber, respectively, compared to the apex), although the magnitude of its enrichment was not as strong as *LusCTL1*. These two *LusCTLs* had high sequence similarity to each other (91.9% amino acid identity) and had similar patterns of expression as compared to each other in the various flax tissues. A subset of *LusCTL* genes (*LusCTL10*, *LusCTL11, LusCTL19, LusCTL20, LusCTL21, LusCTL23, LusCTL24, LusCTL26*) had high relative expression in tissues that contained phloem fibers (MID, peel, fiber) but low relative expression in xylem. Three of these genes (*LusCTL19*, *LusCTL20, LusCTL21*) were enriched \>40-fold in fibers compared to the apical part of stem. These three genes had high similarity to one another (76% identity between *LusCTL19* and *LusCTL20* as well as between *LusCTL19* and *LusCTL21*; 91% identity between *LusCTL20* and *LusCTL21*). ## *LusCESA* Phylogenetic and Expression Characterization To provide context for the expression patterns of *LusCTLs,* and to test whether the expression pattern of cellulose synthase (*LusCESA*) genes differed between gelatinous fibers and cells with a xylan type of secondary cell wall, expression of *LusCESAs* in different flax tissues was analyzed. We identified 14 predicted *LusCESAs* in the flax whole genome assembly by searching predicted proteins for the conserved cellulose synthase domain (Pfam PF03552). No putative *LusCESA7* genes were found in the original published genome published (v1.0). However, though BLAST alignment of the CDS of *Arabidopsis* and poplar CESA7 sequences, two scaffolds (scaffold_57 and scaffold_464) of the flax genome assembly were identified as encoding CESA7 homologs, and these were annotated using the Augustus server (<http://bioinf.uni-greifswald.de/augustus/>). Thus, all 16 predicted *LusCESAs* were aligned with well-characterized *AtCESAs* from *A. thaliana* and *PtiCESAs* from *P. trichocarpa*. This alignment was used to construct a phylogenetic tree and annotate the *LusCESAs*, which were named according to the established *A. thaliana* and *P. trichocarpa* nomenclature systems. The number of *LusCESAs* and *PtiCESAs* isoforms identified for each of the eight major types of CESAs was similar except in the case of CESA3, where one more gene was identified in *P. trichocarpa* than in *L. usitatissimum*. The *LusCESA* appeared to be typical of other genes in this family in that they were large integral membrane proteins with eight predicted transmembrane domains, a hydrophilic domain that faces the cytosol, and a zinc finger domain at the N-terminus of proteins with the characteristic amino acid motif “CxxC” (specific for CESAs only). Relative differential expression of *LusCESA* genes in different tissues of the flax stem was estimated. *LusCESA4, LusCESA8-A, LusCESA8-B, LusCESA7-A, LusCESA7-B* had high expression in tissue that produce secondary walls (TOP, MID, Xylem, Fiber, Root). Transcripts of these *LusCESA* isoforms were the most enriched in Xylem, which contained cells with xylan-type cell walls, and in roots, where secondary vascular tissue (xylem) was also well-developed. These secondary cell wall type *LusCESAs* had also high relative expression in cellulosic fibers, although it was not as strong as for xylem. Changes in expression of the *LusCESA4, 7, 8* isoforms and “xylem-specific” *LusCTL1* and *LusCTL2* were well-correlated in different flax tissues. This group of genes was highly expressed in tissues with secondary cell walls (MID, Xylem and Roots). In contrast, the “fiber-specific” *LusCTLs* had very different patterns of expression in the same tissues: these had low level of expression in xylem, but high level of relative expression in tissues with gelatinous fibers (peel and fiber). # Discussion Certain fibers of many plant species form G-type cell walls, which are rich in crystalline cellulose. Expression of CTLs has been previously reported to be enriched during development of G-type cell walls, along with specific FLAs, LTPs and BGALs. In this work, we analyzed expression of all *LusCTL* genes of GH 19 in different flax tissues and compared this expression with *LusCESAs* and to their inferred phylogenies. In the flax genome, 16 predicted *LusCESAs* were identified. Previously only partial sequences of some flax CESAs were published. All 16 flax CESAs could be placed in discrete clades with *Arabidopsis* and *Populus* CESA homologs. We generally numbered *LusCESAs* in a way that reflects the association of each flax gene with its nearest relative in the *Arabidopsis* genome, as was done for CESAs of *Populus*. Following this pattern, the *LusCESA6A–F* genes we named as a group, similar to *PtiCESA6A–F* and were not distinguished as *CESA2/9/5/6* as in *Arabidopsis* clade. Most of the flax and *Populus CESA* genes are present as pairs of paralogs in their respective genomes, although there were three *LusCESA3* genes (*LusCESA3A–C*) for only two *Populus* genes and one Arabidopsis gene. *AtCESA1* and *AtCESA10* were represented by only one pair of genes (*LusCESA1A*, *B*) in flax. It is well established that proteins encoded by different sets of three CESA genes (CESA1, 3, 6 and CESA4, 7, 8) are required for cellulose synthesis during primary and secondary wall formation, respectively. The functional relationships of the various paralogs of *LusCESA*s (except *LusCESA4*) are presently unclear. According to the data obtained here, secondary cell wall *LusCESA4, LusCESA7-A, B* and *LusCESA8-A, B* were highly expressed both in the xylem cells with lignified cell walls (i.e. xylan type) and in the phloem fibers with thick gelatinous cell wall. This suggests that phloem fibers and xylem may use similar, rather than specialized rosettes. This is consistent with observations from poplar showing only minor differences in expression of cellulose biosynthetic genes in tension wood as compared to normal wood. The different properties of gelatinous and xylan type cell walls are therefore likely determined not by CESAs, but by other proteins associated with cellulose synthesis, which could include specific CTLs. We observed two *LusCTLs* that were expressed more strongly in xylem tissue than in any other tissue surveyed (*LusCTL1, LusCTL2*). The co-expression of certain isoforms of *LusCTL1, LusCTL2* and the secondary wall *LusCESAs* (*CESA4, 7, 8*) suggested a role for these LusCTLs in secondary cell wall development. As noted above, *LusCTL1* and *2* are highly similar to *AtCTL2* of *A. thaliana* and *GhCTL1*, *GhCTL2*, of *G. hirsutum*. The role of *CTL2*, and its close homolog *CTL1*, in cell wall biosynthesis is especially intriguing since associations between CTLs and primary or secondary cell wall synthesis have been reported in different plant species. *CTL2* is strongly upregulated during secondary wall formation in interfascicular fibers in *A. thaliana*. Reduction in crystalline cellulose content in *ctl1 ctl2* mutants was demonstrated, leading to the to the suggestion that *AtCTLs* are involved in cellulose assembly. Furthermore, in *P. trichocarpa*, expression of chitinase genes related to *AtCTL1*, *AtCTL2*, and *GhCTLVII* are highly correlated with secondary wall formation of xylem. It has therefore been proposed that CTL1 and CTL2 work in conjunction with primary- and secondary-cell wall CESAs, respectively. One of the hypotheses for CTL1/2 function is regulation of cellulose assembly and of interaction with hemicelluloses via binding to emerging cellulose microfibrils. However, the mechanism of CTL action in cell wall biosynthesis as well as substrates of catalytic activity (if any) remains unknown. It was suggested that the likely substrates of plant chitinases may be arabinogalactan proteins, chitooligosaccharides and other GlcNAc-containing glycoproteins or glycolipids, and the mechanism by which CTLs act is more likely to involve binding of chitin oligosaccharides than catalysis. Also, it is assumed that chitinases may participate in the generation of such signal molecules that regulate the organogenesis process. Although relative expression of *LusCESA* (*4, 7, 8*) and *LusCTL1, LusCTL2* in xylem tissue was higher compared with phloem fibers, we cannot exclude involvement of these LusCTLs in phloem fiber cell wall development. At the same time, a distinct group of *LusCTLs* (*LusCTL19, LusCTL20, LusCTL21*) had very high enrichment in samples with phloem fibers (MID, peel, fiber) with a low level of expression in xylem. According to the phylogenetic tree, these LusCTLs (group C) were most similar to the previously defined Class IV chitinases. High constitutive expression of Class IV (along with Class I) in most organs of *A. thaliana* under normal growth conditions has been previously noted. Detailed bioinformatic characterization of genes of LusCTL distinct group should be conducted in future. Probably *LusCTLs* that are highly expressed in fibers may be specific to the gelatinous cell wall, while *LusCTL1* and *LusCTL2* are essential for wall thickening in general. # Conclusion High expression of specific *LusCTLs* was observed in different types of thick cell wall producing tissues. *LusCTL1* and *LusCTL2* were preferentially expressed during secondary wall deposition of xylem and were coexpressed with secondary cell wall *CESAs* (*4, 7, 8*). Another group of *LusCTLs*, (especially *LusCTL19, LusCTL20, LusCTL21*) were highly expressed in bast fibers, which have cellulose-rich, gelatinous walls. The group of fiber-enriched LusCTLs was expanded in flax compared to species that do not produce bast fibers, suggesting that these genes might play a unique role during gelatinous cell wall development in general and cellulose synthesis in particular. It is possible that the presence of fiber-specific *LusCTLs*, along with other key participants, determines differences in mechanisms of xylan and gelatinous cell wall formation. To establish the functions of these *LusCTLs* further characterization, including analysis of enzyme activity and structure, is necessary. Chitinase-like proteins remain one the most mysterious proteins in the plant cell wall. This study provides further evidence of their involvement in the process, and distinguishes between groups of CTLs involved in different type of cell wall development. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NM TG MKD. Performed the experiments: NM. Analyzed the data: NM TG MKD. Contributed reagents/materials/analysis tools: NM TG MKD. Wrote the paper: NM TG MKD.

# Introduction Hypospadias (H) is a defect in the urethral development occurring between the 8^th and the 16^th week of pregnancy causing the meatus to be ill-located on the ventral side of the penis. While H is a frequent congenital malformation that needs early correction by urethroplasty, its epidemiology has remained imprecise. Since the early 2000s, H incidence has shown large and unexplained differences across populations and time periods. In 2003, H incidence was estimated to be 8/10,000 live male births in the US and 18.6/ 10,000 in a study of 23 European national registries. H incidence had increased until 1999 according to 36 years of Swedish data, but no variation was observed between 2004 and 2010 across European registries. Since less than 1% of H cases are due to rare monogenic syndromes, the vast majority are still called “idiopathic” and are presumably due to unknown developmental interaction between gene variants and environmental factors. Genetic predisposition is important, as shown by the 12 to 20-fold augmentation of H) among first-degree relatives. Genome-wide studies have identified a few low-risk genomic variants associated with H, while epigenetic studies have started to explore the association of H with DNA methylation. The contribution of environmental factors is likely to be prominent, but has been difficult to assess. A link of H with decreasing male fertility suggests the existence of a “testicular dysgenetic syndrome” that has emerged in developed countries. Agricultural pesticides—prominently suspected endocrine disruptors—remain the main environmental suspects, but studies have yielded variable results across countries, regions and time periods, Case-control studies of parental occupation did not show an increased risk of H in sons of women working in farming and gardening, but none had specifically focused on people working in vineyards. A direct way to assess the exposures of cases and controls to agricultural pesticides is offered by Geographical Information Systems (GIS) and the increasing availability of environmental data bases informing on land use. For example, the comparison of exposures of 354 cases with 727 controls with regard to 38 pesticides within 500 m of pregnant mother’s address in Arkansas found a slight association of H with diclofop-methyl (OR = 1.08), and surprisingly a stronger protective effect of other pesticides (e.g. permethrin: OR = 0.37). Among 304,906 singletons born in North Carolina, the estimated amounts of pesticides used within 500 meters of maternal residence yielded only a marginal association with H in 856 cases. In France, an increase of H incidence at the proximity of Montpellier vineyards has been debated. The fact that many vineyards are located near houses in many countries calls for robust epidemiological evidence based on reliable controls. Notably, a reliable control is”someone who–if he had developed the disease- would have been recruited among the cases of the study”. Surgical centers do not usually record epidemiological information on the exposure history of their patients, but all of them record basic demographic information for billing purposes, including mother’s address at childbirth. This low-cost hospital information crossed with available geographic environmental databases was used for the current study. We have used two independent approaches to define appropriate controls to which the cases should be compared. First we used as controls the cases of C recruited in the same network of surgical centers than those used for H. These C patients could be considered as H controls since all included cases of C had a normal penis and meatus, while none of the studied H cases had mal-descended testes. We did a second independent case-control analysis, where controls were “virtual” controls (VC) defined through an algorithmic approach described below. # Material and methods ## Patients To be included in this study, H cases were the clinically significant cases that had reconstructive surgery in the line of the current practice, and C cases had to have orchidopexy. Minimal glandular hypospadias cases that are considered clinically not significant, thus not operated, were not part of our study. Under coordination by L.Esterle and Pr P.Mouriquand (Lyon), the following 17 French centers of urological pediatric surgery collected retrospectively the date of birth and address at birth of the parents of all boys born after January 1st 1980 who had H (coded as ICD Q54) and C (coded as ICD Q53): Angers (Pr G. Podevin), Bordeaux (Pr B. Frémond), Caen (Pr P. Ravasse), Colmar (Pr S. Geiss), Grenoble (Pr B. Boillot), Lyon (Pr P. Mouriquand), Marseille and Toulon (Pr P. Alessandrini), Vandoeuvre-lès-Nancy (Pr JL. Lemelle), Nantes (Pr MD. Leclair), Reims (Pr PJ. Lefebvre), Rennes (Pr E. Dobremez), Saint Étienne (Pr F. Varlet), Strasbourg (Pr R. Moog), Saint Vincent de Paul at Paris (Pr F. Bargy), Toulouse (Pr J. Moscovici), Valence (Pr B.Defauw). All children operated for hypospadias in the University hospitals participating to this study had all been carefully examined at a pediatric endocrinology unit belonging to the Centre National for Disorders of Sex Development (DSD). Cases resulting from a genetic disorder of sex differentiation or hypogonadotropic hypogonadism were excluded. At total, 8,766 H and 13,105 C cases were recorded by May 2013.. We computed the distance of each case to the nearest vineyard (see below), and constituted the database which is analyzed in this paper. The distribution of the year of birth of the H and C cases is shown in. ## Ethics, data confidentiality The research protocol was approved by the Ethics committee of Ile de France (DC-2009-1041). The parents of children provided written informed consent. To protect the confidentiality and data of participants, all were assigned a unique identification number without identifying information. The security conditions of the database received the agreement of the “national information science and liberties commission” (ref CNIL 912529). Under this agreement, CNIL can allow external researchers to get access to the original databases. Requests must be sent to the board of the HCFCG (HypoCrypto French Collaborative Group, Inserm, Batiment Pincus, Hopital de Bicêtre, 94276- Le Kremlin Bicêtre). ## Assessment of the residential proximity of people to vineyards We used the Land Parcel Identification System (LPIS) to assess the residential proximity of the homes of the children to vineyards. The origin of LPIS is a decision of the European Union Council Regulations in 2003 to create an information system on the land use of European agricultural parcels. In LPIS, an agricultural parcel is defined as a continuous area of land on which an individual farmer cultivates a single crop. The data is encoded using the shapefile format, which is a GIS file format describing vector features such as points, lines, and polygons. Polygons are used in order to specify the borders of the parcels. The geographic projection system used to specify the edges of these polygons is Lambert-93. Each item element in shapefile has attributes. The attribute of a polygon is a number that identifies the crop being grown on that parcel. LPIS information was made available for years 2010 to 2014. We used in this work the 2012 database that was freely downloadable from the French data.gouv.fr web portal (<https://www.data.gouv.fr/fr/datasets/registre- parcellaire-graphique-2012-contours-des-ilots-culturaux-et-leur-groupe-de- cultures-majorita/>). At total the current study is based on the 340,336 LPIS parcels with vineyards, for a total surface of 381,395 Ha. The exposure metric that we use in this work is the individual’s distance to the nearest parcel of vineyard. In practice, we had to calculate these distances for more than 50,000 individuals (cases plus controls) to 340,336 parcels of vineyard. Each parcel was modeled as a polygon. To avoid calculating exact distances to polygons that are sufficiently far away for having any chance to induce a risk, we chose a cutoff at 10,000 meters and pooled all distances greater than 10,000 m in a single class. ## Statistical analysis of the risk associated with distance to the nearest vineyard Two analyses were performed. The first one was based on patients whose home was located in three large classes of distances to a vineyard: 0–100 m, 100–200 m and 200–500 m. The second analysis studied the risk at distances of less than 50 m from a vineyard, despite the fact that only 82 cases were from homes located there. The three classes of “small” distances studied were 0–10 m, 10–20 m and 20–50 m. In both analyses we studied H risk in two independent case-control studies, taking the 5,000–10,000 m class of distance as reference. In the first analysis, we compared the distances to the nearest vineyard in the two groups of H and C cases. For this comparison, we computed the variation of the odds ratio of the H risk, by reference to the C risk, in the different zones of distance to a vineyard, taking the 5,000 to 10,000 m distance as the reference. In the second analysis, we computed the variation of the odds ratio (OR) for the H risk of people living within zones close to vineyards, taking as reference the risk of children born to mothers living between 5,000 and 10,000 m, i.e. roughly 21% of the patients. The specificity of our approach is that the controls of this case-control study were “virtual controls” (VC). A “virtual control” is a geographical point of the map which represents a real physical person that would have been chosen as control, if it was practically possible to sample randomly a subject within the same source population as the “case”. We modeled the source population of each hospital as composed of the population living in the circle around this center, in which there was a proportion p of the cases. In the analyses presented here, p was taken at 80%. The algorithm used to sample the VC is the PPS algorithm implemented by Jack G. Gambino as an R package (PPS stands for probability proportional to size). <https://rdrr.io/cran/pps/>. This algorithm allows to sample people in a territory proportionally to the local density of population. The French National Institute of Statistics (INSEE) 2009 database (<https://www.insee.fr/fr/statistiques/2520034>) provided us with an estimate the population density. This database provides the 2009 French density of population in each age-class within a 200 m resolution grid. From this 2009 information, we can estimate the spatial distribution of children born, say, in 1995: assuming that mobility was low (which is the case in France), this distribution is simply the distribution of the children aged 14 (= 2009–1995) in the 2009 grid. This spatial distribution is then used to randomly chose–thanks to PPS- a cell within all the cells that constitute the source population. The VC is then set at the center of the 200m x 200m cell. We took 5 virtual controls per case by repeating the algorithm 5 times. ## Socioeconomic environment The 2009 version of the geographic information system of the French National Institute of Statistics (INSEE) (<https://www.insee.fr/fr/statistiques/2520034>) provided us with an estimate of the density of population, by age class, in each square of a 200m x 200m grid, **repeat** and an estimate of the local socioeconomic environment by using the Townsend deprivation index (TDI) at the place of birth. A higher TDI score implies more severe social deprivation. ## Statistical packages The odds ratios quantifying the association of disease and distance to the nearest vineyard were estimated using logistic regression adjusted for the coordinating center, to account for difference in the distance distribution between centers. Additional analyses were adjusted for TDI. Subgroup analyses by birth period were also carried out. Analyses were carried out using the R statistical software version 3.6.3, with the `glm` function of the `stats` package. # Results The distribution of the distance of the children affected with H or C to the nearest vineyard is shown in. The two distributions were significantly different (chi square = 38, 9 df, p\<10⁻⁴). H cases were more frequent than C in all classes of distances below 5,000 m to vineyards, with the exception of those located between 0 and 10 m. We first analyzed wide classes of distance (0–100 m, 100–200 m, 200–500 m, 500–1000 m), where the number of H patients was large (between 145 and 376). We found that odds ratios were \> 1 in all these classes of distance, and were significantly \> 1 in the 0–100m class when controls were patients with cryptorchidism (Column 1 of), and when controls were “virtual controls” after adjustment on the Townsend index (column 2b of). In the second part of the analysis, we studied the patients whose houses were located at close distances of vineyards (less than 50 m), although their numbers were small. Again, we estimated the risk of H possibly associated with a short distance from the nearest vineyard by computing the odds ratios of H patients versus C patients, and the odds ratios of patients versus VC. The two analyses gave comparable results. The odds ratio of H patients in the 0–50 m was significantly \> 1 (column 1, where controls are patients with C and column 2b, where controls were “virtual controls”). When we focused on classes of addresses at very short distances of a vineyard (second part of), we found that the odds ratio of H patients belonging to a home located between 10 and 20 m from the closest vineyard was significantly \> 1 (p = 0.029 when controls are C patients, p = 0.007 when controls are “virtual controls”). The results were comparable after adjustment on the Townsend Index. # Discussion-conclusion Common “idiopathic” H results from unknown mechanisms, among which endocrine disruptors are suspected to interfere with gonadal and genital development of the male fetus. We found in this work that the risk of H was higher in children born to mothers living close to a vineyard. The strength of our observation is that it was achieved through two independent methodological approaches. Both methods guarantee that controls were from the same hospital recruitment territory than cases. This avoids the bias of controls made up of children with common pediatric diseases that usually live close to the center, whereas cases of H come from far away for a very specialized intervention. Such bias could create the false impression that patients from rural areas would be more at risk than those from urban areas. The controls who were used in the two approaches match the ideal definition of what must be a control group: "individuals who, if they had had the disease (H), would have been recruited in the same centers as the (H) cases were". In the first approach, the controls were children with C and no H having been operated in the same clinical centers as H. A few studies found a possible positive association between C and exposure to pesticides. However, even if this link is true, it would not invalidate our results. Indeed, the consequence of such a link would be that the “true” odd-ratio is even greater than the one we found. In the second approach, the “virtual controls” were not real persons. However the algorithm that we used guarantees that they are identical to the population of children that would be sampled in the “territory” of the cases. To realize this, we modeled the recruitment “territory” of an hospital as the population living in a virtual circle around this hospital containing a proportion p = 80% of the cases. We also performed the analyses using other models to define hospital territory, and obtained similar results. The increased H prevalence at short distances from vineyards is an indirect indication supporting, but not proving, a harmful effect of pesticides since our study did not qualify nor quantify the loads and nature of pesticides spread in the vineyards. Indeed, this was an impossible task given the 31-year duration of our H case collection. Not only pesticides, such as fungicides, insecticides, herbicides, bactericides, rodenticides, fumigants, but also fertilizers and other toxic chemicals were and still are commonly used for viticulture. During the long studied period providing our data, the complexity, variety and dynamic mix of chemicals used in vineyards were considerable, and no database was available to quantify precisely their nature and load. The found association of vineyard proximity with H risk could also be due to other environmental factors present close to vineyards. Known endocrine disruptors commonly used in viticulture at the time of the study are vinclozolin, procymidone, and linuron, all forbidden since 2007–2008, but still present in some soils close to vines. The fungicide vinclozolin and its metabolites act as androgen receptor antagonists. Procymidone is another anti-androgenic fungicide that can induce hypospadias in rodents. Linuron, a widely used herbicide, is a weaker androgen receptor antagonist than procymidone and vinclozolin. Nitrates are another category of inorganic pollutants that can disrupt gonadal steroidogenesis. It was interesting in this respect to search whether the association of H with vineyard proximity was stronger in the years before 1995, but the small number of cases did not allow a reliable analysis. Agricultural substances may diffuse via surface runoff, leaching to field drains, and atmospheric spray drift. Pesticide drift may occur during ground application and after it, since applied substances have different volatility and local persistence. Our choice of focusing on the narrow zone bordering vineyards was influenced by methods used to apply pesticides, which never include aerial sources such as in the US but only land-based devices. Vineyards are a privileged place for pesticide spraying, undoubtedly one of the most important of all French agriculture. A weakness of our study is that it did not get occupational data for the mothers of cases, although one can expect that a much higher percentage of those living in the vicinity of vineyards are occupationally involved in wine producing activity, thus possibly exposed to pesticide stocks, bisphenol A, phthalates, or other potential endocrine disruptors. It must be noted that the possible risk associated to being a fœtus developed within a short distance of a vineyard concerns only a small proportion of the H cases: H cases born in a house at less than 20 m from a vineyard represent only 0.5% of the total number of H patients (those born at less than 50 m are 1%, and at less than 100 m are 1.8% of the overall H prevalence in France). The VC approach makes the best possible use, at almost no cost, of the huge data resource that is sleeping in hospitals instead of serving epidemiology. It does not require filling questionnaires or track families to get controls, but instead relies on the basic medical information system. In addition, the method avoids some selection bias, in that we did not have to obtain subjects’ consent to participate, as we used only information in the public domain. For all of these reasons, we believe that the VC method could be used increasingly for well-defined diseases using more detailed, and frequently refreshed environmental databases. A limitation of the method is that VC cannot be studied to compare biological exposome with cases, including pesticide and metabolite concentrations in biological fluids or cells. In conclusion, we have found a significant statistical link between the occurrence of H and the residence of pregnant mothers close to vineyards. However we are conscious that generalizations that focus on ‘positive’ findings rather than more comprehensive views that incorporate null findings and study limitations should be avoided. Future environmental research will have to account for the changing nature and loads of pesticides in modern viticulture. This could be done with the method presented, if high-resolution public reliable databases on the geographical use of different pollutants become available to public health research. In addition, further identification of precise hormone- disruptors will require a deep, systematic and specific chemical investigation of vineyards and their close neighborhood. More generally, the technique of “virtual controls” that was used in this work can be applied to search for candidate environmental factors in fetal diseases where the only available information is the location of the patient at birth. # Supporting information We gratefully acknowledge the administrative and medical staff of the surgical centers for providing the address of patients used in this study. [^1]: The authors have declared that no competing interests exist.

# Introduction Telomeres are nucleoprotein complexes that protect the physical ends of linear eukaryotic chromosomes, playing a crucial role in maintaining chromosome stability and integrity. In all vertebrates the DNA component of telomeres consists of the non-coding (TTAGGG)_n motif, which produce long tandem repetitions varying greatly in size between species, individuals and even cell types. The telomeric motif demonstrates a remarkable evolutionary conservation across vertebrate species. The telomere-specific complex associated with the telomeric sequence has been described as "shelterin". Shelterin is composed of three proteins (TRF1, TRF2 and POT1) that directly recognize the (TTAGGG)_n motif, and are interconnected by three additional proteins (TIN2, TPP1 and Rap1) to form a duplex structure (for a review see). The telomeric motif is synthesized by telomerase, a reverse transcriptase-like enzyme, which contains an RNA subunit and a catalytic protein subunit called telomerase reverse transcriptase. Telomerase uses the RNA template to add additional sequences directly to the telomeres. In humans telomerase is expressed in embryonic tissues and specific germline cells whilst in adults, the enzyme can be detected mainly in the testis, and is absent in most normal somatic cells, in non-dividing oocytes and mature spermatozoa. The main role of telomeres is to protect the edges of the linear chromosomes from degradation, recombination or fusion, preventing the chromosomal ends from being recognized as double-strand brakes by DNA repair machinery. Furthermore, the DNA replication machinery cannot completely replicate the ends of linear chromosomes as there would not be any template strands to guide its synthesis ("end replication problem"). In each cell division 50–200 bp are erased from the edges of the chromosomes decreasing the chromosome length and eventually affecting the inner genetic loci. Telomerase preserves the edge of the chromosomes by adding "expendable" telomeric motifs *de novo*. However, in several cell types, such as human somatic cells, telomeres become shorter after subsequent replications, resulting in a minimum amount of telomeric sequence, leading to replicative senescence and ultimately cell death. This phenomenon has been described as the "telomere hypothesis of cellular aging", a theory that proposes that telomeres serve as a "mitotic clock" controlling lifespan. An additional role of telomeres is the maintenance of the chromosome topology in the nucleus matrix and the correct alignment of chromosomes for recombination during the first meiotic prophase. Another important function of telomeres is the silencing of adjacent genes, a phenomenon known as "telomere position effect". As well as the crucial role of telomeres at the edges of chromosomes, non- terminal telomeric motifs known as interstitial telomeric sequences (ITSs) or interstitial telomeric repeats (ITRs), have been observed in many species. The pioneer publication by provided the first cytogenetic evidence of this, reporting that 55 out of the 100 studied species of vertebrates had ITSs. Many more cases were described in the following years in vertebrates, including amphibians, fish, birds, rodents, marsupials and primates. Based on sequence organization and genomic location, Ruiz-Herrera et al. identified two different types of ITSs: short ITSs (s-ITSs) and heterochromatic ITSs (het-ITSs). Other authors have classified ITSs in more detailed categories as short ITSs, long subtelomeric ITSs, fusion ITSs and pericentromeric ITSs. S-ITSs are short sized telomeric repetitions located in internal sites of chromosomes, present in all completely sequenced mammalian genomes (at least 83 in human, 79 in chimpanzee, 244 in mouse and 250 in rat), but often not detectable by cytogenetic techniques such as fluorescent *in situ* hybridization (FISH). It was initially thought that s-ITSs were derived from the telomeric fusion of ancestral chromosomes. However, recent studies concluded that s-ITSs are not in fact associated with chromosomal rearrangements but instead were probably inserted by telomerase during the repair of DNA double-strand breaks. This hypothesis is supported by the frequent association of transposable elements such as SINEs and LINEs with s-ITSs. Het-ITSs are large stretches of telomeric sequences (up to hundreds of kb) localized mainly in heterochromatic chromosomal regions such as in centromeric or pericentromeric areas or within the chromosome arms. In contrast to s-ITSs, het-ITSs are only present in a limited number of species and it is widely believed that they correspond to the remnants of ancestral chromosomal rearrangements which occurred during karyotype evolution. As far as we know, ITSs have been described in only 22 lizard species and never in snakes. In general, squamate reptiles are often considered as a group with evolutionary conserved karyotypes. This view has been supported by classical cytogenetics techniques such as conventional staining and C-banding as well as by chromosome painting, gene mapping, and qPCR mapping of genes linked to sex chromosomes. Considering that particular types of ITSs represent relics of chromosome rearrangements, the conservation of karyotypes in squamates suggests that ITSs should be relatively rare in this group. In order to test this hypothesis, we reviewed published data on the occurrence of ITSs and supplemented it with our novel description of ITSs distribution in 13 species of squamates based on FISH experiments. # Material and Methods ## Specimens and chromosomal preparations The distribution pattern of telomeric motifs was studied in the karyotypes of 30 species of squamate reptiles (28 lizards and 2 snakes), belonging to 17 families from our collection of metaphase chromosome spreads. The specimens originated from pet trade (the companies Animalfarm CZ, Zoopet Sandy, Happy Reptiles, B.A.R. and Zoo Shop Želvička) and were maintained in the reptile breeding laboratory of the Faculty of Science, Charles University in Prague, Czech Republic (accreditation No. 24773/2008-10001). Blood samples were taken from caudal or brachial vessels. The animal procedures were carried out under the supervision and with the approval of the Ethics Committee of the Faculty of Science, Charles University in Prague followed by the Committee for Animal Welfare of the Ministry of Agriculture of the Czech Republic (permission No. 29555/2006-30). Metaphase chromosome spreads were prepared from whole blood cell cultures following the previously described protocol with slight modifications. Briefly, the small amount (approx. 40 μl) of the peripheral blood was cultured for a week at 30°C in Dulbecco’s Modified Eagle’s Medium (Sigma-Aldrich), enriched with 10% fetal bovine serum (Baria), 0.5% penicillin/streptomycin solution (Gibco), 1% L-glutamine (Sigma-Aldrich), 3% phytohaemagglutinin (Gibco), and 1% lipopolysaccharide (Sigma-Aldrich). Chromosome preparations were made following standard procedures including a 3.5 hours colcemid treatment, hypotonization, and fixation in 3: 1 methanol:acetic acid. ## Fluorescent in situ hybridization (FISH) A specific probe for the telomeric motif (TTAGGG)_n was produced and labelled with dUTP-biotin by PCR, using the primers (TTAGGG)₅ and (CCCTAA)₅, without a DNA template, according to the methodology of Ijdo et al.. Briefly, a PCR reaction was performed in 50 μl final volume, including 0.4 μl of each primer (5 pmol/μl), 5 μl of 10× PCR buffer (Bioline), 2.5 μl MgCl2 (50mM), 1 μl dATP, dCTP, dGTP (10 mM each), 0.7 μl dTTP (10 mM), 1 μl dUTP-biotin (1 mM) and 1 μl BioTaq DNA polymerase (5 U/μl, Bioline). The PCR cycling conditions were as follows: 5 min at 94°C, 10 cycles of 1 min at 94°C, 30 sec at 55°C and 1 min at 72°C, followed by 30 cycles of 1 min at 94°C, 30 sec at 60°C and 30 sec at 72°C, with a final step of 5 min at 72°C. The PCR product was precipitated and re-suspended in 300 μl of hybridization buffer (50% formamide/2× SSC). Prior to *in situ* hybridization, 10 μl of the telomeric probe per slide was denatured at 75°C for 10 min and then chilled on ice for 10 min. In parallel, the metaphase slides were subsequently treated with RNase, pepsin, fixed with 4% formaldehyde, dehydrated through a series of 70%, 85% and 100% ethanol washes, denatured in 70% formamide/2× SSC at 75°C for 4 min, dehydrated again and air dried. Afterwards, the probe was applied to each slide and incubated at 37°C for 16–24 hours. Post-hybridization washes were subsequently carried out in 50% formamide/2× SSC at 42°C (3 × 5 min) and in 2× SSC (2 × 5 min). The slides were incubated in 100 μl of 4× SSC/5% blocking reagent (RocheAρχήφόρμαςTέλοςφόρμας) at 37°C for 45 min. The telomeric signal was detected using a modified avidin-FITC/biotinylated anti-avidin protocol for FITC signal amplification. In detail, we prepared two different solutions: a primary antibody solution with 300 μl of 4× SSC/5% blocking reagent, including 0.3 μl avidin-FITC per slide (Vector laboratories) and a secondary antibody solution with 200 μl of 4× SSC/5% blocking reagent, including 2μl biotinylated anti-avidin per slide (Vector laboratories). The FITC signal was enhanced by five subsequent applications of the primary (three times) and the secondary (two times) antibody solutions at 37°C for 30 min each, using 100 μl of each antibody solution per slide, with intermediate washes in 4× SSC/0.05% Tween20 (3 × 5 min). Afterwards, the slides were dehydrated through an ethanol series, air dried, counterstained with 4',6-diamidino-2-phenylindole (DAPI) and mounted with Vectashield anti-fade medium (Vector Laboratories). ## Microscopy and image analyses An Olympus Provis AX70 fluorescence microscope with a DP30BW digital camera was used to take grayscale images that were processed with DP manager imaging software (Olympus) to record the pattern of the telomeric repeats within the chromosomal metaphases. ## Phylogenetic distribution The phylogenetic distribution of the presence/absence of ITSs across squamate reptiles was visualized using Mesquite v.2.75, based on the phylogenetic tree topology of Pyron et al.. # Results FISH with telomeric probe proved to be a valuable tool in revealing the topology of the telomeric motif (TTAGGG)_n in the karyotypes of squamate reptiles. Based on the distribution and the putative origin of the telomeric sequences within the chromosomes we distinguished the following topologies in the karyotypes: ## Karyotypes with only terminal distribution of telomeres In 17 species we observed telomeric sequences only at the expected terminal positions at the ends of the chromosomes. Specifically, this group includes the species *Acrantophis dumerili* (Boidae), *Cordylus tropidosternum* (Cordylidae), *Correlophus ciliatus*, *Oedura monilis* (Diplodactylidae), *Aeluroscalabotes felinus*, *Goniurosaurus lichtenfelderi*, *G*. *luii*, *G*. *splendens* (Eublepharidae), *Bunopus spatalurus*, *Homopholis fasciata*, *Stenodactylus sthenodactylus* (Gekkonidae), *Eremias velox* (Lacertidae), *Asaccus elisae*, *Tarentola annularis* (Phyllodactylidae), *Trogonophis wiegmanni* (Trogonophidae), *Varanus acanthurus* (Varanidae) and *Lepidophyma smithii* (Xantusiidae). ## Karyotypes with ITSs in centromeric regions Five species had karyotypes with telomeric motifs at the terminal positions of all chromosomes, and additional ITSs in centromeric regions of one or more chromosomal pairs. In detail, ITSs were detected at the centromeres of three submetacentric pairs in *Anguis fragilis* (Anguidae), seven chromosomal pairs in *Anolis distichus*, five chromosomal pairs in *Anolis equestris* (Dactyloidae), two large metacentric pairs in *Gerrhosaurus flavigularis* (Gerrhosauridae) and five chromosomal pairs in *Latastia longicaudata* (Lacertidae). ## Karyotypes with ITSs in pericentromeric regions Two lizard species, *Chamaeleo calyptratus* (Chamaeleonidae) and *Cordylus beraduccii* (Cordylidae), had telomeric motifs at the terminal positions of all chromosomes, and additional ITSs in pericentromeric regions of the largest metacentric chromosome pair. ## Karyotypes with ITSs within chromosome arms Two species exhibit ITSs signals between terminal telomeres and the centromeric/pericentromeric regions. In *Trioceros bitaeniatus* (Chamaeleonidae), ITSs are present at intermediate positions on six chromosomal pairs, and in *Pantherophis guttatus* (Colubridae) there is an interstitial telomeric band on a medium sized metacentric chromosome. ## Karyotypes with ITSs in numerous positions Finally, in four species we observed ITSs in numerous positions on chromosomes including all of the above mentioned categories (centromeric, pericentromeric and within the chromosome arms). This extensive accumulation was observed in *Aspidoscelis deppei* (Teiidae), *Lialis burtonis* (Pygopodidae), *Cyrtopodion scabrum* and *Agamura persica* (Gekkonidae). The phylogenetic distribution of ITSs suggests that in general ITS emergence/loss is evolutionary dynamic across squamates. A high incidence of ITSs is present in the sister families Teiidae and Gymnophthalmidae, while Iguania also possess the tendency to accumulate ITSs in their genomes, yet ITSs appear to have a rather random distribution across the other squamate lineages. # Discussion ITSs are present in members of all major lineages of squamates. Taking into account previous publications and our results, only 48.5% of squamates (n = 33 species) demonstrate the normal, expected distribution of telomeres at the edges of all chromosomes. Surprisingly, around half of the studied squamate species (51.5%, n = 35) show ITSs in centromeric regions, pericentromeric regions and/or within chromosomal arms. It therefore appears that the existence of ITSs in squamate genomes is not an exception, but rather a common event. The species with karyotypes demonstrating ITSs seem to be more or less randomly distributed across the phylogeny of squamates, with several families including species with both normal terminal telomeres and ITSs, while a higher incidence of ITSs typifies sister families Teiidae and Gymnophthalmidae and the lineage Iguania. Nevertheless, it should be noted that although our sampling includes members of most major lineages of Squamata, the total number of species tested for the presence of ITSs is still quite small and somewhat patchy, which precludes detailed statistical analyses of ITSs correlations and phylogenetic distribution. ITSs are commonly observed in centromeric regions of both bi-armed and acrocentric chromosomes. ITSs in the centromeres of bi-armed chromosomes might have originated from the remains of “old” terminal telomeres after Robertsonian fusion (e.g. in the gecko *Gymnodactylus amarali*;), while the ITSs on the centromeres of acrocentrics may be the result of extensive amplification due to their proximity to satellite sequences. It is well-documented that telomeres can be part of centromere repetitive elements. In a recent study, it was demonstrated that all centromeres of a vole species exhibit co-localization of ITSs with three other satellite sequences. These ITSs were cloned and sequenced, demonstrating 87% to 94% similarity to the terminal telomeric motif (TTAGGG)_n. The extensive amplification of ITSs in the centromeres of acrocentric chromosomes covering a large part of the centromeric region can be observed in several squamates, such as *Latastia longicaudata*. However, further studies are needed to reveal the type of repetitive elements co-localized with ITSs in squamates and to show if species with ITSs share a similar content of satellite sequences within their centromeres. Extended ITSs in intermediate positions in several lizard species may reflect remnants of past intrachromosomal rearrangements. Although sauropsids possess a relatively low rate of interchromosomal rearrangements, it has been shown in birds that intrachromosomal rearrangements occur rather frequently. In fact, whereas no interchromosomal rearrangements have been documented in the microchromosomes of the chicken, turkey or zebra finch, there have been numerous intrachromosomal rearrangements recorded in these species. In the same context, several pericentromeric inversions have been discovered in the chromosomal pairs 1–4 of *Anolis carolinensis* using *in situ* hybridization with BACs. Furthermore, the comparison between the homologous part of chromosome 15 of the chicken and chromosome X of *Anolis carolinensis* revealed extensive synteny of the gene content, and numerous intrachromosomal, but few interchromosomal rearrangements in the studied chromosomal region. Evidence for numerous intrachromosomal, but rare interchromosomal rearrangements based on interspecific chromosome painting was recently presented in geckos. In some cases telomeric-like sequences appear to accumulate at the heterochromatic part of sex chromosomes. The exact role of the accumulation of ITSs and satellite sequences (for a review see) on the highly heterochromatic W (e.g. in the gecko *Underwoodisaurus milii*;) or Y chromosomes remains unclear. It has been speculated that the accumulation of repetitive sequences on one pair facilitates the suppression of recombination between sex chromosome homologues, enabling the accumulation of sexually beneficial mutations on respective sex chromosomes. Some authors however suggest that the repetitive sequences may accumulate near the sex determining locus as a result of the suppression of recombination rather than inducing it ( and references within). Finally, closely related species with similar chromosome morphology seem to possess different patterns of ITSs distribution, e.g. species of the genus *Anolis*, *Cordylus*, *Paroedura* and *Leiolepis*. Such differences could be explained either by the dynamic nature of ITSs (e.g. as part of satellite DNA or transposable elements) or cryptic rearrangements. Many reptile linages, with the exception of their avian clade, show persistent telomerase activity in the somatic tissues of adults which might not only explain their extensive tissue regeneration potential, but also the existence of ITSs accumulation at numerous positions in their genome. In fact, telomerase appears to be active in all of the tissues of adult *Aspidoscelis sexlineata*, a species with ITSs accumulation. Furthermore, skin fibroblasts from the blue racer snake (*Coluber constrictor*) show increased telomerase activity after a high number of generations *in vitro*. Moreover, telomere length does not decrease with age in the water python (*Liasis fuscus*), but instead increases from approximately 7 kb at hatching to 28 kb at adult age, providing another exception to the hypothesis of “cellular aging”. In summary, we detected ITSs for the first time in the genomes of 13 species of squamate reptiles and documented that ITSs were observed in approximately half (35 out of 68) of the species of lizards, snakes and amphisbaenians, e.g., studied so far. Therefore we can conclude that the occurrence of ITSs is surprisingly high in this group of vertebrates which has otherwise stable and conserved karyotypes. This discrepancy suggests that, similar to birds, squamate reptiles may have a rather high rate of intrachromosomal rearrangement and a low rate of interchromosomal rearrangement. The origin of ITSs in some species of squamate reptiles may however be attributed to other factors such as high telomerase activity and/or the repair mechanisms of double-strand breaks, e.g. triggered by the activity of transposable elements. Future studies should be devoted to increasing the taxonomic scope of the testing of ITSs distribution across squamates and to address questions regarding the functional importance of these unusually frequent elements in squamate genomes. The authors would like to express their gratitude to Jan Červenka for the assistance with animal handling, to Petr Ráb for his continuous support and to Ettore Olmo, Marta Svartman and the anonymous reviewer for helpful comments. Christopher Johnson provided valuable language guidance. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MR LK MA MJP. Performed the experiments: MR MA MJP. Analyzed the data: MR LK MA MJP. Contributed reagents/materials/analysis tools: MR LK MA MJP. Wrote the paper: MR LK MA MJP.

# Introduction Around 3.0 billion years ago, cyanobacteria evolved the ability to perform oxygenic photosynthesis. This invention caused a rise in atmospheric oxygen that changed the diversity and complexity of species on earth dramatically. The plastids of all eukaryotic photosynthetic organisms are descendants of a cyanobacterial endosymbiont. Four kinds of cellular organizations are recognized in cyanobacteria, namely single celled species, linear filaments, linear filaments with heterocysts and branched filaments. Single celled and filamentous cyanobacteria can move across surfaces using a gliding mechanism. For members of the Nostocales, i.e. filamentous cyanobacteria with heterocysts, this motility is often restricted to hormogonia, a cell type formed during environmental stress or in symbioses with plants. Most genetic and molecular studies on cyanobacterial motility have been performed with the single celled species *Synechocystis* sp. PCC 6803 and *Synechococcus elongatus* PCC 7924 and one member of Nostocales, *Nostoc punctiforme*. Type IV pili act as driving complexes for gliding motility (often referred to as twitching motility) and have been intensely investigated in other bacteria such as *Pseudomonas aeruginosa* or *Neisseria meningitidis*. The type IV pilus driven movement is based on a retraction and extrusion of protein filaments. The PilA protein, formed from the PilA precursor protein *via* cleavage by the protease PilD, concatenates to build extracellular pili that are expelled or retracted by motive force generated inside the cell. The ATPases PilB and PilT function as motor proteins for expelling and retraction, respectively, and PilQ forms channels that facilitate PilA translocation through the outer membrane. PilC, PilM and PilN are involved in the assembly process at the inner membrane. Depending on the organism, several other Pil proteins may contribute to channel formation or assembly. Cyanobacteria can express several PilA homologs. In *Synechocystis* sp. PCC 6803, PilA1 is required for motility, as gene knockouts showed a nonmotile phenotype. *Synechocystis* sp. PCC 6803 expresses nine PilA homologs, which are termed minor pilins and serve different roles. Inactivation of some of these also led to non-motile cells. Knockouts of *pilC*, *pilD* and *pilT1* also had a nonmotile phenotype, whereas *pilA2* and *pilT2* knockouts were still motile. For *Nostoc punctiforme* it was found that motility is lost in mutants in which *pilB*, *pilN* or *pilQ* were defective. Type IV pili are also regarded as important machinery for DNA uptake in natural transformation and in this context, cyanobacterial genomes have recently been screened for pilin genes. Almost all out of 400 analyzed cyanobacteria were found to have a complete set of type IV pili genes. This wide distribution suggests that type IV pili could be the apparatus that drives gliding motility in all cyanobacteria, however, in filamentous species belonging to the Oscillatoriales order the role of type IV pili has as yet not been analyzed and other mechanisms for movement have been proposed. Besides gliding movement and DNA uptake, other functions have been described in cyanobacteria in which type IV pili are involved, such as flocculation or floating in water column. What is the evolutionary background of the gliding movement of cyanobacteria? Many cyanobacteria move either towards the light or away from the light in a process termed phototaxis. This effect maximizes photosynthetic light capture or protects the cyanobacterium from damage of the photosynthetic apparatus if the light is too strong, respectively. Phototaxis requires one or several photoreceptors for light sensing and a mechanism to transform the light direction into directional action. It can be difficult to distinguish between photoreceptor, the first protein in the signal transduction chain, and light sensing modulating proteins. Loss of both kinds of proteins by mutagenesis result in a modified phototaxis. A regulation of phototaxis through photosynthesis has also been postulated for filamentous cyanobacteria. In *Synechocystis* sp. PCC 6803 Cph2, PixJ, PixD and UirS have been identified as photoreceptors or as modulating proteins of phototaxis. Cph2, PixJ and UirS are cyanobacteriochromes, which are defined by one or several GAF domains with a bilin chromophore. PixJ has also a methylaccepting chemotaxis domain, like chemosensors in the chemotaxis response. UirS carries a C-terminal histidine kinase. PixD is a flavin-binding BLUF protein. This protein interacts with PixE, a response regulator-like protein, which in turn interacts with PilB and could thereby modulate phototaxis. Knockout of *cph2*, *pixD*, *uirS* or *pixE* results in an induction of phototaxis or reversion of the phototaxis direction at particular wavelengths. In *Synechocystis* sp. PCC6803, a light focusing effect results in a high light intensity on the light avoiding side of the cell. This focusing mechanism has been confirmed for phototaxis of *Synechococcus elongatus* PCC 7924. In this species, a PixJ homolog with 5 GAF domains is the proposed photoreceptor; phototaxis is lost in the *pixJ* mutant. Regular cyanobacterial phytochromes such as Cph1 of *Synechocystis* sp. PCC 6803, have not been reported to be involved in phototaxis. The difference between phytochromes and cyanobacteriochromes such as PixJ lies in their domain arrangements and their spectral properties. A regular phytochrome consists of PAS, GAF and PHY domains and a C-terminal output module such as a histidine kinase. Cyanobacteriochromes lack the PAS or the PHY domain or both, but carry other domains in high variability. In phytochromes and in cyanobacteriochromes, a bilin chromophore is bound to the GAF domain. This chromophore undergoes light-triggered changes between two stable spectral forms. In phytochromes these are red- and far-red absorbing forms, whereas in cyanobacteriochromes, different spectral forms are realized. Our group has isolated several strains of *Phormidium lacuna* from marine rockpools, characterized growth and other parameters and sequenced the genomes. *Phormidium lacuna* is a filamentous species without heterocysts that belongs to Oscillatoriales. According to 16S rRNA based phylogenetic studies, this species is located in the clade C4 according to. *Phormidium lacuna* is only distantly related to the model species *Synechocystis* sp. PCC 6803 or *Nostoc punctiforme*. Directly after isolation from the wild, it became obvious that *Phormidium lacuna* filaments are motile on agar surfaces, but a phototaxis response was not observed. *Phormidium lacuna* contains one typical phytochrome with PAS, GAF and PHY domains and 16 cyanobacteriochrome-like proteins that have one or more GAF domains with a chromophore binding cysteine. Five of these reveal close partial BLAST homology with PixJ of *Synechocystis* sp. PCC 6803. The closest homolog (WP_087710364.1) has four GAF domains and a C-terminal MCP (methyl-accepting chemotaxis protein) domain, but the others have no MCP domain. No PixD or PixE homologs were found. *Phormidium lacuna* harbors also all genes for type IV pili. Molecular studies on filamentous cyanobacteria are often hampered by the difficulty to perform genetic manipulations. During trials for gene transformation by electroporation, we found that *Phormidium lacuna* cells can take up DNA naturally and integrate it into chromosomes by homologous recombination. This was the first successful natural transformation of an Oscillatoriales member. After this discovery we aimed at disruption of pilin genes and the phytochrome gene in *Phormidium lacuna* by insertional mutagenesis. We made disruptant mutants of *pilA1*, *pilB*, *pilD*, *pilM*, *pilN*, *pilQ* and *pilT*, and the phytochrome gene *cphA*. These mutants were investigated for longitudinal movement on agar and lateral movement in liquid medium. During our studies we found experimental conditions for observing phototaxis towards vertical light, also termed photophobotaxis, so that we could compare the tactic responses of mutants and wild type. Type IV pili are clearly involved in all three kinds of movements of *Phormidium lacuna* and CphA could play a photoreceptor role or modulating role in phototaxis and surface attachment. # Methods ## Culture of *Phormidium lacuna* The strain HE10DO of *Phormidium* lacuna was used for all experiments. The genome of the strain HE10JO has been sequenced earlier. The recently sequenced genome of HE10DO differed in only ca. 1000 bases from the published HE10JO sequence. The filaments were cultivated in f/2 seawater medium in 20 ml or 50 ml culture flaks under continuous white LED light of 300 μmol m^-2 s^-1 under agitation at 25°C. ## Construction of disruption vectors and transformation The construction of the *chwA* mutant was described earlier. The gene was previously termed “7_37”. The name ChwA stands for a protein with a “Clostridial hydrophobic with conserved tryptophan (ChW)” domain. For construction of the other mutants, ca. 2000 bp of each coding sequence (plus, when necessary, upstream sequence) were amplified from genomic DNA by PCR using the primers as listed in. Each sequence was cloned into the pGEMT easy *E*. *coli* vector (Promega, Madison, WI, USA) and the plasmid linearized by PCR using primers that bind to the center of the cloned sequence. The KanR cassette was cloned into this sequence by sticky end cloning using restriction enzymes that can be recognized in the primer sequences. The positions where the sequences are interrupted 3’ of the start codon are given in. For transformation of *Phormidium lacuna*, the respective *E*. *coli* vector was purified by midi prep (Macherey Nagel, Düren, Germany). *Phormidium lacuna* was cultivated in 100 ml f/2 medium until A_{750 nm} reached 0.35 (ca. 5 d). The cell culture was centrifuged at 6000 × g for 3 min and the major part of the supernatant was discarded. A small portion of the supernatant was used to bring the volume of the pelleted cells to 800 μl. Ten μg vector DNA were mixed with 100 μl cell suspension and pipetted in the center of an f/2 agar plate (1.5% Bacto agar and 75 μg/ml Kn). This procedure was repeated to get 8 Petri dishes with DNA and *Phormidium cells*. The Petri dishes were sealed with Parafilm. After 2 weeks, filaments were typically transferred to plates with 500 μg/ml kanamycin (Kn). After another 2 weeks, single growing filaments were isolated and cultivated on fresh agar plates with 500 μg/ml Kn. Propagation continued in liquid medium with 100 μg/ml Kn. Kn concentrations and time schedules varied slightly between transformation experiments. To test for integration into the homologous site and completeness of segregation, we used inner and outer primer pairs. The inner primers were identical with the primers used for vector construction and bind at the 5’ and 3’ ends of the integrated sequence, the outer primers bind just 5’ and 3’ outside the integrated sequence, respectively (see for primers). When segregation was complete (no more wild type band visible), the mutant was used for physiological assays. Segregation of *pilA1’* was incomplete over the entire observation time of several months and during the motility experiments. Disruption of *pilD* was not possible. Details are given in the Results section. For cultivation, mutants were always grown in medium with 100 μg/ml Kn, for experiments the cells were always transferred into medium without antibiotics ( for PCR data). ## Motility experiments For all experiments on motility, *Phormidium lacuna* was transferred to Kn-free liquid f/2⁺ medium and cultivated for 5 days until A_{750 nm} reached 0.35–0.4. The PCR pattern of the integration site remained unchanged during this non-selective growth. The filaments were then treated with an Ultraturrax rotating knife (Silent Crusher, Heidolph, Schwabach, Germany) at 10000 rpm for 3 min. For motility studies on agar, filaments were usually concentrated 10-fold by centrifugation and 100 μl of the solution were dispersed on a Bacto agar f/2 Petri dish with a diameter of 5 cm. Images were recorded in 1 min intervals using a conventional microscope and a Bressler (Rhede, Germany) ocular “full HD” camera. The intensity of the microscope light was 250 μmol m^-2 s^-1. In some experiments, the Ultraturrax treated filaments were directly pipetted onto a 5 cm petri dish with agar medium, kept for 4 h in a growth chamber and photographed through a microscope. For studying movement in liquid f/2 culture, 8 ml filament suspension (OD _{750 nm} = 0.4) were transferred into a 5 cm Petri dish without agar. The movement of the filaments was observed through a Leica DM750 microscope and recorded with an EC3 camera for 30 s (no time lapse). Of each sample, recordings at 3 to 10 different characteristic positions were taken. For photophobotaxis experiments, we constructed a plastic holder in which a 5 mm wide, round LED is mounted at one end of a 15 mm long vertical tube with the light beam pointing upward. For blue, green and red light, LEDs with maximum emissions at 470 nm, 520 nm and 680 nm were chosen, respectively. The light intensity was 15 μmol m^-2 s^-1 unless indicated otherwise. Eight ml Ultraturrax treated *Phormidium* suspension (OD _{750 nm} = 0.4) were pipetted into a 5 cm Petri dish which was then placed on the LED- holder so that the LED is in the middle of the petri dish. The samples were kept for 1, 2 or 3 d in a dark room. Typically, the LED irradiation results in the formation of a circular area that is covered with circle of filaments in the center of the Petri dish. After LED irradiation, the Petri dish was photographed at its original position, a movement of the Petri dish could partially destroy the circular shape. The Petri dish was then shaken manually for 3 s and photographed again. The diameters of the circles were measured with ImageJ (NCBI). To this end, the image was loaded with ImageJ and a measuring line was drawn through the center of the petri dish. From the profile of the line, the diameter of the central circle was determined by the distance between half- maximal changes of pixel intensities. Sometimes filaments formed aggregates that were not the result of photophobotaxis. These were not considered in our evaluations. If only aggregates were observed and no evidence for photophobotaxis was found, the diameter of sample was recorded as zero. Photophobotaxis based aggregation could be clearly distinguished from other aggregations by the position within the Petri dish and by the fact that only photophobotaxis based aggregation results in surface attachment. # Results ## Generation of insertion mutants Type IV pili are involved in gliding motility of many bacteria including single celled cyanobacteria and members of the Nostocales (filamentous species with heterocysts). To see how type IV pili are involved in a member of Oscillatoriales (filamentous species without heterocysts), we aimed at disrupting genes of PilA1, PilB, PilD, PilM, PilN, PilQ and PilT in the genome of *Phormidium lacuna*. The sequences were identified as BLAST homologs of *Synechocystis* sp. PCC 6803 and other bacteria. The degree of homology between *Synechocystis* sp. PCC 6803 and *Phormidium lacuna* proteins is given in. We also disrupted the phytochrome gene of *Phormidium lacuna* in order to test for a possible role of this photoreceptor in motion. Phytochromes are abundant in most cyanobacteria, but their biological role in these organisms is unclear. *Phormidium lacuna* contains one phytochrome, which is denominated CphA here, in accordance with CphA from *Fremyella diplosiphon*. In the present study, we did not address other possible photoreceptor proteins such as Cph2 or PixJ. As a control for possible effects on the KanR cassette we used a mutant in which the gene of ChwA is disrupted. *chwA*, originally termed 7_37 (open reading frame 37 on scaffold 7), was the first gene of *Phormidium lacuna* for which natural transformation and homologous integration of a KanR cassette was successful. The protein is annotated in the NCBI database as hypothetical protein. It has 3 ChW repeats, hence the name ChwA in the present study. For gene disruptions, cloning vectors were constructed that carried a sequence of 2000 bases derived from the gene of interest and the region upstream of the start codon. The KanR cassette was placed in the center of this sequence, so that there are 1000 bp of homologous sequence on both sides of the resistance cassette. In this way, we built disruption vectors for *pilA1*, *pilB*, *pilD*, *pilM*, *pilN*, *pilQ*, *pilT* and *cphA* and obtained resistant transformants for all insertion constructs. For *pilB*, *pilM*, *pilN*, *pilQ*, *pilT* and *cphA* genes, the integration into the expected site and complete segregation into all chromosomes was confirmed by PCR using inner and outer primers that bind to chromosomal sequences inside and outside the integration vector, respectively (see for list of primers and for PCR with inner and outer primers). The PCR pattern remained during 7 days on non-selective medium, the maximum time that was used for physiological experiments. In case of *pilA1*, 7 out of 8 resistant strains showed a PCR double band indicative of partial integration, *i*.*e*. one part of chromosomes with interrupted sequences and the other part with wild type *pilA1*. The incomplete segregation pattern remained unchanged over many months in selective growth medium. Complete segregation of *pilA1* insertion is thus impossible, indicating that the complete loss of *pilA1* would be lethal. Although only limited information can be gained from such mutants, we performed also experiments with one of these strains. This strain is termed *pilA1’*. Another resistant strain had only the wild type *pilA1* band. In this strain, the KanR sequence must have integrated into another site. All these mutants show that disruption of *pilA1* and complete segregation is lethal. *pilD* knockout trials yielded only two resistant strains out of 4 transformations, although the same transformation effort was undertaken as with the other genes. During the first two weeks of selection, both mutants grew very slowly, but after another 2 weeks, the filaments recovered and grew at normal speed. In these strains, the *pilD* gene was not interrupted. We assume that the slow growth is caused by disruption of *pilD* genes in a certain proportion of the chromosomes and that by a second recombination event, the KanR cassette moved to another site. This suggests that the loss of *pilD* is lethal. The *pilD* knockout of *Synechocystis* sp. PCC 6803 is also lethal, although the cells can survive on glucose. We did not try selection on glucose medium. ## Motility on agar The motility on agar of *Phormidium lacuna* was usually investigated with filaments that were treated with an Ultraturrax rotating knife. This treatment separates adjacent filaments of the liquid preculture and cuts long filaments into shorter ones. Filaments were then brought onto agar surface and images were recorded in 1 min intervals through a microscope for several h. These microscope observations were made with white light at an intensity of 250 μmol m^-2 s^-1. After placement on the agar surface, all wild type filaments started to move instantaneously. The direction of movement was always longitudinal towards one end of the filament. The polarity of movement, indicating whether a filament moves towards one or the other tip, switched frequently in an apparently random manner. The movement of the filaments was not completely straight but in circle segments with variable diameter. We often observed filaments that described a complete circle on which they moved for several rounds. Each filament left a visible trace of extruded material on the agar surface (examples can be seen in in the areas that are not covered by filaments) which could serve as guide for other filaments or the same filament when it returned to the same position. When a filament touched another filament, it aligned its longitudinal direction so that finally long parts of both filaments were parallel and adjacent. Generally, filaments remained together, and additional filaments joined so that bundles of many filaments were formed. These filament bundles were usually one cell layer thick, sometimes a second layer of few filaments located above the lower one. More than two layers were not observed. The parallel leaf like filament arrangement of the bundles was highly flexible. All filaments moved continuously, also versus each other, bundles joined and separated, and higher ordered structures like circles were formed. Clearly, bundle formation and bundle dynamics are important functions of filament movement and communication between the filaments is required for such coordinated movements. The speed of movement of wild type and mutant filaments was estimated from data obtained during the early stage after Ultraturrax treatment, during which single filaments or few parallel filaments predominate. The distances covered during 1 min were analyzed for 50 to 150 filaments of each strain. During this short time interval, the switching of direction was rare and the measurements are thus more reliable than for longer time intervals. As can be seen in the histogram in, the speed of wild type filaments was variable between 2 and 60 μm / min. The interval between 0 and 2 μm could not be resolved under these conditions. About 30% of wild type filaments were in the range from 5 to 10 μm, the range that contained most filaments. The mutants could be divided into two groups, one group consisting of *chwA*, *pilA1’* and *cphA* which showed slightly reduced wild type like movements and the other group comprising *pilB*, *pilM*, *pilN* and *pilQ* which showed drastically reduced movements or no movements. Since none of the data follow a Gaussian distribution, we estimated the significance of differences by the non-parametric Kruskal Wallis ANOVA test. In this way, results from all mutants were compared with each other. All strains were found to be significantly different from each other (p \< 0.05) and from the wild type with the exceptions of the three pairs *chwA-cphA*, *chwA-pilA1’* and *cphA- pilA1’*, for which the differences were not significant. This is indicated by the lines between panels in. As noted above, the speeds of *chwA*, *cphA* and *pilA1’* were reduced as compared to the wild type. This effect could have been caused by the insertion of the KanR cassette in all three cases, and the possibility that KanR has a similar impact on other mutants must be considered. It is also possible that the motilities of these three mutants are affected by the knockout or by the partial knockout. Note that during these tests all filaments were in Kn free medium. The mutants *pilB*, *pilM*, *pilN*, *pilQ* and *pilT* exhibited significantly reduced motility. This shows clearly that gliding motility is based on the action of type IV pili as shown for single celled cyanobacteria, for Nostocales, and for many bacteria. The pilus motor proteins PilB and PilT, the inner membrane pore forming proteins PilM and PilN and the outer membrane core protein PilQ are all required for proper functioning of type IV pili and for gliding motility of *Phormidium lacuna*. However, only *pilT* mutants have completely lost their motility, whereas in all other mutants there was a small fraction of filaments that remained motile or at least slightly motile. In case of *pilB*, the fraction of motile filaments was ca. 10%. For *pilM*, a fraction of 4% remained motile with a distribution like the wild type. This is shown in the inset of. For *pilN* and *pilQ*, the movement was also measured during 1 h time intervals. In this way, also slow movements were detected (inset of). Such 1 h measurements were performed with *pilT* as well, but no movement was detected at all. shows Ultraturrax treated filaments that were kept for 4 h on agar. Wild type filaments aligned parallel to each other as described above, forming bundles of up to 20. Bundles were also formed by *chwA*, *pilA1’* and *cphA*, and no difference to the wild type was observed. With *pilB*, also bundles were formed that could not be distinguished from the wild type. As noted above, only a fraction of 10% of *pilB* filaments seems to move, but the long-term observations in shows that all filaments of *pilB* can move. We confirmed a “wake up” of previously immobile *pilB* filaments when time lapse recordings were re-investigated. *pilM*, *pilN*, *pilQ* and *pilT* did not form bundles, but filaments remained single or crossed over. Since in the pilin mutants there is a clear correlation between lost bundle formation and reduced motility, the type IV pili are required for bundle formation. Bundle formation requires also that filaments recognize each other and that contacts are formed. It is possible that type IV pili play important roles in these processes. ## Motility in liquid medium A largely unnoticed kind of movement that is different from the gliding movement has been described for Oscillatoriales members. This movement occurs in liquid culture in lateral direction, *i*.*e*. perpendicular to the filament axis. We found such movements when *Phormidium lacuna* filaments were observed with a microscope directly in liquid cultures. The lateral movement was more rapid than the longitudinal gliding movement described above. It could be observed directly without time lapse recordings. After Ultraturrax treatment, some filaments started immediately with lateral movements. Within few minutes, hedgehog-like bushes were formed, with ends of the filaments pointing away from the center towards the medium. Filaments are apparently connected by an extracellular matrix based on a network of polysaccharide and proteins. In our samples, a matrix could be observed by placing a needle tip into the liquid medium next to filaments. A move of the needle caused a move of the nearby filaments in the same direction, although there was no contact visible between filament and needle. Through the network comprising adjacent filaments and the intermediate extracellular matrix, force could be generated between filaments for lateral movement, without the need for surface contact as in gliding motility. For wild type and each mutant, we made recordings of five or more different views and for each view about 5 recordings in series. Each recording was 30 s long. We tried to quantify movement speed or other parameters by different analysis software, but could not obtain reliable results. However, clear differences between strains became evident through qualitative comparisons. Here, we provide selected videos (– Movies) and merged images for each strain. The merged images are composed of one picture that is displayed in red and a second one taken after 10 s that is displayed in green. Moving filaments are thus displayed in red or green. Lateral movements of wild type filaments in liquid culture can be seen in and. All filaments made lateral movements, although at a certain time point only a subfraction moved, as seen also in. The *chwA* strain was indistinguishable from the wild type ( and Movie). Also, *phrA1’* and *cphA* displayed movements similar to the wild type (and Movies). The results of the *pilB* mutant differed qualitatively from one sample to another. In 3 of the 6 sample series, only few filaments moved with slow speed (as). In the other 3 sample series, almost all filaments moved (as in and Movie), although still less as compared to wild type. The reason for this difference among the *pilB* mutant samples is not known. As in the gliding motility experiments, PilB is required for efficient movement, but dispensable for movement in general. With *pilM* mutants, we also observed that few filaments moved, but most filaments remained silent. This result is again equivalent to the gliding motility study where very few filaments showed normal movement. For *pilN*, *pilQ* and *pilT* mutants, no lateral movements were observed (– Movies). The *pil* mutants show clearly that type IV pili are required for lateral movement in liquid. The mechanism for lateral motility must be based on connections between filaments *via* an extracellular network. ## Photophobotaxis In the phototactic response, motile organisms move either towards the light or away from it. This response can most often be induced by unilateral light, *i*.*e*. by light that hits the organism within the area of possible moving directions. We performed many trials with *Phormidium lacuna* to induce phototactic movement by unilateral light (as outlined) but the filaments remained equally distributed in the Petri dish (as). Neither the use of agar plates nor the use of different irradiation wavelengths nor the variation of intensities resulted in a directional response. However, in time lapse studies with *Phormidium filaments* we observed that filaments sometimes gather in the white microscope light that hit the sample from below. An example for this effect is shown in. We therefore suggested that light projected vertically to the movement area can induce light-directed movement of *Phormidium lacuna*. In our subsequent studies, *Phormidium lacuna* filaments were irradiated with light emitting diodes (LEDs) from below. In standard assays, 5 cm Petri dishes (without agar) that were filled with 8 ml Ultraturrax-treated filaments in f/2 medium were placed on a holder in which a light emitting diode shines light through a short tunnel on the specimen from below, at the position of the center of the Petri dish. We observed that red light, blue light and green light of around 5–50 μmol m^-2 s^-1 induce a clear gathering of wild type filaments towards the light in the center. Gathering could be observed after several hours. During 2 d the effect became stronger. Thus, the movement of *Phormidium lacuna* in light follows rather a light-intensity gradient than the light direction and behaves therefore different from single celled cyanobacteria. Directional effects by vertical light have been described for other organisms such as the alga *Euglena gracilis* or *Phormidium uncinatum*. Responses towards unilateral light and responses towards light gradients induced by vertical light have been termed phototaxis. Responses towards vertical light have also been termed photophobotaxis, because the organisms avoid darkness. In order to emphasize the difference between the responses of *Phormidium lacuna* and of single-celled cyanobacteria, we use the term photophobotaxis here. It should be stressed that this movement takes place in Petri dishes with or without agar and that the majority of photophobotaxis experiments were performed without agar. Comparative analyses of wild type and mutants were performed with red light at a light intensity of 15 μmol m^-2 s^-1 for 2 d. Thereafter, a photograph was recorded from the Petri dish. We routinely made a second photograph after a gentle shaking of the Petri dish. The diameter of the covered circle was measured between the two positions of half maximal pixel intensity along a line through the center of the circle, this value was used to calculate the circle area. The area value stands as a measure for the magnitude of the photophobotactic response. Shaking reduced the size of the densely covered area. In this way, the data distinguish between all filaments that have moved to the center and those filaments that were more tightly bound to the surface. Shaking did not increase the error values to a relevant extent, which supports the reliability of the approach. The circle area of the light beam is 20 mm² and in a clear response the entire area is completely covered with filaments. We experienced that a response could result in a covered area even larger than 20 mm², and any smaller covered areas were also often observed. In samples where no accumulation in the center was obvious, as in the dark control or *pilT*^*-* examples shown in, we set the area value to 0. The data processing showed a broad distribution of covered areas for wild type, *cphA*, *pilA1’*, *chwA*, *pilB*. The ranges were from no response to intermediate and strong responses with covered areas above 25 mm². The mean values are presented in. We estimated the significances of differences between all possible pairs of samples according to Kruskal and Wallis. Significance of sample pairs before shaking is presented by colored squares in the lower left triangle of, the pairs of samples after shaking are presented by the colored squares in the upper right triangle of and the significance of each sample before and after shaking are given by the colored squares in the diagonal in of. The wild type covered an average area of 25 mm². Shaking reduced the wild type area to a mean value of 20 mm², i.e. the irradiated area. With ca. 18 mm², the mean covered area of *chwA* was reduced as compared to the wild type; shaking reduced the area to 12 mm². We assume again that the reduced response of *chwA* vs. wild type results from the KanR cassette. In case of *pilA1‘*, the covered area was 25 mm², the same as in the wild type. The mutation does not affect the magnitude of photophobotaxis. If the KanR effect would be the same as in the *chwA* mutant, the partial loss of *pilA1’* would compensate for this effect. The response of *cphA* was further reduced compared to wild type, *chwA* and *pilA1’*. Before and after shaking the covered areas were 12 mm² and 2 mm², respectively. These low values of *cphA* imply that in the wild type, CphA modulates photophobotaxis or acts as a photoreceptor for this reponse. The *pilB* mutant had a covered area of 12 mm² and an area of 5 mm² after shaking. The pattern is comparable with *cphA*. The cause for reduction could be the reduced motility of *pilB*, but also a reduced adhesion, which could result from partially defective type IV pili. The other pilin mutants *pilN*, *pilM*, *pilQ and pilT* had no detectable phototactic response. Only in exceptional cases, a response was obtained. That for a photophobotactic response the integrity of type IV pili is required is no surprise, as photophobotaxis without motion is not possible. # Discussion We describe here for the first time mutant effects on the motility of a member of the Oscillatoriales. *Phormidium lacuna* is an ideal model cyanobacterium for such studies, because natural transformation is established and because the filaments are continuously in motion. The motility of the genera *Oscillatoria* or *Phormidium* have been scientifically investigated since the 19^th century. Cellular and molecular studies on motility of cyanobacteria have, however, concentrated on the model organism *Synechocystis* sp. PCC 6803, a unicellular species. Few experiments followed with *Nostoc punctiforme* (Nostocales) and *Synechococcus elongatus* (single celled). There are three major differences between motilities of *Phormidium lacuna* and single celled cyanobacteria. (i) The gliding motility of *Phormidium lacuna* is characterized by continued change of direction, combined with bending movements, and thus essentially different from the straight movement of single celled cyanobacteria. (ii) *Phormidium lacuna* makes lateral motions in liquid medium (– Movies), such motions are not known for single celled species. (iii) Finally, the photophobotactic movement of *Phormidium lacuna* is induced by vertical light, in contrast to the movement towards unilateral light in phototaxis of single celled species. The mechanisms of light perception and detection of light direction must be completely different between these species. Although type IV pili are widely distributed in cyanobacteria and present in *Phormidium lacuna*, it was initially not clear whether these pili are involved in motions of *Phormidium lacuna*. Lateral motion, independent of surface contacts, is rarely described in cyanobacteria. The direction of lateral movement is opposed to the longitudinal direction in gliding motility. Because *pilM*, *pilN*, *pilQ* and *pilT* were neither motile in gliding motility nor in lateral movements, both kinds of movement must be driven by type IV pili. For lateral motility, this implies that the outer ends of the pili are connected with the extracellular matrix or with pili of other cells, in order to generate force for the motions. Alternatively, another motile structure which is dependent on intact type IV pili, could perform lateral movement. What evolutionary advantage can be found behind lateral motions in liquid? The back and forth movements do not allow the filaments to reach another place. However, filaments join neighboring filaments and form aggregates, and this kind of biofilm formation could protect against predators. The counterpart of these aggregates are the aligned filaments on agar that form leaflike structures. Also in this case, the getting together could result in protection against predators. The results from all three assays, gliding motility, lateral motions in liquid and photophobotaxis indicate that PilM, PilN, PilQ and PilT are essential for the function of the type IV pili—without these any kind of motion is inhibited. PilN and PilM are pore forming components in the inner membrane and PilQ is involved in pore forming at the outer membrane. PilT is an ATPase for retraction of the pilus. Clearly, each of the tested components of the Type IV pilus apparatus and its assembly components are required for motility. We would like to stress here that it is very unlikely that the immobile phenotypes result from second site mutations. We have performed many *Phormidium* transformations in which other sites were targeted. The immobile phenotype was never observed. In *pilB* mutants motility is only partially blocked in the three motion assays. PilB is the ATPase for pilus expelling. The PCR results indicate complete segregation of the interrupted gene. As the other mutants, this strain was kept continuously on Kn medium until it was used for experiments. There is thus no reason to assume that wild type *pilB* copies are still abundant in the cells. In recent RAST annotations we found four proteins with weak homology to the PilB ATPase domain in the genome of *Phormidium lacuna*. Within ca. 300 amino acids of the ATPase region, PilB and the homologous proteins revealed ca. 20% identity. These proteins could replace PilB in the *pilB* mutant, although the N-terminal T2S domain of PilB is lacking in these homologous proteins. Alternatively, another yet unsequenced *pilB* could replace the regular one, but this option is unlikely. We have sequenced draft genomes of altogether 5 strains, and more than 90% of each genome are completely sequenced. A second *pilB* was found in none of the strains. For *pilA1’* we obtained only partially segregated disruption strains, which were almost as motile as the wild type. In case of *pilD* it was even not possible to select strains in which the disrupted gene is partially segregated. The loss of PilD in *Synechocystis* sp. PCC 6803 is lethal because PilA1 prepilin proteins accumulate in the membranes thereby inhibiting synthesis of photosystem II proteins. The same effect could explain why *pilD* of *Phormidium lacuna* is lethal. A *pilA1* mutant would, however, not cause such effects. It is difficult to understand why *Phormidium lacuna* cells cannot survive without PilA1. *Phormidium lacuna* has 2 homologs of PilA2 (WP_140409148, 30% identities of overlapping region, WP_087711780.1, 31%), but no other PilA1 homolog. For *Phormidium lacuna* we could imagine that misassembly of the pilus due to lack of PilA1 results in accumulation of PilA2 proteins in the plasma membrane and in cell death of completely segregated *pilA1* mutants. The *cphA* mutant shows motility on agar comparable with the wild type, *chwA* and *pila1*. The photophobotaxis response of *cphA* was, however, clearly diminished as compared to the other three strains. CphA operates either as a photophobotaxis photoreceptor in *Phormidium lacuna* or as a modulating photoreceptor that acts on the photophobotaxis signal transduction pathway. Typical phytochromes are red light sensors and our relevant phototaxis experiments have been performed in red light. The weak phototaxis response in the *cphA* mutant must be regulated by a different photoreceptor. Phototaxis of *Phormidium lacuna* is observed in blue, green and red light. Therefore, it is very likely that more than one single photoreceptor modulates phototaxis. The role of CphA in photophobotaxis is another major difference between *Synechocystis* sp. PCC 6803 and *Phormidium lacuna*, since knockout mutants of the homologous Cph1 in *Synechocystis* were not affected in phototaxis. CphA could also act indirectly on photophobotaxis, for example if photophobotaxis is mediated through photosynthesis, as it has been hypothesized for filamentous cyanobacteria. Further mutant experiments are required to clarify the photophobotaxis photoreceptor situation in *Phormidium lacuna*. What is the mechanism behind *Phormidium lacuna* photophobotaxis? Random movement in the dark and stop of movement by surface attachment in the light could explain a gathering in the light zone. Surface attachment would be part of this mechanism. However, the time lapse studies (as) do not support such “move in the dark or attach in the light” response, because movement of wild type filaments and filaments of motile mutants is observed in the microscope light. We therefore also believe that the sticking to the surface is not part of the phototactic response. A simple model for photophobotaxis of *Phormidium lacuna* could be as follows: Filaments move in darkness and in light in random directions, but the moving direction switches more often when they leave the light. In this way, filaments would finally accumulate in the light. For a detailed investigation of the mechanism of light sensing, experiments on an infrared microscope with which the movements along light—dark transition can be observed are desired. We also got the impression that stray light induces a gathering response towards the high light intensities (see e.g. bundle). The lack of phototaxis in unilateral light is at contrast with such an observation. Although details on the mechanism need to be investigated, it is clear that the photophobotaxis of *Phormidium lacuna* differs in many aspects from the phototaxis of *Synechocystis* sp. PCC6803 or other single-celled cyanobacteria. # Supporting information [^1]: The authors have declared that no conflict of interest exists.

# Introduction Periods of restricted food intake induce a loss of body mass that is often followed by rapid regaining of the lost weight when the restriction ends, during which physiological regulations associated with either energy intake or expenditure, or the both are reported to be involved. However, the results related to energy budget and behaviors in response to food restriction and refeeding remain controversial. For example, the energy spent for the rate of resting metabolism (RMR) and activity behavior decreased in food-restricted laboratory mice and rats. In contrary, Siberian hamsters (*Phodopus sungorus*) and other hamster species increased activity associated foraging and food hoarding behaviors in response to food shortage. During refeeding, laboratory rats regained body mass and fat mass, showing a “compensatory growth”,. Some wild rodents also showed “compensatory growth”, but to much less extent compared with that observed in laboratory animals. This paradox may reflect different energy strategy and behavioral patterns in wild animals from that in laboratory rodents. Leptin, the product of the *ob* gene, is mainly expressed in adipose tissue and plays important roles in the regulation of both energy intake and expenditure. It was reported that serum leptin level reduced during food restriction and increased during refeeding. Leptin administration to food-restricted laboratory rats reduced food intake and prevented the regain of body mass. In addition, exogenous leptin inhibited food-deprivation-induced increases in food intake and food hoarding in Siberian hamsters. These results make leptin to be a possible candidate involved in the regulations of energy budget and behavior in response to food restriction and refeeding in both laboratory and wild animals. The striped hamster (*Cricetulus barabensis*) is a major rodent in northern China and is also distributed in Russia, Mongolia, and Korea. The hamsters feed on stems and leaves of plant during summer and on foraging crop seeds in winter,. Thus the species must experience great seasonal fluctuations in food quality and availability. Whereas, unlike other wild rodents, such as Djungarian hamsters (*Phodopus sungorus*), Brandt's voles (*Lasiopodomys brandtii*) and Mongolian gerbils (*Meriones unguiculatus*), striped hamsters do not show significant changes in body masses after being maintained in an outside enclosure over a year (Zhao ZJ, unpublished data). We previously found a significant decrease in body mass in stochastic food-restricted hamsters, followed by a slower regaining of body mass during refeeding than that in Swiss mice. It suggests that striped hamster, showing different patterns of body mass regulation from both lab mice and other wild rodents, may become a potential model that is suitable for studying the resistance to over-weight when food restriction ends. In the present study, energy budget and activity behavior were measured in striped hamsters subjected to a successive food restriction for four weeks and refeeding for another four weeks. The effect of leptin supplement on energy budget and activity behavior was examined during both food restriction and refeeding. We hypothesized that regulations of energy budget and activity behavior would be employed to cope with the changes in food availability, but failing to regain the lost weight when the restriction ended. Leptin might be involved in changes in energy budget and activity, and consequently played a key role in the resistance to over-weight in striped hamsters experiencing food restriction and refeeding. # Materials and Methods ## Ethics Statement This study was in compliance with the Animal Care and Use Committee of Liaocheng University. The experiment procedure and protocol were approved by the Committee (Permit Number: 11-0219-011). ## Animals and experiment protocol Striped hamsters were obtained from a laboratory-breeding colony started with animals that were initially trapped from farmland at the center of Hebei province (115°13′E, 38°12′S), North China Plain. Environmental temperature was kept constant at 21±1°C with a 12 h∶ 12 h light∶ dark cycle (lights on at 0800 h). Food (standard rodent chow; produced by Beijing KeAo Feed Co.) and water were provided *ad libitum*. The macronutrient composition of the diet was 6.2% crude fat, 20.8% crude protein, 23.1% neutral detergent fiber, 12.5% acid detergent fiber, and 10.0% ash, and the caloric value is 17.5 kJ/g. Adult male hamsters, 4–5 months old, were singly housed in plastic cages (29×18×16 cm) with fresh saw dust bedding for two weeks before the experiments. ### Experiment 1: Effects of food restriction (FR) and refeeding (Re) on body mass and food intake Twenty four male hamsters were assigned randomly into either control group (Con, *n* = 12) that animals were fed *ad libitum* for 8 weeks, or FR and Re group (FR-Re, *n* = 12) in which each hamster was restricted to 85% of initial food intake for 28 days and refed *ad libitum* for another 28-days. Body mass was measured every three days and food intake was determined on a daily basis. Before animals were restricted, food intake was calculated as the mass of food missing from the hopper every day, subtracting orts mixed in the bedding. Prior to the initiation of food restriction, initial food intake for each animal was calculated as the average of daily food intake over 7 days. Each hamster in FR- Re group was provided with 85% of initial food intake only during FR period, making food-restricted hamster had a 15% reduction of caloric intake. Food was given the same time each day at 1900 h following body mass measurements. ### Experiment 2: Effects of FR and Re on behavior, energy budget, body composition Fifty six hamsters were assigned randomly into one of the following 7 groups (*n* = 8 in each group): controls that were fed *ad libitum* for 8 weeks; FR- d 1, FR- d 7 and FR- d 28 groups, animals were restricted to 85% of initial food intake for 1, 7 and 28 days, respectively; and Re-d 1, Re-d 7 and Re-d 28 groups, during which animals were restricted to 85% of initial food intake for 4 weeks and were then refed *ad libitum* for 1, 7 and 28 days, respectively. At the end of the experiment, behavior observation was made, and RMR and energy budget were measured. ## Behavior observation Behavior observations were made in 4 hamsters from each group over a day (24 h). Observations were performed using computer-connected infrared monitors (SONY, 420 TV line) and were automatically stored in computer, which were then subjected to operator analysis. General activity included any active movement such as walking around the cage and climbing on the cage bars. The time spent on activity was recorded and expressed as min/h and min/24 h, respectively. ## RMR RMR was quantified as the rate of oxygen consumption, using a computerized open- flow respirometry system (Sable system, USA). Air was pumped at a rate of 750–850 ml/min through a cylindrical sealed Perspex chamber at 29±0.5°C (within the thermal neutral zone of this species,. Gases leaving the chamber were dried (silica gel) and sampled using an oxygen analyzer at a flow rate of 150–175 ml/min. The data were averaged and collected every 10 s by a computer connected analogue-to-digital converter (STD-UI2, Sable system), and analyzed using a standard software (Sable system). RMR was measured for 2.5 hours between 11: 00 and 17: 00, and calculated from the lowest rate of oxygen consumption over 5 min, using the equation: VO₂ = Flow rate×(FiO₂−FeO₂)/(1−FiO₂×(1−RQ)), where FiO₂ is input fractional concentration of O₂ to the chamber; FeO₂ is excurrent fractional concentration of O₂ from the chamber; and RQ is respiratory quotient. Here, RQ was assumed to be 0.85. RMR was then corrected to the standard temperature and air pressure (STP) conditions. ## Energy budget Food was provided quantitatively, and the spillage of food mixed with bedding and feces were collected from each cage over the last 2 days in control, FR- d 7, FR- d 28, as well as Re-d 7 and Re-d 28 groups, but over one day in FR- d 1 and Re-d 1 groups. The spillage of food and feces were sorted and separated manually after they were dried at 60°C to constant mass. Gross energy contents of the diet and feces were determined using a Parr 1281 oxygen bomb calorimeter (Parr Instrument, Moline, IL, USA). Gross energy intake (GEI), digestive energy intake (DEI), and apparent energy assimilation efficiency (digestibility) were calculated as follows –: GEI (kJ·d⁻¹) = food intake (g·d⁻¹)×dry matter content of the diet (%)×energy content of food (kJ·g⁻¹); DEI (kJ·d⁻¹) = GEI−(dry mass of feces (g·d⁻¹)×energy content of feces (kJ·g⁻¹)); Digestibility (%) = DEI/GEI×100%. ## Serum leptin levels Animals were euthanized by decapitation between 0900 and 1100 h on the day next to RMR measurements. Trunk blood was collected for serum leptin measurements. Serum leptin level was quantified by radio-immunoassay (RIA) using the Linco ¹²⁵I Multi-species Kit (Cat. No. XL-85K, Linco Research Inc.), following the standard kit instructions. The lower and upper limits of the assay kit were 1 and 50 ng/ml, and the inter- and intra-assay variations were \<3.6% and 8.7%, respectively. ## Body composition After trunk blood was collected, the gastrointestinal tracts were separated, and liver, heart, lung, spleen pancreas and kidneys were also removed. The remaining carcass (including the brain, but excluding the thyroid and urinary bladder) was weighed (to 0.001 g) to determine wet mass, dried in an oven at 60°C for 10 days to a constant mass, and then weighed (to 0.001 g) again to determine dry mass. Total body fat was extracted from the dried carcass by ether extraction in a Soxhlet apparatus. ### Experiment 3: Effect of leptin supplement on food intake and behavior during FR and Re Sixteen hamsters were randomly assigned into one of the four groups: Ad-PBS, hamsters that were fed *ad libitum* and treated with PBS; Ad-leptin, Ad hamsters that were treated with leptin; FR-PBS, FR hamsters that were treated with PBS; FR-leptin, FR hamsters that were treated with leptin. Animals were fed *ad libitum* for 14 days in Ad groups. FR hamsters were restricted to 85% of initial food intake for 10 days, and then refed *ad libitum* for 4 days. On day 8, hamsters were anesthetized with isoflurane and implanted subcutaneously on the dorsal side with a miniosmotic pump (Alzet model 1007D; capacity, 100 µl; release rate, 0.5 µl/h; duration, 7 days; Durect, Cupertino, CA) containing either recombinant murine leptin (100 µg dissolved in 100 µl phosphate-buffered saline \[PBS\], purchased from Peprotech, USA) or PBS. Body mass and food intake were measured daily according the method mentioned in experiment 1. Activity observation was performed as described in experiment 2 and the time spent on activity was recorded and expressed as min/24 h. Animals were euthanized by decapitation and trunk blood was collected for serum leptin measurements as the same methods mentioned in “Serum leptin levels”. ## Statistics Data were expressed as the means ± SE and analyzed using SPSS 13.0 statistic software. Experiment 1, changes in body mass and food intake throughout FR and Re period were analyzed using repeated one-way ANOVA measurements, and differences between the two groups on any day points were examined using independent *t*-tests. Experiment 2, differences in activity behavior, RMR, energy budget, serum leptin levels and body composition between the seven groups were examined using one-way ANOVA or ANCOVA with body mass or carcass mass as a covariate, followed by Tukey's HSD post-hoc tests where appropriate. Experiment 3, body mass change throughout the experiment was examined using repeated measurements. Differences in body mass change, food intake and activity on any day points as well as serum leptin levels were examined using two-way ANOVA (FR×leptin), followed by Tukey's HSD post-hoc tests where required. Correlations of leptin with fat content and gross energy intake were examined using a Pearson correlation analyses. The level of significance was set at *P*\<0.05. # Results ## Effects of food restriction (FR) and refeeding (Re) on body mass and food intake ### Food intake Food intake was not different between Con and FR-Re groups prior to the experiment (d 0, *t*₂₁ = 0.16, *P*\>0.05). There were no changes in food intake throughout the experiment in Con group (d 1–56, *F*_55,605 = 0.62, *P*\>0.05), while significant changes were observed in FR-Re group (d 1–56, *F*_55,550 = 13.81, *P*\<0.01). During restriction, FR-Re animals were provided with 85% of initial food intake only, which was lower than that of control animals (d 1, *t*₂₁ = 2.20, *P*\<0.05, d 28, *t*₂₁ = 2.22, *P*\<0.05). During refeeding, FR-Re animals consumed more food than control animals (d 29, Con, 4.0±0.2 g/d, FR-Re, 5.3±0.5 g/d, *t*₂₁ = 2.71, *P*\<0.05), whereas food intake was not statistically different between the two groups on day 30 and thereafter (d 30, *t*₂₁ = 1.53, *P*\>0.05, d 56, *t*₂₁ = 1.10, *P*\>0.05). ### Body mass There was no difference in body mass between Con and FR-Re groups before the experiment started (d 0, *t*₂₁ = 0.21, *P*\>0.05). Control hamsters increased their weight from 33.1±0.9 g on day 0 to 34.4±1.2 g on day 56 (days 0–56, *F*_20,220 = 6.75, *P*\<0.05). Body mass significantly decreased in FR-Re animals during restriction, which decreased by 16% on day 18 compared with on day 0 (days 1–18, *F*_6,60 = 41.52, *P*\<0.01), and then lowered to a minimum of around 27 g between days 21 and 28. On the first few days of refeeding, body mass shortly increased in FR-Re groups (days 34–56, *F*_8,80 = 7.40, *P*\<0.01). FR-Re animals showed lower body mass than control animals on day 6 till day 34 (d 6, *t*₂₁ = 2.39, *P*\<0.05, d 34, *t*₂₁ = 2.05, *P*\<0.05). Body mass was not statistically different between the two groups on day 37 and thereafter (d 37, *t*₂₁ = 1.63, *P*\>0.05, d 56, *t*₂₁ = 0.80, *P*\>0.05). ## Effects of FR and Re on behavior, energy budget, body composition ### Activity Activity behavior usually occurred during the dark phase in control hamsters, while during the day phase they spent almost all the time on the rest. During food restriction, FR-Re hamsters spent significantly more time on activity both during the dark and the light phase than controls. During refeeding, FR-Re hamsters still showed high activity behavior on day 1 (Re d 1), whereas they decreased the time spent on activity on day 7 (Re d 7) and thereafter. Activity behavior was affected by FR-Re (*F*_6,27 = 6.27, *P*\<0.01), by which FR-Re animals spent more time on activity during food restriction than controls (post Hoc, *P*\<0.05). On day Re 28, the time spent activity was significantly less in FR-Re group than controls (Re d 28, post Hoc, *P*\<0.05). ### RMR FR-Re had a significant effect on RMR when expressed either per mouse (mlO₂/h, *F*_6,48 = 2.53, *P*\<0.05) or per gram body mass (mlO₂/g • h, *F*_6,49 = 3.28, *P*\<0.01,). RMR in FR-d 7 group was higher by 20% and 36% than controls when expressed per mouse and per weight, respectively (post hoc, *P*\<0.05), while it was not statistically different between FR-d 28 group and controls (post hoc, *P*\>0.05). During refeeding, RMR was significantly lower in Re-d 7 group than FR-d 7 group (post Hoc, *P*\<0.05), whereas the differences between Re-d 1, Re-d 7, Re-d 28 groups and controls were not statistically different (post Hoc, *P*\>0.05). ### Energy budget GEI was significantly affected by FR-Re (*F*_6,49 = 8.95, *P*\<0.01), FR-Re hamsters had lower GEI during restriction than controls (post Hoc, *P*\<0.05). GEI was significantly higher in Re-d 1 group than control and FR-d 1, d 7 and d 28 groups (post Hoc, *P*\<0.05), while it was not different between Re-d 7, Re-d 28 and control groups (post Hoc, *P*\>0.05). DEI was similar to the changes observed in GEI, by which DEI was lower in FR-d 1, d 7 and d 28 groups, and higher in Re-d 1 group (*F*_6,49 = 8.05, *P*\<0.01, post Hoc, *P*\<0.05). Digestibility was not affected by FR-Re, and no difference was observed between the 7 groups (*F*_6,49 = 1.18, *P*\>0.05, post Hoc, *P*\>0.05). ### Carcass mass and fat content Wet and dry masses of carcass were significantly affected by FR-Re, which were lower in FR-d 28 group than that in Con group (post Hoc, *P*\<0.05). Fat mass and fat content were also affected by FR-Re. Fat mass and fat content were significantly lower in FR-d 28 groups than controls (post Hoc, *P*\<0.05), while the difference between Con, Re-d 7 and Re-d 28 groups was not significant (post Hoc, P\>0.05). ### Serum leptin Serum leptin level was significantly affected by FR-Re, which was significantly lower in FR-d 1, d 7 and d 28 groups than controls. Serum Leptin was still lower in Re-d 1 group compared with controls, but it increased significantly in Re-d 7 and Re-d 28 groups, which were similar to that observed in control group. There was a positive correlation between serum leptin and fat content in controls, this correlation was also observed in other six groups. No correlation was observed between serum leptin and GEI in control hamsters. Serum leptin was positively correlated with GEI in FR-d 28 group, but no correlations were found in FR-d 1 and FR-d 7 groups. Hamsters in Re-d 7 and Re-d 28 groups showed significantly negative correlations between serum leptin and GEI in. ## Effect of leptin supplement on food intake and behavior during FR and Re ### Body mass change Body mass was not different between the four groups prior to the experiment (d 0, FR, *F*_1,12 = 0.84, *P*\>0.05; leptin, *F*_1,12 = 0.01, *P*\>0.05). Food restriction had a significant effect on body mass change on day 1 till day 10, and restricted hamsters showed lower body mass than Ad animals (d 1, *F*_1,12 = 12.25, *P*\<0.01, d 10, *F*_1,12 = 34.87, *P*\<0.01). Leptin supplement had no effect on body mass change during food restriction (d 8, *F*_1,12 = 0.03, *P*\>0.05, d 10, *F*_1,12 = 0.01, *P*\>0.05), while had a significant impact on body mass change during refeeding (d 12, *F*_1,12 = 5.62, *P*\<0.01, d 14, *F*_1,12 = 20.84, *P*\<0.01). During refeeding phase, body mass increased from −13.6±2.4% on day 10 to −2.5±0.7% on day 14 in FR-PBS group (d 10–14, *F*_4,12 = 8.43, *P*\<0.01), while it did not change in FR- leptin group between these days (d 10, −12.1±2.7%, d 14, −11.2±2.3%, d 10–14, post hoc, *P*\>0.05,). ### Effect of leptin administration on food intake Food intake did not differ between the four groups prior to the initiation of food restriction (d 0, FR, *F*_1,12 = 0.82, *P*\>0.05; leptin, *F*_1,12 = 0.02, *P*\>0.05). During food restriction, food-restricted hamsters consumed 15% less food than *ad libitum* animals (d 1, *F*_1,12 = 3.52, *P* = 0.09). 0n day 8 till 10, leptin supplement did not affect food intake in either *ad libitum* or food-restricted hamsters (d 8, *F*_1,12 = 0.18, *P*\>0.05; d 10, *F*_1,12 = 1.48, *P*\>0.05). During refeeding phase, food intake was higher in FR-PBS hamsters than Ad-PBS hamsters (FR, d 11, *F*_1,12 = 146.12, *P*\<0.01, d 14, *F*_1,12 = 18.78, *P*\<0.01). Leptin supplement had a significant effect on food intake on day 11 till 13 (d11, *F*_1,12 = 4.91, *P*\<0.05), by which food intake increased by 79.1%, 55.8% and 21.3% on day 11, 12 and 13 in FR-PBS group relative to Ad-PBS group, respectively, but elevated by only 52.7%, 18.6% and 5.6% in FR-leptin group (post hoc, *P*\<0.05). However, the significant effect of leptin supplement on food intake disappeared on day 14 (*F*_1,12 = 0.72, *P*\>0.05). No difference in food intake was observed between Ad-PBS and Ad-leptin groups (post hoc, *P*\>0.05). ### Effect of leptin administration on activity There was no group difference in time spent on activity on day 0 (FR, *F*_1,12 = 0.11, *P*\>0.05; leptin, *F*_1,12 = 0.08, *P*\>0.05). The time spent on activity was significantly affected by food restriction on day 4 till 10, and restricted animals spent more time on activity than Ad animals (d 4, *F*_1,12 = 11.65, *P* = 0.01, post hoc, *P*\<0.05; d 10, *F*_1,12 = 23.24, *P*\<0.01, post hoc, *P*\<0.05). During refeeding, effect of restriction on activity was not significant on day 12 (*F*_1,12 = 0.74, *P*\>0.05) and day 14 (*F*_1,12 = 0.94, *P*\>0.05). Leptin supplement resulted in a significant reduction in activity, and hamsters spent 71% and 91% less time on activity in FR-leptin group on day 10 and day 14, respectively, than in FR-PBS group (d 10, *F*_1,12 = 24.50, *P*\<0.01, post hoc, *P*\<0.05; d 14, *F*_1,12 = 16.96, *P*\<0.01, post hoc, *P*\<0.05). The time spent on activity decreased by 36% and 45% in Ad-leptin than Ad-PBS groups on day 10 and 14, respectively, while the difference was not statistically different (d 10, post hoc, *P*\>0.05, d 14, post hoc, *P*\>0.05). ### Effect of leptin administration on serum leptin Serum leptin levels averaged 2.37±0.34 and 3.97±0.49 ng/ml in Ad-PBS and Ad- leptin groups, and 1.95±0.17 and 4.20±0.60 ng/ml in FR-PBS and FR-leptin groups, respectively. No effect of food restriction and refeeding on serum leptin was observed on the day following a 4-day's refeeding (*F*_1,12 = 0.05, *P*\>0.05). Leptin supplement resulted in significant increases in serum leptin for both hamsters fed *ad libitum* and hamsters under food restriction and refeeding (*F*_1,12 = 19.90, *P*\<0.001). # Discussion The change in food availability has been found to affect body mass in small mammals,. In the present study, we observed significant reductions in body mass, carcass mass, and body fat content in striped hamsters restricted to 85% of initial food intake. Weight losses were also observed in food-restricted C57/B6 mice, Swiss mice, golden spine mice (*Acomys russatus, Muridae*), and Mongolian gerbils. Inconsistently, body mass did not decrease in MF1 mice restricted to 80% of *ad libitum* food intake, and rats restricted to 75% of initial food intake. The inconsistency may partly due to the different extent of restriction between the different studies above, since animals under severe food restriction often lose more weight than animals at softer restriction. Here, striped hamster lost weight more rapidly and significantly after restricted to 85% of initial food intake than either laboratory mice or rats, or other field rodents. This may suggest that striped hamsters, showing seasonal foraging behavior, are more sensitive to food shortage than the animals mentioned above. After being refed *ad libitum*, striped hamsters showed rapidly regaining of lost weight, showing “compensatory growth”, whereas the regaining was less and not followed by overweight compared with controls. Laboratory rats subjected to FR-Re, however, showed not only “compensatory growth” but also fatter than *ad libitum* controls. The inconsistent results may be due to the species-specific energy budget strategy in response to the change of food availability. In the present study striped hamsters consumed less food during food restriction than controls. When given free access to unlimited diet, they increased food intake by 33% compared with their counterpart controls (*P*\<0.05). However, this increase was observed only on the first one to three days during refeeding, and then returned to the levels of controls. Inconsistently, when restricted rats were allowed *ad libitum* access to food, the food intake increased to twice control levels for 6 days before returning to control levels. One reason for these disparate results may be the length and severity of restriction before refeeding, and animals at a few weeks of severe food restriction will increase food more intake when allowed to eat *ad libitum*. Another reason may be that food intake during refeeding is proportional to the amount of depletion in energy stores caused by food restriction,. Here we allowed striped hamsters to restrict to 85% of initial food intake, but fat mass decreased by 56%, indicating that the two explains above might not be the case. It may reflect a special energy strategy in response to food restriction and refeeding in striped hamsters. In the present study, digestibility did not change in striped hamsters during food restriction and refeeding, indicating that restricted hamsters were not able to enhance their digestive efficiency to extract more energy from digested diet. This suggests that adaptive regulation of energy expenditure is more important than energy intake in the trade-off of the energy strategy in food- restricted animals. The maintenance requirements include the energy exported for RMR and activity. Some food-restricted animals, like MF1 mice, deer mice (*Peromyscus*) and chipmunks (*Eutamias minimus*) are reported to decrease RMR and activity to completely compensate for the restricted energy intake, and consequently to prevent weight loss. This is largely different from the results from striped hamsters. Here, we found significant increases in RMR and the time spent on activity in food-restricted hamsters, which was consistent with Syrian hamsters (*Mesocricetus auratus*) and house mice (*Mus musculus*). This may reflect a different strategy associated with activity for coping with food restriction between different rodent species. An increase in activity in food- restricted animals may indicate an increased effort in foraging, food hoarding or migratory behavior,. Further, an increase in time spent on activity was attenuated in restricted hamsters on day 28, and increased RMR was observed on day 7 but not on day 28, suggesting time-dependent responses to food restriction. It has been well established that leptin plays a crucial role in the regulations of energy balance. Here, we found significant reductions in serum leptin level in food-restricted hamsters, which was in parallel with the marked decreases in body fat, consistent with the results from other rodents. The body fat loss was 1.3 g in FR-d 28 groups compared with their counterpart controls. Since 1 g adipose tissue contains about 0.8 g lipid (39 kJ/g) and thus contains 31.2 kJ energy, 40.6 kJ energy would be mobilized in hamsters during a 4-week's food restriction. On average, the accumulative energy intake of hamsters during the 4-week's food restriction was 1540 kJ, (the accumulative food intake between day 1 and 28 (g)×energy content of the diet (kJ/g)). Thus, the contribution of the body fat loss to the total energy budget would be 2.6%, making us to assume that the fat reduction may induce a lower leptin levels rather than energy provision. Inconsistent with the reductions in leptin level, the time spent on activity increased in food-restricted hamsters. When these hamsters were subjected to a chronic administration of leptin, a significant reduction in activity was observed. Similarly, leptin administration to food-restricted rats, mice and Siberian hamsters attenuated or prevented running wheel activity or food hoarding behavior. These findings may suggest that leptin functioned as a starvation signal to induce an increase in activity levels, making animals to forage, food hoarding or migrate. Leptin is previously assumed to be an important signal for the switch between fed and fasted states, allowing leptin to function both as a starvation and satiety signal. Here, we also observed significant increases in serum leptin level in striped hamsters during refeeding. These hamsters showed short “compensatory growth” on the first few days during refeeding and recovered body mass and fat mass to the levels of controls, while these animals did also exhibit resistance to overweight relative to their counterparts. An increase in fat storage would enhance the probability of surviving the period of food shortage, but probably simultaneously increases the probability of being killed by a predator. The risks of predation would be a possible interpretation for this resistance to overweight in striped hamsters. Like other rodents, \], striped hamsters show hyperphagia after being refed *ad libitum*, but it is so short. Leptin supplement attenuated the increase of food intake during refeeding, and leptin was negatively correlated with energy intake in hamsters refed for 7 and 28 days, indicating that leptin presence might attenuate the hyperphagia when food was plentiful, consequently preventing over-weight and also decreasing the risk of predation. In detail, we observed that attenuation of food-intake during refeeding period was transient, and food intake on day 14 was similar in both groups. We also found a lack of leptin effect on time spent on activity on day 12 compared to day 14. Thus a short vs long-term effect of leptin supplement during refeeding period was of interest and needed to be carefully addressed in the further study. In addition to striped hamsters, exogenous leptin completely inhibits food deprivation-induced increased food hoarding and intake in Siberian hamsters. Leptin administration has a similar effect on food intake in rats and mice. These findings may suggest that leptin plays a crucial role in controlling food intake in animals with physiological hyperphagia induced by food restriction and refeeding as that taking place in striped hamsters. Based on the findings of this study, there were two possible explanations of the resistance to overweight or obesity. First, this strain of hamster only showed a transient increase in food intake when food restriction ended, and did not develop hyperphagia. Second, energy expenditure associated with activity and RMR did not decrease in refed hamsters compared with their *ad libitum* fed counterparts. Refed hamsters characterized by the lack of hyperphagia and decreases in energy expenditure were likely reach a new energy balance, consequently resulting in a resistance to overweight or obesity. In the present study leptin administration to ad libitum hamsters unexpectedly did not significantly affect either food intake or activity behavior. It is unclear why there is a different response to leptin supplement between *ad libitum* and food-restricted hamsters. The roles of leptin are dependent on both circulatory leptin levels and brain leptin transport. Leptin has been shown to be transported into the rodent brain by a saturable process. A possible explanation for this discrepancy is that the transport may be saturated in *ad libitum* hamsters regardless of the exogenous leptin and consequently show a resistance to peripheral leptin injection. Consistently, leptin treatment has a minimal effect on normal humans. In addition, several orexigenic peptides expressed in arcuate hypothalamic neurons including neuropeptide Y (NPY) and agouti-regulated peptide (AgRp), and anorexigenic peptides, e.g., pro- opiomelanocortin (POMC) and cocaine- and amphetamine- regulated transcript (CART) are found to mediate leptin action on energy balance and behavior. A further study on the response of these neuropeptide to exogenous leptin would be needed to explain the discrepancy of the roles of leptin *in ad libitum* hamsters and animals under food restriction and refeeding. # Conclusion Striped hamsters showed significant reductions in body mass, body fat content and serum leptin level, and exhibited increases in RMR and activity after being restricted to 85% of initial food intake. After being refed *ad libitum*, hamsters returned body mass, fat mass as well as serum leptin to the levels of controls, showing a “compensatory growth”, rather than overweight. In addition, striped hamsters showed a short hyperphagia on the first few days during refeeding. Leptin supplement decreased activity and attenuated the increase in energy intake. These findings suggest that the decreased leptin level during food shortage perhaps functions as a starvation signal to increase activity behavior, and when food is plentiful the increased serum leptin serves as a satiety signal to prevent activity. Finally, leptin may play a crucial role in controlling food intake and consequently preventing overweight and obesity in animals with physiological hyperphagia caused by food restriction and refeeding. We thank the reviewers for the helpful and constructive comments on this manuscript. We also thank Xian-Bin Liu, School of Agricultural Science, Liaocheng University, for his assistance with this study and care of the animals. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ZJZ. Performed the experiments: QXZ KXC YKW. Analyzed the data: ZJZ JC. Contributed reagents/materials/analysis tools: JC. Wrote the paper: ZJZ JC.

# Introduction Marine species exhibit extreme heterogeneity in dispersal ability as estimated from genetic data but despite decades of study, the underlying factors are not yet fully understood. One factor that has received a great deal of attention is life-history, since contrasting strategies can either facilitate or hinder dispersal, leading to predictions about the extent to which different species are likely to exhibit population structure. Specifically, direct developers with restricted dispersal capabilities are hypothesised to be more structured than otherwise ecologically equivalent species with pelagic larvae. In addition, pelagic larval duration is expected to correlate negatively with genetic differentiation. Although several recent meta-analytical and comparative studies lend support to these theoretical expectations, others have not been able to establish clear links between life-history traits and the strength of population structure. Such discrepancies could result from any of a large number of potentially confounding factors including larval or adult behaviour, the degree of ecological specialisation, differences in effective population size or sweepstakes-like reproductive strategies. Moreover, technical factors such as the choice of genetic marker, sample size and the completeness of geographic sampling may also play an important role. In addition to the above factors, our ability to draw general conclusions in respect of the relationship between life history and population structure is also hindered by a bias in the literature towards species from low latitudes. To obtain a more representative view of global patterns therefore requires a broadening of focus to include under-represented geographical and ecological regions such as the poles. Polar species are of particular interest because their development rates are often around 5–10 times slower than equivalent temperate species, leading to exceptionally long larval durations in broadcast reproducers. Furthermore the effects of uniquely polar ecological factors such as ice disturbance on population structure have been little studied despite their pervasive and frequent impact on shallow sites. Antarctica provides an unparalleled opportunity to undertake studies of the origins and maintenance of biological diversity, both at the levels of species and populations. Millions of years of isolation from warmer waters to the North by the Antarctic Circumpolar Current have led to the evolution of a diverse and abundant benthic fauna that is both highly endemic and cold adapted. However, parts of Antarctica are now experiencing anthropogenically induced warming at a rate that may soon outstrip the ability of many species to adapt physiologically. For example, sea temperatures have increased by 1°C to the west of the Antarctic Peninsula in the last 50 years and in parallel mean annual air temperatures on the Antarctic Peninsula have increased by as much as 3°C over the same period. This has driven the widespread retreat of glaciers, ice shelf collapse and the exposure of new habitats in both terrestrial and marine locations. In turn, extensive areas of new intense biological productivity have been generated together with associated nearshore benthic ecosystems. This places a premium on genetic studies capable of documenting ‘baseline’ patterns of population structure, which in turn may help to predict the capacity of species or populations to adapt to environmental change. Consistent with predictions based on life history, several Antarctic brooding species have been found to be so highly structured as to invoke cryptic speciation. However, other brooding species appear unexpectedly homogenous or at least show evidence of gene flow over relatively large geographic scales. Moreover, contradictory findings have also been obtained for several different Antarctic broadcast spawning species. This suggests the need for comparative studies that are able to control for as many incidental factors as possible, for instance by sampling co-distributed species from the same geographic localities, collecting equal sample sizes of individuals and using the same class of genetic marker. To evaluate the role of life-history variation within an Antarctic setting, Hoffman *et al*. recently used Amplified Fragment Length Polymorphisms (AFLPs) to analyse paired populations of the broadcast-spawning Antarctic limpet *Nacella concinna* and the brooding topshell *Margarella antarctica* sampled at five locations along the Antarctic Peninsula (see for the sampling sites). The broadcast-spawner was found to be panmictic across most of the peninsula, with only two populations from the extremes of the range, Adelaide and Signy Islands, being genetically differentiated. In contrast, population structure was seven times stronger overall in *M.antarctica*, with model-based clustering approaches assigning all of the individuals correctly to their source populations based solely on their multilocus genotypes. An intriguing aspect of the above study was the finding that *N.concinna* populations from Adelaide Island, at the base of the Antarctic Peninsula, were significantly differentiated from the other Antarctic peninsula populations despite being connected by continuous coastline and this species possessing long-lived planktotrophic veliger larvae. To explore this further, the original AFLP study was extended to include additional *N. concinna* populations from three localities within Ryder Bay, Adelaide Island (Rothera Point, Leonie and Anchorage Islands), each of which was in turn sub-sampled three times to provide a fine-scale geographic perspective. Surprisingly, limpets sampled from Rothera Point and Leonie Island could not be distinguished on the basis of their AFLP profiles from populations sampled further afield along the Antarctic Peninsula, whereas all the three sub-populations from Anchorage Island showed slight but significant genetic differences. One potential explanation for this is that coastal eddies around Anchorage Island could be advecting larvae back towards the shore, a mechanism invoked to explain local ‘hot spots’ of larval retention encountered in computer simulations. Alternatively, the exposed location of Anchorage Island on the outermost edge of Ryder Bay could predispose it to occasional sporadic larval input from sites to the south. Another possibility relates to the fact that, due to the uneven retreat of coastal glaciers and ice shelves, habitats present at Anchorage Island may be thousands of years older than those around Rothera Point. Consequently, limpet populations across Ryder Bay could have experienced markedly different demographic histories. The above studies raise additional questions in a comparative context. For example, are *M. antarctica* populations within Ryder Bay more structured than those of *N.concinna* as predicted by differences in their life history? Also, does life history influence the pattern of microgeographic population structure, or is this shaped in both species by the same underlying processes? To address these questions, we expanded our previous macrogeographic *M. antarctica* dataset to incorporate 174 additional individuals sampled from four sites over a 1–6 km spatial scale within Ryder Bay. All 414 individuals were genotyped at 228 polymorphic AFLP loci. # Materials and Methods ## Tissue sample collection *M. antarctica* samples were collected by SCUBA divers during the austral summer of 1999 from the shallow sublittoral zone off Rothera Point (East Beach), Leonie Island (Leonie North North East) and from Anchorage North and Trolval on Anchorage Island (see and for details). Samples were stored in 95% ethanol, initially for four months at −20°C and thereafter at room temperature. Previously published AFLP data from Rose Garden (Anchorage Island), Galindez, Dobrowolski, Snow and Signy Islands were also available for comparison. The combined dataset allowed us to investigate population structure over three spatial scales as follows: (i) Fine-scale (up to 2 km, within Anchorage Island) comprising samples from Rose Garden, Anchorage North and Trolval; (ii) Medium- scale (up to 6 km, among islands within Ryder Bay) comprising samples from Anchorage Island, Rothera Point and Leonie Island; (iii) Large scale (up to 1350 km, along the Antarctic Peninsula) comprising samples from Ryder Bay, Galindez Island, Dobrolowski Island, Snow Island and Signy Island. Geographic distances among these localities are given in. ## DNA extraction and AFLP genotyping Total genomic DNA was extracted from a small piece of foot tissue from each individual using a Qiagen DNeasy tissue extraction kit following the manufacturer's recommended protocols. We then used an AFLP protocol adapted from Vos et al. as detailed by Hoffman et al. that employed seven different selective primer combinations. PCR products were resolved by electrophoresis through 6% polyacrylamide gels and exposed to X-ray film for five days. These were developed using a universal X-ray developer (Xograph Healthcare Ltd.) and, if required, a second exposure was made for an adjusted time period. All bands in the approximate size range of 75–300 bp were scored manually by an experienced operator (JIH). Only clear bands with minimal size variation were included, these being recorded as 1 = present and 0 = absent. Pairs of bands that were clearly non-independent were scored as single traits. It was assumed that AFLP bands that were the same size across individuals represented homologous markers. In combining our new data with those previously published, we also revisited the original autoradiographs and were able to score a small number of additional loci across all of the samples. ## Data analysis The program A<span class="smallcaps">flp</span>-S<span class="smallcaps">urv</span> was used to calculate global *F*_st together with pairwise *F*_st values among all of the populations. Statistical significance was determined using permutation tests based on 10,000 randomisations of the dataset. Geographic distances among populations were calculated using a Geographic Information System (ESRI ArcGis v 9.2) as described in detail by Hoffman et al.. The significance of correlations between genetic and geographic distance was assessed using Mantel tests with 999 iterations implemented in G<span class="smallcaps">enalex</span> v6. # Results ## Overall population structure To a previous dataset comprising five *Margarella antarctica* populations spanning the Antarctic Peninsula we added new AFLP data for four populations within Ryder Bay, Adelaide Island (See and for details). The resulting dataset comprised 414 individuals scored for 276 AFLP bands, of which 228 (82.6%) were polymorphic. A permutation test for genetic differentiation among all nine populations based on 10, 000 randomisations of the dataset indicated a strong deviation from the null hypothesis of no genetic structure (*F*_st = 0.0872, *P*\<0.0001). Restricting the analysis to the five populations sampled within Ryder Bay, global *F*_st was lower (0.0181) but still highly significant (*P*\<0.0001), indicating the presence of fine- scale population genetic structuring. The relationship between geographic and genetic distance was highly significant overall (Mantel's *r* = 0.00, *n* = 9, *P*\<0.001), with a power regression fitting the data considerably better (*r*² = 0.811) than linear, logarithmic or second order polynomial regressions (*r*² = 0.641, 0.777 and 0.782 respectively). Restricting the analysis to populations within Ryder Bay, the isolation-by-distance relationship remained positive but failed to reach statistical significance (Mantel's *r* = 0.09, *n* = 5, *P* = 0.383). ## Population structure within Ryder Bay *F*_st values for each of the pairwise population comparisons within Ryder Bay ranged from 0.003 to 0.034. All but one of these values were individually significant, six out of ten being so at *P*\<0.0001. The only comparison failing to reach significance was that between Anchorage North and Trolval, the two geographically closest populations sampled from a single Island, Anchorage. Following table-wide Bonferroni correction for multiple statistical tests, *F*_st between Leonie Island and Rothera Point also became no longer significant (*P* = 0.076). # Discussion Although studies of the population genetic structure of marine species are commonplace, relatively few have explicitly compared direct with indirect- developers, particularly over multiple geographic scales and using the same methodologies with samples from the same sites. Consequently, we extended previous work on the brooding top shell *M. antarctica* to allow a comparison over three hierarchical spatial scales with the broadcast-spawning limpet *N. concinna*. Consistent with expectations, *M. antarctica* populations were structured throughout Ryder Bay, with the overall magnitude of genetic differentiation being greater than previously found in equivalent comparisons involving *N. concinna*. This lends further support to the notion that life history could be an important determinant of population structure in many benthic marine species. ## Strength and patterns of population genetic structure Because the previous study of *M. antarctica* sampled only five populations from the Western Antarctic Peninsula, this did not allow inferences to be drawn in respect of geographic scales below 100 km. Specifically, it was unclear whether the isolation-by-distance relationship could be linear, which would imply strong microgeographic population structure, or whether this could instead break down over finer geographic scales. By adding data from four populations within Ryder Bay, we were able to resolve a non-linear profile, with a power regression providing a greatly improved fit to the data relative to a linear one (*r*² = 0.811 versus 0.661). By implication, population structure appears to be far weaker over microgeographic than macrogeographic scales. This is consistent with a recent study of a brooding sea urchin that found genetic differences between patches separated by around ten metres, but little in the way of genetic structure within patches. In contrast, however, distances of several kilometres appear to be just as effective a barrier to dispersal as far larger regions of unsuitable habitat in a direct-developing cushion star. The reasons underlying such interspecific differences remain unclear, partly due to a paucity of empirical data. These may be specific to the species studied or could instead result from generic regional effects such as markedly stronger seasonality and temporally restricted resources, ice scour in polar regions, or the slowed development and delayed maturity characteristic of low temperature species. Despite population structure being relatively weak within Ryder Bay, global *F*_st was still significantly non-zero (0.0181, *P*\<0.0001) and around twenty times larger than the equivalent value for *N. concinna* based on the same five populations (0.0009, *P = *0.024). Such a marked difference could potentially reflect the influence of Antarctic conditions on the two species' life histories. For example, larval lifetimes tend to be longer in species adapted to cold climates, thereby extending the duration of the dispersal phase and hence the scope for gene flow. This leads to the prediction that, if all other factors could be controlled for, differences in population structure between brooders and pelagic developers could be greater at high latitudes than low latitudes. However, set against this, polar marine species tend to have disproportionately long lifespans (e.g. *Laternula elliptica* can live for up to 36 years and *Adamussium colbecki* for over 100 years). This could conceivably allow more time for adults of direct-developing gastropod species to migrate from site to site where suitable substratum and depth allow. A related observation is that *M. antarctica* and *N. concinna* not only differed in the strength of population structure, but also in the way in which genetic variation was partitioned within Ryder Bay. All but two pairwise comparisons in *M. antarctica* were statistically significant following Bonferroni correction for multiple tests, suggesting that population structure although weak may extend across much if not all of the bay. In contrast, most of the *N. concinna* populations from Ryder Bay were indistinguishable from one another, with only three populations from Anchorage Island revealing significant genetic differences. These contrasting patterns suggest that different factors may be influencing local population structure in the two species. One possibility is that coastal eddies or sporadic larval input from outside the bay could be disproportionately important in the broadcast spawner. To allow generalisation beyond the single comparison we have drawn here, it would be interesting to sample additional brooding and broadcasting species from across the bay, the expectation being that broad similarities should be found within, but not between the two classes of organism. Another factor that could potentially contribute to population structure in *M. antarctica* is benthic topology. Although Ryder Bay is only 15–20 km across, it attains a maximum depth of over 500 metres, which is beyond the distribution depth of many shallow-water marine species. There is also considerable variation in benthic topology and substratum. The latter could be important because the capacity of *M. antarctica* to disperse over soft sediments is greatly diminished relative to hard substrata such as bed rock or loose rubble. This hypothesis would be amenable to testing using a ‘landscape genetics’ approach if sufficiently fine-scale data could be collected on depth and substrate distribution across Ryder Bay. It is noteworthy in this context that the only comparison in which we did not obtain a significant *F*_st value prior to Bonferroni correction involved the two geographically closest *M. antarctica* populations, Trolval and Anchorage North, which were situated less than one kilometre apart on the coast of Anchorage Island. These two sites are separated by rocky coast and around 100 m of soft sediment at depths of 30–40 m (L. Peck pers. obs). However both sites are on the north side of Anchorage Island where steep to vertical rocky slopes descend to beyond 300 m. It is therefore possible that gene transfer between these two sites could take place along the rocky continuum beyond 40 m. ## Caveats and future directions This study was made feasible through the use of AFLPs because these markers are capable of generating large numbers of genome-wide distributed bands in virtually any organism, including diverse Antarctic marine taxa, with little need for optimisation. Although we were able to score a large number of bands, however, the dominant nature of these markers leads to increased variance in the estimation of allele frequencies. Furthermore, AFLP bands are typically assumed to be independent whereas in reality mutations, insertions or deletions can generate bands of different size that are linked. This is difficult to control for experimentally, although we attempted to do so by scoring only bands that were clearly independent from one another. Similarly, not all bands of the same size are likely to be homologous, although manual scoring may help to alleviate this problem because size homoplasious bands representing different loci may show varying intensities as well as some conformation or sequence dependent differences in electrophoretic mobility. Many of the above factors could be eliminated in future studies by deploying Single Nucleotide Polymorphisms (SNPs), since these markers are codominant and can be mapped to a reference genome to ensure selection of an unlinked, genome-wide distributed panel. Moreover, fully automated genotyping and scoring methods allow SNPs to be genotyped with minimal error. As with any type of genetic marker, AFLPs can also be difficult to compare across species boundaries because different subsets of bands will be generated and these will invariably differ both in number and polymorphic information content. However, we believe this is unlikely to account for the contrasting strength and patterns of genetic structure observed in this study, since the number of informative bands obtained for *M. antarctica* and *N. concinna* individuals from Ryder Bay was fairly similar at 189 and 155 respectively, and the smaller panel is still reasonably large. Classical analyses of population structure also assume that all loci are selectively neutral, but some studies have reported conflicting results for different types of marker, suggesting that this may not always be the case. This is implicit in the widespread use of AFLPs for conducting genome scans for ‘outlier loci’ that may be influenced by natural selection, although this approach is rapidly being superceded by Restriction Site Associated DNA (RAD) sequencing and allied approaches that draw upon emerging high-throughput sequencing technologies. These could potentially be employed in the future to determine whether outlier loci could be important in the context of this particular study. Elsewhere, several factors other than life-history have also been invoked to explain the population structure of marine species. For example, Galarza et al. found no relationship between either egg type or larval duration and the strength of genetic structure among seven littoral fish species, leading these authors to suggest that genetic connectivity could be influenced by larval or adult behaviour. Similarly, evidence from two comparative studies, suggests that habitat specificity may also have a strong impact upon population structure. Another influential factor could be effective population size since genetic drift occurs more quickly in small populations, suggesting that highly abundant species could be relatively predisposed towards being unstructured. Finally, local extinctions and recolonisations or extreme heterogeneity in the reproductive success of individuals may also lead to temporal fluctuations in population structure. To account for these and other factors presents a major challenge, since this would require both temporal sampling and the inclusion of additional phylogenetically diverse species. Finally, although our study design allowed us to control for technical factors such as the sampling scheme, choice of genetic marker and the genotyping and scoring protocols, we were not able to include further replicates at the same spatial scale from elsewhere that would help disassociate our findings from any conditions that could be specific to Ryder Bay. Detailed lichenographic studies suggest that the retreat of ice has been uneven across the bay, leaving behind a patchwork of habitats of varying ages, so we cannot exclude the possibility that the two species recolonised at different rates or from different sources, or that they could have responded differently to demographic challenges such as bottlenecks induced by the ebb and flow of sea ice. This criticism could be partly addressed by incorporating multi-level sampling from other geographical regions. However, Ryder Bay was an ideal choice for this particular study due to its proximity to the British Antarctic Survey research base at Rothera Point together with detailed knowledge of the local area. # Conclusion This study extended previous work on the direct developing top shell *M. antarctica* to include a microgeographic component. Overall, a strong but nonlinear isolation-by-distance pattern was detected, indicating relatively weak population structure over scales of 1–6 km. Nevertheless, this was consistent across the bay and greater in magnitude than previously documented for the broadcast spawner *N. concinna*. By implication, life history variation, specifically protected versus broadcast reproductive modes, may impact the population structure of benthic marine species over multiple geographic scales. This could have implications for understanding how Antarctic marine invertebrates may respond to climate change, since dispersal is a key means by which locally extirpated populations might be replenished from adjacent locations. Samples were collected during the British Antarctic Survey (BAS) Peninsula Geneflow cruise in 1999. We thank the dive team, officers and crew of RRS Bransfield for their support in the collection of these samples and the late Martin White for considerable help and support with sample design and logistical aspects. We are also grateful to Peter Fretwell for providing a matrix of geographic distances among populations and for generating the map of sampling locations. We finally thank two anonymous reviewers for comments that improved the final manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JIH AC MSC LSP. Performed the experiments: JIH AC. Analyzed the data: JIH. Contributed reagents/materials/analysis tools: JIH AC MSC LSP. Wrote the paper: JIH MSC LSP.

# Introduction With the expansion of cities, an effective transit system become increasingly important for sustainable land use planning under the influence of the travel demand. The bus route has been extended and many stations are on the bus route. Due to the unbalance of passenger demand along the bus route, overload and inadequate utilization of bus space constantly appear. This is caused by the unified frequency of bus serving for each station on the entire bus route and severely affects the transit service and satisfaction of passengers. Although an increase of the priori frequency is widely adopted to meet the high passenger demand, the serving resource of buses would be wasted at stations with low demand and the number of buses within the bus fleet is limited. Thus, the unbalance of passenger demand should be considered and it is necessary to judge whether a traditional service is suitable for the unbalanced patterns of passengers. Buses should serve the segment with high demand intensively, while stations with low demand can be skipped to balance the passenger demand and improve transit service. Previous studies presented many operational strategies to deal with the different demand patterns and the unbalanced distribution of passengers along the bus route. Tétreault and El-Geneidy considered the boarding and alighting behavior of passengers and the skipped action with limited-stop service to estimate the running time of buses. Their limited-stop service consists of a normal line and a limited-stop line. Leiva et al. minimized the total cost based on waiting time, travel time, and operators’ costs to optimize the limited-stop service and to determine their frequencies by a given origin–destination matrix. Freyss et al. implemented a continuous approximation model to determine the skipping stations based on a regular station density and affluence. Chiraphadhanakul and Barnhart optimized a limited-stop line which runs parallel to the normal line by maximizing the benefit, i.e., the total in-vehicle time savings minus total increase in waiting time. Chen et al. considered vehicle capacity and stochastic travel times and optimized the stopping design of a limited-stop bus service. Moreover, short turning and deadheading strategies have also been proposed to concentrate on segments with high demand of two directions and to decrease the running time of buses between low-demand stations. Furth and Ceder suggested that the frequency of short-turn line is equal to the parameter *n* multiplied by the frequency of the all-stop line. Cortés et al. implemented an integrated deadheading and short-turn strategy with a given OD matrix and by minimizing the total cost of user and operator. Liu et al. considered the influence of random travel time and improved the reliability of the transit system with stop- skipping service. In addition, several real-time control strategies have been studied using GPS data and smart card data. Bin et al. evaluated the reliability of the transit service to determine whether implementing the proposed partway deadheading strategy is meaningful. Bie et al. developed a new algorithm based on GPS data to partition bus operating hours into time of day intervals. Yu et al. used the variation of the intervals between a vehicle and its following vehicle to estimate the irregularity of transit operation and consequently presented a dynamic extra buses scheduling strategy. These previous studies mainly addressed the optimal scheme in limited-stop service or other stop-skipping strategy and provide appropriate skipping stations via OD with unbalanced demand. There are few studies that evaluate the degree of unbalanced passenger demand and judge which demand pattern is required to implement the stop-skipping strategy. This paper focuses on the degree assessment of unbalanced passenger demand for an origin–destination matrix and optimization of a limited-stop service for a bus fleet. The degree assessment of unbalanced passenger demand considers both the different passenger demand between stations and the unbalance of passengers within each station. Variances of both parts determine the degree of unbalanced passenger demand for the bus route, which is used to judge whether a limited-stop service should be implemented. The threshold is calibrated via numerical test. If the degree of unbalanced passenger demand is high, the limited-stop service is optimized when considering the capacity of buses and allows buses to skip several proper skipping stations. Caused by stop skipping action and bus capacity, the waiting passengers who cannot board the buses are classified into four cases: (1) left over due to bus capacity, (2) left over due to skipped action, (3) left over due to bus capacity and skipped action, and (4) left over of the last skipped bus in the bus fleet. The proper skipping stations are determined by minimizing the total cost of the transit system, including the waiting time and in-vehicle time of passengers and the running time of buses. Following this introduction, Part 2 describes the problem setting for the unbalanced passenger demand. Part 3 evaluates the degree of unbalanced passenger demand. The optimization model for the limited-stop service is established and presented in Part 4 and Part 5 presents a solution method. Part 6 analyzes the results of our numerical test and the conclusions are presented in Part 7. # Problem setting The limited-stop service can be implemented based on an unbalance of passenger demand and such an unbalance of passenger demand exists in most bus routes. Figs and show a typical load profile for passenger boarding and alighting demand on a bus corridor with 10 stations. It can be seen that the load of buses is different in the interval between consecutive stations and the boarding and alighting demand of each station are also unbalanced. Furthermore, the segment of unbalanced load is not identical to the boarding and alighting demand. The high demand of the load is concentrated on the middle of the bus line, between stations 5 and 7, while there is low boarding and alighting demand at station 7. This is because there is a high demand of passengers that want to board the bus from preceding stations and would pass station 7, compared to the demand of passenger boarding or alighting at station 7. This means that the unbalance of passenger demand is not only caused by the different patterns of boarding and alighting demand between stations, but also results from the origin–destination demand within each station. Therefore, the proposed degree assessment of the unbalanced passenger demand for the bus route should consider the different passenger demand between stations and the unbalance of passengers within the station. A bus route with *N* stations is studied, operated by a bus fleet of size *m*. The index of vehicles is denoted as *i* and each station is denoted as *j*. *λ*_*l*,*j* represents the demand of passengers between station *l* and station *j*. With a traditional service, buses serve entire stations where passengers can board and alight. However, under limited-stop service, buses would skip several stations where there is no boarding and alighting. # Degree assessment of unbalanced passenger demand Unbalance of passenger demand results from the difference of origin–destination of passengers. Due to the environment around the stations, stations along the bus route gather passengers at different demand. At a station, passengers boarding a same bus will alight at different stations. In this paper, the degree assessment of unbalanced passenger demand will be determined from two aspects: the different passenger demand between stations and the unbalance of passengers within a station. ## Degree assessment of unbalanced passenger demand between stations The limited-stop service allows buses to skip proper stations where passengers consequently cannot board or alight from the bus. Therefore, to save waiting time and avoid unnecessary transfer of passengers, the skipped stations should gather little passengers and there should be low alighting demand of passenger from preceding stations. Thus, passenger demand between stations includes both boarding and alighting demand at a station. The variance of passenger demand is used to estimate the degree of unbalanced passenger demand between stations. $$\lambda_{k}^{+} = {\sum\limits_{j = k + 1}^{N}\lambda_{l = k,j}}$$ $$\lambda_{k}^{-} = {\sum\limits_{l = 1}^{k - 1}\lambda_{l,j = k}}$$ $$\overline{p} = \frac{1}{N} \cdot {\sum\limits_{k = 1}^{N}{(\lambda_{k}^{+} + \lambda_{k}^{-})}}$$ $$\sigma_{}^{Between} = \sqrt{\frac{\sum\limits_{k = 1}^{N}{(\lambda_{k}^{+} + \lambda_{k}^{-} - \overline{p})}^{2}}{N - 1}}$$ where $\lambda_{k}^{+}$ represents the boarding demand at station *k*. $\lambda_{k}^{-}$ represents the alighting demand at station *k*. $\lambda_{k}^{+} + \lambda_{k}^{-}$ describes the passenger demand at station *k*. *σ*^*Between* represents the variance of passenger demand, which presents the degree of unbalanced passenger demand between stations. Small values of this parameter indicate a high degree of balanced passenger demand, while large values indicate that passenger demand along the bus route is unevenly distributed. ## Degree assessment of unbalanced passenger demand within the station At each station, the boarding passengers have different destinations. Skipping several stations for the high origin–destination demand of passengers can save in-vehicle time during their trips. Thus, the degree of unbalanced passenger demand within the station was also studied. $${\overline{\lambda}}_{k} = \frac{1}{N - k} \cdot {\sum\limits_{j = k - 1}^{N}\lambda_{l = k,j}}$$ $$\sigma_{k}^{Within} = \sqrt{\frac{\sum\limits_{j = k - 1}^{N}{(\lambda_{l = k,j} - {\overline{\lambda}}_{k})}^{2}}{N - k - 1}}$$ where ${\overline{\lambda}}_{k}$ represents the mean passenger demand from station *k* to the subsequent stations within station *k*. $\sigma_{k}^{Within}$ represents the variance of passenger demand within the station. ## Determination for implementing limited-stop service To comprehensively reflect the degree of unbalanced passenger demand of the transit system, both the variance of passenger demand between and within stations are considered. The degree of unbalanced passenger demand along the bus route can be defined as follows: $$Ε = \sigma_{}^{Between}/\overline{p} + {\sum\limits_{k = 1}^{N - 1}{\sigma_{k}^{Within}/{\overline{\lambda}}_{k}}}$$ If the degree of unbalanced passenger demand of the transit system is higher than a given threshold (E^*threshold*), this indicates an uneven distribution along the bus route and a larger differentiated OD demand of passengers. Buses should intensively serve stations with high demand. For this condition, a limited-stop service will be implemented to save the cost of the transit system and to fully utilize the resources of the operator. <img src="info:doi/10.1371/journal.pone.0193855.e013" id="pone.0193855.e013g" /> E ≥ E t h r e s h o l d # Optimization model ## Assumptions and definition of variables For the condition that the unbalanced passenger demand of the transit system is higher than the threshold and that the limited-stop service needs to be implemented, proper skipping stations in the limited-stop service should be determined. The limited-stop service can increase the speed of buses, save in- vehicle time, and decrease the waiting time of passengers at serving stations; however, it increases the waiting time of passengers at skipping stations because those passengers have to wait for the next bus. Therefore, to determine the appropriate skipping stations, the optimization of limited-stop service should integrate the cost of the waiting time and the in-vehicle time of passengers as well as the running time of buses. Whether Bus *i* serves Station *j* is denoted as *y*_*i*,*j* (*y*_*i*,*j* ∈ {0, 1}). If the bus stops at the station, *y*_*i*,*j* = 1; otherwise, *y*_*i*,*j* = 0. We assume that consecutive buses cannot skip the same stations to avoid a long waiting time for passengers, namely, *y*_*i*,*j*+ *y*_*i-*1,*j* = 1. Fast buses and the last bus are not allowed to skip the same station, thus: *y*_1,*j*+ *y*_*m*,*j* = 1. The definition of variables used throughout the model formulation is as follows: 1. *m* Total number of buses in the bus fleet. 2. $A_{i,j}^{+}$ Number of arriving passengers during the interval between the bus *i* and *i*+1. 3. *h*_*i*,*j* Interval between bus *i* and *i*+1. 4. $T_{i,j}^{A}$ Arrival time of bus *i* at station *j*. 5. *λ*_{*i*,*j*→*k*} Arrival rate of passengers riding bus *i* from station *j* to station *k*. 6. $\Delta P_{i,j}^{+}$ Number of passenger left over by the bus capacity of bus *i*-1 to board the bus *i*. 7. $T_{i,j}^{W,\Delta P}$ Waiting time of passengers $\Delta P_{i,j}^{+}$. 8. $\Delta S_{i,j}^{+}$ Number of passenger left over by the skipped action of bus *i*-1 to board bus *i*. 9. $T_{i,j}^{W,\Delta S}$ Waiting time of passengers $\Delta S_{i,j}^{+}$. 10. $\Delta I_{i,j}^{+}$ Number of passengers left over due to bus capacity and skipped action to board the bus *i*. 11. $T_{i,j}^{W,\Delta I}$ Waiting time of passengers $\Delta I_{i,j}^{+}$. 12. $\Delta L_{m,j}^{+}$ Number of passengers left over due to the last skipped bus in the bus fleet. 13. $T_{m,j}^{W,\Delta L}$ Waiting time of passengers $\Delta L_{m,j}^{+}$. 14. *L*_{*i*-1,*j*−1} Load of bus *i*-1 departing from stop *j*-1. 15. $N_{i - 1,j}^{-}$ Number of passengers alighting from bus *i*-1. 16. *N*_*c* Capacity of the bus. 17. *H* Average departure interval. 18. $N_{i,j}^{+}$ Number of passengers that can board the bus *i* at station *j*. 19. $T_{i,j}^{W}$ Waiting time of passengers for bus i at station j. 20. $T_{i,j}^{E}$ Dwelling time of bus *i* at station *j*. 21. *a* Boarding time per passenger. 22. *b* Alighting time per passenger. 23. $T_{i,j}^{IN}$ In-vehicle time of passengers between station *j* and station *j*+1. 24. $T_{i}^{R}$ Running time of bus *i* that completes the entire bus route. 25. $T_{i,j}^{D}$ Departing time of bus *i* from station *j*. 26. *C* Total cost of the transit system. 27. *C*_*W* Cost of the waiting time of passengers. 28. *C*_*IN* Cost of the in-vehicle time of passengers. 29. *C*_*R* Cost of running time of buses. ## Limited-stop service formulation The cost of waiting time relates to the number of passengers and the average waiting time of passengers that can board buses. When a bus arrives at a station, the boarding passengers include passenger that arrived during the interval between serving buses and passengers left by the preceding bus. The number of passengers during interval between bus *i* and *i*+1 can be expressed as: $$A_{i,j}^{+} = h_{i,j}^{} \cdot {\sum\limits_{k = j + 1}^{N}\lambda_{i,j\rightarrow k}^{}}$$ $$h_{i,j}^{} = T_{i,j}^{A} - T_{i - 1,j}^{A}$$ where $A_{i,j}^{+}$ represents the number of arrival passengers during the interval between the bus *i* and *i*+1, which is denoted as *h*_*i*,*j*, at station *j*. $T_{i,j}^{A}$ represents the arrival time of bus *i* at station *j*. Under the influence of both stop skipping action and bus capacity, part of the waiting passengers may not be able to board the buses arriving at the station. This leaves passengers for the next serving bus if the skipping buses do not serve the station or if the number of waiting passengers exceeds the vehicle capacity. If the last bus skips several stations in the optimization scheme, the waiting passengers at those stations will be left for the next optimization scheme and the waiting time will be extended (other studies always ignore this part). In this paper, we classify the waiting passengers who cannot board the buses into four cases corresponding to. The number of waiting passengers and the waiting time for Case (1), Case (2), Case (3), and Case (4) will be explained and calculated as follows: - Case (1): waiting passengers left over by the bus capacity As shown in, Case (1) describes that both the preceding bus *i*-1 and the bus *i* serve the station *j* and there are some passengers who cannot board the bus *i*-1 and consequently they have to wait for the bus *i* because the bus *i*-1 reached capacity. The number of left-over passenger due to the bus capacity of bus *i*-1 to board the bus *i* can be obtained by: $$\Delta P_{i,j}^{+} = \begin{cases} {L_{i - 1,j - 1} + A_{i - 1,j}^{+} - N_{i - 1,j}^{-} - N_{c}} & {if\mspace{180mu} L_{i - 1,j - 1} + W_{i - 1,j}^{+} - N_{i - 1,j}^{-} \geq N_{c}\mspace{180mu} \cap y_{i - 1,j} = y_{i,j} = 1} \\ 0 & {otherwise} \\ \end{cases}$$ $$L_{i,k} = {\sum\limits_{j = 1}^{k}{(N_{i,j}^{+} - N_{i,j}^{-})}}$$ $$N_{i,j}^{-} = {\sum\limits_{k = 1}^{j - 1}N_{i,k}^{+}}$$ where *L*_{*i*-1,*j*−1} represents the load of the bus *i*-1 departing from stop *j*-1. $A_{i - 1,j}^{+}$ represents the number of passengers during the interval between bus *i*-1 and the preceding bus. $N_{i - 1,j}^{-}$ represents the number of passengers alighting from bus *i*-1. *N*_*c* represents the capacity of the bus. $N_{i,k}^{+}$ represents the number of passengers that can board bus *i* at station *k*, which is determined by the remaining space in buses. The waiting time of passengers $\Delta P_{i,j}^{+}$ can be computed by: $$T_{i,j}^{W,\Delta P} = \left\{ \begin{array}{lc} {\Delta P_{i,j}^{+} \cdot (\frac{1}{2}h_{i - 1,j} + h_{i,j})} & {if\mspace{180mu} L_{i - 1,j - 1} + W_{i - 1,j}^{+} - N_{i - 1,j}^{-} \geq N_{c}\mspace{180mu} \cap y_{i - 1,j} = y_{i,j} = 1} \\ 0 & {otherwise} \\ \end{array} \right.$$ - Case (2): the waiting passengers left over due to skipped action In Case (2), bus *i*-1 skips the station *j* and bus *i* serves the station *j*. Arriving passengers during the interval between bus *i*-1 and *i*-2 cannot board the bus *i*-1. The number of passenger left over due to the skipped action of the bus *i*-1 to board the bus *i* can be obtained by: $$\Delta S_{i,j}^{+} = \begin{cases} A_{i - 1,j}^{+} & {if\mspace{180mu} y_{i - 1,j} = 0 \cap y_{i,j} = 1} \\ 0 & {otherwise} \\ \end{cases}$$ The waiting time of passengers $\Delta S_{i,j}^{+}$ can be obtained by: $$T_{i,j}^{W,\Delta S} = \begin{cases} {\Delta S_{i,j}^{+} \cdot (\frac{1}{2}h_{i - 1,j} + h_{i,j})} & {if\mspace{180mu} y_{i - 1,j} = 0 \cap y_{i,j} = 1} \\ \begin{array}{lc} 0 & {otherwise} \\ \end{array} & \\ \end{cases}$$ - Case (3): waiting passengers left over due to bus capacity and skipped action Case (3) shows that the bus *i*-2 reaches capacity, which is similar to Case (1), and the bus *i*-1 skips the station *j*. Then, the passengers left over by the bus i-2 should wait for bus *i*. The number of passengers left over due to bus capacity and the skipped action to board the bus *i* can be obtained by: $$\Delta I_{i,j}^{+} = \begin{cases} {L_{i - 2,j - 1} + A_{i - 2,j}^{+} - N_{i - 2,j}^{-} - N_{c}} & {if\mspace{180mu} L_{i - 2,j - 1} + W_{i - 2,j}^{+} - N_{i - 2,j}^{-} \geq N_{c} \cap y_{i - 2,j} = y_{i,j} = 1 \cap y_{i - 1,j} = 0} \\ 0 & {otherwise} \\ \end{cases}$$ The waiting time of passengers $\Delta I_{i,j}^{+}$ can be obtained by: $$T_{i,j}^{W,\Delta I} = \begin{cases} {\Delta I_{i,j}^{+} \cdot (\frac{1}{2}h_{i - 2,j} + h_{i - 1,j} + h_{i,j})} & {\mspace{180mu} if\mspace{180mu} L_{i - 2,j - 1} + W_{i - 2,j}^{+} - N_{i - 2,j}^{-} \geq N_{c} \cap y_{i - 2,j} = y_{i,j} = 1 \cap y_{i - 1,j} = 0} \\ \begin{array}{lc} 0 & {otherwise} \\ \end{array} & \\ \end{cases}$$ - Case (4): waiting passengers left over from the last skipped bus in the bus fleet In this paper, the number of buses in the bus fleet is set as m. Waiting passengers who are skipped by the last bus *m* in the optimization scheme at those stations will be left to the next optimization scheme and this waiting time should be integrated into the optimization of the last bus *m*. The number of passengers left over from the last skipped bus in the bus fleet can be obtained by: $$\Delta L_{m,j}^{+} = \begin{cases} A_{m,j}^{+} & {if\mspace{180mu} i = m \cap y_{m,j} = 0} \\ 0 & {otherwise} \\ \end{cases}$$ The waiting time of passengers $\Delta L_{m,j}^{+}$ can be obtained by: $$T_{m,j}^{W,\Delta L} = \begin{cases} {\Delta L_{m,j}^{+} \cdot (\frac{1}{2}h_{m,j} + H)} & {if\mspace{180mu} i = m \cap y_{m,j} = 0} \\ \begin{array}{lc} 0 & {otherwise} \\ \end{array} & \\ \end{cases}$$ where *H* is introduced to consider the waiting passengers left over from the last skipped bus *m* and is set as the average departure interval. To sum up the above four cases, the number of passengers that can board the bus *i* at station *j* can be obtained via: $$N_{i,j}^{+} = \begin{cases} {N_{c} - N_{i,j}^{-} + L_{i,j - 1}} & {if\mspace{180mu} L_{i,j - 1} + W_{i,j}^{+} - N_{i,j}^{-} \geq N_{c}} \\ {\Delta P_{i,j}^{+} + \Delta S_{i,j}^{+} + \Delta I_{i,j}^{+} + A_{i,j}^{+} \cdot y_{i,j}} & {otherwise} \\ \end{cases}$$ Then, in addition to the above four waiting times of the remaining passengers, the waiting time of passengers for bus *i* at station j should also include the waiting time of arriving passengers during the interval between bus *i* and *i*+1 $$T_{i,j}^{W} = T_{i,j}^{W,\Delta P} + T_{i,j}^{W,\Delta S} + T_{i,j}^{W,\Delta I} + T_{m,j}^{W,\Delta L} + A_{i,j}^{+} \cdot h_{i,j}^{} \cdot y_{i,j}$$ Regarding the cost of in-vehicle time, passengers spend their in-vehicle time on running time between stations and dwelling time at stations. The running time between stations can be obtained from historical data and the running time between station *j* and station *j*+1 can be denoted as $T_{j}^{R}$. The dwelling time depends on the behavior of boarding and alighting, which is the maximum value between boarding time and alighting time. The dwelling time of bus *i* at station *j* can be obtained as follows: $$T_{i,j}^{E} = \begin{cases} {\text{max}\lbrack a \cdot N_{i,j}^{+},b \cdot N_{i,j}^{-}\rbrack} & {if\mspace{180mu} y_{i,j} = 1} \\ 0 & {if\mspace{180mu} y_{i,j} = 0} \\ \end{cases}$$ where *a* is the boarding time per passenger and *b* is the alighting time per passenger. Thus, the in-vehicle time of passengers between station *j* and station *j*+1 can be obtained by: $$T_{i,j}^{IN} = (T_{j}^{R} + y_{i,j} \cdot T_{i,j}^{E}) \cdot L_{i,j}$$ The running time of buses includes the running time between stations and the dwelling time at stations, which is similar to the in-vehicle time of passengers. The running time of bus *i* that completes the entire bus route can be obtained by: $$T_{i}^{R} = {\sum\limits_{j = 1}^{N - 1}{(T_{i,j}^{A} - T_{i - 1,j}^{A})}}$$ $$T_{i,j}^{A} = T_{i,j - 1}^{D} + T_{j - 1}^{R}$$ $$T_{i,j}^{D} = T_{i,j}^{A} + T_{i,j}^{E} \cdot y_{i,j}$$ where $T_{i,j}^{D}$ is the departing time of bus *i* from station *j*. The total cost of the transit system includes the cost of waiting time and in- vehicle time of passengers and the cost of the running time of buses. The object of the optimization model is to minimize the total cost of the transit system. $$\begin{array}{ll} {\text{min}C} & {= C_{W} + C_{IN} + C_{R}} \\ & {= \sum\limits_{i}^{m}(c_{W} \cdot \sum\limits_{j = 1}^{N - 1}T_{i,j}^{W} + c_{IN} \cdot \sum\limits_{j = 1}^{N - 1}T_{i,j}^{IN} + c_{R} \cdot T_{i}^{R})} \\ \end{array}$$ where *c*_*W*, *c*_*IN*, and *c*_*R* are values of waiting time, in-vehicle time, and running time, respectively. # Solution methods To solve the objective function, a solution algorithm is presented that evaluates the degree of unbalanced passenger demand in the bus route and determines the proper skipped stations for the limited-stop service. The solution algorithm includes two parts that determine whether it is required to implement the limited-stop service and to output the optimization scheme. Considering that the design of the limited-stop service is a nonlinear problem \[0, 1\], the genetic algorithm is used to yield a minimum-cost transit operation to optimize transit routes. Genetic algorithms can efficiently cope with mixed-integer non-linear problems and the objective function gradient does not require calculation, thus reducing computational effort. illustrates the solution algorithm for the proposed strategy. The specific steps of this solution method, corresponding to, are as follows: 1. Input the passenger demand (*λ*_*i*,*j*) and the number of the entire bus fleet (*m*) and determine the threshold (E^*threshold*); 2. Evaluate the current degree of unbalanced passenger demand (E); Step 2.1Compute the degree assessment of the unbalanced passenger demand between stations (*σ*^*Between*), according to Eqs –; Step 2.2Evaluate the degree of unbalanced passenger demand within the station ($\sigma_{k}^{Within}$), according to Eqs and ; Step 2.3Determine the degree of unbalanced passenger demand along the bus route (E), according to. If E ≥ E^*threshold*, go to Step 3; otherwise go to Step 5; 3. Optimize the limited-stop strategy; Step 3.1Initiate skipping schemes (*y*_*i*,*j*); set the number of iterations for the genetic algorithm, the crossover probability *P*_*c*, and the mutation probability *P*_*m*. A chromosome consists of the entire bus fleet, as shown below: Step 3.2Calculate the value of the objective function (*C*), according to. Because the objective function is formed to minimize benefits, the objective function is equal to the reciprocal of the fitness function; Step 3.3The selection, crossover, and mutation of chromosomes are applied to produce new chromosomes. The crossover probability is set to 0.5, the mutation probability is set to 0.01; Step 3.4An elitist preservation strategy is adopted to improve both search speed and search precision. A sample after Step 4 is replaced with the best sample before Step 3.3; Step 3.5If the maximal number of generations is exceeded, then stop; Otherwise, go to Step 3.2; 4. Output the optimization program; # Numerical test The proposed strategy is tested on the bus route number 6 of Changchun City in China, as shown in. Twenty stations are on the bus route with a length of 9.9 km. The origin of the bus route is mainly within residential areas and the destination is located downtown. The running time between the origin and destination of the bus route is about 28 min. The travel speed of buses is about 13 km/h. The average passenger boarding and alighting times are 1.5 sec and 3.0 sec, respectively, referring to the practicalities of reinvestigation and a previously published study (Chen et al. 2015). The time value is assumed as \$15/pax-h for waiting time, \$10/pax-h for in-vehicle time, and \$50/veh-h for running time. The number of the entire bus fleet *m* is six vehicles. Historical data of passenger demand was collected during 30 days. We obtained the bus OD data, which records the boarding and alighting station of each passenger, and the running time between stations via on-board surveys of the entire bus route during the peak time period. Furthermore, the running time between stations of the bus route number 6 is shown in. At this time, the average passenger demand shows an unbalanced distribution, as shown in. In the solution method, the number of iterations of the genetic algorithm NG is set to 100. The crossover probability is set to 0.5, and the mutation probability is set to 0.01. depicts the convergence of the calculation with experimentation for 10 times. The algorithm has good convergence and optimal solutions of the objective function are found within 70 iterations. The mathematical software MATLAB R2011a is applied to simulate bus operations, which is expected to provide a fairly reliable environment to test the proposed strategy. Furthermore, the data is used to determine the threshold of degree of unbalanced passenger demand to judge which services are suitable for the bus route. ## Determination of the threshold of the degree of unbalanced passenger demand The threshold of the degree of unbalanced passenger demand can reveal whether it is necessary to implement the limited-stop service. A high degree of unbalanced passenger demand indicates potential skipping stations which can save the total cost of the transit system. Test data is used to construct an examination. In this examination, the degree of unbalanced passenger demand is evaluated and the corresponding total costs of transit system with and without the limited-stop service are computed. shows the total cost of the transit system both with and without the limited- stop service under different degrees of unbalanced passenger demand. This shows that the degree of unbalanced passenger demand concentrates between stations 4 and 8 and the cost of the transit system with traditional service is random under different degrees of unbalanced passenger demand. Furthermore, when the degree of unbalanced passenger demand is below 6.0, the limited-stop service causes little difference to the traditional service. This is because the distribution of passenger demand is relatively flat. Skipping several stations has little influence on saving the total cost of the transit system and skipping stations are even completely removed through the optimization. However, when the degree of unbalanced passenger demand is higher than 6.0 (indicating that there are potential skipping stations), limited-stop service exerts a strong effect on saving the total cost. Therefore, the threshold of the degree of unbalanced passenger demand is set to 6.0. ## Results To evaluate the performance of the proposed strategy, a dataset with a degree of unbalanced passenger demand of 6.3, is optimized to determine the skipping stations. shows the optimization scheme of a transit system for six vehicles. The segment with low demand of passengers (including boarding demand and alighting demand) concentrates on stations 4, 9, and 14. Each bus skips several stations and consecutive buses cannot skip the same station. Furthermore, performances of traditional service and limited-stop service are compared. Under traditional service, buses stop at each station along the entire bus route and each *y*_*i*,*j* is set to 1. The design of a limited- stop service needs to adopt the above solution method, which searches *y*_*i*,*j*, and then determines the proper skipping stations. The data used to test the effect of the traditional service and limited-stop service uses the same passenger demand and running time of buses between consecutive stations. The results of the total cost per hour in response to two services and the reduction in the total cost are shown in. shows that the limited-stop service has a strong effect on saving the total cost of the transit system with a high degree of unbalanced passenger demand. However, there are exceptions for data 3 and 9, where the degree of unbalanced passenger demand is 4.1 and 3.4, respectively, and where no proper skipping stations exist and the optimization scheme is stopping at entire stations. The reductions in total cost for data 3 and 9 are both zero. Besides the data 3 and 9, the limited-stop service can be implemented under a high degree of unbalanced passenger demand. Based on the tested data, there is at least a 7% reduction and the highest reduction is 13%. Therefore, the limited-stop service is not always valid for every pattern of passenger demand and it is necessary to determine the degree of unbalanced passenger demand to judge whether the limited-stop service should be implemented to decrease the total cost of the transit system. # Conclusions Our study is assuming an unbalanced distribution of passenger demand along the bus route, which most transit systems have to accommodate, and provides a limited-stop service for the entire bus fleet. To determine the condition of implementing the limited-stop service, it is useful to consider the degree of unbalanced passenger demand as a judgment condition. In this study, the degree assessment of unbalanced passenger demand reaches the unbalanced distribution of passenger demand between stations and within the station. The variance of the unbalanced distribution of passenger demand is computed as the degree of unbalanced passenger demand. A high degree of unbalanced passenger demand indicates that the bus route needs to implement a limited-stop service. The design of the limited-stop service provides an optimization scheme that includes proper skipping stations for a bus fleet. The limited-stop service allows buses to skip several stations with low passenger demand and consecutive buses cannot skip the same station, considering the waiting time of passengers at those stations. As a result of the test, several conclusions can be formulated: When comparing the total cost, including waiting time cost, in-vehicle time cost, and running time cost, between both with and without limited-stop service and under different degree of unbalanced passenger demand, changes in the total cost with limited-stop service are more obvious in response to different degrees of unbalanced passenger demand. The threshold should be set to 6.0. Under a low degree of unbalanced passenger demand, traditional service and implementing limited-stop service result in little difference. When the degree of unbalanced passenger demand is higher than the threshold, implementing limited-stop service can save the total cost of the transit system. The design of a limited-stop service can respond well to a change in degree of unbalanced passenger demand under different distribution patterns. If applicable, a limited-stop service can save a total cost of about 10%. Furthermore, there are several limitations of this study, which need to be addressed. The limited-stop service needs to be extended to branching corridors that are composed of several feeder lines and a trunk line. Real-time changes of passenger demand can also be integrated into the optimization. Other operational strategies, such as deadheading and short turning, could be combined to deal with the unbalance of passenger demand. # Supporting information This work was supported by the National Science Foundation for Young Scientists of China (grant No. 51608224) and the National Natural Science Foundation of China (grant No. 51378237). [^1]: The authors have declared that no competing interests exist.

# Introduction During mammalian evolution, integration of retroviral RNA into a germ line cell may have led into a formation of a provirus that is transmitted vertically and inherited in a Mendelian manner. In humans, these endogenous retroviruses (HERV) comprise ca. 8% of our genome. While it is known that some retroelements of the human genome are still capable of retrotransposition, DNA sequences of the HERVs have accumulated mutations to the point where retrotransposition or formation of viral particles is not taking place anymore. Despite this mutation-driven functional inactivation, there are hundreds of publications demonstrating associations between HERV expression and various disease states (malignancies, infections, neurological and autoimmune diseases), however, the causal relationship has remained enigmatic. Since the mechanism of action cannot be explained by *de novo* insertional mutagenesis nor with the formation of viral particles, it has been proposed, that potential pathogenicity of the HERVs could simply underlie in the presence of proviral DNA, acting as a transcriptional regulatory sequence, modifying the expression of neighboring and even more distant genes. HERVs can do this for example by acting as transcription factor binding sites. From this hypothesis it naturally follows, that potential effects of the HERVs would be restricted in some genomic window around the primary proviral insertion site. However, there is also evidence supporting more global mode of action as HERVs have been shown to activate immune and inflammatory responses of the body directly. For example, their RNA could be recognized as a pathogen-associated molecular pattern (PAMP) by Toll-like receptors and this would induce type I interferon production contributing to the pathogenesis of autoinflammatory diseases. Some HERVs are still able to encode an intact envelope protein (Env) and its presence has been observed in some viral infections or in autoimmune diseases. It has been proposed that the mechanism of action of Env is based on the antigenicity of the molecule, possibly causing a polyclonal activation of lymphocytes, i.e. functioning as a “superantigen”. As the diseases, where HERV-associations have been observed, demonstrate some of the fundamental and characteristic aspects of aging, e.g. increased level of inflammation and changes in the proportions of the various lymphocyte subsets, we now quantitated the RNA levels of all previously characterized proviruses of HERV-K (HML-2) and HERV-W families in peripheral blood mononuclear cells (PBMC) derived from young and 90-year old individuals. Aging-associated increase in the expression of several HERV families has been reported previously using quantitative PCR. However, qPCR approach utilizes degenerate primers for each HERV family, thus missing the information regarding individual proviruses. RNA- sequencing possesses the capability to obtain this crucial data and hence it was the method of choice. The most recent entrants to our genome are represented by HERV-K (HML-2) family (ca. 0.2–2 million years ago), of which Subramanian et al. have identified 91 full-length proviral sequences. HERV-W represents an older group of HERVs (primary infection ca. 40 million years ago) and it contains 213 full-length or near full-length elements. # Methods ## Study populations Two populations, representing young and elderly individuals, were used. The young ones consisted of healthy laboratory personnel, all female, aged 26 to 32 years (n = 7, median age 28) who did not have any medically diagnosed chronic illnesses, were non-smokers and had not had any infections or received any vaccinations within the two weeks prior to blood sample collection. The elderly individuals (n = 7) were selected among relatively healthy, community living, non-frail, nonagenarian females, without any severe aging-associated diseases, that were participants in The Vitality 90+ study. The nonagenarians were born in 1920 and the samples were collected in 2014. The recruitment and characterization of participants were performed as has been reported previously. The study participants provided their written informed consent. This study was conducted according to the principles expressed in the declaration of Helsinki, and the study protocol was approved by the ethics committee of the city of Tampere (1592/403/1996). ## Sample collection Blood samples were collected by a trained laboratory technician in the laboratory facilities. All blood samples were drawn between 8 am and 12 am and collected into EDTA containing tubes. Samples were directly subjected to leucocyte separation on a Ficoll-Paque density gradient (Ficoll-Paque Premium, cat. no. 17-5442-03, GE Healthcare Bio-Sciences AB, Uppsala, Sweden). The PBMC layer was collected and cells used for RNA extraction were suspended in 150 μl of RNAlater solution (Ambion Inc., Austin, TX, USA). Nonagenarian and control samples were collected at the same time. ## RNA extraction RNA used for RNA sequencing was purified using a miRNeasy mini kit (Qiagen, CA, USA) and the RNA used for PCR analysis using RNeasy mini kit (Qiagen, CA, USA) according to manufacturer’s protocol with on-column DNA digestion (Qiagen). The concentration and quality of the RNA was assessed with a NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE, USA). ## RNA sequencing Agilent Bioanalyzer RNA nano chips (Agilent) were used to evaluate the integrity of total RNA and Qubit RNA–kit (Life Technologies) to quantitate RNA in samples. 1 μg of total RNA was used for ScriptSeq Complete Gold System (Epicentre) to ribodeplete rRNA and further for RNA-seq library preparation. SPRI beads (Agencourt AMPure XP, Beckman Coulter) were used for purification of RNAseq libraries. The library QC was evaluated on High Sensitivity chips by Agilent Bioanalyzer (Agilent). Paired-end sequencing of RNAseq libraries was done using Illumina HiSeq technology with a minimum of 60 million 2x100bp paired-end reads per sample. ## Data preprocessing and analysis Raw reads were aligned to human genome reference build hg19 using TopHat v2.0.13 with the default parameters. Only uniquely mapped reads were considered in the transcript abundance estimation and to this end SAMtools was used to filter out reads mapping to multiple regions of the genome. The downstream analyses were all conducted using the tools in cufflinks2 v. 2.2.1. The raw expression estimates were calculated using cuffquant and the expression were normalized using cuffnorm, which gives the normalized read counts and the fragments per kilobase per million values (FPKM) for each gene as an output. The geometric normalization method was used which scales the read counts as well as the FPKM values according to procedure described in. The annotation data for HERV-K (HML-2) was from Subramanian et al. and that for HERV-W from Grandi et al. To ensure the robustness of the normalization the expressions of HERV elements were quantified and normalized together with ENSEMBL v. 82 gene reference set. For each individual study subject, a given HERV element was considered significantly expressed when the individual expression level exceeded normalized read count of 16. ## Cluster analysis Hierarchical clustering of the samples based on normalized read counts was done separately for both HERV-K (HML-2) and HERV-W. Spearman correlation was used as the distance metric, which is robust against outliers and non-Gaussian distributions, and can capture nonlinear relationships. Ward's minimum increase of sum-of-squares was used as the linkage method, which has been reported to perform better with gene expression data than the more traditional methods of average and complete linkage. Multistep-multiscale bootstrap resampling was done to evaluate the uncertainty involved in the clustering. Thousands of samples of varying sizes are randomly created from the data and then clustered. An approximately unbiased (AU) p-value is obtained, which indicates the bias corrected percentage of dendrogram variants where the specific cluster was observed. # Results The results of the RNA-seq analysis indicated that 33 HERV-K (HML-2) loci out of 91 had a detectable expression, but often at low level and not in all individuals. The expression levels of PBMCs derived from young and elderly individuals were generally similar. Only at three loci (1q22, 10p14 and 12q24.33) the difference was statistically significant as shown in. In the case of HERV-W, the results were similar, in 45 proviruses out of 213 the read count was \>16 at least in one individual, and in the case of Xp11.21 the difference in expression levels between the young and old was significant as shown in. Hierarchical clustering of the samples based on normalized provirus read counts was done to investigate expression patterns. Clustering of samples based on HERV-K (HML-2) expression resulted in two groups separated along the age group lines. There were two deviations from this, with one nonagenarian in the predominantly young sample cluster and one young sample in the nonagenarian cluster. Heatmaps of the clustering of the HERV-K (HML-2) and HERV-W provirus expression levels are shown in Figs and, respectively. Bootstrap resampling of the clustering was done to quantify the certainty of the clustering. Both clusters have an approximately unbiased (AU) p-value of 97, which is the bias corrected percentage of resampling dendrogram variants where the specific cluster was observed. AU p-value of 97 is equivalent to a p-value of 0.03, indicating statistical significance. The same significant clusters resulted even if the significantly differentially expressed 1q22, 10p14 and 12q24.33 were excluded from clustering. HERV-W expression based clustering of samples resulted in one statistically significant cluster (p-value of 0.04), which contains the same six nonagenarian samples that are grouped together in the HERV-K (HML-2) based clustering. # Conclusions The RNA levels of individual proviruses varied considerably between samples. It was not the case that some individuals would have been more active producers than others, but instead different proviruses seemed to be expressing non- systematically within and between individuals. A total of eight HERV-K proviruses and nine HERV-W proviruses were found to be expressed in all 14 samples and consequently these proviruses were expressed with highest RNA levels. This suggests that some individual proviruses could be less restricted in terms of their expression potential, that is brought by the regulation machinery of the cell. Several proviruses were expressed only in small part of individuals and it is tempting to think that these could be the ones behind potential adverse effects, especially if they would be mainly expressed in nonagenarians, as they are probably silenced for a reason. There were no proviruses that were expressed exclusively in nonagenarians, but for example HERV-K 8p23.1a was expressed in 6 nonagenarians and only in 1 young individual. Furthermore, only 4 aging-associated differentially expressed proviruses were identified (in HERV-K (HML-2) 1q22 and 10p14 having a higher and 12q24.33 a lower expression in the elderly and in HERV-W Xp11.21 a lower expression). Putting all this together, it seems to be the case, that aging has only moderate effect on the expression levels of individual proviruses. However, the hierarchical clustering of the expression data indicated that the expression profiles of the young and elderly subpopulations were different. The simplest way to achieve this kind of difference would be if, for example, all the proviruses were expressed systematically slightly up or down in one of the groups. This kind of behavior could be attributed to some kind of common regulator that has only one simple mode of action. However, this was not the case, as different proviruses were up- and downregulated equally in the nonagenarians. This requires more complex regulation and is possibly reflecting multilayered epigenetic regulation machinery involving, among other things, DNA methylation and histone modifications, and inducing distinguishable aging- associated expression profile. Due to Spearman correlation based distance metric in the clustering, each provirus has identical weight in the clustering result, regardless of level of expression. Therefore this result would indicate that there are differences between the age groups that are revealed when the proviral expression profiles are examined as a whole. The underlying cause behind observed expression profile difference thus has to affect the expression of many different proviruses. Understanding what causes this difference could increase knowledge of HERV expression associated disease states and of age-related decline. Since the same nonagenarian samples are clustered by both HERV-K (HML-2) and HERV-W expression, this phenomenon may not be limited to these families, and could be present in other HERV families as well. It is noteworthy, that our analysis is only limited to HERV-K (HML-2) and HERV-W families. Previous studies have indicated that upregulation of some other HERV subclasses might also have implications in tumor immunity. Therefore, it is possible that these HERVs could contribute to aging more than HERV-K (HML-2) and HERV-W. This remains to be explored in future studies. There is a general agreement that the expression of HERVs should be under a strict control, i.e. allowing their expression in the germ line but silencing in most somatic cells, where their activity could disrupt normal gene expression or transcript processing. Several of these control mechanisms have been characterized in detail. As human aging is associated with dramatic epigenetic changes, e.g. DNA methylation, it is maybe surprising that expression levels between the young and old individuals were not strikingly different. However, it is possible that this epigenetic regulation is responsible for the observed differences and the expression profiles would be due to differential sensitivity of the individual proviruses to these aging-associated epigenetic changes. The general expression profile of HERV-K (HML-2) in the resting blood cells used here, was dominated by a few loci, i.e. 3q12.3, 19q13.12b and 1q22., resembling the situation in in vitro pre-activated lymphocytes, suggesting that the proliferative state of the cells has probably only a minor effect. This far, no similar data in the case of HERV-W is available. In conclusion, transcriptional regulation of the proviruses belonging to HERV-K (HML-2) and HERV-W families appears to be two-dimensional in the PBMCs; a subset of HERVs are expressed constantly in age-independent manner having only slight aging-associated differences in the expression levels. These differences might be explained by a fine-tuning of transcriptional regulation that is brought by DNA methylation and is known to be heavily altered in aging. On the other hand, proviruses in another subset of HERVs were characterized by total lack of expression in some individuals. This could be the result of some more drastic mode of regulation such as that of H3K4me3, that is also known to be altered in aging. Aging-dependent HERV profile found with clustering might reflect this aspect of regulations and it is also possible that adverse effects of HERVs are driven by those proviruses that undergo more radical transcriptional relaxation or restriction, that is not necessarily seen in the median RNA levels (for example HERV-K 8p23.1a in and HERV-W 11q14.2). This finding might have some practical consequences. In the clinical studies demonstrating associations with HERV-expression the expression of only one or a few proviruses have been used as the indicator. In these studies, analysis of the whole HERV profile would help in finding the true pathogenic provirus. Small number of samples is a limitation of this study, and more comprehensive studies with bigger sample populations are needed for confident evaluation of the RNA levels. We would like to thank Sinikka Repo-Koskinen for her skillful technical assistance. [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Institute of Ageing and Chronic Disease, University of Liverpool, Liverpool, United Kingdom

# Highlights The determinants of conversion from benign fatty liver disease to non-alcoholic steatohepatitis are not well understood. These studies show that lack of the transcription factor ARNT in myeloid cells predisposes mice to NASH. Type 2 diabetes (T2D) and liver disease are commonly associated. Liver ARNT is decreased in people with T2D and with liver disease. ARNT may be a common pathogenic factor in diabetes and liver disease. # Introduction Non-alcoholic fatty liver disease (NAFLD) is defined as the accumulation of excess, microscopically visible lipid in hepatocytes in the absence of excessive alcohol consumption and is now the most common chronic liver disease in developed countries. The risk of NAFLD is increased with obesity, insulin resistance and type 2 diabetes mellitus (T2DM); the incidence of these risk factors has been increasing in recent years. Amongst overweight and obese people with type 2 diabetes, NAFLD is present in over 50% of cases. Importantly, \~20% of patients with NAFLD progress to develop non-alcoholic steatohepatitis (NASH), which in addition to steatosis is characterised by lobular inflammation, hepatocyte ballooning and fibrosis. NAFLD is relatively benign, however, NASH increases the risk of cirrhosis, liver failure and hepatocellular cancer, with rates of cirrhosis estimated at 5–20% over 10 years. Factors known to influence the progression from NAFLD to NASH include greater hepatocyte lipid accumulation, insulin resistance, oxidative stress leading to lipid peroxidation, production of pro-inflammatory cytokines, and mitochondrial dysfunction. NASH also occurs in the setting of altered systemic concentrations of adipokines (such as leptin, interleukin-6 and adiponectin) which influence hepatic lipid accumulation and insulin sensitivity. Myeloid cells play a key role in NASH progression. In the liver, macrophages contribute to the development of NASH and fibrosis in both the inflammatory and resolution phases. In obesity, excess M1 type macrophages not only accumulate in adipose tissue but also in liver, where they produce the chemokine (C-C motif) ligand 2/monocyte chemotactic protein-1 (CCL2/MCP-1). Macrophages also influence whole-body insulin sensitivity and glucose metabolism, as demonstrated by the striking phenotypes of a number of myeloid- and macrophage-specific conditional knockout mice. Chronic intermittent hypoxia also contributes to hyperlipidaemia, lipid peroxidation and the development of NASH in mouse models. This relationship is supported by cross-sectional studies in humans with obstructive sleep apnoea. The transcriptional response to hypoxia is regulated by hypoxia-inducible factors (HIFs), active heterodimeric transcriptional complexes that can respond to a variety of environmental signals. HIF1 is a heterodimer containing HIF-1α and the Aryl hydrocarbon Receptor Nuclear Translocator (ARNT, also known as HIF-1β); while HIF-2α and ARNT comprise the HIF-2 complex. HIF-1α activity is reduced at high glucose concentrations in human fibroblasts and diabetic animals. ARNT with the aryl hydrocarbon receptor (AhR) regulates response to environmental toxins including dioxin and other cyclic hydrocarbons. Accordingly, deletion of HIFs in myeloid cells decreases NASH progression while myeloid cell-specific HIF-1α or HIF-2α deletion impairs immune function. In contrast, deletion of macrophage AhR leads to opposing effects with an increased acute inflammatory response. With regard to ARNT, expression decreases in the liver and pancreatic islets of patients with type 2 diabetes, and deletion of ARNT in these tissues results in impairment of metabolism. We therefore hypothesised that deletion of myeloid ARNT, which binds to HIF-2α, AhR and SIM2 as well as HIF-1α, would influence steatohepatitis progression during high-fat diet (HFD) feeding. # Materials and methods ## Animal studies All animals received humane care according to the criteria outlined in the “*Australian code of practice for the care and use of animals for scientific purposes*”. Procedures were approved by the Garvan/St. Vincent’s Hospital Animal Ethics Committee. Floxed ARNT mice were created as previously described and bred with LysM-Cre mice, to produce myeloid cell ARNT knockout mice (LAR) and floxed- control (FC) littermates. All mice were maintained on an inbred C57Bl/6 background for at least 12 generations. ## Housing and high-fat diet feeding Mice were housed with a 12-hour light/dark cycle at room temperature, with *ad libitum* access to food and water. From 10–12 weeks of age, mice were fed a high-fat diet (HFD, 45% of energy from fat, based on D12451, Research Diets, New Brunswick, NJ, USA) for 20 weeks until the conclusion of the experiment. ## Metabolic testing For glucose tolerance tests (GTT) and insulin tolerance tests (ITT), mice were fasted for 6 hours then dextrose (2g/kg body weight) or insulin (0.25U/kg) were given by intraperitoneal (IP) injection. For the pyruvate tolerance tests (PTT), mice were fasted overnight (16 hours) prior to administration of pyruvate (2g/kg, *i*.*p*.). In all tests, blood glucose measurements were taken from tail blood using an Optium glucometer (Abbot Diabetes Care, Doncaster, Australia). ## Tissue collection Mice were sacrificed after a 6 hour fast at least 1 week after the last metabolic test. Under anaesthesia (2,2,2-tribromoethanol or ketamine/xylazine), blood was collected by cardiac puncture into tubes containing 20 μl of 0.5M EDTA. Plasma supernatant was stored at -80°C. Livers were collected and divided for formalin fixation (histology) and the remainder was snap-frozen in liquid nitrogen for gene expression and lipid studies. Epigonadal and subcutaneous fat depots were collected, weighed and formalin-fixed prior to histology. ## Macrophage isolation Four days prior to sacrifice, mice were injected with 2ml of 3% thioglycollate IP (Difco, Melbourne, Australia). At sacrifice, after anaesthesia (isofluorane, 5% induction and 2% maintenance), macrophages were isolated by IP injection of 10ml of sterile ice cold PBS. Cells were cultured in RPMI medium (Gibco, Melbourne, Australia) with 10% fetal calf serum and 1% L-glutamine. Two hours later, cells were washed twice with PBS and cultured for a further 24 hours before RNA extraction. ## Gene expression analysis Samples were lysed in RLT buffer (Qiagen, USA). RNA was isolated and cDNA was synthesized as previously described. Real-time PCR was performed using an ABI7900 instrument in combination with SYBR Green PCR master mix (both from Applied Biosystems, Melbourne Australia). Sequences of PCR primers are available on request. For each gene, mRNA expression was corrected to that of TATA-box binding-protein (*Tbp*) using the 2^ΔΔCT method. As described previously expression of 86 genes related to fatty liver was measured using Mouse Fatty Liver RT² Profiler PCR Arrays (#330231 PAMM-157ZA) and RT² SYBR Green qPCR Mastermix (Qiagen, Doncaster, VIC, Australia). Results were analysed using RT² Profiler software, and expression was normalised to *B2m* (beta-2 macroglubulin), *Gapdh* (glyceraldehyde-3-phosphate dehydrogenase) and *Gusb* (beta glucuronidase) as housekeeping genes. ## Histology NAFLD / NASH grade was scored by A.C. masked to genotype according to Kleiner *et al* 2005. Tissues were fixed in 10% buffered formalin, paraffin-embedded and cut into 5μm sections. Sections were stained with haemotoxylin and eosin (H&E), Perl’s stain, Sirius Red or Milligan’s Trichrome staining according to standard protocols. F4/80 staining was performed using the DakoCytomation EnVision+ Dual Link System-HRP (DAB+) Kit (Dako, North Sydney, NSW, Australia) as per manufacturer’s instructions. Following antigen retrieval, rat F4/80 monoclonal antibody was diluted 1:100 in Antibody Diluent (Dako). After staining, slides were counterstained using a standard Haematoxylin protocol. Adipose fat cell size was calculated using ImageJ software by tracing all adipocytes within the microscope field of view and calculating the average size. ## Liver function and triglyceride (TG) content Plasma [Alanine transaminase](http://en.wikipedia.org/wiki/Alanine_transaminase) (ALT) and Aspartate transaminase (AST) levels were measured by the St Vincent’s Hospital Pathology department. Liver TG was extracted from 30-40mg of homogenized liver using a modified Folch method, and measured using a Roche TG kit (GPO-PAP, Mannheim, Germany). ## Western blotting Fifty mg of snap-frozen liver tissue was homogenised in ice-cold RIPA buffer containing 1 mg/l aprotinin, 1 mg/l leupeptin, 10 mmol/l NaF, 1 mmol/l Na3VO4 and 1 mmol/l PMSF. Cleared lysates were electrophoresed in 12% polyacrylamide gels and transferred to PVDF membranes (Thermo Scientific, \#88518). Primary antibodies used included those directed against ADIPOR2 (Sigma SAB1102579), FABP1 (Thermo Scientific, \#720242), and MAPK8/JNK (Cell Signaling, \#9252S). For mitochondrial oxidative phosphorylation components, we used Total OXPHOS Rodent WB Antibody Cocktail (Abcam, ab110413). Membranes were washed and probed with appropriate HRP-conjugated secondary antibodies (Bio-Rad, Cell Signaling), and proteins were visualised using Super Signal West Pico Chemiluminscent Substrate (Thermo) and a ChemiDoc Imaging System (Bio-Rad). For a loading control, we stripped the membranes with 0.2 M NaOH for 10 minutes and reprobed with an antibody directed against the regulatory molecule (14-3-3 Santa Cruz Biotechnology, catalog \#sc-1657). Densitometry was performed using ImageJ software (NIH freeware), and comparisons between FC and LAR livers were performed using unpaired two-tailed t-tests. ## Human liver samples Human liver samples were obtained from patients with NAFLD who had undergone liver biopsy at Westmead Hospital and had stored liver tissue. Patient characteristics are shown in for this group. Non-steatotic (control) liver samples were collected from healthy regions of the livers from donors who had undergone liver resection for benign liver tumours for whom all other causes of liver disease were excluded, as previously described. The ethics approval for the normal liver de-identifies the subjects and therefore patient information is not available for the normal group. Liver biopsies were scored by an expert liver pathologist unaware of clinical data. Histological scoring was based on the system proposed by Kleiner et al.. Steatosis was graded from 0 to 3, lobular inflammation from 0 to 3 and hepatocellular ballooning from 0 to 2. Fibrosis was staged from 0 to 4 with 4 representing cirrhosis. The NAFLD Activity Score (NAS) was calculated to quantify disease activity. All biopsies were of appropriate size and included enough portal tracts for pathological grading and staging of the histological features. Samples with a NAS ≥ 5 or a NAS of 3–4 but with fibrosis were included in the NASH group. Subjects with evidence of secondary causes of steatosis or alternative diagnoses were excluded including alcohol (men, \>30 g/day; women, \>20 g/day), total parenteral nutrition, chronic viral hepatitis (hepatitis B and hepatitis C), autoimmune liver diseases, hereditary hemochromatosis, α1-antitrypsin deficiency, Wilson’s disease and drug-induced liver injury. Ethics approval was obtained from the Human Research Ethics Committees of the Sydney West Local Health District and the University of Sydney. Written informed consent was obtained from all participants. PCR of human liver was done as above, except that RNA from biopsies was assessed using the Agilent 2100 Bioanalyser (Agilent, Waldbronn, Germany) before use. cDNA was prepared using qscript (Quanta Biosciences, Gaithersburg, MD, USA) in a Mastercycler gradient 5331 (Eppendorf AG, Hamburg, Germany). Gene expression for ARNT was measured by qPCR as above. GAPDH was used as the house-keeping gene. ## Statistical analysis Results were analysed using GraphPad Prism software. LAR and FC mice were compared using two-tailed Student’s t-tests, or nonparametric Kruskal-Wallis Test, as appropriate. Time-course experiments were analysed using repeated measures ANOVA (rmANOVA). A p-value \<0.05 was considered significant. Unless indicated otherwise, all data is shown as mean ± SEM. # Results ## Deletion of ARNT from myeloid cells leads to decreased liver weight and increased subcutaneous fat mass ARNT deletion efficiency in this line is \>80%, as previously reported. To increase hepatic lipid load, LAR and FC mice were fed a high-fat diet from 10–12 weeks of age (HFD) and continued for 20 weeks. LAR and FC mice had equivalent weight gain (**,** males). However, by 20 weeks feeding, liver weight was reduced in LAR compared to FC mice (1.42g versus 1.71g respectively, or 3.1% versus 3.6% of total body weight, **,** p = 0.046). This was despite a 23% increase in liver triglyceride content. At sacrifice, despite similar body weight, LAR mice had significantly more subcutaneous fat (3.6±0.2g versus 2.7±0.4g respectively), which equated to 7.6% versus 5.8% of total body weight (p = 0.038). This was accompanied by a significant increase in adipocyte size in this depot. In contrast, overall epigonadal fat mass was not increased with only a trend to increased mass in LAR mice in this depot (p = 0.075). ## Deletion of ARNT from myeloid cells in mice increases the prevalence of NASH NASH was scored histologically in HFD-fed LAR and FC mice by a histo-pathologist masked to mouse genotype (A.C.). All mice, regardless of genotype, developed either NAFLD or steatohepatitis at 20 weeks of HFD. Steatohepatitis developed in 9 of 14 LAR mice (64%) versus 3 of 10 FC (30%). Consistent with this, the NAS (NAFLD activity score) was higher in LAR mice at 4.7±0.5 versus 3.2±0.5, p\<0.05 (**)**. Individual components of the NAS score showed a trend to greater steatosis (2.7 versus 2.2, p = 0.09), no difference in hepatocyte ballooning (1.1 versus 0.7, p\>0.2) and a significant increase in lobular inflammation (0.9±0.3 versus 0.3±0.2, p\<0.05, **).** Representative histological sections for FC and LAR mice are shown in. Livers from LAR mice had greater TG accumulation than those from FC mice ****. Global Sirius red staining to assess fibrosis did not show a significant difference. However, LAR mice had localised areas of increased fibrosis. shows a higher- power example of localised fibrosis in a LAR mouse. There was no difference in iron status (Perl’s stain, **).** Increased macrophage infiltration in the livers of LAR mice was indicated histologically by F4/80 staining (**)**. ## Changes in gene expression in livers from LAR mice Consistent with the histological findings, liver mRNA expression of the macrophage marker *F4/80* was more than two fold higher in LAR HFD mice (**,** p\<0.00001). Examination of genes reported to differentiate between M1 and M2 macrophage phenotype showed that both groups were elevated without any obvious preponderance of M1/M2. Among pro-fibrogenic genes, there was a significant increase in collagen type 1α1 (*Col1A1* mRNA) and a trend to increased matrix metallopeptidase 9 (*Mmp9)* expression in LAR livers (p\<0.0003 and p = 0.052 respectively). In addition, there was increased expression of both Bcl2-antagonist/killer 1 (*Bak1)* and B-cell lymphoma-extra-large *(Bcl-xl)*, (p = 0.0015 and p\<0.0005 respectively). After HFD there was no difference in liver expression of the insulin-signaling genes *Akt2*, insulin receptor *(Ir)*, or insulin receptor substrate 1(*Irs-1)*. There was however a significant decrease in insulin receptor substrate 2 *(Irs-2)* which has been linked to loss of ARNT signalling in other tissues (p = 0.036). Although steatosis with lobular inflammation was evident histologically there was no significant difference in serum aspartate transaminase (AST) or [alanine transaminase](http://en.wikipedia.org/wiki/Alanine_transaminase) (ALT), with wide variability between mice. Macrophages were isolated from mice using thioglycollate. Similar numbers were isolated from FC and LAR mice. LAR macrophages had increased basal expression of *Ccl2*, *Cxcl10*, *Il6*, *Mcp1* and a trend to increased *Tnfα* (p = 0.057). Four hours after treatment with LPS, LAR macrophages had significantly greater expression of *Cxcl10*, *Mcp1* and *Tnfα* **.** Next, we examined changes in gene expression in livers from LAR and FC mice using PCR arrays. Significantly altered genes are named on the figure. Together these changes in gene expression would contribute to impaired metabolism of lipid, as seen in the livers of the LAR mice. Protein levels of some of the genes were assessed by Western immunoblotting. Expression of the smaller (46kDa) JNK isoform was decreased, giving an overall decrease in total JNK in LAR livers (p\<0.05). Despite significant changes on the PCR arrays, FABP1 and Adiponectin receptor 2 (AdipoR2) protein levels did not differ between groups (**)**. Disrupted mitochondrial function is proposed to play a role in development of steatohepatitis, so the mitochondrial complexes were assessed as shown in **.** There was a significant decrease in protein levels for Complexes II, III and V (SDHB, UQCRC2 and ATP5A respectively). Taken together, these results suggest that the increased TG content in LAR livers compared to FC livers may result from impaired fatty acid oxidation. ## LAR mice and glucose homeostasis Prior to HFD feeding, LAR and FC mice had equivalent weight (**)**. As alterations in macrophage function have been shown to influence whole-body glucose metabolism, and diabetes increases risk of NAFLD / NASH, we assessed effects on glucose tolerance. Glucose tolerance tests (GTTs) were slightly worse in LAR mice at baseline, achieving significance at the 60 minute time point (p = 0.039). Following 5 weeks of HFD feeding LAR mice had significantly worse glucose tolerance than\\ FC littermates. Glucose intolerance did not deteriorate further in floxed controls by 20 weeks and improved slightly in LAR mice, and at 20 weeks the two groups had equivalent glucose tolerance. The early differences in glucose tolerance were not related to differences in whole-body insulin sensitivity as assessed by insulin tolerance testing after 11 weeks of HFD. Male LAR mice tended to have lower blood glucose concentrations at 60 minutes after insulin injection (p = 0.095). Similar results were obtained after 20 weeks (data not shown). LAR and FC mice underwent a pyruvate tolerance test to assess endogenous glucose production. There was only a small difference in PTT results between FC and LAR mice with LAR mice having slightly but significantly lower fasting glucose ****. The incremental area under the curve did not differ significantly between groups (p = 0.43). ## Liver ARNT expression was decreased in humans with NASH Finally, to examine whether ARNT might play a role in human NASH, *ARNT* mRNA was measured in liver samples from people with normal liver or NASH. *ARNT* was \~80% lower in the livers of individuals with NASH than in the livers of healthy subjects. # Conclusions In this study, the response of mice with myeloid cell ARNT deletion to HFD challenge was investigated. The key findings were a more than doubled prevalence of steatohepatitis (68% vs 30%), with increased lobular inflammation. This increased prevalence of steatohepatitis was accompanied by macrophage infiltration in the liver and expression of inflammatory cytokine mRNAs. No specific pro-fibrotic stimuli were given (e.g. thioacetamide, CCl4, or methionine choline deficient diet). The study used genetic controls (floxed control mice) and did not study diet-controls i.e. mice receiving normal chow. However, C57Bl/6 mice eating normal chow do not develop NAFLD. For that reason, in the ARNT-deleted mice it is not likely that there would be a strong phenotype on chow diet. Hydroxyproline was not measured to assess global liver collagen content, which is a study weakness. Myeloid ARNT deletion increased subcutaneous fat weight, liver triglyceride content and adipocyte size and worsened glucose tolerance. The alterations in whole body glucose / insulin homeostasis may result from changes in the liver, or adipose depots, or possibly skeletal muscle which was not studied in this work. It is plausible that there is also cross-talk between liver and fat to lead to NASH. A well-described example is adiponectin. In the liver of LAR mice, Adiponectin receptor 2 had lower expression, but this was not confirmed at the protein level. Human NASH is associated with infiltration of mononuclear phagocytes and increased liver expression of *CXCL1*, *MCP-1*, *TNF-α* and *TGF-β1*. In mice fed a methionine/choline-deficient diet likewise, Kupffer cells play an important role in the initiation of liver inflammation, while recruited macrophages are important for disease perpetuation. MCP-1 and TNF-α in particular have been shown to play important roles in the recruitment of inflammatory macrophages in NASH models, with antagonism or reduction of either associated with decreased monocyte recruitment and reduced inflammation. Consistently, in LAR mice, increased numbers of cells of the macrophage lineage (*F4/80* mRNA expression and F4/80+ cells) were observed while the combination of increased mRNA expression for *Mcp-1*, *Cxcl-1*, *Tnf-α* and *Tgf-b1* and recruitment of macrophages appear to have perpetuated inflammation. The finding of increased cytokine expression and inflammation in the liver of LAR mice after HFD is perhaps unexpected considering that mice lacking myeloid HIF-1α or HIF-2α have decreased inflammation in acute immune models. Consistent with those reports, we recently reported decreased skin inflammation and wound healing in LAR animals alongside reduced cytokine mRNA in macrophages. However, it has recently also been reported that mice lacking ARNT/ HIF-1α signalling in myeloid cells have increased allergic response in both a house dust mite model and an OVA murine asthma model, which may be driven by decreased IL-10 production. HIF activation through myeloid cell Vhl deletion reduced inflammation in a model of chronic kidney disease, and macrophage HIF-1α deletion increased the immune response to cancer through de-repression of infiltrating cytotoxic T-cell activity. These results demonstrate that in certain situations myeloid cell ARNT/ HIF-1α function can reduce inflammation, but that decreased myeloid cell expression of cytokines like IL-10 may contribute to increased inflammation and NASH as observed in LAR mice. In support of this, *Il-10* mRNA was increased in LAR livers after HFD (x 2.3 fold), although not to the same extent as *Mcp-1* (x 5.2 fold) and *Tnf-α* (x 3.2 fold). In support of a potential role in humans we have found ARNT mRNA in isolated human monocytes inversely correlates with serum cytokine levels of IL-6, IL-8, MCP-1 and TNF-α. There was decreased *ARNT* expression in human liver from people with NASH compared to normal liver biopsies. A limitation of the study is the lack of a fatty liver without steatohepatitis group, which we were not able to access. There are few good mouse models of NASH. Ideally, a model should recapitulate features of the human disease with fatty liver, inflammation, perisinusoidal fibrosis, and ideally metabolic abnormalities such as obesity and increased blood glucose. All of these features were observed in LAR mice fed 20 weeks of high fat diet. Excitingly, there was also a substantial decrease in expression of *ARNT* in livers of people with NASH suggesting potential human relevance of these findings. It is interesting to note that environmental toxins which activate the ARNT-partner aryl hydrocarbon receptor are associated with fatty liver disease. ARNT may down-regulate fatty liver disease induction by the aryl hydrocarbon receptor. The present results add ARNT to the list of myeloid cell perturbations which result in altered metabolic function. We note that fasting blood glucose levels were decreased at 20 weeks. This was accompanied by a trend to reduced glucose during PTT, which suggested the possibility of impaired hepatic glucose production in LAR mice. In patients with cirrhosis it has been shown that while gluconeogenesis is increased basally, glycogenolysis is decreased. In addition the liver has impaired gluconeogenesis in response to gluconeogenic substrates and glucagon. It is possible therefore that liver dysfunction in LAR livers contributed to the decreased fasting glucose. The proposed mechanisms are shown in. Lack of ARNT in macrophages leads to increased macrophage recruitment to the liver. In association with this there are changes in gene expression, including cytokines, and the proteins for mitochondrial class II, III and V are decreased. The results of this research show that deletion of ARNT in myeloid cells causes liver inflammation and steatohepatitis after high-fat diet. We also found mice lacking myeloid ARNT displayed alterations in metabolism and in fat deposition. ARNT is decreased in livers from people with type 2 diabetes, and this may contribute to their increased risk of NASH. These results suggest that myeloid cell ARNT may be a therapeutic target to reduce liver injury in patients with NAFLD and NASH. ARNT Aryl hydrocarbon receptor nuclear translocator GTT glucose tolerance test HFD high fat diet HIFs hypoxia-inducible factors HIF-1α Hypoxia Inducible Factor 1α ITT insulin tolerance test NAFLD Non-alcoholic fatty liver disease NASH non-alcoholic steatohepatitis PTT pyruvate tolerance test TG triacylglycerides T2DM = type 2 diabetes mellitus VHL von Hippel Lindau [^1]: The authors have no competing interests.

# Introduction For three decades, retinoic acid (RA) differentiation therapy has been tantamount to transforming acute promyelocytic leukemia (APL) from a fatal diagnosis into a manageable disease. RA induces remission in 80–90% of APL PML- RARα-positive patients. However, remission is not durable and relapsed cases exhibit emergent RA resistance. Meanwhile similar success stories have yet to be achieved for other cancer cell types. Parallel to the clinical use of RA in APL treatment, intense research has focused on understanding the source of cancer treatment relapse, and exploring the effectiveness of RA in other cancers. Historically RA resistance in APL has been associated with mutation(s) in the PML-RARα fusion protein, rendering it unresponsive to RA. However, in some APL patients, PML-RARα mutations emerge months after termination of RA therapy, suggesting the existence of other defects. In the patient-derived APL cell line NB4, RA resistance may or may not be correlated with mutant PML-RARα. RA- resistant NB4 cells often remain partially RA-responsive in that they can upregulate RA-inducible differentiation markers, such as CD38 or CD18. HL-60, another patient-derived leukemia cell line, does not harbor the t(15;17) translocation pathognomonic for APL and thus lacks PML-RARα, but is nevertheless RA-responsive. Like NB4 cells, *in vitro* maturation of HL-60 cells is consistent with that of primary APL cells in culture and with clinical RA differentiation therapy progression. Ectopic expression of RARα in RA-resistant HL-60 cells in which mutant RARα was found also does not necessarily restore RA responsiveness, again suggesting the presence of other defects. There is great interest in employing differentiation-promoting agents in combination with RA treatment to overcome resistance, and improve therapy and prognosis in APL and other cancer types. The active form of vitamin D₃, 1,25-dihydroxyvitamin D₃ (D₃), which acts through vitamin D receptor (VDR), is capable of inducing differentiation in myelo-monocytic precursor cells, but has been less widespread as a clinical treatment since D₃ also induces hypercalcemia and hyperphosphatemia. However, co-administration of RA with D₃ is a potential therapeutic strategy to mitigate the side effects and limitations of each individual inducer. Bipotent human acute myeloblastic leukemia (FAB M2) HL-60 cells can be induced to terminally differentiate *in vitro* along the granulocytic lineage toward neutrophil-like cells using RA, while differentiation along the monocytic lineage can be achieved with D₃. RA-treated HL-60 undergoing granulocytic differentiation display early increased surface expression of CD38, followed by CD11b expression. D₃-treated HL-60 cells undergoing monocytic differentiation express CD38, higher levels of CD11b, and the monocytic surface marker CD14. Induced terminal differentiation is accompanied by G1/G0 cell cycle arrest, and the development of inducible oxidative metabolism (respiratory burst), a function of mature granulocytes and monocytes. For the RA-treated case, differentiation requires sustained activation of mitogen-activated protein kinase (MAPK) signaling along the Raf/MEK/ERK axis, and a cascade of signaling regulatory events involving a putative signalosome containing c-Cbl, Vav1, and the Src-family kinases Lyn and Fgr. This is due in part to retinoic acid response elements (RAREs) in the promoter regions of CD38 and BLR1. Both of these proteins are rapidly upregulated by RA; CD38 is the nexus for the putative signalosome while BLR1 drives a prolonged MAPK signal though its relationship with c-Raf. However, D₃-induced differentiation also requires sustained MAPK signaling and results in upregulation of CD38 and CD38-associated factors. Onset of G1/G0 arrest and terminal differentiation is slow requiring approximately 48 h of treatment, during which HL-60 cells undergo two sequential, functionally discernible stages. With a doubling time of approximately 20–24 h, induced HL-60 cells first become primed for differentiation (precommitment phase) and undergo early differentiation events. During the subsequent 24 h, HL-60 complete a second cell division that results in terminally differentiating cells which are committed to a specific lineage determined by the inducer present, e.g. RA or vitamin D₃. Although lineage-specific events, such as CD14 expression, can in fact occur during the first 24–48 h of D₃ treatment in HL-60, the final inducer present is nonetheless the determining factor for lineage selection and subsequent terminal differentiation into that lineage. It has also been shown that HL-60 cells treated with RA for 24 h followed by washing and no retreatment results in a still-proliferating population that retains a “memory” for differentiation that lasts 4–6 cell divisions. During this time, cells proliferate until retreatment in which short RA doses can induce complete granulocytic differentiation. We previously isolated two emergent RA-resistant HL-60 cell lines after chronic RA exposure. These RA-resistant lines do not express CD11b, exhibit G1/G0 arrest, nor develop oxidative metabolism after RA treatment. One resistant line (R38+) retains RA-inducible CD38 expression while the other (R38-) has lost this ability. The R38- line, which sequentially emerged from R38+, thus appears to have an earlier defect which blocks the RA-induced differentiation sequence before the expression of CD38. Signaling events that define the wild-type response are compromised in both R38+ and R38-, which include RA-induced c-Raf expression and phosphorylation, c-Cbl and Vav1 expression, expression of Src- family kinases (SFKs) Lyn and Fgr and Y416 SFK phosphorylation. In this study we examined whether the RA resistance defect segregates with lineage specificity, or with early or late stages of induced differentiation. An early defect might compromise both lineages, whereas a late defect might only affect the granulocytic lineage. Here we report that an RA-resistant cell line that retains partial RA-responsiveness (R38+) is more amenable to D₃-induced differentiation, while the more resistant cell line (R38-) is only partially responsive to D₃. We conclude that the defect in RA response is not necessarily compensated for by D₃ treatment to enable myeloid differentiation, and the RA defect is apparently early and late, possibly reflecting dysfunctions in proper prolonged signaling during early and late stages. The signaling dysfunction notably involves reduced Fgr, c-Raf, and Vav1 expression. Overall, the results are of potential significance to the use of differentiation-inducing agents for overcoming RA resistance. # Results We examined if 1,25-dihydroxyvitamin D₃ (D₃) is able to upregulate differentiation markers in either R38+ or R38- retinoic acid (RA)-resistant HL-60 cells, and by exploiting the biphasic characteristics of HL-60 differentiation, whether D₃ could compensate for RA response dysfunction during the precommitment (first 24 h) or lineage-commitment stage. HL-60 cells were treated with one inducer (RA or D₃) for 24 h, then washed and retreated with either the same, different, or no inducer. The expectation was that the cells receiving the same differentiation agent will behave as if in continuous exposure, while the cells whose differentiation agent switched will reveal dependence of various differentiation markers on the precommitment vs. lineage-commitment stages. Cells receiving no retreatment will reveal what aspects of the differentiation program are precommitment-dependent and will provide a control for the retreated cases. For each treatment regimen, 1 µM RA or 0.5 µM D₃ was used as these doses were previously shown to yield comparable levels of differentiation in HL-60. ## R38- exhibit diminished CD38 expression in response to D₃ compared to R38+ and WT HL-60 We assessed CD38 expression at 24 h to probe resistance in early or precommitment events prior to retreatment with a second inducer, and hence prior to the lineage-commitment phase. At 24 h, both RA and D₃ induced CD38 expression in WT HL-60. Three RA and D₃ cases are shown since each served as the initial 24 h treatment for the subsequent timepoints. An average of 94% and 70% of the WT cells were CD38 positive for RA and D₃ respectively. At 24 h, R38+ and R38- respond to RA-treatment as similarly reported for 48 h, with R38+ expressing, and R38- failing to express, CD38 after RA treatment. CD38 expression in RA-treated R38+ was similar to that of RA- treated WT HL-60, while the CD38 expression level in RA-treated R38- was similar to untreated WT HL-60, as expected. D₃ induced CD38 expression in both resistant cell lines. In R38- the expression was about 50% of the expression in the WT cells, and in R38+ the expression was significantly (p\<0.05) higher than in WT. CD38 expression in D₃-treated R38+ cells is similar to CD38 expression in RA-treated WT HL-60, despite expression in RA- treated WT HL-60 significantly exceeding D₃-treated WT HL-60 (p\<0.0001). Thus R38- cells have precommitment resistance to D₃ but R38+ cells do not, and R38+ actually have higher CD38 expression than D₃-treated WT HL-60. 48 h and 72 h timepoints probed resistance in later events. 48 h post treatment (24 h in precommitment and 24 h in lineage-commitment stages), R38- cells treated with RA/D₃ or D₃/D₃ have comparable CD38 expression levels (38% vs. 40%). R38- cells treated with D₃/RA are 31% positive for CD38, indicating that in R38-, D₃ presented early or late could elicit CD38 expression at comparable levels, although it was short of that of WT cells. By 72 h, CD38 expression induced by D₃ in the resistant cells does not increase significantly. R38+ cells present two interesting behaviors. First, in the cells receiving RA in the precommitment phase and not receiving any differentiation agent in the lineage-commitment phase (RA/-), a decrease in CD38 expression is more abrupt in R38+ (75% at 48 h vs. 54% at 72 h) than in WT HL-60 (92% vs. 85%), showing an impaired ability of R38+ to maintain CD38 expression. Second, D₃/D₃ and D₃/- treatments in R38+ have higher levels of CD38 than WT HL-60 cells (p \<0.005 and p\<0.05 respectively). Overall, WT HL-60 cells behave as expected, with CD38 expression increasing over time during all treatment patterns, except for RA/- and D₃/-, in which CD38 expression decreases as the cells revert to a proliferating state. R38+ have more CD38 expression than WT when treated with D₃ first (but not with RA first), and show a more rapid decrease in CD38 expression than WT during RA/- and D₃/-. R38- have half the induced CD38 expression compared to WT during RA/D₃ and D₃/D₃, slightly less for D₃/RA, no CD38 expression during RA/RA or RA/-, and decreasing CD38 expression during D₃/-. Hence compared to R38+, R38- cells have a diminished response to D₃ in terms of CD38 expression, and they have a more pronounced early defect not apparent in R38+. ## WT, R38+ and R38- HL-60 cells comparatively display decreasing D₃-induced CD11b expression CD38 is the only marker significantly expressed during the precommitment stage (first 24 h). We next focused on 48 h and 72 h to probe resistance in late events. CD11b is an integrin component expressed in granulocytes and monocytes. RA does not induce CD11b expression in either RA-resistant HL-60 cell line. D₃ rescues CD11b expression in both R38+ and R38- cells when administered early or late, with R38+ being slightly more responsive. The effect of D₃ on CD11b expression appeared to be more potent if administrated during the lineage-commitment stage. Comparing RA/D₃ and D₃/RA at 48 h, p\<0.004 for WT HL-60, p = 0.03 for R38+ and p = 0.01 for R38-. By 72 h, the D₃/RA-treated R38+ and R38- cells have similar levels of CD11b to cells treated with D₃/-, indicating that retreatment with RA was comparable to no retreatment for both the resistant cell lines. Overall, WT HL-60 behaved as expected. WT HL-60 exhibited increasing (over time) CD11b expression for all treatment regimens save for RA/- and D₃/-, in which CD11b expression levels dropped by 72 h. Since D₃/RA treatment, compared to D₃/D₃, did not fully restore a WT-like response in the RA-resistant cells, the data suggest a late RA defect(s), which is putatively more pronounced in R38- than R38+. ## D₃-induced CD14 expression occurs in WT, R38+ and R38- HL-60 cells if administered during the lineage-commitment phase CD14 is a glycosylphosphatidylinositol-anchored membrane protein expressed by monocytes, but not by granulocytes. CD14 is a monocytic-specific marker for detecting a differentiation response to D₃ treatment, and here its expression reveals whether defects are lineage unrestricted or restricted (i.e. early or late). We expect WT HL-60 cells to exhibit CD14 expression only during D₃/D₃ and RA/D₃ treatments. By 72 h, all three cell lines treated with D₃/D₃ expressed CD14 at significantly higher levels than all other treatments, with R38+ expressing slightly higher levels (43% positive cells) than the WT HL-60 cells (33% positive cells). At 72 h there is no significant difference between RA/D₃ and D₃/D₃ treated cells for either R38+ or R38− individually, although R38+ cells had higher CD14 expression than R38− cells. This indicates that monocytic differentiation can potentially occur in the RA-resistant cells, and that monocytic differentiation can occur if the differentiation agent present during the lineage-commitment phase is D₃. The R38− response was weaker than the R38+ response, consistent with progressive attenuation of D₃ response as cells become more resistant. ## D₃ cannot rescue respiratory burst activity in either R38+ or R38− Respiratory burst (oxidative metabolism) is the ability of mature neutrophils and macrophages to respond to bacterial infections, and is considered a final functional marker of maturity. None of the treatment regimens were able to significantly rescue this late differentiation marker in the RA-resistant HL-60 (also see). Although D₃/D₃ treatment tended to increase the respiratory burst activity, this did not reach statistical significance. The WT HL-60 behaved as expected, exhibiting a strong respiratory burst in all treatment cases except for RA/- or D₃/-. ## R38- cells comparatively display the lowest level of D₃-induced G1/G0 cell cycle arrest We examined G1/G0 cell cycle arrest, which is also a relatively late attribute of induced differentiation. In WT HL-60, all treatments except RA/- and D₃/- induced a significant (p\<0.005) G1/G0 enrichment at 48 h. RA- treated RA-resistant cells do not exhibit G1/G0 arrest. For both R38+ and R38-, the only treatments that significantly (p\<0.05) increased the proportion of cells in G1/G0 were RA/D₃ and D₃/D_3. Differences between the individual responses of R38+ and R38- were not yet significant at 48 h. By 72 h, the WT cells treated with RA/RA, RA/D₃, D₃/D₃ or D₃/RA were significantly (p\<0.005) arrested in G1/G0_. Also at 72 h, R38+ cells were arrested by RA/D₃ and D₃/D₃ treatments, while the R38- resistant cells were arrested just by D₃/D₃ treatment (p = 0.03). Thus for the R38+ cells, D₃ had to be administrated at least in the lineage-commitment stage, whereas for the more severely resistant R38- cells, D₃ had to be administrated in both precommitment and lineage-commitment stages to obtain significant growth arrest by 72 h. This is consistent with a late differentiation dysfunction for R38+ and early and late dysfunction for R38-. ## D₃ rescues early (24 h) expression of Fgr, Vav1, and p47^phox in RA-resistant cells Significant signaling components that are upregulated during the first 48 h of RA treatment in HL-60 cells have been identified,. Knowing the behavior of this ensemble during RA-induced differentiation, we sought deviations from this in the resistant cells to gain insight into the potential molecular basis of the cellular phenotypic behavior above. We examined the CD38-associated proteins c-Cbl, Vav1, and Slp76; the Src-family kinases (SFKs) Lyn and Fgr, and the Y416 SFK phosphorylation site; c-Raf and its RA-induced phosphorylation sites S259, S621 and S289/296/301; p47^phox, one of many proteins related to oxidative metabolism; and aryl hydrocarbon receptor (AhR) which we reported drives differentiation. We first investigated proteins known to exhibit increased expression in WT HL-60 by 24 h. Fgr is upregulated by RA after 24 h in WT HL-60, but not in R38+ or R38- resistant cells. Fgr could still be upregulated by D₃ in R38+ cells at 24 h.. But R38− cells were slower and required 48 h before Fgr upregulation was discernible. This correlates with the putative more profound resistance of R38− cells. Vav1 is upregulated by RA and D₃ in WT HL-60, and in contrast to Fgr, D₃ results in higher Vav1 expression during the first 24 h compared to RA treatment. But in resistant cells, RA did not cause any appreciable Vav1 upregulation, whereas D₃ is able to increase Vav1 expression in both RA-resistant cell lines. p47^phox is comparably induced by RA and D₃ in WT HL-60 during the first 24 h. But in R38+ cells RA induces only a small increase in p47^phox and none detectable in R38−. In contrast D₃ treatment prominently increases expression of p47^phox in both RA-resistant lines. Slp76 is upregulated by both RA and D₃ in WT HL-60, but expression in resistant cells hardly changed with either RA or D₃ treatment at 24 h. Meanwhile c-Cbl does not exhibit much change by the end of the early 24 h timepoint. There are thus early defects in RA-induced upregulation of select signaling molecules such as Fgr, Vav1, and p47^phox, whose expression in the resistant cells is, to some degree, rescued by D₃. ## WT, R38+ and R38− HL-60 cells comparatively display decreasing D₃-induced expression and phosphorylation of differentiation-associated signaling factors at 48 h We examined signaling factor expression at 48 h to probe for resistance-related aberrations during the late, lineage-commitment phase. Representative blots for 48 h signaling data are shown in. To address slight variations across repeats, all repeats where quantified and average fold change ± S.E.M. is presented in. Also, regraphs the same data separated by cell line to further clarify the expression differences. At 48 h, WT HL-60 treated with RA/RA behaved as expected, with RA exposure resulting in increased Fgr, Lyn, Vav1, c-Cbl, Slp76, c-Raf, AhR and p47^phox expression compared to untreated control. However, RA-resistant cells (R38+ and R38−) treated with RA/RA displayed little to no upregulation of these signaling proteins (excluding Slp76, as found previously) compared to WT HL-60. For the RA/D₃ treatment case, WT HL-60 again show upregulated expression of these proteins. Interestingly, R38+ and R38− treated with RA/D₃ exhibited increased Vav1 and p47^phox expression compared to the RA/RA case, although expression in R38− was diminished relative to R38+. R38+ cells also exhibited c-Cbl expression with RA/D₃ treatment, but R38− tended not to. Lyn and c-Raf were minimally increased in the RA-resistant cells during RA/D₃ treatment. In the resistant cells, Fgr expression is only slightly increased during the RA/D₃ case, with the R38+ line exhibiting less Fgr expression compared to WT HL-60, and R38− even less compared to R38+. Overall this indicates that D₃ treatment, despite the previous RA treatment, was able to predominantly rescue expression of Vav1 and p47^phox signaling proteins in R38+ and less so in R38−, consistent with the more apparent resistance to D₃ of R38− compared to R38+. As most evidenced by Fgr, WT cells treated with RA/- (i.e. RA followed by no retreatment) displayed similar but diminished expression of signaling proteins compared to the RA/RA case, consistent with a need for continuous early and late exposure to drive expression. For D₃/D₃ treated cells, WT and R38+ lines displayed upregulated Fgr, Lyn, Vav1, c-Cbl, c-Raf and p47^phox expression compared to untreated controls, while R38− had notably diminished expression of these proteins, again consistent with greater D₃ response dysfunction in R38− compared to R38+. For the D₃/RA treatment pattern, WT HL-60 still show upregulation of these proteins. But for the RA-resistant R38+ and R38− cells, D₃ followed by RA treatment resulted in less Fgr, Vav1, c-Cbl, c-Raf and p47^phox expression compared to the D₃/D₃ case, consistent with a putative late defect in both resistant cell lines. During RA/RA treatment in WT HL-60, phosphorylation at c-Raf sites S259, S621, and S289/296/301 is increased. For the resistant cells, this response was largely lost. For RA/D₃ treatment, the pS259c-Raf and pS289/296/301c-Raf responses were recovered in R38+ cells, but not in R38− cells, consistent with their putative greater resistance. In both resistant cells, the pS621c-Raf response was lost in RA/D₃ treatments, consistent with the importance of this phosphorylation event in driving differentiation as suggested earlier. D₃/D₃ treatment induced phosphorylation at all c-Raf sites checked in WT and R38+ cells. Both R38+ and R38− retained S621c-Raf during D₃/D₃ treatment, but unlike R38+, phosphorylation at S259 and S289/296/301 sites on c-Raf was lost in R38− cells. Increased loss of induced c-Raf phosphorylation thus correlated with increased resistance. c-Raf phosphorylation was generally reduced in D₃/RA-treated RA-resistant cells compared to the D₃/D₃ case. Overall these c-Raf phosphorylation changes did not necessarily correlate with the changes in c-Raf expression level. In the different treatment regimens, RA and D₃, administered singly or in sequential combination, caused increased c-Raf expression in WT HL-60. The response in each instance was diminished in R38+ and even more diminished in R38−. The Y416 SFK (Src-family kinase) site also showed increased phosphorylation in RA-treated WT HL-60 cells, and this response was largely abrogated in both resistant cells. D₃ administered early or late tended to cause, albeit much smaller, an increase in Y416 SFK phosphorylation in the R38+ but not R38− cells. Taken together the above data motivate the notion that this ensemble of signaling molecules and events support differentiation and that progressive resistance is concomitant with their decreased expression and phosphorylation. Cluster analysis reveals that in WT HL-60 cells there is a tight coupling between the responses of the ensemble of signaling molecules for different treatment regimens, but the coupling is degraded as the cells become progressively more resistant. # Discussion ## Phenotypic Differences between WT, R38+ and R38− HL-60 cells To our knowledge, this is the first study that analyzes how RA resistance depends on early vs. late lineage-commitment events in lineage-bipotent myeloid cells and relates to lineage cross-resistance. Taking the cellular response data together, the responses of R38+ and R38− RA-resistant HL-60 cells to the combinatorial sequences of RA and D₃ treatment distills to two basic results. First, in both R38+ and R38− resistant cells, D₃/RA treatment does not restore RA response, while RA/D₃ could not fully restore D₃ response. Thus D₃ cannot necessarily abrogate the temporally segregated early or late RA defect(s). Second, as the resistance to RA became more pronounced with progression from R38+ to R38−, there was progressive emergence of worse D₃ response. That is, the response to D₃ administered early or late in combination with RA, or administered both early and late, was less effective in R38− than R38+ cells. Although both R38+ and R38− cells equally failed to develop a significant oxidative metabolism, D₃ treatment can nevertheless rescue expression of the respiratory burst-associated protein p47^phox in the resistant cells, with the greatest expression occurring during D₃/D₃ treatment and slightly lower during RA/D₃ treatment, and greater expression always occurring in R38+ compared to R38−. When treated with D₃ during the lineage-commitment phase, R38+ cells always exhibited higher CD14 expression than R38− cells. R38− cells consistently displayed lower CD38 and CD11b expression, lower differentiation-associated signaling factor expression and phosphorylation, and notably had lower G1/G0 cell cycle arrest compared to both WT and R38+ HL-60. Therefore we found that RA differentiation therapy resistance can develop in stages, with initial partial RA resistance and moderate D₃ responsiveness (unilineage maturation block), followed by subsequent pronounced RA resistance and partial D₃ resistance (bilineage maturation block). RA can inhibit monocyte/macrophage activity, and other differentiation programs can also be suppressed by a RAR-dependent process. In the case of WT HL-60 cells, although the precommitment stage can be induced by RA or D₃, the later stages of monopoiesis are inhibited by RA,. If enhancing the differentiation process toward one lineage may inhibit another, then it may be plausible that cells resistant to one induced lineage can respond more strongly to another induced lineage (i.e., the “repressive” pathway is removed). This could be one explanation for why early D₃ treatment induced a slightly stronger response in the R38+ RA-resistant cells than the WT cells in terms of CD38 and CD14 expression. We performed hierarchical clustering analysis between the cell lines across all treatments and results, and interestingly found that WT and R38+ clustered more closely than R38−. Agglomerative hierarchical clustering analysis across all cell lines and treatments vs. signaling components is diagramed in. The treatment cluster family for WT HL-60 separates into two clusters: those treated with RA first and those treated with D₃ first. The untreated control samples exist in a cluster with R38− RA/RA and R38− RA/-. This is consistent with the notion that R38− is the most resistant cell line and consequentially the least dissimilar from untreated WT cells. Allowing that R38− RA/- represents the least responsive case, then the cluster analysis reveals a progression of cases that become more distal to and deviate from the most unresponsive case, namely R38− RA/−, R38− RA/RA, R38+ RA/−, R38+ RA/RA, and finally the RA/D₃ cases for both resistant lines. This clustering conforms to the anticipation that R38− are less responsive than R38+, and that RA is generally less effective than D₃ in eliciting response in the resistant cells. The cases for early D₃-treated resistant cells group together further away in the clustering analysis, consistent with weaker resistance to D₃ compared to RA posited earlier. When comparing both the signaling results and the cellular phenotypic results, hierarchical clustering across all treatments for WT, R38+ , and R38− reveals the increasing distances (lower correlations) as cells become more resistant compared to the WT HL-60 cells. A progressive uncoupling of the signaling molecules thus occurs as WT HL-60 change to R38+ and then to R38−. Thus the repertoire of signaling proteins surveyed may have a seminal role in effecting differentiation. Progressive degradation of the clustering of an ensemble of putative signalosome molecules as resistance increases supports the importance of an intimate co-regulated clustering of those molecules to drive differentiation. We investigated RARα and VDR protein levels at 24 and 48 h and were unable to attribute decreasing resistance to loss of either receptor. ## Vav1, Fgr and c-Raf emerge as prominent differentiation-associated factors A potential suite of molecular dysfunctions is seminal to the progression of observed resistance phenotypes. Vav1 is required for RA-induced granulocytic differentiation as well as TPA-induced monocytic differentiation of HL-60. Vav1, along with c-Cbl and Slp76, exhibit increased expression and exist in a CD38-associated complex during RA-induced differentiation of WT HL-60. These signaling factors are also upregulated along with CD38 during D₃ treatment in WT HL-60, as well as in RA-resistant HL-60. A cohort of molecules known to interact with CD38 is evidently expressed along with CD38 during either monocytic or granulocytic differentiation. D₃ induced Vav1 expression in R38+ and R38− during the first 24 h. If the two RA-resistant lines were retreated with D₃, then Vav1 expression persisted. However if the second treatment was RA, Vav1 expression tended to diminish by 48 h. A similar outcome occurs for c-Cbl, and p47^phox. Thus, although ectopic overexpression of Vav1 or c-Cbl can enhance RA-induced differentiation in WT HL-60, early-induced expression of these signaling factors in resistant cells is not enough to propel RA-induced differentiation during the lineage-commitment stage, which may reflect the co- existence of other potential defects. The data suggest that a late Vav1-dependent function may be disrupted in resistance, and lesser Vav1 expression in R38− compared to R38+ cells may contribute to the increased D₃ resistance in R38− cells. We have previously reported that the Src-family kinases (SFKs) Lyn and Fgr are upregulated with RA treatment in WT HL-60 cells. The D₃-induced upregulation of Lyn and Fgr has been noted by us and others. Lyn and Fgr are the predominant SFKs expressed in myeloid cells. However of the two, only Lyn appears to be the predominantly active (phosphorylated) kinase in RA-induced HL-60, as well as in RA-treated NB4 cells. Lyn and Fgr have been found to exert their functional roles in distinct subcellular compartments. When the aryl hydrocarbon receptor (AhR) ligand 6-formylindolo(3,2-b)carbazole (FICZ) enhances Lyn and Fgr expression, as well as Vav1, c-Cbl, and p47^phox expression, it also enhances RA-induced differentiation in WT HL-60 cells. In R38+ and R38− RA-resistant HL-60 cells, Fgr expression was not induced by RA at 24 h, as expected, yet was only minimally rescued either early or late by D₃ compared to WT HL-60 cells. Thus Fgr may be a signaling component important to the non-resistant phenotype, and dysfunctional early RA regulation of Fgr emerges as a prominent feature of resistance that correlates with loss of cellular phenotypic response. Phosphorylation at the Y416 SFK site seems to be primarily an RA-driven event in the WT HL-60 cells, as the highest pY416 SFK phosphorylation occurs during RA/RA, RA/D₃ and D₃/RA treatments. In contrast, results for Lyn and AhR were not as striking. Overall, there appeared to be higher Lyn expression in WT HL-60 cells across all treatment patterns. But upregulation by RA/RA, RA/D₃, D₃/D₃ or D₃/RA for all cells lines remained similar; hence it is not apparent whether Lyn expression is specific to any inducing agent or phase of differentiation. Similar results were obtained for AhR, with the exception of RA/RA treated WT HL-60, which had the highest AhR expression among all treatment cases and cell lines. c-Raf phosphorylation appears disrupted in resistance. Phosphorylation at the putative inhibitory site S259, the stability site S621, and the functionally ambiguous S289/296/301 site has been found to be induced by RA. Here we show that S259 c-Raf phosphorylation may be an early (but not late) RA driven event. Enhanced pS259c-Raf is observed in WT cells during RA/RA, RA/D₃, and RA/− treatment, but not during D₃/RA treatment (despite higher phosphorylation for the D₃/D₃ case). Also p259c-Raf is increased in the R38+ HL-60 during RA/D₃ and D₃/D₃, but not D₃/RA, treatment. Thus for both the WT and R38+ cells, pS259 was higher during RA/D₃ than D₃/RA treatment, consistent with being an early RA-driven event. Interestingly, the phosphorylation site S289/296/301 was significantly increased (as high as in WT cells) in R38+ during RA/D₃ treatment, but when D₃ was used first, R38+ had less pS289/296/301c-Raf than WT. Overall for c-Raf expression and phosphorylation, the R38− cells tended not to show as great a response to D₃ compared to WT or R38+, whether D₃ was treated first or last. This is again indicative of the greater degree of disrupted c-Raf-dependent signaling in these cells. Consistent with this, c-Raf expression was similarly progressively less during all treatments across WT, R38+ and R38− HL-60 cells. Like Vav1 and Fgr, c-Raf emerges as putatively a key component of the non-resistant phenotype. p47^phox and c-Cbl expression may be correlated with CD38 and/or CD14, since these two signaling factors were also more highly expressed in early D₃-treated R38+ cells compared to WT HL-60. But Fgr, Vav1 and c-Raf showed decreasing (across WT, R38+, R38−) induced expression for all treatments, similar to CD11b expression and G1/G0 arrest, notably implicating their dysfunction in progressive resistance. ## D₃ treatments and RA resistance in other studies Retinoic acid (RA) and the active form of vitamin D₃, 1,25-dihydroxyvitamin D₃ (D₃), are dietary factors that demonstrate chemotherapeutic efficacy in inducing maturation in leukemia cells. RA is the current treatment for acute promyelocytic leukemia (APL), and retinoids serve preventative and therapeutic roles in other cancers and diseases. D₃ is able to exert anti-proliferative effects in other myeloid cells and other cancer cell types. It has been shown that analogs of D₃ can induce differentiation of myeloid cells with minimal calcium toxicity. Like D₃, D₃ analogs have shown efficacy in inducing differentiation not only in myeloid lines, but in prostate and breast cancer cells. Co-administration of RA with D₃ or analogs thereof is a potential therapeutic strategy to mitigate the side effects of each individual inducer (RA syndrome, hypercalcemia, RA or D₃ resistance). One group found that RA and analogs of D₃ can act synergistically in WT HL-60 to promote differentiation and inhibit cell growth, and that a RA- resistant HL-60 cell line is more sensitive to D₃ treatment than the parental WT cells. We previously developed an HL-60 cell line resistant to sodium butyrate (a monocytic inducer) that was also cross-resistant to RA; however, this line remained responsive to monocytic differentiation by D₃. Interestingly, D₃ was shown to induce granulocytic (not monocytic) differentiation in a RA-resistant APL cell line. In another case, D₃ treatment did not induce differentiation in RA-resistant HL-60. RA resistance in HL-60 has been historically attributed to mutation of RARα. However, in some RA-resistant myeloid lines where a mutation in RARα was found, expression of wild-type RARα did not fully restore RA responsiveness. It is clear that other defects arise, which most likely vary across resistant sublines developed in different laboratories. This may account for the varying reports seen in the literature regarding a response (or lack of response) of RA- resistant cells to D₃ treatment. In fact, in a single study one group developed two RA-resistant HL-60 cell lines, one of which was D₃-responsive and harbored a RARα mutation, while the other was D₃-resistant and had intact RARα. # Conclusions The present study shows that induced signaling and phenotypic conversion progressively degrade in discernible stages of resistance. We showed that D₃ cannot necessarily abrogate temporally segregated early or late RA-resistance defect(s). Nonetheless, D₃ can induce extensive, but not complete, functional monocytic differentiation in the RA-resistant cells compared to WT HL-60. Therefore although the segregation of unilineage vs. bilineage resistance is not exact, we show that an RA-resistant cell line that retains partial RA-responsiveness (R38+) is more amenable to D₃-induced differentiation, while a sequentially emergent cell line more resistant to RA (R38−) is less responsive to D₃. We found that the emergent R38− RA-resistant HL-60 cell line was more dissimilar from WT and R38+ than WT and R38+ where from each other. An ensemble of signaling molecules that are co-regulated in WT HL-60 become progressively more uncoupled as resistance becomes more pronounced, a trend involving increasing loss of response to RA and then D₃. There was a putative early Fgr expression dysfunction and a late Vav1-dependent dysfunction correlated with progressive resistance, as well as dysfunctional c-Raf expression. HL-60 are negative for the t(15;17) mutation, making RA-induced mechanisms in these cells potentially applicable to other cancers. Overall RA resistance may thus result from dysfunction of multiple pathways, rather than single genetic defects. # Materials and Methods ## Cell culture and treatments HL-60 human myeloblastic leukemia cells, derived from the original patient isolates, were a generous gift of Dr. Robert Gallagher, and were maintained in this laboratory and published previously (– and others). HL-60 wild-type (WT), and the two RA-resistant HL-60 (R38+ and R38−) cells subsequently isolated in our laboratory were grown in RPMI 1640 supplemented with 5% fetal bovine serum (both: Invitrogen, Carlsbad, CA) and 1x antibiotic/antimycotic (Sigma, St. Louis, MO) in a 5% CO₂ humidified atmosphere at 37°C. Cells were cultured in constant exponential growth as previously described. Viability was monitored by 0.2% trypan blue (Invitrogen, Calsbad, CA) exclusion and routinely exceeded 95%. Experimental cultures were initiated at a density of 0.2×10⁶ cells/ml. There were seven treatment regimens studied: (1) untreated, (2) RA/RA, (3) RA/D₃, (4) RA/−, (5) D₃/D₃, (6) D₃/RA, and (7) D₃,/−. The first agent, RA or D₃, was added for the first 24 h (precommitment phase) followed by wash and retreatment with either the same, different, or no inducing agent (−) for the second 24 h (lineage-commitment phase) and beyond, for a total of 48 and 72 h. After 24 h of initial treatment, cultures underwent two washes of 10 min each in 15 ml of RPMI 1640 supplemented with 5% fetal bovine serum and 1× antibiotic/antimycotic before resuspension in fresh complete media and retreatment. The reported results indicate the total timepoint, encompassing both prewash and postwash treatments. All-*trans* retinoic acid (RA) (Sigma, St. Louis, MO) was added from a 5 mM stock solution in 100% ethanol to a final concentration of 1 µM in culture. 1,25-dihydroxyvitamin D₃ (D₃) (Cayman, Ann Arbor MI) was added from a 1 mM stock solution in 100% ethanol to a final concentration of 0.5 µM in culture. All other reagents were purchased from Sigma (St Louis, MO) unless otherwise indicated. ## CD38, CD11b and CD14 quantification 1×10⁶ cells were collected from cultures and centrifuged at 700 rpm for 5 min. Cell pellets were resuspended in 200 µl 37°C PBS containing 2.5 µl of either APC-conjugated CD11b antibody, PE-conjugated CD38 antibody, or PE- conjugated CD14 antibody (all from BD Biosciences, San Jose, CA). Following 1 h incubation at 37°C, cell surface expression levels were analyzed with a BD LSRII flow cytometer (BD Biosciences, San Jose, CA). APC fluorescence (excitation at 633 nm) was collected with a 660/20 band pass filter and PE fluorescence (excitation at 488 nm) was collected with a 576/26 band pass filter. Undifferentiated control cells were used to determine the fluorescence intensity of cells negative for the respective surface antigen. The gate to determine percent increase of expression was set to exclude 95% of the control population. ## Respiratory burst quantification 1×10⁶ cells were collected and centrifuged at 700 rpm for 5 min. Cell pellets were resuspended in 500 µl 37°C PBS containing 5 µM 5-(and-6)-chloromethyl-2′,7′-dichlorodihydro–fluorescein diacetate acetyl ester (H₂-DCF, Molecular Probes, Eugene, OR) and 0.2 µg/ml 12-o-tetradecanoylphorbol-13-acetate (TPA, Sigma, St. Louis, MO). Both, H₂-DCF and TPA stock solutions were made in DMSO at concentrations of 0.2 mg/ml and 5 mM, respectively. A control group incubated in H₂-DCF and DMSO without TPA was included. Cells were incubated for 20 min at 37°C prior to analysis by flow cytometry. Oxidized DCF was excited by a 488 nm laser and emission collected with a 530/30 nm band pass filter. The shift in fluorescence intensity in response to TPA was used to determine the percent cells with the capability to generate inducible oxidative metabolites. Gates to determine percent positive cells were set to exclude 95% of control cells not stimulated with TPA. ## Cell cycle quantification 1×10⁶ cells were collected by centrifugation at 700 rpm for 5 min and resuspended in 200 µl of cold propidium iodide (PI) hypotonic staining solution containing 50 µg/ml propidium iodine, 1 µl/ml Triton X-100, and 1 mg/ml sodium citrate. Cells were incubated at room temperature for 1 h and analyzed by flow cytometry using 488-nm excitation and collected with a 575/26 band-pass filter. Doublets were identified by a PI signal width versus area plot and excluded from the analysis. ## Protein detection by Western Blotting 2×10⁷ cells were lysed using 350–400 µL lysis buffer (Pierce, Rockford, IL) supplemented with protease and phosphatase inhibitors (Sigma, St. Louis, MO), and lysates were cleared by centrifugation at 13,000 rpm for 30 min at 4°C. Equal amounts of total protein lysates (15 µg) were resolved by SDS- PAGE, transferred onto PVDF membranes and probed with antibodies. c-Cbl (C-15) antibody was from Santa Cruz Biotechnology (Santa Cruz, CA). pS621c-Raf antibody was from Pierce Thermo Scientific (Lafayette, CO). Lyn, Fgr, pY416-SFK, AhR, Vav1, Slp76, p47^phox, c-Raf, pS259c-Raf, pS289/296/301c-Raf, VDR, RARα, GAPDH, horseradish peroxidase anti-mouse and horseradish peroxidase anti- rabbit were from Cell Signaling (Danvers, MA, USA). Enhanced chemiluminescence ECL reagent (GE Healthcare, Pittsburg, PA) was used for detection. ## Statistical analysis Treatment group means were compared using the Paired-Samples T-Test. The data represent the means of three repeats ± S.E.M. A p-value of \<0.05 was considered significant (using GraphPad software and Excel). For agglomerative hierarchical clustering of signaling data, average quantified Western blot data was clustered using Cluster 3.0 and visualized with TreeView. To assess the correlation of the expression patterns of both the phenotypic markers and signaling molecules, a hierarchical cluster analysis was conducted by single linkage method (nearest neighbor) and marker similarity metrics based on the Pearson correlation using SYSTAT 8.0 software. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HAJ RPB. Performed the experiments: HAJ RPB CNI RM. Analyzed the data: HAJ RPB CNI RM AY. Contributed reagents/materials/analysis tools: JDV AY. Wrote the paper: HAJ RPB JDV AY.

# Introduction Asthma is characterized by airways with inflammatory infiltrates, hyper- responsiveness, remodeling, and limited airflow, i.e. shortness of breath. Asthma can range from mild to life-threatening. Almost 10% of children in industrialized countries develop asthma. This places a disproportionate burden on health care systems, with incremental costs estimated at \$56 billion in the US in 2007. The domestic cat is the only animal species that spontaneously develops a syndrome of asthma which replicates the hallmark features of the human disease. Feline asthma is itself a major veterinary concern that affects between 1 and 5% of the 74 million pet cats in the U.S. The similarities between human and feline allergic asthma led to the establishment and characterization of a feline experimental model of allergic asthma that mimics the hallmark clinicopathologic features in humans. Allergic asthma is experimentally induced in cats using the clinically relevant Bermuda grass allergen (BGA) as aeroallergen to induce airway eosinophilia, airway hyperresponsiveness and histologic evidence of airway remodeling. Additionally, cats develop spontaneous clinical signs after allergen challenge (cough, wheeze and/or labored breathing on exhalation), BGA- specific IgE and a T helper 2 cytokine profile. This model is robust and has been used to investigate a variety of relevant clinical therapies including inhibitors of neurogenic inflammation, antihistaminics / antiserotonergics, enantiomers of albuterol, allergen-specific immunotherapy, small molecule inhibitors, and adipose-derived mesenchymal stem cells, among others. Since eosinophilic airway inflammation drives airway hyperresponsiveness and remodeling, it is the primary outcome measure evaluated in this model. Aeroallergen-induced asthma is utilized herein to develop a noninvasive diagnostic approach to feline asthma. Current clinical diagnostic methods are imprecise, low in sensitivity, and frustrated by the overlapping symptoms of other lower airway disorders. Consequently, there has been a clear need for accurate and objective early detection of asthma, as well as for means of ongoing monitoring of asthma and its management. The clearest indication of allergic asthma in cats, together with clinical signs, is ≥ 17% eosinophils in the cytology of bronchoalveolar lavage fluid (BALF). Several barriers have been encountered in the development of disease-specific biomarkers for asthma diagnosis: Standard clinical tests for asthma such as the cytology of BALF are accurate but invasive and impart some risk, making the approach unsuitable for serial sample collection from children and pet cats. NMR spectra of urine samples of children, interpreted using supervised statistics, showed high accuracy in separating stable, chronic asthma from health or from acute asthma. In serum and urine, no abnormalities have thus far been found to be definitively diagnostic for asthma in cats however. Although concentrations of metabolites in exhaled breath condensate (EBC) samples are much lower and harder to detect than in serum or urine, EBC collection is noninvasive and directly samples airway fluids and the lung microenvironment. Metabolomics of body fluids, including EBC, often relies on measurements by mass spectrometry or NMR spectroscopy, as they are sensitive enough to detect unique spectral “fingerprints” of a wide variety of metabolites. NMR has the advantages of quick and quantitative measurement of body fluids, good resolution of spectral peaks, and no manipulation or destruction of samples. NMR spectra of EBC interpreted statistically attained a high degree of discrimination of healthy from asthmatic children. NMR assay of EBC from 79 asthma patients discriminated those with neutrophil-rich sputum or using inhaled corticosteroids with high accuracies of 79% and 85%, respectively. This relied upon use of multiple regions of the NMR spectra in the statistical analyses. Even higher discrimination of health from the mild asthma of a smaller number of patients was obtained by interpreting NMR spectra of EBC using supervised statistics, regardless of the EBC being collected at -5 or -27°C. Gas chromatography–mass spectrometry (GC-MS) detection of volatile organic compounds in EBC with statistics classified asthma with an accuracy of 96%. Statistical models from the same study successfully classified the patients with well-controlled asthma or eosinophilia or neutrophilia in sputum. Our purposes described below have been to investigate procedures, performance, and possibilities for analyzing NMR-based “breathprints” for noninvasive “breathomics” of asthma, using the feline model of allergic asthma. Metabolomics studies have often proceeded by untargeted or targeted approaches. The former divides each NMR spectrum into segments called bins or buckets for investigation of the *proportions* that the bins contribute to the total of the statistical variances across the measurements compared. The more recent targeted approach estimates *absolute concentrations* by comparison with a library of reference spectra of the individual metabolites found in the body fluid. Both approaches proved highly effective in discriminating mild asthma from health from NMR spectra of EBC analyzed with supervised statistics. The concentration- dependent targeted strategy proved to be an enhancement with tighter clustering of the patient groups. However, the composition of EBC has been unclear. No libraries of spectra of NMR-observable compounds in EBC have appeared at this writing for estimating their concentrations, as evident from the latest software for targeted profiling. Consequently, the established untargeted approach using spectral binning was adopted for the current study. The sparseness of EBC spectra and high resolution at 800 MHz moderate the concern of peak overlap in bins. Choice of preprocessing steps is crucial and heavily influences the perceived importance of metabolites. Normalization of spectra by their integral addresses the highly variable concentrations of specimens of urine and EBC, but can be skewed by strong signals to introduce artifacts. This systematic error can be overcome by calculating the most probable dilution factor using probabilistic quotient normalization (PQN). After suitable preprocessing, multivariate statistics are applied to find patterns associated with biological status. The popular unsupervised method of principal component analysis (PCA) simplifies or projects the many measured features (variables) in spectra down to fewer uncorrelated principal components (PCs) that suggest trends shared among measured variables in the spectra. Variations on partial least squares (PLS) are popular to predict the class or phenotype of the sample, after an initial phase of training the statistical model. PLS uses multiple linear regression to derive from the data matrix **X** (here, the NMR spectra) the significant components related to the categories or phenotype **Y** (here, the presence or absence of allergic asthma in the feline patients). This makes use of “latent variables” not only to represent **X**, which PCA also represents, but also to correlate **X** with **Y**. A function called discriminant analysis (DA) is typically applied to clarify separation between categories. However, PLS-DA can be hampered by variation unrelated to the categories of interest. In order to overcome large variations between subjects or patients, multi-level PLS-DA was proposed for sets of data where each subject has a before and after condition (“crossover design”), e.g., before and after asthma induction in this work. By separating the between-subject variation from within- subject variation, the within-subject variation accompanying disease or treatment can be analyzed by PLS-DA with potentially improved classification. As an alternative to improve PLS-DA classification performance, orthogonal signal correction (OSC) can be applied to the data to remove the contributions of variances unrelated (orthogonal) to the class **Y**, and thereby improve categorization. This work investigates the potential of untargeted metabolomics using NMR spectra of EBC, interpreted using multivariate statistics, to diagnose asthma noninvasively in cats. The study exploits the feline model of allergic asthma induced by Bermuda grass allergen by comparing the EBC of cats in a state of health with a subsequent state of allergic asthma early after induction (six weeks after initial exposure to antigen). The choice of preprocessing with PQN normalization and *glog* transformation proved pivotal. Classification of health or early asthma among 106 samples using statistically validated OSC-PLS-DA or multi-level PLS-DA appears superior, with excellent sensitivity and specificity. OSC-PLS-DA points out the trends of increase in acetone, increase in two metabolites of unclear identity, and decrease of phthalate in EBC from cats with allergic asthma. # Materials and Methods ## Animal Care All animals were domestic shorthair, purpose-bred cats. 27 were from a commercial vendor (Liberty Research, Inc, Waverly, NY). The remainder were bred from a high-responder asthmatic cat colony (Comparative Internal Medicine Laboratory, University of Missouri, Columbia MO). There were 39 males and 14 females. The care of the 53 cats in the study followed the NIH Guide for the Care and Use of Laboratory Animals. The University of Missouri Animal Care and Use Committee approved the study design (ACUC protocols \#6912 and 7891). A commercial mixture of kitten and adult maintenance dry kibble diet was fed *ad libitum*. The research cats were housed in groups in large runs, with a variety of enrichment toys (e.g, hanging hammocks, elevated platforms to climb, baby toys, toy mice, balls with bells, etc.). Additionally, the cats were socialized and monitored at least daily by members of the research team, in addition to being monitored by members of the Office of Laboratory Animal Care. None of the cats became ill or died during this study. All of the cats were adopted to private homes after the study. ## Induction of Allergic Asthma In order to induce allergic asthma, on day 0, cats \< 1 year of age were administered 12 μg of Bermuda grass allergen (BGA) in 10 mg of alum and 100 ng *Bordetella pertussis* toxin (both subcutaneously). On day 14, intranasal BGA (75 μg of BGA in 200 μL of phosphate-buffered saline) was given. On day 21, 12 μg of BGA in 10 mg alum was injected subcutaneously. On day 28, the formation of wheals on an intradermal test confirmed sensitization to BGA. For the next 2 weeks, intensive aerosol challenges were administered to all cats with BGA (500 μg of BGA in 4 mL of phosphate-buffered saline) over 10 min in awake and unrestrained cats in a sealed plastic chamber. A nebulizer (Acorn nebulizer, model 646, Devilbis Health Care, Somerset, PA) delivered an aerosol of the BGA solution at an air flow rate of 9.3 l /min. An air compressor (Easy Air 15, Precision Medical, Inc., Northampton, PA) supplied the compressed air flow at a pressure of 2.93 kg/cm². At week 6, samples were collected 24 h after challenge with BGA aerosol (see below). ## Collection of Body Fluids and Evaluation of Airway Eosinophilia In all cats, EBC and BALF were collected before (day 0) and after (week 6) sensitization to BGA. The EBC was collected non-invasively just prior to BALF collection by placing cats in a 25 L plexiglass chamber for 20 to 30 min according to a previously published design with modifications. BALF was collected in a blind fashion under anesthesia using an 8 French red rubber catheter passed through an endotracheal tube according to a previously described protocol. The percentage of eosinophils in each BALF modified Wright’s stained cytospin was determined by counting 200 nucleated cells, with an asthmatic phenotype defined as \>17% eosinophils. The EBC and the supernatant of centrifuged BALF remaining were promptly stored at -80°C until further analysis. NMR spectra of the EBC were measured for use in training and evaluating statistical approaches and biomarkers. ## Collection of NMR Spectra of EBC To maximize the sensitivity and resolution of NMR measurements of inherently dilute EBC specimens, we collected their spectra with a Bruker Avance III 800 MHz spectrometer with 5 mm TCI cryogenic probe. EBC samples were prepared with 7% D₂O and 20 μM trimethylsilyl propanoic acid as 0 ppm reference standard. 1D ¹H NMR spectra were collected using W5-WATERGATE water suppression supplemented by presaturation of the water resonance at minimal power. 32,768 transients were normally averaged. After data collection, the ¹H NMR free induction decays were zero filled to 32,768 points, apodized with 2 Hz exponential broadening to enhance sensitivity, and Fourier transformed into spectra spanning from -2 to 12 ppm, with correction of phases and baseline using Bruker Topspin 3.1. ## Spectral Preprocessing About 30,000 points from 0.02 to 10 pm were retained for analyses, while the region around the suppressed water peak from 4.5 to 5.38 ppm was omitted. Preprocessing steps were conducted with Topspin and Origin. To accommodate the large variability in overall concentrations of biomolecules condensed in EBC, the amplitudes of the NMR spectra were normalized to the mean spectrum by probabilistic quotient normalization (PQN). The first step of the PQN was unit normalization of the integral of each spectrum to 1.0. The median quotient was calculated from the quotients of all the spectral positions (variables) to those of the mean spectrum. Then all variables of each original spectrum were divided by the median quotient of that spectrum. The NMR spectra of EBC specimens were next binned into segments to accommodate slight inequities in NMR peak positions and line shapes as recommended. (EBC spectra are sparse enough to avoid the greater overlap within buckets of spectra of serum and urine. Collecting the spectra at 800 MHz resolved the overlap of peaks enough to improve the reliability of binning. Shortcomings of bins have been discussed). Each spectrum was divided into 455 bins each 0.02 ppm (16 Hz) wide and was then scaled with either Pareto scaling or *glog* transformation for comparison. Pareto scaling used the following expression where ${\widetilde{x}}_{mn}$ is the scaled peak height: $${\widetilde{x}}_{mn} = \frac{{x\prime}_{mn}}{\sqrt{s_{n}}},\, where\ s_{n} = \sqrt{\frac{\sum_{m = 1}^{M}\left( {{x\prime}_{mn} - {\acute{x}}_{n}} \right)^{2}}{m - 1}}$$ Scaling by *glog* transformation instead proved much better at increasing the weighting of the smaller variances. It uses the relationship: $${\widetilde{x}}_{mn} = {\ln\left( {x^{\prime}}_{mn} + \sqrt{\left. {x^{\prime}}_{mn}^{2} + \lambda \right)} \right.}$$ where λis the *glog* transformation parameter. λwas obtained by dividing the same sample into eight replicates (aliquots). Each replicate was handled independently for the NMR analysis. The spectra of the replicates were preprocessed identically to ensure that the variances from the replicates arise solely from “technical” variations. In order to avoid the scaling effect caused by the transformation, the modified Jacobian of the *glog* function proposed was applied. λwas then calibrated using the Maximum likelihood criterion and Nelder-Mead minimization algorithm in MATLAB, using the equation and script provided in ref. Data were mean-centered at the outset of multivariate statistical analysis. ## Multivariate Statistical Analyses and Validation PCA decomposes the spectra (data matrix **X**) into PCs, which are a linear combination of weighted variables observed in the spectra. The optimized weights generate a loading plot. Each point of the score plot represents a spectrum’s projection onto the PC. The PCs were computed using eigenvectors by singular value decomposition. PCA was performed using the MATLAB Statistics Toolbox. (The R-script princomp is also suitable; see <http://www.inside-r.org/r-doc/stats/princomp>). PLS generated components instead using multiple linear regression, not only to model the spectral data matrix **X**, but also to calculate the projection matrix that maximizes the covariance between **X** and the class matrix **Y** of “response variables” or asthmatic phenotypes. Discriminant analysis (DA) was then applied to seek separation between the categories of absence and presence of allergic asthma. PLS-DA was performed with MATLAB using libraries of H. Li available from <http://www.mathworks.com/matlabcentral/fileexchange/47767-libpls-1-95-zip>. (R script plsda is also suitable: <http://www.inside-r.org/packages/cran/mixOmics/docs/plsda>). The collection of EBC before and after allergic induction of asthma was exploited in order to evaluate the multilevel PLS-DA approach to this crossover design in the data. Multilevel PLS-DA was implemented in MATLAB with M-files developed at the Univ. of Amsterdam and available at: <http://www.bdagroup.nl/content/Downloads/software/software.php> Preprocessing by filtering out one OSC component to remove the largest variance of **X** orthogonal to **Y** (again, presence or absence of asthma) was evaluated as well. OSC-PLS-DA calculations used the R script at: <https://gist.github.com/dgrapov/5166570>. To avoid overfitting, leave-one-out cross-validation was performed to select the optimal number of the components for use in model prediction. The prediction power of the statistical modeling was evaluated and validated by intensive Monte Carlo cross-validation and permutation testing. ## Identification of Pertinent Metabolites and Implications For identifying dilute small molecules trapped in EBC, high sensitivity was obtained using the 800 MHZ NMR system with cryoprobe and overnight signal averaging of sensitivity-enhanced ¹³C heteronuclear single quantum coherence (HSQC). For broad bandwidth, the HSQC used chirp adiabatic inversion and refocusing pulses. Reference NMR spectra were found at the Human Metabolome Database (HMDB) and Madison-Qingdao Metabolomics Consortium Database (MMCD). The database and server named Complex Mixture Analysis by NMR (COLMAR) ¹³C-¹H HSQC was used in semi-automated comparisons of these databases’ reference spectra with the natural abundance ¹³C HSQC in order to identify multiple carboxylic acids detected. Statistical total correlation spectroscopy (STOCSY) was proposed for identifying biomarkers by calculating the correlation matrix of the NMR data sets. A STOCSY correlation map was prepared from the 106 NMR spectra once preprocessed, segmented into 0.02 ppm bins, and *glog*-transformed. A ¹H-¹H correlation matrix (455X455 points) was generated from these preprocessed spectra. Covariance with a correlation coefficient r ≥ 0.7 was considered as significant and plotted in the STOCSY correlation map. Variable Importance in Projection (VIP) from the OSC-PLS-DA and OSC-corrected multi-level analyses was calculated in order to evaluate the diagnostic significance of each spectral bin. Spectral bins with VIP value larger than 1.0 are considered to discriminate between groups. Potential pathways affected by induction of the early stage of allergic asthma were anticipated by submitting the biomarker candidates to the MetaboAnalyst 3.0 server at: <http://www.metaboanalyst.ca/faces/ModuleView.xhtml> # Results EBC and BALF specimens were collected from a research cohort of 53 cats when healthy, and six weeks after beginning the protocol of exposing them to Bermuda grass antigen in order to induce allergic asthma. The most trusted evidence of allergic asthma in cats is the elevation of eosinophils above 17% in BALF. Before sensitization, the eosinophils averaged 3.9 ± 3.0% among the 53 animals. By the end of the six week period of sensitization to the aeroallergen, each cat had developed allergic asthma according to eosinophil counts elevated to an average of 54.9 ± 23.9% and exceeding 17% in all 53 cats. ## Preprocessing of NMR Spectra of EBC for Statistical Comparisons In pursuit of metabolomics discrimination and biomarkers of early allergic asthma, 1D NMR spectra were acquired for the EBC specimens of each of the 53 before and after inducing the asthma. Overall concentrations of solutes in EBC samples varied over a 5-fold range. To remove the impact of this large variation in overall metabolite concentrations among the samples, the peak heights in the NMR spectra were normalized by PQN. Normalization is part of the spectral binning (untargeted) strategy and enables statistical comparison of the *proportions* that the NMR peaks represent in the samples (in contrast to *absolute concentrations* of selected metabolites). PQN suppresses artifacts introduced by normalizing to the integral. This improves the reliability of statistical comparisons of EBC specimens collected from patients with potential variations in exhaled volume, pulmonary gas exchange, and ambient humidity in the condensation chamber. The untargeted approach was also motivated by the biomarkers being unknown during the undertaking and development of the project. Since scaling methods influence statistical results, the effects of Pareto scaling and *glog* transformation were compared. Without scaling, the bins with highest intensities (which tend to have higher variances) dominate the total variances and biased the statistics towards the tallest NMR peaks. Pareto scaling improved the situation but still resulted in a large range of variance among the bins. The *glog* transformation gives greater weight to small variances that most likely result from small peak heights ( and Figs). Preprocessing with PQN and *glog* transformation proved essential to discriminating asthma from health by enabling many spectral regions besides those of concentrated lactate to be considered by multivariate statistical analyses. ## Principal Component Analysis of the NMR Spectra PCA was applied to the 106 preprocessed NMR spectra. The first two PCs account for about 38% of the total variance among the spectra. The loading plot of PC1 and PC2 identifies many spectral bins with large variances among the 106 spectra. Metabolites represented by some of these spectral bins, identified by efforts described below, are labeled on the loading plots. The 3D plot of scores using the first three PCs does not adequately separate the healthy and asthmatic states of the cats. This is due to unsupervised PCA placing high weights on the high variance of NMR peaks that are uncorrelated with the disease state of the cats. ## Partial Least Squares Improves Separation of Asthma from Health Supervised PLS-DA methods were applied in a quest for better separation of asthmatic and healthy states of the cats and to recognize diagnostic spectral features. Five PLS components optimized the performance and reliability of the separation of the cats when healthy or asthmatic, as judged by the decrease of the root mean square error of the prediction (RMSEP) and the increase of the Q² value from leave-one-out cross-validation. PLS-DA achieved much better separation than PCA in the score plot between cats before and after developing early allergic asthma, presumably due to increased weights on variables that distinguish the health of the cats. Multi-level PLS-DA was introduced to paired data sets to separate between- subject variation (e.g., due to genetic background) and within-subject variation (effect of “treatment”) in the data. The crossover design of our data set (before and after induction of asthma) allowed us to test if multi-level PLS-DA could improve performance. After removing the between-subject variation, the loading plot is better dispersed than that of PLS-DA. The score plot from multi- level PLS-DA better separates the groups, provided that the PQN type of normalization has been used. The spectra were preprocessed further using orthogonal signal correction (OSC) that removed two OSC components. This filtered out variances between spectra unrelated to separation between health and asthma in subsequent PLS-DA and multi-level PLS-DA calculations. Plots of RMSEP and Q² generated by leave-one-out cross-validation showed that OSC-PLS-DA and multi-level PLS-DA perform clearly better in predictions than PLS-DA. Multi-level OSC-PLS-DA performed the best of all by these criteria. OSC-PLS-DA and multi-level PLS-DA reduced the RMSEP and increased Q² more significantly than did PLS-DA when using the same number of components. The first three to four components are enough to optimize the performance OSC-PLS-DA, which is a simplifying advantage over PLS-DA. Two to three components suffice for multi-level OSC-PLS-DA to offer superior predictive power. The score plot from OSC-PLS-DA using the first three components shows group separation that is much improved over PCA and PLS-DA. In the 2D score plot form OSC-PLS-DA, around three spectra of EBC from asthma are not distinguished from health. Multi-level OSC-PLS-DA fully resolves the specimens from heath and asthma in the score plot, and needs only two components to do so. In order to test the prediction power of the model, two-thirds of the 106 samples were randomly selected as the training set, and the remaining third was used as the test set. Monte Carlo cross-validation suggests that the OSC-PLS-DA model with four components has very good predictive power with Q² = 0.70, contrasting the negative control of random permutation testing with Q² = -0.07. Cross-validation suggests even better predictive power from the multi-level OSC-PLS-DA model with three components where Q² = 0.84 and the randomly permuted negative control has Q² = -0.15. The prediction for the training set (a third of the specimens) using the OSC-PLS-DA model with four components has sensitivity of 94.1% and specificity of 94.1%. The multi-level OSC-PLS-DA separation of the specimens into healthy and asthmatic classes is annotated with the names of the cats on a 2D score plot in. The prediction from multi-level OSC-PLS-DA model using three components has sensitivity of 94.1% and specificity of 100%. ## NMR Identification of Metabolites in EBC from Cats Independently from the untargeted statistical process of discriminating asthma from heath, we sought to identify more of the metabolites in feline EBC using two strategies. The more informative strategy proved to be a ¹³C HSQC spectrum of a relatively concentrated EBC sample from a healthy cat. Compounds present in the natural abundance ¹³C HSQC spectrum were analyzed using the COLMAR ¹³C HSQC database and server that matches the spectrum against the MMCD and HMDB databases of NMR spectra of 555 compounds. Additional matching was done manually against reference NMR spectra available in HMDB and MMCD databases. These efforts identified the amino acids alanine, aspartic acid, glycine, isoleucine, leucine, phenylalanine, serine, threonine, tyrosine, and valine. The ¹³C HSQC identified additional metabolically significant carboxylic acids of acetate, lactate, and pyruvate. Also present are the dicarboxylic acids phthalate, hexanedioic acid (adipate) and either or both of heptanedioic acid (pimelate) and octanedioic acid (suberate), for which the NMR peaks overlap. Weak peaks for most but not all of the groups of sucrose or a related sugar are also observed. An anomeric proton doublet at 5.41 ppm in 1D spectra is consistent with the presence of a sucrose- like molecule. A weak peak matching trimethylamine is also observed in the ¹³C HSQC (not shown). Overall, carboxylic acids predominate in NMR spectra of feline EBC. The second strategy to gain more information to identify and confirm molecules was to ascertain co-variation among the 106 preprocessed 1D NMR spectra and to plot the co-variation as a correlation map known as STOCSY. Two or more co- varying NMR peaks lying away from the diagonal constitute a ‘spin system’ represented by spots along a row or column, which better define the identity of a molecule than a single peak can. Comparison of the positions of the correlations in the STOCSY against the database of TOCSY spectra at the COLMAR server identified niacinamide, benzoate, and another aromatic metabolite that probably contains a hydroxyphenyl moiety. The STOCSY also confirmed phthalate and tyrosine. The aliphatic region of the STOCSY and comparison of spectra of EBC from cats with the HMDB database identifies isopropanol, propionate, and butyrate. (The latter’s potential alternative assignment as methyl butanoate appears unlikely due to the absence from the STOCSY of a correlation to the methyl peak expected in the range of 3.35 to 3.65 ppm. Nor is there a correlation for isobutyrate.) The STOCSY confirms the assignments of lactate and hexanedioate based on the ¹³C HSQC. The butyrate is confirmed by TOCSY spectra and multiplet line shapes in 1D NMR spectra. The propionate and isopropanol are also confirmed by 1D spectra. A number of spin systems evident in the STOCSY failed to match the online databases, suggesting the absence from the databases of multiple compounds present in EBC. ## Metabolites and Pathways Perturbed by Allergic Asthma The VIP plots from OSC-PLS-DA and multi-level OSC-PLS-DA suggest the spectral features that best distinguish the EBC of cats in states of health and early allergic asthma. The trends in the six best markers in the VIP plots were manually inspected in the 106 PQN-normalized NMR spectra. The highest VIP score suggests acetone to be the best biomarker identified. Acetone in EBC is clearly elevated by early allergic asthma in 74% of the cats, in whom acetone was scarcely detectable during health. The unidentified metabolites giving rise to broad, overlapped peaks near 5.8 ppm are elevated by early allergic asthma in 70% of the cats. The unidentified, probably hydroxyphenyl-containing aromatic metabolite is increased by the experimental asthma in 62% of the cats. Isopropanol is increased in 51% of the cats when asthmatic, and niacinamide in 43% of them. Leucine levels are not affected by asthma in 73% of the cats, rendering it unreliable as a marker despite VIP scores \> 1. Decreased phthalate upon onset of experimental allergic asthma in 60% of the cats suggests its diagnostic value, despite its origins in environmental sources. The primary marker of acetone and secondary markers of isopropanol and niacinamide were submitted to metabolic pathway analysis by MetaboAnalyst 3.0. A *P*-value of 0.05 suggests that nicotinate and nicotinamide (niacinamide) metabolism could be perturbed in the asthmatic animals that exhale more niacinamide. NAD⁺, NADP⁺, nicotinamide ribotide, and nicotinamide riboside each are precursors to nicotinamide. The frequent elevation of acetone in early allergic asthma in cats implicates synthesis and degradation of ketone bodies with *P*-value of 0.007. Inclusion of isopropanol in the pathway analysis implies perturbation of propanoate metabolism by the experimental asthma, with the confidence of a *P*-value of 0.0006. Acetone accumulates in the majority of cats with allergic asthma presumably because of active synthesis of ketone bodies. This can be regarded as belonging to a larger set of pathways of propanoate metabolism. Isopropanol, which appears elevated in half of the cats with experimental asthma, is an immediate precursor to acetone. Acetoacetate is the better known precursor of acetone and lies on the pathway from acetyl-CoA, which is central in metabolism. # Discussion Novel noninvasive means for diagnosing asthma and monitoring its management present a significant need and opportunity in both pediatric care and veterinary care of pet cats. Earlier studies reported statistically promising diagnostic potential for human asthma by collecting EBC and assaying it by NMR for asthma in both children and adults, but did not identify biomarkers. The metabolomics results in the feline asthma model bolster these evidences for EBC being a noninvasively accessible fluid that is measurable by NMR for diagnostic value for asthma. Increases of acetone, a group of overlapped peaks near 5.8 ppm, and a hydroxyphenyl-like metabolite, as well as decreases of phthalate, have emerged as the best NMR-detectable markers in EBC of early allergic asthma in cats (Figs and). Increases of isopropanol and niacinamide in EBC are confirmatory of allergic asthma in part of the cats. These NMR-detected observations complement the volatile organics detected by GC-MS or electronic nose. ## Ketosis in Allergic Asthma Clinically relevant ketones in cats produced during states of decreased glucose utilization or negative energy balance include acetone, β-hydroxybutyrate, and acetoacetate. In 74% of the cats of this study, exhaled acetone was increased with induction of allergic asthma, implying increased ketogenesis. The acetone may have been produced locally by microbiota in the lung or emanated from the circulation of these cats. Acetone is one of the serum metabolites that was found to be related to the severity of eosinophilic asthma and airflow limitation in adult humans. In cats, increased serum β-hydroxybutyrate, is characteristic of diabetes, diabetic ketoacidosis, hepatic lipidosis, and less frequently other conditions such as chronic kidney disease and hyperthyroidism in which fat is used as an energy source. Ketonemia in cats is typically associated with a shift from glucose utilization to β-oxidation of fatty acids and negative energy balance. No evidence of negative energy balance was noted in the 53 cats of this study however. Similarly in humans, increases of acetone in serum accompanied decreases of % forced expiratory volume in 1 s (FEV₁%), an important manifestation of asthma. The lower FEV₁% was also correlated with much increased very low and low density lipoprotein (VLDL/LDL), which is strongly suggestive of altered lipid metabolism. The authors suggested the increases of acetone and histamine release result from the elevation of VLDL/LDL. This could be related to the increased exhalation of volatile organic compounds in asthma, which has been attributed to increased peroxidation of lipids. ## Other Metabolic Biomarkers of Asthma The decrease of glucose and increase of lactate in serum from adult humans with asthma, detected by NMR-based metabolomics, may be suggestive of hypoxia. Those changes could be consistent with the increase of exhaled acetone and apparent ketogenesis in asthmatic cats. At the early stage of feline allergic asthma, the changes in lactate levels were, however, highly heterogeneous and without clear trend (not shown). This may be attributable to variability in the recent exercise and nutrition of the cats. An NMR-detected study of EBC from adults suggested mild asthma to decrease short-chain fatty acids, valine, formate, hippurate and urocanic acid, as well as to increase proline, propionate, isobutyrate, and phenylalanine, Such changes cannot, however, be corroborated in cat EBC upon onset of experimental asthma. The human study normalized the EBC spectra to their spectral area, a common practice that can introduce artifacts of some metabolites *appearing* to decrease. The study of EBC of adult humans made no mention of an increase of acetone by asthma. These differences in prospective biomarkers of asthma suggest the need for caution and limits to generalizing across very different cohorts of asthma patients. Why was phthalate observed in this feline study? Phthalates are found in many consumer products. Humans consume phthalate esters introduced to the diet by the processing of foods, especially food packaging films and meats, which contain di-2-ethylhexylphthalate. The dry kibble diet of the cats was highly processed and contained animal fats, i.e. likely sources of phthalate esters by analogy with the contamination documented in human diets. Perhaps exhalation could be a route of elimination of phthalates that was impaired by the experimental asthma in the cats. The most serious technical impediment to use of EBC from cats is the inherent large variability, which appears unrelated to the pulmonary disease state. This stymied PCA from distinguishing early experimental asthma from health. Fortunately, probabilistic quotient normalization enabled successful discrimination using supervised statistical methods based on PLS-DA. For the best diagnostic separations, multi-level PLS-DA or preprocessing with orthogonal signal correction suppressed much variability that interfered in diagnostic classification. With OSC-PLS-DA, the predictive model was simplified to only three to four components and attained predictive power of 94% sensitivity and 94% specificity. Multi-level suppression of between-subject variability combined in novel fashion with orthogonal signal correction notably increased the specificity in predicting asthma to 100%. Consequently, the combination of OSC and multi-level enhancements to PLS-DA appears promising for comparative metabolomic and clinical research studies where specimens can be assayed from each subject before and after a treatment or change in clinical status. # Conclusions After preprocessing of NMR spectra with probabilistic quotient normalization and *glog* transformation, OSC-PLS-DA and multi-level PLS-DA overcame confounding variability among EBC samples to distinguish asthma from health noninvasively in a cohort of research cats. The predictive ability of the OSC-PLS-DA model is promising as both the sensitivity and specificity are 94%. The promising biomarkers of allergic asthma in cats to emerge in their EBC are increases in acetone, unidentified metabolite(s) with broad NMR peaks near 5.8 ppm, and an aromatic compound probably containing a hydroxyphenyl group. Also promising is the decrease of phthalate in 60% of the cats with asthma. The noninvasive, untargeted approach utilizing properly preprocessed NMR spectra of EBC interpreted with OSC-PLS-DA appears worth further evaluation for translational research. Reliable differential diagnosis of early asthma in children is especially needed. # Supporting Information We are grateful to Richard H. Barton for recommending the PQN method and for comments on the manuscript. BALF bronchoalveolar lavage fluid COLMAR Complex Mixture Analysis by NMR EBC exhaled breath condensate GC-MS gas chromatography–mass spectrometry HMDB human metabolomics database HSQC heteronuclear single quantum coherence MMCD Madison-Qingdao Metabolomics Consortium Database OSC orthogonal signal correction PC principal component PCA principal component analysis PLS-DA partial least squares with discriminant analysis OSC-PLS-DA orthogonal signal correction partial least squares with discriminant analysis PQN probabilistic quotient normalization RMSEP root-mean-square-error-of-prediction STOCSY statistical total correlation spectroscopy VIP variable importance in projection [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** CRR SRV. **Data curation:** YGF CRR SRV. **Formal analysis:** YGF SRV. **Funding acquisition:** CRR SRV. **Investigation:** YGF MF CHC HR CRR SRV. **Methodology:** YGF SRV CRR. **Project administration:** SRV CRR. **Resources:** CRR CHC HR SRV YGF. **Software:** YGF. **Supervision:** SRV CRR. **Validation:** YGF SRV CRR. **Visualization:** YGF SRV. **Writing – original draft:** SRV YGF. **Writing – review & editing:** CRR SRV YGF CHC HR. [^3]: Current address: Department of Small Animal Clinical Sciences, College of Veterinary Medicine, Texas A&M University, 4474 TAMU, College Station, TX, 77843–4474, United States of America

# Introduction *Plasmodium* parasites cause malaria, a mosquito-borne disease responsible for hundreds of thousands of deaths and hundreds of millions of clinical cases annually. After transmission from feeding Anopheline mosquitoes, injected sporozoite-stage parasites travel through the skin and eventually make their way to the liver where they initiate an asymptomatic hepatocyte infection that culminates days later in release of erythrocyte-stage merozoites. The merozoites initiate a cyclical infection of erythrocytes that leads to all the clinical manifestations of malaria, including fevers and chills and sometimes progressing to severe anemia, coma and death. The single greatest advance in the fight against malaria would be the production of a safe and effective vaccine that induces complete protection against infection with *Plasmodium* sporozoites. To achieve this goal, a major focus has been placed on vaccines targeting the pre- erythrocytic stages of development (the transmitted sporozoite stage and the subsequent liver stage). Experimental vaccination of humans and mice with attenuated sporozoites protects against challenge with wild-type sporozoites. Antibodies can block hepatocyte invasion, while CD8⁺ cytotoxic T cells recognize parasite-infected hepatocytes and can provide sterile protection in several mouse models. Furthermore, CD8⁺ T cells play a role in protection in primates and likely in humans as well and are induced in humans experimentally immunized with attenuated sporozoites. Parasite-specific CD8⁺ T cells may kill infected cells by one or more proposed mechanisms (reviewed in). Thus, inclusion of CD8⁺ T cell target antigens is likely to be critical for any sterile protective malaria vaccine. There are two basic approaches to pre-erythrocytic malaria vaccine development–manufacture of an attenuated ‘whole organism’ vaccine or identification and manufacture of a single subunit antigen or set of antigens that can provide complete protection. However, the extremely large number of genes expressed by *Plasmodium* has thus far prevented the efficient identification of broadly protective antigens suitable either for inclusion within a subunit vaccine or for targeting by transgenic parasite vaccines with improved protective efficacy. For subunit vaccines, there are two sequential phases of development: antigen discovery and formulation testing. Antigen discovery typically involves screening of immune cells from pathogen-exposed subjects for responses against small numbers of laboriously cloned, pathogen-derived gene products or predicted peptide epitopes. Antigen discovery is then followed by an equally laborious process of production and evaluation of experimental vaccines targeting the newly discovered antigenic proteins–many vaccines fail here because the antigens discovered simply do not induce protective responses. As the number of proteins increases, each step requires considerable investment of time and money, with no guarantee that the antigens discovered will ultimately confer protection when formulated as a vaccine. *Plasmodium* species each encode \~5,300 proteins but only a few pre- erythrocytic *Plasmodium* proteins (i.e., CSP, TRAP, LSA1, Exp1/Hep17, CelTOS, L3, Pf16, STARP) have been studied as T cell antigens. Amongst those tested in humans (CSP, LSA-1, Exp-1, TRAP, CelTOS), most have not reliably induced complete protection (reviewed in), although CSP protected 85% of subjects in one study. Because it is nearly impossible to systematically study all potential T cell antigens using conventional methods, a number of higher throughput approaches have been applied to malaria T cell antigen discovery. Synthetic *Plasmodium* peptides or antigen presenting cells (APCs) transfected with *Plasmodium* protein-expressing plasmids have all been used to screen for T cell interferon-γ (IFNγ) responses *in vitro*, but such approaches are limited by the cost of large peptide libraries and the complexity of cloning A/T-rich Plasmodial genes, respectively. Recently, a peptide screening approach identified multiple new antigens targeted by T cell responses from RAS-immunized humans—protection afforded by these antigens is not yet known. We previously utilized APCs transfected with minigene libraries encoding long *Plasmodium* peptides in an effort to capture a larger portion of the proteome using synthetic biology techniques. New methods that accelerate discovery of vaccine subunits for pathogens with thousands of genes would be a major step forward in the fight against *Plasmodium* and other complex intracellular pathogens. DNA vaccination and screening is ideally suited for such higher-complexity evaluation of vaccine candidates. In malaria, DNA vaccination against *P*. *yoelii* CSP achieves CD8⁺ T cell-dependent protection in the rodent model. In addition, DNA vaccination of humans against *P*. *falciparum* CSP can induce CD8⁺ T cell responses. Finally, small numbers of *Plasmodium* protein-coding genes have been shown to be immunogenic when used in a four-gene DNA vaccine in mice or non-human primates. A five-gene DNA vaccine administered to humans was also able to induce CD8⁺ T cell responses that could be further boosted by exposure to sporozoites. Even though there was no evidence of protection in this human study, the study showed that a DNA prime / *Plasmodium* boost protocol can increase responses. Studies from outside the malaria field further indicate that complex DNA vaccines delivered by biolistic (gene gun) delivery can be physically segregated onto distinct gold beads and that this serves to maintain diversity of both antibody and T cell responses. We previously used a high-throughput synthetic minigene technology for *in vitro* detection of murine CD8⁺ T cells primed with live *Plasmodium* parasites *in vivo*. This work led us to hypothesize that these same minigenes could be used to both initiate specific immune responses *in vivo* as a DNA- based vaccine, and to subsequently screen DNA vaccine-induced T cell responses to identify antigenic targets. This approach potentially combines the antigenic complexity achieved by whole organism vaccines with the feasibility of multi- subunit vaccination. Using a microarray-based oligonucleotide synthesis technology, we rapidly produced two complex minigene vaccines, each encoding over \>1,000 peptides derived from 36 (vaccine 1) and 53 (vaccine 2) liver-stage *P*. *yoelii* proteins. Targets were a set of pre-erythrocytic proteins containing signal peptides and/or transmembrane domains. Putatively secreted or transmembrane proteins were selected based on their potential to cross both the parasite membrane and the parasitophorous vacuolar membrane to enter the hepatocyte MHC class I pathway. The assumption that such parasite proteins may be exported into the hepatocyte is based on the observation that exported erythrocyte-stage PEXEL/HT domain-containing proteins also contain a signal peptide or a transmembrane domain. Following vaccine production, mice were repeatedly gene gun vaccinated with the minigene libraries followed, in some mice, by sporozoite exposures. T cell IFNγ responses were evaluated and several novel responses were identified including a strong response to *P*. *yoelii* malate dehydrogenase (PyMDH), which was further characterized and found to bear similarities to another recently described *Plasmodium* liver-stage T cell response. This preliminary study indicates that T cell responses against genuine Plasmodium antigens can be induced and detected using highly complex DNA minigene vaccines. As described in the discussion, additional improvements designed to increase the breadth and magnitude of responses, coupled with improved screening sensitivity may allow minigene vaccination/screening technology to be used for high-throughput identification of protective T cell antigens. # Materials and Methods ## Vaccine design Eighty-nine liver-stage *P*. *yoelii* 17XNL proteins predicted to contain a signal peptide and/or have transmembrane domains were selected using filters built into PlasmoDB. Coding sequences were downloaded and broken into sequential 33 codon segments overlapping by 14 codons. In regions of variability, all permutations of closely-spaced variations were included as alternate minigenes. A pool-specific primer unique to groups of 10 minigenes followed by a start codon were appended onto the 5’ end of each segment, while a common primer sequence was appended onto the 3’ end. The reverse compliment of each 150 bp minigene template was ordered as a single oligo-pool synthesis (CustomArrays, Inc., Bothell, WA). Gene identifiers and product descriptions from [PlasmoDB.org](http://PlasmoDB.org) as well as pool-specific primers, minigene sequences and encoded peptides are listed in. ## Vaccine assembly Each pool of 10 minigenes was amplified using individual pool-specific primers plus the common primer. Pool-specific primers included a T7 promoter sequence on the 5’ end. Equivalent amounts of each pool were then combined and subjected to dial-out error correction as described. Briefly, random tags were added by amplifying the combined library using stepout primers hybridizing to the T7 and common primer sequences. The tagged library was sequenced using base paired-end reads on an Illumina miSeq by a commercial vendor. Reads were aligned with the original library design and dialout primer pairs flanking accurate minigenes were selected. Dialout primer pairs for each minigene were ordered from IDT, Inc. (Coralville, IA). Primers could not be adequately designed for some minigenes and such minigenes were omitted from the final library post-dial out. Error-corrected minigenes were amplified using the selected dial-out primers. Each minigene was individually recombined as an amino-terminal fusion with the mouse LC3 coding sequence in a modified pNGVL3 vector using SLiCE ligation- independent cloning. Recombined plasmids were transformed into DH10G *E*. *coli* hosts by electroporation using an AMAXA 96-well shuttle device (Lonza, Walkersville, MD) and cultured individually in 1.2 mL LB broth in deep-well 96-well plates. Groups of 10 cultures were pooled, centrifuged and frozen until purification. Bacterial pellets were processed using a 96-Plus Endotoxin-free kit (Qiagen, Valencia, CA) according to manufacturer instructions. Typical yields were between 100 and 300 ng/μL for each pool. ## Loading of gene gun cartridges Nine μL of each plasmid pool, each containing 10 plasmids, were combined with 1 μL of 200 ng LT adjuvant plasmid in 96-well V-bottom plates. Each well was diluted with 10 μL 50 mM spermidine and 1 mg of 1 μm gold beads (Inbios Gold, Hurstbridge, Australia). Plates were agitated on a horizontal shaker while 10 μL 10% CaCl₂ was added to each well. DNA-coated gold particles were centrifuged and washed thrice with 100% ethanol. Gold particles were suspended in 10 μL 100% ethanol with 50 μg/mL polyvinylpyrrolidone. One-quarter of all pools in a vaccine (the equivalent of twelve 40-kD proteins) were combined and loaded into Tefzel cartridges using a custom-made tube turner. Gene gun cartridge tubes were dried, sliced and stored desiccated at 4°C until use. ## Mice All animal studies were approved by the University of Washington Institutional Animal Care and Use Committee (protocol 4317–01). BALB/cj mice were obtained from Jackson Laboratories (Bar Harbor, ME) and housed in approved facilities at the University of Washington. Thy1.1⁺ BALB/c mice were bred at the University of Washington from a pair originally obtained from Jackson Laboratories. Humane sacrifice was performed by flow-metered carbon dioxide overdose. ## Gene gun vaccinations Female Balb/cj (6–8 wk old) mice were shaved on the abdomen and administered the DNA vaccine corresponding to 0.5–1.0 μg total DNA using a PowderJect XR1 research device. On Days 0, 21 and 49, each mouse received four separate cartridges, one cartridge corresponding to each quarter of the vaccine. At later time points, vaccinated mice were humanely sacrificed and splenocytes harvested and pooled for screening as described later. ## Generation of screening templates One μL of each vaccine pool was diluted 1:100 and further amplified using Rolling Circle amplification using a TempliPhi kit (GE BioSciences, Pittsburgh, PA). Rolling circle-amplified DNA from each minigene pool was used as a template to amplify short linear expression cassettes from the CMV promoter through the human beta-globin 3’UTR/polyA sequence. PCR reactions utilized 20 ng template combined with a CMV promoter-specific primer and a 3’ human beta-globin UTR antisense primer and Phusion polymerase. Cycling conditions were 35 cycles of 98°C (15 sec), 55°C (15 sec) and 72°C (2 min) each. Products were purified using 96-well filter plates (Qiagen, Valencia, CA) and resuspended in 20 μL 10 mM Tris (pH 7.8). ## Minigene library ELISPOT screening of sensitized splenocytes Ready-Set-Go mouse IFNγ ELISPOT kits (eBioscience, San Diego, CA) were used as previously described. Briefly, 2 μL of PCR-amplified expression cassettes from each minigene pool in each vaccine were transfected into 1x10⁶ freshly-dividing P815 cells using the RAW264.7 program of an AMAXA 96-well shuttle. Splenocytes (1x10⁶/well) were added as responders, and plates were incubated overnight. ConA was included as positive control and mock- transfected P815 cells were used as negative controls. ELISPOT plates were developed according to the manufacturer’s instructions and scanned and analyzed using an ImmunoSPOT counter (Cellular Technology Limited, Shaker Heights, OH). Transfection efficiency was also controlled by checking GFP expression and viability by flow cytometry of separate wells of P815 cells transfected on each plate of wells. ## Sporozoite immunizations Where indicated, BALB/cj mice were singly or doubly immunized with 1x10⁴ wild-type *P*. *yoelii* sporozoites and were then administered azithromycin i.p. (0.8 mg/d) on Days 1–3 post-immunization. In other experiments, mice were immunized once or twice with combinations of PyRAS and/or genetically-attenuated Pyfabb/f- sporozoites as described. Sporozoites were obtained from the Center for Mosquito Production and Malaria Infection Research (CeMPMIR, Center for Infectious Disease Research, Seattle, WA). ## MHC binding assay RMA/S cells expressing H2-K^d were used to test MHC binding as described. ## *In vivo* cellular cytotoxicity assay The *in vivo* killing assay was performed to measure cellular cytotoxicity as previously reported although three populations of target cells were differentially labeled here with Cell Proliferation Dye eFluor® 670 (eBioscience) at 5 μM (PyMDH), 0.75 μM (PyCSP) and 0.1 μM (mock). ## Immunization-challenge experiments To generate DNA and Listeria-based vaccines for single antigen immunization, minigene-encoded antigens were isolated and further cloned into the delivery vectors. Briefly, minigene expression constructs encoding the dominant PyCSP epitope SYVPSAEQI and the dominant PyMDH epitope SYQKSINNI were cloned separately into the pNGVL3.LC3 vector for use as DNA vaccines. Recombinant PyCSP-expressing *actA-/inlB- L*. *monocytogenes* was a kind gift from Aduro Biotech. PyMDH-expressing *actA- L*. *monocytogenes* was constructed by cloning a minigene encoding the PyMDH epitope into the pPL2-N4 vector. The resulting pPL2-N4.MDH plasmid was transformed into SM10 hosts and conjugated into *actA-* Lm10403S. Using these materials, a gene gun DNA prime/recombinant *Listeria* boost protocol was used to vaccinate mice against single antigens prior to sporozoite challenge. For each arm of the experiment, mice were primed against either PyCSP or PyMDH using two gene gun cartridges on Days 0 and 2. Three weeks later, animals were boosted i.v. with 5x10⁶ CFU recombinant *Lm* expressing the cognate antigen and treated with 2 mg/mL ampicillin in the drinking water for 3 days thereafter. Animals were challenged three weeks later with 1x10⁴ wild-type *P*. *yoelii* sporozoites i.v. Forty-four hr later, animals were humanely sacrificed and livers perfused with PBS. Livers were harvested and emulsified in a bead beater with NucliSens lysis buffer (bioMérieux, Durham, NC) to preserve RNA. Total nucleic acid was extracted using a NucliSens EasyMag (bioMérieux) and *Plasmodium* 18S rRNA and murine GAPDH were measured by RT-PCR as described. Liver burden is reported as log₁₀ changes between mice for *Plasmodium* 18S rRNA copy number normalized to mouse GAPDH. At the time of humane sacrifice, spleens were also harvested for ELISPOT analysis as a measure of T cell frequency at the time of challenge (no T cell expansion during 0–44 hr post-challenge). # Results ## Synthetic biology methods enable rapid production of multi-protein synthetic minigene libraries suitable for T cell vaccination and subsequent screening Two minigene libraries encoding peptides from 36 (Vaccine 1) or 53 (Vaccine 2) *P*. *yoelii* proteins each were synthesized. Error-correction including Dial- out tagging, sequencing, analysis and re-amplification was completed in 4 weeks. The libraries were designed to contain the complete peptide compliment of all included proteins, but after error correction by dial-out PCR, a fraction of minigenes could not be recovered as error-free sequences. Vaccine 1 contained 91.9% of the intended proteome (95%CI: 87.4–96.4%) and Vaccine 2 contained 91.1% (95%CI: 87.0–95.3%) as indicated in. Plasmid insertion and plasmid culture/isolation were accomplished in three weeks. Spot-checking individual cultures by sequencing showed that \>90% of clones encoded the intended sequence. The proteins included in the libraries are listed in. Minigenes were adjuvanted with the LT-encoding plasmid and loaded onto individual aliquots of gold particles for biolistic delivery. ## Library vaccination and minigene screening enables identification of library vaccine-induced IFNγ-producing T cell responses BALB/cj mice (n = 5/group) were vaccinated with Vaccine 1 or Vaccine 2 on Days 0, 21 and 49 using a PowderJect research device. Aside from mild erythema at the gene gun vaccination site for 1–2 days post-vaccination, the mice tolerated the DNA vaccination without any obvious signs of distress. IFNγ responses were evaluated usingpooled splenocytes harvested six days after the final gene gun vaccination. To evaluate the \>100 minigene pools per vaccine, duplicate wells of target cells (P815) were each transfected with PCR-amplified expression cassettes corresponding to each minigene pool in a vaccine (10 minigenes/pool). The transfected target cells (1x10⁶/well) were combined with 1x10⁶ splenocytes from library-vaccinated mice in overnight IFNγ ELISPOT assays as previously reported. Compared to previous minigene-transfected ELISPOT screens of sporozoite-immunized mice, background production of IFNγ in these experiments on DNA minigene library-vaccinated mice was extremely low (\<3 SFU/million splenocytes). Such low background signal is critical for detection of responses in the setting of multi-protein vaccination. Two antigens in each vaccine induced responses significantly above baseline. Each of the antigenic pools contained minigenes from a single protein such that the initial screen identified responses to the products of genes PY00619 and PY00638 in Vaccine 1 and PY01906 and PY03376 in Vaccine 2. In Vaccine 2, the known epitope from the circumsporozoite protein (CSP, pool 55) did not yield detectable responses. There are several possible explanations for this result. First, a particular minigene may not express in an optimal manner for several reasons including failure to clone, presence of a PCR-induced mutation, inclusion of a suboptimal 5’ UTR which is dictated by the pool- specific primer, efficient shunting into the autophagosome via the LC3 tag, and/or poor representation relative to other minigenes in that pool. Another possible explanation is that ELISPOT screening using minigene-transfected P815 cells as APCs is less sensitive than peptide ELISPOTs since the minigene- transfected APCs express lower quantities of antigen, typically yielding spots of smaller diameter and lower intensity. To evaluate these possibilities, we subcloned the CSP epitope-containing minigene from Vaccine 2 using minigene specific primers and verified its presence and primary sequence in the library (data not shown). In a mixed minigene vaccination experiment using immunization with a single pool consisting of 20 linear minigenes with and without the *library-derived* LC3-tagged CSP minigene, we observed relatively low frequency responses (20–30 spots/million splenocytes) using CSP peptide as a recall antigen. When the CSP epitope was tested as a single minigene vaccine with a ubiquitin tag, this minigene could elicit strong CSP responses in IFNγ ELISPOTs (500–1000 spots/million splenocytes). Thus, it appears that the minigene encoding the major CSP epitope as a fusion to LC3 was functional but was unable to induce a response when formulated to at extremely high complexities. In addition, since we were unable to detect CSP responses in these experiments, we do not regard the lack of response against any library-encoded peptide as evidence that the target is not immunogenic or protective. Optimized delivery and increased dose are likely to induce responses against a larger percentage of the proteins targeted in these vaccines. ## A subset of library vaccine-induced responses are recalled by immunization with irradiated *P*. *yoelii* sporozoites but only one such response is primed by sporozoites alone A subset of the minigene library vaccinated animals were administered 2x10⁴ irradiated *P*. *yoelii* sporozoites (PyRAS) 16 days after the final gene gun vaccination and splenocytes were harvested 7 days after RAS exposure. Minigene-transfected P815 ELISPOT screening was again conducted to identify vaccine-primed, RAS-recalled responses. Recall responses were observed against pools encoding PY00619, PY00638 and PY03376 whereas responses against the PY01906 pool were undetectable. To determine if RAS alone could induce these responses, completely naïve BALB/cj mice were administered 2x10⁴ PyRAS and evaluated by minigene-transfected P815 ELISPOT one week later. In mice exposed only to PyRAS, the only detectable response was to the pool for PY03376, which encodes *P*. *yoelii* malate dehydrogenase (PyMDH). ## Identification of a Class I epitope in *P*. *yoelii* malate dehydrogenase Three adjacent minigene pools from PyMDH were targeted in the primary screens. These pools included a total of 23 overlapping minigenes that were required to encode all permutations of four closely-spaced, non-synonymous sequence variants reported in the genome of *P*. *yoelii* strain 17XNL. It is unknown whether these reported variations are genuine or are the result of sequencing errors. In an attempt to identify a targeted epitope, the translated sequence of each variant minigene was evaluated for predicted Class I MHC binding using the NetMHC Pred 3.4 tool (available at <http://www.cbs.dtu.dk/services/NetMHC-3.4/>). Prediction was limited to MHC Class I H2^d molecules since mastocytoma P815 targets used in the screens do not express MHC class II. The peptide SYQKSINNI yielded strong predicted binding characteristics for H2-K^d and corresponded to a homologous, invariant sequence reported for *P*. *berghei* (PBANKA_1117700) MDH. This peptide was synthesized and tested by ELISPOT using frozen splenocytes from animals vaccinated with Vaccine 2 and from animals exposed to one or two doses of 1x10⁴ Py wild-type sporozoites with concurrent prophylactic azithromycin drug treatment to prevent onset of erythrocyte-stage infection. Splenocytes from all such mice responded to the PyMDH peptide, although the response in animals exposed to two doses of Py sporozoites was markedly reduced compared to a single dose of sporozoites. Sequencing of a genomic PCR product amplified from the *P*. *yoelii* 17XNL isolate used for PyRAS production confirmed that SYQKSINNI was the encoded peptide. Despite the aforementioned ambiguity in the Plasmodb database, sequencing revealed no evidence of non- synonymous variation within the SYQKINNI-coding region. The same approach was used to bioinformatically predict H2^d and H2^b binders for PY00619, PY00638, and PY01906. However to date we have not identified any specific peptide epitopes from these proteins by ELISPOT screening (data not shown). Nonetheless, the PY00619, PY00638 and PY01906 minigene pools were indeed antigenic as they were each able to induce \>100 SFU/million when administered to mice as a single pool gene gun immunization (data not shown). It is possible that some of the minigene-induced responses are directed at fusion peptides formed by the junction of the pathogen-derived peptide and the C-terminal LC3 fusion partner. Such antigens could not be present in the parasite itself and would therefore not be boosted upon challenge, however fusion peptides of this nature have not been formally tested. These data highlight the complexity of epitope prediction and suggest that the reactive peptide in the minigene screening stage was a different peptide than that predicted bioinformatically. For the purposes of this study, the peptides from PY00619, PY00638 and PY01906 were not pursued further. When ELISPOT assays were conducted in the presence or absence of anti-CD8a antibodies, PyCSP- and PyMDH-specific T cell responses were comparably reduced, indicating IFNγ production in both antigen-specific responses was CD8⁺ T cell-dependent. ## The MDH epitope SYQKINNI binds to murine H2-K^d The affinity of SYQKSINNI for H2-K^d was determined using an RMA/S lymphoma cell line expressing H2-Kd as previously described. H2-K^d was stabilized by a known binder (PyCSP-derived SYVPSAEQI) and with similar affinity by PyMDH-derived SYQKSINNI. The apparent K_d for the peptides was 11.6 nM (PyMDH) compared to 9.6 nM (PyCSP). ## PyMDH-specific CD8⁺ T cells are functionally cytotoxic To evaluate the functional capacity of PyMDH-specific T cells, BALB/cj (Thy1.2⁺) mice were primed i.v. with 2x10⁶ GM-CSF-bone marrow-derived dendritic cells activated and pulsed overnight with lipopolysaccharide (0.1 μg/mL) and either 1 μg/mL PyCSP or PyMDH or no peptide. Six days later, splenocytes from Thy1.1 BALB/c mice were pulsed with PyCSP or PyMDH peptides or with no peptide and were differentially stained, mixed and injected into animals primed against individual peptides. Eighteen hr later, spleens were harvested and peptide-specific killing measured by flow cytometry. Animals immunized against the PyCSP peptide specifically killed 84% of PyCSP- pulsed targets, and animals sensitized to the PyMDH peptide killed 91% of PyMDH- pulsed targets. ## High-frequency PyMDH-specific CD8⁺ T cells do not protect mice from *P*. *yoelii* sporozoite challenge Five animals per group were vaccinated with either PyCSP or PyMDH using a DNA prime/*Listeria* boost protocol. At a memory time point, all vaccinated mice plus five naïve infectivity control mice were challenged with 1x10⁴ wild-type *P*. *yoelii* sporozoites. ELISPOT performed at the time of liver harvest demonstrated antigen-specific, high frequency IFNγ-producing responses to PyCSP and PyMDH in appropriately immunized mice. Compared to infectivity control mice, 5/5 PyCSP-vaccinated mice showed significant vaccine-induced protection, with one animal showing undetectable Py 18S RNA in the liver indicating sterile protection. In contrast, the Py 18S rRNA concentration in animals vaccinated with PyMDH was indistinguishable from that of the infectivity controls, indicating no protection despite high frequency PyMDH-specific T cell responses. ## Heterologous cross-species immunization with a late-arresting sporozoite re-expands the PyMDH-specific T cell population more than homologous immunizations Previous studies showed that T cell responses to pre-erythrocytic antigens could either be expanded by repeated homologous (same-species) sporozoite immunization (e.g., CSP-specific CD8 T cells) or failed to re-expand (e.g., L3-specific CD8 T cells). Recent unpublished work has shown that immunization with two different murine-infecting *Plasmodium* species (*P*. *yoelii* 17XNL and *P*. *berghei* ANKA) could recall the L3-specific response to the epitope shared between species if the secondary immunization was with a late-arresting, genetically- attenuated sporozoite (S. Murphy, pers. comm.–manuscript in review). Since the PyMDH epitope was also conserved between *P*. *yoelii* and *P*. *berghei*, mice underwent homologous or heterologous immunizations with combinations of PyRAS, PbRAS and Pyfabb/f- parasites. As observed for L3, when RAS parasites were used, neither homologous (PyRAS→PyRAS) nor heterologous (PbRAS→PyRAS) regimens re- expanded the PyMDH-specific cells. However, when the secondary immunization was switched to the less-attenuated Pyfabb/f- strain, the heterologous regimen (PbRAS→Pyfabb/f-) expanded PyMDH-specific responses more so than the homologous regimen (PyRAS→Pyfabb/f-), although not to the level observed with a single Pyfabb/f- immunization. This finding suggests that PyMDH-specific responses and those against other proteins with similar protein expression kinetics may be boosted by heterologous more so than homologous species immunizations. # Discussion Development of antibody-based subunit vaccines has been aided by the inherent stability, sensitivity and soluble nature of antibodies in serum. In contrast, the development of T cell-based vaccines requires evaluation of live cellular MHC-restricted interactions between T cells and MHC-matched APCs displaying the cognate target antigen. For intracellular pathogens expressing thousands of potential targets such as malaria, this requirement has inhibited discovery of protective vaccine subunits. The synthetic minigene vaccine technology described here is designed to enable the rapid evaluation of T-cell responses against large numbers of pathogen-derived proteins, in order to identify both dominant and sub-dominant antigens. In these proof-of-principle studies, minigene vaccination resulted in the identification of four novel T cell antigens from among 89 pre-erythrocytic stage *P*. *yoelii* proteins. These included one dominant antigen encoded by PY03376 (PyMDH), and three antigens encoded by PY00619, PY0638 and PY01906 minigene pools. We identified the peptide SYQKSINNI as a major PyMDH epitope. Robust responses against PyMDH were detected in animals receiving the vaccine alone and in response to sporozoite immunization in the absence of DNA vaccination. PyMDH-specific responses in DNA-vaccinated animals could be recalled by sporozoites. However, despite recent reports of a protective T cell response against *Toxoplasma gondii* malate dehydrogenase, no protection against PyMDH was observed following DNA prime/*Listeria* boost vaccination despite induction of extremely high frequency, cytotoxic PyMDH-specific CD8⁺ T cell responses. The PyMDH response was also evaluated in mice exposed to one or two rounds of sporozoite immunizations. The PyMDH-specific response was not potently recalled by homologous sporozoite immunizations, although the epitope is shared with *P*. *berghei* ANKA MDH and recall could be partially augmented by immunization with two different species of murine-infecting *Plasmodium* sporozoites. Thus, PyMDH appears to be an antigen in the same category as that of the recently described pre-erythrocytic *P*. *yoelii* ribosomal protein L3, which also displayed robust pre-erythrocytic immunogenicity but no protective efficacy and poor recall by repeated homologous sporozoite exposures. The other antigenic minigene pools have not been characterized to the same extent as PyMDH. Animals receiving PyRAS 16 days after the final DNA library vaccine exhibited low level responses against PY00619 and PY00638 minigene pools but not to PY01906, and additional studies will be required to determine whether these responses represent true parasite recall or simply residual, declining responses from the initial DNA vaccination. Each of these three minigene pools was able to induce IFNγ responses in mice vaccinated with single pool DNA vaccines. The antigenic epitope(s) for PY00619, PY00638 and PY01906 are currently unknown and additional studies are ongoing to evaluate these antigens further. The fact that the robust responses against PyMDH are not protective suggests that the epitope may not be displayed on the surface of an infected hepatocyte harboring live proliferating parasites, or may be presented by the hepatocyte at a time point that is too late to provide protection. To be displayed on the surface of the hepatocyte, PyMDH would probably need to be exported from the parasite and across the hepatocyte-derived vacuole membrane to the hepatocyte cytoplasm. However unlike erythrocyte-stage parasites, there is currently no known consensus motif that directs pre-erythrocytic protein export from the vacuolar parasite to the hepatocyte cytoplasm. Responses to PyMDH-like epitopes may be driven either entirely or in part by cross-presentation of dead or dying parasites. We and others have identified other antigens with similar response profiles. This type of antigen may represent an immune evasion strategy employed by the parasite whereby epitopes derived from proteins that are not protective serve as “decoys”, directing immune system resources toward non-protective targets. There is increasing evidence that this decoy antigen phenomena also occurs in bacterial and viral infections. Such antigens may be under selective pressure to remain invariant as T cell responses against them could promote survival of the parasite. In contrast, T cell epitopes from proteins that are protective are likely to be expressed and presented by hepatocytes during liver- stage infection and therefore may be under selective pressure to alter their primary sequence to avoid immune detection. Separating antigenic protective antigens (the “wheat”) from dispensable antigens such as PyMDH and *P*. *yoelii* ribosomal protein L3 (the “chaff”) represents a major hurdle to our understanding of the *Plasmodium*-specific T cell repertoire and to development of effective subunit vaccines for malaria. Both multi-gene DNA library vaccines elicited responses against just two novel antigens each. Thus, a major question concerning our ‘highly parallel immunization’ strategy is whether immunization with highly complex DNA vaccines is capable of stimulating responses against a highly diverse set of targets, even when small subsets of minigenes are segregated and delivered to different APCs via different gold bead carriers. The malaria literature indicates that mixtures of small numbers of *Plasmodium* genes encoding *bona fide* antigens are immunogenic in mice and non-human primates and humans, although the quantity of DNA and the administration methods differ from the gene gun approach reported here. In one mouse study, administration of a mixture of five *P*. *falciparum* antigen-coding plasmids led to alterations (CSP and Exp1 decrease and LSA1 increase) in T cell immunogenicity compared to immunization with single gene plasmids only. During construction of the vaccines reported here, each of the more than 100 minigene pools (10 cloned minigenes/pool) in each vaccine were individually precipitated onto separate aliquots of gold beads in parallel in an effort to promote diverse responses against multiple vaccine-encoded antigens. Such physical segregation of antigens onto different pools of gold beads is known to promote both antibody and T cell response diversity in other studies of gene gun-delivered antigens. However, the minigene library vaccine represents a much larger set of antigens than studied in the aforementioned references, and we are uncertain how the degree of diversity and the absolute quantity of any individual antigen-encoding DNA ultimately affect immune response outcomes. In addition, a small portion of some proteins were not included in the library because correct sequences were not recovered following dial-out error correction. Therefore, some bona fide antigens could also have been encoded in the omitted sequences. Given these caveats, the library vaccination-screening strategy described herein can *rule in* T cell responses, but is not sufficiently sensitive in its current form to *rule out* immunogenicity or protective effects of targets where responses are not detected or for peptides not covered for specific proteins. Given the suboptimal performance of the CSP control, we are currently exploring several methods to increase the intensity and breadth of responses induced by complex DNA library vaccines. First, as mentioned earlier, we have adopted a standard 5’ UTR and a ubiquitin fusion design that have proven superior to the design used in the LC3-based vaccines. Ubiquitin-tagged proteins have shown increased CD8⁺ T cell immunogenicity in many systems including for some malaria DNA vaccine antigens. Simply increasing the dose of DNA on gold beads is unlikely to yield a significant effect as previous titration studies with the gene gun suggest that the dose of DNA required to stimulate a response is well within the range of each construct achieved in this protocol. In contrast, increasing the number of DNA vaccination cartridges administered per animal would result in an increase in the number of DNA-coated particles delivered to dermal APCs, potentially boosting the number of responders induced against multiple antigens. In addition, using the sub-optimal CSP minigene, we have tested novel adjuvants targeting B7 family inhibitory receptors and have observed a significant increase in the number of responders induced (B. Stone, S. Murphy, unpublished observation). Furthermore, we tested a P815 cell line that overexpresses the co-stimulatory molecule B7.1 as APCs in minigene- transfected ELISPOT screening assays and observed five-fold higher sensitivity for IFNγ producing cells, indicating that additional antigens may be discoverable by ELISPOT with optimized transfection-compatible APCs. Vaccination studies testing these improvements using these and additional minigene libraries are currently underway. # Supporting Information We thank Emma Fritzen, Heather Kain, Matt Fishbaugher, Will Betz and Stefan Kappe of CeMPMIR (CID Research) for sporozoite production. B.C.S. and S.C.M. have filed provisional patents through the University of Washington with the US Patent and Trademark Office for novel methods of multi-gene DNA vaccination using minigene libraries and for novel antigens. [^1]: Please note that Drs. Stone and Murphy filed provisional patents through the University of Washington with the US Patent and Trademark Office for novel methods of multi-gene DNA vaccination using minigene libraries and novel antigens (Patent Applications 62/200,487, System and Method for Highly Parallel Immunization and Antigen Discovery and 62/281,343, Identification of Plasmodium Malate Dehydrogenase as a Protective Malaria Vaccine Antigen). These filings do not alter the authors' adherence to PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: BCS SCM. Performed the experiments: BCS AK ZPB SCM. Analyzed the data: BCS ZPB SCM. Contributed reagents/materials/analysis tools: DHF JTF JS. Wrote the paper: BCS SCM.

# Introduction The rat has long been an invaluable animal model in many biomedical research fields, including behavioral studies, cardiovascular disease, immunology, transplantation, toxicology and pharmacology. However the use of rats has been hindered by the lack of embryonic stem (ES) cells and the consequent difficulty in generating animals with precise genetic modifications. Derivation of germline competent rat ES cells, and their functional equivalent, induced pluripotent stem cells, thus represents a major step forward. Induced pluripotent stem (iPS) cells can be derived from somatic cells by forced expression of exogenous transcription factors, notably Oct4, Sox2, c-Myc and Klf4. These activate endogenous pluripotency genes and reprogram the cell to a pluripotent state. ES and iPS cells have now enabled the generation of rats with gene targeted inactivation of p53, protease-activated receptor-2 and hypoxanthine phosphoribosyltransferase. Directed differentiation of rat pluripotent stem cells also provides a source of cell types such as cardiomyocytes, which are useful for toxicological screening, research into tissue regeneration and development of organ repair procedures. Rat iPS cell lines have been derived from different rat strains and a variety of somatic cells, including embryonic fibroblasts, neural precursor cells, bone marrow cells, liver progenitor cells, and ear fibroblasts. However this work has been based on retroviral or lentiviral vectors that have drawbacks. Establishment of a true self-sustaining pluripotent state, independent of exogenous reprogramming factor expression, requires epigenetic silencing of the exogenous genes. Virally transduced genes are frequently silenced in the host cell, but persistence or reactivation of factor expression interferes with differentiation, and c-Myc expression has led to tumor formation in iPS derived offspring in mice. Next generation iPS cells must therefore incorporate tight control of transgene expression. The limited capacity of retroviral vectors also means that individual retroviruses are commonly used to introduce reprogramming factors. Multiple independent infections of each cell are thus required to deliver the full complement of factors and many cells may not receive an ideal, equimolar ratio of each gene. Here we report iPS cell generation using a non-viral vector containing the murine reprogramming factors Oct4, Sox2, c-Myc and Klf4 controlled by a bidirectional doxycycline-inducible promoter. # Materials and Methods Animal experiments were approved by the Government of Upper Bavaria and performed according to the German Animal Welfare Act and European Union Normative for Care and Use of Experimental Animals (permit number: Az. 209.1/211-2531-114-03). Chemicals were obtained from Sigma Aldrich, and cell culture media and supplements from PAA Laboratories or Gibco Life Technologies unless otherwise specified. Oligonucleotide sequences are shown in. ## Plasmids The reprogramming cassette of pReproII-attB was assembled by standard cloning and PCR methods using the P_bi-1 promoter from pBI-5 (Clontech) and cDNAs for murine Oct4, Sox2, Klf4 and c-Myc (ImaGenes). Two coding regions, linked by a T2A peptide sequence, were placed on either side of the P_bi-1 promoter. Each bicistronic coding region (Oct4-T2A-c-Myc or Sox2-T2A-Klf4) is preceded by a hybrid intron and terminated by the bovine growth hormone (bGH) polyadenylation signal. The reprogramming unit was combined with a CAG (cytomegalovirus early enhancer element and chicken beta-actin promoter) promoter driven expression cassette for rtTA2(S)-M2 and tTS^KRAB from plasmid pRTS-1, in which the Eμ enhancer was replaced by the cytomegalovirus early enhancer. A synthetic DNA element containing a 53 bp attB site, which can be recognized by φC31 integrase, was added to generate pReproII-attB. The φC31 integrase expression plasmid pCAG-C31Int(NLS) has been described previously. ## Generation of Rat iPS Cells Adipose tissue-derived mesenchymal stem cells (rADMSC) were isolated from subcutaneous fat, and fibroblasts from ear tissue (rEFs) of Fischer and Wistar rats according to standard methods. The passage number of the source cells was between P2 and P3, both showed typical fibroblast-like morphology. For the generation of iPS cells, 5×10⁵ rADMSCs were nucleofected with 1 µg reprogramming vector pReproII-attB and 1 µg φC31 integrase expression vector pCAG-C31Int(NLS) using the Nucleofector II device (Lonza) with program U-023 and the Human MSC Nucleofector Kit (Lonza). 5×10⁵ rEFs were nucleofected with 3 µg pReproII-attB and 3 µg pCAG-C31Int(NLS) using program A-024 and the Basic Nucleofector Kit for Primary Mammalian Fibroblasts (Lonza). Cells were plated onto tissue culture flasks on day 0 and nucleofection repeated on day 3. Transfected cells were transferred onto mitomycin C inactivated mouse embryonic fibroblasts (MEFs) on day 6. A schematic overview is shown in. Rat iPS cells were derived in two different media. N2B27-3i medium: 1∶1 mixture of N2 medium (DMEM/F12, 1×N2 supplement, 100 µg/ml BSA fraction V) and B27 medium (Neurobasal medium, 1×B27 supplement w/o retinoic acid, 2 mM Glutamax) supplemented with 0.1 mM 2-mercaptoethanol, 20% Knockout Serum Replacement, 1000 U/ml hLIF (produced in house), 3 µM GSK3β inhibitor CHIR99021 (Axon Medchem), 0.5 µM MEK1/2 inhibitor PD0325901 (Axon Medchem), 0.5 µM ALK5 inhibitor A83-01 (Biotrend). N2B27-2i medium: 1∶1 mixture of N2 medium and B27 medium (see above) supplemented with 0.1 mM 2-mercaptoethanol, 1000 U/ml hLIF, 3 µM GSK3β inhibitor CHIR99021, 0.5 µM MEK1/2 inhibitor PD0325901. Doxycycline (1.5 µg/ml) was added to culture media to induce expression of reprogramming factors. ## General Cell Culture Rat adipose tissue-derived mesenchymal stem cells (rADMSC) were cultured in MSC medium (MEM α, 10% FCS). Rat ear fibroblasts (rEF) were cultured in EF medium (DMEM, 15% FCS, 1× non-essential amino acids, 1× sodium pyruvate, 2 mM Glutamax, 5 ng/ml bFGF (Promokine)). Mouse embryonic fibroblasts (MEF) were cultured in DMEM+ medium (DMEM, 10% FCS, 1× non-essential amino acids, 2 mM Glutamax, 1× sodium pyruvate). Rat iPS cells were maintained on mitomycin C inactivated MEFs in N2B27-2i or N2B27-3i medium. iPS cells were routinely subcultured every 3 to 4 days by flushing loosely attached colonies off the feeder layer and dissociation by Accutase in suspension. For feeder-free monolayer culture, plates were coated with either 0.1% gelatin from bovine skin in PBS, 5 µg/ml fibronectin in PBS, 200 µg/ml rat tail collagen type I (Serva) in H₂O, 2% growth factor reduced Matrigel (BD Biosciences) in DMEM/F12, 2% Geltrex (Invitrogen) in DMEM/F12 or 4.2 µg/ml laminin (Roche) in PBS for 2 h at 37°C. ## Isolation of Genomic DNA and RNA Genomic DNA for bisulfite sequencing and PCR was isolated using the GenElute Mammalian Genomic DNA Miniprep Kit (Sigma). DNA for Southern blot analysis was obtained by standard phenol/chloroform extraction. RNA was isolated using either Trizol (Invitrogen), or the High Pure RNA Isolation Kit (Roche), and genomic DNA removed by treatment with the Turbo DNA-free Kit (Ambion) according to the manufacturer’s instructions. ## RT-PCR RNA was reverse transcribed with random hexamer oligonucleotides using SuperScript III (Invitrogen) according to the manufacturer’s protocol. PCR with GoTaq DNA polymerase (Promega) was performed using oligonucleotides listed in. Thermal cycling conditions were: 94°C, 2 min; 30 cycles of 94°C for 30 s, 58°C for 30 s, 72°C for 1 min; then final elongation 72°C for 5 min. Quantitative PCR was performed using a 7500 Fast Real-Time PCR System and the SYBR Green PCR Master Mix (Applied Biosystems) with oligonucleotides listed in, according to the manufacturer’s instructions. Thermal cycling conditions were: 95°C, 10 min; 40 cycles of 95°C for 15 s, 60°C for 1 min. Expression of the exogenous reprogramming factors was calculated with the ΔΔCT method and expressed as fold change relative to the corresponding cell line with doxycycline induction. ## Southern Blot Analysis 10 µg genomic DNA was digested with *BglII*, separated by gel electrophoresis and transferred to Hybond-N+ membrane by capillary blotting. The hybridization probe was generated by PCR with GoTaq DNA polymerase (Promega) incorporating alkali-labile digoxigenin-11-dUTP (Roche). The 1308 bp Klf4 probe was amplified from pReproII-attB with primers prKlf4_F and prKlf4_R. Thermal cycling conditions were: 94°C, 2 min; 35 cycles of 94°C for 30 s, 58°C for 30 s, 72°C for 90 s; then final elongation 72°C for 5 min. Hybridization using DIG Easy Hyb (Roche) and probe detection using anti-digoxigenin antibody Fab fragments conjugated with alkaline phosphatase (Roche) were performed according to the manufacturer’s instructions. ## PCR for Pseudo-attP Sites Pseudo-attP site PCR analysis was performed as described using published oligonucleotides for rat pseudo-attP sites rps1, rps2 and rps3. Each PCR assay used one primer within the reprogramming vector and one primer in the genomic DNA sequence upstream or downstream of rps1, 2 or 3 sites. PCR was performed with GoTaq DNA polymerase (Promega) according to the manufacturer’s instructions. Thermal cycling conditions were: 94°C, 2 min; 35 cycles of 94°C for 30 s, 58°C for 30 s, 72°C for 2 min; then final elongation 72°C for 5 min. PCR products from rat iPS cell lines T1 and T13 were subcloned into pJet1.2/blunt (Fermentas) and the DNA sequence determined. ## Alkaline Phosphatase and Immunocytochemistry Cells were fixed with 4% paraformaldehyde. Alkaline phosphatase staining was performed with SIGMA*FAST* BCIP/NBT according to the manufacturer’s instructions. Immunocytochemistry of undifferentiated and differentiated iPS cells was performed with primary antibodies against Oct4 (1∶100; sc-8628, Santa Cruz), SSEA1 (1∶200; sc-21702, Santa Cruz), SSEA4 (1∶100; sc-21704, Santa Cruz), albumin (1∶100; A0001, Dako), sarcomeric α-actinin (1∶250; EA-53, Sigma) or βIII-tubulin (1∶250; SDL.3D10, Sigma) followed by goat anti-mouse IgM-FITC (1∶200; sc-2082, Santa Cruz), chicken anti-goat IgG-FITC (1∶200; sc-2988, Santa cruz), goat anti-mouse IgG-FITC (1∶200; sc-2010, Santa Cruz), Alexa Fluor 594 goat anti-rabbit IgG (1∶750; A11012, Invitrogen) secondary antibodies. Nuclei were stained with Hoechst 33258. ## Bisulfite Sequencing 500 ng genomic DNA was treated with the Epitect Bisulfite Kit (Qiagen) or Epimark Bisulfite Conversion Kit (NEB) according to the manufacturer’s instructions. A 206 bp region of the endogenous rat Oct4 promoter (−1495 to −1290) was amplified by PCR from bisulfite converted genomic DNA using primers BS-Oct4_F and BS-Oct4_R. PCR was performed with GoTaq DNA polymerase (Promega). Thermal cycling conditions were: 94°C, 2 min; 35 cycles of 94°C for 30 s, 55°C for 30 s, 72°C for 1 min; then final elongation 72°C for 5 min. PCR fragments were subcloned into the vector pJet1.2/blunt (Fermentas) and the DNA sequence of five individual clones determined. Bisulfite sequencing data were analyzed with the online tool QUMA. ## Karyotype Analysis Rat iPS cells in log phase were treated with 10 µg/ml colcemid for 4 h. Cells were collected, treated with Accutase to obtain a single cell suspension, incubated for 12 min at room temperature in 75 mM KCl and fixed with ice cold methanol/acetic acid (3∶1). Metaphase preparation and chromosome counting was performed by CHROM*Bios* GmbH (Nussdorf, Germany). ## Embryoid Body (EB) Formation Embryoid bodies were generated either by growth in suspension, or “colony EB” culture. For suspension culture, iPS cells were dissociated with Accutase, resuspended at 4×10⁶ cells per 15 ml EB medium I (50% N2B27-2i, 50% DMEM+) and cultured in 10 cm non-adhesive culture dishes. For colony EB culture, loosely attached iPS colonies were flushed off the feeder layer and transferred into 10 cm non-adhesive culture dishes in EB medium I. For both methods, the medium was changed to EB medium II (30% N2B27-2i, 70% DMEM+) after 48 h. A further 48 h later, medium was changed to DMEM+ and EBs cultured for an additional 4 days in non-adhesive culture dishes. After 8 days EBs were analyzed or allowed to attach to gelatin-coated tissue culture plates in DMEM+ medium. ## Teratoma Formation 4–5×10⁶ rat iPS cells from line T1/64 were resuspended in N2B27-2i, mixed with high density Matrigel (BD Bioscience) and injected subcutaneously into NOD scid gamma (NSG) mice. Teratomas were harvested after 25 days, fixed in 4% paraformaldehyde, embedded in paraffin and sectioned. Sections were stained with hematoxylin and eosin (H&E) according to standard protocols. ## Transfection of Rat iPS Cells Rat iPS cells were transfected with Nanofectin (PAA), or Lipofectamine 2000 (Invitrogen) as monolayer cultures on 2% Geltrex (Invitrogen) in 12 well plates according to the manufacturer’s instructions using the GFP expression plasmid pmaxGFP (Lonza). Nucleofection was performed using the Nucleofector II device (Lonza) and the Mouse Embryonic Stem Cell Kit (Lonza) with program A-024 according to the manufacturer’s instructions. ## Production of Recombinant NLS-Cherry-9R Protein and Protein Transduction The expression vector pTriEx-Cherry encodes the red fluorescent protein NLS- Cherry-9R. NLS-Cherry-9R contains a 6xHis tag, the SV40 Large-T nuclear localization signal (NLS) at the N-terminus and a protein transduction domain consisting of 9 arginine residues (9R) at the C-terminus of the mCherry red fluorescent protein. The pTriEx-Cherry expression cassette was assembled by standard PCR methods. Recognition sites for the restriction enzyme *BspHI* and *XhoI*, the 6xHis tag, NLS sequence and 9R were added to the mCherry fluorescent reporter gene (Clontech) coding region with long oligonucleotides using Phusion Polymerase (Finnzymes). The mCherry translational start codon was mutated from ATG to GTG, which also encodes the C-terminal valine of the NLS to avoid translation of untagged protein. The cassette was inserted between the *NcoI* and *XhoI* restriction sites of pTriEx-HTNC (Addgene plasmid 13763) to generate pTriEx-Cherry. Expression in bacteria and purification of NLS-Cherry-9R was performed according to. Protein transduction was performed with iPS cells on MEF feeder cells, in suspension culture in 15 ml Falcon tubes, or in monolayer culture on 2% Geltrex using 5 µM recombinant protein for 4 or 24 h. # Results ## Generation of Doxycycline-dependent Rat iPS Cells Two cell types from two rat strains were used to generate iPS cells: adipose tissue-derived mesenchymal stem cells (rADMSC) and ear fibroblasts (rEF) from Fischer and Wistar rats. Cells were cotransfected with the reprogramming vector pReproII-attB and the φC31 integrase expression vector pCAG-C31Int(NLS). pReproII-attB contains the minimal bidirectional doxycycline-inducible promoter P_bi-1, which directs expression of the murine reprogramming factors Oct4, c-Myc, Klf4 and Sox2 as bicistronic mRNAs. The pReproII-attB vector also encodes other necessary components of the Tet-On system: the tetracycline- controlled transactivator rtTA2(S)-M2 and the tetracycline-regulated repressor tTS^KRAB under the control of the constitutive CAG promoter. An attB site is also included to facilitate φC31 integrase-mediated integration at pseudo-attP sites in the host genome. We found that double nucleofection (on days 0 and 3) resulted in more efficient generation of rat iPS cells than a single nucleofection step. Six days after initial nucleofection, cells were transferred onto a feeder layer in N2B27-3i medium containing doxycycline to induce expression of the exogenous reprogramming factors. Colonies with rat ES/iPS cell-like morphology appeared 8 to 10 days after doxycycline induction. Individual colonies were manually picked between day 14 and 20, transferred to multiwell plates and expanded. The outline scheme is shown in. A total of 37 iPS cell lines were established from Fischer rat ADMSCs and EFs and 18 lines from Wistar rat ADMSCs and EFs. Results and efficiencies are summarized in. At this stage, growth of each line relied on continued presence of doxycycline and expression of exogenous reprogramming factors. ## Analysis of Vector Integration We used a non-viral vector equipped with an attB phage φC31 integrase recognition site to facilitate integration into the rat genome. Three preferred integration sites, so called pseudo-attP sites rps1, rps2 and rps3, have been identified in the rat genome. Two Fischer iPS cell lines, T1 and T13, and two Wistar iPS cell lines, E9 and E14, were analyzed by Southern blot to determine the number of vector integrations in each. We identified two (T1, T13) or four (E9, E14) integrations as shown in. To determine whether these were at known pseudo-attP sites (rps) we performed PCR analysis similar to that previously described. φC31-mediated integration can occur in either forward or reverse orientation, we therefore analyzed integration in both directions. None were found at sites rps1 or rps3. Screening of rps2 sites revealed integration in reverse orientation in lines T1 and T13. The wild-type fragment spanning rps1, 2 and 3 was amplified in all lines, indicating that the rps2 integration in lines T1 and T13 had occurred at one allele. Sequence analysis of the PCR fragment from lines T1 and T13 confirmed vector integration into the rps2 site on chromosome 1q41. The other integrations did not occur at the known three sites. ## Generation of Doxycycline-independent Rat iPS Cell Lines Two doxycycline-dependent iPS lines from Fischer ADMSCs (T1, T13), and two from Wistar EF (E9, E14) were used for further experiments. Initially iPS cells were cultured in N2B27-3i medium with doxycycline for 14 days, then doxycycline induction withdrawn. 263 colonies were isolated between days 14 and 24 after doxycycline removal from lines T1 and T13 and expanded in N2B27-3i medium. 10 doxycycline-independent iPS cell lines were established. This rather low efficiency (\<4%) led us to investigate alternative culture media. Twelve media compositions were tested, these differed in the presence or absence of Knockout Serum Replacement, inhibitor combinations, addition of Thiazovivin, ROCK inhibitor Y27632 or ascorbic acid. Best results were obtained with N2B27-2i medium, which differs from N2B27-3i in that it lacks ALK5 inhibitor A83-01 and Knockout Serum Replacement. Using N2B27-2i, 15 of 40 (37.5%) colonies from Fischer line T13 and 39 of 114 colonies from Wistar lines E9 and E14 (22.2% to 36.5%) were established as doxycycline-independent iPS cell lines, representing a considerable improvement over N2B27-3i. Detailed results and efficiencies are shown in. This, together with reduced spontaneous differentiation and a more stable ES-like cell morphology led us to use N2B27-2i medium for further culture. Examples of cell morphology before reprogramming and after doxycycline removal are shown in. Comparison of our reprogramming efficiencies with rat iPS cells generated by viral methods or by φC31 integrase-based reprogramming of murine and human cells is shown in. ## Expression of Exogenous Reprogramming Factors in Doxycycline-dependent and -Independent Rat iPS Cell Lines We compared the expression of exogenous reprogramming factors in the parental doxycycline-dependent rat iPS cell lines (T1, T13, E9 and E14) in the presence of doxycycline, with six doxycycline-independent subclones (T1/64, T13/3, E9/19, E9/54, E14/5 and E14/15). Quantitative RT-PCR specific for bicistronic Oct4-T2A-c-Myc or Sox2-T2A-Klf4 mRNAs showed high exogenous factor expression in doxycycline-dependent lines, and 100- to 1,000,000-fold less in doxycycline- independent subclones, see. This, together with an undifferentiated morphology, indicated successful activation of the endogenous pluripotency program. ## Pluripotency Markers, Methylation Status and Karyotype Analysis We analyzed the expression of endogenous pluripotency marker genes in two doxycycline-independent iPS cell lines: Fischer iPS line T1/64 and Wistar iPS line E9/54. Oct4, Sox2, Nanog, FGF4 and Rex1 expression was analyzed by RT-PCR using rat-specific primers. As shows, both iPS lines expressed these key markers, while the source rADMSC and rEF cells did not. Both lines also expressed the stem cell marker alkaline phosphatase (AP). Immunocytochemical analysis revealed that both lines T1/64 and E9/54 were Oct4,SSEA1 positive and SSEA4 negative, which accords with previous descriptions of rat ES and iPS cells. Source cells were Oct4, SSEA1, SSEA4 and AP negative. Another hallmark of successful reprogramming is the demethylation of promoter regions of key pluripotency genes such as Oct4 and Nanog. We performed bisulfite sequencing of a 206 bp region (−1495 to −1290) of the rat Oct4 promoter, which contains 7 CpG sites. In rADMSCs 6 to 7 CpG sites were methylated, in rEFs 3 to 5 were methylated, whereas none were methylated in iPS cell lines T1/64 and E9/54. This further confirmed reprogramming to a self-sustaining pluripotent state. We then performed karyotype analysis. Rat iPS cell line T1/64 showed a normal karyotype (2n = 42) in 80% metaphase spreads examined. Line E9/54 was mostly polyploid. Diploid metaphase spreads were analyzed and 20% showed normal karyotype (2n = 42). These results accord with previous reports of rat ES and iPS cell karyotypic variability. ## Differentiation of Rat iPS Cells *in vitro* and *in vivo* Rat iPS cell differentiation was investigated *in vitro* by embryoid body (EB) formation, and *in vivo* by teratoma formation in NSG mice. EB procedures developed for mouse ES cells were unsuccessful when applied unmodified to rat iPS cells. Rat iPS cells cultured as single cells in suspension in medium containing serum typically died after two days. We therefore developed an alternative protocol. Cell suspensions were cultured in 50% DMEM+, 50% N2B27-2i medium (EB medium I) for two days, during which cells spontaneously aggregated. Medium was changed to 70% DMEM+, 30% N2B27-2i (EB medium II) for two days, then to DMEM+. Under these conditions rat iPS cells reliably generated EBs. These findings are consistent with reports that the GSK3β inhibitor CHIR99021 is necessary for EB formation in rat ES cells. We also developed a technique, termed colony EB culture, that produced EBs with high efficiency by taking advantage of the poor attachment of undifferentiated rat iPS cells. Colonies 100 to 200 µm in diameter were flushed off the feeder layer, resuspended in EB medium I and cultured further as described above. This improved cell survival, but produced EBs of diverse size. We used colony EB culture with iPS cell line T1/64 and E9/54 and analyzed expression of genes characteristic of the three embryonic germ layers by RT-PCR. After 8 days in suspension, mesodermal (Nkx2.5, Flk1, SM22a), endodermal (Sox17, Gata4, Gata6) and ectodermal (Nestin, NCAM) marker genes were expressed, whereas Oct4 and Nanog expression was markedly reduced. EBs were allowed to attach to gelatin-coated plates, cultured for 12 days in medium containing serum (DMEM+) and outgrowths examined. Immunocytochemical analysis revealed differentiation to neurons as detected by expression of βIII-tubulin, cardiomyocytes by sarcomeric α-actinin, and hepatocytes by albumin, see. Undifferentiated rat iPS cells did not express these markers, see. The ability of rat iPS cells to generate differentiated tissues within a teratoma was tested by injecting undifferentiated T1/64 cells into NSG immunodeficient mice. All injection sites generated tumors up to 1.5 cm in diameter after 25 days. One was sectioned and examined histologically. Staining with H&E revealed complex organized structures and identifiable derivatives of the three embryonic germ layers, including intestinal epithelium and pancreatic cells (endoderm), cartilage and blood vessels (mesoderm), also neural rosettes and epidermis (ectoderm), as shown in. These data demonstrate that our rat iPS cells are pluripotent. ## Rat iPS Cells in Feeder-free Monolayer Culture We investigated whether our rat iPS cells could be cultured as monolayers without feeder cells. This would allow more control over differentiation than EB based methods. Rat iPS and ES cells are known to attach poorly even on feeder cells, but success has been reported with rat ES cells on laminin coated plates. We compared laminin, bovine gelatin, collagen type I, fibronectin, and the Engelbreth-Holm-Swarm tumor basement membrane extracts Matrigel and Geltrex. No attachment was observed on gelatin, collagen type I or fibronectin. In contrast, rat iPS cells attached to Matrigel, Geltrex or laminin coated plates, proliferated and formed colonies of morphologically undifferentiated cells. Cells on Matrigel, Geltrex and laminin were fixed after 5 to 7 days and characterized by immunocytochemistry and alkaline phosphatase (AP) staining. Similar to iPS cells on feeder layers, they were positive for Oct4, SSEA1, AP and negative for SSEA4. Immunocytochemical detection of differentiation markers albumin, βIII-tubulin and sarcomeric α-actinin was negative, confirming that rat iPS cells on Matrigel, Geltrex and laminin remain undifferentiated (see). Cumulative cell numbers and population doubling times were compared between cells grown on Matrigel, Geltrex, laminin and on feeder cells. No differences were observed after 48 h, but after 96 h iPS cell numbers were higher on Matrigel, Geltrex and laminin (mean 1.75×10⁶ cells per well) than on feeders (1.17×10⁶ cells per well). Doubling times based on exponential cell growth were 16.4 h (laminin), 19.8 h (Matrigel), 20.6 h (Geltrex) and 27.8 h (feeders). ## Transfection and Protein Transduction of Rat iPS Cells The ability of pluripotent cells to undergo genetic and other manipulations in culture is fundamental to their practical usefulness. We therefore assessed rat iPS cells for their ability to undergo DNA transfection by three different methods: Nanofection and Lipofection of feeder-free monolayers and Nucleofection of suspended cells. A GFP reporter plasmid was used as a convenient indicator. Transient transfection efficiency determined after 24 h showed more than 10% of cells transfected with Lipofectamine 2000, ∼9% with Nanofectin and ∼4% with Nucleofection, see. Cells were replated onto feeders two days after transfection and no change in iPS morphology was observed. Lipofectamine 2000 and Nucleofection slightly reduced cell viability, but Nanofectin had no detectable effect. We investigated if our rat iPS cells can be efficiently transduced with recombinant proteins. This would for example allow direct delivery of recombinases such as Cre, or transcription factors for directed differentiation without genetic manipulation. A cell penetrating red fluorescent protein, NLS- Cherry-9R, was generated by fusing the mCherry coding region to a protein transduction domain consisting of 9 arginine residues (9R) and the SV40 Large-T nuclear localization signal (NLS),. NLS-Cherry-9R was expressed in bacteria, and rat iPS cells transduced for either 4 or 24 hours on feeder cells, in suspension or as feeder-free monolayers. As shown in, 24 hours transduction in suspension or monolayer culture resulted in up to 80% red fluorescent cells. These experiments demonstrate that our rat iPS cells are amenable to genetic and non-genetic manipulation, enabling the use of a wide range of molecular tools. # Discussion We describe a significant advance in rat iPS technology, the efficient generation of rat iPS cells using a single non-viral vector that allows tight control over reprogramming factor expression. The established iPS lines were self-sustaining and had activated the endogenous pluripotency program. Methods have been developed to improve rat iPS cell viability and successful generation of differentiated iPS derivatives *in vitro*. Teratoma formation *in vivo* and differentiation *in vitro* demonstrate pluripotency. Our rat iPS cells can be cultured as monolayers free of feeder cells, at least for short periods, and readily undergo DNA transfection and protein transduction. To date rat iPS cells have been generated using retroviral or lentiviral vectors based on the original breakthrough by Yamanaka. However, the viral approach has shortcomings. To circumvent the need for multiple viral infections, Hamanaka *et al.* (2011) developed a polycistronic inducible lentiviral system that encodes all factors. This enables control of factor expression and introduces individual factors in equivalent ratio. Nevertheless lentiviral transduction still requires high viral titers for efficient transduction and viral production must be under strict biosafety conditions. Our plasmid based reprogramming approach avoids these issues. A single vector contains the reprogramming factors under doxycycline-inducible control and all necessary Tet-On regulatory components. Because transfected DNA integrates into the host genome less frequently than infecting retro- or lentiviruses, an attB site was included to facilitate φC31 integrase-mediated integration. Cotransfection of the reprogramming vector with φC31 integrase enabled efficient generation of doxycycline-dependent rat iPS cells. Direct comparison of the efficiency of our approach with that of others is not always straightforward because many reports base their calculations on different parameters, such as the number of stable iPS cell lines, alkaline phosphatase positive colonies, SSEA1 positive colonies or cell morphology. Reprogramming efficiency, calculated as the number of stable iPS cell lines obtained relative to starting cell number, reveals a range of 0.00174% to 0.014% in previous reports. We obtained 0.0027% to 0.0028% for doxycycline-dependent and 0.005% to 0.0078% for doxycycline-independent rat iPS cells. This compares favorably with methods based on viral vectors. Generation of self-sustaining rat iPS cells after doxycycline removal depended on the culture medium. N2B27-2i medium was superior to N2B27-3i at this stage, highlighting a negative influence of the ALK5 inhibitor A83-01 or Knockout Serum Replacement on iPS generation. This indicates that one or both reduces colony formation, maintenance and expansion and accords with previous reports that N2B27-2i conditions are suitable for rat ES and iPS culture. Addition of Thiazovivin, ROCK inhibitor Y27632 or ascorbic acid did not increase efficiency compared to N2B27-3i or N2B27-2i. Traditional methods of EB formation such as hanging drops, or aggregation in single cell suspension are difficult with rat pluripotent stem cells. Our colony EB method is a significant improvement that reduces cell death and allows reliable generation of EBs without the need for additional small molecules such as the ROCK inhibitor Y27632. We also investigated appropriate conditions for feeder-free monolayer culture and found that rat iPS cells attached to Matrigel, Geltrex and laminin substrates. Previously only attachment to laminin has been shown. Rat iPS cells grown on Matrigel, Geltrex or laminin maintained expression of Oct4, SSEA1 and alkaline phosphatase. Cumulative cell numbers were similar on the three substrates but slight differences were observed in population doubling time. The increased range of coating matrices opens new possibilities for feeder-free iPS cell culture, transfection and cell differentiation without EB formation. Future applications such as gene targeting require genetic manipulation, preferably free of feeder cells. The established monolayer conditions enabled transfection with Lipofectamine 2000 and Nanofectin, resulting in up to 10% transfection efficiency, useful alternatives to the most commonly used techniques, nucleofection and electroporation. Directed differentiation of iPS cells may require the addition of inductive transcription factors for only a short time. Transduction of protein rather than DNA is a useful means of achieving this, and also leaves the genome unaltered. Our rat iPS cells can be efficiently transduced with recombinant proteins in suspension or monolayer culture. Rat iPS cells can thus be cultured on feeder cells, transfected/transduced in monolayer culture and then transferred back onto feeder cells as required. We have demonstrated that rat iPS cells can be generated by a safe and simple non-viral approach. Improved methods for differentiation and the use of coating substrates for monolayer culture will facilitate derivation of many different cell types from rat iPS and ES cells, a goal not yet achieved. These will provide valuable resources for diverse biomedical research. # Supporting Information The authors would like to thank Peggy Mueller and Margret Bahnweg for excellent technical assistance with cell culture and molecular biology, Jana Tretter and Krzysztof Flisikowski for assistance with qPCR, Jessica Haas for expression of recombinant NLS-Cherry-9R protein and Michael Hennig for support of the project. [^1]: AI is an employee of Small Molecules Research - Discovery Technologies, Pharma Research and Early Development, F. Hoffmann-La Roche Ltd. - a pharmaceutical company involved in biomedical research and manufacture of medical products. This work was supported, in part, by F. Hoffmann-La Roche Ltd. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: CM A. Saalfrank AI A. Schnieke. Performed the experiments: CM A. Saalfrank NR SE MH. Analyzed the data: CM A. Saalfrank NR SE DS. Contributed reagents/materials/analysis tools: AP RK WW. Wrote the paper: CM AK A. Schnieke. [^3]: Current address: Stem Cell Platform, F. Hoffmann-La Roche Ltd., Basle, Switzerland

# Introduction Consumption of a high fat diet is associated with obesity, insulin resistance and the development of chronic health conditions, such as atherosclerosis and Type II diabetes. The rising rates of obesity in some populations have been referred to as an “epidemic”, and there is indication that current trends may worsen in the future. The deleterious health effects associated with obesity have been linked to several metabolic abnormalities, but increasingly, it has been recognized that inflammation is a key process that contributes to negative health outcomes. High fat diets induce a state of heightened inflammation, with elevated levels of C-reactive protein, interleukin-6, and other systemic inflammation markers. Additionally, strong localized inflammation responses to obesity are known to occur within individual tissues. In white adipose tissue, for example, high fat diet can increase macrophage infiltration into fat depots, while also promoting infiltration by T-lymphocytes and macrophages. In cardiac tissue, localized recruitment of monocytes is an initiating step in the development of atherosclerotic plaques, which drives development of heart disease in association with high fat diet. In other tissues, including liver, muscle and pancreas, high fat diet can drive excessive accumulation of lipids and their derivatives, leading to a state of lipotoxicity that augments pro- inflammatory signals to promote accumulation and activation of macrophages. High fat diet and obesity are thus associated with systemic inflammation as well as localized inflammatory responses that affect both adipose and non-adipose tissues. This inflammation response may represent a useful target for intervention strategies aimed at combating deleterious health effects of high fat diet and obesity. Inhibitors of the monocyte chemoattractant protein 1 (MCP-1/CCR2) pathway, for instance, have been shown to attenuate insulin resistance and other negative consequences of excessive energy intake in laboratory rodents. The inflammation response associated with high fat diet is accompanied by shifts in tissue composition, activation of apoptotic pathways, disruption of lipid metabolism, and altered sensitivity to insulin. Recently, the investigation of these complex effects has been facilitated by a large-scale study of hepatic gene expression patterns in male and female mice of 12 mouse strains, in which whole-genome microarrays were used to profile expression in mice that had been maintained on either a high fat (30% fat) or control diet (6% fat) for a period of four weeks. This unique dataset, which includes whole-genome expression patterns from 144 mice, provides a valuable resource for the investigation of gene expression patterns associated with a single dietary intervention. Mechanistic studies of high fat diet in mice have often focused on a single mouse strain, usually C57BL/6J, which is known to exhibit a response to high fat diet that is at least partly idiosyncratic. The data resource provided by Shockley et al. have greatly expanded this focus, providing a tool for identifying aspects of the response to high fat diet that are shared among multiple strains, as well as aspects that are specific to individual strains. An initial analysis of these data highlighted salient features of the hepatic response to high fat diet, including responses that are invariant among all strains and mice, regardless of gender (e.g., increased *Abcg5* expression). Interestingly, a group of 557 immune-response genes was identified, which were both induced by high fat diet in multiple strains and also correlated with total cholesterol levels. Such genes were disproportionately associated with antigen processing and presentation, and included a number of histocompatibility antigens (e.g., *H2-Ab1*, *H2-Eb1*, *H2-Aa* and *Cd74*). It is possible that, for some immune-associated genes, elevated hepatic expression is a localized response of hepatocytes to damage associated with high fat diet. Alternatively, increased expression of such transcripts in liver may reflect an influx of monocytes, macrophages, T-cells and other white blood cells, which express at high levels genes encoding antigens and proteins central to immune processes. This latter possibility suggests that, in mice fed a high fat diet, hepatic gene expression patterns can be used to gauge inflammation intensity, and that targeted analytical methods can be developed to exploit this information to gain insight into the tissue remodeling that accompanies excessive intake of dietary fat. This study provides a microarray-based characterization of hepatic inflammation responses to high fat diet in male and female mice of 12 different mouse strains. A data-mining algorithm is developed and applied, which leads to the *in silico* calculation of “inflammation profiles” that highlight leukocyte subsets best able to explain gene expression patterns associated with high fat diet. The algorithm utilizes genome-wide expression profiles for leukocyte populations, which are available in public microarray data depositories, to identify sets of “signature transcripts” that represent a molecular fingerprint for particular populations of leukocytes (e.g., CD4+ T cells). For each set of signature transcripts identified, the group-wise response of the set to high fat diet in liver is evaluated, and this response is used to infer whether the associated leukocyte population is likely to infiltrate liver tissue of mice provided a high fat diet. This approach leads to inferences with extremely strong statistical support, and in the present context, generates mechanistic hypotheses useful for building a model of hepatic response to high fat diet, and for understanding how this response is similar and different among mouse strains. The analytical approach developed in this study can also be applied in other *in vivo* settings to better understand inflammation processes on the basis of microarray data. # Results ## Overview of the transcriptional response to high fat diet in mouse liver Microarray data from the Novartis strain-diet-sex survey (GSE10493) was used to evaluate hepatic responses to high fat (HF) diet in male and female mice of 12 mouse strains (129S1/SvImJ, A/J, C57BL/6J, BALB/cJ, C3H/HeJ, CAST/EiJ, DBA/2J, I/LnJ, MRL/MpJ-Tnfrs6lpr/J, NZB/BINJ, PERA/Ei, SM/J). We compared transcript levels between HF-fed (*n* = 3) and control-fed mice (*n* = 3) for each of 24 strain-gender combinations, where HF-fed mice received 30% Kcal from fat over 4 weeks and control-fed mice received 6% Kcal from fat over the same time period. Response to HF diet varied considerably in magnitude among the 24 strain-gender combinations. In some cases, following FDR adjustment for multiple hypothesis testing, more than 1000 unique genes were either increased or decreased by HF diet (e.g., NZB males and females, A/J females; see). In other cases, fewer than 50 unique genes were increased or decreased by HF diet (e.g., females of the SM, DBA, B6 and 129 strains; see). Cluster analysis identified two groups among the strain-gender combinations, with one group exhibiting lower response to HF diet (mostly females; see), and the other group exhibiting a heightened response to HF diet (mostly males; see). In some cases, the “strain effect” appeared to dominate the “gender effect”, and males and females of the same strain clustered together (e.g., see NZB, A, PERA, I, CAST), although this was not true for some strains (e.g., see females from the BALB, 129 and MRL strains). Gene ontology analyses revealed that biological processes associated with immune responses (e.g., antigen processing and presentation, defense response to Gram- positive bacteria, activated T cell proliferation, chemotaxis and cell adhesion) were the most frequently overrepresented gene ontology terms among the genes increased by HF diet. For instance, in males, genes increased by HF diet were disproportionately associated with antigen processing and presentation in 9 of the 12 mouse strains. Among females, the trend was less strong, and genes increased by HF diet were disproportionately associated with antigen processing and presentation in 5 of 12 mouse strains, although other frequently overrepresented processes included neutrophil chemotaxis (7 of 12 strains) and inflammatory response (7 of 12 strains). These results are in agreement with those presented by Shockley et al., and suggest that increased expression of genes involved in immune system processes is a major feature of the hepatic response to HF diet, which occurs robustly in male and female mice and among multiple mouse strains. ## An *in silico* hepatic inflammation profile associated with high fat diet in B6 Males We hypothesized that, among transcripts elevated by HF diet, over-abundance of transcripts related to immune system processes was at least partly due to an inflammatory response, in which circulating leukocyte populations infiltrate hepatic tissue of mice provided a HF diet, thereby driving increased expression of genes that exhibit high expression in leukocytes. To evaluate this possibility, we developed a data mining procedure, which searches among gene expression profiles of leukocyte populations and their subsets, identifies signature transcripts for each population, and then evaluates whether such signature transcripts are disproportionately elevated in hepatic tissue of mice provided a HF diet. The approach leads to the generation of an *in silico* “inflammation profile” associated with response to HF diet, which provides indication of the degree to which effects of HF diet may be attributable to leukocyte infiltration, and also points to particular cell populations that are likely to be prominent components of the inflammatory infiltrate. The method was first applied to calculate an inflammation profile associated with response to HF diet in B6 male mice, which is perhaps the most commonly investigated strain and gender in physiological studies of excessive dietary fat intake. This revealed that, in B6 males, signature transcripts associated with many leukocyte populations are disproportionately elevated in hepatic tissue of B6 males provided a HF diet, including CD4+ and CD8+ T cells, B cells, NK cells, dendritic cells (DCs), macrophages (Mφ), neutrophils, monocytes, granulocytes and natural helper cells. Among all cell types, however, those from the monocyte- Mφ and monocyte-DC lineages scored most highly. For instance, the strongest trend was present among signature transcripts associated with a population of CD8A- myeloid dendritic cells (Gene Expression Omnibus samples GSM258647 and GSM258648). With respect to this population, the procedure identified a total of 454 signature transcripts, and nearly all of these were elevated in mice provided a HF diet relative to mice provided a control diet. In particular, of the 454 signature transcripts, 446 were increased by HF diet while only 8 were decreased (ratio = 446/8 = 55.7; FDR-adjusted P = 7.15×10⁻¹⁶¹). Moreover, 37 of the 454 signature transcripts were increased significantly (FDR-adjusted P\<0.05), while none were significantly decreased (FDR-adjusted P = 7.71×10⁻²⁴). Transcripts most strongly associated with this dendritic cell population included *H2-Aa*, *Cytip* and *Ms4a4c*, and each of these transcripts was elevated in hepatic tissue of mice provided a HF diet. These results provided strong indication that, as a group, DC-associated transcripts are disproportionately elevated in liver of B6 males provided a HF diet, consistent with the hypothesis that leukocyte populations, particularly DCs, infiltrate hepatic tissue of B6 males as part of an inflammatory response that accompanies excessive fat intake. ## *In silico* hepatic inflammation profiles associated with high fat diet in male and female mice of 12 mouse strains The above analyses focus on B6 male mice, but the most interesting aspect of the Novartis strain-diet-sex survey data is the possibility of examining how patterns compare across a diversity of mouse strains and between both genders. We thus calculated a complete inflammation profile for each of the other 23 strain-gender combinations. A summary of all inflammation profiles is provided in. This analysis revealed a number of consistencies among the strain-gender combinations, as well as certain aberrant patterns. The most consistent pattern associated with HF diet was increased expression of transcripts associated with bone-marrow derived Mφ. This effect was detected with respect to all strain-gender combinations, with the exception of DBA males (see below). Aside from bone-marrow derived Mφ, HF diet also increased transcripts associated with thioglycollate-elicited peritoneal Mφ, conventional DCs, regulatory T cells and white adipose tissue. There was strong correspondence between the magnitude of inflammation profile ratios (HF- increased/HF-decreased) and the clustering patterns observed in. In particular, ratios were especially large for NZB mice (males and females), A/J mice (males and females), 129 males and B6 males, and in, these strain-gender combinations clustered together in a single group. This result suggests that the genome-wide response to HF diet is correspondent with, and perhaps partly dependent upon, the leukocyte-infiltration signatures detected by inflammation profiles. The single cell population that, across all strains and genders, consistently scored most highly on inflammation profiles was a population of thioglycollate- elicited peritoneal Mφ (Gene Expression Omnibus samples 258701 and 258702). For this Mφ population, 570 signature transcripts were identified (e.g., *Mmp12*, *Gpnmb*, *Bex1* and *Il7r*), and among the 24 strain-gender combinations, HF diet usually increased 70–90% of these signature transcripts in liver. For 23 of 24 strain-gender combinations, the proportion of the 570 transcripts increased by HF diet was statistically significant. An extreme case was NZB/BINJ females, for which 553 of the 570 signature transcripts (97.01%) were increased by HF diet, with 406 signature transcripts exhibiting a statistically significant increase (FDR-adjusted P\<0.05). At the other end of the spectrum, transcripts associated with thioglycollate-elicited peritoneal Mφ were not strongly elevated by HF diet in DBA mice. In DBA females, only 57.37% of signature transcripts (327 of 570) were elevated by HF diet. In DBA males, however, 61.2% of signature transcripts (349 of 570) were *decreased* by HF diet, and this proportion of HF-*decreased* transcripts was statistically significant (FDR-adjusted P = 1.57×10⁻⁸;). There was considerable variation in terms of the degree to which inflammation profile results correlated between males and females of the same mouse strain. For some strains, leukocyte populations that scored highly in males and also tended to score highly in females, with an overall correlation between inflammation profile ratios that exceeded 0.90 (e.g., strains A, C3H, I and K; ). On the other hand, for the DBA and SM mouse strains, the correlation between males and females was relatively low (0.16 and 0.36, respectively), suggesting potential gender differences in the degree and type of hepatic inflammation that develops in response to HF diet. ## DBA mice exhibit resistance to inflammatory gene expression patterns associated with high fat diet The DBA mice were aberrant relative to other mouse strains and, based upon inflammation profiles, there was little or no indication that HF diet strongly elevated expression of leukocyte-associated transcripts. With regard to the 570 transcripts associated with thioglycollate-elicited peritoneal Mφ (see above;), there was no correlation between DBA and B6 males in terms of how these transcripts responded to HF diet (*r* = −0.051; see). A close inspection of inflammation profiles calculated for DBA mice reveals further aspects of the strain that are unique. In DBA females, ratios of HF-increased to HF-decreased signature transcripts were much lower relative to all other strain-gender combinations (HF-increased/HF-decreased ratio\<2;). In DBA males, trends were especially striking, with leukocyte-associated transcripts exhibiting a response to HF diet that was opposite to that in other mouse strains. For instance, 282 signature transcripts were identified with respect to one CD8+ T cell population, and 208 (73.8%) of these transcripts were decreased in hepatic tissue of DBA males provided the HF diet (FDR-adjusted P = 4.15×10⁻¹⁵;). These results suggest that DBA mice, and particularly DBA males, exhibit resistance to the elevated expression of leukocyte-associated transcripts with HF diet that is characteristic of other mouse strains included in our analysis. ## The gene expression phenotype of high-scoring cell populations: Toll-like receptor signaling as a potential mediator of hepatic infiltration The ability of a cell population to infiltrate hepatic tissue may depend upon expression of key receptors that facilitate homing to liver, attachment to endothelial surface, or transendothelial migration into the tissue. We therefore characterized the gene expression phenotype of high-scoring cell populations to identify potential unifying characteristics. Populations with high scores on inflammation profiles were more likely to highly express transcripts associated with the toll-like receptor (TLR) signaling pathway, degradation of glycan structures, leukocyte transendothelial migration, Jak-STAT signaling, cytokine- cytokine receptor interaction, and cell adhesion molecules. The most robust characteristic of high-scoring cell populations was elevated expression of genes associated with TLR signaling (e.g., *Tlr2*, *Tlr13*, *Irf5*, *Myd88*, *Il10rb*). For 19 of the 24 strain-gender combinations that we evaluated, transcripts associated with TLR signaling were overrepresented among the 200 transcripts for which expression was most strongly correlated with population scores on inflammation profiles. The *Irf5* gene, for example, encodes a transcription factor (interferon regulatory factor 5) that is activated downstream of the TLR-Myd88 signaling pathway. Our analyses revealed that, with respect to C3H females as well as males of the 129, B6 and I strains, expression of *Irf5* in leukocyte populations was a better predictor of inflammation profile score than any other single transcript. This trend was, in part, attributable to the high expression of *Irf5* in cell types usually assigned high scores on inflammation profiles (e.g., Mφ, DCs and monocytes; see). Nevertheless, there were some Mφ and monocyte populations with low inflammation profile scores, and in these populations, *Irf5* expression was comparatively low. Likewise, among other high-scoring populations, apart from those associated with the monocyte-Mφ and monocyte-DC lineages, expression of *Irf5* expression was comparatively high. These results indicate that high expression of *Irf5* (and other genes associated with TLR signaling) is a common characteristic of cell populations that appear to infiltrate hepatic tissue in mice provided a HF diet. ## Expression of macrophage-associated transcripts is correlated with total cholesterol and other phenotypic characteristics We hypothesized that leukocyte-associated transcripts associated with high scoring populations from inflammation profiles may be linked to certain phenotypic characteristics that are sensitive to HF diet (e.g., cholesterol, triglycerides, glucose). For instance, as shown in, the transcript *H2-Aa* was a signature transcript of CD8A- myeloid dendritic cells, and previously, Shockley et al. reported that this and other antigen-associated transcripts exhibited a strong correlation with total cholesterol levels among 120 mice evaluated in the Novartis strain-diet-sex study. To determine if this was a general characteristic of leukocyte-associated transcripts, we focused on the 570 signature transcripts associated with the consistently high-scoring population of thioglycollate-elicited peritoneal Mφ, and evaluated whether these transcripts tended to have a strong correlation with total cholesterol levels. As predicted, there was an unusually large correlation between the expression of these 570 signature transcripts and total cholesterol levels. On average among the 570 transcripts, the correlation with total cholesterol was 0.556, which was significantly larger than correlations observed for other probe sets represented on the array (*r* = −0.079; P = 7.71×10⁻²⁹¹ based on t-test comparison; see). Correlations were especially strong for those transcripts that were most increased by HF diet. We next evaluated average correlations among the 570 transcripts with respect to other phenotypic characteristics (mean *r* = 0.349, high density lipoprotein; mean *r* = 0.314, glutamate dehydrogenase; mean *r* = −0.280, triglycerides; mean *r* = −0.200, glucose; mean *r* = −0.173, nonesterified fatty acids; mean *r* = 0.136, calcium; mean *r* = 0.109, blood urine nitrogen; mean *r* = 0.097, body weight), and in each case the average correlation differed significantly relative to all those represented on the array platform (P\<2.07×10⁻⁵⁶). These analyses indicate that hepatic expression of Mφ-associated transcripts is correlated with multiple phenotypic measures that are linked to pathological aspects of the HF diet. ## Greater than 50% of high fat-increased transcripts can be explained by hepatic infiltration of monocytes, macrophages or dendritic cells in some mouse strains What proportion of transcripts elevated by HF diet in liver might be attributable to infiltration by monocytes, Mφ, or DCs as part of an inflammatory response? To address this question, we evaluated the 500 transcripts most strongly increased by HF diet for each of the 24 strain-gender combinations. For each strain-gender combination, we evaluated the top 500 transcripts individually, and for a given transcript, we determined whether that transcript had been designated as a signature transcript of a high-scoring (statistically significant) cell population in the inflammation profile calculated for the corresponding strain-gender combination. If the transcript was indeed a signature transcript of a significant population, that transcript was “assigned” to that cell population, and this served as an explanation for why that transcript was elevated by HF diet in liver. In males, the proportion of HF- increased transcripts explained by monocytes, Mφ, or DCs varied between 12.4% (MRL males) and 65.6% (CAST males). An exception was DBA males, for which none of the top 500 transcripts were explained on this basis (because no significant cell populations were identified in the inflammation profile for DBA male mice; see). In females, we estimated that between 7.2% (DBA females) and 64.8% (CAST females) of the top 500 transcripts were, potentially, explainable on the basis of monocytes, Mφ, or DC infiltration. # Discussion High fat diets can promote systemic inflammation and may lead to a metabolic state associated with negative long-term health outcomes. Microarray experiments have provided a systems-level view of the major effects of high fat (HF) diet within individual tissues, and have shown that HF diet often increases the expression of genes disproportionately associated with immune system processes, such as the processing and presentation of antigens,. We hypothesized that, in liver, this effect of HF diet is due to infiltration of hepatic tissue by white blood cells, and we have applied an algorithm to test whether this hypothesis is consistent with data from the Novartis strain-diet-sex survey (GSE10493). This hypothesis was supported by our analyses and we show that “signature transcripts” highly expressed in certain leukocyte populations (e.g., thioglycollate-elicited peritoneal Mφ) are often overwhelmingly elevated in hepatic tissue of mice provided the high fat diet. This effect, which appears to reflect the degree of localized hepatic inflammation, has a genetic component and differs in intensity among males and females of 12 mouse strains. In some strains (NZB male and female mice, B6 males and A/J females), we estimate that 50–60% of genes elevated by HF diet can be explained on the basis of leukocyte infiltration. On the other hand, in DBA males and females, there was little or no indication that hepatic gene expression patterns associated with HF diet were shaped by leukocyte infiltration. These results, taken together, suggest that inflammatory processes and resultant tissue remodeling partly explain genome- wide expression patterns associated with HF diet, that microarray data can be exploited to gain biological insights into these patterns, and that the intensity of hepatic inflammation in response to HF diet is genetically regulated and heterogeneous among inbred mouse strains. The HF diet perturbs the hepatocyte-leukocyte balance that normally exists within the liver microenvironment, and there are multiple (non-mutually exclusive) models that may explain how the inflammatory cascade is initiated by HF diet to ultimately promote an influx of leukocytes into the liver. Healthy liver contains a balance of hepatocytes and other intrahepatic cell populations, including resident macrophages (Kupffer cells), in addition to immune cells entering from circulation. This balance can be altered by triggers that increase generation of chemoattractant compounds by hepatocytes to enhance the recruitment of immune cells, or by cellular modifications intrinsic to circulating leukocyte populations that increase their capacity for adhesion or migration into endothelial tissues. In the present context, *in silico* inflammation profiles have provided a tool for gauging the relative intensity of inflammatory gene expression in different mouse strains, and also provide an unbiased method for identifying populations that appear most likely to explain observed effects of HF diet on hepatic gene expression. For most mouse strains, there was strong evidence that HF diet increased the hepatic expression of transcripts highly expressed in bone marrow-derived cells from the monocyte- Mφ and monocyte-DC lineages. This inference is based upon marked and statistically significant patterns in the data (e.g., see ; FDR-adjusted P-value less than 10⁻¹⁵⁰), and is also in agreement with previous immunohistochemical investigations, which have shown that HF diet increases abundance of these cell types in mice of multiple genetic backgrounds. We characterized a gene expression phenotype associated with high-scoring cell populations, with the expectation that the capacity of cells to home to liver tissue and attach to endothelial substrates would depend upon expression of receptors or key cellular components. This analysis revealed that a unifying characteristic of cell populations assigned high scores on inflammation profiles was high expression of genes encoding components of the toll-like receptor (TLR) signaling pathway (e.g., *Irf5* and *Myd88*). Our findings, therefore, add to the growing recognition of this pathway as a contributor to obesity-associated inflammation,, and suggest that heightened expression of TLR genes in several mouse strains may predispose leukocytes to respond to *in vivo* signals that arise due to excessive fat intake (e.g., elevated fatty acids). This possibility is consistent with findings from recent studies, which have shown that deletion of toll-like receptor 4 (*Tlr4*) in hematopoietic cells abrogates HF diet- associated inflammation in liver and adipose tissue, and that liver damage due to adoptive transfer of immature myeloid cells (CD11b⁺Ly6C^hiLy6G⁻) from HF-fed mice occurs only when donor cells carry the gene encoding the Myd88 intracellular adaptor protein. Development of hepatic inflammation depends upon several parallel mechanisms and the interplay between circulating cytokines and concurrent processes in both hepatocytes and nonparenchymal cells. For instance, HF diet may stimulate hepatocytes to increase local production of monocyte chemoattractant protein-1 (MCP-1/CCL2), which serves as a chemoattractant signal that would draw monocytes from circulation, which could then differentiate locally into Mφ or DC progeny cells. Indeed, inspection of the Novartis strain-diet-sex database reveals that, on average among all 24 strain-gender combinations, expression of *Mcp-1*/*Ccl2* is increased by 58% (range: 4% decrease in SM females; 3-fold increase in NZB males and A females; significant increase in 16 of 24 cases). This strengthening of a chemoattractant gradient could, moreover, be accompanied by increased fractional abundance of the circulating inflammatory Ly-6C^hi monocyte subset, which express specialized ligands that facilitate tethering of monocytes to endothelial substrates. Hepatic inflammation may also be heighted by damage to liver tissue associated with HF diet. The hepatic blood supply from the portal vein, for example, is highly sensitive to alterations in the absorption capacity of the intestinal mucosa. Along these lines, systemically high cytokine levels associated with HF diet may disrupt tight junction complexes in the intestine to compromise intestinal permeability, ultimately leading to elevated endotoxin levels in the portal blood supply, which may damage liver tissue and promote release of inflammatory compounds by hepatic stellate cells and Kupffer cells. Lastly, given a HF diet, inflammatory signals emitted by interstitial fat deposits in hepatic tissue may be enhanced, which would serve to tighten the association between such fat deposits and white blood cells. Inflammation profiles, for instance, suggested increased presence of T cell-associated transcripts in liver of HF-fed mice of most strains, and this could arise from a closer association with intra-hepatic adipose tissue and T lymphocytes,. At the same time, signature transcripts of white adipose tissue were also disproportionately elevated by HF diet, suggesting that some degree of fat expansion occurred with HF diet, which could augment still further the abundance of adipose-associated T cell subsets and proinflammatory signals derived from adipose, leading to accelerated lymphocyte recruitment and an overall reinforcement of the inflammatory cascade. It is increasingly recognized that genotype is an important factor shaping response to dietary intervention in mice. To understand such genotype-by- environment interactions, the diversity of existing mouse strains provides a valuable research tool. Following 4 weeks of the HF diet, nearly all strains exhibited increased abundance of leukocyte-associated transcripts in liver, indicative of localized inflammation, but this response was markedly attenuated in DBA mice. In DBA males, transcripts associated with certain leukocyte populations were disproportionately *decreased* by HF diet (e.g., CD8+ T cells, B cells, DCs and Mφ), providing evidence that the diet had possible anti- inflammatory effects in mice of this strain and gender. This observation is supported by an independent analysis recently reported by Zhu et al., which showed that after 1–21 weeks of HF diet (0.5% cholic acid, 1.25% cholesterol, 15% fat), hepatic tissue of female DBA mice exhibited decreased expression of genes associated with “immune response and inflammation”, while the opposite pattern was observed in B6 mice. Additionally, Zhu et al. noted that, relative to B6 mice, apoptosis of hepatic cells was reduced in DBA females. The present study, therefore, in combination with the results of Zhu et al., suggests that for HF diets of 1–21 weeks in duration, DBA mice of both genders exhibit some level of resistance to hepatic inflammation. It is uncertain whether this response of DBA mice to HF diet is strictly liver-specific, or whether it is indicative of an aberrant pattern by which systemic inflammation progresses in this mouse strain. Previously, Kirk et al. compared nine inbred mouse strains and showed that, in DBA mice, plasma cholesterol levels were unusually hyporesponsive to multiple HF diets, which may have been due to decreased absorption of dietary cholesterol,. In other studies, however, in which slightly different HF diets were evaluated, plasma cholesterol levels have been observed to increase in DBA mice. DBA mice also appear susceptible to increased body weight when provided a HF diet. A recent study, in fact, evaluated 42 inbred mouse strains provided an atherogenic diet for 18 weeks, and showed that DBA mice gained more body weight than any other strain. Lastly, DBA mice may exhibit a differential response to healthy diets, such as caloric restriction, which is a well-studied anti-inflammatory diet known to increase lifespan in most inbred mouse strains. In particular, Forster et al. reported a series of experiments in which long-term caloric restriction increased lifespan in B6 mice and B6D2F1 hybrids, but led to a slight lifespan *decrease* in mice of the DBA strain. The present analysis indicates that, following 4 weeks of HF diet, inflammatory gene expression patterns in hepatic tissue of different mouse strains can vary. It is important to note, however, that these observations correspond to a fixed time point following the start of HF diet (i.e., 4 weeks), and that development of hepatic inflammation may proceed along a time course, with a more acute short-term phase followed by a longer-term chronic phase in which inflammation is partly attenuated. A recent study of hepatic gene expression data in Apo3Leiden mice (B6 background), for example, has indicated that an acute inflammatory response to HF diet develops in the short-term (1 day–1 week), but that this response partly attenuates between 8 to 16 weeks with the progression of steatosis. In our analysis, the 4 weeks of HF diet is neither a short-term or long-term response, but can be viewed as an intermediate time point, which may lie at the transition between distinct phases of the hepatic or systemic responses to HF diet. Potentially, differences between mouse strains with respect to inflammation intensity that we observed may not represent variations in the intensity of hepatic inflammation developing in each strain, but rather HF-response patterns that are temporally out-of-step between strains. In the case of DBA mice, one possibility is that hepatic inflammation occurs more rapidly relative to other strains, such that after 4 weeks of HF diet, DBA mice are at an advanced and less acute inflammatory stage and thus falsely appear to exhibit resistance to hepatic inflammation. However, the study of Zhu et al. (see above) argues against this possibility, since Zhu et al. showed that inflammatory gene expression in DBA females was hypo-responsive to HF diet across a range of time points between 1 and 21 weeks. Nevertheless, we note that data analyzed in this analysis are specific to the 4-week time point, and thus do not permit evaluation of the temporal progression of HF-associated inflammatory gene expression patterns among the 12 inbred strains evaluated. The generation of a more comprehensive dataset, involving a diversity of strains and a time series of gene expression profiles, would be a challenging task, but would provide a basis for understanding how progression of hepatic inflammation in response to HF diet may differ among inbred mouse strains. This study has provided a targeted analysis of hepatic gene expression patterns in HF-fed and control mice, and results are consistent with the hypothesis that hepatic infiltration of white blood cells is a robust response to HF diet in mice that occurs in multiple strains and both genders. These findings provide a reference point for future studies investigating effects of HF diet in various mouse strains and both genders. Further work should be directed at evaluating whether HF diet promotes a similar microarray-based inflammation signature in non-hepatic tissues, evaluating how strongly inflammation signatures correlate with results generated from immunohistochemical analyses, and determining whether supplementation of HF diets with cholic acid augments inflammatory processes. While our results indicate that development of hepatic inflammation is a robust response to 4 weeks of HF diet, we have also shown that this response is not universal and appears attenuated in DBA mice. This result highlights the strain-specificity of dietary responses, which challenges efforts to construct general models of dietary response that will have applicability to multiple mouse strains, to other species, and potentially, to humans. For this reason, further generation of comprehensive multi-strain datasets is needed to ensure that conceptual models of dietary response are not genotype-specific or dependent upon the properties of a single mouse genotype. Finally, findings from this study demonstrate that a complex inflammatory process, involving shifts in cellular composition combined with altered transcription within cells, is indeed amenable to computational analysis guided by a biological rationale. This *in silico* strategy is likely to be equally useful in other contexts in which gene expression patterns are the cumulative product of intracellular processes and a broader inflammatory response (e.g., cancer, neurodegeneration, psoriasis, atherosclerosis, aging). # Methods ## Novartis strain-diet-sex microarray database The effects of high fat (HF) diet were analyzed using the Novartis strain-diet- sex survey database, which can be obtained from Gene Expression Omnibus under accession numberGSE10493. The complete dataset, consisting of 144 CEL files, was downloaded and expression scores were subsequently calculated using the robust multichip average (RMA) method. We note that a web-based query tool for exploring these data has been made available online at <http://cgd- array.jax/org/DietStrainSurvey>. The experimental procedures associated with animal care and tissue processing have been described by Shockley et al.. In brief, for HF treatments, mice were provided a diet with 30% Kcal from dairy fat, which contained 1% cholesterol by weight and 0.5% cholic acid. For control treatments, mice were provided with a standard diet containing only 6% fat (Cat. No. 5K52, Lab Diets, St. Louis, MO). Animals were housed in specific pathogen free facilities prior to sacrifice at 10–13 weeks of age (20–55 weeks of age in the case of CAST/EiJ and PERA/Ei mice). Mice were fasted approximately 5 hours before sacrifice and perfused with saline before dissection. Phenotypic data collected from mice utilized in these experiments can be accessed and downloaded from <http://cgd.jax.org/datasets/expression/10strain.shtml>. Further details regarding general phenotypic characteristics of each mouse strain can be obtained from the Mouse Phenome Database (MPD) and a convenient webpage with MPD links for each of the 12 mouse strains considered in our analysis is available (see: <http://phenome.jax.org/db/q?rtn=projects/strainlist&projsym=GX- Shockley1>). ## Algorithm for calculation of inflammation profiles Inflammation profiles were calculated for each of 12 mouse strains and both genders (i.e., 24 strain-gender combinations). Inflammation profiles are an *in silico* device that can be used to interpret gene expression patterns, and in the present context, are used to gauge intensity of inflammation in response to HF diet, and to infer which types of leukocytes best explain hepatic gene expression differences between HF-fed and control mice. The overall process is summarized in. In step 1, a set of *n* signature transcripts is identified for each of *j* = 1,…, *N* leukocyte populations. This step required two sets of reference microarray data (i.e., treatments *A* and *B*), both of which correspond to data obtained from Gene Expression Omnibus, which are further described in and. All data samples listed in and were generated using the Affymetrix 430 2.0 oligonucleotide microarray, which is the same platform utilized in the Novartis strain-diet-sex survey. The first set of reference data (i.e., treatment *A*) corresponds to a batch of *n*_A = 65 CEL files obtained from GEO, where each CEL file was generated from an array hybridization with source material isolated from mouse liver.The second set of reference data (i.e., treatment *B*) corresponds to a batch of *n_B* CEL files from the *j*th leukocyte population. For each leukocyte population evaluated, there were at least two replicate samples available (i.e., *n_B*≥2), and on average, there were 2.80 replicates available per population. To identify *n* signature transcripts for any one leukocyte population, *n*_B CEL files associated with that population were jointly normalized with the liver reference dataset (i.e., the *n*_A = 65 CEL files). The LIMMA algorithm (linear models for microarray data) was then used to identify transcripts differentially expressed between the leukocyte population and the liver reference data. We utilized a high threshold for the identification of transcripts with significantly higher expression in the leukocyte population (i.e., FDR-adjusted P-value\<10⁻⁴ and fold-change≥16), with FDR- adjustment of p-values carried out using the Benjamini-Hochberg method. Given this threshold, there was, on average, *n* = 594 signature transcripts identified for each leukocyte population evaluated in our analysis (range: 152≤*n*≤1311). The set of *n* signature transcripts is then analyzed, in step 2, to determine if the set is disproportionately elevated by HF diet in liver tissue. If a leukocyte population invades hepatic tissue with HF diet, the *n*₁ signature transcripts increased by HF diet should greatly exceed the *n*₂ signature transcripts decreased by HF diet, and the ratio *n*₁/*n*₂ should be large (where *n* = *n*₁+*n*₂). Inflammation profiles show, for each strain-gender combination, the ratio *n*₁/*n*₂ generated from steps 1 and 2 (as shown). However, it is expected that, for any given population, some of the *n* signature transcripts will also be signature transcripts of other populations. For instance, there were 19,900 pairwise combinations among the *N* = 200 cell populations evaluated in our analysis. For each pairwise combination, we evaluated the overlap of signature transcripts and found that the median level of overlap, among all pairwise combinations, was 25.39% (1st quartile: 19.54%; 3rd quartile: 32.41%). Such overlap of signature transcripts can be problematic, since it may be that a given cell population does not invade hepatic tissue of HF-fed mice, but nevertheless, the cell type shares many signature transcripts with another population that actually does invade hepatic tissue of HF-fed mice. In this case, the non-invading population may have a large *n*₁/*n*₂ ratio, and thus appear to invade hepatic tissue, when in fact it does not. To address this potential problem, we also evaluate a second ratio for each population, *n*_1\*/*n*_2\*, which is generated from steps 3 and 4. The ratio *n*_1\*/*n*_2\* is calculated based upon a reduced set of *n*\* signature transcripts, which for each population, is obtained by filtering the *n* signature transcripts, in order to exclude any transcripts that are also signature transcripts for another population with a larger *n*₁/*n*₂ ratio (i.e., step 3). The *n*\* signature transcripts therefore represent transcripts highly expressed in a particular leukocyte population (relative to liver), which are uniquely expressed in that population relative to any other populations for which the ratio *n*₁/*n*₂ is larger. For nearly all populations, therefore, the value of *n*\* is less than *n*, with the exception of the highest-ranked population (with the largest *n*₁/*n*₂ ratio), for which no sub-setting can be performed and *n* = *n*\*. The advantage of considering both *n*₁/*n*₂ and *n*_1\*/*n*_2\* ratios is that, for populations not invading hepatic tissue but sharing many signature transcripts with other populations that do invade hepatic tissue, the ratio *n*₁/*n*₂ would be large but the ratio *n*_1\*/*n*_2\* would not be. The consideration of both *n*₁/*n*₂ and *n*_1\*/*n*_2\* ratios to evaluate statistical significance (see below) thus serves to de-correlate patterns of statistical significance and to reduce the chance that some populations will spuriously emerge as significant because their signature transcripts overlap with those of another high-scoring population. ## Leukocyte populations evaluated in inflammation profiles Most of the *N* = 200 populations evaluated in our procedure were white blood cell populations. For convenience, we have collectively referred to these as “leukocyte populations” throughout this paper. We note, however, that not all populations represented among the 200 we evaluate are in fact leukocytes. Some populations correspond to RNA extracted from certain types of progenitor cells isolated from bone marrow, which are precursors to circulating blood cells (e.g., hematopoietic stem cells, common lymphoid progenitors, erythroblasts; see ). In other cases, populations correspond to RNA extracted from whole organs (e.g., thymus, spleen, white adipose tissue; see). These populations were evaluated in parallel with true white blood cells subsets because of their role in leukocyte development (e.g., thymus, spleen, progenitors). In the case of adipose tissue, it was suspected that HF diet could augment adipose tissue deposits in liver, and thus promote elevated expression of adipose-associated transcripts by a mechanism comparable to the elevated expression of immune- associated transcripts due to white blood cell infiltration. ## Statistical significance criteria for inflammation profiles The inflammation profiles presented in our analysis include red symbols that correspond to populations for which statistical significance criteria were satisfied, along with black symbols, which represent populations for which significance criteria were not satisfied. There were two criteria for statistical significance. First, the ratio *n*₁/*n*₂ needed to be significantly large, as compared to the ratio of HF-increased to HF-decreased transcripts observed among all 45,101 transcripts represented on the Affymetrix 430 Mouse Genome 2.0 array. The significance of this ratio was evaluated using a hypergeometric test to determine the likelihood of observing *n*₁ HF-increased transcripts within a sample of *n*₁+*n*₂ transcripts sampled from all those evaluated on the Affymetrix 430 Mouse Genome 2.0 array. P-values generated from this test were then adjusted for multiple testing among all 200 populations using the Hochberg method, which is a conservative p-value adjustment method that is valid for the case in which p-values are non-negatively correlated. We note that, while the value *n*₁ is the test statistic considered by the hypergeometric test, the value of *n*₁ is directly proportional to *n*₁/*n*₂, and thus we view this approach as a test of the *n*₁/*n*₂ ratio. The second criteria for statistical significance was the same test, except applied to the ratio *n*_1\*/*n*_2\* rather than *n*₁/*n*₂. For this test, the hypergeometric distribution was used to determine, for each population, whether the observed value of *n*_1\* was significantly large, given the null scenario in which a random sample of *n*_1\*+*n*_2\* transcripts are chosen from all those evaluated on the array platform. As above, p-values generated from this test were adjusted using the conservative Hochberg correction to account for multiple testing among the 200 leukocyte populations. For inflammation profiles presented in this study, red symbols correspond to populations for which both hypergeometric tests (as applied to *n*₁/*n*₂ and *n*_1\*/*n*_2\*) are significant (FDR-adjusted P\<0.05). The main advantage of using the hypergeometric distribution to evaluate significance of *n*₁/*n*₂ and *n*_1\*/*n*_2\* ratios is that the approach is straightforward and does not require the repeated simulation of a null distribution for every leukocyte population considered, which would have been computationally expensive in the present context. We note, however, that our algorithm could be adapted to utilize alternative statistical measures of correspondence between leukocyte-associated and HF-associated gene expression patterns, such as the test statistic used in the gene set enrichment analysis (GSEA) algorithm proposed by Subramanian et al.. # Supporting Information We acknowledge and thank investigators that have contributed their microarray data to the Gene Expression Omnibus database. Two reviewers provided helpful comments on this manuscript. [^1]: Analyzed the data: WRS. Wrote the paper: WRS AJ JG. [^2]: The authors have declared that no competing interests exist.

# Introduction Foot and ankle function are compromised when an individual suffers from diabetes mellitus (DM), and foot and ankle functioning can have costly outcomes if not prevented or treated. According to the International Working Group on the Diabetic Foot, strategies have been suggested to encourage foot care and self-management, in addition to using therapeutic footwear when severe diabetic polyneuropathy (DPN) is present. However, although it is believed that muscle weakness and joint limitations in DM and DPN patients are irreversible, specific therapeutic foot and ankle exercises may contribute to preventing and controlling musculoskeletal and structural deficits that may affect foot function and balance and increase the risk for ulcers if not treated\[–\]. Up to now, there have been 10 clinical trials–some of them with low risk of bias–that have demonstrated the beneficial effects of foot and ankle–related exercises for improving DPN symptoms and sensory deficits and for reducing peak plantar pressure\[–\], in addition, these studies have showed the ability for patients to improve foot–ankle range of motion\[,–\] and foot–ankle muscle strength and functioning. Supervised and unsupervised therapeutic foot-related exercises performed by patients with low and moderate neuropathy have been shown to reduce plantar pressure distribution during gait\[,–\]. Likewise, in one study, a personalised therapeutic exercise protocol was followed for 12 weeks to rehabilitate small joints and foot–ankle muscles. The results showed satisfactory changes in gait biomechanics with an improved distribution of plantar pressure, resulting in a better physiological pattern in foot–ankle rollover. Besides exercises, educational and self-care actions are essential for preventing late consequences and help patients identify clinical situations earlier before a late diagnosis with complications, such as DPN^,. The use of technology by health providers has not only improved patient monitoring and adherence, but has also reduced the demands on healthcare facilities. A number of studies with DM patients have been conducted using e-health technologies that allow people to engage in activities in their preferred environment, thereby taking up less of the health professional’s time and decreasing demands on health centres; they provide a means for people to better monitor themselves, having them depend less on face-to-face care and reducing human and financial costs. So far, e-health technologies used for diabetes have focused on general, whole- body exercises or have had other purposes, such as glucose monitoring. Software for specific foot and ankle exercises is not available, and mostly, these programmes have not been able to personalise the exercises progression following the user’s individual physical capacities. In this context, the current study aimed to develop and validate free web software that can be accessed through computers or smartphones, here targeting people with DM and low or moderate DPN; this had the potential of enabling self-management and customised care through a personalised foot–ankle exercise routine. # Methods ## Software development and structure We used an application layer that provides online services accessed through the World Wide Web or through mobile applications (Android and iOS). The system was developed using hypertext markup language and JavaScript for the interface and for the usability of the hypertext preprocessor tool employed to analyse user data. Structured query language (SQL) was used for the database and MySQL version 5.0.51, as well as SQLITE for systems management because it requires few hardware resources. For the user requirements, an HTML 5/CSS 3 compatible browser is required and must be able to navigate in a web environment with a minimum resolution of 1200 x 780 pixels. To use this web software version, no installation is required: the user only has to enter the link [www.usp.br/labimph/soped](http://www.usp.br/labimph/soped). To run the app, the user needs to have Android 4.3 or higher and to download the application from the same link, in the ‘download our application’. No operating system requirement is made. In both cases, users need Internet access. The software is available in English and Portuguese, but it can be translated into any language. The software was created with the intention of being used independently by a person with DM at his or her own convenience, but also has the potential to be a tool that facilitates primary and secondary health services worldwide. The software is in its first version. We intend to revise it when including other languages while keeping the software free and public. All access codes and algorithms used in the software are available as supplemental material. In the development of this software (**Figs and**), three main aspects were considered: (i) foot care recommendations and information about DM and DPN; (ii) self-assessment of feet according to the main foot alterations of DM and DPN (calluses, cracks, deformities and soft-tissue lesions, among others); and (iii) customised foot–ankle exercises to strengthen muscles, increase range of motion and improve functionality. ## Website sections and features The main sections and features of the website include the following: - Informative webpages about foot care recommendations and the disease’s main complications. - Self-assessment of the feet, which is included to encourage users to assess their feet regularly and stimulate an investigation of their health conditions. Some validated instruments were included: Michigan Neuropathy Screening Instrument for the self-assessment of DPN signs and symptoms; Foot Health Status Questionnaire; and a brief investigation of fall occurrences. To guarantee the users’ physical integrity, a physical examination of the user’s feet was included to check for the presence of foot alterations/deformations, such as calluses, cracks, mycoses, deformed toes, ulcers and amputations. A general feedback of the user’s health status is provided, based on the answers given. If there are any signs or symptoms of severe health conditions, a clear recommendation is made to seek for medical assistance, such as severe polyneuropathy, increased risk for falls and bad foot health status. Users with preulcerative lesions are referred to contact a foot care specialist. - A custom exercises section was made available only after the user had responded to all self-assessments in the software. The programme will personalise the exercises progression, according to each individual physical capacity. An effort scale that the user fulfil will determine the progression or not to other levels of difficulties. - Because of its potential to increase engagement, gamification components were employed throughout the user environment to encourage and motivate the patients to use the tool. The user’s panel was designed with dynamic features and with game functions to stimulate the users to practice the exercises and navigate throughout the software. Information about diabetes and a physical examination of the feet come with a 2D animation and interactive menu. We inserted a reward system for the completion of each step of the software and after the sessions and progression of the exercises. Even if real progression in exercise difficulty had not occurred, users would be rewarded for their dedication and persistence, not just for their physical ability. Details of the reward system is presented in. - Possibility of interactions between users and researchers. A forum section was included to facilitate the exchange of information. ## Exercise protocol A therapeutic exercise protocol was developed to provide autonomy for the individuals during exercise without the need for professional supervision. The protocol is simple, contains clear written instructions (as well as video and audio) and preserves the safety of the target population during exercise. Furthermore, it establishes the training volume, progression criteria and guidelines for discontinuing the protocol. A singular feature of this tool is that it personalises the progress of a foot–ankle exercise programme based on individual capabilities, which is similar to conventional physiotherapy. To include this feature, we incorporated a visual analogue scale, which is represented by a ruler that quantifies the level of effort required to perform each exercise so that daily progress can be customised based on one’s results. To personalise the exercises, a progression algorithm was created from the perceived effort of each user, who could progress in the exercise programme’s difficulty, maintain it or return to the previous stage in accordance with the following criteria: with a 0.0 to 2.0 score in the visual scale, the user progresses to the next level of effort on the following day; from 2.1 to 7.0, the user advances to the next level after 2 days at the present level; from 7.1 to 10, the user returns to the previous level. The physiotherapeutic foot–ankle exercise protocol is based on previous clinical trials. It was designed following three criteria established in a supervised, face-to-face intervention: (a) muscle stretching (20 exercises); (b) strengthening of the intrinsic foot muscles (40 exercises); and (c) strengthening of the extrinsic foot–ankle muscles and functional exercises such as balance and gait training (44 exercises). In total, 39 different exercises were chosen, and when including their sublevels of progression, a total of 104 different exercises can be completed. For each session, only eight exercises are combined to provide the three previously described criteria. The exercises are recommended to be performed twice or three times weekly. To avoid monotony and enhance motivation, the exercises always change from session to session, and the maximum duration of a session is no longer than 20 minutes. Those features also prevent the users from doing excessive effort, because they also limit the uncontrolled progression, as explained: the exercises should only be done twice or three times a week; no more than eight exercises each day is allowed; and the individual difficulty is regulated by the effort scale. Some exercises have sublevels that correspond to increases in the load, number of repetitions or time duration. However, each exercise is different and may contain only one level or up to five levels of difficulty. For each exercise, the user attributes the effort, and the progression is made for this exercise. So if a session is composed of eight exercises, the user may progress in two of these exercises but may stay at the same level for six of the exercises in the next session. Therefore, individual physical capacities are respected, and one exercise will not block the progression of other exercises that are easier to perform. Therefore, the user’s overall progression is not classified as levels but rather follows the reward system of trophies and items presented in. The following muscle groups are targeted in the protocol: medial-plantar aspect: abductor hallucis, flexor halluces brevis and adductor hallucis; lateral plantar aspect: abductor digiti minimi, flexor digiti minimi brevis and opponens digiti minimi; middle-plantar aspect: flexor digitorum brevis, quadratus plantae, lumbrical muscles, plantar interosseous and dorsal interosseous muscles; dorsal- foot aspect: extensor digitorum brevis and extensor halluces brevis. The following joints are targeted in the protocol: talocrural, tarsometatarsals, interphalangeals and metatarsophalangeals. ## Tool validation The Delphi method was used for validation. The process occurred concurrently with a judging panel of 20 people with DM and another panel of nine health professionals specialising in treating people with DM and DPN. The judges had access to the desktop and mobile versions of the software, and their responses were given for both applications. The judges used the software for a period between 30 and 45 days, twice or three times a week. This time period was chosen because at least a usage of 30 days represent one third of the entire protocol, and is representative enough of the functionality of the software. During this time, judges could properly verify all sessions and features of the SOPED that are: fulfill personal information; read all the instructions and information about the disease; fulfill the feet physical inspection; read and practice many exercises; use the effort scale in different situations and with different exercises; verify the progression, maintenance or regression of the exercises according to their individual capacities; use the forum; receive the rewards; verify the safety information; and navigate through different sessions of the SOPED. In addition, all the judges received an attached file with detailed descriptions of all the exercises included in the software, whit their respective training volume and progression, and therefore could analyze all the protocol without the need to perform every exercises for three months. The panel of users comprised six men (30%) and 14 women (70%) with a mean age of 41.4 years (21–65 years), mean DM diagnosis of 14 years (1–33 years) and normal cognitive performance as assessed by the Mini Mental State Examination (score media of 28.5). Depression level was assessed using Beck’s Depression Inventory, in which one of 20 subjects exhibited a score of 27 (moderate level), six a score of more than 4 (minimum level) and four a score of more than 12 (mild level). Educational level was 20% (4/20) for high school and 80% (16/20) for university level. In addition, 80% (16/20) of the patients were working. There were 12 patients with Diabetes but without neuropathy, confirmed by the Michigan Neuropathy Screening Instrument (MNSI) and 8 patients with previous confirmed diagnostic of neuropathy (MNSI and physical examination from the database of the research center). Other eligibility criteria for users were the following: both sexes: having type 1 or type 2 DM; 18 years old or more: free of tissue injury at the time of execution of the exercises; be able to use the software alone; or have someone to help at all times; and having completed schooling equal to or higher than the fourth year of elementary school. The panel of specialists comprised nine women, including one psychologist, three physiotherapists (two specialised in clinical diabetes care, one specialised in biomechanics and musculoskeletal function of diabetic neuropathy), one podiatrist nurse (specialised in podiatric care and diabetic foot), one physician/endocrinologist (specialised in diabetic foot), one occupational therapist and two physical education professionals (one specialised in clinical and gait analysis of diabetic foot). The mean age was 45.44 years (35–59 years), and the mean experience in the treatment of people with DM was 18.8 years. For specialist selection, résumés were assessed using an adaptation of the Fehring criteria, which allows for a minimum score of 5 points. The mean adapted Fehring score was 9.7 out of 14. The other eligibility criteria for the specialists are described in and are based on adapted Fehring criteria. The first round of assessment consisted of a questionnaire containing 16 items that was based on a 5-point primar scale (I completely agree, I agree, I neither agree nor disagree, I disagree and I totally disagree) in which the comments considered important to each member of the panel were obtained for each item. The instrument that was built to discern the specialists’ opinions addressed the following matters: objective of the web software; fitness of the language to the population; amount and quality of information; contribution of exercises to decreasing foot deficits caused by DM; and whether the tool promoted daily exercise. The instrument that was built to elicit the user’s opinion addressed the same matters, but also aimed to determine whether the user correctly understood the exercise performance. For the first round, both suggestions from specialists and users were evaluated only by the researchers, and their suggestions and recommended changes were incorporated into the software by the researchers. After incorporating the suggestions, the new version of the software was submitted again, in a second round, to the same panel members who evaluated the modifications in the software. At this stage, the changes should be approved or not, until a minimum of 70% was reached, and then the final version of the web software was determined. ## Statistical analysis The data were analysed using descriptive statistics, means, relative and absolute frequencies and the content validation index (CVI). The CVI measures the proportion of items that the judges are in agreement. The content validity is determined by the proportion of judges that score items as being relevant or representative. That correspond to a score of 4 or 5 (‘agree’ and ‘strongly agree’) on a Likert scale out of all possible answers (the others being ‘disagree’, ‘strongly disagree’, and ‘don’t agree nor disagree’). The score is calculated by the sum of the agreement of the items marked 4 or 5. The CVI was calculated only for the first round. For the final validation, after the second round, we used a 70% approval consensus criterion for all modifications implemented in the software. ## Ethical approval This research was approved by the Research Ethics Committee of the School of Medicine of the University of São Paulo (Approval No. 2.262.357). Informed consent was obtained from all participants included in the design process. # Results In the first round of the Delphi technique, we obtained an accurate and satisfactory result regarding all aspects of the web software that were queried, all of which showed a high degree of agreement. shows that 90.3% of the specialists agreed with the web software particulars, 5.6% neither agreed nor disagreed, and only 4.2% disagreed with some aspects. The results indicate that the initial version achieved satisfactory CVIs, given that all the CVIs of the individual items (n = 16) obtained values of more than 0.78, except for item 13, the question on accessibility (which was influenced by the presence of retinopathy in users, a common complication of persons with DM). We included an explanatory audio during the exercise-video demonstration to facilitate accessibility. The overall CVI of the first round was 0.902. A similar result in the first round was observed in the assessment conducted by patients with DM, in which 89.7% agreed, 5.9% neither agreed nor disagreed, and only 4.4% disagreed with some aspects. The CVI results of the individual items showed values higher than 0.78 for 14 of the 16 items in the instrument, and the overall CVI was 0.903. Although there was a high degree of agreement in the first round, the panels suggested a number of changes that were incorporated, producing the first version of SOPED–Educational Diabetic Foot Software. All the suggestions made by specialists and users were analysed by the researchers and were incorporated and implemented in the software. This new version of the software was presented once again to the panels in the second round of the validation process. Some of the suggestions could not be implemented but were justified and resubmitted to the panel for their approval. In this second round, the judges could only ‘approve’ or ‘disapprove’ of the changes. In general, the main suggestions involved an indication of the sequence of steps on the homepage. The starting point was unclear, so we made it more evident where to begin. Also, the health professionals questioned the usability and correct interpretation of the effort scale. We added a clearer explanation in the tutorial and also next to the scale icon. In the second round of the web software assessment, one health professional and five users with DM dropped out of the study, but we still had a sufficient number of judges. Therefore, eight diabetes specialists and 15 anonymous panel members with diabetes approved or disapproved of the changes. None of the items were rejected by either group (given the approval index of 70%). We obtained an approval rate of 100% from the DM specialists and 97% from the software users with DM, as detailed in. Consequently, no further rounds were needed. # Discussion SOPED was developed and validated with a high degree of agreement between DM specialists and people with DM at 100% and 97%, respectively. The software allows self-management and personalised care for patients, which is recommended by international consensus. The main feature and innovation of the software is the customised treatment that respects the physical capacities of each user by ensuring that the software users recognise the purpose and importance of completing the appropriate effort scale. One concern that had an average disapproval rating from health professionals of 15% and from people with diabetes of 22% was the lack of audio in the videos, which made it difficult or impossible for people with visual problems to use the software (42.9% of people with DM have some type of retinopathy). It was important to make the software intuitive and easy to use because diabetes is prevalent in people who are 20–79 years old. SOPED encompasses some recommendations given by the International Consensus on the Diabetic Foot (2015) for the care and prevention of diabetic foot complications: (1) inspect and examine the affected foot; (2) identify the affected foot; and (3) educate people, family and/or caregivers and health professionals. In accordance with previous recommendations for the prevention of foot ulcers, as a safety function, we added in the periodic examination (every 30 days) of feet to assess tissue integrity. It is mandatory for the continued use of the tool, and access is blocked if the user exhibits any preulcerative signs, such as sores, blisters or a developing ulcer. Similarly, users are not allowed upon the first use of the software to see the exercise instructions if they present any sign of tissue damage. If preulcerative signs are present, a clear recommendation is made to seek urgent medical care. For future versions of SOPED, daily inspection will be recommended in a more evident way, besides the already given recommendation that is included in the instructions. Since many aspects of the feet can change after initiating the exercises, a quick questionnaire can be included just before starting the next session. To avoid a repetitive and monotonous exercise sequence, each session was designed to be conducted in a short period of time. The variation in exercises was also planned with gamification concepts. The main component of the gamification aspects was the system created to reward each successful exercise execution, regardless of individual physical capacity. Despite the questions raised and discussed in an attempt to increase adherence to the software, it will be important to conduct an intervention with the target population to analyse whether the stimuli will be effective for improving foot–ankle mobility and functionality and strengthening foot–ankle muscles. There is also a need to verify the long-term effectiveness of the proposed exercises in a controlled, randomised clinical trial. Nevertheless, positive results are expected because the effect of this type of intervention has already been proven to be efficient in promoting changes in DPN-related deficits. This tool complements the traditional recommended interventions of foot inspection, podiatric care, shoes and prescriptions and can be suggested by ay health professional because of its multiprofessional characteristic. The software was designed to be used at health centres as a self-explanatory tool validated by professionals from various areas, hence making its use interdisciplinary. The final version of SOPED is available at the following link: \<<http://www.usp.br/labimph/soped/>\> for the desktop and also to download the mobile application. # Conclusion SOPED was developed based on scientific evidence and on a high level of agreement between health experts and users with diabetes. SOPED can be recommended by an interdisciplinary team and is a free preventive model that can be implemented in primary and secondary care as a complementary treatment for DPN. Further steps to validate the software in a larger population are recommended. # Supporting information We are thankful to the health professionals and people with DM for their contributions and suggestions to improve the tool. [^1]: The authors have declared that no competing interests exist.

# Introduction Human strength training is well known to increase skeletal muscle mass and induce muscle phenotypic changes. Increase in muscle strength resulting from skeletal muscle hypertrophy is of great interest to people including elite power athletes, patients rehabilitating from disease-induced atrophy and the elderly who have diminished mobility due to muscular weakness. Muscle hypertrophy is induced by cellular and molecular mechanisms including a number of signaling pathways leading to an increase in protein synthesis and a decrease in protein breakdown. Skeletal muscle satellite cells (SCs) are a group of quiescent cells located between the basal lamina and plasma membrane of the myofibers in mature muscles. These cells are mainly responsible for postnatal muscle growth by hypertrophy, as well as for exercise- or injury-induced muscle regeneration. Indeed, resistance/strength training can increase SC activity and/or the number of myonuclei. Although SCs are key regulators of muscle growth during development and muscle adaptation following exercise, the cellular regulation of the SC function remains largely unexplored. Recently, interleukin-6 (IL-6) has been implicated as part of the activation of human SCs in response to damaging eccentric contractions,. Traditionally, IL-6 is considered as a pleiotropic pro-inflammatory cytokine associated with the control and coordination of immune responses. Increasing evidence indicates that skeletal muscle cells are an additional important source of IL-6 after a single bout of endurance exercise in humans or overload induced hypertrophy in rodent, at least in part under the dependence of the serum responsible factor (SRF). Interestingly, IL-6 knock-out (IL-6⁻/⁻) mice demonstrated a blunted hypertrophic response and a lower SC-related myonuclear accretion compared to wild-type mice following compensatory hypertrophy. Furthermore, SC from IL-6⁻/⁻ mice demonstrated an impaired proliferative capacity, both in vivo and in vitro. This impairment was related to a lack of IL-6 mediated activation of signal transducer and activator of transcription-3 (STAT3) signaling. The activation of Janus tyrosine kinases (JAKs) by IL-6 leads to STAT3 phosphorylation (pSTAT3) and activation which elicits dimerization and translocation of pSTAT3 into nucleus. pSTAT3 induces the transcription of downstream genes involved in several biological functions including cell proliferation, differentiation, and survival of myoblasts. These responses are mediated by the expression of cell cycle regulators *c-myc* and *cyclinD1*, the antiapoptotic genes *Bcl-2* and *Bcl-xL* – and intermediate early response genes such as *c-fos* and *junB*, as well as the angiogenic factor (*VEGF*) and the suppressor of cytokine signaling 3 *(SOCS3)*. Moreover, Yang *et al.* (2009) reported that STAT3 could interact with MyoD, the STAT3-MyoD complex being responsible for the stimulatory effect of STAT3 on myogenic differentiation. During recovery from exercise, the activation of STAT3 signaling has been demonstrated in human skeletal muscle. However, few studies have explored the link between the IL-6/JAK/STAT pathway and SC behavior in the muscular hypertrophy induced by strength or resistance training both in animals or humans. For example, whether or not the muscle IL-6 response still persists after several weeks of training has not yet been investigated. Moreover, the precise mechanisms of the IL-6/JAK/STAT pathway on satellite cell behavior via the regulation of the known myogenic regulatory factors have to be defined in resistance training-induced skeletal muscle hypertrophy. For that purpose, we modified a physiological exercise model from Lee *et al.* (2004) to investigate the underlying molecular and cellular events related to the IL-6/JAK/STAT3 pathway of the rat forearm limb muscle Flexor Digitorum Profundus (FDP) after either a single bout of exercise or after 2, 4 and 10 weeks of voluntary resistance training. We hypothesized that 10 weeks of intense resistance training would lead to hypertrophy linked to repeated muscle IL-6/STAT3-dependent gene stimulation, particularly those genes related to satellite cell behavior, after each single bout of resistance exercise. # Materials and Methods ## Ethics statement This study was approved by the Committee on the Ethics of Animal Experiments of the Languedoc Roussillon in accordance with the guidelines from the French National Research Council for the Care and Use of Laboratory Animals. (Permit Number: CEEA-LR-1069). All surgery was performed under sodium pentobarbital anesthesia, and all efforts were made to minimize suffering. ## Animals 48 Male Wistar Han rats, weighing around 220 g, were purchased from Charles River (Charles River Laboratories International, Wilmington, MA) and housed at a constant room temperature and humidity and maintained at a 12:12 h light-dark cycle. Rats had access to standard rat chow and water *ad libitum*. ## Experimental design Rats were exercised in apparatus adapted from Lee *et al.* (2004). A 1-m ladder with 2-cm grid steps and inclined at 85° was made in our laboratory. Initially, rats were familiarized with the ladder by practicing voluntary climbing of the ladder from the bottom to the top cage for one week, after which the strength- training or exercise regimen started. Cloth bags containing weights were attached to the base of the tail with a Velcro strap. ### Resistance training Protocol After one week of adaptation, 36 rats were randomly divided into six groups: CTL2, CTL4, CTL10 (CTL = non-training controls, n = 6 in each group) and TR2, TR4 and TR10, which were rats trained for 2, 4 and 10 weeks (Training Resistance n = 6 in each group). The initial weight attached to the tail of each animal was 50% of its body weight (bw). Rats were positioned at the bottom of the climbing apparatus and when they reached the top of the ladder, they were allowed to rest in a simulated home cage for 2 min. Rats performed 10 repetitions or climbs, five times a week during 2 (TR2), 4 (TR4) or 10 (TR10) weeks. Training was performed every afternoon. Loads were increased by 10% every 2 days but only if the rat was able to perform 10 climbs per set. After 2 weeks of training, the load reached 120% of bw, 150% after 4 weeks and 210% after 10 weeks of training. Maximal repetition was determined as the maximum weight carried up the exercise ladder by the rats in one climb and was only measured on the last day after 10 weeks training. 72 hours after the last training bout, rats were killed via an intraperitoneal injection of pentobarbital 50 mg.kg⁻¹ (Penthotal®). The forearm muscle, Flexor Digitorum Profundus (FDP), was dissected, frozen in isopentane chilled in liquid nitrogen and stored at −80°C for later use. ### Single resistance exercise protocol After one week of adaptation, rats were randomly divided into three groups: REST (n = 4) with rats sacrificed just before exercise, E2H (n = 4) and E6H (n = 4) where rats were sacrificed 2 and 6 hours after the single bout of exercise respectively. Twenty-four hours before sacrifice, the animals of each group were injected intraperitoneally with 100 mg.kg⁻¹ of bromodeoxyuridine (BrdU, B5002, Sigma Aldrich, St Louis, MO) to identify cells in proliferation. In the afternoon, rats of the E2H and E6H groups made 4 climbs with a load that reached 25% of bw, 4 climbs at 50%, 4 climbs at 75% and 6 climbs at 100% of bw. Between each climb, rats were allowed to rest for 2 min. Rats were anesthetized via an intraperitoneal injection of pentobarbital 50 mg.kg⁻¹ (Penthotal®). The forearm FDP muscles were harvested, frozen in isopentane chilled in liquid nitrogen and stored at −80°C for later use. The animals were then killed by an overdose of pentobarbital. ## Myosin Heavy Chain Immunohistochemistry For fiber type analysis, transverse serial sections of FDP muscles (10 µm thick) were obtained using a cryostat at −20°C. Frozen sections were fixed with acetone solution for 10 min, washed and incubated 30 min in phosphate buffered saline (PBS) blocking solution with 2% bovine serum albumin (BSA). Sections were then incubated 2 hours at room temperature with mouse monoclonal antibody (DSHB, Iowa City, Iowa) directed against MHC-I (#A4-971), MHC IIa (#2F7), MHC IIx (#6H1) and MHC-IIb (#10F5). Sections were then washed three times with PBS and incubated one hour at 37°C with peroxidase-conjugated rabbit anti-mouse IgG secondary antibody (A-9044, Sigma-Aldrich, St Louis, MO). MHC staining was revealed with NovaRed™ (Vector® Lab, Burlingame, CA) and slides mounted with Mowiol. Images were captured with a microscope (ZEISS Axiophot) coupled with a CCD camera connected to a computer. MHC analysis was realized with Image-J® software. Muscle fiber Cross-Sectional Area (CSA) was obtained in ×10 magnification images from 1500 fibers per group (250 fibers per muscle from 6 rats). ## Western Blotting 20 mg FDP samples were homogenized in 10 volumes of lysis buffer (Tris 20 mM pH 6,8, NaCl 100 mM, EGTA 1 mM, NaF 100 mM, Triton X-100 0,5%, Na₃VO₄ 5 mM) with inhibitory protease cocktails (P8340, Sigma Aldrich, St Louis, MO). The homogenate was rotated 10 min at 4°C and the supernatant collected. Protein samples (50 µg) were denatured and separated on 10% SDS-PAGE. The proteins were transferred onto a nitrocellulose membrane and blocked in 5% wt/vol BSA for phosphoSTAT1 (Tyr 701), phosphoSTAT3 (Tyr 705), phosphoErk1/2 (Tyr 202/204) and 5% wt/vol dry milk for STAT1, STAT3, Erk1/2 and α-tubulin antibodies in Tris-buffered saline with 0.1% vol/vol Tween 20 (TBST) for one hour at room temperature. Primary antibodies pSTAT1 (1/1000), pSTAT3 (1/1000), pErk1/2 (1/1000), STAT1 (1/1000), STAT3 (1/1000), Erk1/2 (1/1000) and α-tubulin (1/2500) diluted in blocking buffer were applied and incubated overnight at 4°C. Membranes were subsequently washed three times with TBST and incubated one hour at room temperature with a secondary antibody conjugated to horseradish peroxidase, donkey anti-rabbit IgG (1/4000) for pSTAT1, pSTAT3, Erk1/2, STAT1 and STAT3 and rabbit anti-mouse IgG (1/4000) for pErk1/2 and α-tubulin. Proteins were visualized by enhanced chemiluminescence (32106, Pierce®, Rockford, IL) and quantified with ImageJ® software. Levels of pSTAT1, pSTAT3 and pErk1/2 proteins were expressed relative to total STAT1, STAT3, Erk1/2 respectively. Antibodies pSTAT1 (#9171), pSTAT3 (#9131), pErk1/2 (#9106), STAT1 (#9172), STAT3 (#9132) and Erk1/2 (#9102) were purchased from Cell Signaling Technology (Danvers, MA), α-tubulin (#T6199) and secondary antibody rabbit anti-mouse (#A9044) from Sigma-Aldrich (St Louis, MO) and secondary antibody donkey anti-rabbit (#NA934V) from GE Healthcare (Buckinghamshire, UK). ## pSTAT3, Pax7 and BrdU immunohistochemistry 12 µm thick FDP muscle sections were either co-stained for pSTAT3 and Pax7 antibodies to distinguish the cellular localization of STAT3 activation in nuclei of myocytes or in nuclei of both quiescent and activated satellite cells (Pax7⁺ cell) or co-stained for BrdU and laminin antibodies to assess SCs mitotic activity. All cryosections were rehydrated in phosphate buffered saline (PBS). For pSTAT3 and Pax7 staining, sections were blocked in 1% BSA for one hour at room temperature. Primary antibodies were incubated with 0.1% Triton X-100 and 1% BSA overnight at 4°C. PBS was used for all washing steps throughout staining, with the exception of Pax7, where Tris-buffered saline (TBS) was used to wash the sections. Muscle sections were first stained for pSTAT3 and followed by a second staining for Pax7 (DSHB, Iowa City, Iowa). After incubation with pSTAT3 primary antibody (1/50), the sections were washed in PBS, and goat anti- rabbit IgG secondary antibody (1/1000, Alexa Fluor® 488, \#A-11034, Invitrogen Life Technologies, Renfrew, UK) was applied for 30 min at 37°C, post-fixed in 1% paraformaldehyde (PAF) for 10 min, incubated with the second primary antibody Pax7 (undiluted) overnight at 4°C. Sections were washed in TBS and goat anti- mouse IgG secondary antibody (1/1000; Alexa Fluor® 568, \#A-11031, Invitrogen Life technologies, Renfrew, UK) was applied for 30 min at 37°C, post-stained with Hoescht and fixed in 1% PAF. For BrdU and laminin labeling, sections were fixed in 4% PAF (pH 7.2) for 5 min, washed in PBS and placed in 2M HCl at 56°C for 30 min in order to denaturate double-stranded DNA. After neutralization with 0.1M sodium borate (pH 8.5) for 10 min, sections were washed in PBS and blocked with normal goat serum (1/30, G9023, Sigma) for 1 h at 25°C. Muscle sections were first stained for BrdU and followed by a second staining for laminin. After incubation in monoclonal BrdU antibody (1/50, B-2531, Sigma) in PBS with 0.2% BSA and 0.05% Tween-20 for 2 h at 25°C, the sections were washed in PBS and goat anti-mouse Alexa Fluor® 568 secondary antibody (1/250) was applied for 1H30 at 25°C. Sections were post-fixed in 4% PAF and incubated with second primary polyclonal antibody laminin (1/50, L9393, Sigma) diluted in 1% BSA for 30 min at 37°C. After washing, goat anti-rabbit Alexa Fluor® 488 secondary antibody (1/1000) was applied on sections for 30 min at 37°C. Sections were washed and post-fixed in 4% PAF. All sections were mounted with Dako Fluorescent Mounting Medium (Dako, \#S3023, Carpinteria, CA) and images were collected using a microscope (ZEISS Axiophot) coupled with a CCD camera and analyzed using ImageJ® software. BrdU-stained cells were counted as SCs when located intra-laminin staining and correlated to the number of fibers (∼500 fibers per muscle). ## RNA Extraction and Real Time Polymerase Chain Reaction RNA was isolated from homogenate muscle samples using the RNeasy mini Kit following the manufacturer's instructions (Qiagen, Germantown, MD). The RNA was quantified with a spectrophotometer (Eppendorf AG, Hamburg, Germany). RNA integrity was electrophoretically verified by ethidium bromide staining and by OD₂₆₀/OD₂₈₀ nm absorption ratio \>1.95. 2 µg RNA of each sample was reverse transcribed to cDNA in 20 µl reactions using a commercially available kit (High Capacity cDNA Reverse Transcription Kit; Applied Biosystem Life Technologies Carlsbad, CA) according to the manufacturer's instructions. The cDNA synthesis reaction was carried out using thermal cycler (MiniCycler™, MJ Research, St Bruno, Canada) and followed by 10 times dilution with ultra-pure water containing denaturated salmon sperm DNA. Forward (F) and Reverse (R) primers used to amplify genes are listed in. Quantitative real-time PCR was performed in a 20 µl final volume with 250 nM of each primer using iQ SYBR Green Supermix (Bio-rad, Hercules, CA). After incubation at 95°C for 10 min, the cycling protocol was performed in MiniOpticon™ (Bio-rad) as follows for IL-6, LIF, SOCS3 and cMyc: 10 s at 95°C for denaturation, 30 s at 60°C for annealing. For the reaction of Myogenin, the annealing temperature was set at 61°C, for Rpl32, CycloA and CyclinD1 at 63°C and for MyoD, Myf5 and Pax7 at 64°C. After 40 cycles of PCR, melting curve analysis was performed to check primer specificity. All Cq values were analyzed using a comparative critical threshold method previously described by Pfaffl (2001). Transcription levels were normalized using Cq arithmetic mean of two reference genes: CyclophilinA and Rpl32. ## Statistical analysis All values are expressed as means ± SEM. A one-way ANOVA was employed to compare data. When a significant effect was indicated, a Fisher significant difference post hoc test was performed. Significance was set at p\<0.05. # Results ## Resistance training induces phenotypic changes and fiber hypertrophy of FDP muscle This resistance training protocol induced an alteration in fiber type composition with marked changes from 4 weeks of training (TR4). The proportion of type I and IIx fibers decreased whereas that of IIa increased by 47% in TR4 and TR10 groups. To confirm that the voluntary resistance training protocol could promote muscular hypertrophy, fiber cross-sectional areas (µm²) were measured at different time points after 2 (TR2), 4 (TR4) or 10 (TR10) weeks of training. Two and four weeks of training caused significant hypertrophy of type IIx, respectively +49% and +88% compared to respective control group (n = 6; p\<0.05). As all FDP muscle fiber types were hypertrophied after 10 weeks of resistance training (TR10), we choose to focus all further analyses at this point in time. ## 10 weeks of resistance training did not affect the myonuclear domain We found that the increased cross-sectional area of muscle fiber after 10 weeks of training was not accompanied by any variation in the myonuclear domain value suggesting incorporation of new nuclei into the fibers. Indeed, we noted a significant increase in the number of myonuclei per fiber cross-section (data not shown). ## STAT1 and STAT3 are phosphorylated in rat skeletal muscle following acute resistance exercise The activation of STAT1 and STAT3 were assessed by pSTAT1 tyrosine 701 and pSTAT3 tyrosine 705 in FDP muscle samples at REST, 2 hours (E2H) and 6 hours (E6H) after a single bout of resistance exercise as well as after 10 weeks of resistance training (TR10). 2 and 6 hours post-exercise, pSTAT3 (n = 4, p\<0.01) and pSTAT1 had increased significantly (n = 4, p\<0.05) from resting value (REST). To confirm that the increased phosphorylation observed was not due to an increase in STAT3 and STAT1 protein levels, the density of the pSTAT3 and pSTAT1 band was normalized against total STAT3 and STAT1 proteins respectively, which remained constant across the samples. In contrast, pSTAT3 had decreased (n = 6, p\<0.05) after 10 weeks of resistance training (TR10) compared to resting value (CTL10). Moreover, immunofluorescence staining revealed that pSTAT3 co-localized with Pax7⁺ cells only at E2H, with no-detectable pSTAT3 at REST, E6H, CTL10 and TR10 indicating that STAT3 signaling was transiently active within SCs at E2H. ## Erk1/2 is phosphorylated in rat skeletal muscle following acute resistance exercise Erk1/2 phosphorylation to total levels were significantly increased 6 hours after a single bout of exercise (E6H;) compared to resting values (REST, p\<0.05). There was no change of Erk1/2 phosphorylation after 10 weeks of training (TR10) compared to resting values (CTL10) ## Satellite cell proliferation after acute resistance exercise To assess the involvement of SCs following acute resistance exercise, BrdU- positive cells situated between basal lamina and plasma membrane (intra-laminin staining) were counted to quantify satellite cell proliferating state. When expressed in percentage of fibers, BrdU-positive SCs increased from 0.3% (REST) to 6%, 2 hours after exercise (E2H; p\<0.05) and reached to 3%, 6 hours after exercise (E6H, p\<0.05).Thus, resistance exercise contributes to activate SCs into proliferating state as early as 2 hours post-exercise. ## Impact of resistance exercise and training on STAT3-responsive genes Downstream genes of STAT3 were investigated including IL-6, LIF, SOCS3, myogenic regulatory factors (MyoD, Myf5, Pax7, Myogenin) and markers of cell proliferation (CyclinD1, c-Myc). ## Myocyte mRNA expression of IL-6, LIF and SOCS3 The mRNA expression for IL-6 increased significantly 2 and 6 hours after a single bout of exercise, respectively 2.2-fold (n = 4, p\<0.05) and 3.2-fold (n = 4, p\<0.01), compared to resting values. No significant change was obtained for LIF mRNA expression after acute exercise. After 10 weeks of training, IL-6 mRNA expression also increased 1.57-fold (n = 6, p\<0.01) compared to CTL10 values whereas LIF mRNA expression is downregulated (n = 6; p\<0.05). The mRNA expression for SOCS3 significantly increased 6 hours after a single bout of resistance exercise, 3.7-fold (n = 4, p\<0.05), but tended to increase 2 hours after exercise (p = 0.08). No change was observed in SOCS3 mRNA expression after 10 weeks of resistance training. IL-6 and SOCS3 mRNA levels after a single bout of exercise are significantly correlated (R² = 0.55, p\<0.05) at 2 and 6 hours after exercise. ## mRNA expression of CyclinD1 and c-Myc are upregulated and correlate with IL-6 gene expression The mRNA expression for CyclinD1 and c-Myc genes, two markers of cell proliferation, increased significantly in E2H group, respectively 1.5-fold (n = 4, p\<0.05) and 3-fold (n = 4, p\<0.05). CyclinD1 mRNA expression also increased 1.8-fold (n = 4, p\<0.05) in E6H but it did not reach significance for c-Myc in E6H group (n = 4, p\<0.05;). As for SOCS3, CyclinD1 mRNA expression correlated with IL-6 mRNA after a single bout of exercise (R² = 0.64, p\<0.05). On the other hand, significant decreases were observed in TR10 group for mRNA expression of CyclinD1 and c-Myc, respectively 0.78-fold (n = 6, p\<0.05) and 0.70-fold (n = 6, p\<0.05). ## Myogenic Regulatory factor mRNA expression (Pax7, MyoD, Myf5, Myogenin) The mRNA expression of Pax7, a marker of quiescent and activated satellite cells, MyoD and Myf5, markers of active satellite cell proliferation, significantly decreased 2 hours (E2H) after a single bout of exercise, respectively 0.50 (p\<0.01), 0.56 (p\<0.01) and 0.63-fold (p\<0.01). MyoD and Myf5 mRNA also decreased 6 hours after the same exercise (E6H), respectively 0.21 (p\<0.01) and 0.54-fold (p\<0.01) whereas Myogenin mRNA increased 2.24-fold (p\<0.05). No significant change in Pax7 and MyoD mRNA expressions was observed in TR10 group, whereas Myf5 and Myogenin mRNA expressions significantly decreased in TR10 group respectively 0.56 and 0.62-fold (n = 6, p\<0.05). The ratio Pax7/MyoD mRNA, a marker of satellite cell self-renewal, significantly increased in E6H, 4.08-fold (n = 4, p\<0.01) # Discussion In the present study, we modified a physiological exercise model from Lee *et al.* (2002) to induce hypertrophy and explored the link existing between the IL-6/STAT3 pathway and the acute and chronic SC activation through the myogenic regulatory factor (MRF) kinetic response. To our knowledge, no studies have focused on skeletal muscle hypertrophy after voluntary resistance training exercise in rats (i.e. without electric stimulation to force the animals, or compensatory hypertrophy with tonic-like acquired neuronal activity). For example, the squat model apparatus by Tamaki *et al.* (1992) did not produce any skeletal muscle hypertrophy although the strength training was performed for 12 weeks (3 sets of 10 repetitions at 75% of 1 maximum repetition). With our physiological resistance training model, skeletal muscle hypertrophy occurred mainly after 10 weeks of resistance training in rats. Similar to previous studies in humans, we reported an increase in CSA for all fiber types but to a greater extent for type IIa (+92%) and type IIx (+100%) fibers in FDP muscle. It is well documented that resistance training could induce fiber hypertrophy through an enhancement of protein synthesis occurring just after the training session and lasting up to 24–48 h in humans. In parallel, depending on the exercise stimulus, the recruitment of additional nuclei derived from SC incorporated into muscle fibers could occur. Indeed, several works in humans have evidenced an increase in the number of myonuclei per fiber when fiber size increases approximately more than 25%. Thus, the myonuclear domain (i.e. the theoretical amount of cytoplasm supported by a single myonucleus in a muscle fiber) remained constant although a large increase in fiber CSA via the addition of SC-derived nuclei occurs. However, according to McCarthy *et al.* (2011), in a SC depleted mice muscle, a robust fiber hypertrophy can occur after synergist muscle ablation. Such hypertrophy is associated with an expansion of myonuclear domain suggesting that SCs are not required to sustain hypertrophy in this model. Yet, in our study we found up to 100% of fiber size increase after 10 weeks of resistance training (TR10) which is associated to a constant myonuclear domain, suggesting that in addition to the upregulation of protein synthesis, the SC population participates in fiber hypertrophy. ## Acute exercise Serrano *et al.* (2008) have shown a blunted hypertrophic response in skeletal muscle of IL-6^−/− mice following compensatory hypertrophy, suggesting that IL-6 may play a role in skeletal muscle hypertrophy. To further explore the link between the IL-6 pathway and the intervention of SCs in the hypertrophic response, our study has focused on MRFs and proliferating capacity through activation of the IL-6/STAT3 signaling pathway after resistance exercise in rats. Under specific conditions, the IL-6/STAT3 pathway could be relevant in SCs as it could mediate the hypertrophic response after resistance training in humans. In our animal model, we observed an increase in pSTAT3 at both 2 and 6 hours after resistance exercise, which was closely associated with an increase in IL-6, SOCS3, c-Myc and CyclinD1 mRNAs after a single bout of exercise. One can argue that IL-6 gene up-regulation may come from local inflammation. However, we performed Hematoxiline/Eosine staining and did not find any presence of inflammatory cells (*data not shown*). The exercise-induced cytokine response differs from that in classical infectious context where TNF-α and IL-1β are the first secreted pro-inflammatory cytokines. In humans, after exercise, TNF-α and IL-1β do not increase, and IL-6 is usually the first cytokine present in the circulation. There is now a growing evidence that acute exercise related IL-6 response act as an anti-inflammatory cytokine since IL-6 can exert inhibitory effects on TNF-α and IL-1 production and stimulate the production of both anti- inflammatory cytokines IL-1ra and IL-10 STAT3 is activated in SCs in a transient manner as only pSTAT3 was detectable in SCs (Pax7⁺) after 2 hours of resistance exercise. Moreover, the number of mitotically BrdU positive SCs was significantly increased at both 2 and 6 hours after acute resistance exercise which is concomitant with the cell cycle markers CyclinD1 and c-Myc mRNAs. These cells cycle markers are known as IL-6/STAT3-responding genes and have a critical role in cell growth and cell- cycle transition from G1 to S phase. Altogether these results sustain IL-6 dependent SC proliferation. Others members of the IL-6 cytokine family, particularly the Leukemia Inhibitory Factor (LIF), could also contribute to STAT3 activation. However, contrary to IL-6, the LIF gene stimulation was not significant 2 hours (E2H) or 6 hours (E6H) post-exercise in the present study. Most of the studies looking at MRFs both after injury in rodent and resistance training in humans, suggested that the increase in MRF mRNA expression occurred at later time point (12 hours to 2 days). When focus is made on the early mRNA regulation of MRFs, we observed a downregulation of MyoD and Myf5 mRNAs, 2 and 6 hours after acute resistance exercise whereas an increase in Myogenin mRNA at 6 hours post-exercise was noted. Similar results were obtained in humans by Costa *et al.* (2007) that reported a 45% decrease in MyoD mRNA, an absence of increase in Myf5 mRNA but an increase in Myogenin mRNA, 3 days after an eccentric training program. Moreover, looking at the protein level, Sakuma *et al.* (1999) have also shown decreased plantaris MyoD content in the first 5–6 days after the ablation of both synergist soleus and gastrocnemius muscles, leading to compensatory plantaris hypertrophy in rats. The significant up regulation of the Myogenin gene 6 hours (E6H) after acute resistance exercise suggests that some SCs are going to differentiate. As STAT1/STAT3 signaling pathway is early activated after our exercise, we first hypothesized that it could promote myoblast proliferation and prevent myoblast differentiation by inhibiting MyoD transcription. The increased Myogenin mRNA shows that SCs differentiation is not completely abolished but suggests that different pool of SCs come into different states after exercise as suggested by Schultz et al.. One part of SCs is specifying to become reserve cells as MyoD and Myf5 mRNAs are decreased along with the Pax7/MyoD ratio upregulation, but some SCs are able to engage differentiation early after exercise and fuse with existing myofibers to give their material. Thus, the SC population can be phenotypically and functionally divided into several compartments allowing different engagement. Thus, early after resistance exercise there might be a preferred proliferative phase of the first undifferentiated satellite cells in order to build reserve cells before engaging in a myogenic lineage, as suggested by Yoshida *et al.* (1998). Precisely, this mechanism may result from activation of the JAK1/STAT1/STAT3 signaling pathway. Indeed, the STAT1/STAT3 signaling pathway activation could promote myoblast proliferation and prevent premature myoblast differentiation by inhibiting MyoD transcription whereas the STAT2/STAT3 pathway is required for myogenic differentiation. Interestingly, we showed a concomitant increase in pSTAT1, peaking at 2 h after exercise, and pSTAT3 at both 2 and 6 hours after exercise (; p\<0.05). Thus, the early downregulation of MyoD and Myf5 mRNA could be mediated in part by the STAT1/STAT3 pathway in order to promote cell proliferation by STAT3 activation and repress cell differentiation via STAT1. This hypothesis is sustained by the significant upregulation of the Pax7/MyoD ratio observed at 6 hours after exercise (p\<0.01) suggesting that SCs return in a quiescent state. Moreover, the concomitant increase in pErk1/2 level could strengthen this proliferative phase, as the Erk1/2 pathway activation has been shown to inhibit differentiation at the early stage of differentiation but promote myocyte fusion in the late stage of differentiation. As suggested by Fukuda *et al.* (1996) in pro-B cells lines, Erk1/2 phosphorylation at 6 hours after exercise (E6H) may come from IL-6 signaling which is also required for proliferation of satellite cells mediated by STAT3. Finally, these data are in accordance with those of Sun *et al.* (2007) who pointed out a dual role of STAT1 and STAT3 in myoblast proliferation and differentiation. ## IL-6/STAT3 response to resistance training After 10 weeks of resistance training, the IL-6 mRNA was still 1.4-fold higher than for the resting condition. However, this increase was not accompanied by the upregulation of STAT3 target genes (SOCS3 mRNA) but instead by a downregulation of CyclinD1 and c-Myc mRNAs (p\<0.05;). Accordingly, STAT3 was significantly less phosphorylated compared to the resting conditions. The decrease of pSTAT3 content could not be explained by an increase in the negative feedback loop initiated by the upregulation of SOCS3 since SOCS3 mRNA was not altered. Similarly to the results obtained with acute exercise, Myf5 (p\<0.05) and MyoD (p = 0.1) mRNAs were reduced after resistance training and contrary to acute exercise also Myogenin mRNA was reduced. Consistently, the decreased basal pSTAT3 content noted after 10 weeks of training along with, CyclinD1, c-Myc, Myf5 and Myogenin mRNAs could be due to impairment in satellite cell proliferation. Previous studies showed that activated, proliferating satellite cells express both Pax7 and MyoD. Once activated, some cells then downregulate Pax7, maintain MyoD and differentiate, contrary to others which downregulate MyoD, maintain Pax7 expression and remain undifferentiated. We suggest that the heavy resistance training proposed here may have acutely decreased the number of proliferative satellite cells in order to increase the quiescent satellite cell pool as the Pax7/MyoD mRNA ratio was increased 6 hours after exercise compared to the resting condition. Moreover, the decreased basal pSTAT3 content after several weeks of heavy resistance training might be part of a protective mechanism from excessive muscle mass upregulation, as some fibers had already reached up to 100% hypertrophy As depicted recently by Chakkalakal et al., it would be interesting to verify the Fgf2 signaling and/or sprouty1 (Spry1) expression in SC niche of resistance trained rats as an adaptive mechanism to limit hypertrophy. In conclusion, the hypertrophic effect obtained after 10 weeks of resistance training in rat FDP muscle is acutely associated with the upregulation of the IL-6/STAT3 signaling pathway and the early downregulation of differentiating related MRF gene expressions. In fact, each acute resistance exercise bout in rat induces an increase in SCs IL-6 signaling through the activation of pSTAT3 and its dependent genes, CyclinD1 and c-Myc. The fast decrease in MRF mRNAs could reflect a proliferative phase of the satellite cell population mediated by STAT1/STAT3 activation in order to first rebuild the pool of reserve cells. After 10 weeks of resistance training, the huge training-induced increase in muscle fiber cross-sectional area, up to 100%, is to be linked to the decrease of Pax7/MyoD mRNA ratio. This could be an adaptive mechanism to protect skeletal muscle from excessive hypertrophy. We thank G. Hugon (INSERM, ERI 25, Montpellier) for his microtome technical assistance, Drs. F. Favier and A. Bonnieu for their precious assistance and helpful comments and Miss B. Bonafos for taking care of the animals used in the present study. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: GB GP. Performed the experiments: GB AD OG BR BV GP. Analyzed the data: GB GP. Contributed reagents/materials/analysis tools: GB AD OG BR BV GP. Wrote the paper: GB GP RC.

# Introduction Australia is a major global beef producer. In 2019, Australia was the second- largest beef and veal exporter after Brazil and accounted for 14% of total global beef export. The northern region of Australia encompassing parts of three states including Queensland, Northern Territory and Western Australia accounted for 45, 9 and 8% of the Australian national cattle herd, respectively. Beef cattle are routinely backgrounded on extensive grazing systems and finished on pasture for the lean beef market, or energy-dense grain-fed prime beef markets. Northern Australian tropical beef cattle rely mainly on native grass with few sown grass and legume pastures. In these summer rainfall-dominant dry tropics and sub-tropics, cattle are able to selectively graze in the early wet season, but often lose body condition, experience slow growth and struggle to attain maintenance weight in the other seasons due to low diet crude protein (CP) and digestible energy, pasture senescence, frost and overall poor pasture quality. Augmenting grass pastures with legumes has been reported to improve diet CP and energy digestibility. Besides, legumes improve the yield and nutritive value of grass-based pastures since the nutritive value of the resultant diet is higher compared to grass-only diet, with specifically profound effects in winter and spring. In northern Australian light-textured soils, tropical pasture legumes came into general use after 1960. Benefits on the heavier textured soils (the Vertosols) are being evaluated after the development of suitable legumes. Among the new legume pastures developed is desmanthus, a legume native to the Americas. Desmanthus is reported to have the potential for use as a forage legume in extensive grazing systems and crop rotations. Two studies examining growth performance of livestock fed desmanthus-grass diet reported higher liveweight gain in cattle and goats compared to *Cenchrus ciliaris* (buffel) and *Brachiaria mulato* (mulato) grass only diets. In contrast, growing goats fed *Sorghum bicolor* (sudan grass) and supplemented with *D*. *bicornutus* leaves gained less weight than those supplemented with *Leucaena leucocephala* (leucaena), *Medicago sativa* (lucerne) or *Lablab purpureus* (lablab). Beef cattle supplemented with incremental *D*. *leptophyllus* or *D*. *bicornutus* levels up to 31% dry matter (DM) had similar weight gain with their counterparts fed Rhodes grass (*Chloris gayana*) only diet, although desmanthus supplementation improved rumen fermentation. These studies indicate that there are discrepancies and inconsistencies in animal growth in response to supplementation with desmanthus. Therefore, more studies are required to determine desmanthus’ effect on beef cattle growth performance and change in rumen and plasma metabolites. Our previous study, demonstrated that in an extensive grazing system typical of Central Queensland, Brigalow region, backgrounding 400 beef cattle steers during the dry season for 147 days on buffel grass alone or buffel-desmanthus mixed pastures with desmanthus accounting for 11.5% pasture biomass, did not produce any significant differences in liveweight, daily weight gain and plasma metabolites. The lack of difference was thought to be due to similar dietary crude protein levels for steers on both paddocks accessed through browsing on shrubs and forbs, necessitating the need for feeding steers isonitrogenous diets with varying levels of desmanthus inclusion in a controlled pen trial. Therefore, the aim of the current study was to evaluate the effect of increasing levels of desmanthus in isonitrogenous diets on beef cattle growth rate, rumen fermentation and plasma metabolites of tropical crossbred beef cattle. We hypothesized that cattle fed isonitrogenous diets supplemented with incremental levels of desmanthus would have similar growth rates, rumen fermentation and plasma metabolites concentration. # Materials and methods This study was carried out at the CSIRO Lansdown Research Station, Queensland, Australia (19.59° S, 146.84° E) between March and July 2020. The station receives 861 mm mean total annual rainfall with mean annual min and max temperatures of 16.8 and 26.1°C, respectively. All procedures in this study followed the CSIRO Animal Ethics Committee approved guidelines (approval number 2019–38) and the Australian code of practice for the care and use of animals for scientific purposes. ## Forage establishment and management Soil analysis was carried out in a 12 ha plot of land before forage establishment to examine suitability of the plot for desmanthus production. Pure stands of three desmanthus cultivars, namely *D*. *virgatus* cv. JCU2, *D*. *bicornutus* cv. JCU4 and *D*. *leptophyllus* cv. JCU7 (Agrimix Pastures Pty Ltd, Ferny Hills DC, QLD, Australia), were planted on 29 November 2019 in 4 ha plot per cultivar at a sowing rate of 2 kg/ha. The plots were irrigated as follows; 15–22 mm/m² every two days for the first 10 days, every three to four days from day 11 to 30 and once a week onwards. Each cultivar plot was divided into three and harvesting was staggered between them to ensure similar maturity stage of the forage at harvest. Overgrown forage was slashed and used as mulch to encourage new regrowth. Plots were not fertilized during desmanthus establishment but were top dressed with Natramin S (4.7% Ca, 0.07 P, 6.3S, 2.8 K, 2.3 Mg, 23.3 Si, 5.1 Fe, 3255 ppm C, 930 ppm Mn, 140 ppm Zn, 56 ppm Cu, 23 ppm Co, 18 ppm B and 6 ppm Mo; Ag Solutions, Gympie, QLD, Australia) at 400 kg/ha in February and urea at 100 kg/ha in March 2020. Weeds were controlled by spraying the plots with 3 L/ha glyphosate-based herbicide (Roundup; Monsanto, Kilda Road, Melbourne, Australia) mixed with 0.7 L/ha 2,4-dichlorophenoxyacetic acid (Titan amine; Titan Ag Pty Ltd, Princes Street NSW, Australia) four weeks pre-planting. Fluazifop-P (Fusilade Forte; Sygenta Australia Pty Ltd, Lyonpark Road, NSW, Australia) and clethodim (Select; Arysta Life Science Australia Pty Ltd, Hindmarsh Square, Adelaide, South Australia, Australia) were sprayed at 0.8 and 0.4 L/ha, respectively, three weeks after planting and twice more during the trial when forage was slashed back to prevent competition from grasses. All herbicides were mixed in 150 L/ha of water. ## Animal management and diets An *a priori* power analysis was conducted using G-Power to determine the appropriate sample size. A total sample size of 48 steers was required to achieve statistical power of 95% with a critical F-value of 4.06 for a medium effect size and a significance level of 0.05. Therefore, forty-eight tropically adapted 24–28 months old Brahman, Charbray and Droughtmaster crossbred steers weighing 332 ± 21 kg, were used for the study. The steers were fitted with insecticidal cattle ear tags containing a synergized formulation of zeta cypermethrin (Y.Tex Corporation, Cody, Wyoming, USA) and treated with an Ivermectin based pour-on parasite control (BAYMEC, Bayer Australia Ltd, Pymble NSW, Australia) at 1 ml/10 kg liveweight (LW) dosage at the beginning of the study for internal and external parasites control. Steers were group-housed in 12 open pens (4 steers per pen) with three pens per experimental diet in a completely randomized design. Steers were allocated to four groups based on their initial liveweight to ensure similar mean liveweight per pen. The pens were then randomly assigned to one of four diets. Each pen measured 60 m², fitted with 18 m² shade and 4 m by 1 m feed trough. Steers were allocated to one of the following four experimental diets: 0, 15%, 30% or 45% desmanthus on DM basis with Rhodes grass hay as the basal diet. Diet CP was adjusted by the inclusion of lucerne hay in the 0, 15% and 30% desmanthus diets, based on forage CP and not DM basis, to ensure all diets were isonitrogenous. The 0 desmanthus diet (Rhodes grass and lucerne) was used as a positive control. Lucerne was selected because it is the most widely grown perennial legume globally and has been extensively studied. Desmanthus consisted of the three species, namely *D*. *virgatus*, *D*. *bicornutus* and *D*. *leptophyllus* fed in equal proportions. A mixture of these three species was used because *d*esmanthus is commonly marketed as a mixed product, e.g. Jaribu desmanthus comprising *D*. *virgatus cv Marc*, *D*. *leptophyllus* cv Bayamo and *D*. *pubescens cv Uman* in the 1990s and Progardes^® consisting of *D*. *biconutus cv JCU4*, *D*. *leptophyllus cv JCU7* and *D*. *virgatus* cv JCU2 and cv JCU5 cultivars to ensure that the best adapted cultivars eventually dominate while the less adapted cultivars take advantage of seasonal, land types and climate variation. Steers were gradually adapted to the experimental diet within ten days. Throughout the study, steers had unlimited access to clean water and mineral block (Trace Element Northern, Ollson’s, Yennora, NSW, Australia). Desmanthus was harvested at the late bloom to full seed maturity stages and chopped with a flail type forage harvester (New Holland model 38 Crop-Chopper, Haryana, India) every morning, while Rhodes grass and lucerne hay were chopped with a tub grinder (Roto Grind model 760, Burrows Enterprises, LLC, Greeley, CO, USA) once a week. Experimental diets were mixed by hand and offered daily between 8:30 and 9:30 am after residuals were collected. Feed offered was adjusted to allow for 5–10% refusals. Refusals were weighed daily and a sample per pen stored at -20°C, from which a weekly composite bulked sample per pen was obtained for DM and chemical analysis. Weekly samples of the desmanthus, lucerne and Rhodes grass were obtained throughout the study for DM and chemical analysis. ## Feed intake, liveweight and body condition scores Dry matter intake (DMI) per pen was determined by the weight difference between feed offered and refusals collected 24 hrs after feeding. Steers were weighed at the start and end of the study, fortnightly for the first six weeks and monthly thereafter until the end of the study. The frequency of weighing changed after six weeks due to labour limitations resulting from COVID-19 pandemic related restrictions. All steers were weighed before feeding to reduce variation due to gut fill. Unfasted LW were recorded automatically (Gallagher 65 Scanlon Drive, Epping, Victoria 3076, Australia) and the average daily gain calculated by regressing all fortnightly and monthly LW by time in days. Body condition scores (BCS) were recorded monthly using the five-point (1–5) scoring system. ## Forage and refusals analysis The forage offered and refusals DM, CP, neutral detergent fibre (NDF), acid detergent fibre (ADF), hemicellulose, dry matter digestibility (DMD) nutritive values were estimated using near infrared reflectance (NIR) spectroscopy at the CSIRO Floreat laboratory (Floreat, WA, Australia). The samples were dried in a forced-air oven at 60°C for 48 h and ground to pass through a 1 mm mesh with a Christy and Norris grinder (Christy Turner Ltd, Suffolk, England), the spectra collected using the Unity Spectrastar 2500X rotating top window system (Unity Scientific, Milford, MA, USA) and predictions were generated using the chemometric software package Ucal (Unity Scientific) as described by Norman et al.. ## Rumen fluid collection and analysis Rumen fluid samples were collected in the morning prior to feeding at the start, middle and end of the experimental period (week 0, 10 and 20) from 24 steers to determine the effect of diet on rumen pH, ammonia nitrogen (NH₃-N) and volatile fatty acids (VFA). About 200 ml of rumen fluid samples were collected from the ventral sac via oro-ruminal tubing using a reinforced plastic suction tube and a hand pump. The pH of each sample was taken immediately using a portable pH meter (Aqua-pH, TPS Pty Ltd, Brendale, QLD, Australia). Rumen fluid sub-samples of 8 ml were stabilized by adding 2ml of 25% metaphosphoric acid and stored at -80°C awaiting NH₃-N and VFA analysis. The rumen NH₃-N was analyzed using the colorimetric method of Chaney and Marbach, while VFA were determined by gas chromatography (Shimadzu Corporation, Kyoto, Japan) as described by Gagen et al.. Since rumen fluid with viscous appearance and pH of 7.5 or above indicate contamination with large volumes of saliva, samples with these characteristics were excluded in pH, total VFA and NH₃-N analysis. ## Blood collection and plasma metabolites analysis Blood samples were collected in the morning prior to feeding at the start, middle and end of the experimental period (weeks 0, 10 and 20) via jugular venipuncture into 10 ml sodium heparin blood Vacutainer tubes (BD, Sydney, Australia), centrifuged at 1425 x g for 20 min at 4 °C (Beckman Coulter, Inc. California, USA) to separate the plasma from the serum and stored at -80°C prior to analysis. All plasma metabolites were analyzed using the AU480 chemistry analyzer6 (Beckman Coulter, Inc. California, USA) according to the manufacturer’s procedures. Plasma non-esterified fatty acids (NEFA) were analyzed using colorimetric method, beta-hydroxybutyrate (BHB) and glucose analyzed by the 3-hydroxybutyrate dehydrogenase and hexokinase methods, respectively, total bilirubin by the modified diazo method and creatinine by the kinetic modified Jaffe method. ## Statistical analysis Data were analyzed using the SAS software version 9.4 (SAS Institute, Cary, North Carolina, USA), with an initial screening for data entry errors, outliers and data distribution done for all data sets. Mixed model (PROC MIXED) restricted maximum likelihood (REML) procedures in SAS fitted the fixed effect of diet and pen nested within diet as a random effect in the statistical model. Sampling week was analyzed as a repeated measure and covariance structures were specified. Final LW and BCS were analyzed by including the initial measurements as covariates. P values were deemed significant when below 0.05. When there was significant effect of diet, orthogonal polynomial contrasts were performed to test for linear, quadratic and cubic responses to increasing desmanthus proportion. The quadratic and cubic responses were dropped from the model because they were not significant for all variables tested. PROC CORR procedure fitted with Spearman’s ρ test was used to calculate the residual correlations between diet, rumen and plasma metabolite parameters. Baseline rumen and plasma metabolite values were excluded in the correlation analysis because the quality of the pasture that cattle grazed before the study commenced was not analyzed. # Results ## Diet quality, intake and growth performance The nutritive values of the forages are shown in and the four experimental diets are presented in. Rhodes grass had the lowest CP and highest fibre contents compared to the legume forages. The lowest DMD and ME values were obtained in the Rhodes grass and JCU7 cultivar. The diets were formulated to be isonitrogenous hence the similar CP (*P* = 0.84). Diet NDF levels were similar (*P* = 0.40) but significant differences were observed in diets ADF, hemicellulose and ME levels (*P ≤* 0.001). There was a linear increase in ADF and decrease in hemicellulose and ME with increase in desmanthus proportion. The DMI and nutrient intake data are presented in. Increasing the desmanthus levels in the diet decreased the DMI and subsequently CP, NDF, ADF, hemicellulose and ME intake (*P* ≤ 0.008). The DMI decreased from 8.8 to 7.6 kg/day with increasing desmanthus proportion from 0 to 45% (*P* = 0.009). There were no significant differences in LW, BCS, ADG and feed to gain ratio observed between diets. At the end of the study steers weighed between 419 to 434 kg (*P* = 0.21) with feed to gain ratio of 12.9 to 14.6 kg DMI per kg weight gained (*P* = 0.31). ## Rumen and plasma metabolite parameters Rumen pH and metabolites data of steers fed increasing desmanthus level*s* in the diet are presented in. A linear decrease in total VFA with incremental levels of desmanthus in the diet was observed (*P* = 0.026). The rumen pH and all the other metabolites measured were similar for all the diets (*P* ≥ 0.076). Proportion of desmanthus in the diet had no effect on plasma NEFA, BHB, glucose, creatinine and total bilirubin (*P* ≥ 0.067). Medium to high residual correlations (0.41–0.83) were observed between diet parameters, rumen VFA and plasma metabolites. Diet ADF was correlated to all rumen metabolites except NH₃-N and n-caproate. A negative correlation between diet ADF with total bilirubin and NEFA was observed. Diet CP correlated positively with rumen NH₃-N and propionate but negatively with acetate/propionate ratio, n-caproate and creatinine (*P* \< 0.05). # Discussion This study evaluated the effect of incremental levels of desmanthus in isonitrogenous diets on the feed intake, rumen fermentation, plasma metabolites and growth rate of tropical crossbred beef steers. ## Diet quality, intake and growth performance Rhodes grass and lucerne quality indicators were within the range reported in other studies. Rhodes grass had lower CP and higher fibre content compared to legume forages. The observed difference agrees with previous studies that had reported grass to have lower quality than legume forages. The CP value of *D*. *virgatus* and *D*. *bicornutus* were within the values reported in previous studies carried out in northern Australia. The CP content for *D*. *leptophyllus* was comparatively lower than the CP of *D*. *virgatus* and *D*. *bicornutus* but similar to that reported by Suybeng et al.. Differences in values between studies may be due to soil fertility, climate, plant fraction and stage of maturity at harvest. The quality differences between species may be due to differences in plant characteristics. For instance, desmanthus ranges widely from early to late maturing and herbaceous to suffruticose plant types. The fibre content was lower than that reported by Suybeng et al. but higher than that reported by Durmic et al.. This difference may be due to the maturity stage and plant part collected. Unlike this study where the whole plant was used, Durmic et al. collected the leaves and only 5cm long stems to mimic cattle grazing. Significant variation in plant part nutritive value has been widely reported. Diets were formulated to ensure that they were isonitrogenous but increasing desmanthus proportion in the diet increased ADF content and reduced NDF and DM intake. High indigestible fibre content reduces voluntary feed intake in ruminants due to low digestibility and high rumen fill effect. Dairy cows fed grass silage of varying maturity were observed to eat less as the ADF content increased. Dietary fibre content has been widely used to predict digestibility. A negative correlation exists between indigestible fibre content and digestibility. However, DMD is reported to restrict rumen fermentation and DM intake when the DMD:CP ratio exceeds 8 to 10, but in this study, the ratio ranged between 8.5 and 8.7 indicating that DMD was not likely to limit rumen fermentation and feed intake. Confined *Bos indicus* crossbred beef cattle weighing 400kg fed an 8 MJ/kg DM diet require 40 MJ ME daily for maintenance and 69 MJ ME for 0.5 kg ADG. In this study, the crossbred steers consumed 61–72 MJ ME daily; hence weight gain was expected. Steers in all experimental diets gained at least 0.5 kg/day. Typically, the annual ADG of cattle grazing northern Australian native grass pastures is 0.3 kg/day. Poppi et al. reported that ADG ranging between 0.4 and 0.6 kg can be achieved using improved forage, and these gains are adequate to meet the target slaughter weight required for the prime beef market within 2.5 years of age. The observed significant drop in feed intake as the desmanthus level increased did not result in a significant difference in LW. The reduced feed intake may have improved digestibility by reducing the passage rate of digesta from the reticulorumen. Besides, desmanthus is a tannin-containing legume. Condensed tannins can form complexes with proteins that protect CP from microbial degradation in the rumen, thus enhancing lower gastrointestinal tract digestion. Lack of significant difference between diets indicates that desmanthus alone or mixed with other high-quality legume forages can be used to supplement grass-based diets of beef cattle in northern Australia without any detrimental impact on productivity. Supplementation of grass-based grazing cattle with desmanthus was reported to improve growth rate. In contrast to our findings, incremental levels of up to 31% desmanthus resulted in only 0.2 kg ADG in confined steers even when 11.8% dietary CP was achieved. The authors attributed the low weight gain to low DM intake. The steers consumed 1.4% DM/LW compared to at least 2% in this study. The feed to gain ratio in this study (12–14) was similar to values reported for cattle fed high forage diets of similar CP. Continental crossbred steers fed a 70:30 forage to concentrate diet with 13% CP had feed to gain ratio of 10.4–16.9, while *Bos indicus-Bos taurus* crossbred cattle fed 11% CP diet consisting of 5–24% concentrate supplement had 9.4–13.5 feed to gain ratio. ## Rumen and plasma metabolites VFAs play a major role in ruminant nutrition. They contribute 60–70% of metabolizable energy. The VFA levels vary with animal feeding patterns and diet composition. Highly fermentable diets may raise rumen VFA levels to 200 mM, with the peak occurring 2 to 4 h after feeding. However, hay diets produce smaller fluctuations throughout the day and concentrations of less than 100 mM are common. Previous studies reported increases in total rumen VFA with increases in diet digestibility. This agrees with our findings of a high negative correlation between ADF, a marker of diet digestibility, and total VFA. Total VFA concentrations decreased with an increase in desmanthus level in the diet. This could be due to several factors: a) The decrease in DMI associated with higher proportion of desmanthus in the diet increased ADF and reduced feed intake. The increase in ADF was associated with a decrease in total VFA concentration, in line with previous observations in a dual-flow continuous culture system where high levels of indigestible fibre resulted in low concentrations of rapidly fermentable carbohydrates and reduced VFA concentration. b) Condensed tannins may have formed complexes with proteins availing proteins for digestion in the lower gastrointestinal tract. This may increase microbial protein synthesis efficiency in the rumen. On the other hand, lucerne is highly fermentable in the rumen, hence the high total VFA. Rumen NH₃-N levels of 10 mg/dl are required to maintain effective rumen microbial activity and maximize voluntary DMI in cattle fed low-quality tropical forage. These levels (10.4–12.2 mg/dl) were attained in all the diets in this study. The molar proportion of individual VFA is influenced by dietary composition and DMI. An increase in the proportion of dietary forage or fibre leads to an increase in acetate and decrease in propionate, butyrate, iso-butyrate, valerate and iso-valerate molar concentrations. Similar to our study, cattle grazing pasture of varying digestibility had similar rumen molar percentages of acetate, propionate, butyrate, iso-valerate and valerate. However, in contrast to our study, they reported an increase in iso-butyrate concentrations with increased diet DMD. However, the increase was numerically small and considered to be of no biological significance. All plasma metabolites measured in this study were not affected by the dietary desmanthus levels. Cattle fed diets of similar CP as in this study had similar glucose and NEFA concentrations. Plasma NEFA originates from the mobilisation of stored fat. The similar NEFA concentration between diets may indicate that the steers had comparable energy balance. In ruminants, negligible levels of glucose are absorbed across the portal-drained viscera from dietary sources. Propionate plays a major role as a precursor for glucose synthesis, contributing 46–73% to hepatic gluconeogenesis in cattle. The similar molar proportion of propionate in the current study may have resulted to similar glucose levels between the diets. All plasma metabolites measured were within the normal range reported for beef cattle. Similar levels of plasma metabolites indicate that an inclusion level of up to 45% desmanthus in the diet does not cause adverse effects on energy metabolism and health status of supplemented steers. Overall, increasing the proportion of desmanthus in the diet did not influence LW, ADG, VFA molar proportion and plasma metabolites of steers fed isonitrogenous diets. These findings indicate that desmanthus can be used as a renewable protein source for beef cattle in the northern Australian beef cattle production system, particularly in the tropical semi-arid clay soil environments where lucerne is not adapted. The northern Australian beef industry serves both the pasture-fed and grain- finished beef markets. It is common practice for cattle backgrounded on pastures to be finished in the feedlot on energy-dense grain diets for short periods of about 100 days before slaughter. Further studies are required to determine the effect of desmanthus supplementation on feedlot performance and carcass quality of both pasture and feedlot finished cattle. # Conclusion This study aimed to evaluate the weight gain, rumen fermentation and plasma metabolites of tropical crossbred steers in response to supplementation with incremental levels of desmanthus forage legume in isonitrogenous diets. The results showed similar weight gains, VFA molar proportion and plasma metabolites, but a decrease in total VFA with increase in dietary desmanthus levels. Hence, the hypothesis that cattle fed isonitrogenous diets supplemented with different desmanthus level*s* will have similar growth rates, rumen fermentation and plasma metabolites was accepted. The results indicate that desmanthus alone or in combination with other high-quality legume forages can be used to supplement grass-based diets of beef cattle in northern Australia. However, more studies are required to examine the effect of desmanthus supplementation on cattle feedlot performance and carcass quality. The authors thankfully acknowledge the support provided by Wayne Flintham, Heitor Fleury, Melissa Mathews, Holly Reid, Steve Austin, Jess Simington, Stefania Maffei, Khalu Tomachy, Paulo Delbone, Shedrach Pewan and Ewerton Delbone during cattle management, feeding and sampling. The technical support provided by Wendy Smith and Stuart Denman for rumen metabolites analysis, Elizabeth Hulm for pasture analysis and Jemma Green for plasma metabolites analysis is highly appreciated. The authors also acknowledge the College of Public Health, Medical and Veterinary Sciences of the James Cook University, Department of Industry, Innovation and Science, Agrimix Pastures Pty Ltd, CSIRO and Meat & Livestock Australia Ltd. [^1]: The authors have declared that no competing interests exist.

# Background Serum uric acid (SUA) has long been associated with dyslipidemia, diabetes, hypertension, coronary calcifications and renal failure. It has been suggested that elevated SUA may be associated with mortality in high risk patients with asymptomatic carotid atherosclerotic disease, acute myocardial infarction, heart failure, diabetes mellitus and hypertension. The role of SUA as an independent risk factor for CV mortality is questionable. It is not clear whether SUA has a causal role in the development of CV disease and death or whether the association is circumstantial. It is possible that SUA correlates with CV risk factors and may reflect an underlying CV disease. Data regarding the association between SUA and long-term mortality are mainly based on a single baseline SUA measurement. Recently, we have shown an association between baseline SUA and long-term mortality from stroke, coronary heart disease (CHD) and all causes mortality. These associations were based on a single baseline SUA measurement. It has been suggested that elevated SUA levels may blunt renal auto-regulation of blood pressure and cause endothelial dysfunction. If SUA is related to blood pressure levels, then SUA levels may fluctuate in parallel to blood pressure fluctuations. Several studies have shown an association between blood pressure variability and CV morbidity and mortality. We therefore designed this study to further evaluate the association between the visit-by-visit variability of SUA levels and CV and all-cause mortality. # Methods The current study is based on the IIHD study, which was conducted as a collaborative project of the National Heart and Lung Institute, NIH, USA, the Israel Civil Service Commission and the Hadassah Medical Organization, at the beginning of the 1960 sec, a time when Ethical Review Boards did not yet exist in Israel. However, all participants had given their oral consent to take part in the study upon their recruitment in 1963 following explanations regarding the study objectives and the long-term follow up. In addition, the Tel Aviv University Ethical Review Board approved of the linkage of the IIHD database with the Israel Population Registry. Furthermore, Prof. Goldbourt, an author in this publication, is legally responsible for the IIHD study database and has approved its use for the purpose of this study. ## Study participants The original cohort included 10,059 individuals, who were recruited by stratified sampling of civil servants and municipal employees in 1963, based on the following inclusion criteria: 1. Men aged 40 years or older. 2. Work place limited to the three largest urban areas in Israel. The sample size was aimed at obtaining sufficient number of participants from six areas of birth which were proportional to the Israeli male population of this age. Participants underwent clinical, dietary, psychological and blood biochemical evaluations in 1963, 1965 and 1968. Further details of the study have previously been described. Patients with missing measurement of SUA in one of the evaluations (n = 1237) were not included in the study. Thus, the present analysis includes 8822 participants who had 3 measurements of SUA and baseline assessment of diabetic and coronary morbidity status. ## Mortality data Follow-up since the last evaluation on 1968 lasted 18 years. Data regarding death was derived from the Israeli Mortality Registry. The underlying cause of death was documented on the basis of case-by-case determinations by a review panel until 1970 and by the use of International Classification of Diseases (ICD) codes thereafter. ## Uric acid measurements Blood for measurement of SUA levels was drawn on each evaluation. SUA was measured by Fister's adaptation of the colorimetric method using phosphotungstic acid in the presence of cyanide and urea. All determinations were performed in duplicate and the mean result of the two tests was used for analysis. SUA variability was defined as the standard deviations (SD) of Z-scores of SUA across study visits. ## Risk factor assessment Blood pressure (BP) was measured with a standard mercury sphygmomanometer, taken twice in the supine position, with a time interval of 15–30 between both measurements. The second measurement was used for the analysis. Non-fasting cholesterol was measured using the Anderson and Keys modification of the Abel method. Cigarette smoking was self-reported, and was classified as ever-smoked or not. Additional variables included diabetes mellitus (DM), body mass index (BMI) and history of CHD, defined as confirmed angina pectoris, documented hospitalization for myocardial infarction, or electrocardiographic pattern of an old infarction. ## Statistical analysis We conducted analysis examining whether the standard deviations (SD) of Z-scores of SUA across study visits predicted CHD- and stroke- as well as all-cause mortality. SUA-Z was defined as the difference between the individual SUA and the mean of SUA, divided by the standard deviation (SD) for the pertinent examination, namely separately for the 1963, 1965 and 1968 means and SD. Hazard ratios (HR) were calculated applying Cox proportional hazard models. The proportional hazard assumption was assessed applying Schoenfeld residuals. HR associated with the SD of SUA-Z were calculated for 18-year stroke-, CHD- and all-cause mortality associated with quartiles of the above variability. The lower quartile served as the referent, adjusting for age. A subsequent model adjusted additionally for the baseline value of SUA as well as for baseline frequency of diabetes mellitus and CHD. Stratified rate ratios were calculated using a Mantel-Haenszel-type method. This was used to carry out trend tests for an increment of one quartile of SD of SUA-Z. The extent by which including the visit-by-visit SUA variability as a predicting factor of mortality adds to the prediction of all-cause and CHD mortality was estimated by Harrell's C concordance index and by Somers' D index. Statistical analysis was carried out using the STATA statis-tical package, version 15 (STATA, College Station, TX). Significance was considered when the p value was \< 0.05. # Results ## Subject characteristics A total of 8822 men were included in the analysis. Participants were divided according to quartiles of SUA variability during the years 1963–1968. Baseline characteristics are presented in. About two thirds of the participants had smoked at any time. Participants had a mean BMI of 25.7 ± 3.3 Kg/m², their baseline SUA was relatively low and the baseline rates of diabetes mellitus and CHD were low. Participants in the 4^th quartile had higher body mass index and initial SUA levels. ## Association between serum uric acid variability and long-term mortality During the follow up 2893 subjects died, 830 of them died from CHD and 292 died from stroke. Rates of all-cause mortality and mortality from stroke and CHD per 10000 person years are given in. Stroke mortality was not associated with SUA variability. However, the rate of CHD mortality and all-cause mortality was higher in the 4^th quartile of the SUA variability. Multivariate analysis of 18-year CHD mortality yielded a significant association between SUA variability during the years 1963–1968. Similarly, the results for all-cause mortality showed increasing age-adjusted mortality with increasing SUA variability. Kaplan-Meier survival curves for all-cause mortality and CHD mortality as a function of SUA vaiability are presented in. In tests for trend, the average risk ratio associated with a rise in one quartile was 1.14 (95%CI; 1.06–1.23) for CHD mortality and 1.10 (95%CI; 1.06–1.23). Sensitivity analysis, incorporating the last (1968) SUA levels assessed, rather than the 1963 ones, yielded virtually identical HR. Estimating the differentiation advantage associated with inclusion of the visit-by-visit SUA variability yielded the following results: Harrell's C concordance index rose from 0.703 for all-cause mortality and 0.718 for CHD mortality, excluding the above variability, to 0.720 and 0.743 for all-cause and CHD mortality, respectively. Corresponding results for the Somers' D index yielded estimated rises from 0.406 and 0.436 to 0.440 and 0.486, respectively. # Discussion In the present study, we showed for the first time that variability in SUA levels is associated with long term CHD and all-cause mortality. It is well established that SUA is associated with cardiovascular risk factors, including hypertension, hyperlipidemia, obesity and renal failure. It has also been shown that SUA is associated with CHD. The role of SUA as an independent risk factor for CVD and death is controversial. Several studies have investigated this issue, revealing conflicting results. Culleton et al. found that among women, SUA was predictive of CHD, death from CV disease and death from all causes. However, after adjustment for CV risk factors, SUA was no longer associated with CHD, death from CV disease, or death from all causes. These findings indicate that SUA does not have a causal role in the development of CHD or CV mortality. Other studies have demonstrated that SUA levels are an independent predictor of CV mortality. In the present study, we aimed to investigate the association of SUA variability with long-term all-cause and specific-cause mortality. The meaning of risk factors variability has been previously investigated with regard to several parameters. Blood pressure variability was associated with CV morbidity and mortality. Further studies have demonstrated that glycemic and blood pressure variability may be independent risk factors for the development of albuminuria and for decreased glomerular filtration rate (GFR) in patients with type 2 DM. Fluctuations in body weight are associated with coronary death. A recent study showed an association between body weight variability and CV events in patients with coronary events. Apparently it seems that fluctuations in hemodynamic and metabolic parameters are associated with poor outcome. A possible explanation for the association of variation in blood pressure and body weight with increased CV risk is that the need to accommodate to blood pressure or body weight fluctuations requires energy recruitment through activation of the sympathetic nervous system and the renin angiotensin system. In the present study we have shown that SUA variability between the years 1963–1968 was associated with CHD mortality and all-cause mortality beyond and above SUA levels properly. Data regarding the significance of SUA variability are scarce. Ceriello et al. investigated the impact of variability of HbA1c, systolic and diastolic blood pressure, cholesterol, triglycerides and SUA levels on developing chronic kidney disease among diabetic patients. They found that high variability in all the aforementioned parameters predicted the decline in GFR. High variability in SUA levels conferred the highest risk of decline in GFR. Microalbuminuria and reduced kidney function serve as surrogate markers for CV disease in diabetes, as both are independent risk factors for CV events among patients with type 2 DM. These findings are consistent with our findings that SUA variability is an independent risk factor for CV events. It is unclear whether SUA is an independent CV risk factor or reflects an underlying CV disease. If SUA was an independent risk factor for CV events, then thiazide diuretics, which increase SUA, should increase, and drugs, which lower SUA, should decrease CV morbidity and mortality. Unfortunately, there is no evidence for the detrimental effects of diuretic and beneficial effects of SUA lowering agents on CV outcomes. Thus, it is more likely that SUA reflects an underlying CV disease. If we assume that SUA is a marker of metabolic and hemodynamic changes, then we can understand why SUA variability is associated with CHD and all-cause mortality. Several additional explanations for the effect of SUA variability may be suggested. Uric acid emerged as an inflammatory factor that increases oxidative stress. Therefore, varying levels of SUA may be associated with increased oxidative stress, which can contribute to excess CV risk. Another plausible explanation relies on the anti-oxidant role of uric acid. It has been found that uric acid has anti-oxidant effects, suggesting that varying SUA levels might reflect a compensatory mechanism to counter oxidative stress which constitutes a major risk factor for the development of CV disease. It is likely that SUA variability is parallel to variability of cardiovascular risk factors such as hypertension, diabetes mellitus, hyperlipidemia and renal failure. Therefore variation in SUA levels at a relatively young age may reflect development of other CV risk factors, more significantly than persistent hyperuricemia by itself. The practical implication for the findings in our study is that when assessing SUA levels, consecutive annual measurement should be observed. Patients with varying levels should be followed up more closely for the existence and development of CV risk factors. Our study has several limitations. First, our study was comprised solely of men. Therefore, the conclusion of our study may not apply to women. The findings in our study apply only to a heterogeneous relatively young population. Another limitation is the lack of information regarding drug therapy during follow up which may have affected SUA levels. We were forced to use the SD of SUA-Z score as UA variability measure rather than the SD or coefficient of variance of UA, since we observed a left shift of SUA at 1963 compared to other two observations. We also did not adjust SUA variability to alcohol and diuretic use which can affect SUA levels. However, the rate of alcohol intake and diuretic use were very low. The major strength of our study is the long-term follow up of the cohort. In addition, to the best of our knowledge, this is the first time that clear association between SUA variability and long term CHD and all-cause mortality was assessed. In conclusion, SUA variability may add to the risk assessment of CHD and all-cause mortality. # Supporting information [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Department of Internal Medicine F And the Rheumatology Unit, The Chaim Sheba Medical Center, Tel-Hahsomer, Israel

# Introduction *Leishmania infantum* is the etiological agent of zoonotic visceral leishmaniasis in the Mediterranean basin, where dogs are the main reservoirs. A recent outbreak in humans has been described in Spain, where the main vector is *Phletobomus perniciosus* (Psychodidae: Phlebotominae). Sand flies are the blood-feeding vector hosts in the life cycle of the parasite. Promastigote development takes place within the gut of the sand fly simultaneously to migration towards the anterior gut, whereas blood components are progressively digested, leading to nutrient depletion. Chemotaxis and osmotaxis promote directed migration. After development, promastigotes are released into the dermis of the mammalian host during blood feedings. Then, they differentiate to the amastigote stage within host phagocytic cells. Eventually, a sand fly feeds from the infected mammalian host and amastigotes are released to the mid gut, where they become motile promastigotes. The promastigote stage is generally cultured in complex undefined liquid media at 26–27°C, which mimics to some extent the conditions of the sand fly gut microenvironment. Mammalian serum provides complex nutrients in cultures, thus improving growth kinetics. Inactivation of serum is performed by heating at 56°C for 1 h. This procedure avoids lysis of promastigotes by proteins of the complement system. The parasite is metabolically versatile because it is able to use amino acids, fatty acids or glucose as the major carbon source and consequently, adapt to different environments. However, culture may affect differentiation in some aspects. The aim of this study is the evaluation of general and specific consequences of serum deprivation for cultured promastigotes, including growth rate, ploidy, infectivity and differential gene expression. # Materials and Methods ## Promastigote cultures The *L*. *infantum* MCAN/ES/98/10445 (zymodeme MON-1) isolate was used in this study. Promastigotes were cultured in triplicate at 27°C in complete medium (CM) or in heat inactivated fetal bovine serum (HIFBS)-depleted medium for 4 days. CM consists of RPMI 1640 supplemented with 2 mM glutamine (Lonza, Karlskoga, Sweden), 10% HIFBS (Lonza) and 100 UI/ml penicillin-100 μg/ml streptomycin (Life Technologies, Carlsbad, CA). Cell recovery from cultures was performed by centrifugation at 2,000g for 10 min. Morphology was routinely evaluated at the light microscope (40X). For this purpose, 10⁷ promastigotes were harvested, washed in PBS and resuspended in 1 ml PBS. A 10 μl aliquot was deposited between a slide and a coverslip. Cell counting was performed at the light microscope in a Neubauer chamber (40 X) after diluting 20 times an aliquot of promastigote culture in 0.5 M EDTA. Differences in growth of HIFBS-depleted and CM promastigote cultures were compared by the Student's t-test. ## Cell cycle analysis by flow cytometry Samples of 50 x 10⁶ promastigotes were harvested for cell cycle analysis. Three biological replicates of the experiment were performed. G1 arrest was achieved by 6h treatment with 0.8 mg/ml hydroxyurea in fresh CM or HIFBS-depleted medium (Sigma-Aldrich, Basel, Switzerland). Thereafter, the cells were centrifuged, washed three times with PBS and fixed with 1 ml cold 70% ethanol at -20°C for 30 min. Next, promastigotes were harvested, washed twice with PBS and incubated for 30 min in 0.5 ml of a solution containing 50 μg/ml propidium iodide (PI) and 100 μg/ml RNase A (Sigma-Aldrich) in PBS. PI uptake was analyzed in a FACSCalibur^™ flow cytometer using CELLQuest^™ software (Becton Dickinson, Franklin Lakes, NJ) by gating promastigotes at forward-angle versus side-angle light scatter and registering fluorescent emission collected in the FL2-A detector through a 585 nm band pass filter. ## Evaluation of *in vitro* infection of stimulated U937 cells with promastigotes. The non-adherent human myeloid cell line U937 (ATCC^® CRL1593.2), originally obtained from pleural effusions of a patient with histiocytic leukemia, was cultured at 37°C in complete medium in the presence of 5% CO₂. After 72 h, 2 x 10⁵ cells/cm² were harvested at 250g for 10 min and stimulated with 20 ng/ml phorbol 12-myristate 13-acetate (Sigma-Aldrich) in CM over 8-well cell chamber slides (LabTek, New York, NY). This treatment allows differentiation to macrophage-like cells. Then, the wells were mildly rinsed with CM and the infection step performed by incubating cells at 37°C for 2 h in 400 μl of CM containing *L*. *infantum* promastigotes. The promastigote:cell ratio was 5:1. Next, cells were washed three times with CM to remove remaining promastigotes. Infected cells were then incubated in CM at 37°C, 5% CO₂ for 72 h. Three final washes were performed before treatment with hypotonic solution (180 μl CM diluted with 220 μl water per well) for 5 min. Four washes were carried out with 150 μl ethanol- acetic acid 3:1 after removing the hypotonic solution. Fixation was carried out with the same solution for 10 min and this step was repeated three times. Finally, cells were allowed to air dry and the wells removed from the slide. Staining was performed with Diff-Quick^® Stain Solution I and II (Dade Behring, Marburg, Germany). The preparations were washed with distilled water, air dried and mounted with Entellan^® Neu (Merck, Darmstadt, Germany). The average infection rate and the number of amastigotes per infected cell were assessed at the light microscope (40X) and contrasted by the Student’s t-test. ## Isolation of total RNA and protein Total RNA extractions were immediately performed with TRizol^® reagent (Life Technologies) following the manufacturer´s instructions. Whole protein was obtained from 10⁸ cells by mild agitation at 4°C during 30 min in 50 mM Tris-HCl pH 7.4, 2 mM EDTA and 0.2% TritonX-100 in the presence of a cocktail of protease inhibitors (Roche, Mannheim, Germany). Then, samples were centrifuged at 8,000g at 4°C for 10 min. Protein concentration was estimated by the Bradford method. ## Western blot SDS-PAGE of protein extracts was performed at 12 mA for 30 min, then at 30 mA for 90 min, in 8% slab gels casted in a MiniProtean II Cell system (BioRad, Hercules, CA). Twenty μg protein extract was loaded per well including 1 μl Benzonase Nuclease HC (Novagen, Madison, WI). Proteins were then transferred to 0.45 μm nitrocellulose membranes (BioRad) at 100 V for 1 h in a Mini Trans-Blot Cell wet transfer system (BioRad). Then, the membranes were blocked with 5% skimmed milk in PBS-0.1% Tween 20 (Sigma) for 1 h and washed three times with PBS-0.1% Tween 20 for 15, 5 and 5 min respectively. After that, membranes were incubated with 1∶500 of rabbit anti-LACK polyclonal serum diluted 1:500 in blocking solution for 2 h. A monoclonal mouse anti-*L*. *mexicana* gGAPDH antibody kindly provided by Paul Michels (University of Edinburg) was also included in the mixture at 1:10,000 dilution. The washing steps were repeated. Next, 90 min incubation was carried out with 1∶2,000 HRP-conjugated goat anti- rabbit IgG (DAKO, Ely, UK) and the membrane was washed again. Finally, ECL-based detection (GE Healthcare, Pittsburg, PA) and development was performed in X-ray film according to the manufacturer's instructions. Densitometry was performed with Gel Doc XR System and Quantity One version 4.6. software (BioRad) and the intensity data of both groups were compared by the Student's t-test including three biological replicates. ## Microarray hybridization analysis Differential gene expression was analyzed by shotgun DNA microarray hybridization experiments as described. Briefly, mRNA was amplified from three replicate samples using MessageAmp^™ II aRNA Amplification Kit (Life Technologies) and cyanine-labelled cDNA was synthesized (Life Technologies). Then, combined Cy5/Cy3 cDNA microarray hybridizations were carried out in triplicate (HIFBS-depletion/CM). The slides were scanned with a GenePix 4100A instrument (Axon, Foster City, CA) and raw medians of fluorescence intensity values obtained were normalized by the LOWESS per pin method. Differential transcript abundance was inferred by the paired Student’s t test. The spot selection criteria were: fold-change (F) \> 1.7 (Cy5/Cy3; Cy5 \> Cy3) or \< -1.7 (-Cy3/Cy5; Cy5 \< Cy3); fluorescence intensity value 10-fold higher than the substracted background; p \< 0.05. Clones that fulfilled these cutoff values were recovered from the genomic library that was used for microarray construction. The clone ends sequenced with the M13 universal primers and they were assembled by alignment with the whole-genome *L*. *infantum* sequence as detailed in. Depending on mapping and assembling outcomes, clones were classified in type a (congruent alignments, unique pair of alignments), b (congruent alignments, more than one pair of alignments) and c (uncongruent alignments or lack of one insert end read). ## Real time quantitative RT-PCR analyses (qRT-PCR) Unlabelled single-stranded cDNA was synthesized following the same procedure as for labelled cDNA but using a mixture of unlabelled dNTPs (10 mM each). Custom TaqMan^® FAM-MGB Assay-by-Design primers and probes (Life Technologies) are provided in the and amplification with TaqMan^® Universal Master Mix 2x (Life Technologies) was run in a 7900HT Fast Real Time PCR system using SDS 3.1 software (Life Technologies) following the manufacturer’s instructions. The gGAPDH was the reference gene and the fold- change values were calculated with efficiency-corrected normalized quantities. ## BCAT activity assay The assay was performed with the substrate 3-methyl-2-oxopentanoate and L-glutamate was added as the co-substrate. All reagents and enzymes were purchased from Sigma-Aldrich. Incubations were carried out in quadruplicate at 25°C for 30 min in 1ml buffered (0.1M Tris-HCl, pH8.3) reaction mixture containing 2.1 mM 3-methyl-2-oxopentanoate, 300 mM L-glutamate, 0.2 mM NADH, 0.1 mM pyridoxalphosphate (PLP), 200 mM L-aspartate, 200 μmol/min/l L-aspartate aminotransferase (ASAT), 1 mmol/min/l L-malate dehydrogenase (MDH) (both enzymes from porcine heart) and 150 μg protein extract. Negative controls without NADH were also set. The time course consumption of NADH was monitored by measuring the decay of absorbance at 334 nm continuously with a Cary 4000 spectrophotometer (Agilent Technologies, Santa Clara, CA). The results were contrasted by the paired Student's t test. # Results ## Serum depletion considerably affects growth rate and ploidy As expected, pronounced decrease of proliferation is observed when the promastigote culture does not contain HIFBS. Changes in promastigote morphology were not detected, except for more frequent lack of elongation of the fusiform cell body in most HIFBS-depleted promastigotes. Tetraploidy was observed also in this case. ## *In vitro* infection ability of HIFBS-depleted and CM *L*. *infantum* promastigotes The average infection rate and the number of amastigotes per infected U937 cell were evaluated at 24 and 48 h after infection with CM and HIFBS-depleted promastigotes. The average infection rate is 54% with CM and 43% with HIFBS- depleted promastigotes after 24 and 48 h. Cells infected with CM promastigotes contained 3.3 ± 0.3 amastigotes at 24 h and 5.0 ± 0.0 at 48 h, whereas the results were 2.1 ± 0.1 at 24 h and 4.2 ± 0.1 at 48 h in the case of cells infected with HIFBS-depleted promastigotes. These differences are statistically significant according to the paired Student's t-test outcome. Therefore, serum depletion decreases the infection ability of promastigotes in terms of infection rate and number of amastigotes per infected cell. ## Serum depletion slightly affects gene expression at the transcript level in *L*. *infantum* promastigotes Total RNA isolated from HIFBS-depleted and CM promastigotes was not degraded. mRNA from all samples was successfully amplified. Differential expression of the control genes included in the microarrays was not detected. The A2 gene is not differentially regulated and the flagellum remains emergent from the cell body. These facts indicate that HIFBS-depleted promastigotes do not undergo the developmental process to the amastigote stage, as well as CM control promastigotes. Also, expression of the LACK antigen remains constant in promastigotes at the transcript and protein levels when they are depleted from serum. The differential gene expression rate is 0.4% (33 genes out of 8154 coding genes) in HIFBS-depleted promastigotes with respect to CM. Genes involved in DNA repair, gene expression regulation, protein folding, proteolysis, metabolism, detoxification, signalling, the flagellum and the surface coat are up-regulated in HIFBS-depleted promastigotes. Differential regulation of five genes was validated by qRT-PCR, which also solved two clones that represent more than one gene sequence. By contrast, two genes tested by qRT-PCR are negative for differential expression (clones Lin138C1 and Lin309D1). Consequently, at least one of the remaining genes represented in these clones are differentially regulated. ## The differentially regulated metabolic genes up-regulated under serum depletion code for enzymes catalyzing rate-limiting reactions The glycosomal phosphoenolpyruvate carboxykinase (gPEPCK) and the arginase are down-regulated under HIFBS-depletion, whereas the genes coding for the 4-coumarate-CoA ligase (4CCL), the fatty acyl-CoA synthetase 1 (FAS1), the methionine synthase reductase (MTRR), the p-nitrophenylphosphatase (PNPP) and the methylmalonyl-CoA epimerase (MMCE) are up-regulated. ## Branched-chain amino acid aminotransferase (BCAT) activity slightly but significantly decreases in HIFBS-depleted promastigotes BCAT activity was measured using 3-methyl-oxopentanoate as the substrate and glutamate as the co-substrate in direct relation to NADH (A_334nm) decay. Background activity was not observed in negative control reactions set in the absence of NADH. The enzyme activity (EA) measured as nmol/min/mg soluble protein is about 1.6-fold higher in HIFBS-depleted promastigotes that in reference CM promastigotes. The differences observed are statistically significant (paired Student's t test, p = 0.022). # Discussion The presence of complement proteins contained in serum drastically decreases viability of promastigotes. In fact, only about 3% of stationary phase promastigotes survive when they are transferred to normal human serum. Therefore, heat inactivation of serum is essential for appropriate culturing of *Leishmania* promastigotes. Serum provides complex nutrients for appropriate growth, despite heat inactivation affects thermolabile nutrients such certain vitamins and amino acids. However, complete medium for promastigote culturing includes the defined medium RPMI 1640, which contains all protein amino acids and required vitamins. Some proteins are probably denatured at 56°C, whereas some others are not (e.g. albumin, which coagulates at higher temperatures). Lipids are denatured at higher temperatures than proteins. In fact, the major effect observed below the denaturation temperature range is not conformational transition of a single molecule but a change in the conformation of the supramolecular structure. For example, lipid monolayers of bilayers may be disintegrated without a conformational change in each single molecule, i.e. the gel-to-liquid transition due to temperature increase. Anyway, both native and denatured proteins and lipids are also source of their backbone residues as nutrients. Color does not change either, which indicates that the Fe-hemin complex is mantained intact after heating at 56°C. This is very important, as Fe is essential for proper growth of promastigotes. HIFBS-depleted promastigotes are not differentiated to an amastigote-like stage, as revealed by morphology and constant abundance of the LACK protein and the A2 transcript. The reduced growth rate of HIFBS-depleted promastigotes may be related with the down-regulation of the GINS Psf3 gene. The relationship between this expression profile and tetraploidy might be explained by hypothetical impairment of mitosis with cytokinesis. The G1 and G2 peaks are slightly displaced respectively from the 200 and 400 values in the FL2 axis in the case of CM and from the 400 and 800 values in the HIFBS-depletion plot. This is consistent with constitutive aneuploidy observed in the genus *Leishmania*. An unusually low differential expression rate has been reported in all the stages of different *Leishmania* species compared to other organisms. In this context, the differential gene expression rate of HIFBS-depleted promastigotes is even more reduced than usual compared to CM control promastigotes. However, HIFBS-depletion affects important metabolic genes involved in limiting or crucial steps. One of them is the gPEPCK, which is down-regulated not only by HIFBS depletion, but also in metacyclic promastigotes isolated from the stomodeal valve of the sand fly (unpublished result), in amastigotes with respect to cultured promastigotes and by the specific effect of temperature increase plus acidification towards differentiation. These stages and experimental conditions involve nutrient depletion. The glycolytic genes remain constantly expressed at the transcript level in HIFBS-depleted promastigotes and gluconeogenesis is probably less active due to gPEPCK down-regulation. One of the most important energy and carbon sources for amastigotes is glucose obtained from the host, although glucolysis is more active in promastigotes in the absence of starvation and β-oxidation of fatty acids in amastigotes. RPMI is a rich medium containing plenty of glucose (11 mM), which is in agreement with the gPEPCK expression profile found because apparently, gluconeogenesis is not required under these conditions. In contrast, the FAS1 gene is up-regulated in HIFBS-depleted promastigotes, probably because of the absence of the complex lipoid substances provided by serum. MMCE up-regulation suggests that the rate of branched chain amino acid and/or odd-chain fatty acid degradation is higher in HIFBS-depleted promastigotes than in CM. In fact, promastigotes and amastigotes are able to use amino acids as their major carbon sources. No gene directly involved in branched-chain amino acid catabolism is differentially regulated in HIFBS-depleted promastigotes, whereas the MMCE gene is involved in both branched-chain amino acid and odd-chain fatty acid lipid catabolism. For this reason, we aimed to evaluate the BCAT activity as an indication of relative activity of branched-chain amino acid degradation in culture (HIFBS-depletion versus CM promastigotes). The BCAT activity decreases under serum depletion in promastigotes. This finding suggests that MMCE up-regulation is not related to an increase of activity of branched chain amino acid catabolism but to degradation of odd-chain fatty acids. This activity may be linked to the 4CCL. These findings suggest that odd-chain fatty acids may be degraded to allow the biosynthesis of common fatty acids that may be required under depletion of complex lipids contained in serum. Methionine biosynthesis would be also required in HIFBS-depleted promastigotes, as the MTRR is up-regulated in these conditions. The cofactor S-adenosylcobalamine is oxidized over time when it is coupled to the methionine synthase, thus inactivating its activity. The role of the MTRR is keeping the methionine synthase (MTR) active by reverting oxidation of the cofactor. In addition to the amino acids provided in the HIFBS-depleted medium (i.e. RPMI), a possible source would be protein turn-over via the ubiquitin proteasome, which is suggested on the basis of the up-regulation of PSMD8 and PSMB3. To summarize, according to the gene expression profiles, catabolism of sugar, fatty acids and amino acids may remain constant under serum depletion, whereas fatty acid and methionine biosynthesis may be favored and gluconeogenesis may decrease. The PNPP-encoding gene is up-regulated in HIFBS-depleted promastigotes, as well as in axenic amastigotes with respect to promastigotes. The PNPP participates in this cycle and bears the phosphoglycolate phosphatase activity (E.C.3.1.3.18). The glyoxylate cycle is present in *Leishmania* spp. and is probably related with glucolysis, gluconeogenesis and glycine biosynthesis. In this case, PNPP up-regulation may be linked to glucolysis rather than gluconeogenesis because the gPEPCK is down-regulated in HIFBS-depleted promastigotes. Additionally, it may favor survival under nutrient depletion, as the glyoxylate pathway may be present in *Leishmania* to accelerate oxidative catabolism, given the low efficiency of the Krebs cycle in these organisms. As a consequence of the unusual gene expression mechanisms in *Leishmania* spp., post-transcriptional, translational and post-translational regulation are specially important processes in these organisms. The translation factor SUI1 is one of the up-regulated genes in HIFBS-depleted promastigotes. This gene was also found to be up-regulated in amastigote-like forms obtained by increasing temperature and lowering pH but probably not by the presence of heavy metals (e.g. cadmium) Therefore, this translation factor may influence translation control under specific stress situations like nutrient stress, temperature decrease and acidification. The hsp20 and the cyclophilin might participate in post-translational regulation processes provided their up-regulation in HIFBS- depleted promastigotes. The GLO1 is located in the kinetoplast and it is essential for survival. This gene is involved in the ketoaldehyde detoxification pathway. The up-regulation of the GLO1 gene under serum depletion may be one of the mechanisms the parasite displays to maintain certain growth rate, which is actually much slower than in the CM control. It has been described that nutrient depletion is associated to metacyclogenesis (reviewed in). Therefore, up-regulation of the HASPB and the amastin LinJ.34.2600 under serum depletion suggests that the differentiation state of HIFBS-depleted promastigotes is more advanced. In fact, the HASPB is associated to differentiation of promastigotes and amastins constitute a superfamily of proteins of unknown function basically expressed in the amastigote stage (reviewed by). However, the following considerations are in disagreement with a more advanced differentiation stage of promastigotes under serum depletion: i) morphology (i.e. higher frequency of stumpy instead of slender promastigotes); ii) the differentiation process encompasses proper growth in culture as well as within the sand fly gut it mimics, which is not observed in HIFBS-depleted promastigotes; iii) ploidy alteration; iv) down-regulation of the arginase in HIFBS-depleted promastigotes; and v) decreased infectivity *in vitro*. High expression levels of the arginase increases the chance of survival of amastigotes within the host phagocytes. The pre-adaptation hypothesis is essential to understand amastin and arginase expression in promastigotes. This hypothesis consists of a phenotype prepared in advance for differentiation of promastigotes to amastigotes, i.e. invasion of the mammalian host phagocyte. Therefore, an unsuccessful differentiation process takes place in HIFBS-depleted promastigotes, which are less infective than CM promastigotes. A possible explanation is the down-regulation of the arginase gene. # Conclusions In general, the axenic culture model is used to perform biological and biomedical studies concerning *Leishmania* promastigotes. As an insight into the role of inactivated serum in the culture medium, this study has revealed that serum depletion considerably decreases the growth rate of promastigote cultures and leads to reduced infectivity and ploidy alteration. Consequently, only mediums containing the complex nutrients of serum are appropriate in axenic cultures in order to mimic to some extent the natural developmental processes of promastigotes. The effect on the transcriptome is slight in terms of differentially regulation rate but important when gene function is considered (i.e. GINS Psf3, gPEPCK, FAS1, PNPP, MTTR, MMCE, HASPB, arginase and amastin). Down-regulation of the arginase in HIFBS-depleted promastigotes contributes to explain their reduced infectivity. The results discussed herein have provided clues to understand processes and to establish new hypotheses and observations that may be studied in the future. For example, the role of the glyoxylate cycle in these organisms or the elucidation of signal transduction pathways and their connection with stimuli and effector gene expression regulation mechanisms that are completely unknown in these organisms so far. # Supporting Information Our gratitude to Alfredo Toraño, Mercedes Domínguez, Víctor Parro, Manuel J. Gómez and Juan F. Giménez for their support in different experimental procedures. Thanks to Paul Michels for lending us kindly the anti-*L*. *mexicana* gGAPDH antibody. PA thanks CSIC for the I3P-BPD2003-1 grant and two contracts of employment at a position included in the A1 group (respectively developed from January 16^th to July 23^rd 2008 and from October 16^th 2008 to April 15^th 2009). MAD thanks the Spanish Ministry of Economy and Competitiveness for the FPI Pre-doctoral fellowship BES-2011-047361. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: PA AA VL. Performed the experiments: PA AA MM MD IM. Analyzed the data: PA AA. Contributed reagents/materials/analysis tools: PA IM VL. Wrote the paper: PA AA VL.

# Introduction Approximately 1.5 million people in the U.S. have used cocaine within the past month. HIV prevalence among individuals who use cocaine (4–22%) is many times higher than the national average (0.4%). Only about ten percent of individuals who use cocaine inject, and HIV rates among injecting vs. non-injecting individuals who use cocaine are similar, suggesting that risky sexual behavior is the most prominent HIV transmission vector among individuals who use cocaine. Most sexual HIV risk reduction interventions for individuals who use cocaine target HIV risk reduction knowledge and condom use skills. Although these interventions increase knowledge and skills, meta-analyses have demonstrated less robust effectiveness in reducing risk behavior. The observation that individuals who use cocaine continue to engage in risky sexual behavior despite knowledge/skills improvements prompts examination of other factors that may underlie risk behavior, including decision-making processes. Delay discounting provides a useful framework for examining relations between decision making and risk behavior. Delay discounting is a concept from the field of behavioral economics describing how delaying an event reduces that event’s value or impact on behavior. This is shown, for example, by the observation that individuals typically prefer immediate over delayed rewards. Most delay discounting studies ask participants to make choices between receiving smaller amounts of money available immediately vs. larger amounts available after various delays. Steeper discounting of monetary rewards is related to cocaine use (e.g.,) and use of other substances, as well as a variety of non-drug- related problem behaviors, including pathological gambling, obesity, and failing to engage in preventive health behaviors. However, choices between immediate and delayed outcomes involve a variety of reinforcers other than money. For example, in a casual sex scenario, one may prefer to use a condom because it decreases the risk of sexually transmitted infection (STI). However, if a condom is not readily available, the same person might prefer immediate unprotected sex over waiting to obtain a condom. In other words, the value of condom protection may be discounted due to delay. The Sexual Delay Discounting Task (previously referred to as the “Sexual Discounting Task”) was developed to assess the influence of delay on choices related to condom use in casual sex scenarios. Studies using the task in individuals with cocaine use disorders reported several findings. First, individuals with cocaine use disorders generally indicated that they would be less likely to use condoms as the delay to condom availability increased. Second, participants discounted condom-protected sex more steeply for partners with whom they most vs. least wanted to have sex, and for partners they judged least vs. most likely to have an STI. Third, steeper discounting of condom- protected sex was significantly associated with higher rates of self-reported sexual HIV risk behavior. Fourth, the Sexual Delay Discounting Task showed good 1-week test-retest reliability. Together, these findings suggest the Sexual Delay Discounting Task has both external and internal validity among individuals with cocaine use disorders. Despite the reliability and validity of the Sexual Delay Discounting Task within individuals with cocaine use disorders, there are no reports comparing delay discounting of condom-protected sex between individuals with cocaine use disorders and those who do not use cocaine. A recent study using the Sexual Delay Discounting Task demonstrated that opioid-dependent women discounted delayed condom-protected sex and monetary rewards more steeply than non-drug- using control women, and that participants in both groups discounted in an orderly manner that was sensitive to partner characteristics. Moreover, a recent study in 18–24 year old youth found increased delay discounting of condom- protected sex to be significantly associated with greater self-reported drug use. These findings suggest that steeper discounting may be related to higher rates of sexual HIV risk in drug-using populations, but it is unknown whether steeper discounting is related to the higher rates of sexual HIV risk behavior observed specifically among individuals who use cocaine. It is worthwhile to examine the relationship between the discounting of sexual outcomes and cocaine use because individuals who use cocaine have shown higher rates of sexual risk behavior and greater delay discounting of monetary rewards than individuals who use heroin. The present study therefore compared discounting of delayed sexual and monetary outcomes between individuals with cocaine use disorders and matched non-cocaine-using controls. Beyond delay, at least one additional factor that may influence condom use is the probability of contracting an STI with unprotected sex. Indeed, one reason condoms are ever preferred is likely because they decrease the probability of aversive outcomes (i.e., STIs or unwanted pregnancy). Therefore, we also compared probability discounting of sexual and monetary outcomes between the two groups. With methods analogous to delay discounting, probability discounting tasks systematically examine how uncertainty influences an event’s value or impact on behavior. While delay and probability discounting are only weakly correlated, the two processes have shown independent associations with clinically relevant behavior (i.e., gambling) and show differential effects of experimental manipulations (i.e. reward magnitude manipulations show directionally opposite effects on delay and probability discounting). Therefore, delay and probability discounting likely represent separate behavioral processes. With respect to the influence of probability on sexual outcomes, previous studies using the Sexual Delay Discounting Task showed that a partner’s perceived likelihood of having an STI influenced delay discounting of condom use. However, these studies did not explicitly manipulate the probability of STI contraction. We developed the Sexual Probability Discounting Task to quantitatively examine how specified risk of STI contraction resulting from unprotected sex influences condom use. The reliably weak correlations observed in previous studies between delay and probability discounting of monetary rewards also prompted us to examine correlations between discounting in the Sexual Delay and Sexual Probability Discounting tasks. # Materials and Methods ## Ethics Statement Study procedures were approved by the Johns Hopkins Medicine Institutional Review Board 3 (Office for Human Research Protections Registration \#00001656). The study was conducted according to the principles expressed in the Declaration of Helsinki. Written informed consent was obtained from the participants. ## Participants Volunteers were recruited using flyers, Internet, newspaper, and radio advertisements, and word of mouth referral. Inclusion criteria for both the cocaine use disorder (Cocaine) and non-cocaine-using (Control) groups included being at least 18 years of age, having at least an 8^th grade reading level, and reporting having vaginal or anal intercourse with another person during their lifetime. Participants in the Cocaine group met Diagnostic and Statistical Manual of Mental Disorders (4^th edition, DSM-IV) criteria for cocaine abuse or dependence, whereas participants in the Control group reported no lifetime use of cocaine. Participants in both groups could meet criteria for abuse for drugs other than cocaine, but could not meet dependence criteria for other drugs (excluding nicotine and caffeine). Exclusion criteria for both groups included self-reported serious head trauma, dementia, significant cognitive impairment, or diagnosis of major psychiatric disorder besides substance abuse/dependence. ## Procedure After an initial telephone screening assessing basic inclusion/exclusion criteria, initially qualified participants were scheduled for an in-person screening. If qualified, participants remained in the laboratory for approximately four hours to complete a variety of behavioral tasks. During the in-person screening, participants provided informed consent and a urine sample to test for drug use. Participants also completed a demographic questionnaire, a verbal intelligence assessment (Quick Test), a reading comprehension assessment (Wide Range Achievement Test), a lifetime drug use questionnaire, and a checklist to assess current and past drug abuse and dependence. Occurrence and frequency of HIV risk behaviors in the past month were assessed using the HIV Risk-Taking Behavior Scale (HRBS). The HRBS is a psychometrically reliable and valid questionnaire featuring 11 items scored on a 6-point scale (scores of 0–5, with higher scores indicating higher risk) pertaining to injection drug use (6 items) and sexual risk behavior (5 items). Only scores on the sexual risk behavior subscale, which assessed participants’ number of sexual partners in the past month, frequency of condom use with regular and casual partners and when paid for sex, and frequency of anal sex, were compared between groups. Several personality measures and behavioral tasks were also obtained but are not relevant to the present analyses. As in our previous sexual discounting studies, participants then used a computer to view 60 individually-presented color photographs of diverse, clothed people (30 male, 30 female) and were asked to select photographs of individuals that they would consider having casual sex with based on physical appearance. The photographs were assembled from a variety of publicly available online repositories to provide a range of physical appearances that would be conducive to the multiple hypothetical partner conditions described below. Before viewing photographs (and before subsequent sexual tasks), participants were instructed to pretend that they were not in a committed relationship. Next, participants identified from the subset of initially selected photographs the person they (1) most wanted to have sex with, (2) least wanted to have sex with, (3) judged was most likely to have an STI, and (4) judged was least likely to have an STI. A single photograph could be assigned to multiple partner conditions, but not for “most” and “least” categories within one dimension. Participants who selected fewer than two photographs (*n* = 2 potential Control group participants) were disqualified from further participation. Female participants who selected only photographs of females would have been disqualified from further participation because the risk of HIV infection from female-female sex is extremely low, although this criterion resulted in no exclusions for the current study. Participants who were disqualified from further participation were compensated \$30. Qualified participants were then trained on using a visual analog scale. Next, they completed the four discounting tasks described below, in addition to other decision-making tasks not relevant to the present analyses. Monetary tasks were administered before sexual tasks, and delay tasks were administered before probability tasks. ### Sexual Delay Discounting Task Delay discounting of condom-protected sex was assessed using a computerized version of the Sexual Delay Discounting Task. At the beginning of each of the four partner conditions (presented in a pseudo-randomized order), the participant was shown the relevant photograph and instructed to imagine the person was interested in having sex now, that there was no chance of pregnancy, and that a condom was readily and immediately available. The participant indicated his/her likelihood of using a condom by clicking on a visual analog scale that ranged from “I will definitely have sex with this person without a condom” (0%) to “I will definitely have sex with this person with a condom” (100%). In subsequent trials, the participant was asked to rate his/her likelihood of waiting a given delay (ascending order; 1 hour, 3 hours, 6 hours, 1 day, 1 week, 1 month, and 3 months) to have sex with a condom. The visual analog scales for these trials ranged from “I will definitely have sex with this person now without a condom” (0%) to “I will definitely wait \[delay\] to have sex with this person with a condom” (100%). In the event that a photograph was assigned to multiple partner conditions, the participant completed the 8-trial series only once for that photograph. ### Monetary Delay Discounting Task Delay discounting of hypothetical money was assessed using a computerized task used previously. Participants made choices between smaller amounts of money delivered immediately vs. a larger amount (\$100) delivered after a delay. Unlike the Sexual Delay Discounting Task, which compared groups with respect to decisions involving each of four hypothetical partners, for the Monetary Delay Discounting only a single condition (i.e., reinforcer magnitude) was evaluated for group differences. Participants were instructed to treat choices as if the outcomes were real and that they should take their financial circumstances into account when making their choices. Based on the pattern of a participant’s choices, the task algorithm calculated an indifference point (i.e., a smaller amount of money subjectively equivalent to delayed \$100) at each of 7 delays: 1 day, 1 week, 1 month, 6 months, 1 year, 5 years, and 25 years (see and for a description of the algorithm used to determine indifference points). Although some delays were common to both discounting tasks (i.e., 1 day, 1 week, and 1 month), the standard range of delays typically assessed in the Monetary Delay Discounting Task exceeded that of the Sexual Delay Discounting Task. The order in which delays were assessed (ascending or descending) was randomly determined. ### Sexual Probability Discounting Task In this task, participants were asked to imagine in each decision that having sex with a photographed individual was associated with a specified risk of contracting an STI. We administered the task only for the “most want to have sex with” and “least want to have sex with” partner conditions to avoid explicitly confounding our experimental manipulation of risk with the perceived risk of partners. Presentation order was identical to that of the Sexual Delay Discounting Task. At the beginning of each partner condition, a research assistant placed a printed copy of the relevant photograph (21.59 cm x 27.94 cm) on the desk in front of the participant and reminded her or him to imagine the person was interested in having sex now, and that there was no chance of pregnancy. For the first trial, the research assistant read text shown to the participant specifying that if he/she did not use a condom, then there was a 1 in 1 (100%) chance of contracting an STI from the photographed individual. A visual analog scale located below the prompt ranged from “I will definitely have sex with this person without a condom” to “I will definitely have sex with this person with a condom” for this and all other probability trials (see for use of a visual analog scale to assess probability discounting). No trials involved delays. In all trials, risk was described both as odds in favor and percent chance of contracting an STI. Other risk values assessed (in descending order) were 1 in 3 (33%), 1 in 13 (8%), 1 in 100 (1%), 1 in 400 (0.25%), 1 in 700 (0.14%), 1 in 2,000 (0.05%), and 1 in 10,000 (0.01%). ### Monetary Probability Discounting Task Probability discounting of hypothetical money (\$100) was assessed using a computerized task used in a previous study. Choices were between smaller amounts of money delivered immediately with 100% certainty vs. a larger amount (\$100) delivered immediately, but with a specified probability of delivery. Indifference points were obtained at each of 7 probabilities of receiving \$100: 99%, 90%, 75%, 50%, 25%, 10%, and 1%. Hypothetical rewards instructions and the choice-adjustment algorithm were identical to the Monetary Delay Discounting Task. The order in which probability values were assessed (ascending or descending) was matched to the Monetary Delay Discounting Task for each participant. ### HIV testing Upon completion of all experimental tasks, a 3.5 cc blood specimen was collected and sent to a commercial laboratory (Quest Diagnostics Incorporated, Baltimore, MD) for HIV-1/HIV-2 antibody testing. Immediately after the blood draw, participants were discharged and compensated \$75 for study completion. ### Evaluation of group characteristics Group characteristics were compared using independent-samples *t*-tests for continuous variables and Fisher’s exact tests for categorical variables. Groups were matched (i.e., no significant or trend-level differences, *t*s ≤ 1.62, *p*s ≥. 11) on the following characteristics: age, sex, race, ethnicity, marital status, educational attainment, monthly income, Quick Test score, and cigarettes per day. ### Orderliness of discounting data Likelihood values from the sexual tasks were calculated as a proportion of the total length of the visual analog scale and indifference points from the monetary tasks were expressed as a proportion of \$100. Nonsystematic discounting data were identified according to criteria adapted from a previously established algorithm. For the sexual tasks, two criteria were used to identify nonsystematic data. First, starting with the second delay value (1 hour) or probability value (33%), the likelihood of condom use at a given delay or probability could not exceed the immediately preceding likelihood by more than 0.2. Second, the likelihood of condom use at the longest delay (3 months) or smallest probability (0.01%) could not exceed the likelihood when a condom was immediately available or the risk of contracting an STI was 100% by more than 0.1. For the monetary tasks, the analogous two criteria were applied as well as an additional criterion, which specified that the final indifference point could not be greater than 0.9. Participants whose data violated one or more of the aforementioned criteria in a condition were excluded in analyses of that condition. ### Discounting data analysis Groups were compared using extra sums-of-squares *F* tests analyzing all individual participant data as previously described (GraphPad Prism version 6.05 for Windows, GraphPad Software, La Jolla, CA). For delay tasks, these analyses regressed proportion likelihood of condom use (sexual tasks) or indifference points (monetary tasks) against delay (hours). For probability tasks, these measures were regressed against odds against, which were calculated as (1/*p*)-1, wherein *p* is the probability in favor of an event’s occurrence. Nonlinear regressions were performed using a two-parameter hyperbolic discounting equation: Proportion likelihood or indifference point = 1/(1+*r*X)^*s*, wherein X is the task-specific independent variable (i.e., either hours or odds against), *r* is a parameter proportional to discounting rate, and *s* is a parameter describing the nonlinear scaling of the dependent variable (monetary amount or likelihood of condom use) and delay or odds against. The free parameters *r* and *s* were unconstrained in regression analyses. The *F* tests compared the difference in nonlinear regression model error when free parameters were shared (i.e., one curve best-fit to all discounting data collapsed across groups) vs. when free parameters were unshared (i.e., a separate best-fit for each group). A significant *p* value indicates significantly less error when model parameters are unshared, indicating that the groups differ. In addition to analyzing raw group likelihood data from the Sexual Delay Discounting and Sexual Probability Discounting tasks, we also compared 0-delay (and 100%-probability) likelihood values, which were non-normally distributed, between Cocaine and Control groups using Mann-Whitney *U* tests. For the Sexual Delay Discounting Task we also isolated the effect of delay on likelihood of waiting to use delayed condoms (i.e., discounting) from differences in preference for using immediately available condoms (0-delay). Specifically, individual participant raw likelihood values for each condition were standardized by dividing each non-zero delay trial likelihood value by its respective 0-delay trial value. In the event that a standardized likelihood value exceeded 1 (i.e., a non-zero delay trial likelihood exceeded its respective 0-delay trial value), the value was replaced with a value of 1. Participants who reported zero likelihood of using an immediately available condom were excluded from standardized data analyses because the relative effect of delay on reward value is undefined if the initial value is zero. An identical procedure was employed to standardize the Sexual Probability Discounting Task data with respect to 100%-probability trial data. Extra sums-of-squares *F* tests were used to compare standardized discounting data between Cocaine and Control groups in each condition. Correlations among all discounting measures (using standardized likelihood in the sexual tasks) were conducted using an area-under-the-curve (AUC;) metric. Spearman’s rank-order correlations were calculated because AUC values were non- normally distributed. The criterion for significance in all tests was *p* \<. 05. # Results ## Sample characteristics presents characteristics of the Cocaine (*n* = 23) and Control (*n* = 24) groups. There were no significant differences between the groups with the exception of substance use, self-reported sexual risk behavior, and HIV variables. All 23 participants in the Cocaine group met criteria for cocaine abuse, and 20 participants also met criteria for cocaine dependence. Eighteen participants (78%) reported inhalation (smoking) as their preferred route of cocaine administration, and 4 (17%) and 1 (4%) participants preferred intranasal (snorting) and intravenous routes, respectively. Compared to the Control group, the Cocaine group reported significantly greater alcohol and cannabis use (number of past-year users; days used per month), and included significantly more participants meeting diagnostic abuse criteria for these two drugs. There were no significant group differences in measures of opioid and cigarette use. Participants in the Cocaine group also had significantly higher HRBS sexual risk subscale scores, and had a greater proportion of HIV-positive participants, than the Control group. Five of six Cocaine group participants with HIV knew they were HIV-positive prior to participating in the study and completing discounting assessments. Excluding these five HIV-positive participants did not alter whether 13 of 14 between-groups comparisons of discounting data met significance or not (the exception being a significant between-group difference in the standardized analysis of the “least likely to have an STI” partner condition, one that was not significant in the full sample). We therefore we report results from the full sample. In preparation for the sexual discounting tasks, 13 men in the Control group and 11 men in the Cocaine group selected exclusively female partners, while 7 women in the Control group and 10 women in the Cocaine group selected exclusively male partners. One man in the Control group and 2 men in the Cocaine group selected exclusively male partners, and 1 man in the Control group selected both male and female partners. Two women in the Control group selected both male and female partners. Sexual discounting data from partner conditions in which female participants selected a female partner (*n* = 4; 1 partner condition for 1 participant and 3 partner conditions for 1 participant) were excluded prior to analysis because the risk of female-female HIV transmission via sexual behavior is extremely low. ## Orderliness of data Across all four partner conditions of the Sexual Delay Discounting Task, 77% and 85% of discounting functions were systematic for Cocaine and Control groups, respectively. Similarly, 87% (Cocaine) and 88% (Control) of Monetary Delay Discounting Task functions were systematic. Across both partner conditions of the Sexual Probability Discounting Task, 98% (Cocaine) and 100% (Control) of discounting functions were systematic and 100% (Cocaine) and 96% (Control) of Monetary Probability Discounting Task functions were systematic. Excluding nonsystematic data and data that could not be standardized (because of a zero initial likelihood of condom use) resulted in *n* that differed somewhat from analysis to analysis, however, these exclusions did not alter whether any group difference reported in reached significance or a trend toward significance. ## Between-group comparisons of discounting data ### Sexual Delay Discounting\\ Reported likelihood of waiting to use a condom decreased as a function of delay to condom availability in all four Sexual Delay Discounting Task partner conditions for both groups. Significant group differences (i.e., lower likelihood of waiting to use a condom among Cocaine participants than among Control participants) were observed in two partner conditions: the “most want to have sex with” partner condition \[Cocaine *n* = 19, Control *n* = 19, *F*(2, 300) = 5.81, *p* \<. 01\], and the “least want to have sex with” partner condition \[Cocaine *n* = 13, Control *n* = 20, *F*(2, 260) = 6.38, *p* \<. 01\]. (left column) shows best-fit curves to mean standardized likelihood of condom use in Cocaine and Control groups within each partner condition of the Sexual Delay Discounting Task. The right column of displays these same data, except with delay to condom availability expressed ordinally to facilitate visual inspection at short delays. When data were standardized to isolate the effect of delay, participants in the Cocaine group discounted significantly more steeply than those in the Control group in two of four partner conditions: the “most want to have sex with” partner condition \[Cocaine *n* = 14, Control *n* = 15, *F*(2, 228) = 5.51, *p* \<. 01\], and the “least want to have sex with” partner condition \[Cocaine *n* = 13, Control *n* = 19, *F*(2, 252) = 8.56, *p* \<. 001\]. Discounting did not differ significantly between Cocaine and Control groups with the “most likely to have an STI” partner condition \[Cocaine *n* = 18, Control *n* = 19, *F*(2, 292) = 0.55, *p* =. 58\], or the “least likely to have an STI” partner condition \[Cocaine *n* = 15, Control *n* = 16, *F*(2, 244) = 2.02, *p* =. 13\]. No significant group differences in likelihood of using an immediately available condom were observed (*U*s ≥ 116.5, *p*s ≥. 25). ### Monetary Delay Discounting The top row of shows best-fit curves to delay discounting data for hypothetical money (left graph; right graph shows data with delays expressed ordinally). Cocaine participants (*n* = 20) discounted delayed \$100 significantly more steeply than Control participants \[*n* = 21, *F*(2, 283) = 18.29, *p* \<. 0001\]. ### Sexual Probability Discounting Reported likelihood of using an immediately available condom decreased with increased odds against contracting an STI for both partner conditions in both groups. However, no significant differences in discounting between Cocaine and Control groups were detected in unstandardized likelihood of condom use in the Sexual Probability Discounting Task for either the “most want to have sex with” partner condition \[Cocaine *n* = 23, Control *n* = 24, *F*(2, 372) = 2.30, *p* =. 10\] or for the “least want to have sex with” partner condition \[Cocaine *n* = 22, Control *n* = 23, *F*(2, 356) = 1.88, *p* =. 16\]. Fig (left column) shows best-fit curves to mean standardized likelihood of condom use in Cocaine and Control groups within each partner condition of the Sexual Probability Discounting Task. The right column of displays these same data, except with odds against contracting an STI expressed ordinally. Analyses of standardized data suggested a trend for those in the Cocaine group to discount condom-protected sex less steeply as STI risk decreased for the “most want to have sex with” partner \[Cocaine *n* = 23, Control *n* = 24, *F*(2, 372) = 2.95, *p* =. 054\], with no discounting difference for the “least want to have sex with” partner \[Cocaine *n* = 22, Control *n* = 23, *F*(2, 356) = 1.72, *p* =. 18\]. The groups also did not differ statistically with respect to likelihood of condom use when the risk of contracting an STI was 100% (*U*s ≥ 226, *p*s ≥. 23). ### Monetary Probability Discounting The bottom row of shows best-fit curves to probability discounting data for hypothetical money (left graph; right graph shows data with odds against expressed ordinally). Cocaine (*n* = 23) and Control (*n* = 23) group participants showed a significantly different pattern of discounting for a probabilistic \$100 reward \[*F*(2, 318) = 5.24, *p* \<. 01\]. However, the bottom right panel of shows that one group did not consistently discount to a greater or lesser extent than the other group. At lower odds against receiving \$100 (i.e., 0.01, 0.11, 0.33, 1), mean indifference points for the Control group tended to be higher than mean indifference points for the Cocaine group, whereas at higher odds against receiving \$100 (i.e., 9, 99), the Cocaine group tended to be higher than the Control group. ## Correlations Among Discounting Tasks shows Spearman rank correlations for sexual (using standardized likelihood) and monetary discounting tasks. Upon elimination of nonsystematic discounting data sets and sexual data sets showing zero likelihood of condom use under 0-delay or 100% probability of STI conditions, *n* ranged from 22 to 46 across these correlations. Within the Sexual Delay Discounting Task, discounting measures among partner conditions were positively and significantly correlated in 4 of 6 instances. Similarly, within the Sexual Probability Discounting Task, discounting between the two partner conditions was positively and significantly correlated. The Sexual Delay Discounting Task and the Sexual Probability Discounting Task were significantly and positively correlated in 7 of 8 partner conditions. Conversely, the Monetary Delay Discounting Task and Monetary Probability Discounting Task were not significantly correlated. The Monetary Delay Discounting Task was not significantly correlated with the Sexual Delay Discounting Task for any of the 4 partner conditions. Similarly, the Monetary Probability Discounting Task was not significantly correlated with either partner condition in the Sexual Probability Discounting Task. Finally, although the Monetary Probability Discounting Task was significantly and positively correlated with the Sexual Delay Discounting Task for 2 of the 4 partner conditions, no significant relation was found between the Monetary Delay Discounting Task and either partner condition in the Sexual Probability Discounting Task. # Discussion This study systematically examined discounting of delayed and probabilistic sexual and monetary outcomes among individuals with cocaine use disorders and demographically-matched controls. First, we found that individuals with cocaine use disorders discounted significantly more steeply than controls in two of the four Sexual Delay Discounting Task partner conditions, as well as in the Monetary Delay Discounting Task. Second, in the novel Sexual Probability Discounting Task, both groups showed an orderly effect in which odds against contracting an STI systematically decreased the likelihood of using an immediately available condom. Third, no robust group differences in probability discounting of sexual outcomes or monetary rewards were found. Finally, correlations showed sexual and monetary results were unrelated, for both delay and probability discounting tasks. Each finding will be discussed in turn. To date, no other study has examined whether individuals with cocaine use disorders discount delayed condom-protected sex more than matched controls. After controlling for each individual’s likelihood of condom use when no delay was involved, individuals with cocaine use disorders discounted delayed condom- protected sex significantly steeper than controls in two of four partner conditions. We suspect that the relatively high rates of reported condom use in the “most likely to have an STI” partner condition were responsible for the inability to detect group differences. Although likelihood of condom use was relatively lower in the “least likely to have an STI” partner condition, a significant group difference was not obtained, perhaps due in part to similar ratings of condom use likelihood between groups at delays shorter than 1 day. There were no significant differences between individuals with cocaine use disorders and controls in likelihood of using immediately available condoms in any of the four Sexual Delay Discounting Task partner conditions, demonstrating that the between-group differences in likelihood of using delayed condoms were truly driven by differential responses to delay (i.e., delay discounting). Had this study examined only preferences about using immediately available condoms, an entire dimension of increased HIV risk behavior would have been overlooked. This highlights the importance of delay discounting as a contributor to the high rates of HIV risk behavior. For the Monetary Delay Discounting Task, the finding of steeper discounting in individuals with cocaine use disorders replicates several previous findings, contributing to overall confidence in our study findings. In the novel Sexual Probability Discounting Task, decreased odds of contracting an STI systematically decreased the likelihood of using an immediately available condom in both groups, suggesting that perceived STI risk has a lawful effect on condom use regardless of cocaine use history. Probability discounting of sexual outcomes has been shown in college students using tasks assessing choices between certain shorter durations vs. uncertain longer durations of sexual activity or choices between certain “less than ideal” and uncertain “ideal” sexual outcomes, with idealness represented visually by line length. These tasks assessed a theoretically important issue, the effects of uncertainty of a sexual act on its value, while our task assessed the clinically-relevant effect of STI uncertainty on condom use. Orderly probability effects in all three tasks speak to the robust effect of probability on sexual outcomes, regardless of whether a probabilistic reward or a probabilistic punishment (hypothetical STI contraction in the current study) is being assessed. In contrast to the delay discounting results, no robust between-group differences in probability discounting of either sexual or monetary outcomes were observed. The only between-group difference detected was in the shape of the monetary probability discounting function; there was no reliable between- group difference across the range of different probabilities assessed. A prior study showed no association between drug use and probability discounting of sexual outcomes (although low drug use in the sample was a limitation). Our comparisons between individuals with cocaine use disorders and matched controls regarding STI risk and condom use further suggest no robust relations between drug use and probability discounting of sexual outcomes. This conclusion is consistent with the larger literature on probability discounting, which shows mixed results regarding relations between drug use and probability discounting of monetary gains, and no apparent relation between drug use and probability discounting of monetary losses. Our findings therefore support the somewhat paradoxical conclusion that behavioral processes underlying risk-taking may be unrelated to drug use, and specifically cocaine use in the present study. Likewise, loss aversion (i.e., the tendency to overweight losses relative to equivalent gains), or negativity bias may be similarly unrelated to cocaine use disorders, given our observation that individuals with cocaine use disorders did not differ from matched controls in terms of their sensitivity to STI contraction risk (i.e., a loss), but did differ significantly in their sensitivity to delayed condom-protected sex (i.e., a gain). Collectively, our differential findings between delay and probability discounting highlight the importance of examining multiple behavioral processes in relation to clinically relevant behavior, and suggests that, with respect to sexual outcomes, steeper delay discounting of condom-protected sex, but not reduced sensitivity to STI probability, contributes to increased sexual HIV risk among individuals with cocaine use disorders. Correlations showed sexual and monetary results were unrelated, for both delay and probability discounting. The sexual partner conditions were generally positively related within a single type of sexual task (delay or probability). Moreover, delay and probability discounting in the sexual tasks were generally positively related. These findings were in contrast to the lack of significant association between delay and probability discounting of monetary rewards, and in contrast to the lack of significant associations between the monetary and sexual results for either delay or probability discounting. The nonsignificant correlation (in the positive direction) between delay and probability discounting of money should be viewed in the context of mixed evidence on this relationship, with studies typically showing correlations in the positive direction but varying from weak to strong in correlation strength, and varying between significance and nonsignificance \[–, –\]. The predominantly nonsignificant correlations between discounting of delayed condom-protected sex and delayed money replicates previous findings. One potential explanation for the relation between the delay and probability sexual tasks is that choice behavior involving waiting for a delayed condom and avoidance of STI contraction recruit related processes. This seems especially plausible given the risk of STI contraction implicit in all sexual situations, including the “most/least want to have sex with” partner conditions in the Sexual Delay Discounting Task. With respect to probability discounting, it should be noted that, unlike the delay discounting tasks, the probability tasks assess events of opposite valences, with the risk of contracting an STI or receiving \$100 representing a probabilistic loss (punishment) and gain (reward), respectively. As a result of this difference, a general tendency toward risk-taking would be evidenced by steep discounting of STI risk and shallow discounting of the monetary reward (and vice versa for risk-aversion), resulting in a negative correlation between these measures. The fact that we did not observe significant negative correlations between these measures lends some support to the conclusion that sexual outcomes are discounted uniquely relative to the discounting of monetary outcomes. These results add to growing evidence showing domain specificity in discounting results. In other words, discounting results differ depending upon the type of outcome studied. Most human discounting studies have examined only money as the outcome, with the implicit but questionable assumption that decision making for monetary outcomes is indicative of decision making in clinically relevant domains such as substance use disorders. Replicating previous findings with respect to opioid-dependent women, the present results suggest that delay discounting of condom-protected sex and delay discounting of money are different processes, although individuals with cocaine use disorders discount both more steeply than matched controls. Similarly, results suggest that probability discounting of STI contraction and probability discounting of money are different processes, but unlike with delay discounting, these different probability discounting processes do not differ between individuals with cocaine use disorders and matched controls who did not use cocaine. One limitation is that despite matching participants demographically with respect to age, sex, race, ethnicity, marital status, education, monthly income, intelligence test score, and cigarettes per day, participants in the Cocaine group and the Control group differed on substance use variables other than cocaine use. Specifically, participants in the Cocaine group showed significantly higher rates of alcohol and cannabis use relative to participants in the Control group. Moreover, participants in the Cocaine group showed significantly higher self-reported sexual risk behavior and were significantly more likely to be HIV positive. We believe these differences are to be expected and indicate that we studied a representative sample of individuals with cocaine use disorders \[2–6; 18–21\]. Indeed, the higher rates of sexual risk behavior and HIV infection in the Cocaine group highlight the behavioral problems associated with cocaine use disorders, which prompted the present study. Further, there is some evidence to suggest that cocaine use may be a stronger predictor of discounting than alcohol and cannabis use. For example, individuals with problematic cocaine use discount more steeply than those with problematic alcohol use, and the relationship between cannabis use and discounting is less robust than previous studies with other drugs. Finally, differences in other drug use between the Cocaine group and Control group are comparable to differences that may have been present in previous studies examining discounting in individuals who use cocaine relative to a matched control group. An additional limitation is that the various discounting tasks involved hypothetical rather than real outcomes. However, delay and probability discounting studies have generally shown similar results when using real and hypothetical money (\[,, –\] but see). Moreover, the Sexual Delay Discounting Task has demonstrated reliability and relationships with self-reported sexual risk. Another potential limitation is that discounting in the sexual tasks may have been affected not only by delay or probability, but also by the effort associated with condom use, whereas this was not the case in the monetary tasks. Although a potential effort-related confound could have contributed to the lack of significant correlations between the task types, attempts to control for this variable could jeopardize the external validity of the sexual tasks. Moreover, it is unlikely that effort influenced group differences in the Sexual Delay Discounting Task, because the Sexual Probability Discounting Task also involves the same potential effort-related confound, yet did not show robust group differences. Another shortcoming was our relatively small sample size. However, the orderliness of the data and the detection of between-groups differences suggest the sample was sufficient to detect meaningful results. Finally, although the present findings suggest an association with cocaine use disorders, the etiology of increased sexual HIV risk still remains unclear. To examine the potential contribution of cocaine pharmacology on sexual risk behavior, future research should examine the effects of acute cocaine administration on the Sexual Delay Discounting Task and the Sexual Probability Discounting Task in individuals who use cocaine. The translational nature of this research contributes a novel perspective to the prevention of HIV sexual risk behavior among individuals with and without cocaine use disorders. Both delay and probability discounting may serve a diagnostic role, identifying individuals who are at risk for HIV or STI contraction (or transmission to others, as highlighted by the non-trivial percentage of our sample that was HIV-positive). Among individuals who exhibit high rates of sexual risk behavior and steeply discount delayed condom-protected sex, behavioral treatment strategies aimed at reinforcing condom carrying or training delay tolerance may reduce the likelihood of HIV and other STI transmission. Among individuals who exhibit high rates of sexual risk behavior and steeply discount the possibility of uncertain STI contraction, training that increases the perceived likelihood that partners may have an STI may also decrease HIV and other STI transmission. The present translational research on basic behavioral processes using clinically relevant decisions may therefore be leveraged to improve public health. The authors would like to thank Natalie R. Bruner, Ph.D., Crystal Fridy, and Grant Glatfelter for assistance in conducting this study, and Leticia Nanda, CRNP for collecting blood samples and providing HIV counseling. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MWJ PSJ. Performed the experiments: MWJ PSJ ESH. Analyzed the data: MWJ PSJ. Contributed reagents/materials/analysis tools: MWJ. Wrote the paper: MWJ PSJ ESH MMS.

# Introduction Within the human body, there exists multitudes of microorganisms which colonize mucosal surfaces to form distinct microbial communities within body site niches. Together, these microorganisms outnumber human cells by a factor of 10 and are collectively termed the *microbiome*. Bacterial communities existing in anatomic niches such as the gastrointestinal tract, oral cavity and genitourinary system can be classified as mutualists (symbiotically effective microbes), commensals (microbes that are neither harmful nor helpful to the host) and pathogens (colonizing microbes which may cause potential harm to the host). Mechanisms by which microbiota exert their influences on human health are not well-defined, but under certain circumstances the relative abundance of certain bacterial communities can become altered, thereby disrupting normal homeostasis and resulting in human disease. Microbial disequilibrium can result in endothelial dysfunction, alteration of host immune systems, bacterial translocation, chronic inflammation, and global genomic instability, all of which are hallmarks of cancer. Indeed multiple studies have linked alteration of the normal micriobiome with development of human cancers. While disruption of the genitourinary microbiome may potentially promote gynecologic carcinogenesis, the exact role of the microbiome in gynecologic cancers remains unclear. The vagina is a complicated microbial niche which allows for survival and proliferation of a number of both beneficial bacteria as well as opportunistic pathogens. The “ascending infection theory” postulates that disruption in the vaginal microbiome can allow pathogenic bacteria to ascend into the upper genital tract and cause polymicrobial infections, resulting in inflammation of the uterus, fallopian tubes and ovaries, a condition known as pelvic inflammatory disease (PID). As chronic inflammation is an etiologic factor in most cancers, this establishes a plausible hypothesis for how microbial dysregulation may promote tumor formation. Furthermore, endometrial cancer is promoted by obesity, hormonal imbalances, diabetes and metabolic syndrome, all of which may promote changes in the microbiome. The objective of the present study was to characterize the endometrial, cervicovaginal and anorectal microbiota of postmenopausal women undergoing hysterectomy for endometrioid and serous uterine cancers relative to controls undergoing hysterectomy for benign conditions. # Materials and methods ## Subject recruitment This prospective study was approved by the Albert Einstein College of Medicine Institutional Review Board and the Protocol Review and Monitoring Committee. Eligible subjects scheduled for hysterectomy were recruited from a gynecologic oncology clinic from September 2016 thru April 2019. Patients meeting inclusion criteria were approached initially by a gynecologic oncologist; those expressing interest in participation then met with a clinical study coordinator and were given the option of enrollment in the study and those willing provided informed written consent. Women were considered eligible for inclusion if they were English or Spanish-speaking, post-menopausal (defined as 50 years of age or older who have not had menses for 12 consecutive months or more) with biopsy proven well-differentiated endometrioid endometrial adenocarcinoma (EAC), uterine serous carcinoma (USC) or with non-cancerous conditions requiring hysterectomy. Exclusion criteria included history of prior cancer, prior chemotherapy or radiation therapy, prior bariatric surgery, history of human immunodeficiency virus or PID, use of douching, hormone therapy, systemic or local antibiotics, pro-biotics or anti-fungal medications within 2 weeks of initial consultation, or confirmed urinary tract or vaginal infection such as bacterial vaginosis, sexually transmitted disease or candidiasis within 1 month of initial consultation. A total of 55 potentially eligible patients were evaluated during the study period. Of these, 11 women had unexpected final pathology results rendering them ineligible, 8 lacked specimen collection and 1 was determined after enrollment to be pre-menopausal. These women were excluded from the study. ## Sampling procedures After induction of anesthesia and prior to antibiotic administration and vaginal preparation, the patient was placed in dorsal lithotomy position. A sterile bi- valve speculum was inserted inside the vagina to expose the uterine cervix and vaginal canal. A sterile q-tip applicator was used to swab the vaginal fornices and ectocervix and was then placed into a 1 mL tube of Specimen Transport Medium (STM) (Digene Female Swab Specimen Collection Kit, Qiagen, CA). Anorectal swabs were obtained by inserting a nylon q-tip applicator approximately 1 cm beyond the anal verge and then placed in a tube containing STM. The patient was then surgically prepped and the hysterectomy proceeded as planned. After removal of the uterus, the specimen was placed on a sterile field and opened coronally using a sterile scalpel in order to expose the endometrial cavity. Swabs were obtained from the endometrial cavity and placed in STM as described above. The surgical specimen was then handed over to pathology for routine clinical diagnostic testing. All research samples were stored in a -20° freezer within one hour of collection. ## Clinicopathologic data After pathologic confirmation of hysterectomy specimens as containing either non-cancerous tissue, well-differentiated endometrioid carcinoma or uterine serous cancer, clinical and pathologic information was abstracted from the medical record and stored in a secure database. Collected variables included patient age, race, ethnicity, body mass index (BMI in kg/m²), parity, chart-indicated diagnosis of medical comorbidities (diabetes, hypertension, cardiovascular disease, smoking status), stage of cancer, tumor size, and presence of lymphovascular space invasion. Continuous variables were reported as means ± standard deviations. Categorical data were presented as number of patients with percentages. Bivariate analysis was performed to assess differences of covariates between groups (benign, endometrioid and serous). Continuous variables were assessed using one way analysis of variance (ANOVA) or Kruskal-Wallis tests as appropriate, whereas categorical and dichotomous variables were examined using x² or Fisher exact tests as appropriate. Data analysis was performed using Stata 14.2. ## DNA extraction and next-generation sequencing DNA from samples stored in STM was extracted using the QIAamp DNA Kit (Qiagen, CA) following their standard protocol, modified by pre-incubation with Proteinase K and agitation with glass beads. Sterility was maintained by processing these samples in a sterile Biosafety Cabinet in an isolated extraction room. An aliquot of DNA from each STM sample was PCR amplified using barcoded primers annealing to conserved sequences within the V4 region (\~250 base pairs) of the 16S ribosomal RNA (rRNA) gene. The 16SV4 rRNA region was chosen because of it’s robust ability to effectively distinguish between vaginal bacterial species when compared to alternate 16S regions. Barcoded and purified DNA libraries were sequenced using an Illumina MiSeq (Illumina Inc., San Diego, CA) using paired-end reads. ## Bioinformatics analysis Paired end Illumina reads were left trimmed to remove bases that had a PHREAD quality score of 25 or lower using prinseq-lite. Quality controlled reads were then demultiplexed using dual Golay barcodes using denovobarcode. Reads were then merged using PANDAseq and processed using the QIIME2 platform. Briefly, VSEARCH was used to cluster sequences into operational taxonomic units (OTUs) and to assign taxonomy using the Greengenes database. The *phyloseq* package was used to import data into R for final processing. Alpha and beta diversity was calculated using *phyloseq* and *ggplot2* was used to make final figures. ANCOM was used to identify bacterial markers for endometrial cancer with adjustment for multiple testing (FDR\<0.05) as well as adjustment for age and race. Ward.D2 algorithm was used to perform hierarchical clustering using identified biomarkers. ANCOM was utilized in order to overcome the compositional limitations inherent to the study of the microbiome. For more details on this issue and how ANCOM addresses it see the cited reference. # Results A total of 35 patients had microbiome specimens collected and analyzed (14 with well-differentiated EAC, 11 with USC, and 10 controls). The median age of the final cohort was 63.6 ± 9.2 years with no significant differences between groups (p = 0.47). Most patients were obese with a median BMI of 33.6 (IQR 26.3, 38.8) and there were no significant differences in BMI between groups (p = 0.10). There were no significant differences in race, ethnicity, medical comorbidities (diabetes, hypertension, hyperlipidemia) or smoking status across histologic groups. Of the 25 women with cancer, most (88.0%) had early stage (I or II) cancer with no significant differences between groups in terms of extent of disease (p = 0.57), tumor size (p = 0.68), or presence of lymphovascular space invasion (p = 0.13). Of the control women, 6 (60%) had uterine fibroids, 2 (20%) had ovarian cystadenomata, 1 (10%) had adenomyosis and 1 (10%) had pelvic organ prolapse. The top 20 bacterial genera within the cervicovaginal, uterine and anorectal microbiota revealed significant community separation based on anatomical site. The cervicovaginal microbiome was found to be dominated by *Lactobacillus* with additional elevated abundances of *Prevotella* and *Gardnerella* relative to other genera. There was a clear separation of the uterine microbiome from the other two niches and this body site was the most even in terms of genus diversity with a dominance of *Flavobacterium*. The anorectal microbiome was dominated by either *Prevotella* or *Bacteroides* (depending on the individual sample). PERMANOVA analysis using weighted unifrac distances revealed significant separation between the three anatomical niches (R2 = 0.25, p \< 0.001). Microbial alpha diversity analyses demonstrated that serous cancers were characterized by a significant reduction of diversity within taxa of the control uterine microbiome based on the Chao1 (p = 0.004) and Fisher (p = 0.007) measures with a reduction in Shannon diversity (p = 0.094). The anorectal and cervicovaginal microbiome alpha diversity was not correlated with case status with the exception of Chao1, which was reduced in endometrioid cancers (p = 0.026). ANCOM analysis was performed to identify biomarkers that can distinguish patients across the three groups following adjustment for patient age, race and BMI. shows composite hierarchical clustering performed using all biomarkers identified in the ANCOM analysis where W-stat\>10 and FDR\<0.05. Two distinct microbiome clusters were observed with cluster indicated as “Biomarker Cluster 1” having a significantly higher prevalence of serous cancer patients (7/10) as compared to “Biomarker Cluster 2” where the majority of the benign cases are found (n = 6/7) (p = 0.042). No significant differences were found when comparing the presence of benign cases vs. endometroid (p = 0.30) or endometroid vs. serous cases (p = 0.68). Biomarker cluster 1 had several bacterial genera elevated including uterine *Pseudomonas* while Biomarker cluster 2 was distinguished by a dominance of cervicovaginal *Lactobacillus* and *Clostridium* genera. In order to identify potentially relevant functional pathways that may be associated with the serous cancer associated microbiome clusters, PICRUSt analysis was performed at KEGG level 3. shows PCA within the cervicovaginal region. A total of 49 pathways were identified to be significant. In order to present the data in a concise manner, only pathways with a corrected FDR\<0.05 and a fold change of 3 or greater are shown in. Similarly shows the PCA performed for the uterine samples with significantly differentially abundant pathways after correction for multiple testing (FDR\<0.05) in. # Discussion Our study demonstrates significant differences in bacterial diversity within the uterine, cervicovaginal and anorectal microbiota of women with EAC, USC, and non-malignant controls suggesting a potential role of disruption of the normal microbiome in these histologic types of uterine cancer. Walther-António and colleagues performed a similar study examining microbial differences in women with benign conditions, endometrial hyperplasia and endometrial cancer and demonstrated that endometrial cancers were enriched for Firmicutes, Spriochaetes, Actinobacteria, and Proteobacteria. They also found an association between *Atopobium vaginae* and *Porphyromonas* along with a high vaginal pH and the presence of endometrial cancer. Although their study did contain 3 patients with serous cancer, their analysis did not differentiate between EAC and more aggressive uterine histologies. Our study demonstrates significant reduction in microbial diversity of USC specimens relative to EAC specimens or controls. Our study is also important in that it provides a characterization of a postmenopausal population. Until the 1980s, it was thought that a healthy uterus was sterile. The endometrial cavity is a body niche with low bacterial abundance with 100–10,000 fold fewer bacteria compared with the vagina. The bacterial paucity of the endometrium compared with difficulty culturing most uterine bacteria has made it difficult to characterize the uterine microbiome until the advent of 16S rRNA next generation sequencing technologies. Nevertheless, the majority of extant literature regarding the endometrial microbiome focuses on pre-menopausal women and the association between microbiome dysregulation and adverse pregnancy outcomes. Our study shows that the uterine microbiome of postmenopausal women was comparable across histologies in terms of genus diversity with a dominance of *Flavobacterium*. Franasiak et al. investigated the uterine microbiome at the time of IVF and embryo transfer and also found *Flavobacterium* to be an abundant taxa of the uterine microbiome. Walsh et al also examined the endometrial microbiome composition in patients with and without endometrial cancer and identified significantly increased alpha diversity amongst post-menopausal patients. In their study, the post-menopausal endometrium demonstrated enrichment of *Anaerococcus*, *Peptoniphilus* and *Porphyromonas*, but only 7 uterine specimens were examined of post-menopausal women. All 35 women in our study were post-menopausal and each of them had sufficient endometrial samples available for analysis. In addition to a reduction of uterine bacterial diversity, we were also able to demonstrate a significant correlation between lower vaginal *Lactobacillus* and elevated uterine *Pseudomonas* associated with USC case status. Vaginal *Lactobacillus* has been previously associated with several gynecological cancers such as cervical, ovarian and endometrial cancers as well as general health of the cervicovaginal tract. *Pseudomonas* has also been recently implicated in association with endometrial cancer by Winters as well as Walther- António and colleagues using 16S rRNA amplicon sequencing of uterine samples. *Lactobacillus* may be acting to limit carcinogenesis by reducing local inflammation through cytokine modulation. Other studies have demonstrated a positive association between vaginal *Lactobacillus* and genitourinary health. Furthermore we have integrated biomarkers from across several sampled regions to show that we can effectively separate Serous cancers from controls using unsupervised clustering. This may suggest that the use of disease related biomarkers from across multiple anatomical regions may provide more clinically relevant groupings than the more broad CST categories. Additionally, we identified several pathways previously associated with several cancer subtypes. Particularly interesting may be our identification of the elevation of the chaperone and folding pathway within the uterine microbiome within the USC associated microbial biomarker clade. Overexpression of the HSF1 protein in particular has been linked with significantly lower survival in endometrial cancer patients and given our finding, this may have a correlation to the composition of the urogenital microbiome. Strengths of our study include the fact that we have characterized microbial composition of anogenital body sites in a racially and ethnically diverse population of postmenopausal women with uterine cancer and contrasted these findings with a control group of postmenopausal women without cancer. Although our cohort is small, it is among the largest published reports characterizing the uterine microbiome in postmenopausal women. Our histologic groups were well balanced in terms of their clinical and pathologic covariates. We also excluded enrollment of premenopausal women and those with recent infections or use of probiotics and antibiotics in order to reduce confounding of our results. It can therefore be assumed that the differences seen between groups can be attributed to histologic differences rather than lifestyle factors that would be difficult to control for in our analyses. It should also be noted that the functional pathways analysis, although shown to be highly correlated with true functional composition of sampled communities, is an estimate based on the 16S rRNA sequencing and should be verified using shotgun metagenomic approaches. Nevertheless there are limitations of our study. Although we made all attempts to collect specimens in a sterile fashion, sterility can never be completely ensured. It is difficult to obtain samples from the endometrial cavity without passing through the cervical os or bivalving the uterus from the side as we have done in this study. Furthermore, the use of a uterine manipulator during hysterectomy may influence the results. Additional prospective studies are required to longitudinally sample the microbiota of postmenopausal women and determine if progressive disruption of the microbiome contributes to endometrial carcinogenesis. Moreover prospective clinical trials will help determine if interventions such as probiotic administration may mitigate risk of gynecologic malignancy. # Conclusion The microbial diversity of anatomical ecological niches in postmenopausal women with EAC and USC is different compared to benign controls. This difference is both in terms of community structure as defined by a reduction of microbial alpha diversity within the uterus and by differences of bacterial taxa within the cervicovaginal and uterine regions. Microbial composition and presence of specific functional pathways may be necessary for development of endometrial cancer and further investigation using prospective datasets is warranted. # Supporting information 10.1371/journal.pone.0259188.r001 Decision Letter 0 Ishaq Suzanne L. Academic Editor 2021 Suzanne L. Ishaq This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 22 Jun 2021 PONE-D-21-05233 Characterization of the endometrial, cervicovaginal and anorectal microbiota in post-menopausal women with endometrioid and serous endometrial cancers PLOS ONE Dear Dr. Gressel, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Thank you for your patience during this unusually long review period, due to the pandemic, it has been difficult to find enough available reviewers and acquire timely reviews.  As such, I was only able to secure one set of reviews for your manuscript.  However, I agree with the reviewer's assessment of the submission, and with their recommendations to improve the manuscript. Please submit your revised manuscript by Aug 06 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at <plosone@plos.org>. When you're ready to submit your revision, log on to <https://www.editorialmanager.com/pone/> and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript: A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: <http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory- protocols>. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at <https://plos.org/protocols?utm_medium=editorial- email&utm_source=authorletters&utm_campaign=protocols>. We look forward to receiving your revised manuscript. Kind regards, Suzanne L. Ishaq, PhD Academic Editor PLOS ONE Journal Requirements: Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice. When submitting your revision, we need you to address these additional requirements. 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at <u><https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_m ain_body.pdf></u> and [https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_titl e_authors_affiliations.pdf](about:blank) \[Note: HTML markup is below. Please do not edit.\] Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 3\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No \*\*\*\*\*\*\*\*\*\* 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 5\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: This study aimed to discover if the microbiome at three body sites is different among post-menopausal women undergoing hysterectomy. They swabbed these three sites, which then underwent 16S rRNA sequencing on the V4 region. QIIME2 was used for taxonomic identification at the genus level, and PICRUST was used for functional annotation. Overall, this study found differences in alpha diversity at these three body sites across three clinical groups (endometrial cancer, uterine serous cancer, and other benign conditions). They also discovered potential metabolic pathways associated with these different clinical groups that shotgun metagenomics can confirm. Authors do a good job of stating their findings (without overstating) and tying them to biological mechanisms and clinical importance. This study is unique in that it focuses on post-menopausal women, whereas much of the literature on endometrial microbiome focuses on those of reproductive age. Page 9 line 58: use of the word dysbiosis: dysbiosis is a vague word that means a lot of different things depending on who you ask. It seems as though the phrase “the microbiome” could easily replace dysbiosis with the same meaning. I suggest either replacing the phrase “dysbiosis” or defining its use in the statement of translational relevance. Page 10 line 78, introduction: manuscript citations 5 through 9 are studies linking “dysbiosis” to human cancer, however, upon further inspection, these citations appear to link specific taxa to the development of cancer. I don’t think these citations fit that manuscript’s definition of “dysbiosis” so I would change the word “dysbiosis” in this line. Page 10 line 79 states “some authors postulate”; it is my understanding that we “should” be stating findings as “the research suggests/has evidence for”, etc, so as to not “call out” other authors in a manuscript. Citation 11 on page 10 line 84: “vaginal dysbiosis has been link\[ed\] to bacterial vaginosis” This citation discusses PID, not BV. I’m assuming that this is the wrong citation here? I would also be careful about what citation you use, because the diverse community state type is associated with bacterial vaginosis, but is NOT itself a dysbiotic state. Authors appear to use “dysbiosis”, “microbial disequilibrium”, and “microbial dysregulation” interchangeably. If all three terms are used to refer to the same phenomena, this needs to be explicitly stated. I would lean toward removing “dysbiosis” from the manuscript in favor of the other two terms due to the baggage that comes with the term “dysbiosis” Page 12 line 151: it would be great to include rationale for why the V4 region was selected for 16S rRNA sequencing Citation 27: page 12 line 163: I read the ANCOM pub just now, and this seems like a very interesting and statistically sound approach. It’s not clear to me from the manuscript if this is a program (like QIIME), a package (like phyloseq), or just a method that one implements themselves. Page 13 line 193: a p value greater than 0.05 is not marginally significant; 0.45 would be marginally significant. I would change this language prior to publication. Page 14 lines 201-203: It’s not clear to me what the significance of the “biological cluster \[1\|2\]” is, and this doesn’t appear to be further discussed in the discussion section. I would either remove this, include it in the discussion/state it’s relevance, or, if I missed this within the manuscript, state it more clearly. Page 15 lines 245-246: I don't remember reading in the methods/results that depleted vaginal Lactobacillus \>\> elevated uterine Pseudomonas in those with USC. I like the discussion of this (lots of lit, proposed biological mechanisms with clinical links), but don’t remember this in methods. Is this related to the “biological clustering”? If so, this isn’t currently clear in the manuscript. Should be more explicit in the manuscript prior to publication I like your discussion and justifications in lines 265-271 I appreciate that there is not a discussion of community state types - it would be hard to draw conclusions of CSTs with the V4 region alone and in general, the CST does not provide clinically relevant information. I am explicitly stating this since there is often pressure to include CSTs in analysis despite the reasons stated here. Some figures (fig 1, 4) may be hard to read for those with red-green colorblindness, it is worth adjusting the color palette before publication since this is a common condition. Manuscript does not state where the data (fastq/fasta files) are available. \*\*\*\*\*\*\*\*\*\* 6\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Emily F. Wissel \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, <https://pacev2.apexcovantage.com/>. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at <figures@plos.org>. Please note that Supporting Information files do not need this step. 10.1371/journal.pone.0259188.r002 Author response to Decision Letter 0 7 Sep 2021 Journal Requirements: 1\. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. \*\*\*We confirm that we have met the style requirements outlined in these documents. Comments to the Author: 1\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes 3\. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: No \*\*\*We have added the 16S V4 amplicon sequencing data to SRA as described below. 4\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*Thank you for reviewing our manuscript and for your insightful comments. We have done our best to address each of the comments in our revised manuscript. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: This study aimed to discover if the microbiome at three body sites is different among post-menopausal women undergoing hysterectomy. They swabbed these three sites, which then underwent 16S rRNA sequencing on the V4 region. QIIME2 was used for taxonomic identification at the genus level, and PICRUST was used for functional annotation. Overall, this study found differences in alpha diversity at these three body sites across three clinical groups (endometrial cancer, uterine serous cancer, and other benign conditions). They also discovered potential metabolic pathways associated with these different clinical groups that shotgun metagenomics can confirm. Authors do a good job of stating their findings (without overstating) and tying them to biological mechanisms and clinical importance. This study is unique in that it focuses on post-menopausal women, whereas much of the literature on endometrial microbiome focuses on those of reproductive age. Page 9 line 58: use of the word dysbiosis: dysbiosis is a vague word that means a lot of different things depending on who you ask. It seems as though the phrase “the microbiome” could easily replace dysbiosis with the same meaning. I suggest either replacing the phrase “dysbiosis” or defining its use in the statement of translational relevance. \*\*\*Thank you for drawing attention to the semantic differences in the literature surrounding this term “dysbiosis.” While there is certainly precedent for this term in the literature regarding the gynecologic microbiome, you are correct that the term is not often well-defined or consistent across studies. We initially had planned to define this term in the “statement of translational relevance,” in light of your other comments below, we have replaced this term with “disruption in the normal gynecologic microbiome” which is our intended meaning in using the word. Page 10 line 78, introduction: manuscript citations 5 through 9 are studies linking “dysbiosis” to human cancer, however, upon further inspection, these citations appear to link specific taxa to the development of cancer. I don’t think these citations fit that manuscript’s definition of “dysbiosis” so I would change the word “dysbiosis” in this line. \*\*\*Thank you. We have replaced all uses of the term “dysbiosis” in favor of “disruption in the normal gynecologic microbiome.” Page 10 line 79 states “some authors postulate”; it is my understanding that we “should” be stating findings as “the research suggests/has evidence for”, etc, so as to not “call out” other authors in a manuscript. \*\*\*Thank you for pointing this out. We have revised the manuscript to correct this verbiage. Citation 11 on page 10 line 84: “vaginal dysbiosis has been link\[ed\] to bacterial vaginosis” This citation discusses PID, not BV. I’m assuming that this is the wrong citation here? I would also be careful about what citation you use, because the diverse community state type is associated with bacterial vaginosis, but is NOT itself a dysbiotic state. \*\*\*We thank the reviewers for the opportunity to clarify this point. The cited reference: Wang Y, Zhang Y, Zhang Q, Chen H, Feng Y. Characterization of pelvic and cervical microbiotas from patients with pelvic inflammatory disease. J Med Microbiol 2018; refers to a study from Zhejiang university in which 38 patients with pelvic inflammatory disease and 19 control patients without pelvic inflammatory disease had 16s rRNA amplicon profiling to test the hypothesis that microbes in the vagina and cervix can spread to the upper genital tract and cause PID. You are absolutely correct that this reference does not refer to bacterial vaginosis and we have corrected the preceding statement as follows: “The vagina is a complicated microbial niche that allows for survival and proliferation of a number of both beneficial bacteria as well as opportunistic pathogens. The “ascending infection theory” postulates that disruption in the vaginal microbiome can allow pathogenic bacteria to ascend into the upper genital tract and cause polymicrobial infections, resulting in inflammation of the uterus, fallopian tubes and ovaries, a condition known as pelvic inflammatory disease.” Authors appear to use “dysbiosis”, “microbial disequilibrium”, and “microbial dysregulation” interchangeably. If all three terms are used to refer to the same phenomena, this needs to be explicitly stated. I would lean toward removing “dysbiosis” from the manuscript in favor of the other two terms due to the baggage that comes with the term “dysbiosis” \*\*\*Thank you. We have replaced all uses of the term “dysbiosis” in favor of “disruption in the normal gynecologic microbiome.” Page 12 line 151: it would be great to include rationale for why the V4 region was selected for 16S rRNA sequencing \*\*\*We thank the reviewer for the opportunity to clarify the use of 16S V4 rRNA sequencing in this project. The V4 region is the most commonly used region, which allows for more streamlined comparison’s across studies, and was recently demonstrate to show the best taxonomic resolution between vaginal flora (see PMID: 30535155, Willian Van Der Pol 2019). We have added justification along with the relevant citing text in the methods section: \*\*\*Lines 199-201: The 16SV4 rRNA gene region was chosen because of it’s robust ability to effectively distinguish between vaginal bacterial species when compared to alternate 16S regions. ANCOM is a method. In a simple explanation it just uses all combinations of microbial ratios to overcome compositional limitations of the microbiome when characterized using standard next generation sequencing. It does a few additional things like dealing with structural zeros within the data, but the testing of ratios is the core of this method. We have added an additional reference that delves into the use of microbial ratios as a means of dealing with compositionality that discusses ANCOM and several other similar approaches: (PMID 31222023, Morton 2019). \*\*\*Lines 214-216: ANCOM was utilized in order to overcome the compositional limitations inherent to the study of the microbiome. For more details on this issue and how ANCOM addresses it see: (PMID 31222023, Morton 2019). Page 13 line 193: a p value greater than 0.05 is not marginally significant; 0.45 would be marginally significant. I would change this language prior to publication. \*\*\*We completely agree. We have removed the term “marginally significant”. Page 14 lines 201-203: It’s not clear to me what the significance of the “biological cluster \[1\|2\]” is, and this doesn’t appear to be further discussed in the discussion section. I would either remove this, include it in the discussion/state it’s relevance, or, if I missed this within the manuscript, state it more clearly. \*\*\*In Figure 4 we utilized the biomarkers identified with ANCOM from across the three sampling sites to perform unsupervised clustering. This resulted in the two distinct sample clusters that appeared to consistently segregate Serous cases from controls. The observation of these clusters is relevant as the use of disease relevant markers across several body sites may provide more clinically relevant groups as opposed to CST clusters that are often reported in the literature. We have added a statement regarding this in the discussion: Lines 273-274: Furthermore we have integrated biomarkers from across several sampled regions to show that we can effectively separate Serous cancers from controls using unsupervised clustering. This may suggest that the use of disease related biomarkers from across multiple anatomical regions may provide more clinically relevant groupings than the more broad CST categories. Page 15 lines 245-246: I don't remember reading in the methods/results that depleted vaginal Lactobacillus \>\> elevated uterine Pseudomonas in those with USC. I like the discussion of this (lots of lit, proposed biological mechanisms with clinical links), but don’t remember this in methods. Is this related to the “biological clustering”? If so, this isn’t currently clear in the manuscript. Should be more explicit in the manuscript prior to publication \*\*\*We thank the reviewer for the chance to clarify our results as these are important points within the discussion. The Lactobacillus/Pseudomonas relationship is presented in figure 4 and mentioned in the figure caption “ “Biomarker Cluster 1” has a significant reduction of cervicovaginal Lactobacillus and Clostridium and a higher overall abundance of uterine Pseudomonas. ” We have added additional text in the results section in order to make this more clear: \*\*\*Lines 223-225: Biomarker cluster 1 had several bacterial genera elevated including uterine Pseudomonas while Biomarker cluster 2 was distinguished by a dominance of cervicovaginal Lactobacillus and Clostridium genera. I like your discussion and justifications in lines 265-271 \*\*\*Thank you very much. I appreciate that there is not a discussion of community state types - it would be hard to draw conclusions of CSTs with the V4 region alone and in general, the CST does not provide clinically relevant information. I am explicitly stating this since there is often pressure to include CSTs in analysis despite the reasons stated here. \*\*\*We thank the reviewer for this statement and agree that CSTs do not tend to offer clinically useful categories. This was in part what prompted us to use identified biomarkers to see whether we can generate clusters that are more directly relevant to endometrial cancer using the three sampling sites. We have added additional comments regarding this as stated in the response above. Some figures (fig 1, 4) may be hard to read for those with red-green colorblindness, it is worth adjusting the color palette before publication since this is a common condition. Manuscript does not state where the data (fastq/fasta files) are available. \*\*\*We have uploaded the sequence fastqs to SRA (PRJNA758386) and added a link to the project in the new “Data Availability” section. We have also modified the figures to allow them to be interpreted by people with color blindness using the <https://for-hue.herokuapp.com> web site. 10.1371/journal.pone.0259188.r003 Decision Letter 1 Ishaq Suzanne L. Academic Editor 2021 Suzanne L. Ishaq This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 15 Oct 2021 Characterization of the endometrial, cervicovaginal and anorectal microbiota in post-menopausal women with endometrioid and serous endometrial cancers PONE-D-21-05233R1 Dear Dr. Gressel, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at <http://www.editorialmanager.com/pone/>, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to- date. If you have any billing related questions, please contact our Author Billing department directly at <authorbilling@plos.org>. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact <onepress@plos.org>. Kind regards, Suzanne L. Ishaq, PhD Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions **Comments to the Author** 1\. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation. Reviewer \#1: All comments have been addressed \*\*\*\*\*\*\*\*\*\* 2\. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 4\. Have the authors made all data underlying the findings in their manuscript fully available? The [PLOS Data policy](http://www.plosone.org/static/policies.action#sharing) requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 5\. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here. Reviewer \#1: Yes \*\*\*\*\*\*\*\*\*\* 6\. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters) Reviewer \#1: Authors addressed all comments fully. They added clarifications to the manuscript about why they selected the V4 region for 16S sequencing, clarified that "dysbiotic" referred to a disruption of the gynecological microbiome, clarified some of the background citations, further describes the biomarker clusters and their clinical relevance, and have made all their data publicly available. This paper has novel findings about the microbiome at three body sites across three clinical conditions. These findings can pave they way for future clinical interventions and studies around these conditions through further evaluation of the identified biomarker clusters. \*\*\*\*\*\*\*\*\*\* 7\. PLOS authors have the option to publish the peer review history of their article ([what does this mean?](https://journals.plos.org/plosone/s/editorial- and-peer-review-process#loc-peer-review-history)). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. **Do you want your identity to be public for this peer review?** For information about this choice, including consent withdrawal, please see our [Privacy Policy](https://www.plos.org/privacy-policy). Reviewer \#1: **Yes: **Emily F. Wissel 10.1371/journal.pone.0259188.r004 Acceptance letter Ishaq Suzanne L. Academic Editor 2021 Suzanne L. Ishaq This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. 20 Oct 2021 PONE-D-21-05233R1 Characterization of the endometrial, cervicovaginal and anorectal microbiota in post-menopausal women with endometrioid and serous endometrial cancers Dear Dr. Gressel: I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department. If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact <onepress@plos.org>. If we can help with anything else, please email us at <plosone@plos.org>. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Dr. Suzanne L. Ishaq Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Salinity and water deficit are the principal environmental drivers of stress in mangrove tree species. Despite these limiting factors, mangrove forests are able to form successfully through the adoption of unique ecological strategies by the tree species that make up this system, with the tolerance of specific conditions being determined by the optimal range within an entire gradient of conditions. In mangrove systems, trees of the genus *Avicennia* L. are considered to be the most tolerant of salinity, although the potential of these trees for growth and the assimilation of carbon is reduced with increasing salinity. In comparison with the other Neotropical mangrove species, *Avicenna germinans* (L.) L. is one of the most resilient forms, capable of tolerating an extensive gradient of salinity. The species has achieved this through the development of a number of morphological and ecophysiological adaptations. The stress generated by conditions of extreme salinity affects the structure of the mangrove, reducing its stature, trunk diameter, and the size of the leaves, transforming the forest into a scrub mangrove, which is unlike the dwarf mangrove, where reduced stature is not accompanied by a reduction in leaf size, for example. The mangroves on the Brazilian Amazon coast occupy a number of distinct gradients of soil salinity and topography. On the Ajuruteua Peninsula, in the state of Pará, for example, the patches of scrub mangrove forest are dominated almost entirely by *A*. *germinans*, which occupies the sites with the highest salinity and topography, forming a gradient from shrub-like trees to short mangrove trees. These environmental drivers also have a direct influence on the production of biomass and carbon storage, with the trees distributing their nutritional resources as efficiently as possible in response to these conditions. Studies of biomass production and carbon storage have focused on different types of mangrove forest around the world, including fringe, basin, riverine, and scrub forest types, which have generated an ample range of estimates (\~ 8–460 Mg ha^-1), reflecting the diversity of environmental conditions. Most estimates of the production of biomass and carbon storage by mangrove ecosystems have focused on well-developed forests, while the effects of stressful conditions have been largely overlooked. As few studies have focused specifically on these stressed forests of short stature, it is important to develop allometric equations that provide reliable estimates of their biomass, not only because this vegetation is characterized by considerable morphological variation and is widely distributed in tropical and subtropical regions, but also because these equations will help to minimize the uncertainties intrinsic to the estimates of productivity available for the world’s mangrove ecosystems as a whole. In general, studies of the production of biomass by mangrove forests, including stressed forests, have focused on the above-ground biomass (AGB), while only a few studies have analyzed the below-ground biomass (BGB), and none have focused specifically on the BGB of scrub and/or dwarf mangrove forests. Below-ground biomass is considered to be one of the five primary carbon reserves in forested areas, and represents one of the largest carbon stocks in the tropical region. To fill this knowledge gap, we designed a study to assess the estimates of the above- and below-ground biomass and carbon storage in the different compartments and height classes of these forests dominated by *A*. *germinans* on the Ajuruteua Peninsula, on the Brazilian Amazon coast. The collection of data on the topography and salinity of these sites allowed us to identify the principal environmental drivers of the variation in the forest height classes. Similarly, the collection of data on root biomass directly through the excavation of specimens allowed us to develop allometric models to estimate the BGB of these forests and to assess the estimates of the total biomass (AGB+BGB) of scrub mangrove forests. Through this approach, we aimed to provide the means for the calculation of reliable biomass estimates that can be extrapolated to other, similar mangrove forests around the world. In addition to our estimates, derived from different height classes of scrub mangrove trees, we developed a general model comprising all height classes to assess the possibility of minimizing uncertainties associated with allometric models for biomass/carbon estimates. # Materials and methods ## Study site The study site is located on the Ajuruteua Peninsula (00°45’–01°07’ S, 46°50’–46°30’ W), in the northeastern extreme of the state of Pará, on the Brazilian Amazon coast, a region dominated by a hot and humid equatorial climate. The climatic data for the past 40 years reveal a mean annual temperature of 26.5ºC, mean annual precipitation of 2,348.5 mm, and relative humidity of 85%. The region has two well-defined climatic periods, with the timing of the rainy season being influenced primarily by the location of the Intertropical Convergence Zone, or ITCZ. The rainy season occurs when the ITCZ shifts southward between January and June, whereas the dry season (monthly precipitation of less than 100 mm) lasts from July through December. On this peninsula, a total area of approximately 16,465.5 ha (= 164.65 km²) is covered by mangrove forest, which is formed by three tree species: *Rhizophora mangle* L., *Laguncularia racemosa* (L.) C. F. Gaertn., and *A*. *germinans*. The region is characterized by semidiurnal macrotides, with a tidal range of 4–6 m. Most input of freshwater comes from the Caeté River which, during the rainy season, reduces salinity to zero in the local tidal creeks, such as the Taici Creek, which traverse the mangroves bordering the upland forests. This figure also shows the central portion of the peninsula, at kilometer 21 of the PA-458 state highway, where scrub mangrove forests dominate the landscape as a result of both high topography and the low flooding frequency, which is reflected in a hydrological deficit and increased salinity (\> 100). ## Soil salinity and topography The topographic gradient was surveyed using a differential Trimble R4 GNSS handheld GPS and a Topcon ES series total station. The total station was used to sample 45 points, with the angles and distances being measured using an electronic optical rangefinder and an electronic angle scanner. The geographic coordinates were recorded using the differential GPS in static post-processed mode, with a posteriori correction by triangulation using the geodesic stations of the Brazilian Continuous Monitoring Network, or RBMC. The data collected using the GNSS receptor were processed using the Trimble Business Center software, while the coordinates collected by the total station were corrected using the Spectrum link software. The salinity of the soil was measured subsequently along the topographic gradient. For this, samples of interstitial water were extracted from a depth of 30 cm using a pipette inserted through a 200 mm diameter PVC tube, which was buried in the ground. ## Tree allometric dataset The mangrove tree species, *A*. *germinans*, forms the scrub mangrove forests located on the highest part of the peninsula, that is, at 3.5 m a.s.l., covering an area of approximately 812 ha, which represents around 5% of the total area of mangroves on the peninsula. These stressed mangrove forests present a gradient of structural features, with shrub-like trees of heights as low as 30 cm, many twisted branches resulting from regrowth, and stunted trees, with some individuals of up to 800 cm in height. The structural characteristics of a forest can provide important parameters for the development of allometric equations, although the results of destructive sampling can provide more realistic values. This supports the destructive sampling procedures adopted in the present study, which provide more accurate parameters for the application of the allometric equations than the data available from other sites. Thus, mangrove trees with different classes of height were cut, according to the Instituto Chico Mendes de Conservação da Biodiversidade (ICMBio), Licence Nº 60471/2017. The scrub mangrove forest was divided in three well-defined strata, with tree heights of (i) 30–120 cm, (ii) \>120–250 cm, and (iii) \> 250, and emergent trees of up to 800 cm. Considering the destructive approach of the sampling method, we measured 15 trees for each height class, which is a representative sample number for BGB studies, with the equivalence of the samples being the basis for the comparative effect between the height classes. The choice of the sampled trees was based on the variation existing in the range of each height class, that is, the height of the collected trees was well distributed within each class, reducing the sampling bias. The selected trees were well-separated from their neighbors along the scrub mangrove forest that covers an area of approximately 812 hectares, in order to facilitate the excavation of their roots. A series of measurements were taken from each tree: (i) total height (h; m); (ii) diameter at breast height (DBH), that is, 130 cm above the ground, in the case of trees that were at least 3.5 m in height; (iii) basal diameter (bd), that is, at 30 cm above the ground, in the case of trees that were less than 3.5 m in height; (iv) area of the tree crown, given by the formula: crown area = \[(R1)/2)\*(R2)/2)\]\*π\], where R1 = the greatest radius and R2 = the smallest radius, and (v) crown volume, given by the formula: crown volume = crown area\*h. We obtained a disk sample of each tree at base height (30 cm above the ground), from which a transverse section was extracted to determine the density of the wood (*ρ*; g cm^-3), based on the water displacement method. The dry mass was obtained after drying the sample in an oven at 70ºC for 72 hours or until reaching a constant weight. The basic density of the wood was obtained using the equation: *ρ* = M/V, where M = the dry mass (g) and V = the volume (cm³). The BGB was quantified by root sampling. The below-ground portion of the *A*. *germinans* trees was divided into three compartments: (i) root crown, (ii) primary roots, which originate from the root crown, and (iii) secondary roots, which originate from the primary roots. Trees were cut at a height of 15 cm above the ground to facilitate the excavation and removal of the root crown. The primary roots were exposed from their origin at the root crown to their deepest extremity. We extracted two primary roots from each tree, including the pneumatophores, only when these were buried, and their associated secondary roots. The roots were then taken to the Mangrove Ecology Laboratory on the Bragança campus of the Federal University of Pará, where the basal diameter and length of the primary and secondary roots were measured. The samples were then washed throughly and carefully, and their fresh weight was determined using a digital precision balance (0.02 kg). This material was divided into subsamples that were weighed to determine their fresh weight using a second digital precision balance (0.01 g). The subsamples were dried in an oven at 105ºC until they reached a constant weight, and they were then weighed again to determine their dry weight. The dry (Dry = *D*) to fresh (Fresh = *F*) weight ratio (*D*:*F*) was also calculated for every root compartment of each tree. The biomass values obtained for the excavated primary and secondary roots were used to develop allometric models to estimate the dry weight of the portion of each type of root that was not excavated. For this, we used the basal diameter of each excavated root as the predictor variable. The total dry weight of the primary and secondary roots was determined from the sum of the dry weights recorded for the “excavated roots” and the dry weights estimated for the “unexcavated roots”. The total dry weight of the primary and secondary roots was then added to the dry weight of the root crown to obtain the total below-ground biomass of each tree. For this, we used the structural attributes of each tree (height, trunk diameter, crown area and volume, and the density of the wood) as the predictor variables for the development of the allometric models used to estimate the total BGB for the three tree height classes of the *A*. *germinans* scrub forest. We used a similar approach to estimate the AGB for the scrub forest, using the allometric models developed previously for the same study site. Finally, to transform the BGB values into carbon, we used the carbon concentration (42.6%) recorded for *Avicennia schaueriana* Stapf & Leechman *ex* Moldenke in the mangroves of the Brazilian Southeast. The AGB was converted based on the carbon concentration (41.9%) estimated by Carneiro for the same scrub *A*. *germinans* forest study area. ## Data analysis The weight of the “unexcavated” primary and secondary roots of each height class was estimated using models developed specifically for this purpose. As the relationship between the dry weight of the primary/secondary roots and the basal diameter was non-linear, since in biomass data it is often a power function of a variable that express tree size, a number of different regressions were applied to describe these relationships. The power (y = a\*x^b) and second- order polynomial (y = ax²+bx+c+e) functions were the best models, where y = the dry weight of the primary or secondary root (kg root^-1), x = the basal diameter of the root (cm), a, b and c = the model parameters, and e = the additive error of the model. The normality of the residuals of each regression was verified using the Shapiro-Wilk test and the homoscedasticity was verified using the Breusch-Pagan test and the graphic analysis of the residuals. We selected the best model to estimate the biomass of the “unexcavated” primary and secondary roots based on the goodness of fit tests: (i) adjusted coefficient of determination (R²_adj), (ii) Root Mean Square Error (RMSE) and (iii) Mean Percentage Error (MPE). Afterwards, we developed linear and non-linear allometric equations to estimate the BGB and calculated R²_adj, RMSE, MPE, and Akaike’s Information Criterion (AIC) to define the best statistical model for each tree height class: 1. R²_adj $$R_{adj}^{2} = 1 - \left\lbrack \frac{\left( {1 - R^{2}} \right)*(n - 1)}{n - k - 1} \right\rbrack$$ 2. Root Mean Square Error (RMSE): $$RMSE = \sqrt{\frac{\sum e^{2}}{n}}$$ 3. Mean Percentage Error (MPE). $$MPE = \left( \frac{\sum{(e)/n}}{M_{obs}} \right)*100$$ 4. Akaike’s Information Criterion (AIC): $$AIC = n*\left( ln\left( \frac{\sum\left( e^{2} \right)}{n} \right) \right) + 2*\left( {k + 1} \right) + c$$ where: *n* is the number of samples, *k* is the number of independent variables present in the model, *R*^*2* is the coefficient of determination, the term “*e*” refers to the residuals, that is, the difference between the observed and predicted values, *M*_*obs* is the average observed dry weight, and *c* is the constant. The total BGB values estimated for each height class were compared using the nonparametric Kruskal-Wallis analysis of variance (H), with Dunn’s *post hoc* test. This same procedure was used to compare the contribution of each compartment to the BGB among the different height classes. The variation in the BGB values between different compartments \[root crown and roots (primary+secondary)\] in each height class was verified using the t test. All the analyses were run in the R studio 3.6.0 platform. The non-linear allometric equations were developed using the *nls2* package and the *post hoc* test in *FSA* package, both in the R platform. # Results ## Soil salinity and topography The scrub mangrove forest varied considerably in height across its distribution on the Ajuruteua Peninsula as a result of the variation in both topography and the salinity gradient. Our results revealed that topography-driven salinity reduces tree height by approximately 7 meters, and changes the habit (i.e., shape and growth) of the *A*. *germinans* individuals, with individuals in height class 1 presenting a bushy habit, with multiple stems. The reduction in height was inversely related to both the increasing topographic gradient (elevation increasing 0.13 m, from 3.39 m to 3.52 m a.s.l.) and interstitial salinity, which increased 55 ppt, from 45 to 100 ppt, the maximum reading of the RHS-10/ATC refractometer. ## Belowground biomass allometry The three height classes of the scrub mangrove presented different patterns of BGB according to the models developed for the prediction of the dry weight of the unexcavated primary and secondary roots, with all parameters estimated being significantly different at the 1% level. Similarly, the values of all the selection criteria of the models developed using the residuals for validation, indicated that the equations selected have high predictive power. In general, the models generated for height class 3 were the most accurate in comparison with the other two classes. The models developed for the primary and secondary roots of this class were the best adjusted (R²_adj = 0.98). However, when the two models are compared, the lowest RMSE value (0.005) was recorded for the secondary root model. The coefficients of determination (R²_adj) explained between 86% and 98% of the variance in the biomass observed in each height class analyzed and in each type of root. All the models presented low MPE values (-0.75–0.23%), where the negative values indicate underestimates, and the positive values, overestimates. The residuals of all the models were normally distributed and had homogeneous variances. As for the models used to estimate the unexcavated roots, the coefficients of all the models selected to estimate the total BGB and that of the different compartments were significant. These models also had a high degree of predictive power, and the residuals were also distributed normally and had homogeneous variances. While a number of linear and non-linear relationships were tested, the linear and power equations were the most adequate. In the case of height classes 1 and 2, the best equations were found for the crown, based on the selection criteria used (RMSE = 0.002 kg; AIC = -80.74; R²_adj = 0.98 for class 1, and RMSE = 0.03 kg; AIC = -27.50; R²_adj = 0.95 for class 2). In the case of height class 3, the best equation was that developed for the total biomass, considering the same criteria (RMSE = 0.16 kg; AIC = 4.93; R²_adj = 0.99). The MPE values ranged from -0.23% to 0.01% among the different classes, with the values related to class 3 representing underestimates, while those referring to class 1 represented overestimates. ## Biomass storage and compartments contributions The contribution of the crown to the BGB increased proportionately with increasing height class, whereas the values recorded for the roots follow the opposite pattern, that is, decreasing with increasing height. In height class 1, the contribution of the crown to the total biomass represents only 43% of that of class 3, whereas the contribution of the roots decline 7.4% between classes 1 and 3. A similar pattern was observed when the two compartments were analyzed together, that is, the percentage difference of the crown+roots decreased 13% between classes 1 and 3. The estimates of the mean BGB, AGB, total biomass, and the BGB:AGB ratio for each of the three scrub *A*. *germinans* height classes are shown in. Class 3, which includes the largest trees, had higher estimates of biomass, and was approximately four times more productive than class 2 and 20 times more productive than class 1. The production of biomass varied significantly among the height classes (BGB: *H* = 67.13, d.f. = 2, p \< 0.001; AGB: *H* = 64.01, d.f. = 2, p \< 0.001). In all cases, the *post hoc* analysis indicated that these differences were related primarily to the extremely low values recorded for class 1. No significant variation was observed when each height class was analyzed separately, however, with the production of biomass being directly proportional to the height of the vegetation. The results of the present study indicate an inverse relationship between the relative proportions of the BGB and AGB, and the height classes, that is, larger trees tend to produce higher BGB values that are proportionally more similar to the AGB values as a result of the increase in the percentage production of BGB and the reduction in the production of AGB. The estimate of the total biomass stored in the scrub *A*. *germinans* forest revealed a production of around 84 Mg.ha^-1 of BGB (48.6% of the total biomass) and 88 Mg.ha^-1 of AGB (51.4%). On the Ajuruteua Peninsula, the scrub mangrove forest covers an area of approximately 812 hectares, which implies a total production of approximately 139.7 Gg of biomass, and 59 Gg of carbon. Slightly more of the biomass (71.6 Gg) and the carbon (30.0 Gg) were allocated to the AGB in comparison with the BGB, with 68.0 Gg of the biomass and 28.9 Gg of the carbon. Similarly, the biomass (kg ind^-1) and respective C values (kg C ind^-1) were also calculated for an average *A*. *germinans* individual of each height class. ## Generalist allometric models *vs*. Size-specific models A generalist model (which includes all height classes) was generated to evaluate the effects of structural differences on the estimates of the BGB of the scrub mangrove forests of the Ajuruteua Peninsula, which was y = 0.07577 \* D^1.98745 (R²_adj = 0.96; AIC = 58.69; RMSE = 0.85, and MPE = -0.77). This model underestimates by 22.6 Mg ha^-1 (27%) the total BGB (83.8 Mg ha^-1) derived from the sum of the estimates of the three specific equations. # Discussion The principal aim of the present study was to develop reliable allometric models to estimate the BGB of scrub mangrove forests, to cover an important lacuna for the understanding of the production of biomass and carbon storage in mangrove forests, given that the available models refer only to the production of AGB in this type of forest. The models we developed to estimate the BGB for the different height classes found in the scrub *A*. *germinans* forest followed both linear and power trends, which is typical of the models used to estimate the biomass of tropical forests, including mangroves. Tree height and diameter are the structural attributes included most often in our models, with no major differences in comparison with the allometric equations developed to estimate the BGB of other, non-scrub mangrove forests around the world, in either species-specific or generalist models. The estimated BGB values for the different height classes of scrub mangrove forest are within the range of values reported for mangrove forests at other localities, such as the Everglades National Park (24–47 Mg ha^-1) and Rookery Bay (29–284 Mg ha^-1), both in Florida, in the United States, and the Endings Lagoon (9.8 Mg ha^-1) in Mexico. Although a number of studies have provided estimates of the BGB of non-scrub mangrove forests in different parts of the world, data are still relatively scarce overall. Even so, broad comparisons show that our values are higher than the estimates available from the vast majority (89%) of sites in the 71 countries that have mangrove forests, as well as the mean value of approximately 27 Mg ha^-1 estimated for other types of forest around the world. It is important to note, however, that much of this discrepancy may be related to the effects of the application of different sampling methods, which reinforces the need for caution when comparing the results of studies based on distinct approaches. The BGB estimates available for mangrove forests around the world have been obtained using a range of both direct and indirect approaches, such as the trench method, extraction (pull up), and the analysis of soil cores, resulting in highly diverse biomass estimates. The root sampling method adopted in the present study has been applied successfully in previous studies of mangroves and other types of tropical forest around the world. The discrepancies resulting from the application of different sampling methods become especially apparent when our values for the scrub mangrove forest are compared with those obtained for other, non-scrub mangrove forest, although they present similar biomass values. However, the estimates of BGB available for hypersaline environments are relatively low overall, as was the case in the present study. This reinforces the conclusion that the soil core sampling method, which focuses only on the fine roots (20 mm), will likely underestimate the BGB of mangroves. Much higher values have been recorded, by contrast, in studies in which the roots are excavated, either completely (total excavation) or partially (trench, pull up, sampling method), in comparison with sampling methods that do not incorporate the roots of larger diameter. Some studies have indicated that methods in which the roots are excavated provide relatively reliable estimates of the BGB, despite the fact that some of the smaller and finer parts of the root are lost during extraction. Based on the models developed to estimate the AGB of the scrub mangrove forests of the Ajuruteua Peninsula, it was possible to estimate the total biomass of this type of mangrove. Our findings also indicate that the below-ground compartment of the scrub mangrove forests contributes a larger proportion of the biomass (48%) than that estimated for mangrove forests under minimal environmental stress, such as those studied in Tanzania, where the BGB constituted 41% of the total biomass. Some studies have concluded that as much as 80% of the live forest biomass worldwide is located in the above-ground compartment, and only 20% below ground, with slightly higher percentages (\~25%) being found in tropical forests. Our results are also consistent with those of previous research which indicates that the mangroves are characterized by a relatively high percentage of BGB in comparison with other tropical forests. This means that half of the total biomass and carbon stocks of the scrub mangrove forests on the Ajuruteua Peninsula is allocated to the below-ground compartment. We observed an inversely proportional relationship between the abiotic factors (topography and salinity) and the biotic variables (tree height and percentage BGB), that is, the greater the topography and the higher the salinity, the lower the height of the trees and their production of BGB, which is the exactly opposite pattern observed in the production of AGB. However, our estimates of the production of BGB indicated a pattern that contrasted absolutely with that recorded by Saintilan, who found an increase in the contribution of the BGB with increasing salinity. The results of the present study also indicate higher AGB and BGB values than those recorded in a dwarf *Kandelia obovata* (S. L.) Yong forest in Japan. In the present study, the percentage estimates of the biomass for height class 2 were relatively similar to those recorded in the Japanese mangrove. This is almost certainly a reflection of the structural similarities of the two types of stunted mangrove trees, given that the *A*. *germinans* trees of height class 2 were the same size as the dwarf *Kandelia* trees. However, the absolute biomass recorded in this study in Japan were similar to those of height class 3 in the present study, which may be accounted for by both the differences in the methodological approaches to the estimation of the BGB and the varying responses of the trees to the different local environmental factors. The allocation of the biomass in the scrub mangrove forest is influenced by a range of factors, including the diameter of the tree. Other factors, such as the local frequency of inundation, may also contribute to the dynamics of the compartmentalization of the biomass in these forests, in particular in response to environmental stressors. This implies that fluctuations in flooding patterns also play an important role in the hydrological and/or saline stress of these environments, leading to an increase in the proportion of the BGB. Our findings are consistent with this conclusion when the BGB estimates of the three height classes are compared. The significant variation found among classes in the BGB values may be explained by the variation in the salinity of the soil within the study site. The highest salinity was recorded in the areas dominated by shrubby *A*. *germinans* individuals from height class 1, and the lowest in areas dominated by the taller individuals from class 3. A similar tendency was found in *Avicennia marina* (Forssk.) Vierh. forests under hypersaline conditions in Australia, further reinforcing this finding. Overall, our results cover a major lacuna in the models available to estimate the production of below-ground biomass and carbon storage by scrub mangrove forests, and also contribute to the refinement of the approach used to estimate total biomass in this environment. The findings of the present study indicate an inverse relationship between the stature of the vegetation and the production of BGB in the scrub mangrove forests of the Ajuruteua Peninsula, which contributes to the correction of uncertainties on the compartmentalization of the biomass and carbon in this forest. It is particularly important, in this context, to take into consideration the systematic errors in the allometric models used to estimate the BGB and AGB, given that the differences among studies can be accounted for primarily by the uncertainties intrinsic to the different models. As the selection of the model is an important source of uncertainty, models developed specifically for each height class of the mangrove forest provide more accurate estimates, reducing the uncertainties intrinsic to the different biomass estimates (total, above- and below-ground). As major differences exist in the carbon stocks among different height classes, the specific allometric models developed for each height class should normally be applied rather than generalist models. The results of the present study describe the effects of the gradient of topography and salinity on the production of biomass and carbon storage of scrub mangrove forests. All the models were based on direct measurements of the size and weight of the trees, which permitted the systematic calibration of the data, which reinforced the accuracy of the calculation of the tree biomass and carbon stocks of this type of mangrove forest. This implies that the site-specific models developed in the present study may be a valid option for the analysis of the extensive tract of mangrove found on the Brazilian Amazon coast, and similar coastal environments in other parts of the world, where scrub mangroves dominate much of the landscape. Ultimately, improved accuracy in the biomass estimates will be fundamental for the systematic evaluation of the process of carbon storage, and will be essential for the development of effective strategies for the conservation and management of the mangrove, as well providing potentially valuable indicators for the analysis of the impacts of climate change. # Supporting information We gratefully acknowledge the Laboratório de Ecologia de Manguezal (LAMA) for logistical and technical support. The authors are also grateful to the anonymous referees who contributed to improve this paper. [^1]: The authors have declared that no competing interests exist.

# Introduction Studies of local high-frequency network oscillations beyond the spectral frequency limits of traditional electroencephalogram (EEG), i.e. greater than 40 Hz, have increased dramatically over the last decade. This evolution can be attributed to early animal and human studies on high frequency oscillations (HFOs) in subcortical limbic and neocortical structures, which suggest that HFOs play a role in neurological disease. For epilepsy in particular, HFOs are believed to reflect some basic neural disturbances responsible for epileptogenesis and epileptogenicity. Specifically, different studies in animals and humans have suggested that HFOs reflect abnormal synchronization due to the impairment of neuronal and network processes. Most of these initial studies were carried out using microelectrodes and specialized systems capable of wide bandwidth recordings that could detect HFOs containing spectral frequencies up to 600 Hz. More recently, the use of clinical EEG systems supporting wide bandwidths has revealed that HFOs can be also recorded using conventional depth and grid electrodes. Based on these evidences, some studies have considered the HFO density as a hallmark of the epileptic foci, which has generated a growing interest in the detection and analysis of these events. Despite these advances, the clinical examination of such recordings remains limited mainly because the detection of these events is a challenging procedure. Specifically, HFO events have a relatively low signal to noise ratio compared to other activities (e.g., interictal epileptiform discharges). These events can occur as brief bursts lasting 30 ms or less, and they are found in brain areas capable of generating seizures. In addition, the appropriate identification and analysis of HFOs requires large storage and computational capabilities. This is particularly true for clinical explorations, where subjects are continuously monitored for days, weeks or even months, creating massive amounts of data to be stored and processed. This data is usually acquired at high sampling rates, varying from 10–30 kHz, 1–4 kHz or below 1 kHz for intracranial micro and macro electrodes, or surface EEG respectively, where several locations in the brain are simultaneously acquired (varying from dozens to hundreds of contacts). To complement this, there is not a general criterion for the identification of HFO events, and their demarcation differs between experts. Indeed, during the HFO validation process, the inter-rater reliability should be taken into account, but currently there is not a common procedure to share HFO analyses across different groups. To overcome all these issues, the following two strategies have been proposed: the development of analytic techniques that can reliably detect HFOs in continuous wide bandwidth EEGs recorded from microelectrodes and conventional clinical electrodes and the creation of infrastructure projects that facilitate the sharing of wide bandwidth data. Specifically in terms of the first strategy, several methods for the automatic detection of HFOs have been recently developed. Nevertheless, despite efforts to develop a reliable procedure to correctly identify HFOs, there is currently no consensus about the one method that performs best in all contexts. As solution to this, it has been suggested that different algorithms for HFO detection should be tested across similar datasets to improve the performance of automatic HFO detection methods, and these algorithms should permit the change of multiple parameters on the go if meticulous studies are required. It is important to consider that most of the developed methods require sophisticated mathematical and computational tools for their implementation and operation, which can substantially reduce the detection algorithms utility in contexts where non-technical researchers and clinicians could potentially benefit. Also, within these strategies, it is required the development of analysis tools that can accurately characterize and quantify this type of events including the possibility to share the HFO analyses. Though several open-source computational tools exist for the analysis of neural data (see for instance MEA-Tools, EEGLAB; SigTOOL; BioSig; FieldTrip; AnyWave), none of them have been designed for the analysis of HFOs, and they lack the adequate environment for manual and automatic detection and validation of this type of events as well as the implementation of several HFOs detection algorithms. Consequently, some groups have developed customized interfaces to test their own algorithms, or they have joined professional EEG companies to develop proprietary software. Unfortunately, these applications currently remain insufficiently documented or non-available to the general public, restricting their use and validation and making difficult the further support for comprehensive HFO analyses. To achieve a practical and reliable tool for HFO analysis, we developed RIPPLELAB, a MATLAB open-source application (Mathworks®, Natick, MA, USA). This tool integrates several methods for HFO detection within an easy-to-use graphical user interface (GUI), which assists clinicians and researchers in the identification, selection and validation of HFOs. In particular, RIPPLELAB allows users to implement different quantitative measures to optimally identify and classify HFOs, reducing significantly the time required to visualize and validate putative events. RIPPLELAB also proposes a common procedure to share HFO analyses in order to promote collaborations across research centers. This tool was designed to be manipulated by researchers and clinicians with no programming skills, and users do not require any programming abilities to execute an analysis. More seasoned scientists can expand the possibilities through the configuration of advanced parameters or the inclusion of complementary modules # Materials and Methods ## HFOs Detection Methods HFOs are cortical discharges observed in the EEG. They are usually defined as local field potentials (LFP) of short duration (30–100 ms), with more than three oscillations distinguishable from the background activity. HFOs are usually categorized depending on their main frequency component as *gamma* (60–150 Hz), *ripples* (150–250 Hz) and *fast ripples* (\> 250 Hz), where gamma are essentially considered physiological oscillations, and fast ripples are mainly studied as pathological oscillations. Conversely, ripples can be considered physiological or pathological oscillations according to different factors such as their recording region. In particular, pathologic epileptic HFOs can occur as independent events or together with spikes. Detection methods exploit all these characteristics to identify HFOs from background. Broadly speaking, HFO detection methods can be classified in three groups: manual review, supervised detection and unsupervised detection. Manual review is performed by visual inspection of an expert with required knowledge on electrophysiological recording and signal processing to distinguish putative events from filtered artifacts. Supervised detection relies on detection methods with high sensitivity and low specificity detection, which is complemented with manual review. Finally, unsupervised detection requires methods with high sensitivity and high specificity to be effective, which is difficult to achieve in databases with dissimilar characteristics. HFO detection by manual review is performed by visual inspection of EEG recordings in parallel with a frequency representation of the same recording. Two methods of visual marking are commonly used: the visual inspection of the time-frequency plot in parallel with the focus electrodes, and the simultaneous inspection of raw data, filtered data \> 80 Hz and filtered data \> 250 Hz. By definition, two main issues arise when the visual inspection methodology for HFOs detection is applied: (i) a high subjectivity because the visual marking is influenced by the perception of the reviewer, and (ii) the increased time- consuming of the process that limits the amount of signal to be analyzed by the reviewer, which is usually in the interval between 5–10 minutes. Yet, this method is considered the gold standard when assessing the performance for most of the automated algorithms for HFO detection. In general, the supervised and unsupervised detection algorithms achieve a series of common steps for detection of putative events. The first step is to emphasize the frequency of interest by filtering the raw signal. Then, a detection measure for thresholding is implemented, which can be based on energy, statistical or spectral characteristics of the filtered data. Finally, depending on the method, a supervised or unsupervised mechanism for discriminating HFOs from noise, artifacts and spikes is implemented. The HFO detection methods have been developed according to the needs of each research center: The first supervised method for the automatic detection of HFOs was implemented by Staba et al.. This method analyzes 10-min EEG segments in frequencies between the 80 and 500 Hz by applying a band-pass filter, and subsequently a root mean square (RMS) for the detection of HFO events. The authors reported a sensibility of 84%. Gardner et al. implemented a similar strategy by selecting putative HFOs through the evaluation of the *Line Length* energy from the equalized filtered signal in 3-min epochs. The reported sensitivity was 89.5%. Crèpon et al. applied the Hilbert envelope to select putative events, achieving a reported sensitivity of 100% and a specificity of 90.5%. The first unsupervised method for HFO detection to our knowledge was developed by Blanco et al.. The authors processed 10-min EEG signal identifying putative HFOs using the Staba’s method, and a series of features from each event were computed to develop a K-Medoids/Gap-Statistic clustering to separate different types of events, where true and false HFOs were characterized. In this study, the authors reported no measures of sensitivity or specificity. Zelmann et al. implemented a supervised method to improve HFO detection in EEG signals with continuous high frequency activity. They reported a sensitivity and specificity of 91%. Dümpelmann et al. developed another unsupervised method adjusted to the detection of ripples. This method selects the RMS amplitude, the *Line Length* energy and the instantaneous frequency as main features, and then a radial basis function neural network is used to select real events. In this work, the authors reported a sensitivity of 49% and specificity of 36.3%. Another supervised algorithm was implemented by Birot et al. to detect fast ripples in the 250–600 Hz band. This algorithm uses the RMS amplitudes and a Fourier or Wavelet energy ratio to detect putative events. The authors reported a sensitivity of 93% and a specificity of 95% with optimal parameters. Matsumoto et al. applied a support vector machine with features extracted from putative HFOs recognized by a *Line Length* detection, employing spectral characteristics for training. The authors reported a sensitivity of 83.8% and a specificity of 84.6%. Burnos et al. implemented a Hilbert envelop detection and they included a time-frequency analysis for a spectral characterization of the selected events; the authors reported a total sensitivity of 50% and specificity of 91%. Finally, a recent unsupervised algorithm was developed by Gliske et al., where a parallel detection of artifacts and events is implemented. This study reports a sensitivity of 78.5% and specificity of 88.5%. ### Algorithms for automatic detection of HFOs Four methods for automated detection were implemented in RIPPLELAB. They were selected based on the following three characteristics: (i) the detection method is supervised, (ii) the algorithm has the ability to detect ripples and fast ripples, and (iii) the sensitivity reported is higher than 80%. Nonetheless, because of the RIPPLELAB’s modular structure, other methods for HFO detection can be easily integrated into the application if desired. The first implemented method, Short Time Energy (STE) is the algorithm proposed by Staba et al.. In brief, the wideband EEG signal is band-pass filtered in the high frequency range. The energy from the filtered signal is then computed using the RMS defined by the equation $$E(t) = \sqrt{\frac{1}{N}\,\sum_{k = t - N + 1}^{i}x^{2}(k)}$$ within a N = 3 ms window, and successive RMS values greater than 5 standard deviations (SD) above the overall RMS mean are selected as putative HFO events if they last more than 6 ms. Finally, only the events containing more than 6 peaks greater than 3 SD above the mean value of the rectified band-pass signal are retained. In addition, the events separated by 10 ms or less are marked as a single oscillation. In the original paper, the estimation of the energy threshold depended on the complete analyzed segment, which had a duration of 10-min. In RIPPLELAB, the energy threshold can be computed for the entire signal, as originally proposed by Staba et al., or for shorter segments, as suggested by Gardner et al.. The flowchart of the implemented STE algorithm is shown in. The second implemented algorithm named Short Line Length detector (SLL), was developed by Gardner et al.. In this approach, a preprocessing stage is done with a derivative filter in order to equalize the spectrum of the signal. Next, a band-pass filter is applied. From this, the energy of the signal is calculated by a short time line length measure defined by $$E(t) = \sum_{k = t - N + 2}^{i}\left| {x(k) - x\left( {k - 1} \right)} \right|$$ with window N = 5 ms. An event is valid if its amplitude is greater than the 97.5th percentile of the empirical cumulative distribution function for each 3-min epoch and if it has a minimum duration. This duration is set to 80 ms in, but it is ignored in. We set this parameter to 12 ms by default in order to accept events larger than 6 oscillations at 500 Hz. The flowchart of the implemented SLL algorithm is shown in. The third method we included was proposed by Crépon et al.. In this method, the signal is first filtered between a selected frequency range, and the envelope is then computed with the Hilbert transform. For an event to be considered valid, two conditions must be met: first, for each event, the local maximum must exceed a threshold of 5 SD of the envelope calculated originally over the entire recording or from a time interval. Second, each detected HFO must have a minimal time length of 10 ms. This method is called Hilbert Detector (HIL) in RIPPLELAB, and its flowchart is presented in. As in the STE detector case, we included the possibility to analyze the threshold by epochs specified by the user. The last algorithm incorporated, the MNI detector (MNI) was developed by Zelmann et al.. In this method the signal is first band-pass filtered. Then, a baseline detection procedure based on the wavelet entropy is applied. For this, the signal is divided into segments of 125 ms with 50% overlap. Next, for each segment, the normalized wavelet power of the autocorrelation function is computed using the complex Morlet wavelet. Subsequently, the maximum theoretical wavelet entropy from the segment is obtained for the white noise, and the segment is considered as a baseline interval when the minimum entropy is larger than a threshold. If a sufficient amount of baseline exists, HFO candidates are detected in accordance with the energy, defined as the moving average of the RMS amplitude of the filtered signal. Segments with energy above a threshold and lasting more than 10 ms are considered as HFOs. Similar to other methods, events located less than 10 ms apart are considered as single events. If a sufficient amount of baseline is not present in the signal, an iterative procedure is carried out where the threshold is computed for the band-passed signal. Originally, this detection methodology was implemented with 1-min segments of EEG signal. In addition to this, we included the possibility to process the data thresholds in epochs of time specified by the user. The flowchart of the MNI algorithm is presented in. ## Software Overview RIPPLELAB is a multi-window GUI developed in MATLAB for the analysis of high frequency oscillations. It is intended to be a user-friendly and intuitive tool, where users with technical and non-technical backgrounds can explore and analyze brain oscillations from different types of electrophysiological data, especially at high frequency ranges. RIPPLELAB has been released under GNU Public License version 3, and the source code and documentation can be found in <https://github.com/BSP-Uniandes/RIPPLELAB/>. The code was originally written in MATLAB version 7.12, and it is compatible with later versions. The tool was developed with a modular design, allowing expert users to modify and integrate new developments. RIPPLELAB was developed exclusively using MATLAB scripts; therefore, the compilation of native libraries or external functions is not required. This multi-platform tool can be used on OS X, Linux and Windows 32 and 64-bit architectures, and it can be installed as a MATLAB App in versions equal or later to R2012b. The RIPPLELAB multi-window approach is displayed in. We chose MATLAB as programming platform because of its widely extended usage in the epilepsy and HFOs research environment. The complete procedure for detection and analysis of HFOs through RIPPLELAB consists of several steps that are briefly presented in the following subsections and are summarized in. They include different options for the general display and pre-processing of electroencephalographic data. Specific information about how to run the software is given in detail in the RIPPLELAB’s user manual, which is distributed along with the source code. ### Step 1: Importing files and selecting channels RIPPLELAB can import several file types including European Data Format, (\*.rec, \*.edf), Nicolet files (\*.eeg), Neuralynx files (\*.ncs), EPILEPSIAE file format (\*.data), EEGLAB files (\*.set), Plexon files (\*.plx), Axon binary format (\*.abf), Micromed files (\*.trc) and custom MAT Files (\*.mat). Furthermore, when other MAT files do not meet the requirements, these files can be converted through RIPPLELAB to the appropriate format in order to be loaded into the system. Also, due to its modular structure, users and developers can easily add other file types if needed. The user can simultaneously analyze one or several files of long-term recordings. Moreover, users can create and save custom bipolar or average montages for both display and analysis. Once the channel selection is done, the user can either visualize the data for pre-processing or go directly to implement the HFO detection. If this last option is chosen, RIPPLELAB does not load the entire file in memory; it only keeps the channel information for subsequent analysis. It is important to note that the software is set to load all the selected data into memory for further processing by default. Hence, when working with large files it is recommended to select only specific time segments of the focus channels to avoid overloading the system memory. For this, the user can modify the start and end times of the signal to be loaded. ### Step 2: Displaying and pre-processing data As shown in, RIPPLELAB offers useful options for visualization and pre- processing signals. This tool allows the user to visually explore the raw data and to compute different measures to better tune the HFO detection algorithms. This step is optional, and it is not required for HFO analysis. The pre-processing options include filtering, time-frequency and spectral tools. *1 –Filtering*: Different filter options can be implemented in RIPPLELAB. The user has the possibility to select between causal and non-causal filters to implement low-pass, high-pass and band-pass configurations with custom frequencies to the entire group of electrodes or to a selected group of channels. Likewise, a notch filter can be implemented at 50 or 60Hz with and without harmonics. The *causal filter* implements a Hamming-windowed FIR digital filter of 50th order and a cutoff attenuation specified at -6dB. The *non-causal filter* employs a type-II Chebyshev IIR forward-backward digital filter, which has a passband ripple of no more than 1 dB and a stopband attenuation of at least 20 dB, a cutoff attenuation specified at -3dB, and a second-order section implementation to maintain stability. These filter characteristics allow the user to implement highly selective filters with narrow pass bands. *2—Time-Frequency analysis*: The user has the possibility to perform a time- frequency transform of one selected electrode in a desired range of frequencies for the displayed interval. This time-frequency transform is estimated by the scalogram, which is computed with the continuous Gabor Wavelet and is defined by the equations: $$\mathbf{C}\left( {\mathbf{s},\mathbf{\tau}} \right) = \frac{1}{\sqrt{\mathbf{s}}}\int_{- \infty}^{\infty}\mathbf{x}\left( \mathbf{t} \right)*\mathbf{\psi}\left( \frac{\mathbf{t} - \mathbf{\tau}}{\mathbf{s}} \right)\mathbf{d}\mathbf{t}$$ $$\mathbf{\psi}\left( \mathbf{t} \right) = \frac{1}{\left( {\mathbf{\sigma}^{2}\mathbf{\pi}} \right)^{1/4}}\mathbf{\exp}\left( \frac{- \mathbf{t}^{2}}{2\mathbf{\sigma}^{2}} \right)\mathbf{e}^{\mathbf{i}\mathbf{\eta}_{\mathbf{s}}\mathbf{t}}$$ where *t* indicates time, *s* represents scale, *τ* represents translation, *η*_*s* indicates the angular frequency at scale *s* and *σ* indicates the standard deviation of the Gaussian window in time. We set *σ* = 6/*η*_*s* in order to satisfy the admissibility condition. *3—Cursors and power spectrum*: The user can estimate different measures such as time position, amplitude, time duration, and minimum, maximum and average amplitude of the segment between both cursors. Moreover, the power spectrum density (PSD) using the Welch estimation can be obtained in a new overlapping panel for the segment between the cursors. ### Step 3: HFOs detection methods To provide a comprehensive support for HFO analysis, RIPPLELAB offers several alternatives for manual and automatic detection, visualization and manipulation of events. Each detection method includes a configuration panel, in which users can set the parameters and choose different analysis options. ### Options for manual detection of HFO events The Visual Marking option assists the manual selection of HFO events depending on user criteria, and it can only be performed for a single electrode. This option displays the corresponding time-frequency plot together with the raw and filtered signals such as presented in. In addition, the power spectrum of a selected segment is estimated and plotted. The user can classify selected events as gamma (40–120Hz), ripple (120–240Hz) or fast ripple (\>240Hz). All of the implemented detection methods are configured by default with the parameters proposed in the original papers. However, the customization of these values is possible according to the user’s needs as displayed in. In order to help users to choose the appropriate method, some general considerations must be discussed. First of all, it is important to note that the methods we included in RIPPLELAB were designed to detect HFOs at different frequency bands. In fact, the STE detector was designed to find ripples in the frequency range between 80 and 175 Hz, and fast ripples between 200 and 600 Hz. The SLL method performs the detection over the frequency range \> 80 Hz emphasizing high frequencies throughout the derivative filter. Distinctively, the HIL detector focuses on events at the frequency range between 180 and 400 Hz, and the MNI method centers its detection over the range from 80 to 450 Hz. Another difference across methods is their definition of event length. Specifically, the STE and MNI detectors state 6 ms and 10 ms, respectively, as event minimum lengths. A comparison of the implemented methods with different parameters was already performed in, where it was indicated that each of the detections methods can be optimized according to the database and the HFO characteristics that the researcher requires. The problem of selecting the optimal parameters is not trivial, and it has not been solved yet When selecting a method to implement, we first suggest setting the *Frequency Limits* parameter to the range of interest, and subsequently carry out the detection with all methods over a short-time interval of a well-known signal (e.g. 1-min EEG segment). It is recommended to test several threshold levels and epoch times. For instance, when increasing the threshold of a determined method, the sensitivity decreases, but the specificity rises. Similarly, depending on the epoch time used for thresholding, the estimation of the local background energy differs due to changes in the vigilance state, artifacts and epileptic activity. Usual epoch times for energy thresholding are 10-min, 5-min, 3-min and 1-min, though still, the user can establish a custom epoch time. ### Inclusion of a new detection method in RIPPLELAB RIPPLELAB allows advanced users to include new detection methods. For this, the scripts are written in MATLAB sections named according to the functionality of the code, and the object handles are stored in MATLAB structures. To facilitate the process of edition, the sections to modify when a new method is included are marked with the comment: \[\*\*INSERT!\*\*\], and an example is provided for each case. In general, to include a new detection method the following two features must be set: the visualization of the panel that allows the configuration of different parameters associated with the new method and the selection of the new method for further processing. Additional detailed information on the inclusion of a new detection method is provided in the user manual. ### Step 4: HFOs visual validation For a more detailed analysis of the detected HFOs, a special interface has been created (*HFO Analysis Tool*) to simplify the visual validation procedure performed by the user. This last GUI is presented in, and it is composed of the following sections: - The *General Information* panel presents the name of the analyzed file, the selected method for detection and the currently selected channel. - The left panel presents a short segment of individual events. In the top panel, the raw signal is plotted highlighting the HFO in red. The equivalent filtered signal and the time-frequency plot are presented in middle and bottom panels, respectively. - The *Selection Panel* allows users to visualize event by event. Valid HFOs can be confirmed, and spurious events can be rejected. - The *Frequency Controls* panel allows the user to change the frequency limits for both the filtered signal and the time-frequency plot. - The *Event* Panel allows the user to make manual adjustments to the temporal limits of the detected event through the time cursors provided for this effect. In this panel, the user can define the event type according to the classification: *Gamma*, *Ripple*, *Fast Ripple*, *Spike*, *Artifact or Other*. Additional types can be easily added if needed. Information about the maximum peak frequency and the time duration of the event is also provided. Lastly, the *Fast Ripple Index*, *the Ripple Index and the Gamma Index* are provided for reference, as well as the peak frequency of each band. ### Step 5: Logging, saving and retrieving A main feature of RIPPLELAB is that the result of the analyses carried out during a session can be saved in files for future reviews, validations or sharing. For this purpose, RIPPLELAB proposes a custom-made structure MAT-File with the extension.*mat* modified to.*rhfe* which is created after each analysis. The general organization of this structure is described in. Finally, logs of all operations accomplished during the detection procedure can be saved, which includes general information about the HFO detection method. ## Software validation In order to minimize eventual programming errors in all developed scripts, we extensively and systematically tested each step of all the implemented detection algorithms using different types of data with distinct characteristics of noise, pathological activity and vigilance states. Extensive validations were also carried out on the different functionalities for signal preprocessing and inspection (display, filtering, spectral estimation, etc.). The performance and reliability of the tool for HFO analysis was also exhaustively evaluated for all detection methods on simulated and real data. ### Simulated HFOs dataset To test the HFO detection capabilities of RIPPLELAB, we first created a controlled scenario where the following types of events were simulated: spikes, gamma, ripple and fast ripple as well as different types of artifacts including 50 Hz noise. All events were implemented as sinusoids and Gaussian curves. Gamma, ripples, and fast ripples were reproduced by implementing a sine wave of 125Hz, 225Hz and 325Hz respectively, and they were masked by a 50-percent cosine tapered window in order to obtain a sine train with 8-oscillations within full amplitude. Spikes were reproduced using a Gaussian curve with a temporal duration of 30 ms containing ±3.7 SD. Artifacts were constructed using a discontinuity, and the 50Hz noise comprised harmonics in the 100–500 Hz range with a decreasing power-law amplitude. Both artifacts and noise were also masked by a 50-percent cosine tapered window to avoid discontinuities at the edges. To perform the detection of HFO events, the simulated activities were placed on top of a 30-min background signal of 1024 Hz sampling rate. Two background cases were studied: non-background and real background. A zero- level signal was implemented in the non-background case, and a background activity from an intracranial EEG (iEEG) recording was used for the real background analysis. This electrophysiological interval was carefully chosen so that no high frequency events were present in the selected interval, which was visually confirmed by two experts. Similarly, the amplitude of this interval was normalized to have zero mean and one SD. In order to have a reasonable event amplitude fulfilling the spectral EEG power-law decay, we set the event amplitudes as next: Gamma = 3.5 SD, Ripple = 2.7 SD, Fast Ripple = 2.0 SD, Spike = 10 SD and artifacts and noise = 2 SD. Furthermore, spikes overlaid with fast ripples and spikes close to ripples (\< 30 ms apart) were also simulated maintaining the same amplitude levels. For both simulated background cases, all the detection methods were executed using the same parameters as published, and HFOs were detected for a frequency band between 80 and 500Hz. The sensibility and specificity of the detection was determined by the following equations: $$Sensitivity = \frac{TP}{TP + FN}$$ $$Specificity = \frac{TN}{TN + FP}$$ where TP stands for true positives, TN for true negatives, FP for false positives and FN for False negatives. Detections of gamma, ripples, fast ripples and spike + fast ripples oscillations were considered as true detections, while spikes, artifacts and noise were considered as false detections. ### Real EEG dataset To assess RIPPLELAB’s capability to handle large databases––a common scenario found in clinic and research contexts––we completed an extensive analysis of 16 patients with invasive macro-electrodes from the EPILEPSIAE database. This database contains long-term EEG recordings of epileptic patients (a total of 275) complemented with extensive metadata and standardized annotations of the data sets. These patients were all implanted with invasive macroelectrodes (depth electrodes, or subdural grids and strips) as a part of their clinical procedure to determine the epileptogenic zone. In our work, we first evaluated the automatic detection of HFOs from large amounts of data analyzing long-term EEG records, and then we implemented a visual validation for a group of candidate events. For the first part of the analysis, two nights per patient were selected, one of them without any clinically marked seizure. For two patients, however, only one night was included because they did not present seizure-free nights. Only channels in the seizure onset zone were analyzed (n = 205 from 16 patients). A total of 3471 hours of iEEG sampled at 1024 Hz were processed using the STE and SLL automated detection methods. We defined epochs of 3-min for energy computation and a frequency range from 80 to 500 Hz for both of them. For the second part of this analysis, we evaluated the type of oscillations that were identified for the implemented detection methods. To do this, 20-second channel segments with a high rate of detections, 100 to 1000 per hour, were randomly chosen from the results obtained in the previous analysis (n = 22). For these segments, the four automatic detection methods were applied and a visual validation procedure was performed. As suggested by Worrell et al. and Menendez de la Prida et al., only events with at least four oscillations and minimum duration of 25 ms between adjacent events were considered as HFOs. Events containing high and sharp amplitudes without a clear superimposed HFO were classified as Spikes. The remaining events were categorized as noise or artifacts, and thus labeled as Other. # Results ## Simulated data analysis For both background cases all methods accomplished high sensibility and specificity. Specifically, for the non-background case, all methods accomplished a sensibility and specificity of 100%. In addition, the event detection was consistent among each detection method, meaning that the event boundaries established for each method were uniform across all the events of the same type. It is important to note that for the MNI detector, the analysis was carried out using the detection for channels with baseline (right branch), which is expected because there are not components inside the band of interest besides the simulated events. In the real background case, sensitivities for STE, SLL and MNI detectors were 100%, and 99.33% for the HIL method. For this same case, the specificities were 100% for STE and HIL methods, 99.97% for SLL and 97.79% for the MNI detector. Nevertheless, event detection was also consistent among the implemented methods as can be seen in. As expected, these results reveal that HFO detection is highly affected by the signal background. The results are also consistent with the specifications for SLL and MNI detectors which lose specificity in order to increase their sensibilities. ## Real data analysis As a result of the first part of this analysis, the SLL method detected the largest number of events in comparison to the STE (1.040.437 and 252.642 respectively). On average, the STE method spent 32.07 ± 10.6 seconds analyzing hour-long segments whereas the SLL method performed slightly faster processing one hour segments in 30.52 ± 10.2 seconds. Nevertheless, the proportions of detected events using both methods were comparable for all but three patients. This result suggests that STE and SLL methods identify similar HFO densities on most of the signals even though the characteristics they use to evaluate HFOs are different, which is expected for different methods of high sensitivity. For a complete summary of the results, see. For the second part of the analysis, a total of 14.804 events were visually reviewed for an expert. For all methods, 4542 events were classified as valid HFOs, 5115 as spikes, and 4967 as others. In our study, the MNI detector presented the highest number of HFO detections (n = 1929), but also had the highest level of false detections (true HFOs events represented only 27% of total detections). Conversely, the STE detector presented the lowest number of HFOs detected events but with the highest true detection rate (n = 418 corresponding to 43% of total detections). Moreover, the HIL method obtained the highest spike detection rate (50%) and SLL the highest noise detection rate (49%). For a complete summary of the results see. Examples of different types of detected events are present in. Although the main purpose of our analyses was to evaluate RIPPLELAB’s capabilities and not to perform a rigorous comparison between methods, our results are consistent with a previous study that made a comparison of these four algorithms. In particular for that study, using the default parameters, the STE method achieved a sensitivity of 38.1% and a specificity of 100%, the SLL method achieved sensitivity of 27.6% and a specificity of 92%, the HIL detector achieved sensitivity of 21.1% and a specificity of 90% and the MNI detector achieved sensitivity of 91% and a specificity of 91%. Thus, in concordance with our own results, highest sensitivity and specificity were obtained with the MNI detector presented higher sensitivity and the STE detector presented higher specificity respectively. Altogether, these results demonstrate that RIPPLELAB succeeds in reducing the analysis complexity for the HFO reviewer, and that this application properly identifies putative HFOs. All the above indicates that RIPPLELAB is a valuable tool for HFO analysis. # Discussion HFO detection and analysis is a difficult task that requires the use of different tools such as detection algorithms, diverse signal processing techniques and specific visualization options. Even though several computational tools exist for the analysis of neural data, none of them includes the appropriate graphical environment nor the implementation of any detection method for HFO analysis, for which our software tool is particularly oriented. For this reason RIPPLELAB becomes a unique tool that consolidates a variety of current analytic methodologies for the analysis of these type of events. RIPPLELAB is a free-of-charge and open source software tool that includes both manual and automatic detection of HFO events. It can handle different types of electrophysiological data such as invasive and scalp EEGs through several file formats. The automatic detection incorporates four already-published methods. Furthermore, RIPPLELAB provides a GUI to facilitate the visual validation of detected events. Especially noteworthy are the possibility to automatically analyze large files, and the possibility to save all the analyses. The resulting files can be further reviewed and easily shared by different groups to be compared conveniently. The RIPPLELAB code was developed in a modular manner, making possible the integration of new methods for HFO detection and the continuous development of new characteristics. The RIPPLELAB capabilities were tested through the analysis of simulated and real signals from iEEG recordings of epileptic patients. These results were consistent with previous studies where the detection algorithms were studied and compared. This fact attest to the reliable operation of RIPPLELAB as a tool for HFO analysis. RIPPLELAB was developed for users of different technical backgrounds, so it provides access to powerful methods for HFO detection without the necessity for detailed knowledge of methodological aspects. Due to its characteristics, RIPPLELAB could serve as a standard platform for testing and comparing new or existing HFO detection and analysis methodologies. Because of the relevance of HFOs in epilepsy, we think that this tool will be particularly useful in clinical and research contexts associated with this pathology, and we hope that this tool could promote and simplify the collaboration and exchange of information between centers working in the field of HFOs. We would like to thank Vincent Navarro (Centre de Recherche de L’Institut du Cerveau et de la Moelle Epinière, Paris, France) for valuable suggestions on the interface design, Samuel Ahn, (Department of Neurology, David Geffen School of Medicine at UCLA, Los Angeles, USA) and Gerriet Janssen for helpful comments on the manuscript. This research was partially supported by COLCIENCIAS (Grant 567). [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MN MV MLQ. Performed the experiments: MN CAR. Analyzed the data: MN CAR MLQ MV. Contributed reagents/materials/analysis tools: MN MV. Wrote the paper: MN CAR MLQ MV.

# Introduction Elevated plasma triglyceride levels have well-established links with chronic diseases such as obesity, insulin resistance and cardiovascular disease. There is growing interest in the function of triglycerides during the postprandial state – in large part because postprandial hypertriglycemia is a risk factor for cardiovascular disease. One mechanism involves elevated triglycerides after a meal recruiting monocytes and inflammatory signaling molecules that eventually lead to atherosclerosis. However, there is need for a comprehensive understanding of what regulates postprandial fat disposition. Oral sensation, especially taste perception, plays a primary role in food selection – but also guides the disposition of ingested nutrients. Sweet taste and other orosensations elicit gastric emptying, digestive enzyme secretion, and insulin release. These physiological responses, which are commonly referred to as cephalic phase responses, prepare the gut and other organs for the approaching absorption and distribution of nutrients. Several lines of evidence suggest that orosensation modifies fat disposition. Oral fat stimuli increase plasma triglyceride concentrations in both animal and human studies: In rats, oral exposure to corn oil or sweet taste leads to a more prolonged elevation of plasma triglycerides relative to oral water or no taste exposure. In humans, oral fat elicits a rise of plasma triglycerides at two different time points; a small spike at ∼1 h after fat loading that is derived from intracellular lipids in enterocytes, followed at ∼4 h by a prolonged elevation of triglycerides. Tasting and expectorating is sufficient to augment the rise in postprandial triglycerides by influencing both the production of chylomicrons and the metabolism of very low density lipoproteins. The basic findings that oral stimuli influence fat trafficking have been replicated and extended, but there has been little attention to whether the chemical or hedonic properties of taste are responsible. Preference (i.e., liking) is an important aspect of taste as well as quality (i.e., sweet, bitter, salty, fatty, etc.). The purpose of the present study was to examine how preferred and nonpreferred tastes influence the disposition of fat. To this end, we infused fat directly into the stomach of rats with implanted intragastric and intraoral catheters. Orosensation was manipulated by infusing preferred or non- preferred taste compounds into the oral cavity. The disposition of infused fat was observed by evaluating blood triglyceride and fatty acid concentrations, and the distribution of radioactive ¹⁴C-fat mixed with the gastric load. # Materials and Methods ## Animals & Maintenance Male Sprague–Dawley rats (weighing 351–375 g; Charles River Laboratories, Raleigh, NC) were housed individually in stainless steel cages at 22°C on a 12∶12-h light-dark cycle (lights on at 06∶00). The rats had free access to AIN-76A diet and deionized water, unless otherwise mentioned. The experiment protocol was approved by the Monell Chemical Senses Center Institutional Animal Care and Use Committee \[protocol no. 1149\]. ## Materials As taste stimuli, we used a “sweet solution” consisting of a mixture of 0.125% saccharin and 3% glucose (both Sigma-Aldrich, St Louis, MO) or a “bitter solution” consisting of 0.15% (0.0038 M) quinine hydrochloride (Sigma-Aldrich). The saccharin-glucose mixture is avidly ingested by rats ; the 0.15% quinine is strongly disliked–tasting it elicits negative hedonic responses (i.e., gapes and chin rubs) and it is almost completely avoided in two-bottle preference tests. As an intragastric fat load, 20% intralipid was purchased from Sigma-Aldrich (Cat. No. I-141). Radioactive ¹⁴C-triolein was purchased from American Radiolabeled Chemicals Inc (St Louis, MO) and stored at −20°C until use. The following enzymatic colorimetric kits or ELISA kits were used for the assay of blood components; triglycerides, ketones, glycerol and glucose from Cayman Chemical Co. (Ann Arbor, MI); non-esterified fatty acid from Wako Diagnostics (Richmond, VA); insulin from Alpco Diagnostics (Windham, NH); total GIP, total GLP-1 and leptin from Millipore (Billerica, MA); peptide YY and cholecystokinin from Phoenix Pharmaceuticals (Belmont, CA). ## Surgery At least 5 days after arrival, rats were surgically implanted with an intragastric catheter and an intraoral cannula. The rats were anesthetized with an intraperitoneal injection of 1 ml/kg of the following mixture: ketamine (4.28 mg/ml; Ketaset, Fort Dodge Animal Health, Fort Dodge, IA), xylazine (0.86 mg/ml; AnaSed, Lloyd Laboratories, Shenandoah, IA), and acepromazine (0.14 mg/ml; Aceproject, Butler, Bublin, OH). For the intragastric surgery, a midline incision was made, the stomach was gently retracted, and a Silastic catheter (0.64-mm ID, 1.19-mm OD) was inserted ∼1 cm through a hole poked with an 18-gauge needle through the glandular portion of the stomach. The catheter was fixed to the gastric wall with 2–0 silk suture. The distal end of the catheter was passed under the skin and exteriorized at the back of the neck. It was glued to a 1-cm square piece of Marlex mesh that was mounted under the skin to anchor it, and the exteriorized portion was sheathed in Tygon tubing to protect it from being bitten. The intraoral cannula consisted of polyethylene-90 tubing (Warner Instruments, Hamden, CT) with one end flared and fixed with a small Teflon disc (6-mm diameter, 0.8-mm thickness). The cannula was inserted into the cheek immediately lateral to the first molar. The Teflon disk was placed so as to rest against the inside of the cheek, and the other end of the cannula was exteriorized at the same position as the gastric catheter and fixed there. Shortly after surgery, and again on the following day, the rats were treated with antibiotics (Triple Antibiotic Ointment, Medique, Fort Myers,FL) to prevent infections and with buprenorphine hydrochloride (Buprenex, Reckitt Benckiser Pharmaceuticals Inc., Richmond, VA) to alleviate discomfort. The patency of the intragastric catheter and intraoral cannula was checked every 2 or 3 days by flushing saline; any rat with a blocked or broken catheter or cannula was excluded from the experiments. After at least 7 days to recover from surgery, rats were given three training sessions (one a day) in order to habituate them to the test procedures. To do this, two intraoral infusions, one of 0.5 ml water and one of 0.5 ml sweet solution, were introduced into the oral cavity in a randomized order, with a 5-min interval between them. These training sessions were conducted between 09∶00 and 12∶00 (light period). ## Test Procedure ### Experiment 1 Before the test, some rats were subjected to procedures designed to induce a conditioned taste aversion to the sweet solution (*n* = 10). To do this, 0.5 ml of sweet solution was infused intraorally and immediately followed by an intraperitoneal injection of 4 ml/kg•BW of LiCl (20 mg/ml; Sigma-Aldrich) as the malaise-inducing agent (conditioned group). The same volume of isotonic saline was injected into rats of the unconditioned group (*n* = 9). This injection procedure was repeated after 3-days so that each rat received two taste aversion conditioning trials. The test was started 3 days later. All rats received two tests: one with the sweet solution and one with water presented orally as a control. The order of these tests was randomized (crossover design) and there was a 1-week interval between them. On each test day, the rats were deprived of food beginning 1 h before testing began until the end of the test session. The rats were infused with 5 ml of 20% intralipid through the gastric catheter at a rate of 1 ml/min using a Sage syringe pump (model 351; Orion Research Inc., Cambridge, MA). Immediately after the infusion, the rats were infused with 0.3 ml water or the sweet solution through the intraoral cannula. At 15 min before (−15 min) and then at 30, 120 and 240 min after the gastric infusion, blood was collected from the tip of the tail of awake rats into heparinized capillary tubes (Fisher Scientific, Pittsburgh, PA). After the ∼140 µl sample was withdrawn, one end of the capillary tube was sealed with Critoseal (McCormick Scientific, St. Louis, MO). Within no more than 5 min, the whole blood was centrifuged for 2 min (IEC MB microhematocrit centrifuge; International Equipment Co., Needham Heights, MA), and plasma collected. The plasma samples were used for the assays on the same day as they were prepared. To verify conditioning had occurred successfully, at the end of the experiment two-bottle choice tests were conducted. To do this, the rats were first deprived of food and water for 5 h, and then given two drinking bottles, with one containing water and one sweet solution for 1 h. Intakes were measured by weighing the bottles (±0.1 g) before and after the presentation. During this test, the unconditioned rats drank 12.1±2.7 ml sweet solution and 1.9±0.7 ml water (87% preference); the conditioned group drank 0.2±0.1 ml sweet solution and 2.0±0.8 ml water (9% preference). Thus, the conditioning procedure was successful. ### Experiment 2 Exposure to a preferred sweet taste had no effect on blood fat fuels in *Experiment 1* (see Results, below). This appeared at least superficially discrepant with earlier work, In particular, using procedures similar to ours, Ramirez showed that tasting saccharin elevated blood fat concentrations, particularly when the sweet taste had previously been paired with an intragastric fat load. A methodological concern was that in our Experiment 1 rats received saline injections during conditioning procedures. This additional handling might potentially influence the rats’ subsequent responses. We therefore repeated the blood fat analysis test used in *Experiment 1* in 12 naïve rats, except this cohort did not receive any conditioning procedures. The rats received a gastric infusion of 5 ml of 20% intralipid followed immediately by 0.3 ml intraoral water or sweet solution. Blood samples for analysis of triglycerides and fatty acids were collected at −15, 30, 120 and 240 min. ### Experiment 3 Several studies show that sweet taste receptors are present in the intestines and are functional. To evaluate their potential contribution to the fat disposition observed in *Experiment 1* and *2,*, in *Experiment 3,* the taste solution was infused intragastrically in 11 rats. Intralipid was delivered in the same manner as in *Experiments 1* and *2* (i.e., 5 ml of 20% intralipid at 1 ml/min) and then either 0.3 ml water or sweet solution was infused through the intragastric catheter over 20 sec. Blood was collected from the tail at −15, 30, 120 and 240 min. All rats received two tests: one with the sweet solution and one with water. ### Experiment 4 In this experiment, we determined the effect of an unconditioned avoided taste on fat disposition. Bitter quinine hydrochloride solution was used as a taste solution. The procedure was the same as for *Experiment 1* and *2*: Immediately after the intragastric infusion of 5 ml of 20% intralipid, the rats (*n* = 10) were infused with 0.3 ml water or the bitter solution through the intraoral cannula. Blood was collected from the tail and used for the assays. ### Experiment 5 In this experiment, the organ distribution of fat was traced by the recovery of radioactivity from intragastrically infused ¹⁴C-triolein. We assessed tissue radioactivity in the gastrointestinal tract and in several organs at 4 h after fat infusion, the time at which the largest effect of sweet taste was observed in earlier experiments. One hour before the experiment (at 09∶30–10∶00), each rat was moved to a plastic cage (28 cm×45 cm×20.5 cm) with woodchip bedding. It was infused with 1.0 µCi of ¹⁴C-triolein in 5 ml intralipid into the stomach at a rate of 1 ml/min. Immediately after that, it was given 0.3 ml of water (*n* = 8), sweet taste solution (*n* = 8) or bitter taste solution (*n* = 7) through the intraoral cannula over 20 sec. At 4 h after the infusion, it was deeply anesthetized with isoflurane (AErrane; Baxter, Deerfield, IL) and blood was collected by cardiac puncture. The blood was transferred into a 1.5-ml Eppendorf tube and allowed to clot at room temperature for 30 min. Serum was prepared by centrifugation at 3000 g for 15 min at 4°C. The serum was used for the measurement of radioactivity and the assay of blood components. After the cardiac puncture, each rat was dissected and pertinent organs (stomach, small intestine, colon, heart, liver and kidney) and tissues (femoris muscle and epididymal fat) were excised. The stomach, small intestine and colon were opened and their contents were collected by washing their inner walls three times with 3 ml of phosphate-buffered saline (Mediatech Inc., Herndon, VA). The collected gut contents were weighed and homogenized. Other organs and tissues were weighed and homogenized in 10 ml of phosphate-buffered saline. One milliliter aliquots of each homogenate were added to 10 ml scintillation fluid (Scintiverse; Fisher Scientific), and radioactivity was measured using a Packard Instruments beta scintillation counter to determine tissue uptake. Values were expressed as a percentage of the total radioactivity infused. ## Statistical Analysis Differences between rats given different oral treatments were assessed using analyses of variance with factors of Taste (water, sweet and/or bitter) and Time (if measurements were made at more than one time). Differences between the treatments at particular times were assessed using paired t-tests or Fisher’s LSD post hoc tests (when comparisons of more than 3 groups were required). Results are expressed as means ± S.E.M. # Results ## Experiment 1: Hedonically Aversive Taste Decreases Blood Fat Concentrations In *Experiment 1*, we examined whether sweet taste influenced the disposition of intragastrically infused fat. The hedonic value of the taste was manipulated by eliciting a conditioned taste aversion to sweetness in one group of rats. In the unconditioned group (*n = *9), sweet taste had no significant effects on blood triglycerides or NEFA levels relative to the water control condition. In the conditioned group (*n* = 10), on the other hand, sweet taste significantly decreased blood triglycerides \[main effect of Taste, *F* _1,9 = 8.39, *P* = 0.018, Taste×Time interaction; *F* _3,27 = 4.00, *P* = 0.018; \]. In addition, NEFA levels were also decreased by the sweet taste in a similar pattern \[main effect of Taste; *F* _1,9 = 2.37, *P* = 0.158, Taste×Time interaction; *F* _3,27 = 4.00, *P* = 0.018; \]. For both fat fuels, the difference was evident at 120 min postinfusion, but not at earlier or later times. ## Experiment 2: Replication that a Preferred Sweet Taste does not Significantly Influence Blood Fat Concentrations Replicating the results of *Experiment 1*, animals in this experiment also did not display a significant influence of sweet taste on triglyceride concentrations \[main effect of Taste; *F* _1,11 = 2.50, *P* = 0.142, Taste×Time interaction; *F* _3,33 = 1.95, *P* = 0.140; \]. There was a tendency for sweet taste to elevate triglycerides at 240 min postinfusion, but this was nonsignificant even by paired t-test (*P* = 0.058). ## Experiment 3: Gastrointestinal Sweet Taste Infusions do not Influence Blood Fat Concentrations Infusion of the sweeteners into the stomach had no effect on blood triglycerides \[main effect of Taste; *F* _1,10 = 0.07, *P* = 0.801, Taste×Time interaction; *F* _3,30 = 0.24, *P* = 0.870; \]. ## Experiment 4: Bitter Taste Decreases Blood Fat Levels In *Experiment 4*, we determined whether an innately aversive bitter quinine hydrochloride taste solution, influenced fat disposition. Relative to water taste, bitter taste decreased blood triglyceride levels significantly \[main effect of Taste; *F* _1,9 = 7.25, *P* = 0.025, Taste×Time interaction; *F* _3,27 = 3.81, *P = *0.021; \] and tended to decrease NEFA levels \[main effect of Taste; *F* _1,10 = 2.21, *P = *0.171, Taste×Time interaction; *F* _3,30 = 2.39, *P* = 0.091; \] in a similar pattern to the conditioned aversive sweet taste (*Experiment 1*). ## Experiment 5: Taste Influences Fat Disposition by Altering Gastric Emptying In *Experiment 5*, we compared the tissue distribution of ¹⁴C-triolein, and a panel of blood fuels and hormones at 4 h after rats received oral exposure to water, sweet solution, or bitter solution. There were large and significant differences in stomach contents \[*F* _2,20 = 4.11, *P* = 0.032; \]. Rats exposed to the bitter taste solution had significantly more–about twice as much–radioactivity in the stomach than did rats exposed to the sweet taste solution. There were small, albeit significant differences among the three groups in radioactivity in the colon. The distribution of radioactive fat in the other tissues did not differ. In this experiment, there were significant effects of the sweet taste on triglyceride concentrations \[*F* _2,20 = 3.51, *P* = 0.049\]. Sweet taste significantly increased blood triglycerides compared with water (*P* = 0.017, post-hoc test) but not bitter taste (*P* = 0.101). Blood hormone concentrations were unaffected by taste, with the exception that bitter taste decreased blood GLP-1 levels relative to water (*P* = 0.025) and sweet taste (*P* = 0.036; *F* _2,20 = 3.57, *P* = 0.047;). # Discussion We demonstrate here that the hedonic value of a taste can affect the disposition of an intragastric fat load. When accompanied by the taste of water, intragastric fat infusions transiently elevated blood triglycerides and NEFAs, with a peak occurring at about 120 min postinfusion. Identical fat infusions accompanied by hedonically negative tastes–either an innately avoided bitter taste or a sweet taste that had been associated with malaise–increased blood triglycerides and NEFAs significantly less. A hedonically positive sweet taste had more ephemeral effects: Relative to the taste of water, sweet taste increased blood fat levels significantly in one experiment, had a tendency for an effect in this direction in another, and produced no difference in a third. The decrease in blood fat concentrations produced by exposure to unpleasant taste is most likely secondary to altered absorption processes, especially gastric emptying. Rats that tasted quinine after a fat load had markedly more radioactive fat label remaining in the stomach 4 h later than did rats that tasted water or a sweet solution. The action of unpleasant taste to retard gastric emptying is consistent with other studies. For example, Yamamoto et. al. showed that quinine-containing bitter mash stays longer in the stomach of rats than does unadulterated mash, and Wicks et. al. demonstrated that bitter taste delays gastric emptying in humans. This also makes teleological sense: Unpleasant taste normally signifies a food that is toxic. Slowing gastric emptying reduces the rate of absorption of the toxin and thus minimizes its blood concentrations. There are several potential explanations for why sweet taste produced only ephemeral effects on the disposition of a gastric fat load. These include (i) methodological factors, in particular, the times we sampled blood or/and the deprivation condition we posed on rats may be important. Oral stimuli mobilize the endogenous fat stored in enterocytes and release it into the circulation rapidly, ; it is unlikely that we captured this considering the relatively late time points at which we observed effects (i.e., 2 or 4 h after taste exposure). (ii) Physiological factors may influence the appearance of fat in the blood. For example, the increased rate of fat absorption caused by sweet taste might be accompanied by increased tissue uptake of fat (perhaps mediated by insulin), leading to stable blood fat concentrations despite increased turnover. (iii) Hedonic factors may be involved. Sweet and bitter tastes are at opposite ends of a palatability continuum; the “control” water taste may fall closer to the sweet end than bitter end of the continuum, making the sweet-water contrast smaller than the bitter-water contrast. Indeed, water is sometimes considered to be sweet. (iv) It may be that sweet taste has little effects on fat disposition but instead prepares the body to metabolize carbohydrates. As shown by Ramirez, the association of sweet taste with an intragastric fat load may be necessary for eliciting the maximum effects on fat disposition. It will require additional research to assess these possibilities. But whatever the mechanism, it is clear that the preferred sweet taste never decreased blood fat levels, which contrasts with the effects of nonpreferred tastes. Taste receptors are present in the gastrointestinal tract where they can initiate hormonal and neural responses to chemical stimuli. In fact, Janssen et. al. and Glendinning et. al. have demonstrated that intragastric infusion of a bitter taste can delay gastric emptying. In our study, direct intragastric infusion of sweet taste solution did not show any effects on fat disposition although the same infusion given orally was effective. We suspect that the 0.3-ml volume of taste solution we used in this study was too small to activate the gut taste system, while being easily sufficient to evoke oral sensation. Further studies are needed to elucidate the physiological mechanisms involved in the modification of fat disposition by taste. Like many other cephalic phase responses, it probably involves activation of the vagus, which innervates the gastrointestinal tract and exerts a major influence on gastric emptying. It is also possible that unpleasant taste could produce a stress-like response, inhibiting gastric emptying or otherwise reducing gastrointestinal absorption by activating the sympathetic nervous system. An intriguing issue is how taste exposure–lasting only a few seconds–can have effects on blood fat fuels 2 or even 4 h later. One possibility is that taste stimulation activates neural circuitry, probably in the brain (although possibly in the enteric nervous system), that maintains strong but not complete inhibition of gastric emptying or intestinal absorption until the stomach is nearly empty. Alternatively, secondary effects initiated by nervous responses might be involved, such as the modulation of fat trafficking. There is a growing literature that taste influences fat disposition. Our results suggest that unpleasant tastes reduce the gastric emptying of fat, leading to lowered concentrations of triglycerides and NEFAs in the blood. The effects of a pleasant sweet taste were less clear, but this does not detract from the main implication of this paper. It may be possible to manipulate the taste of food to mitigate postprandial hypertriglycemia which, in turn, could alter the risk of cardiovascular disease (see introduction). The authors thank for their technical advice and helpful discussion: Dr. Mark I. Friedman and Dr. Glen J. Golden of the Monell Chemical Senses Center (Philadelphia, PA), Dr. Yoshihisa Katsuragi, Mr. Naoto Kudo and Mr. Koichi Yasunaga of the Kao Corporation (Tokyo, Japan). [^1]: This study was funded by the Kao Corporation, the employer of Katsuyoshi Saitou. There are no patents, products in development or marketed products to declare. This does not alter the authors' adherence to all the PLOS ONE policies on sharing data and materials, as detailed online in the guide for authors. [^2]: Conceived and designed the experiments: KS MT. Performed the experiments: KS JL. Analyzed the data: KS JL. Wrote the paper: KS MT.

# Introduction Insects are well known for diverse systems of chemical communication . They use chemicals for locating potential food sources, detecting predators and recognizing kin. Moreover, one of the most notable uses of volatile compounds in insects involves mate choice. When seeking mates, chemical signals can convey a wide range of information on the identity and quality of prospective mates, and play a key role in species recognition, reproductive isolation and speciation. In Lepidoptera, sexual communication usually involves multiple signals, including visual, tactile and olfactory cues to identify potential partners. Pheromones of butterflies are multicomponent blends of chemicals that convey information about sex, age, mating status, and species). Lepidopteran pheromones are Volatile Organic Compounds (VOCs), the majority of which belong to families of esters, alcohols, ethers, and heterocyclic compounds. Based upon the composition of the blend, pheromones facilitate male mate choice, and/or expelling other males from the mating area. During nuptial flight and/or courtship, butterflies usually release pheromones when they are in close proximity to their partner, unlike some female moth species that release them to attract males from long distances. The composition of pheromone blends is crucial for species recognition. Chemical blends of closely related species often have the same major chemical component, but chemical messages of each species can differ in minor components or in the ratio between them. Male butterflies secrete sex pheromones from modified wing scales, the androconia, or from small structures called "coremata" and "hair pencils", while females produce them in ductless glands in abdominal segments (or of legs or wings). In addition, butterflies produce cuticular lipids with low volatility that can act as male sex pheromones in some cases. Several butterfly studies have described morphology and localization of scent organs, while others have analyzed the blend composition of scent pheromones to evaluate their roles in mate selection and species recognition. We hypothesize that olfactory communication between *Hipparchia* individuals likely occurs, as the presence of androconia on male forewings, courtship sequences with a ventilation phase (*Fanning*), and contact stimulation (*Bowing*) suggest. Nevertheless, pheromones of this butterfly genus have not been analysed before. Here we analysed the volatile compounds of two *Hipparchia* species: *H*. *fagi* (Scopoli, 1763), the woodland grayling, and *H*. *hermione* (Linnaeus, 1764), the rock grayling. These two closely related species are morphologically and genetically similar, and are sympatric and often syntopic, i.e., sharing similar habitats and flight periods. Even though *H*. *fagi* and *H*. *hermione* differ in courtship sequences which surely play important roles in finalizing mating rituals), differences in sex pheromones may strengthen premating reproductive isolation mechanisms and facilitate species recognition. Most previous studies of pheromones have used gas chromatography—mass spectroscopy (GC-MS) as this is the gold standard in analytical chemistry for VOCs analysis. However, separation of blends into individual compounds strongly depends on the choice of the GC column material and on parameters used in the analysis. Thus, if a sample is made of a mixture of different and unknown compounds, this approach may not be able to detect them all, underestimating the original chemical complexity. In place of GC-MS analyses, we decided to use an array of partially selective gas sensors characterized by a broad sensitivity to a variety of compounds. This analysis, however, cannot separate and identify each relevant compound, but it can separate and compare the global pattern of volatile compounds between species. The broad selectivity of our electronic sensors is comparable to that of animal olfactory receptors, where a limited number (i.e., about 300 of receptors in humans) are able to detect millions of unique odors, i.e., combinatorial selectivity. For this reason, these sensors are known as electronic noses (ENs). They transform VOCs into a pattern of sensor signals, where the identity of each volatile compound remains unknown, but the sensor signals can be compared to identify similar or dissimilar patterns of VOCs between species. The main focus of our study was: i) to assess the production of volatile chemical patterns by males and females of both species, *Hipparchia fagi* and *H*. *hermione*, and ii) to compare the VOCs patterns in order to evaluate whether there were differences between sexes and species that might be involved in premating reproductive isolation mechanisms. We argue that analysis of the olfactory cues used by *H*. *fagi* and *H*. *hermione* males will help us better understand courtship interactions and of the importance of pheromones in mate recognition in butterflies. # Materials and methods No specific permissions were required for the study locations because these are not protected areas, and, therefore, the field studies did not involve endangered or protected species. ## Butterfly samples and study area *Hipparchia fagi* and *H*. *hermione* used for chemical analyses were from Vallemare (Rieti, Italy, WGS84: 42.4836°N—13.1148°E), where the species are sympatric, their habitats overlap and adults fly together in summer. These species show very similar wing patterns, and identifying species can be difficult when using external features only, but the genital morphology provides good diagnostic characters for species identification. Therefore, we first classified individuals based on their external features and confirmed species by examination of their genitalia. In the early stages of our study, we used reared individuals of *H*. *fagi* to finalize the procedures with the electronic nose (see section "Measurement protocol for butterflies"). When the protocol was optimized, we used individuals caught in the wild and performed laboratory experiments in the study area. These populations have been used in previous behavioural studies. ## Electronic nose The VOCs of butterflies were measured with a gas sensor array (Electronic Nose) developed at the University of Rome Tor Vergata. The Electronic Nose (EN) was an ensemble of seven cross-selective quartz microbalance (QMB) gas sensors coated by seven different metalloporphyrins, based on 5,10,15,20-tetrakis-(4-butyloxyphenyl) porphyrins (buti-TPP) and characterized by the following metal ions: (Cu buti-TPP (sensor 1), Zn buti-TPP (sensor 2), Mn buti-TPP, (sensor 3), Fe buti-TPP (sensor 4), Sn buti-TPP (sensor 5), Ru buti- TPP (sensor 6) and Cr buti-TPP (sensor 7). Porphyrins are a versatile molecular framework to develop arrays of cross- selective sensors. The synthetic chemistry of porphyrins is well developed, since most of the periodic table elements have been incorporated into porphyrin rings, and peripheral porphyrin ring positions have been joined with several different functional groups. Porphyrins can interact with airborne molecules via different interaction mechanisms including coordination, van der Waals forces, hydrogen bonds, π- π and cation- π interactions. Furthermore, in the solid state the aggregation motif could also introduce additional sensing properties. In spite of the large spectrum of mechanisms of interaction, these sensors can capture a large variety of volatile compounds, including alkanes, aromatics, amines, aldehydes, alcohols, and organic acids. QMBs are piezoelectric resonators, and a change of the mass (Δm) on the quartz surface produces a change in the frequency (Δf) of the electric output signal of the oscillator circuit connected to the resonator. The quantities Δm and Δf are linearly proportional in the low-perturbation regime. These QMBs have a fundamental frequency of 20 MHz, corresponding to a mass resolution of few nanograms. The sensor signal is the frequency of the output voltage. The sensitivity of porphyrins coated QMB is rather variable, and it depends on the chemical characteristics of the volatile compound. The limit of detection is of the order of 100 ppb for diverse compounds such ethanol and ethyl acetate. The interactions between volatile compounds and metalloporphyrins are reversible. In each measurement, the sensors of the EN were exposed for 2–3 mins to the mixture of volatile compounds to analyse, and then “cleaned” for 10–15 mins with a reference atmosphere, i.e., filtered ambient air. ## Measurement protocol for butterflies In 2007, we tested the EN on *H*. *fagi* only to assess its performance with butterfly VOCs and ascertain optimal environmental conditions for future experiments (see details of the pilot study in Supporting Information, and Figs). Both butterfly species (65 *H*. *fagi* and 63 *H*. *hermione*) were analyzed during three consecutive reproductive seasons: 14–20 July 2008, 13July-8August 2009; 13–26 July 2010. In 2008, we measured VOC emissions of males only, whereas in the following seasons, 2009 and 2010, we measured emissions of individuals of both sexes. Daily butterfly collections were carried out early in the morning, and all butterflies were kept in flight cages (50×50×50 cm) with food resources (over- ripe fruits and water), waiting to be tested. After 1–7 hours, each butterfly was transferred to a 314 ml glass jars for VOCs measurement. To prevent contamination of VOCs, each individual was manipulated as little as possible, using rubber gloves and tweezers, and glass jars were boiled in distilled water and dried before each experimental session. All measurements were taken between 09.00 a.m. and 7.00 p.m., in “not perturbed” conditions, within a temperature range of 19–30°C, with calm and sunny conditions and ambient photoperiod regime. Once each butterfly was in a sealed glass jar, the extraction of VOCs started after 15–20 min, the time needed for a headspace composition to form and sent to the EN. Each butterfly was measured 3 to 5 times and, while all values were used in the initial pilot study, only individual means were used for statistical analyses. The headspace composition of an empty sealed glass jar was used as a control (reference air). ## Statistical analyses The sensor response to a sample was calculated as the difference between the frequency of the sensor output voltage of the reference air and that of the air after sample exposure. The responses of the sensors were organized into a vector of 7 sensor-dimensional space. The group of responses to different samples were then ordered into matrices, whose rows represented the samples and columns the seven sensors. Signal differences between sexes and species and Principal Component Analysis (PCA) scores were evaluated with a parametric Kruskal-Wallis test, followed by Bonferroni correction for multiple comparisons. PCA and Partial Least Squares Discriminant Analysis (PLS-DA) were used to explore and classify the Electronic Nose data, respectively. The PLS-DA algorithm offers a simple and meaningful method to perform discriminant analysis. In PLS-DA, the class membership is represented as a vector, where the number of elements corresponds to the number of classes, and the class membership of data is expressed setting the corresponding vector element to one and all the others to zero. PLS-DA calculates a set of novel variables, called latent variables, to maximize the covariance with the class membership vectors. Like PCA, the latent variables can be plotted to provide a visual representation of the variation of the data. The performance of PLS-DA classifiers depends on the number of latent variables involved in the model. As PLS-DA classifiers tend to overfit the data, the choice of latent variables requires a cross-validation procedure where the prediction of the model is calculated on an independent set of data. We therefore used a Leave-one-out Cross Validation (LOOCV) procedure, where given n samples, the PLS-DA algorithm was trained n times on all data, except for one sample, and the statistics computed for that left-out sample. The average test-error rate over n trials was the estimated error rate. We used the cross-validated PLS-DA that is equivalent to a multivariate ANOVA. The statistical significance of the PLS-DA classification models was evaluated with a permutation test, where the membership class was randomly attributed and a cross-validated PLS-DA model calculated each time. For each permutation, a cross-validated PLS-DA model was calculated, and the Receiver Operating Characteristic (ROC) curve and the area under the ROC curve (AUROC) were evaluated as estimators of the statistical properties of the classification model. The ROC curve is a typical tool to describe the performance of a binary classifier, and is made plotting the true positive rate versus the false positive rate with the threshold between the two group changes. The area under the curve (AUROC) yields the probability that a positive outcome is classified as negative. PCA and PLS-DA were applied to the auto-scaled data matrix (and Tables), and the responses of each sensor were normalized to a zero mean and unitary variance. In some cases, the data were also linearly normalized to mitigate the dependence of the sensor signals on the VOCs concentration. In linear normalization, the signal of each sensor is divided by the sum of the others. In case of an array of N linear sensors the response of the i-th sensor (Δ *fi*) to the i-th compound at concentration c (*cj*) is given by: $$\Delta f_{i}^{*} = \Delta f_{i}/\sum_{j}\Delta f_{i}.$$ All data analyses were performed in Matlab R2011 (Mathworks, Natick, MA, USA). # Results Sensor 6 (Ru buti-TPP) detected significant differences between males and females of *H*. *hermione*, while sensor 2 (Zn buti-TPP) between males and females of *H*. *fagi*, sensor 5 (Sn buti-TPP) and sensor 4 (Fe buti-TPP) between females of *H*. *fagi* and *H*. *hermione*. Details about the distribution of sensors responses are given in Supporting information (and Figs). Since EN ability in detecting scents may exceed that of individual sensors, we assessed the differences between groups also on the whole sensor array data. ## Intraspecific comparisons: Males *vs*. females Both PCA and PLS-DA on sensor array data did not show any difference between sexes in *H*. *fagi*. LOOCV, optimized with four latent variables, achieved only 68% of classification accuracy, showing that the discrimination between males and females in *H*. *fagi* was still rather poor, although sensor 2 (Zn buti- TPP) had a p-value close to 0.01. Partial separation between the sexes in *H*. *hermione* was observed in the PCA. PCA explained only 4% of the total variance due to PC2. While PLS-DA quantified the discrimination between sexes. LOOCV produced a model with 3 latent variables. Due to our limited sample sizes, we could not carry out a proper training and test validation of the classifiers. Thus, the reliability of the PLS-DA model was tested with a permutation test with the ROC curves generated in each iteration of the permutation test. The comparison of the ROC curves indicated that classifier performed better than random class membership most of the time. This behaviour was evident when the frequency distribution of the area under the ROC (AUROC) was analysed. The histogram was fitted with a normal distribution that quantified the probability of assigning a random class membership was more or less accurate than the separation of data into male and female groups. Specifically, in this species, the classification into males and females in *H*. *hermione* with 85% accuracy was achieved at a confidence level above 2σ. ## Interspecific comparison: *H*. *hermione vs*. *H*. *fagi* Multivariate analysis of the sensor responses to the did not show any appreciable difference between *H*. *fagi* and *H*. *hermione* male volatile compounds. The application of PLS-DA to the whole data set did not improve the results: LOOCV resulted in 51% of accuracy similar to random choice. The plot of the first two latent variables of PLS-DA showed an almost perfect overlap between the two groups. Separation between females of the two species is shown in. Quantitative description of the discrimination capabilities of the array was obtained by PLS- DA. LOOCV produced a model with 4 latent variables and a cross-validated predicted accuracy of 85%. The collection of ROCs and the related AUROC distribution are shown in. Under the hypothesis of a normal distribution of the AUROC, the predicted accuracy (85%) was obtained with a confidence interval of 3σ. # Discussion The main purpose of this study was to assess the ability of EN to detect volatile organic compounds patterns produced by *Hipparchia* species and differences between sexes and/or species.We were able to convert *H*. *fagi* and *H*. *hermione* VOCs into seven gas sensor signals and EN revealed significant differences between *H*. *fagi* and *H*. *hermione* females, but not between males. When single species were considered, EN was able to distinguish between VOCs patterns released by the two sexes in *H*. *hermione*, but not in *H*. *fagi*. ## Cues used by males in individual recognition Visual stimuli have a main role in prompting males to flight towards potential mates in several *Hipparchia* species, including *H*. *fagi* and *H*. *hermione*. They can, in fact, use visual stimuli as initial approach for species recognition and mate choice. However, the flight approach is not always strictly selective. Males have a relatively poor ability in identifying conspecific partners from far away, but their ability improves as soon as the partner gets closer. Although *H*. *fagi* and *H*. *hermione* have a similar wing colour patterns, after the initial pursuit (*Flight* phase) interspecific interactions rarely continue. We hypothesize that species recognition occurs at the early stages of sexual interactions. Similarly, chemical and tactile stimuli most likely play an important role in courtship behaviour of these butterflies. Our results show that VOCs patterns from females of *H*. *fagi* and *H*. *hermione* are significantly different as captured by the sensors 4 and 5 of the EN. Sensor 4 showed the largest response when exposed to VOCs released by *H*. *hermione*, and sensor 5 when it was exposed to VOCs from *H*. *fagi*. These responses suggest that the differences between species were not due to of concentration differences, but rather to a change in VOCs profile. Females producing different VOCs suggest that males may recognize conspecific partners on the basis of female smell at an early phase of the sexual interaction, i.e., nuptial flight. This is possible for *H*. *fagi*, as previous behavioral studies have shown that males identify conspecific females at an early stage of the sexual interaction. Indeed, if male *H*. *fagi* accidentally chases a female *H*. *hermione*, he usually abandons the pursuit as soon as they are close to one other. On the contrary, VOCs are probably not the only cue used by *H*. *hermione* males to discriminate between conspecifics and *H*. *fagi* females. In fact, *H*. *hermione* males can undertake the nuptial flight with *H*. *fagi* females, and usually the courtship only ends because the female refuses the copulation attempt. ## Cues used by females in individual recognition This study is consistent with previous observations on sexual behaviour and suggests that female *H*. *fagi* and *H*. *hermione* do not recognize volatile olfactory cues for partner recognition, and instead use their olfactory system when they are in close contact with their partner during courtship. The mating behaviour of contact stimulation (*Bowing*), during which the male captures the female antennae between his forewings and ensures their physical contact with his androconia, develops differently in the two species. *H*. *fagi* females need longer *Bowing* and only a few repetitions of *Bowing* to accept to mate, while *H*. *hermione* females need more intense solicitation by short, highly repeated *Bowing*. These behavioural differences between species suggest that females place greater emphasis on tactile rather than chemical stimuli when accepting mates, and may therefore explain the lack of male VOCs differences in our study. # Conclusions The first use of an Electronic Nose to investigate VOCs in *Hipparchia* butterflies showed that courtship signaling involves an integrated sequence of signals, with at least two modes: sight and scent in addition to behavioral patterns that can involve contact-chemoreception signals, i.e., during *Bowing*. Volatile compounds are produced during sexual interactions of *H*. *fagi* and *H*. *hermione* suggesting an potential role of VOCs in partner recognition during the early stages of the interaction. To better understand the signals involved in species recognition, more manipulative behavioural experiments are needed, and studies on different species would also be useful for further comparisons. The Electronic Nose proved to be sensitive enough to reveal significant differences between the patterns of volatile compounds of different sexes and species. This device does not provide direct information on the identity of the measured VOCs, rather it captures the global differences of the patterns. Further analyses are needed to identify the relevant compounds responsible for the chemical differences between species. We suggest that Electronic Noses provide unique opportunities to measure VOCs released by animal species in their natural environment because they can be operated in the field. # Supporting information We are very grateful to Dr. Richard Hewett (University of Salford, UK) for assistance in English editing of the original submission and Dr. Anna Fabiani (Rome, Italy) for her helpful comments and language review of the final version. [^1]: The authors have declared that no competing interests exist.

# Introduction Glycosylated proteins and other glycoconjugates are major components of cells, defining and modulating several key physiological processes in normal tissues. Many of the effects of the glycoconjugates are mediated by glycan–lectin interactions, that are involved in involved in many normal and pathological processes from cell recognition and communication to pathogen invasion and tumour metastasis. The awareness of the glycan component of glycoconjugates carries biological information has motivated numerous studies of glycans, and significant progress has been made in the past years related to defining the structures and functions of glycans in biological systems. However, the progress within this field is challenged by the complexity and structural variation found in glycoconjugates combined with the high specificity, low affinity, and often multivalent nature of the interactions. There is therefore a need for new experimental techniques to study glycan related biological and medical problems. Optical tweezers (OT) is one of several single-molecule manipulation techniques that have evolved rapidly over the last decades and that are finding an increasing number of applications within life-sciences. This technique is based on the generation of an optical trap, through focusing a laser to a diffraction- limited spot with a high numerical aperture microscope objective. A dielectric particle near the focus will experience a restoring force that keeps the particle near the focus, as further outlined in several reviews. For small displacements of the particle, the optical trap acts as a linear spring. The calibration underlying the conversion from the detected displacement of the particle to the force driving this displacement is straightforward for silica and polystyrene beads, and these are therefore widely used in experiments aiming at determining interaction forces, as handles for the biomolecules of interest. OT have been applied for high resolution studies of forces required to unbind biomolecules, studies of structural dynamics of biomacromolecules studies of individual molecular motors, as well as studies of mechanical properties of biological tissues and cells. Despite these well documented capabilities of the OT, atomic force microscopy (AFM), characterized by a force range going from 5 to 1000 pN, is more frequently used to determine molecular interaction forces. However, the low strength of carbohydrate–protein and carbohydrate–carbohydrate interactions makes OT, capable of determining forces in the range going from 0.5 to 100 pN, an ideal probe for quantification of these interactions. The aim of the present paper is to characterize the unbinding properties of mucins carrying the cancer antigens ST and STn with MGL, including also identifying the carbohydrate moieties mechanistic in these interactions. Mucins are glycoproteins that contain a range of *N-*acetylgalactosamine (GalNAc)-Ser/Thr *O*-linked glycans, and these glycans comprise more than 50 wt% of the molecule. Cell-surface-bound and secreted mucins from epithelial and other mucin-producing cells constitute an important part of the glycome surrounding these cells. Whereas secreted mucins function as a protective layer over the epithelium, the glycans of cell-surface-bound mucins control antigenicity as well as interactions with the environment and bind to mammalian lectins. In this study we focus on the transmembrane, human mucin MUC1. MUC1 contains a variable number of tandem repeats (TRs) (25–125) of 20 amino acid residues with each repeat having five potential sites for O-glycosylation. Certain changes in glycosylation are associated with development of cancer. Cancer cells often express truncated glycan structures including the carbohydrate antigens Tn (GalNAcα1-*O*-Ser/Thr), the sialylated Tn structure (STn; NeuAcα2-6GalNAcα1-*O*-Ser/Thr), and T (Galβ1- 3GalNAcα1-*O*-Ser/Thr). The Tn structure has been described as a tumor-associated antigen in various human tumor entities. It is regarded as a useful biomarker because it is expressed early in transformed cells, both in human and in animal carcinogenesis. Furthermore, a direct correlation has been shown between carcinoma aggressiveness and the density of this antigen. The presence of STn in human tumors can be due to the up-regulation of ST6GalNAc-I transferase or the inactivation of the COSMC chaperone. In addition to these short cancer- associated antigens, MUC1 expressed by breast carcinoma cells also carries the mono- and disialyl core 1 structure (ST, NeuAcα2-3Galβ1–3\[NeuAcα2–6\]+/–GalNAcα1-*O*-Ser/Thr) found widely in normal cells. Cancer-associated aberrant glycosylation can represent altered capacities for interaction with the microenviroment. The interaction between tumor-associated antigens and specialized antigen- presenting cells is critical for the induction of a specific anti-tumour immune response. Glycopeptides corresponding to three tandem repeats of MUC1, glycosylated with 9 or 15 molecules of GalNAc, have been shown to specifically bind to and be internalized by immature monocyte-derived dendritic cells (DCs). The macrophage galactose-type lectin (MGL) expressed by monocytes is a well- studied C-type lectin binding to MUC1. Human MGL is a 40 kDa transmembrane glycoprotein consisting of a 39 amino acid (aa) cytoplasmic region, a 21 aa transmembrane segment and a 256 aa extracellular domain (ECD) with a carbohydrate recognition domain (CRD) and a neck region. It is reported that MGL binds to the Tn antigen present on MUC1, and NMR data indicate that MGL also binds the STn antigen. Furthermore, based on NMR data it has been suggested that the affinity of the STn antigen to MGL is mainly mediated by the GalNAc moiety. The presence of tumor-associated macrophages (TAMs) in the microenvironment of malignant tumors of human carcinomas has been correlated with an adverse prognosis of the patients. A subpopulation of TAMs, the M2 macrophages, appear to be causally involved in the tumor progression, and monocytes can be differentiated into M2 macrophages by addition of the conditioned tumor cell medium. Interestingly, monocytes stimulated in this way express MGL. The interaction of MGL, expressed by M2 macrophages, with Tn and STn exposed by tumor cells, has been suggested to modulate the TAM phenotype and/or activity, and thus affect the progression of human tumors. The importance of gaining further knowledge of the possible role of the STn structure is supported by the fact that the Tn glycan is mostly intracellular and not frequently on the carcinoma surface. Thus, the binding of MUC1 carrying STn to MGL may be more physiologically relevant than the binding to MUC1 carrying Tn. Tumour associated STn is associated with poor prognosis and resistance to chemotherapy in breast carcinomas, inhibition of DC maturation, DC apoptosis and inhibition of NK activity, and the binding of MUC1(STn) to MGL may be in part responsible for some of the characteristics of STn expressing tumours. In this paper we quantify and compare the strength of the molecular interaction between the two cancer associated antigens MUC1(Tn) and MUC1(STn) and the lectin MGL by use of OT. Additionally, we apply the OT based experimental strategy to explore the interaction between a short synthetic polymer carrying GalNAc and MGL. These additional experiments provide information relevant for identifying the chemical groups essential for the observed MUC1—MGL interactions. # Materials and methods ## Samples MUC1-IgG Tn, STn and ST samples were produced using wt and mutant CHO cell expression systems as previously described. The molecules contained the extracellular part of human MUC1, including 16 MUC1 tandem repeats. The molar mass of the core polypeptide chain of the MUC1 molecules was 46 kDa. They also carried an IgG domain with MW of about 50 kDa. Each tandem repeat had 5 glycosylation sites, and their average glycosylations, as determined by mass spectroscopy, were: MUC1(Tn) = 3.4, MUC1(STn) = 3.8, and MUC1(ST) = 4.6. The total molecular weights of the glycoprotein constructs were found to be MUC1-IgG-(Tn) = 107 kDa, MUC1-IgG-(STn) = 127 kDa and MUC1-IgG-(ST) = 147 kDa. The glycan decorations on these mucins are summarized in. α-GalNAc- PEG₃-NH₂, referred to as GalNAc-PEG in the following, was obtained from Sussex Research Laboratories Inc. Macrophage galactose/N-acetylgalactosamine (GalNAc) specific lectin (MGL), also known as CLEC10A, was obtained from R&D Systems R&D Systems Inc.Minneapolis, USA. ## Covalent attachment of molecules to polystyrene beads MUC1 molecules, MGL and GalNAc-PEG were immobilized to colloidal polystyrene beads (Spherotech, Lake Forest, Illinois). The immobilization procedure was based on the introduction of a covalent bond between amino groups on the polystyrene beads and carboxyl groups on the molecule to be immobilized, or vice versa, using the water soluble carbodiimide EDC (1-(3-dimethylaminopropyl) -3-ethylcarbodiimide hydrochloride) as a catalyst of the bond formation between the carboxylic acid and amine groups. The immobilization protocol was previously used for immobilization of proteins including mucins onto amine functionalized glass surfaces. When investigating the interaction between MGL and MUC1(Tn), MUC1(STn) or MUC1(ST), the MGL lectins were dissolved in 100 μl aqueous boric acid (50 mM, pH 5.8, referred to as conjugation buffer) at a concentration of 0.1 mg/ml. Amine-terminated polystyrene (nominal diameter 3.07 μm) and EDC were added to this solution to final concentrations equal to 0.03% w/v and 2.5 mg/ml, respectively. The MUC1 molecules were dissolved in 100 μl of the conjugation buffer to a concentration equal to 0.2 mg/ml, and amine-terminated polystyrene beads (nominal diameter 2.01 μm) and EDC were added to final concentrations equal to 0.03% w/v and 2.5 mg/ml, respectively. When investigating the interaction between MGL and GalNAc-PEG₃-NH₂, the MGL was immobilized via their carboxylic acid groups onto 2.01 μm amine functionalised polystyrene beads. The concentrations used were equal to 0.1 mg/ml, 0.03% w/v and 2.5 mg/ml for the MGL, polystyrene beads and EDC, respectively. GalNAc- PEG₃-NH₂ were immobilized onto carboxylic acid functionalized polystyrene beads (nominal diameter 3.07 μm), using a concentration of GalNAc-PEG₃-NH₂, EDC and polystyrene beads equal to 0.5 mg/ml, 2.5 mg/ml and 0.03% w/v, and they were dissolved in 100 μl of conjugation buffer. Unreacted reagents were removed from the functionalized beads by centrifugation (10000 rpm, 4 min), and the beads were re-suspended in aqueous 100 mM Hepes buffer pH 7.2 containing 1 mM MnCl₂ and 1 mM CaCl₂. The bead functionalization procedures were carried out at room temperature (20°C). Prior to OT experiments, 3 μl of each of the two functionalized bead solutions that was intended studied was diluted in 50 μl of the Hepes buffer and transferred to the sample chamber of the OT. The calcium dependence of the interactions were investigated by diluting the functionalized bead solutions in either 100 mM Hepes buffer pH 7.2 or in 100 mM Hepes buffer pH 7.2 containing 1 mM MnCl₂, 1 mM CaCl₂ and 5 mM EDTA. ## Optical tweezers Optical tweezer measurements were carried out using the dual beam instrument JPK Nanotracker (JPK Instruments, Berlin, Germany). The sample chambers were made from a circular glass slide, two pieces of double-sided tape and a quadratic coverslip. The circular glass slides used as floors in the sample chambers were pre-coated with bovine serum albumin (BSA, Sigma) (1 mg/ml, 20 min incubation) to reduce adhesion of the functionalized polystyrene beads. The solution containing the functionalized beads was introduced in the sample chamber by capillary forces, prior to sealing the sample chamber and mounting it on the sample stage of the OT. Prior to all measurements, one bead of 2.01 μm in diameter and one bead of 3.07 μm in diameter were identified based on their size using the microscope and captured in separate optical traps. The trap stiffness was determined for each trap prior to each experiment from the power spectra obtained by tracking the 3D Brownian motion of the bead. During the experiments the beads are moved in the x-y plane. Based on the detected displacement of the bead relative to the laser focus as well as the trap stiffness and trap sensitivity, the force withheld by the molecular bond prior to rupture is determined. Performing the calibration procedure 13 subsequent times on the same bead revealed a relative uncertainty in the determination of the trap stiffness of 5.8%. The detection of the bead position relative to the laser beam was based on back focal plane interferometry. The OT instrument used has a force resolution of less than 0.1 pN. ## Observation of forced unbinding of MUC1—MGL interactions using optical tweezers The experiments were carried out as outlined in. A MGL-functionalized polystyrene bead was trapped in one of the two optical traps of the dual trap system, and a MUC1 or GalNAc-PEG functionalized bead was trapped in the other. The distance separating the two traps was then reduced until the two polystyrene beads were in contact and pushed each other slightly out of the laser focus, observed as an increase in the force acting on the beads. The beads were left in contact for 0.8 s before increasing the bead separation distance. When increasing the inter-bead distance, intermolecular interactions between the mucins or mucin analogues and MGL, if formed, were broken due to the applied force. Prior to bond rupture, the beads are displaced relative to the center of the optical trap in proportion to the force acting on the beads. Repeated approach—retract cycles were carried out with bead separations in the range 1.5–3 μm. ## Analysis of intermolecular bond rupture events The bond strength and the corresponding force loading rate applied to the bond just prior to rupture were determined for each observed rupture event based on the magnitude and slope of the force jump, respectively. The raw data sampled at 2.1 kHz at the retraction speed of 1 μm/s, were smoothed using a 6 datapoint moving average and analyzed with respect to unbinding events. The occurrence of the force unbinding events were determined based on differentiating the smoothed raw data to make the identification of force jumps more easy to distinguish from signals from the noise. The procedure employed to determine the force loading rate uses a linear approximation of the increase in force just prior to the unbinding event, and is previously reported. The data segments of the force- traces just prior to the unbinding events are selected based on balancing the suppression of effect of noise in the data while conforming to the linear approximation. The experimentally determined energy landscapes of the macromolecular interactions were interpreted based on the theoretical framework outlined in the following. According to the model first proposed by Bell and later elaborated by Evans and coworkers the dissociation rate related to the transition from a bound to an unbound free state is for a molecular pair dependent on the applied force. Key parameters appearing in this model include x_β which is defined as the thermally averaged distance from the bound complex to the transition state projected along the direction of the applied force and k_BT, the thermal energy. Consequently, the rate of dissociation under a constant loading force *f*, *k*_*off**(f)*, exponentially increasing with the force: $$k_{off}(f) = k_{off,0}\exp\left( \frac{x_{\beta}f}{k_{B}T} \right)$$ the probability density *P(f)* for observing a bond rupture between a molecular pair at the force *f* subjected to constant force loading rate *r*_*f* predicted by the Bell-Evans assumption is: $$P(f) = k_{off,0}\exp\left( \frac{x_{\beta}f}{k_{B}T} \right)\exp\left\lbrack {\frac{k_{off,0}k_{B}T}{x_{\beta}r_{f}}\left( {1 - \exp\left( \frac{x_{\beta}f}{k_{B}T} \right)} \right)} \right\rbrack$$ when the applied force along the unbinding pathway exceeds the force *f*_*β* governed by the distance *x*_*β*, i.e., *f*_*β* *= k*_*B**T/x*_*β* an exponential increase in the most likely unbinding force, *f*^\*, is predicted. $$f^{*} = f_{\beta}\ln\left( {r_{f}/r_{f}^{0}} \right)$$ parameter *r*_*f* is the actual force loading rate, and *r*_*f* ^*0* a thermal scale for loading rate, *r*_*f* ^*0* = *f*_*β**/t*_*0* where *t*_*0* is the inverse of the transition rate. Parameters characterizing the interactions between the mucins and lectin were extracted from the data generated by the OT as follows. The set of data of f versus *r_f* for each type of macromolecular pairs were divided into intervals with equal range of Ɗln(*r_f*) for the intervals. The mean value of rf and spread represented by the standard deviation of the set of the data within each interval were estimated and a histogram was estimated. The most probable unbinding force *f*\* within each interval was estimated using a non- linear fit of P(*f*) to histograms centered around a mean force loading rate. Parameter *x_β* was estimated by fitting the linearized version of to the estimated mean *r_f* and *f*\* as outlined above. The uncertainties of *x_β* were estimated based on the uncertainty of the slopes determined in the fitting to the linear version of. Estimates of k_{*off, 0*} were determined from the estimated intercept in the fit of the linear version of from the procedure used to estimate *x_β*. A constrained fit of P(f) keeping the *x_β* parameter constant was used to guide an eventual splitting of the *f*\*\* vs *r_f* data into regions, each conforming more closely to the behavior predicted by, e.g., representing barriers with their particular parameters, than when assuming one barrier. # Results ## Observations of unbinding events for single molecular pairs of MUC1 and MGL The experiments were carried out by trapping a MGL-functionalized polystyrene bead in one of the two optical traps of the OT setup and a MUC1 or GalNAc-PEG functionalized bead in the other. When bringing two polystyrene beads in contact and then increasing the inter bead distance, any intermolecular interactions between the mucins or mucin analogues and MGL, were broken due to the applied force. Frequent force jumps, reflecting the rupture of intermolecular interactions, were observed in the force curves when MGL interacted with either MUC1(Tn) or MUC1(STn), but not when allowing MGL to interact with MUC1(ST). The histogram of the rupture forces for the MUC1(Tn)–MGL or MUC1(STn)–MGL pairs revealed a large spread. Such broad distributions are observed for macromolecular pairs in direct force unbinding assays having a high probability for multiple interactions. The existence of multiple interactions in these experimental series is confirmed by the appearance of the force curves. Typical force curves obtained for MUC1(Tn)–MGL ( and MUC1(STn)–MGL obtained when using the experimental conditions explained in display successive rupture events and/or high rupture forces. Due to the observed inability of MUC1(ST) to interact with MGL, these molecules were in later experimental series used as non-interacting spacer molecules between the interacting MUC1(Tn) and MUC1(STn) molecules on the beads studied using OT. Experimentally, surfaces displaying MUC1(ST) in addition to the MUC1(Tn) or MUC1(STn) were obtained by mixing MUC1(ST) and MUC1(Tn) or MUC(STn) in the conjugation buffer. A mixing ratio of 80% (0.08 mg/ml) MUC1(ST) and 20% (0.02 mg/ml) MUC1(Tn) or MUC1(STn) in the conjugation buffer yielded a significant reduction in the high force tail of the force distribution. For MUC1(Tn) a narrow distribution of rupture forces was observed, with a most probable rupture force located at 12 pN. For MUC1(STn) a broader distribution was observed. Further increasing the fraction of the MUC1(ST) to 90% of the total content of glycoprotein in the solution during the conjugation step did not significantly influence the distribution of interaction forces observed for MUC1(Tn). However, in the case of STn, this reduction resulted in a more narrow distribution of rupture forces and a reduction in the most probable rupture force from 24 to 12 pN. Based on these observations, a concentration equal to 0.01 mg/ml MUC1(Tn or STn) and 0.09 mg/ml MUC1(ST) were identified as the optimal concentrations and were in this study used as basis for assessment of single-molecular pair interactions. The hypothesis that the high rupture forces observed when using the experimental conditions underlying the histogram distributions presented in are due to multiple MUC1-MGL interactions are supported the signatures observed in the force curves obtained when using these experimental conditions. These force curves contain increased probability for successive rupture events compared to what is obtained when decreasing the density of the interacting molecules (obtained at the experimental conditions explained in, respectively). ## Dynamic force spectroscopy of MUC1–MGL interactions The force–distance curves obtained for the MUC1(Tn)—MGL interaction allowed identification of 1268 force jumps. For the MUC1(STn)–MGL interaction 1648 force jumps were collected. For each of these force jumps, the rupture force and the loading rate were determined. The data contained in the dynamic force spectrum obtained for the MUC1(Tn)–MGL interaction were grouped into 7 subgroups along the axis of increasing mean loading rates, from an average loading rate equal to 29 pN/s for the first interval to 134 pN/s for the last interval. The probability density function of unbinding under external force, P(f) was fitted to the distribution of unbinding forces contained in subgroup, using the parameters k_{*off, 0*}, reflecting the lifetime of the interaction, and x_*β* as fitting parameters. The most probable rupture force *f\** was for each subgroup determined based on the peak in the probability function. summarizes the number of observations contained in each subgroup as well as the estimates obtained for k_{*off, 0*}, x_*β*, the average loading rate *r*_*f* and the most probable rupture force *f\**. For the interval characterized by an average loading rate *r*_*f* equal to 29 pN/s, unbinding forces ranging from 5 to 14 pN were observed, with a most probable value equal to 6.8 pN. For this interval, x_β was determined to 0.51±1.1 nm, and k_{*off, 0*} to 2.0 s^-1. The most probable unbinding forces increased with increasing force loading rate and f^\* = 27 pN was determined for the subgroup with average loading rate *r*_*f* = 134 pN/s. For this interval, x_β was estimated to 0.12 nm, and k_{*off, 0*} to 1.9 s^-1. The force jumps observed for the MUC1(STn)—MGL interaction were divided into 6 subgroups. The estimates obtained for the key parameters describing this interaction are presented in. For the lowest loading rate range (mean loading rate 43 pN/s), the most probable rupture force f^\* was equal to 7.1 pN, x_β was determined to 0.31±0.1 nm, and k_{*off, 0*} was determined to be 3.3 s^-1. For the subgroup with the highest mean loading rate, 137 pN/s, the most probable unbinding force was determined to be f^\* = 37 pN, x_β was determined to be 0.09 nm, and k_{*off, 0*} was determined to be 1.8 s^-1. The dynamic force spectra obtained for these two interactions are also similar. The 95% confidence intervals of the fit of *f*\* vs ln (*r*_f) to the linear version of, indicating large overlap of the domains, suggest that parameters of the unbinding barriers at lowest range of *r*_f are similar for the MUC1(Tn)-MGL and MUC1(STn) interactions. In order to investigate the potential Ca²⁺ dependence of the interactions, OT experiments were performed prior to and after adding EDTA to the 100 mM Hepes buffer pH 7.2 containing 1 mM MnCl₂ and 1 mM CaCl₂. Frequent force jumps were observed both prior to and after adding EDTA. The distribution of the rupture forces revealed a similar interaction strength as observed for the MUC1(Tn)–MGL interaction when investigated in the Hepes buffer prios to addition of EDTA. Also in experimental series performed using 100 mM Hepes not containing Ca²⁺, force jumps were observed. ## Observation of interaction between MGL and GalNAc-PEG To further study the role of the sugar residue of MUC1(Tn) involved in the MGL binding, polystyrene beads functionalized with the α-GalNAc- PEG₃-NH₂ were prepared. When bringing these beads in contact with polystyrene beads functionalized with MGL, signatures of rupture of intermolecular bonds were observed upon bead separation. Based on the obtained force–distance curves 35 rupture events were identified. The distribution of rupture forces observed for these interactions are presented in. # Discussion Some of the challenges faced when studying carbohydrate—protein interactions are related to their inherent low strength and multivalency. These challenges have hampered the progress in the emerging field of glycomics, and this field is therefore expected to benefit from the application of new methodologies. Single- molecule manipulation techniques have evolved rapidly over the past decades and are finding an increasing number of applications. However, whereas the number of studies in which AFM is used to quantify intermolecular interactions increased rapidly, OT is still rarely used in such studies. Despite this, the low force range attainable with OT makes this probe a powerful tool for quantifying weak intermolecular interactions. In the current study, OT is used to investigate the interaction between the mucin MUC1 and the lectin MGL. The results obtained reveal that intermolecular bonds form between MGL and MUC1(Tn) or MUC1(STn), but not with MUC1(ST). The observations presented in this paper are thus compatible with a hypothesis where MGL specifically binds to cancer-associated mucins, as previously proposed. When quantifying the strength of single molecular bonds, a low probability of observing an interaction event is essential since this assures a high probability that the rupture events observed reflect the rupture of single molecular bonds. In the current study, good control of the surface density of MUC1(Tn) and MUC1(STn) was obtained by adding the non-interacting MUC1(ST) molecules to the solution used for surface functionalization. When applying this strategy the density of the interacting molecules, and thus the tendency for multiple interactions, was efficiently controlled. An additional benefit of the approach used is the reduction in the possibility for non-specific interactions between areas of the polystyrene beads not displaying MUC1 molecules. The contribution of such non-specific interactions in the data set would otherwise contribute with noise and hamper a correct determination of the interaction force between single molecular pairs of MUC1 and MGL. Following optimization of the immobilization procedure to yield mainly single-molecular pair unbinding events, the most probable rupture force was determined to 12 pN for both MUC1(Tn)—MGL and MUC1(STn)—MGL interactions. The number of reported OT based quantitative studies of carbohydrate-protein interactions is limited. However, the interaction of the *H*. *pylori* adhesin BabA with the antigen Lewis b has been quantified using OT, and the rupture events were observed to be centered at multiples of 12.5 pN. Due to an observed high probability for 25, 50 or 75 pN, with lower probability for the intermediary strengths, the authors concluded that the strength of the BabA-Leb interaction was equal to 25 pN. The parameters x_β and k_{*off, 0*} were determined for both the MUC1(Tn) and the MUC1(STn) interactions with MGL (Tables and). These parameters provide information related to the shape of the energy landscape of the intermolecular interactions as well as the lifetime of the interaction, respectively. Unfortunately, these parameters were not determined in the previous study of BabA–Lewis b interactions. In the current study, very similar lifetimes were determined for the two interactions (Tables). The average lifetimes of the interaction, reflected by the parameters k_{*off, 0*}, were for the inner barriers determined to 0.5 s and 0.3 s for the MUC1(Tn)—MGL and MUC1(STn) systems, respectively. These lifetimes are in the same range as previously determined for other carbohydrate–protein interactions using AFM and OT. As expected it is shorter than the lifetime of antibody-peptide interactions as well as the interaction between cell-surface sulfatase Sulf1 and glycosaminoglycans. Despite the similarity of both the binding strengths and the lifetimes of the MUC1(Tn) and MUC1(STn) interactions with MGL, a slightly lower x_β value was determined for the MUC1(STn)–MGL system (Tables). The values obtained in the current study, being in the range 0.09–0.51 nm, are comparable in size to the values determined in previous single molecule studies on related systems. Previous AFM based quantitative studies of the interaction between porcine submaxillary mucine (PSM) interacting with the lectin SBA, gave x_β values in the range of 0.05–0.12 nm, decreasing with increasing force loading rate. The direct comparison of data obtained by OT and AFM need to take into account the differences in loading rates realized using these techniques. AFM provides data in a higher loading rate range compared to OT. Provided the molecular interaction possesses an inner energy barrier in the energy landscape, the characterization by AFM may yield lower x_β values. The similarity of the rupture force and lifetimes of the MUC1(Tn) and MUC1(STn) interaction with MGL (Tables) is consistent with the main conclusions drawn based on previous NMR data, which indicated a similar binding mode for the Tn and the sialylated Tn antigen when interacting with MGL. The NMR data indicated that the N-acetyl group and the H-2, H-3 and H-4 protons of the GalNAc residue made the major contribution to the interaction. However, the NMR data also indicated a slightly lower affinity of the STn antigen for MGL compared to the Tn antigen. In the current study small variations between the two interactions were also observed: the parameter x_β was slightly lower for the MUC1(STn)–MGL relative to the MUC1(Tn)—MGL interaction. This might indicate a slightly shorter separation distance between the MUC1(STn) and the MGL when bound to each other. However, due to the relatively small variations observed combined with the experimental challenges related to the determination of this parameter clear conclusions should not be drawn based on this difference. The slightly lower affinity of the STn antigen for MGL observed using NMR but not when studying the interaction using OT might be due to the fact that in the NMR study, mucin analogs in which the glycans were linked to a single serine unit were used. It can at this stage not be ruled out that the amino acid portion of the glycoprotein somewhat influences on the properties of the interaction, as recently proposed. The experimental data revealed no Ca²⁺ dependence of the MUC1(Tn)–MGL interaction. Ca²⁺ dependent binding is observed for several other lectin–glycan interactions. The Ca²⁺ dependence is also previously investigated for carbohydrate–carbohydrate interactions, where it is observed for some glycans, whereas in other studies no such dependence, or only a weak dependence, is reported. In a previous study of Ca²⁺-dependent cell adhesion, relatively strong adhesive bonds formed also in calcium-free artificial seawater. The rupture forces were of the same order of magnitude as those obtained for the same molecule in the presence of 10 mM Ca²⁺. However, more detailed analysis of the unbinding process revealed that the lifetime of the bound complex was longer in the presence of Ca²⁺. Thus, based on the AFM force probe observations, the difference in lifetime, and not the magnitude of the unbinding force, was suggested to explain the Ca²⁺ dependence reported for this system using other experimental techniques. In the present study the investigation of the Ca²⁺ dependence of the interaction was not the main scope, and the determination of the lifetime of the interaction could not be performed in a reliable way based on the limited number of force versus distance curves obtained in the present study for the MUC1(Tn)–MGL interaction. However, similar lifetime parameters were determined for the MUC1(Tn)–MGL and the MUC1(STn)–MGL interaction (Tables). The potential importance of the bond lifetime will thus not influence on the comparison of these two systems. Also other studies point to the importance of the lifetime as well as the on rate for bond formation when studying intermolecular bonds. The values of k_on and k_{*off, 0*} as determined using AFM has also previously been reported not to correspond with the determined K_D using thermodynamic methods. One example of such discrepancy is the antibiotic vancomycin interacting with its target in *Staphylococcus aureus* that has been characterized by AFM, Isothermal calorimetry, affinity capillary electrophoresis and competitive titration methods. The value of K_D estimated based on the AFM observations was 3–6 orders of magnitude from the range of bulk solution values. Such a discrepancy might be related to the conditions inherent in conventional single molecule force spectroscopy, which include a suboptimal sampling of slowly formed bonds due to the limited time available for bonds to be formed. Quantification of the intermolecular rupture force between the PEG based mucin analogue carrying GalNAc and MGL revealed that the GalNAc unit binds to the MGL even when not attached to a polypeptide backbone. This is in accordance with previous observations of non-glycosylated MUC1 peptides, which revealed that the sugar residue is essential for MGL binding. GalNAc specific lectins have been shown to bind to mucin mimetic glycopolymers displaying GalNAc attached to synthetic polymer backbones. The results presented in the present paper extend the previous knowledge by affording quantitative information related to the strength of these interactions. The results indicate that the rupture force is similar to this observed for the MUC1(Tn)–MGL interaction. Quantitative data related to the glycan–lectin interaction, as provided in the present paper, complements the information obtained by other experimental tools and illustrates that the force probe approach is an interesting supplement in studies aiming at revealing the binding capabilities of glycans or other biologically important molecules. The OT based identification of the molecular groups involved in the MUC1 –MGL interactions, as presented in the current paper, and its consistency with previously published data related to the MUC1 –MGL interactions, demonstrate the strength and reliability of the OT based approach in studies of such weak intermolecular interactions. It is therefore to be expected that future studies using this methodology will contribute with new insight related to a broad range of molecular binding partners, including but not restricted to the weak glycan–lectin interactions, and their functions in cellular systems. # Conclusions The low force range attainable with OT makes this force probe well suited for studies of weak intermolecular interactions. We have used OT to provide further evidence of the interactions occurring between MGL lectins and the cancer- associated antigens mucins MUC1(Tn) and MUC1(STn). Since both of these structures are expressed in human tumors, this interaction is a likely mechanism explaining how macrophages or dendritic cells expressing MGL lectins may recognize tumor cells expressing these glycans. The interaction strength increased from 6 to 37 pN over the loading rate interval from 29 to 137 pN/s, and no significant differences in binding strength was observed between the two mucins studied. The experimental data obtained related to the Ca²⁺ dependence are considered consistent with reported Ca²⁺ dependent thermodynamics if taking into consideration the limited access to the kinetics in the OT approach. The observed absence of interactions observed between MGL and MUC1(ST), a structure expressed by breast carcinoma cells as well as by normal cells, points to the specificity of the MUC1(Tn) and MUC1(STn) MGL interaction. The results also demonstrate that MGL is able to bind the monosaccharide GalNAc, the carbohydrate moiety of MUC1(Tn), existing in a sialylated version in MUC1(STn), with comparable binding strength as the MUC1(Tn)—MGL and MUC1(STn)—MGL interactions. These observations are consistent with the interpretation that the GalNAc residue is essential for the MUC1(Tn) and MUC1(STn) interactions with MGL. The consistency between the conclusions obtained here and those previously reported related to the MGL–MUC1 interactions validate the proposed OT based approach for studies of a broad range of molecular interactions, including the weak glycan–protein interactions. The work was supported by a project grant from Breast Cancer Now, number 2011NovPR43. We thank Katsiaryna Siarpilina for collecting the data shown in. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** MS GP JB. **Funding acquisition:** JB. **Investigation:** MS SH. **Methodology:** MS SH BTS. **Resources:** GP RB JB. **Software:** MS SH BTS. **Validation:** MS SH BTS. **Visualization:** SH BTS MS. **Writing – original draft:** MS. **Writing – review & editing:** SH GP RB JB BTS MS.

# Introduction Culture-independent molecular microbiology methods have refined and redefined the knowledge of endodontic infections, revealing a diversity of species much broader than previously anticipated by culture. It has been shown that about 40–60% of the endodontic microbiome is composed of as-yet-uncultivated bacterial phylotypes, which are species that remain to be grown and characterized in the laboratory. Endodontic infections are caused by a multispecies community of bacteria, usually organized as biofilms adhered to the root canal walls, and the development of apical periodontitis has been suggested to be the result of the collective pathogenicity of the community. Although DNA-based molecular microbiology methods have allowed to accurately identify and expand the list of microbial species present in endodontic infections and associated with different clinical conditions, it is difficult or even impossible to infer physiology and pathogenicity based on these identification methods. Therefore, there is a growing need to evaluate the products released by the bacterial community members in order to understand their role in the pathogenesis of apical periodontitis. Proteomics technologies have emerged as a large-scale analysis of differentially expressed proteins, allowing a better understanding of the overall physiologic profile of cells and tissues in a given condition. In microbiological studies, proteomics has been commonly used for the study of a pure culture of microorganisms; proteome evaluations of environmental microbial communities have been referred to as either whole community proteomics or metaproteomics, and intend to characterize the entire protein complement of the community at a given point in time. Metaproteomics may help interpret the bacterial biofilm behavior and interaction with the host by building inventories of the final gene products, i.e., proteins, released by the community. Because bacterial communities face numerous challenges in their natural environment, it is important to analyze the products of gene expression directly in samples. Indeed, studies in the area of proteomics have allowed the qualitative and quantitative evaluation of proteins present in certain environments. Improved performance of proteomics relies substantially on previous sequencing and especially metagenome efforts. A combination of liquid chromatography (LC) with mass spectrometry (MS) has become a powerful approach for the identification of proteins occurring in complex mixtures. In methods based on LC, hundreds of proteins or peptides are separated by chromatographic columns, detected, identified and quantified by mass spectrometry in a single operation. Bottom-up or shotgun proteomics is a high-throughput technology that can characterize a very large number of proteins at the same time. In this approach, proteins present in a sample are first digested into peptides by proteolytic enzymes, usually trypsin, which cleaves the protein specifically at the carboxy-terminal region of arginine and lysine residues. Next, peptides are ionized by matrix-assisted laser desorption/ionization (MALDI) or electrospray ionization (ESI), which is coupled to LC, and then analyzed by a mass spectrometer. Detection is based on the mass- to-charge (*m/z*) ratio of the peptides. Afterwards, the peptide-sequencing data are searched against protein databases using specific tools. Over the last decade, there has been an increased interest in high-resolution mass spectrometry with time-of-flight (TOF) and Orbitrap mass analyzers, usually coupled to LC. Different configurations have been used, with the most common being quadrupole-TOF (QTOF) and linear ion trap quadrupole-Orbitrap (LTQ- Orbitrap), which permit mass determination with high accuracy and resolution. Thus far, the only metaproteomic analysis of endodontic infections was performed by Nandakumar et al., who applied reverse-phase nano-liquid chromatography- tandem mass spectrometry (nLC-MS/MS) for the identification of bacterial proteins in 7 cases of primary or persistent infections. They found proteins involved with adhesion, autolysins, proteases, virulence factors, conjugation and antibiotic resistance. The present study intended to expand the knowledge of the metaproteome of endodontic infections using two complementary MS platforms to analyze samples from acute apical abscesses and from infected root canals of teeth with asymptomatic apical periodontitis taken before and after treatment. In addition, human proteins associated with endodontic infections were identified. Knowledge of the functional expression of proteins in the environment has the potential to contribute to the understanding of the ecological and pathogenic behavior of bacterial communities, and may be of value for identification of potential biomarkers of disease activity or persistence, which may be important for diagnostic and therapeutic purposes. # Materials and Methods ## Ethics statement This study was carried out in accordance with the guidelines of, and after approval by, the Ethics Committee at Estácio de Sá University, Rio de Janeiro, Brazil, and written informed consent was obtained from the patients. ## Subjects and Sample Collection Samples were taken from patients who had been referred for root canal treatment or emergency treatment to the Department of Endodontics, Estácio de Sa University. Only teeth from adult patients (ages ranging from 20 to 39 years) with carious lesions, necrotic pulps, and radiographic evidence of apical periodontitis were included in this study. Samples were obtained from the root canals of 12 teeth with asymptomatic apical periodontitis and from aspiration of pus from two cases of acute apical abscess, which showed localized swelling, fever, lymphadenopathy, and malaise. No apparent communication from the abscess to the oral cavity or the skin surface was observed. Selected teeth showed no significant gingival recession and were free of periodontal pockets deeper than 4 mm. Samples from the root canals of teeth with asymptomatic apical periodontitis were taken as follows. After the tooth crown was cleansed with pumice, a rubber dam was placed and the tooth and the surrounding field were decontaminated by a protocol using 3% hydrogen peroxide followed by 2.5% sodium hypochlorite (NaOCl) solution. Complete access preparations were made using sterile burs without water spray. The operative field, including the pulp chamber, was again swabbed with 2.5% NaOCl, which was then inactivated with sterile 5% sodium thiosulfate. If the root canal was dry, a small amount of 10 mM Tris–HCl (pH 8.0) was placed in the canal. A K-type file no. 15 was introduced up to approximately 1 mm short of the root apex, based on radiographs, and used to gently file the canal walls. Afterwards, the fluid in the canal was aspirated using a sterile disposable syringe and transferred to a cryotube containing protease inhibitor phenylmethanesulfonylfluoride (PMSF) and immediately frozen at −80°C. This root canal sample was called S1. Samples from acute apical abscesses were taken by aspiration of the purulent exudate from the swollen mucosa over each abscess. The overlying mucosa was disinfected with 2% chlorhexidine (CHX), and a sterile syringe was used to aspirate pus, which was immediately injected into cryotubes containing PMSF and frozen. Root canal samples from teeth with asymptomatic apical periodontitis were also taken after chemomechanical procedures. Canals were instrumented at the same appointment in all cases by using BioRaCe instruments (FKG Dentaire, La Chaux- de-Fonds, Switzerland) with the working length (WL) established 1 mm short of the radiographic apex. Master apical files ranged from BR5 (40/.04) to BR7 (60/.04), depending on both the root anatomy and the initial diameter of the root canal. Patency of the apical foramen was confirmed with a K-type file no. 20 throughout the procedures. Irrigation was performed with either 2.5% NaOCl (6 teeth) or 2% CHX (6 teeth), using disposable syringes and NaviTip needles (Ultradent, South Jordan, UT, USA) inserted up to 4 mm short of the WL. Post- instrumentation (S2) samples were taken from the root canals as described for S1 samples. ## Polymerase chain reaction for bacterial presence Ten microliters from clinical samples were subjected to DNA extraction by using the QIAamp DNA Mini Kit (Qiagen, Valencia, CA, USA), following the protocol recommended by the manufacturer. Presence/absence of detectable bacteria in clinical samples was determined by using end-point PCR with universal 16S rRNA gene primers 8f (5′- AGA GTT TGA TYM TGG C - 3′) and 519r (5′- GTR TTA CCG CGG CTG CTG - 3′). Positive and negative controls were included in each batch of samples analyzed. Positive controls consisted of DNA extracted from *Enterococcus faecalis* (ATCC 29212). Negative controls consisted of sterile ultrapure water instead of sample. All reactions were run in triplicate. PCR reactions were performed in 50 µL of reaction mixture containing 1 µM of each primer, 5 µL of 10× PCR buffer (Fermentas, Burlington, ON, Canada), 3.8 mM MgCl₂, 2.5 U of *Taq* DNA polymerase (Fermentas) and 0.2 mM of each deoxyribonucleoside triphosphate (Invitrogen Life Technologies, Carlsbad, CA, USA). PCR amplifications were performed in a DNA thermocycler (Mastercycler Personal, Eppendorff, Hamburg, Germany). Cycling conditions consisted of initial denaturation step at 95°C/2 min, followed by 36 cycles at 95°C/30 s, 55°C/1 min, and 72°C/1 min, and final extension at 72°C/10 min. PCR products were subjected to electrophoresis in a 1.5% agarose gel–Tris-borate-EDTA buffer. The gel was stained with GelRed (Biotium, Hayward, CA, USA) and visualized under ultraviolet illumination. ## In-solution trypsin digestion With the purpose to identify the exoproteome, samples were centrifuged at 9.000 *g* for 20 min, the supernatant was collected and the pellet discarded. Because pilot tests showed that individual root canal samples had very low detectable levels of proteins, samples were pooled for metaproteome analyses: one pool for 6 S1 samples and another pool for the 6 respective S2 samples from canals treated with NaOCl as the irrigant; the same was done for canals irrigated with CHX. Twenty microliters of each sample was used in each pool. Samples from acute apical abscesses were analyzed individually. Before enzymatic digestion with trypsin, samples were reduced with 10 mM dithiothreitol for 1 h at 56 °C and alkylated with 40 mM iodoacetamide for 30 min at room temperature in the dark. Fifty micrograms of proteins from clinical samples, previously dosed by the Folin-Lowry method, were digested with trypsin (Promega, Madison, WI, USA) (1∶50, w/w) overnight at 37°C and the peptides were desalted by Zip Tip® (Millipore, Billerica, MA, USA). Samples were vacuum-dried and reconstituted with 0.1% formic acid. ## Nanoflow liquid chromatography coupled with LTQ Velos Orbitrap An aliquot containing 4.5 µL of each pool of root canal samples and the individual samples from abscesses were loaded on LTQ Velos Orbitrap mass spectrometer (Thermo Fisher Scientific, Waltham, MA) connected to nanoflow LC (nLC-MS/MS) by an EASY-nLC system (Proxeon Biosystems, West Palm Beach, FL, USA) through a Proxeon nanoelectrospray ion source. Peptides were separated by a 2%–90% acetonitrile (ACN) gradient in 0.1% formic acid using an analytical column PicoFrit Column (20 cm × ID75 µm, 5 µm particle size, New Objective, Woburn, MA), at a flow of 300 nL/min over 45 min. The nanoelectrospray voltage was set to 1.7 kV and the source temperature was 275°C. All instrument methods for the LTQ-Orbitrap Velos were set up in the data-dependent acquisition mode. The full scan MS spectra (m/z 300–1,600) were acquired in the Orbitrap analyzer after accumulation to a target value of 1e6. Resolution in the Orbitrap was set to r = 60,000 and the 20 most intense peptide ions with charge states ≥2 were sequentially isolated to a target value of 5,000 and fragmented in the linear ion trap by low-energy collision induced dissociation (CID) as dissociation or fragmentation method, normalized collision energy of 35%. The signal threshold for triggering an MS/MS event was set to 500 counts. Dynamic exclusion was enabled with an exclusion size list of 500, exclusion duration of 60 s, and repeat count of 1. An activation q = 0.25 and activation time of 10 ms were used. Samples were analyzed in duplicate. Peak lists (msf) were generated from the raw data files using Proteome Discoverer version 1.3 (Thermo Fisher Scientific) with Sequest search engine and searched against Human (200,740 sequences and 86,640,852 residues) and Bacterial (377,577 sequences and 139,313,674 residues) protein database downloaded from UniProt (<http://www.uniprot.org>) in September 2012, with carbamidomethylation as fixed modification, oxidation of methionine as variable modification, one trypsin missed cleavage and a tolerance of 10 ppm for precursor and 1 Da for fragment ions. All datasets were processed using ScaffoldQ+v.3.3.1 software with false discovery rate less that 1% as cut-off. Gene Ontology annotation for functions category (biological or molecular) were obtained using UniProt tools and database. ## LC-QTOF analysis Samples were vacuum-dried and reconstituted in 20 µL of 0.1% formic acid. A total of 4 µL of reconstituted peptide mixture was injected onto an LC-MS system consisting of a 1260 series liquid chromatograph, HPLC-Chip Cube MS interface, and 6530 QTOF mass spectrometer (all Agilent Technologies, Santa Clara, CA). The system was equipped with an HPLC-Chip (Agilent Technologies) that incorporated a 360 nL enrichment column and a 150 mm ×75 µm reverse phase column was packed with Polaris-C18, 3 µm particles. Three analytical replicates were analyzed by mass spectrometry. For each mass spectrometry experiment, peptides were loaded onto the enrichment column with solvent A (water with 0.1% formic acid). A two- step gradient generated at a flow rate 0.3 µL/min was used for peptide elution. This included a linear gradient from 3% solvent B (acetonitrile with 0.1% formic acid) to 40% B over 40 min followed by a sharp increase to 90% B within 5 min. The total running time, including column reconditioning, was 65 min. The column effluent in all cases was directly analyzed by the 6530 QTOF mass spectrometer that was interfaced with an HPLC-Chip Cube nanospray source. The latter was operated at a capillary voltage of 1950 V with a capillary current of 0.085 µA in extended dynamic range (2 GHz) mode. The MS data were acquired in the positive ionization mode using Agilent MassHunter Workstation QTOF B.04.00. During the course of data acquisition, the fragmentor voltage, skimmer voltage, and octopole RF were set to 150, 65, and 750 V, respectively. Auto-MS/MS was performed using scan speed varied based on precursor abundance option with a maximum cycle time of 6.8 s. In each cycle, MS spectra were acquired at 8 Hz (eight spectra/s) (m/z 295–1700), and the twenty most abundant ions (with charge states 2+, 3+, and \>3+) exceeding 1000 counts were selected for MS/MS (m/z 50–1700). A medium isolation (4 m/z) window was used for precursor isolation. A collision energy table (m/z 300–1500, z1 \[CE 9–69\], z2 \[CE 9–69\], z3 \[CE 9–54\], z\>3 \[CE 9–54\] was used for fragmentation with beam-typefragmentation or high energy collision dissociation (HCD) in 6530 QTOF. Reference mass correction was activated using a reference mass of 1221.9906. Precursors were set in an exclusion list for 0.15 min after one MS/MS spectra. LC-QTOF data was searched against the UniProt micro-organism and human database, using the Agilent Spectrum Mill Server software (Rev B.04.00.127). Data were extracted and peak lists were created with the Spectrum Mill Data Extractor program with the following attribute: scans with the same precursor ±0.03 *m*/*z* were merged within a time frame of ±60 s. The UniProt microorganism and human database was searched for tryptic/non-tryptic peptides with a mass tolerance of 10 ppm for the precursor ions and a tolerance of 50 ppm for the fragment ions. Two missed cleavages, fixed Modification (carbamidomethylation) and variable modifications \[oxidized methionine (M)\] were allowed. Spectrum Mill autovalidation was performed at peptide and protein level (1% FDR). ## Protein classification Identified bacterial proteins were categorized according to biological function as follows: transcription and translation, cell division/peptidoglycan synthesis, chemotaxis, DNA process, energy metabolism, fatty acid metabolism, methanogenesis, membrane, nucleotide metabolism, antibiotic resistance, adhesion, pathogenesis/virulence, proteolysis, stress response, transport, vitamin biosynthesis, and other/unknown. Human proteins were classified as follows: cellular process and metabolism, immune system, circulatory system, extracellular connective matrix, and other/unknown. Keratins and other presumed contaminants were removed from the entire dataset. # Results PCR analysis of the individual clinical samples using universal 16S rRNA gene- based primers revealed the presence of bacteria in all S1 and abscess samples. Of the S2 samples, 4/6 samples from the NaOCl group and 5/6 samples of the CHX group showed positive results for bacteria. Analyses of the metaproteomic data generated by the two mass spectrometers revealed a total of 308 proteins of microbial origin. A larger number of proteins were identified in abscess samples (173 proteins) when compared to S1 samples from teeth with asymptomatic apical periodontitis (88 proteins). A large part of these proteins were classified as having “other/unknown function” (53 proteins from acute infection and 24 from the asymptomatic cases). In the group of canals irrigated with CHX, the number of identified proteins decreased from 74 in S1 to 31 in S2. In the NaOCl group, however, the number of proteins increased from 14 to 35. In abscess samples, proteins involved with transcription and translation processes were more frequent (30 proteins), followed by proteins associated with energy metabolism (21 proteins) and DNA processes (18 proteins). In root canal samples from asymptomatic teeth, proteins involved with transcription and translation processes were also more frequently identified (16 proteins) as compared to other proteins. Two proteins related to antibiotic resistance, TetR and a beta-lactamase, were detected in S1 samples. Proteins linked to pathogenesis/virulence were detected in one abscess sample, one pool of S1 samples and the pool of S2 samples from the NaOCl group. Eight proteins with proteolytic activity were found and most of them (six) occurred in abscesses. Proteins participating in bacterial adhesion were detected in one abscess sample, as well as in S1 and S2 samples from the NaOCl group. Proteins involved in production of biofilm matrix were found in one pool of S1 samples. Two CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats)-associated proteins were found in one pool of S1 samples and S2 from the NaOCl group. An archaeal protein linked to production of methane was detected in one pool of S1 samples. groups the identified microbial proteins according to their potential role in either pathogenicity or resistance/survival. (supplementary material) depicts all proteins of bacterial or archaeal origin found in this study. Analyses of human proteins using the two mass spectrometers revealed 139 proteins, which were classified as follows: cellular process and metabolism (87 proteins), circulatory system (9 proteins), extracellular connective matrix (2 proteins), immune system (24 proteins), and other/unknown functions (17 proteins). Defensins 1 and 3 were identified in one abscess sample, one pool of S1 samples and the pool of S2 samples from the NaOCl group. Proteins involved with the host defense to infection are shown in. All the human proteins identified in this study are detailed in (supplementary material). # Discussion This study evaluated the metaproteome associated with endodontic infections by using two mass spectrometry platforms. The large majority of proteins found were related to metabolic and housekeeping processes, including protein synthesis, energy metabolism and DNA processes. This is indicative of a living active microbial community. Moreover, several other proteins of interest were detected, including some related to pathogenicity and resistance/survival. Because these proteins may assume special relevance in terms of pathogenesis of apical periodontitis and resistance to host defenses and treatment, they deserve a more detailed discussion as follows. ## Microbial proteins involved in pathogenicity Biofilm formation is an important virulence trait of many bacterial pathogens. It has been shown that apical periodontitis is a disease caused by biofilms colonizing the root canal system. In the present study, three proteins involved in bacterial adhesion were detected: glycosyltransferase 1, tight adherence protein G, and coagulation factor 5/8 type domain protein. The enzyme glycosyltransferase 1 was found in initial root canal samples. This enzyme is involved with production of extracellular polysaccharides that are important for bacterial adhesion to surfaces and biofilm formation. Eight proteins with proteolytic activity were found in this study, most of them associated with abscesses. They included a collagenase, a metalloprotease, a serine protease, an extracellular protease, and endopetidases. These enzymes may play an important role in several ecological and pathogenic effects, including tissue invasion, acquisition of nutrients from proteins, destruction of the connective tissue matrix, and inactivation of host defense molecules. Streptopain (SPE B), a protein of streptococcal origin, was found in one pool of S1 samples. This is a cysteine protease-like exotoxin with several biological effects, including cleavage of fibronectin, vitronectin, and interleukin 1-β precursor and participation in the apoptosis of monocytes and epithelial cells. Therefore, this exotoxin can participate in bacterial invasion of connective tissues, induction of inflammation, reduction of phagocytic activity, and inhibition of tissue repair. Another protein involved in tissue invasion, a putative invasin, was detected in an abscess sample. Tissue invasion can be regarded as a crucial initial step for the pathogenicity of several bacteria. ## Microbial proteins involved in resistance/survival processes Several stress proteins were detected in the samples tested. Cells respond to environmental stress by inducing or accelerating the synthesis of specific proteins known as stress proteins, including heat-shock proteins (HSPs), which act as molecular chaperones in the assembly and folding of proteins, as well as proteases when damaged proteins need to be degraded. In addition to being found intracellularly, HSPs can also be located on the cell surface and on outer membrane vesicles released by bacteria. The chaperone DnaK (HSP70) was found in one abscess sample and in S2 samples. Another chaperone, (GroEL or HSP60), was found in S1 samples from asymptomatic root canal infections. HslU, which is another HSP, was also found in this study and has been shown to provide an essential ATPase activity. These findings may indicate a stress response to host defenses and treatment, in an attempt to survive and/or recover from damage. These stress proteins may also act as virulence factors via the following mechanisms: cytotoxicity, adhesion, modulation of host cell activities, and induction of synthesis of pro-inflammatory cytokines. Others stress response proteins detected in this study include alkyl hydroperoxide reductase, whose expression has been shown to be up-regulated in some anaerobes upon oxidative stress ; a periplasmic sensor signal transduction histidine kinase, which senses specific environmental stimuli ; a tetratricopeptide TPR_2 repeat protein, which mediates protein-protein interactions and the assembly of multiprotein complexes, and may be involved in protein folding. Catalase and quinone oxidoreductase are involved in the response to oxidative stress, and were found in post-treatment samples. Catalase also plays an important role in the reactivation of bacteria present in a viable but noncultivable state. A toxin component of the putative toxin-antitoxin system was detected in an abscess sample. This system may mediate the general stress response, and participate in the regulation of biofilm and persister cell formation, which is an important mechanism of reduced susceptibility of biofilms to antibiotics. Proteins involved in DNA repair are also of interest because they may participate in the response to stress conditions induced by treatment or host defenses. Some of the DNA repair-related proteins detected in this study were A/G-specific adenine glycosylase, DNA ligase, ATP-dependent helicase, and UvrABC system proteins. Of the 12 proteins related to DNA repair, 8 were found in abscesses; in these cases, there may be a strong oxidative stress caused by phagocytes during the combat to infection. Two enzymes that confer resistance to antibiotics were detected: beta-lactamase (resistance to penicillins and cephalosporins) and TetR (resistance to tetracyclines). Previous studies have detected the genes for these antibiotic resistance enzymes directly in root canals samples or in isolates from endodontic infections, but in situ expression of the gene products have not been previously determined. In a previous study using a metaproteomic approach, Nandakumar et al. also identified enzymes related to resistance to tetracyclines and beta-lactams. These findings confirm that endodontic bacteria may produce antibiotic resistance enzymes in the root canal environment and this may have some important implications, since antibiotics are used for some acute conditions or have even been proposed as topical irrigants or intracanal medication during root canal treatment. ## Other proteins of interest Another noteworthy finding was the identification of the enzyme methyl coenzyme M reductase of archaeal origin. Methanogenic archaea have been detected in endodontic infections by DNA-based molecular methods, including by an assay directed to *mcrA*, which encodes for the methyl coenzyme M reductase. The presence of products of this microorganism in endodontic infections indicates metabolic activity in a complex network of microbial interactions. CRISPR consists of identical repeated DNA sequences interspaced by highly variable sequences (spacers). CRISPR-associated (cas) genes encode conserved proteins that together with CRISPRs compose the CRISPR/Cas system, which is present in many prokaryotes and confer protection against invasion by phages, plasmids, and transposons. A previous study reported the occurrence of CRISPR- cas in several endodontic isolates of E. faecalis and suggested a role for this system in modulation of interactions in the endodontic polymicrobial biofilm. In the present study, CRISPR-associated proteins were identified in samples taken before and after treatment, confirming that this prokaryote defense system is present in members of endodontic infections. ## Host defense proteins Human proteins were identified in all samples, being more frequent in abscess samples and in S1 samples from asymptomatic teeth. The fact that most proteins identified are intracellular and are involved in cellular and metabolic processes can be possibly explained by rupture of cells during the inflammatory process in response to infection, especially in abscesses. Another possible source of intracellular proteins is the remnants of necrotic pulp tissue in the canal. However, it must be recognized that cellular rupture may also have occurred during sample taking or processing steps. In addition, 24 proteins related to the innate or adaptive immune system were found. They include defensins 1 and 3 and myeloperoxidases, which are produced by polymorphonuclear neutrophils. These proteins are not only involved with defense against microorganisms but also in tissue destruction. Other immunity-related proteins of interest include components of the complement system, imunoglobulins, protease inhibitors, receptors involved in the regulation of T cells, such as CTLA4 (Cytotoxic T-lymphocyte protein 4), and T-LAK cell-originated protein kinase (Lymphokine-activated killer T-cell-originated protein kinase), which is involved in the activation of lymphoid cells. Databases for human proteins are much more complete and accurate, reducing the number of proteins classified as unknown. ## Abscess *versus* asymptomatic cases Metaproteomic analyses revealed a larger number of proteins in the abscess samples when compared to samples from asymptomatic infection, in spite of the fact that samples from asymptomatic teeth came from a larger number of cases. Studies have shown that the number of bacterial cells and species in abscesses are greater than in asymptomatic cases, which may help explain our findings. Also, it is expected that bacterial cells engaged in an acute infection are in active processing of protein expression machinery. However, one should not discard the possibility that in abscesses, samples were taken from pus, in which an acute host response is established and a high killing rate of host cells and bacteria is expected. This would generate more proteins, including cytoplasmic proteins. ## Post-treatment samples Analysis of the total number of proteins detected after treatment revealed a decrease in the CHX group but an increase in the NaOCl group. The purpose of analyzing post-instrumentation samples was two-fold: first, to evaluate the effects of treatment on the metaproteome; second, to evaluate if the proteins expressed after treatment procedures would differ from those present before treatment, considering the possibility that microorganisms resisting to treatment may produce a different pattern of proteins. The reduction in the CHX group was completely expected, but the increased in NaOCl group may be related to the tissue dissolving ability of this substance and its high destructive effects on cells, releasing cytoplasmic proteins to the environment. This is coherent with the types of proteins found in S2 samples from the NaOCl group. In the group of CHX, one stress response protein was found, and this may be related to a response to treatment. However, a more distinct pattern was not evident. This may have been due to one of the following reasons: limitation of the sampling approach; few bacterial cells remaining after treatment; and the short time elapsed between treatment and sampling, not allowing sufficient time for bacteria to change their pattern of protein expression in order to adapt themselves to the altered environment. If the latter is true, further studies should be performed to analyze the metaproteome associated with bacterial persistence after treatment, and for this, researchers should wait a little longer to collect samples after intracanal procedures. ## Technical and analytical considerations Actually, in the present study, efforts were expended towards evaluation of the exoproteome. However, many cytoplasmic proteins were detected. This may be due to cells that recently died and released their protein content in the environment (especially in abscesses and in samples taken after treatment), which are still considered as part of the exoproteome if these proteins are stable and remain in abundance. However, presence of cytoplasmic proteins may also have been a result of sample processing that resulted in cell lysis. Several proteins identified in this study were not linked to endodontic bacteria, which may be related to databases with incomplete information for oral bacteria. This also may help explain why a large number of proteins remained unidentified and were classified as having unknown function. It is salient to point out that the majority of the bacteria found in the endodontic microbiome, especially the as-yet-uncultivated portion, have not had their genomes sequenced. The same is true for bacteria in other human and environmental sites. Although the extent of protein sequence conservation is largely unknown, the possibility exists that most of the peptides attributed to nonoral bacteria actually belong to phylogenetically similar oral relatives of those species, which remain to be sequenced. It is also important to note that proteins from uncultivated or not-yet-sequenced cultivated species that are abundant in the community may potentially pass unnoticed as there is no representative sequence in the database. The number of proteins identified may have also been affected by the principle of parsimony adopted by the softwares used in this study, which reduce the occurrence of redundant sequences. Moreover, the utilization of rigid filters, such as the 1% False Discovery Rate, and filtering of data generated from Proteome Discoverer by Scaffold may have contributed to a limited but more reliable protein identification,. Another limitation of this study is that, unlike DNA, proteins cannot be amplified, therefore less abundant proteins may be undetected. In contrast to bacterial isolates cultivated in the laboratory, human clinical samples, especially from a small environment like the root canal, provide a reduced amount of biomass available for proteomic analysis. Therefore, for root canal samples to be properly analyzed they had to be pooled. In addition, many proteins were found only by one of the methods used. This may be related to the different aliquots used in each identification, which may influence detection of proteins found in low abundance. The peptide mixture generated by the shotgun proteomic strategy is so complex to permit that the mass spectrometer acquires the mass spectra of all peptides in a single LC-MS/MS run. Consequently, these experiments must be frequently repeated so as to increase the number of peptides obtained by mass spectra. This approach was used in the present study. # Conclusions Interrogation of the metaproteome of endodontic microbial communities provides information on the physiology and pathogenicity of the community at the time of sampling. The present study qualitatively described the proteins of microbial and human origin in association with endodontic infections. There is a growing need for expanded and more curated protein databases that permit more accurate identifications of proteins in metaproteomic studies. # Acknowledgments The authors are grateful to Vadi Bhat, from Agilent Technologies, for the technical support with the 6530 QTOF mass spectrometer. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JCP JFS INR. Performed the experiments: JCP INR RRD. Analyzed the data: JCP JFS INR RRD AFPL MRSS. Contributed reagents/materials/analysis tools: JFS INR AFPL MRSS. Wrote the paper: JCP JFS INR AFPL MRSS.

# Introduction Identifying a protein's interaction partners is essential for deciphering protein function. The yeast two-hybrid (Y2H) system is a high-throughput genetic method that enables rapid genome-wide screening to discover a protein's interaction partners. In its traditional and extensively-used form, the Y2H system relies on the ability of chimeric proteins to activate transcription of a reporter gene, an event that takes place in the nucleus. Although the Y2H is a powerful approach, limitations arise from the fact that protein-protein interactions are interrogated at a specific subcellular location, the nucleus. Transcriptional activators represent a class of proteins that cannot readily be studied using the traditional Y2H. These proteins can activate transcription on their own, thereby obscuring the transcriptional readout of the Y2H assay. For these reasons, an assay that examines protein-protein interactions at another subcellular location would provide a powerful complement to Y2H technology. Indeed, several groups have described two-hybrid or protein complementation assays (PCAs) that take place in the secretory pathway, on the cell surface, or in the periplasmic space of bacteria, but none of these methods has yet been widely adopted and each has its drawbacks. Those assays that rely solely on fluorescence-activated cell sorting, necessitate access to special instrumentation. Furthermore, they are not as efficient at analyzing large libraries as assays that depend on survival or conditional growth. Assays that take place in bacteria, may not be suitable for analyzing proteins that require specialized factors or post-translational modifications only present in mammalian cells. To fill these gaps, we report the Golgi two-hybrid (G2H) system, a method for identifying protein-protein interactions in the secretory pathway of yeast. In the design of the G2H system, we took inspiration from key features of the traditional Y2H system. Specifically, we noted that the traditional Y2H system relies on the modularity of a transcription factor, which is separated into two domains. Protein-protein interactions bring together those two domains, reassembling the transcription factor and activating transcription of a reporter gene or genes. In an analogous fashion, our G2H method capitalizes on the modular nature of Golgi-resident glycosyltransferases, which consist of localization (LOC) and catalytic (CAT) domains. Golgi-resident glycosyltransferases add monosaccharides to protein and lipid substrates, a large fraction of which are subsequently trafficked to the cell surface or the extracellular space. In the G2H system, the interaction between two proteins reconstitutes the glycosyltransferase in a null background, thereby restoring wild-type cell surface glycosylation. Selection and screening strategies are critical features of any approach to protein-protein interaction discovery. In the traditional Y2H system, a protein- protein interaction results in activation of a reporter gene that, in turn, confers a survival advantage or enables visual identification. Survival-based reporters, such as HIS3, provide selective pressure, thereby enabling rapid analysis of large libraries of cells. Survival-based selection can be combined with screening using a second reporter, such as LacZ, that induces a color or fluorescence change. By using two reporters, large cDNA libraries can be rapidly screened while eliminating many false positives. In contrast, most non-nuclear two-hybrid assays and PCAs rely on a single readout, thereby limiting throughput and increasing the incidence of false positives. Guided by the Y2H, we chose a Golgi-resident glycosyltransferase whose activity can be detected in multiple ways. In the G2H system, changes in cell surface glycosylation form the basis of both conditional growth and flow cytometry-based assays. Each of these assays could, in principle, be used both to screen individual strains and to select cells containing interacting bait-prey pairs from a mixed population. We describe the fundamental features of the G2H method, a two-hybrid assay that takes place in the secretory pathway of eukaryotic cells. We demonstrate the utility of the G2H system by using it to observe protein-protein interactions that cannot be detected through traditional Y2H methods. We show that the selection strategies incorporated in the G2H method can be used to identify and enrich for cells containing interacting bait-prey pairs. # Results ## The G2H reporter is the yeast glycosyltransferase Och1p The G2H assay is based on the reassembly of a Golgi-resident glycosyltransferase. Most Golgi-resident enzymes are composed of modular LOC and CAT domains. The N-terminal LOC domain dictates localization and is anchored in the membrane, while the C-terminal CAT domain resides in the lumen of Golgi and performs the sugar transfer reaction. When these domains are expressed separately no catalytic activity is observed: the LOC domain is properly localized in the Golgi but lacks catalytic activity, and the CAT domain does not encounter its substrates because it is secreted by default. Proteins to be interrogated, a bait and a prey, are then fused to the LOC and CAT fragments of a reporter glycosyltransferase. Interaction between bait and prey reconstitutes the glycosyltransferase and restores its activity, resulting in a concomitant change in cell surface glycosylation. Robust selection strategies are paramount to the utility of the G2H assay. Therefore, we sought to identify a glycosyltransferase whose activity could form the basis of a conditional growth assay as well as at least one complementary screening method. We examined phenotypic changes caused by the activity of Golgi-resident *S. cerevisiae* glycosyltransferases whose activities are critical to cell wall integrity, by assessing the growth of deletion strains at elevated temperature or in the presence of small molecule stressors. We focused attention on *och1*Δ because it grows only slightly slower than the parental strain under permissive conditions (30°C), but exhibits strong sensitivity to temperature and to Congo red, caffeine, and hygromycin. Och1p is an α1-6-mannosyltransferase that adds mannose to Man₈GlcNAc₂ in *N*-linked glycans to yield Man₉GlcNAc₂. This enzymatic reaction initiates the formation of the mannan outer chain. Once the product Man₉GlcNAc₂ is formed, a cascade of α1-6-mannosyltransferases acts to produce Man_50-100GlcNAc₂, a high mannose structure that covers the cell wall of wild-type yeast. While *och1*Δ yeast exhibited dramatically slowed growth under non-permissive conditions (37°C or 30°C in the presence of Congo red), transformation with an *OCH1* plasmid rescued growth to close to wild-type levels. Although *och1*Δ yeast exhibited reduced growth in both liquid culture and on agar plates, we conducted all subsequent growth assays on agar media because we were concerned that the flocculation tendency of *och1*Δ yeast might interfere with our ability to accurately measure growth in liquid culture. In addition to the conditional growth assays, we assessed the composition of *och1*Δ yeast cell walls by flow cytometry. Since Och1p plays a critical role in mannan biosynthesis, *och1*Δ yeast have dramatically different *N*-linked glycans from their wild-type counterparts. Chitin-binding reagents such as Alexa488-labeled lectin wheat germ agglutinin (WGA) bind strongly to *och1*Δ but weakly to the parental strain. Furthermore, expression of full-length *OCH1* in *och1*Δ complements the WGA binding phenotype. ## The interaction between MyoD and Id2 restores Och1p activity Our next step was to demonstrate that Och1p enzymatic activity could be reassembled from the component LOC and CAT domains. As in previous work , we used sequence alignment of Och1p protein sequences from multiple species to identify the LOC and CAT domains. We designated *S. cerevisiae* Och1p amino acids 1-80 as LOC and Och1p amino acids 78-481 as CAT. DNA sequences encoding LOC and CAT domains were cloned into the yeast expression vectors p425-TEF and p426-TEF, respectively ( and and). Since the catalytic domain contains no localization cue, we also included an N-terminal signal peptide to ensure its entry into the secretory pathway. To evaluate the ability of Och1p activity to be reconstituted from its modular components, we examined a known protein-protein interaction, MyoD binding to Id2. While this interaction has been used as a positive control in Y2H assays, it presents a potential challenge for detection. In addition to forming a heterodimer with Id2, MyoD can also homodimerize through the same interface. We wondered if the G2H assay would be able to detect MyoD:Id2 complex formation in the face of competing MyoD homodimerization. Thus, we prepared plasmids encoding the fusion proteins LOC-MyoD and Id2-CAT. To ensure that any interaction we observed was specific, we also prepared a LOC domain fused to SV40 large T antigen (LOC-SV40TAg) and a CAT domain fused to p53 (p53-CAT). *och1*Δ yeast were transformed with pairs of plasmids. All strains grew at 30°C and only slight growth differences were apparent under these permissive growth conditions. Under non-permissive growth conditions (at 37°C or at 30°C in the presence of Congo red), dramatic differences in strain growth were observed. *och1*Δ yeast transformed with either LOC-MyoD or Id2-CAT displayed *och1*Δ-like growth both at 37°C and at 30°C in the presence of Congo red. Conversely, *och1*Δ co-transformed with both LOC-MyoD and Id2-CAT plasmids grew robustly at elevated temperature and on Congo red plates. The observed growth differences are unlikely to be due to toxicity of the expressed proteins, because none of the plasmids affected growth of the parental yeast strain. We interpret the rescued growth observed in *och1*Δ\[LOC-MyoD\]\[Id2-CAT\] yeast to mean that Och1p activity is reconstituted in these cells. The distinction between *och1*Δ and *och1*Δ\[LOC-MyoD\]\[Id2-CAT\] was most pronounced in the presence of Congo red, suggesting that Congo red provides a stronger selective pressure than elevated temperature. Taken together, these data indicate that Och1p, like mammalian glycosyltransferases, is modular and can be reassembled into an active form via the interaction between MyoD and Id2. To confirm that the growth phenotype observed in LOC-MyoD/Id2-CAT-transformed cells is due to Och1p activity, Alexa488-modified WGA lectin was employed to fluorescently label yeast based on cell wall composition. Flow cytometry analysis of yeast incubated with Alexa488-WGA revealed that the cell wall of *och1*Δ\[LOC-MyoD\]\[Id2-CAT\] is shifted to resemble that of complemented *och1*Δ\[OCH1\], whereas the cell wall of *och1*Δ transformed with only Id2-CAT is indistinguishable from that of *och1*Δ cells. These data provide further evidence that Och1p activity is present in cells containing both LOC-MyoD and Id2-CAT, yet is absent from cells containing only the Id2-CAT fusion. Thus, Och1p is modular, and its inactive modules can be reassembled via the MyoD-Id2 interaction to form a functional enzyme. ## Mutations to the MyoD-Id2 interaction interface increase Congo red sensitivity and WGA binding To confirm that the rescued growth and WGA binding phenotypes observed in *och1*Δ\[LOC-MyoD\]\[Id2-CAT\] yeast are due to the interaction between MyoD and Id2, we designed mutations to disrupt this interaction. Because a crystal structure of the MyoD-Id2 complex is not available, we modeled the MyoD-Id2 interaction using the crystal structure of a dimeric form of MyoD. We reasoned that mutating the amino acids that form the tightly packed helix-helix contacts would interfere with MyoD:Id2 complex formation, preventing reassembly of Och1p and leading to increased sensitivity to Congo red and increased WGA binding. To test this possibility, we constructed plasmids that harbor point mutations in the second helix of the HLH domain of MyoD. These mutations encode changes that replace hydrophobic residues with lysines. We also prepared plasmids in which one or more turns of the second helix of MyoD are deleted. As expected, *och1*Δ yeast expressing Id2-CAT and mutant forms of LOC-MyoD grew more slowly in the presence of Congo red than *och1*Δ yeast transformed with wild-type LOC-MyoD and Id2-CAT and demonstrated increased WGA binding. Examination of the phenotypes observed for MyoD mutants suggests that the Congo red growth assay and the WGA binding assay may report on the relative strength of the bait-prey interaction. The I149K and I157K mutants produced severe growth defects and strong WGA binding. In fact, *och1*Δ yeast transformed with LOC- MyoD(I149K) and Id2-CAT were almost indistinguishable from *och1*Δ yeast transformed with Id2-CAT alone. This observation is expected given the very strong conservation of isoleucine at these positions throughout the bHLH family. On the other hand, the L160K and Q161K mutations had a more modest effect, consistent with the fact that charged and polar amino acids are found at these positions in some bHLH family members. In addition, L160 and Q161 are near the C-terminal end of the HLH motif and may not be as critical to packing as more centrally located residues. One surprise was the mild phenotypic changes observed for the L150K mutant, which appears to have only a small effect on the MyoD-Id2 interaction despite near conservation of leucine at this position throughout the bHLH family. As a control, we also made two point mutations to Id2, both of them outside of the canonical HLH motif: V86K is four amino acids C-terminal to the HLH motif, while L124K is 42 amino acids away. As expected, these mutations had milder effects than those made to the HLH region of MyoD and the severity of the phenotypes correlated with their proximity to the HLH motif. ## The G2H detects interactions of Gal4p's activation domain (AD) Next, we examined whether the G2H can detect interactions that cannot be studied using the traditional, transcription-based Y2H. The yeast transcription factor Gal4p contains a potent acidic AD that interferes with analysis by standard Y2H methods. *In vitro* affinity purification methods have been used to identify binding partners of Gal4p's AD, but the notoriously promiscuous binding properties of this domain have complicated analyses. As an alternative, Kodadek and colleagues used the Sos recruitment system to demonstrate that the Gal4p AD interacts with Gal80p, Hap5p, and Rpt4p. While successfully detecting some Gal4p AD binding partners, their analysis of a yeast cDNA library did not identify Gal11p, a known Gal4p AD binding partner. The absence of Gal11p was hypothesized to be due to the toxicity associated with overexpression of this transcriptional regulator, since Gal11 overexpression has been documented to cause a growth defect. Taken together, the literature data suggested that detecting Gal4p's interactions would be a demanding test for our new assay. To test the G2H's ability to detect Gal4p AD's interactions, *och1*Δ yeast were transformed with a plasmid encoding Gal4p AD fused to the *OCH1* CAT domain (Gal4AD-CAT) and with plasmids encoding the *OCH1* LOC domain fused to potential interaction partners. We observed that *och1*Δ yeast expressing Gal4AD-CAT alone exhibited increased growth in the presence of Congo red and decreased WGA binding, perhaps reflecting the notorious propensity of the acidic AD to interact nonspecifically with a variety of proteins. Despite this increased background signal, we observed a significant growth enhancement and decrease in WGA binding when the yeast were co-transformed with Gal4AD-CAT and LOC-Gal80 or LOC-Gal11, and a more modest effect for yeast co-transformed with LOC-Rpt4 or LOC-Hap5. *och1*Δ yeast containing Gal4AD-CAT and LOC-Rpt6 demonstrated a slight decrease in WGA binding, but no significant change in growth phenotype relative to *och1*Δ yeast containing the Gal4AD-CAT construct alone. These data suggest that the G2H can detect interactions of the Gal4p AD with Gal80p and Gal11p, and, to a lesser extent, Rpt4p and Hap5p. Our inability to detect an interaction between Gal4AD and Rpt6p is consistent with reported cross-linking data and might indicate that these proteins do not interact directly. Using a non- transcriptional readout enabled us to examine the interactions of a transcriptional activator and moving interaction interrogation to the Golgi seems to have relieved the toxicity that is normally observed with Gal11p overexpression. ## Selective conditions enable enrichment of cells containing interacting bait-prey pairs Having established that the G2H can be used to detect known protein-protein interactions, we wished to test whether the phenotypic changes that we observed could form the basis of a selection strategy. Approximately equal quantities of seven strains were mixed together and cultured under the selective pressure of Congo red. Each *och1*Δ strain was co-transformed with Id2-CAT and one of seven LOC plasmids: LOC (no fusion protein), LOC-SV40TAg, LOC-Gal80, LOC-Gal11, LOC- Hap5, LOC-Rpt6, or LOC-MyoD. The plasmid encoding LOC without a fusion protein was included to mimic typical plasmid library construction where some clones lack inserts. Strains were co-cultured for 72 hours and ratios of LOC plasmids in the culture were measured at various times using quantitative PCR. Within 24 hours, a significant enrichment of LOC-MyoD was observed. After 72 hours, the LOC-MyoD and LOC plasmids were dominant (66% and 23%, respectively) and all other LOC fusions were minor (\<5%) constituents. These data indicate that, under selective conditions, yeast containing an interacting bait-prey pair rapidly outcompete those with non-interacting pairs. The representation of LOC alone increased slightly over time, while representation of all non-interacting LOC fusions decreased, suggesting that fusion of LOC to a non-interacting protein actually provides negative selective pressure. Based on these observations, we predict that the G2H method can be used to detect and enrich for yeast that harbor LOC and CAT constructs fused to novel interaction partners. # Discussion We describe the development of the G2H method and show that it can be used to detect interactions among proteins that cannot be studied using transcription- based two-hybrid systems, namely transcription factors. By virtue of its secretory pathway localization, the G2H also has the potential to be used to study the secretome, a class of proteins that remains poorly characterized. Detecting protein interactions via the G2H method relies on robust phenotypic changes that will enable the G2H method to be used in both screening and selection experiments. By interrogating protein-protein interactions in the Golgi, rather than the nucleus, the G2H provides a powerful complement to transcription-based approaches. We demonstrated two specific examples of the utility of Golgi localization. First, Gal4AD is a transcriptional activator and cannot easily be studied using assays that employ a transcriptional readout. By testing Gal4p AD interactions in the Golgi, we avoided off-target transcriptional effects. Furthermore, we were to detect specific Gal4p interactions, even in the face of non-specific binding. Second, by targeting a toxic protein to the secretory pathway, we relieved its negative growth effects. Overexpression of Gal11p normally results in a dramatic decrease in growth, yet cells expressing the Golgi-localized LOC-Gal11 construct did not experience this toxicity. Indeed, the opposite was true: co-transformation with the LOC-Gal11 plasmid improved the growth of Gal4AD-CAT-expressing yeast. In addition to detecting the interactions of Gal4AD with a number of binding partners, we were pleased to observe robust detection of the MyoD:Id2 interaction. Because the LOC-MyoD fusions are sequestered to the luminal face of the Golgi membrane, we were concerned that they might preferentially homodimerize with one another, rather than heterodimerizing with the soluble Id2-CAT fusion. Nonetheless, we are able to observe strong evidence of the LOC- MyoD:Id2-CAT heterodimeric interaction. Competition for binding will also be expected in cases where the prey is endogenously expressed within the Golgi and capable of competing with the LOC fusion protein for binding to the bait-CAT fusion protein. To investigate whether this situation will interfere with interaction detection by the G2H, we plan to conduct experiments to test the ability of the G2H to detect interactions that normally occur within the secretory pathway. Like the traditional, transcriptional-based Y2H, the G2H is likely to have limitations. Proteins not normally localized to the secretory pathway could misfold in the G2H due to glycosylation of cryptic acceptor sequences or abnormal disulfide bond formation, thereby rendering them unable to engage in their normal protein-protein interactions. Misfolded proteins could also be retained in the ER, potentially leading to false positive signals. We observed reconstitution of Och1 in the Golgi through the MyoD:Id2 interaction and through the interactions of Gal4AD with a number of binding partners; therefore, at least some nuclear protein-protein interactions can assemble the secretory pathway environment, but others may be unable to do so due to differences in pH, ion concentration (Ca²⁺ in particular), oxidation state, or protein composition. Because the secretory environment is distinctly different from the nucleus, we predict that the G2H will be able to detect many protein-protein interactions that the classical, transcription-based Y2H cannot. Conversely, we expect that many interactions that are readily detected by the classical Y2H will be inaccessible to the G2H. So far, the only case where the G2H assay failed to detect a well-characterized interaction was the p53:SV40TAg complex (data not shown). We speculate that the dodecameric structure of the SV40TAg:p53 complex may be incompatible with the topology of the LOC and CAT fusions or that the high molecular weight complex interferes with correct trafficking of the Och1p fusion proteins, a phenomenon that was observed when another large oligomeric complex was ectopically localized to the secretory pathway. If one of these hypotheses is correct, modifications to the G2H may be necessary to adapt it to analyses of proteins that oligomerize into very high molecular weight complexes. For example, inserting longer linkers between LOC and CAT domains and the bait and prey proteins may accommodate complex interaction geometry. Alternatively, using an ER-resident glycosyltransferase, rather than a Golgi-resident one, may enable detection of complexes that cannot exit the ER. The experiments presented here describe a qualitative relationship between protein-protein interaction affinity and the signal observed in the G2H assay: the interaction between MyoD and Id2 produces strong signals, while introduction of mutations designed to disrupt this interaction decreases the strength of the phenotypic readouts. More comprehensive analysis will be needed to determine whether the signals observed in the growth and WGA binding assays are directly correlated with interaction affinity. A systematic analysis of interactions with varying affinities will enable us to answer this question and to assess the full dynamic range of this new assay. The use of a glycosyltransferase, rather than a transcription factor, as a reporter enables new screening and selection methods. The Och1p reporter system described here relies on phenotypic changes observed in *och1*Δ yeast. The growth assay is simple to implement and its sensitivity can be adjusted by altering the concentration of Congo red. The WGA binding assay sensitively detects different levels of Och1p activity and, in principle, could be incorporated into a fluorescence-activated cell sorting (FACS) experiment to separate yeast with an active Och1p from those in which the protein is not reassembled. By relying on a glycosyltransferase reporter, we envision that the G2H could also be adapted for use in other eukaryotic cells; all that is required is a modular reporter glycosyltransferase that causes a measurable cell surface change. Large families of glycosyltransferases occur in all eukaryotes, with 171 of these enzymes identified in humans. In addition the yeast *Pichia pastoris* has recently been engineered to have human-like glycosylation patterns and may have a secretory pathway better suited to discovering novel mammalian secretome protein-protein interactions. For example, one could imagine using a G2H assay that incorporates human-like glycosylation to discover protein ligands for orphan cell surface receptors. Indeed, interactions among extracellular and cell surface proteins are poorly represented in existing protein-protein interaction databases and new methods are needed to enable their discovery. More broadly, glycosyltransferase activity has the potential to be more widely exploited for screening and selection experiments. The utility of glycosyltransferases stems from two key features. First, they are modular enzymes that can be reassembled from their component parts. Second, they have the ability to provide an extracellular report of intracellular events: the activity of secretory pathway glycosyltransferases occurs within the cell, but results in dramatic changes on the cell surface. In the same way that the transcription-based Y2H assay has been adapted to new uses, such as the discovery of protease substrates and of protein-protein interactions that depend on post-translational modifications such as acetylation and phosphorylation, we anticipate the G2H has the potential to be used to report on biological events beyond simple protein-protein recognition. # Materials and Methods ## Strains, plasmids, and growth conditions *Saccharomyces cerevisiae* strains (MATa, background BY4741) were purchased from Open Biosystems. *S. cerevisiae* strains were grown on yeast extract, peptone and dextrose (YEPD) or on synthetic dextrose medium lacking leucine and uracil (CSM-Leu-Ura; MP Biomedicals). Detailed methods for plasmid construction are described in. Primers used are shown in, and. ## Yeast transformation All plasmids used are listed in. Plasmids were transformed into *och1*Δ yeast (background strain BY4741, MATa, his3Δ1, leu2Δ0, met15Δ0, ura3Δ0) or wild-type yeast (BY4741 MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) by the lithium acetate/SS carrier DNA/PEG method and selected on CSM-Leu-Ura plates. We used yeast colony PCR protocol to verify transformation. A single colony was transferred to 2 µL of sterile water in a microcentrifuge tube for each PCR reaction. The tubes were microwaved for 30 seconds and the contents used as a template for PCR. PCR cycling conditions were step 1: 94°C for 5 min; step 2: 94°C for 45 s; step 3: 60°C for 1 min; step 4: 72°C for 2 min; step 5: repeat steps 2-4 35 times; step 6: 72°C for 10 min. The yeast strains used are listed in. ## Yeast growth assays on agar plates Growth of yeast strains on agar-based growth medium was scored by dilution plating. Yeast strains were grown in liquid culture at 30°C for two days in YEPD or SD-Leu-Ura, then standardized to an optical density at 600 nm (OD₆₀₀) of 1.5. Strains were serially diluted ten-fold in media in a 96-well plate and then transferred to agar plates supplemented with or without Congo red (100 mg/L for YEPD plates; between 2.5 and 6 mg/L for CSM-Leu-Ura plates) using an inoculating manifold. Plates were incubated at 30°C or 37°C for 2 or 3 days and then imaged using an Alpha Innotech FluorChem HD2 photodocumentation system. ## Yeast growth assays in liquid culture The ability of yeast strains to grow at 30°C or 37°C in liquid YEPD was measured by optical density. Starter cultures of wt, *och1*Δ, and *och1*Δ + *OCH1* were grown at 30°C for two days in YEPD. Fresh YEPD cultures were then inoculated (OD₆₀₀ = 0.01) and grown for 3 days at 30°C or 37°C. The OD₆₀₀ of each culture was measured in triplicate once daily for three days. Data were plotted as the average OD₆₀₀ of each culture, with error bars representing +/− one standard deviation. ## Flow cytometry assay Yeast cell wall chitin was detected using the lectin wheat germ agglutinin (WGA) and analyzed by flow cytometry. Yeast strains were grown at 30°C for two days in CSM-Leu-Ura, diluted to adjust OD₆₀₀ to 0.4, then aliquoted into a 96-well plate with conical bottom (3 wells per strain; 200 µL per well). The yeast were pelleted (1200 *g* for 3 min in a tabletop centrifuge) and washed twice with 100 µL of FACS buffer (0.9% NaCl solution). Yeast were then incubated in 100 µL FACS buffer containing 0.01 or 0.02 mg/mL WGA-Alexa 488 (Invitrogen) for one hour at room temperature in the dark. After the incubation, yeast were washed three times with 100 µL of FACS buffer, then resuspended in 200 µL of FACS buffer and placed on ice. Flow cytometry experiments were performed on a FACSCalibur (BD Biosciences) instrument gating on 10,000 live cells per replicate. Data were analyzed using FlowJo software (Tree Star, Inc.). Representative data from each experiment are shown. ## Enrichment experiment A small library was prepared by mixing together approximately equal quantities of seven *och1*Δ strains that were co-transformed with p426-Id2-CAT and one of seven p425-LOC plasmids: p425-LOC-MyoD, p425-LOC-SV40TAg, p425-LOC-Gal80, p425-LOC-Gal11, p425-LOC-Hap5, p425-LOC-Rpt6, and p425-LOC. The mix, having a starting OD₆₀₀ of 0.2, was cultured in 400 mL CSM-Leu-Ura containing 10 mg/L Congo red. Yeast were harvested at 0, 8, 24, 32, 50, and 72 hr. Plasmid DNA was isolated the Wizard Plus SV Minipreps DNA Purification System (Promega), with the following modification to the manufacturer's protocol: after resuspending the cell pellet in the cell resuspension solution, the same volume acid-washed glass beads (Sigma) were added and mixture was vortexed for 3 min. The concentrations of plasmid DNA isolated for each time point were measured by UV absorbance. Each qPCR reaction contained 10 µL of 2X iQ SYBR green Supermix (Bio-Rad), 1 µL of a 10 µM forward/reverse primer mix, 11.5 ng plasmid DNA and DNase/RNase-free water for a final volume of 20 µL. Cycling conditions were as follows: 95°C for 3.5 min, then 40 repeats of the following steps: 95°C for 30 sec, 60°C for 45 sec and 72°C for 2 min. SYBR green fluorescence was detected with a BioRad MyiQ2 Two-Color Real-Time PCR Detection System. Melting curves were obtained from 55°C to 98°C, with fluorescence measurements taken at every 0.5°C increase in temperature. Copy number of each plasmid was calculated by the iQ5 optical system software. Standard curves for primers and plasmids were obtained by 10-fold dilution of plasmids starting from 3,000,000 copies down to 300 copies. All reactions were carried out in triplicate and a non-template control was performed in each analysis. # Supporting Information We are grateful for the use of the UT Southwestern flow cytometry facility. We thank Jennifer Cochran (Stanford University) for prepro-encoding DNA. [^1]: Conceived and designed the experiments: DHD BL EJG JJK. Performed the experiments: DHD BL JJK. Analyzed the data: DHD BL JJK. Contributed reagents/materials/analysis tools: DHD BL EJG SN AKR. Wrote the paper: DHD BL JJK. [^2]: The authors have declared that no competing interests exist.

# Introduction Determination of unknown species is of interest in many areas of applied genetics. For example, in forensic genetics, determination of species can be used to aid police investigations for the identification of forensic stains, solve cases of poaching and aid the protection against trading with endangered species. In food industry the authenticity of meat or fish species can be monitored and in archeology human remains can be sorted from non-human remains. The majority of existing DNA typing methods for species determination focuses on single-species DNA sources and is based on PCR amplification using species- specific primers. There are, however, many instances were no a priori information about the species is available. For such cases a universal typing method would be valuable, especially if the method has the possibility to detect several species in a mixed source. Example of such issues could be bite marks, meat or fish in food or in fact any contaminated samples where the source in minority is of interest. During the recent years there has been an enormous progress in DNA sequencing technologies. These so called next generation sequencing (NGS) technologies have by massively parallel and clonal sequencing increased the ability to gain sequence information from single molecules within a complex or degraded DNA source. These new technologies have, for example, been used for large genome sequencing projects such as the typing of the Neanderthal and the woolly mammoth genomes. NGS technologies have also been used in more specified projects either using a metagenomic shot gun sequencing approach or a more targeted approach. In applications for species identification, Coghlan and colleagues recently presented a cost-effective and efficient method using a next generation deep sequencing technology for the identification of plants and animals in traditional Chinese medicines. The above mentioned studies have been performed at research laboratories with access to research-intensive high-throughput sequencing technologies. The recent and fast development of bench top next generation sequencers like the 454 GS Junior (Roche), Ion Torrent (Promega) and MiSeq (Illumina) have made NGS technologies applicable and affordable for people working in different fields of applied genetics. Encouraged by these new technologies, we present a DNA-typing method for the determination of mammal species using targeted massively parallel sequencing of a short mitochondrial DNA (mtDNA) sequence utilizing the benefits of the NGS bench top technologies. Some of the major advantages with the method are: 1) Short PCR amplicon, i.e. facilitating the analysis of degraded and poor DNA sample quality; 2) Universal PCR-primers, i.e. there is no need for a priori species information; 3) Economy, i.e. the high capacity for sample multiplexing reduces cost per sample; and 4) Deep sequencing, i.e. the possibility to detect DNA present in minute amounts. The outline of the paper is as follows. The design of the target sequence is presented including a phylogenetic analysis of mitochondrial DNA sequences from more than 300 species belonging to the Class Mammalia After this in silico design, the selected universal PCR primers were tested in vitro for their ability to generate PCR-amplicons with DNA from several different species. Experiments were furthermore performed on artificially mixed DNA samples using the NGS technology 454 GS Junior (Roche) to detect all species within an artificial mixture. Finally, the method was tested on two samples from authentic forensic casework. # Materials and Methods ## Design of the target and phylogenetic analyses The complete genome of the mitochondria for 334 different species belonging to the Class Mammalia were downloaded from GenBank ([<u>http://www.ncbi.nlm.nih.gov /nuccore</u>](http://www.ncbi.nlm.nih.gov/nuccore)) and the DNA sequences for the 16S ribosomal RNA gene were extracted using the FeatureExtract 1.2 Server ([ <u>http://www.cbs.dtu.dk/services/FeatureExtract</u>](http://www.cbs.dtu.dk/serv ices/featureextract)). Multiple alignment was performed by the means of MAFFT ([<u>http://mafft.cbrc.jp</u>](http://mafft.cbrc.jp)) using the default settings. From the resulting alignment, the sequences for the universal PCR primers, designed previously, were extracted together with the DNA sequence between the forward and reverse primer sites (i.e. the target region). A distance matrix was computed by pairwise comparison of the DNA sequences between the primer sites for all 334 mammal species, using MEGA v. 5.1. In order to visualize the ability to separate different mammal species, a tree (UPGMA) was created using MEGA v. 5.1. The tree was constructed based on the total number of nucleotide differences for the complete target region between the 334 mammal reference sequences. Note that only one sequence was included for each species and thus intraspecific variability was not considered (see the discussion for further comments on this). ## Verification tests A number of different analyses were performed on DNA samples of known origin and DNA samples from authentic forensic casework to test the method and its ability to be applied in routine: 1\) In order to verify the specificity of the universal PCR primers, DNA from 15 different species of known origin was amplified and sequenced, individually with Sanger sequencing as described below. The samples were provided by The Swedish National Veterinary Institute and The Swedish Museum of Natural History. 2) Twenty two-species DNA mixtures and one four-species DNA mixture, with known DNA template copy number ratios, were analyzed with deep sequencing using GS 454 Junior (Roche) as described in detail below. 3) The 454 deep sequencing method was applied on five samples from authentic forensic casework. The task was to detect the minor component of assumed DNA mixtures in all samples. The first sample was taken from a case with a man accused of animal cruelty of a dog. The question was if DNA from dog could be detected in the trace sampled from the man. The remaining samples were from a case with a human corpse that had multiple bite marks of an unknown “attacker”; here the question was “which species made the bite marks?” ## DNA extraction, mtDNA copy quantification and artificial mixtures For the reference samples, DNA were extracted based on a previously described method, and the DNA from the two authentic case samples was extracted using a Chelex method. The number of mtDNA copies in a DNA extraction was quantified using an in-house developed assay based on quantitative real time PCR. The quantification method utilized SYBR Green I chemistry using the same primer pair as for the 454 PCR amplification protocol described below. Each PCR reaction contained 12.5 μl 2x QuantiFast SYBR Green PCR Master Mix (Qiagen, UK), 1 µl of each primer (800 nM final conc.), 8.5 μl RNase-free water and 2 μl of DNA template for a total reaction volume of 25 μl. Thermal cycling was performed in an iQ™5 multicolor Real-Time PCR Detection System (Bio-Rad) using conditions recommended in the QuantiFast protocol (Qiagen). Plasmids (GenExpress, Germany), with a known number of copies, containing the targeted DNA sequence for humans, were used as a standard reference. All samples were quantified in duplicate and the mean value was used. Pooled DNA samples from different species were used to create artificial DNA mixtures either in a 1:1 or in a 99:1 mtDNA copy number ratio. ## Sanger sequencing and 454 sequencing For the Sanger sequencing, the targeted DNA sequence was amplified using the PCR primers, presented in, tailed with M13-adaptors. The PCR amplicons were sequenced by standard Sanger sequencing using BigDye® Direct Cycle Sequencing kit (Applied Biosystems) according to the manufacturer’s protocol. Separation, by capillary electrophoresis, was performed on an ABI3500 XL instrument using POP-7 (Applied Biosystems) and the DNA sequences were called using Sequence analysis software v. 5.4 (Applied Biosystems). For the 454 deep sequencing, the HPLC-purified PCR primers (Biomers) were used with adaptors suitable for the 454 sequencing chemistry on the 454 GS Junior (Roche). Barcodes were included in the primers to aid the ability for sample multiplexing. The DNA was amplified using an in-house developed assay as follows. Each PCR reaction contained GeneAmp 10x PCR Buffer (Applied Biosystems), 3 mM MgCl₂, 0.5 mM Nucleotide+Uracil, 2 % Glycerol, 0.016 % BSA, 1.25 U AmpliTaq Gold DNA polymerase (Applied Biosystems), 0.05 U Uracil-DNA Glycosylase (USB corporation), 800 nM of each primer and 1 μl of DNA template (1,000-10,000 mtDNA molecules), for a total reaction volume of 25 μl. Thermal cycling was performed in a GeneAmp^® PCR System 9700 (Applied Biosystems) with touch-down PCR using the following parameters: 10 min at 37 °C, 5 min at 95 °C, followed by 10 cycles of 30 s at 94 °C, 30 s at 67 °C (minus 1 °C per cycle), 30 s at 72 °C and 30 cycles of 30 s at 94 °C, 30 s at 58 °C, 30 s at 72 °C with the final extension for 10 min at 72 °C. The PCR products were purified using AMPure (Beckman Coulter) following the manufacturer’s protocol. The PCR products were then quantified using KAPA SYPR^® FAST Bio-Rad iCycler qPCR kit (KAPABiosystems) according to the manufacturer’s protocol. Thermal cycling was performed in an iQ™5 multicolor Real-Time PCR Detection System (Bio-Rad) using conditions recommended by the manufacturer. All samples were quantified in duplicates and the final PCR product was diluted to 10⁸ molecules per µl. For the emulsion PCR (emPCR) and sequencing on 454 GS Junior, the manufacturer’s protocols were followed with reduced volume of primers (1/2 of the suggested volume) and the addition of 20 molecules per bead compensating for the short length of the PCR amplicon. ## Bioinformatics Sff-files extracted from the 454 GS Junior instrument were converted to FASTQ- files using the web-based tool Galaxy ([<u>https://main.g2.bx.psu.edu/</u>](https://main.g2.bx.psu.edu/)). Reads with low quality were removed followed by the separation of the reads originating from different PCR:s by filtering on the barcode sequences. The barcodes were further removed, together with primer sequences, from the reads using Tagcleaner ([<u>http://edwards.sdsu.edu/cgi- bin/tagcleaner/tc.cgi</u>](http://edwards.sdsu.edu/cgi-bin/tagcleaner/tc.cgi)). The resulting reads were grouped by collapsing identical reads. Sequences with less then 20 identical reads were removed from further analysis unless otherwise stated. The remaining sequences were searched against sequences in GenBank using BLAST ([<u>http://blast.ncbi.nlm.nih.gov/Blast.cgi</u>](http://blast.ncbi.nlm.ni h.gov/blast.cgi)), in order to identify the origin of the unknown DNA-sequences. Data described herein is available from the Sequence Read Archive (SRA) with the accession number SRA107543. # Results ## Design of the target and phylogenetic analyses The DNA sequences for all 334 mammal species were aligned and the primer site sequences were extracted and compared with the universal primer sequences (modified from). There were high homologies among the reference species for the majority of the nucleotide positions, both for the forward and the reverse primer sequences. Further analysis showed that 97 % of the sequences from the reference species had primer sequences that were identical, or only with one nucleotide difference, to the universal primer. The nucleotides in majority were as follows for the forward primer; gacgagaagaccctatggagC (5’-3’) and tccgaggtcAccccaaccTCCG (5’-3’) for the reverse primer, which are identical to the primer sequences used for the PCR amplification. Thus, based on this phylogenetic study, the selected universal primer sequences should be able to produce PCR amplicons for further analysis for the largest proportion of mammal species. Further analysis showed that the length of the DNA sequence between the primer sites varied among the reference species, with an overall mean of 75 base pairs (SD = 3 bp). The mammal reference DNA sequences were further compared for all species in pairs, to count the number of nucleotides that differed between the sequences of different species. displays the distribution of the number of nucleotide differences for the target amplicon with an overall mean number of pairwise differences of 31.5 (SD = 7). For the purpose of this method it was, however, of most interest to measure the ability to separate different species. Based on the reference sequences from the 334 different mammal species, only 0.06 % of the pairwise comparisons resulted in identical reference sequences. Thus, over 99.9 % of the species could be distinguished using the present assay. Species with identical sequences are listed in. Furthermore, a tree based on the total number of pairwise differences, for the complete target sequence, is shown in. Closely related species tend to cluster, with only a few nucleotide differences for the complete target. ## Verification of the universal primers The ability of the universal PCR primer pair to amplify DNA was tested with samples from a variety of different species using PCR amplification followed by Sanger sequencing. DNA sequences were successfully obtained for all 15 mammal species tested. These sequences were confirmed to match the corresponding reference sequences in GenBank. ## Species determination of mixed samples presents the result from experiments with artificial mixed DNA-samples of either 1:1 or 1:99 mtDNA copy number ratios. All species were successfully detected in the 1:1 mixtures, while for some of the 1:99 mixtures, only the DNA from the species in majority was observed. For the non-human DNA mixtures, the ratio of the obtained reads was similar to the expected ratio, although large variations were seen. The mixtures that included DNA from human resulted, however, in a decreased number of human reads, approximately with a factor of 10-20 times in comparison with the expected ratio. ## Error rates Since the reference sequences were known, experiments were performed in order to estimate error rates based on the DNA sequences obtained from the analyses. In the first test with DNA from elk, *Alces alces*, a total of 11,683 reads were obtained. 10,742 of these reads (89.6 %) were identical with the reference sequence for elk, whereas the remaining 1,211 reads (10.4 %) consisted of 211 sequence variants. The mean length of these sequences was 72 bases, thus the error rate was estimated to be 0.0014 errors per base. The error rate is known to be sequence specific, especially for sequences containing homo polymers when using 454 sequencing chemistry. The human, *Homo sapiens*, target reference sequence contains a stretch of five adenosine bases. The second experiment, with DNA from human, resulted in 5,268 reads of which only 66 % were identical with the reference human sequence and out of the erroneous sequences 78 % consisted of a sequence variant with four adenosine bases. In total, the error rate was estimated to 0.0047 per base, thus almost five times higher than for the elk experiment. As discussed in more detail below single nucleotide errors are, however, not important for the purpose of the method described here since the complete sequence of the target amplicon is used for the search and matching. ## Samples from authentic forensic casework In the first authentic case, which consisted of a man accused of animal cruelty, DNA from both human, *Homo sapiens*, and the genus *Canis* (including the subspecies domestic dog and wolf) was detected in the sample ( In the second case, the four different samplings from bite marks, were analyzed in one single 454 sequencing run with equal input of PCR amplicons. DNA from human, *Homo sapiens*, was detected in all samples and DNA identical to the reference sequence for the genus *Canis* was additionally found in two of the samples. The presence of DNA from *Canis* was consistent with other findings in both cases. However, a notable observation was that the presented deep sequencing method enabled detection of small proportions of the minor component (\<1 %). # Discussion Our main goal with this work was to develop a universal method for species determination including the ability to detect all species in samples of mixed DNA from different mammal species. When it comes to species determination, a large number of different DNA typing methods and case issues have been presented, see for a review. The NGS method, with the universal target approach, that we here present has several advantages compared to other typing methods. For example a metagenomic approach might be too inefficient, for the purposes of the method described here, if the analyzed sample contains a major proportion of sequences from bacteria, fungi, algae or unknown species origin. Methods that are species-specific could be useful when having information of what to look for, especially in mixed samples where the species of interest is present in very small amounts. In most contemporary applications one-species-specific methodologies do not, however, allow for simultaneous detection of different DNA components in a mixture. In these methods there must be an a priori hypothesis of what to look for. If casework requires the ability to detect a large variety of different species, specific assays for each species would be needed, making the task tedious, costly and not always applicable compared with having a universal method. There are, however, species-specific methods for the identification of species within a mixture. Tobe et al. designed and validated an assay to identify 18 European mammal species from mixtures. Although limited to a number of specified species, the assay offers both a fast and inexpensive alternative to the herein presented assay, especially when prior information of the unknown species might be available. A universal approach, on the other hand, has the major advantage that no a priori knowledge is needed for a case. Such a method does, however, require more work in designing the target sequence and additional tests to verify that the method really is capable of detecting all the species it is aimed for. Both the phylogenetic study of reference sequences and the verification tests of DNA of known origin showed that the method presented in this paper is suitable for the vast majority of mammal species. Even though the choice of target (16S rRNA) is less variable than other commonly used genes (e.g. cytochrome oxidase 1 and cytochrome b), the target has earlier proved to be useful for species identification. For our assay, the target satisfied the need for homologous primer sites enclosing a short variable region of only approximately 100 base pairs. The short target sequence is well suited for analysis of degraded DNA which has been shown in previous work on artificially degraded samples. This is often the case, since the material in question might come from processed animal tissue, as in food, or from forensic stains destroyed by environmental conditions, or archaeological findings. Too short sequences could, however, give less phylogenetic resolution and further tests might be needed in order to verify the method’s validity for species not included in this study. By redesigning the PCR primers, for example by utilising degenerated primers, the method described herein could be expanded to work for additional species to be detected. Phylogenetic studies (data not shown) of reference DNA sequences from almost 1,600 species belonging to the taxa *Vertebrate* showed for example that approximately 80 % of the species had differences of two nucleotides or less in the primer sequence regions. The target sequence was, for most instances, proven to be variable enough to resolve the vast majority of the 334 reference mammal species used in the phylogenetic analysis 1 should, however, note that the analysis performed did not take any intraspecific variation into account. Tobe et al. previously showed, in a comprehensive study, the relationship between the interspecific and the intraspecific variation for two other commonly used target genes in mtDNA. From our phylogenetic studies the degree of interspecific variation revealed that the chosen target might be sensitive for intraspecific variation for some of the closely related species (for example species belonging to the genus *Cervus*). Thus, further phylogenetic studies of the intraspecific variation are advised to maximize the confidence of a given hit at the species or subspecies level. The main advantage using NGS deep sequencing is the ability to individually sequence each PCR amplicon at a certain depth. This increases the ability to detect traces of DNA within a mixture of species present only in small amounts. In our experimental setup of artificial DNA mixtures with different ratios, both species were always detected for the 1:1 copy number ratios. The minor component of only 1 % could be detected in all cases where the minor component was of non- human origin. Our tests indicated that the human DNA sequence was less efficiently (tenfold) detected in a mixture. The reason for this was not clear but the sequencing part of the analysis can be excluded since we obtain maximum (or very close to) number of reads from those analyses. Also the initial PCR amplification can be excluded since analysis of mixtures using standard pyrosequencing result in pyrograms with expected peak heights. This leaves us with the emPCR step, which hypothetically, for unknown reasons, does not amplify the human sequences as efficiently as the other sequences when present in a mixture. Although this is a technical drawback it can actually be regarded as an advantage in casework where the human component might disturb the detection of a minor component of another species but not vice versa. The analysis of the samples from the two forensic cases indicated that the method also works well for authentic case samples and not only for artificially “clean” samples. In both cases DNA from *Canis* could be detected, which was as expected given other information in the cases. In general our experiments show that using quality requirements of no less than 20 reads for the minor DNA component is robust enough to detect the minor component down to as little as 1 % of the total DNA. Thus the method is sensitive enough to also be vulnerable to contamination. Strict routines for separating the pre- and emPCR amplification are needed and the use of dedicated barcode tags for each case can aid in monitoring possible run to run contamination. Further studies are, however, needed to monitor background levels for casework samples and to set up analysis parameters for the bioinformatic evaluation of the massive sequence data. For the latter, one must address issues of separating a true sequence variant (e.g. heteroplasmy) from a false sequence variant (e.g. amplification error, sequencing error). For the application discussed here, the complete sequence of the target is used for the identification of species, thus single base errors are not vital unless a presumable mixture contains DNA from closely related species. In summary, we present a universal method for species identification of mammals using a targeted massively parallel sequencing approach. Both phylogenetic studies and sequencing experiments confirm the specificity of the universal primer set. Minor DNA components down to 1 % were shown to be detectable in species mixtures using the deep 454 sequencing method. Although promising results were obtained with the current settings, the rapid development of bench top instruments will further improve the method with less “hands-on”, lower detection limit and fewer sequencing errors. # Supporting Information The authors wish to thank Prof Peter Söderkvist for providing laboratory equipment for this study and Ankin Güvencel for providing samples from forensic casework. We also thank the anonymous reviewers for their constructive comments. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AOT BD JW GH. Performed the experiments: AOT BD JW. Analyzed the data: AOT. Wrote the manuscript: AOT GH.

# Introduction Aggressive behaviour (AB), as observed in social disorders such as DBD (including conduct (CD) and oppositional defiant disorder (ODD)), is characterized by a repeated pattern of antisocial behaviour and severe aggression, where the basic rights of others, major age-appropriate norms or societal rules are violated. Such problems can cause significant impairment in social, academic, or occupational functioning. Clinical and subclinical forms of AB are observed in up to 14% of all girls and 16% of all boys. The negative impact of aggression-related problems reaches beyond a patient’s family, ultimately affecting society as a whole (e.g. school-dropouts, delinquency, teen-pregnancies, substance abuse or difficulties integrating into work life). Early conduct problems are key precursors of persistent AB and thus also predictive for ODD, CD and antisocial personality disorder in adulthood. Neurodevelopmental theories and longitudinal studies are in line with these behavioural observations, suggesting that the presence of early brain alterations in individuals with aggressive behaviour may heighten the risk for long-lasting social impairments. In the current paper we particularly focus on adolescents with *aggressive behaviour* (AB), hereby summarizing neuroimaging research in youths with either conduct problems, CD or ODD. In recent years structural (e.g voxel-based/surface-based) and functional (e.g. fMRI/PET) neuroimaging techniques have grown into powerful tools to investigate the neuronal basis of the human brain in typically developing individuals as well as patients. It has been demonstrated that both, brain structure and function, may be modified by experience. Activation-dependant structural plasticity can even occur after as little as seven days of training and it is suggested to play a key role in human adaptation to environmental changes and disease. Even though neuroimaging evidence points toward a neuronal basis of AB, the overall number of research studies within this population remains relatively scarce. Furthermore, it has to be noted that AB characteristics as seen in CD and/or ODD are considered heterogeneous in respect to their pathologies. CD and ODD are frequently associated with comorbidities such as attention-deficit hyperactivity disorder (ADHD) or anxiety). These comorbid disorders can differ in their pathophysiological mechanisms, some of them seem exclusive on a biological level making it possible that different developmental trajectories with varying neurobiological bases lead to the clinical manifestations of AB. The vagueness of the group definition within many of the current studies on AB is thus bound to impact general conclusions drawn from it. Even though the total number of studies is still limited, neuroanatomical and functional variations in youths with AB have been reported with increased frequency since the advent of modern neuroimaging. In particular, brain structure in AB has been investigated using voxel-based morphometry (VBM), diffusion tensor imaging (DTI) or surfaced-based morphometry. VBM studies for example have revealed differences in gray and white matter volume in brain regions including the amygdala, insula, orbitofrontal and dorsomedial prefrontal cortex when comparing adolescents with AB and typically developing controls. Similarly, studies using surface-based morphometry or DTI provide evidence for structural alterations and/or impaired connectivity within brain regions involved in emotion processing, reward and empathy. Functional neuroimaging studies corroborate the structural neuroimaging literature. Cognitive paradigms employed in the investigation of AB have focused on disturbances in the emotion processing and regulation network of the brain. These tasks particularly target emotion processing/regulation, empathy, theory of mind, passive avoidance, decision making or executive functioning. Overall, studies point towards aberrant brain function in AB in key areas of social cognition and emotion, including prefrontal (orbitofrontal, dorsolateral and medial prefrontal cortex), limbic (e.g. amygdala, anterior insula, cingulate cortex) and temporal cortices. Despite increasing evidence about the uniformity of atypical brain structure and function in AB, it has yet to be objectively determined which brain regions are commonly affected. Functional and structural neuroimaging studies are crucial for the understanding of the phenotype and aetiology of AB. However, most results and interpretations are based on individual neuroimaging studies and present various limitations (e.g. small sample sizes, low reliability, dependency on task chosen). Furthermore, very few imaging studies have yet investigated brain structure and function in the same population. Activation likelihood estimation (ALE) meta-analyses allow the identification of consistent findings of brain activation and structure across multiple data sets. Hereby, ALE quantitatively investigates communalities between reported foci based on modelling them as probability distributions centered around the corresponding coordinates. The resulting probability maps mirror the likelihood of morphological change and/or activation on a voxel-wise level across an entire set of studies. ALE has been successfully applied in meta-analyses of various neuropsychiatric disorders to date and provides a promising tool for a more unified investigation of pathophysiologic changes in disease. Therefore, the present paper intends to close this gap in research and aims to aggregate all structural and functional neuroimaging studies conducted in adolescent AB to date. In a first step, we planned to conduct a systematic literature review of neuroimaging findings in adolescents with AB. Secondly two separate meta-analyses looking at gray matter volume reductions as well as hypoactivations during emotion processing tasks in AB were carried out. Finally, we decided to run a conjunction analysis to identify potential overlaps in deviant brain structure and function in adolescents with AB. # Method ## Participants We decided to focus our analysis on adolescents with *aggressive behaviour* (AB) in general as opposed to a specific clinical diagnosis. By including both community samples and clinical samples in the present meta-analyses we adhere to the heterogeneity in juvenile aggression. This heterogeneity is further reflected by different behavioural symptoms of aggression and antisocial tendencies, such as oppositional behaviour, impulsive hot-tempered quarrels or premeditated violent acts, the presence of callous unemotional/psychopathic traits or co-morbid conditions in CD and ODD patients. All studies were conducted during childhood and/or adolescence and share the communality of aggression and antisocial tendencies within the populations studied. Thus, AB as defined here may be considered an umbrella term for children and adolescents with a range of subclinical and clinically relevant symptoms of pathological aggression. ## Study Selection For the structural and functional neuroimaging meta-analyses we used PubMed and Google Scholar to systematically search for neuroimaging literature in AB. Literature searches were conducted and reviewed by several research team members (NMR, WMM, LVF, ET) and adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA;) guidelines and the revised Quality Of Reporting Of Meta-analyses (QUOROM) statement. Our main search conducted through PubMed included the following key words: *“conduct disorder”*, *“conduct problems”*, *“disruptive behaviour disorder”*, *“oppositional defiant disorder”* and *“aggression”*, each in combination with methodologically relevant terms including *“VBM”*, *“fMRI”* and/or *“neuroimaging”*. Moreover, a number of review articles published on conduct disorder, antisocial behaviour and aggression in adolescents were considered. Finally, additional publications were explored by searching the reference list of the articles obtained to assure integration of all data available. Studies were included in our meta-analyses if the following criteria were given: (**I**) included at least one clinical group with described aggressive behaviour, (**II**) in combination with a healthy control sample, (**III**) conducted during adolescence, (**IV**) reported whole brain gray matter volume alterations or whole brain functional neuroimaging data, (**V**) results are described using a standard reference space (Talairach or MNI) and (**VI**) the same threshold was used throughout the whole brain analysis. All structural studies included employed a standard VBM analysis protocol. In both meta-analysis of structural and functional brain alterations in adolescents with AB versus controls, no studies providing results based on a priori region-of-interest analysis only were included (since they violate the assumption, under the null hypothesis, that the likelihood of locating activated foci is equal at every voxel). Similarly, no animal studies or case reports were included in any meta-analysis and only studies from peer-reviewed journals that are written in English were considered. Data is current up to July 2015. Of the 1021 studies identified through our systematic review, we screened 930 (after removal of duplicates) and consequently assessed the full texts of 173 articles. 156 studies had to be excluded from the functional or structural meta- analysis in adolescents with AB, because they did not meet the criteria listed above (for detailed exclusion reasons, see). Looking more closely at our review on <u>*structural research studies*</u> in AB revealed that only five studies reported on gray matter volume increases in AB (four reported de- and increases, one study only reported increases). Therefore we did not conduct a separate meta-analysis for gray matter volume increases in AB. Consequently, eight studies were included in our meta-analysis about gray matter volume reductions, together reporting data from 408 research participants (224 AB, 184 typically developing controls = TD), and 50 foci of gray matter volume decreases in youths with AB. Our systematic literature review of <u>*functional neuroimaging studies*</u> in youths with AB identified experiments targeting emotion processing, empathy, theory of mind, passive avoidance, decision making or executive functioning. We decided to restrict our functional meta-analysis to tasks only including emotionally loaded and visually presented stimuli (e.g. tasks of emotion processing and empathy). In case of sample overlap, the study with the highest subject number meeting all other criteria listed above was selected. In case of comparisons between AB and TD in more than one contrast, only foci from the contrast putting the highest demand on emotion processing, were included. The majority of studies indicated hypoactivations in AB. Only six studies that fulfilled all other criteria listed above reported hyperactivations in AB compared to TD. Therefore, we did not conduct a separate meta-analysis on functional overactivations in AB. Consequently nine studies suggesting hypoactivations in adolescents with AB compared to TD were selected. Together the selected studies report data from 375 research participants (215 AB, 160 TD) and describe 58 foci of hypoactivation in AB compared to TD. ## ALE Meta-Analysis Procedure We conducted two separate meta-analyses on gray matter volume alterations and functional hypoactivations in adolescents with AB. Data analysis was carried out using the revised version of the ALE approach for coordinate-based meta-analysis of neuroimaging data (GingerALE software, version 2.3; available from <http://brainmap.org/ale/>). In short, this new approach implements a random- effects model, a quantitative uncertainty model to determine the FWHM and an exclusive gray matter mask (for further details, see also). Most importantly, instead of testing for an above-chance clustering between foci, the revised ALE algorithm assesses above-chance clustering between experiments. The spatial relationship between foci in a given experiment is now assumed to be fixed and ALE results are assessed against a nulldistribution of random spatial association between experiments. Prior to running any analyses, coordinates reported in Talairach space were transformed to MNI space using the tal2icbm algorithm. The here employed revised ALE approach identifies areas of convergence of activation across various experiments, minimizing the within- groups effects (approach by Turkeltaub and colleagues). Each focus is represented as a centre for 3D Gaussian probability distributions, where the standard deviation depends on group size (capturing spatial uncertainty) rather than single time points. First, the probabilities of all activation foci in a given experiment are combined for each voxel, which is represented in modelled activation maps (fMRI) or modelled anatomical maps (VBM). Secondly, the ALE method combines all modelled maps (fMRI and VBM separately) on a voxel-by-voxel basis to form an ALE image containing all unthresholded voxel ALE values. In the last step, this ALE image is tested against the null hypothesis under the assumption that all activated voxels are homogeneously distributed in the brain, independent of the experiments. This null-hypothesis model (a distribution map made by multiple permutations of random voxel activation) was created using a random-effects statistical method and tested against the original ALE image according to the selected significance threshold. Therefore, the nulldistribution is constructed reflecting a random spatial association between different studies. Comparing the “true” ALE score to this distribution allows a focused inference on convergence between studies while preserving the relationship between individual foci within each study. Critically, this change from fixed- (foci-based) to random-effects (testing between study effects) inference in ALE analysis allows generalisation of the results to the entire population of studies from which the analysed ones were drawn. This more conservative approach with an increased specificity does also accommodate the idea of convergence across heterogeneous studies. We used a statistical threshold of p\<0.05 False Discovery Rate (FDR) corrected for multiple comparisons and a minimum cluster size of 500mm³. ALE maps are overlaid onto a standard brain in MNI space (Colin27 available at <http://www.brainmap.org/ale/>) using the Multi-image Analysis GUI (Mango available at <http://ric.uthscsa.edu/mango/mango.html>) and clusters were anatomically labelled by cross-referencing the Talairach Daemon and aal. In order to further investigate possible overlaps between the structural (VBM) and functional (fMRI) meta-analysis in adolescent AB, a formal conjunction analysis was performed by multiplying binarized versions of the individually thresholded ALE maps. # Results Our meta-analysis of <u>*structural neuroimaging studies*</u> in adolescents with AB revealed 19 clusters of significant convergence between the studies. The largest clusters were found in the right inferior frontal lobe (inferior frontal/precentral gyrus), right precuneus and left-hemispheric insula. Further smaller clusters were found bilaterally in the frontal (e.g. dorsolateral and medial frontal gyrus), parietal (e.g. precuneus) and temporal lobe (e.g. middle/superior temporal gyrus) as well as the cerebellum (e.g. culmen). Our meta-analysis of <u>*functional hypoactivation*</u> in adolescents with AB revealed 8 clusters of significant convergence between the studies with the largest clusters in the right middle/superior frontal gyrus, left thalamus and basal ganglia, as well as left-hemispheric insula. Beyond others, further clusters included the right anterior cingulate, left middle temporal gyrus and right amygdala. A formal conjunction analysis using the thresholded ALE maps from the structural and functional meta-analysis discovered three areas of regional overlap (**)**. The biggest area of functional and structural overlap (128mm³) in adolescents with AB was identified within the right dmPFC. Additionally, the analysis exposed two smaller, close-lying clusters of convergence with a peak in the left claustrum, extending into the insular cortex. # Discussion To our knowledge, the current work provides the first quantitative summary of functional hypoactivations and gray matter volume reductions in adolescents with AB by summarizing findings of eight structural and nine functional neuroimaging studies in a total of 783 participants (408 \[224 AB/184 TD\] and 375 \[215 AB/160 TD\] for structural and functional analysis respectively). Our findings indicate 19 structural and eight functional foci of significant alterations in AB, mainly located within the emotion processing and regulation network of the human brain (including orbitofrontal, dorsolateral/medial prefrontal cortex and limbic brain regions; for reviews on emotion processing and regulation see also). Conjunction analysis reveal that functional and structural alterations in AB overlap in three areas, with the largest cluster centered in the right dmPFC and two smaller clusters that encompass the left insula. In the following sections we will review structural and functional neuroanatomical evidence derived from healthy participants as well as those with aggressive behaviour (e.g. conduct problems, CD, ODD) for the key areas implicated here (orbitofrontal and dorsomedial prefrontal cortex, insula, cingulate cortex, amygdala). ## Orbitofrontal and Dorsomedial Prefrontal Cortex Our findings identify prefrontal brain regions including orbitofrontal and dorsomedial prefrontal cortex as main locations of aberrant brain function and structure in youths with AB. Furthermore, an overlap in the foci representing structural and functional changes that co-localize in AB is centered in the right dmPFC. While the orbitofrontal as well as the dorsomedial prefrontal cortex can be differentiated based on quantitative as well as qualitative markers, both have equally been suggested in emotion processing and working memory/inhibitory control. The medial prefrontal cortex in particular has been implicated in emotional self-regulation, general self-referential activities and emotion-related decision making. Meta-analytic evidence suggests a more generic role of the dmPFC in emotion processing (e.g. appraisal, evaluation, experience, response), non-specific to a particular emotion. In addition, lesion, neurophysiological and neuroimaging evidence have linked the orbitofrontal and dorsomedial prefrontal cortex to stimulus-reinforcement association learning. The ability to rapidly decode and readjust values of different input signals is likely to be crucial to emotional behaviour and may ultimately influence emotional learning. It has been suggested that the observed deficits in decision making may directly result from aberrant emotion processing as for example observed after frontal brain damage. Research has for instance demonstrated that aberrant self-monitoring abilities may be responsible to preclude the generation of social emotions typically associated with the resolution of social mistakes. Finally, a whole line of evidence has linked the prefrontal cortex to aggression. In its extreme, antisocial personality disorder and psychopathy are exemplary for individuals displaying increased aggressive behaviour and studies of both have linked structural and functional changes to the prefrontal cortex. ## Insula Both our functional and structural AB meta-analysis have found significant clusters of hypoactivations or altered brain structure within the insula. In addition to that, two smaller clusters reached significance in the left insular cortex during our conjunction analysis, mapping structural and functional alterations in youths with AB. The insula or insular cortex is part of the cerebral cortex forming the base of the lateral sulcus (or sylvian fissure). From a neurodevelopmental perspective it is the first region of the cortex to develop and differentiate around 6 weeks of fetal life. The insula is bi- directionally connected to various brain regions, including the orbitofrontal cortex, anterior cingulate, supplementary motor areas, parietal and temporal cortices, but also to subcortical structures such as the amygdala, basal ganglia and thalamus. Connectivity to and from the insula is divided, in that the anterior part of the insula has greater connectivity with the frontal lobe, while posterior parts are more strongly connected to the parietal lobe. Neuroimaging evidence has suggested that the insula may play a key role in the awareness of bodily sensations and affective feelings. Meta-analytic data supports this idea, and suggests that the insula is a key player in the evaluation, experience or expression of internally generated emotions. Particularly the left insula, along with frontal and temporal brain regions, is associated with anger. Furthermore, an emotion-specific role of the insula for disgust has been discussed. However, the majority of neuroimaging findings and meta-analytic reviews to date support a generic role of the insula in emotional behaviour. Atypical neuronal functioning of the insula (e.g. during tasks of emotion processing and empathy) are linked to AB. However, so far, both hyper- and hypoactivations are observed during tasks of empathy, face or pain processing. In psychopathy particularly fear conditioning has been linked to aberrant insula activation. Functional atypicalities within the insula are further observed in borderline personality disorder, schizophrenia, depression or anorexia nervosa. Gray matter volume alterations within the insula are associated with various psychiatric conditions beyond antisocial populations, including bipolar disorder, schizophrenia, drug dependence, major depression or anorexia nervosa.Therefore, the neuronal and structural alterations within the insula may reflect a characteristic of psychiatric conditions per se. ## Cingulate Cortex The cingulate cortex showed functional as well as structural foci of significance in each of our two meta-analyses individually. Cytoarchitectonically, the cingulate gyrus may be divided into four functionally independent but interconnected subregions, including the anterior cingulate cortex (emotion), the midcingulate cortex (response selection), the posterior cingulate cortex (personal orientation), and the retrosplenial cortex (memory formation and access). Overall the cingulate cortex has been implicated in the regulation of cognitive as well as emotional processes (e.g. processing of acute pain or affective stimulus material), most likely through an interaction with the prefrontal cortex, anterior insula, premotor area, the striatum and cerebellum. We here particularly identified regions within the bilateral anterior cingulate as foci of interest through both our functional and structural meta-analysis. While dorsal aspects of the anterior cingulate have been linked to tasks of executive functioning, the anterior part of the cingulate is part of the emotion processing network. It is further suggested that the cingulate gyrus may serve as a transition and/or interaction zone between affective and cognitive processing. Studies in AB and antisocial personality disorder have found both gray and white matter increases as well as decreases within the cingulate ; the developmental pathway within this region thus still needs further assessment. Hypoactivation in AB within the cingulate has been reported during tasks of emotion processing, empathy, response inhibition and sustained attention. Similarly, individuals with antisocial personality disorder or psychopathic tendencies show reduced activation within the cingulate during tasks of emotion processing and conflict resolution, as for example observed in moral decision making, deception, frustration and emotion processing. ## Amygdala Both our functional and structural meta-analyses have identified the right and left-hemispheric amygdala as significant foci of interest, even though this area has not reached significance in our conjunction analysis. The amygdala is crucial for the perception and encoding of emotionally loaded stimulus material and has been suggested as the brain locus of fear (e.g. detection, generation, maintenance of fear and coordination of response in the danger of such). To summarize the existing fMRI evidence, neuronal activation within the amygdala has been observed in healthy individuals in tasks that include arousing stimulus material (e.g. emotionally loaded images, facial expressions or words), during tasks of empathy, moral reasoning or when processing potential threats). A range of tasks investigating amygdala responses to different evocative stimulus material led to the suggestion that increased activation within the amygdala may particularly mirror affective processing under acute danger or threat, rather than fear per se. Furthermore, neuronal activation is thought to mirror dispositional affective style, whereby increased amygdala activity correlates with affective reactivity to negative stimuli. Interestingly, amygdala activation in response to emotionally loaded stimuli may be attenuated by task demand or comorbid anxiety and depression symptoms. For example, concurrent goal-directed processing can disrupt amygdala activation that is evoked by emotional images. This is in line with meta-analytic evidence indicating that studies employing a cognitive task during affect processing are less likely to demonstrate amygdala activation. Because of its role in aversive conditioning, instrumental learning and fear processing, the amygdala is often chosen as a region of interest in investigations targeting AB, antisocial personality disorder or psychopathy. Amygdala dysfunction is suggested to be one of the core features in the symptomatology of antisocial disorders (e.g.). Structurally, the amygdala is altered in AB similarly as in antisocial personality disorders and psychopathy. Finally, it is to note that the amygdala is strongly interconnected with the orbitofrontal brain regions and alterations in the connectivity between these two centers have been reported in AB and psychopathy (e.g. connectivity between key regions of the emotion processing and regulation network (e.g., for a further discussion see following section). ## Structure-Function Relationship and Connectivity Findings While neuroplasticity is known to potentially range from synaptic plasticity to more complex changes (e.g. shrinkage in cell size, neural or glial cell genesis, spine density or even changes in blood flow or interstitial fluid), the neurophysiological basis of experience-induced neuroplasticity is still a matter of extensive research. Some studies indicate that functional and structural measures of plasticity may be related. For example it could be hypothesized that experience-related gray matter volume changes correspond to task-specific processing, or, more precisely, synaptic remodelling within specific processing areas. Another possibility may be that impaired connectivity between key regions leads to the functional alterations observed. For example researchers have argued that the social and emotional deficits seen in AB may be mediated by impaired connectivity between the emotion processing and regulation network. These system-specific deficits may be observed by diffusion tensor imaging and tractography measurements. For example, the uncinate fasciculus is a white- matter tract connecting the amygdala and neighbouring anterior temporal lobe with the orbitofrontal cortex and it thus may be involved in facilitating empathy, emotion regulation and socio-cognitive processes. Such models would for example explain why local changes in brain structure cannot always be inferred from purely functional models. For example in individuals with reactive aggression aberrant amygdala activity but intact amygdala structure is observed. In such cases it is possible that impaired fibre connections (e.g. reduced functional anisotropy in the uncinate fasciculus) to and from this area cause the neuronal differences observed. In line with evidence in AB significant differences in the fractional anisotropy (FA) measures of the uncinate fasciculus have been demonstrated in adolescents with conduct disorder as well as in adult psychopathy. Similarly, studies of intrinsic connectivity (resting state) explore functional networks that are non-stimulus driven and may inform about the basic functional brain architecture while implicating anatomical connectivity of the regions involved. In individuals with antisocial personality disorder this intrinsic connectivity between highly interconnected brain centres is disrupted. Independent of the precise neurophysiological nature of structure-function associations, our results have indicated co-localized structural and functional deficits in right dmPFC and left insular cortex. Based on today’s structure- function knowledge we thus hypothesize that decreased synaptic density may have led to a co-localized decrease within the BOLD response measured through fMRI. However, it has to be noted that here we only investigate co-localized structure-function findings that are based on gray matter volume reductions and functional hypoactivations in AB. This limitation (no volume increases or hyperactivity investigated) is due to the nature of the existing neuroimaging evidence, with only five studies reporting gray matter volume increases and six studies providing evidence for functional hyperactivations in individuals with AB. Further studies comparing adolescents with AB compared to controls are needed in order to examine functional hypoactivations and gray matter volume increases more extensively. Furthermore, only longitudinal research studies will be able to show the precise developmental trajectory of these alterations in detail. ## Limitations Meta-analytic approaches such as the current one have a number of limitations in need for discussion. The presented analyses are first of all limited by the detail and quality of the original research studies. This includes problems of variations within the significance threshold of data reported, insufficient information on possible coordinate transformations and variation in group sizes. Additionally, even though psychosocial factors have been significantly linked to brain structure in AB, none of the studies to date systematically studied the influence of these within their designs. Furthermore, only a small number of studies to date have examined brain structure and function in youths with AB on a whole brain level. We decided that a more stringent inclusion criteria is beneficial over the absolute number of studies entering the analyses, especially in regards to the attempt to truly capture the neuronal and structural phenotype of adolescents with AB. The number of studies entering each analysis therefore is on the lower limit. Contrast analyses are ideally contain a minimum of 15 studies in each dataset to obtain sufficient statistical power (<http://brainmap.org/ale/>). Therefore, the current analysis runs the risk of being under-powered. Most of the studies included here consisted of only, or majority of, male participants (see Tables). Some of the included study designs considered sex- matched clinical and control groups, while others applied a gender covariate within their design. Two VBM and one fMRI study included only female participants. These studies were nevertheless included in the current meta- analyses because the structural alterations observed in girls with CD broadly overlapped with those previously reported in male samples only. But while the current population included mirrors the occurrence of AB in the general population (e.g. higher number of males with AB), research has shown that it may be crucial to differentiate clinical cases based on gender in future research studies. Specifically, to determine possible gender related differences of structural and functional characteristics in individuals with AB, a comparison between meta-analyses of studies examining females and those examining males separately would have been of interest, but was not possible due to the small number of studies that are available for each group individually. Another potential caveat is the fact that clinical and subclinical forms of aggressive behaviour are often associated with comorbid diagnoses, most prominently attention-deficit hyperactivity disorder (ADHD; reported in up to 69% of CD patients) and anxiety. To date there is no neuroimaging evidence investigating pure diagnosis of clinical manifestations of aggressive behaviour (e.g. CD or ODD). Researchers argue whether aggressive behaviour in combination with ADHD even posits a distinct subtype or not and common neurobiological pathways are considered. Overall it can be concluded that neuroimaging research studies on aggressive behaviour in children and adolescents to date are characterized by diverse approaches in regards to the sample selection and definition, all of which have their justification and pitfalls. Ultimately, only a comparisons of both, pure and comorbid groups will be able to inform about the specificity and predictive value of either definition. Here we included adolescents with clinical and subclinical forms of aggressive behaviour, most of which have comorbid ADHD symptoms (e.g.. Many of the included studies report no differences in results when controlling for ADHD (through exclusion or a covariate within the study design;). Similar problems are IQ differences, drug use or socioeconomic status, all of which are a characteristic of populations with aggressive behaviour. Studies included in the current meta-analysis have all matched their participants according to IQ measures or used IQ as a covariate within their study design. Drug use and socio-economic status were controlled for in some, but not all, studies and further research is needed using a more careful sample characterisation in order to inform about the impact of these variables on brain structure and function. It is also to consider that the diagnosis of conduct disorder (clinical manifestation of AB) may encompass at least two clinically relevant subgroups. While the first group exhibits callous-unemotional traits (e.g. reduced guilt, callousness, uncaring behaviour and reduced empathy) and heightened risk of persistent antisocial behaviour, the second group is characterized by heightened threat sensitivity and reactive aggression. Callous-unemotional traits are highly heritable, expressed as early as at two years of age and are predictive of the most severe and persistent variant of conduct disorder. Studies also indicate that this severity may significantly impact the neuronal alterations observed. To summarize, while we were unable to constrain the current meta- analysis based on potential subtypification and gender variables, these factors may pose an exciting view on data analysis strategies and interpretations for future studies. For all the reasons noted, the current results have to be interpreted with caution. However, multimodal neuroimaging methods combining two or more functional (fMRI and/or EEG) and structural (MRI and/or DTI) approaches are suggested to provide a more sensitive measure in comparison to unimodal imaging for disease classification. Furthermore, we think that the confounding variables discussed here have influenced the functional and structural meta- analyses similarly. Overall, we could demonstrate that structural and functional alterations in adolescents with AB co-localize within key regions of the emotion processing and regulation network (e.g. prefrontal and insular cortex). Thus, our current analysis, using an activation likelihood estimation approach, provides an important step towards a more focused method of neuroimaging in AB. Future studies need to determine whether the here identified convergent clusters of neuronal and structural alterations may be applicable for clinical purposes (for example an improved pathophysiological description of individuals with AB) or whether a further specification (e.g. based on subtypes and gender) may be needed. However, the coordinates presented here can serve as non-independent regions of interest for future studies in AB, conduct disorder or in individuals with AB or antisocial/psychopathic tendencies. # Summary and Conclusion Aggressive behaviour constitutes a major issue of public health and increased knowledge about the behavioural and neuronal underpinnings of AB are crucial for the development of novel and implementation of existing treatment strategies. However, single site studies often suffer problems of small sample size and thus power issues. Quantitative meta-analysis techniques using activation likelihood estimations as implemented here offer a unique opportunity to investigate consistency of results between several studies investigating the same research question and population. We have implicated several brain regions of the emotion processing and regulation network to show hypoactivations and gray matter volume reductions in adolescents with AB (including prefrontal brain regions, amygdala, insular and cingulate cortex) and demonstrated that functional and structural alterations in AB co-localize within right dmPFC and left insular cortex. Overall, we are in line with meta-analytic work as well as structural, functional and connectivity findings that make a strong point for the involvement of a network of brain areas responsible for emotion processing and regulations. This network is impacted in individuals with AB and antisocial personality disorder/psychopathy. However, much still needs to be investigated. For example, study findings differ in regards to hypo- or hyperactivations and gray matter volume reductions or increases in different regions of the emotion processing and regulation network. Due to power constraints, the current meta- analysis only investigated hypoactivations and gray matter volume reductions in youths with AB and no hyperactivations or increases in brain structure. Future studies implementing longitudinal designs may be able to shed more light on the developmental pathway as well as onto typical and atypical trajectories within the regions reported. Such longitudinal designs will further allow the investigation of the bidirectional influence of biological and psychosocial influences in AB. # Supporting Information We thank Kübra Özoglu and Lea Klüwer for their help during the manuscript preparation. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NMR CS. Performed the experiments: NMR WMM LVF ET. Analyzed the data: NMR WMM LVF. Wrote the paper: NMR WMM LVF ET CS.

# Introduction Tamoxifen has formed the basis of treatment of women with metastatic breast cancer for several decades, resulting in significant improvements in quality of life and overall survival rates in a significant proportion of patients with clinically defined positive estrogen receptor (ER) status. However, both *de novo* and *acquired* resistance to tamoxifen, as well as to other endocrine agents, due to the loss of tumoural ER expression and/or its function presents a major therapeutic challenge, and usually leads to more aggressive disease upon relapse. Cellular transition from an epithelial to a mesenchymal phenotype (epithelial to mesenchymal transition; EMT) has been identified in various disease conditions including breast neoplasia and is associated with endocrine resistance and poor prognosis. Accompanying phenotypic changes include loss of cell-cell adhesion as a result of reduced E-cadherin and expression of catenins within adherens junctions, reduced claudins and occludins expression at tight junctions and reduced expression of epithelial cytokeratins such as KRT8, 18, and 19 which presumably aids in disruption of cytoskeletal connections necessary for maintaining normal tissue architecture. A variety of growth factors and their downstream signaling components have been associated with endocrine resistance and EMT. These include transforming growth factor β (TGFβ), insulin like growth factor 1 receptor (IGF1R), epidermal growth factor receptor (EGFR), PI3K/Akt, ERK/MEK, and MAPK,. We have previously reported the establishment of several cell lines in long term culture that have switched from an estrogen responsive to an endocrine independent state by the depletion of ER, induced by shRNA transfection of MCF-7 cells. Microarray and real time-PCR analysis confirmed a modified gene expression profile in the established transfectant cells indicative of EMT; loss of epithelial markers including E-cadherin, catenin, occludins and claudins, and enhanced expression of genes normally associated with mesenchymal cells such as N-cadherin, vimentin, fibronectin, integrin β4 and α5, and various metalloproteinase. In addition, these ER-depleted cells exhibit a series of changes in morphology and enhanced motility and invasiveness. It was demonstrated that MDA-231 cells (derived from a *de novo* ER negative breast tumor) as well as shRNA mediated ER-depleted cells (pII) (but not the parental ER positive MCF-7) were able to invade through a layer of basement membrane protein extract that is widely used to simulate the extracellular matrix (ECM). We also used agarose spots to simulate the polysaccharide component of the ECM and showed that pII cells could also penetrate into such structures. These experiments were conducted in the presence of serum; the contribution of individual serum components responsible for cell invasion was not determined. In the present study, we used an under-agarose gel assay – to study the invasive movement of two of the ER- depleted cell lines (pII and another designated YS2.5 which displays enhanced ER silencing compared to pII) in comparison to the parental MCF-7 cells, MDA-231, the normal breast epithelial cell line HBL100, as well as a line that was ER-shRNA transfected but failed to exhibit ER down- regulation (YS1.2). We present data on the effect of the growth factors EGF, IGF-1, TGFβ, PDGFC and the chemokine RANTES on cell invasion and proliferation. We demonstrate that EGF is the most effective stimulus for endocrine resistant cells to invade into an agarose matrix, that this is associated with increased phosphorylation of Akt and ERK1/2 and at least in part mediated by increased metalloproteinase (MMP) activity. # Materials and Methods ## Cell Lines HBL100 normal breast epithelial cell line, MCF-7 and MDA-231 human breast carcinoma cell lines were obtained from the ATCC (American Type Culture Collection, VA, USA). pII, YS2.5 and YS1.2 cell lines were established in this laboratory by transfection of MCF-7 with ER directed shRNA plasmid as described previously. For routine culture all cell lines were maintained at 37°C in a humidified atmosphere of 5% CO₂ in advanced DMEM supplemented with 5% fetal bovine serum (FBS), 600 µg/ml L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin and 6 ml/500 ml 100 x non-essential amino acids (all from Invitrogen, CA, USA). For YS2.5 and YS1.2, the maintenance medium also contained G418 (1 mg/ml) but this was omitted during experiments. IGF-1, EGF, TGFβ and the CC chemokine RANTES were purchased from Sigma, USA. PDGFC was purchased from Sino Biologicals. Stock solutions were prepared in sterile PBS and stored in small aliquots at −20°C and diluted in sterile PBS just prior to performing experiments. The EGF inhibitor erlotinib (LC laboratories, USA), the ERK inhibitor PD0325901 (Chemitek), and the Akt inhibitor LY294002 (Tocris Bioscience) were prepared in sterile dH₂O and stored in small aliquots at −20°C and diluted in DMEM just prior to performing experiments. ## Cell Invasion Assay Ultra-pure agarose (Invitrogen) was melted in PBS, then (once cooled below 40°C) supplemented with DMEM with or without 5% FBS or other components (insulin- transferrin-selenium (ITS) (Invitrogen), ITS plus bovine serum albumin (BSA) (Sigma) or IGF-1 to give a final 0.5% solution and allowed to solidify in individual wells of 6 well dishes at room temperature. Once set, 1–3 sample chambers (3.5 mm in diameter) were created in the gel, 2.5 mm apart in a horizontal line, by insertion of a metallic mould. Cells (4×10⁴) were re-suspended in DMEM containing various additives (5% FBS, 600 µg/ml L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin and 6 ml/500 ml 100 x non-essential amino acids) and loaded into formed chambers as appropriate for the experiment. Plates were incubated at 37°C in 5% CO₂ humidified atmosphere. After 24 h, cells that had penetrated into the agarose were manually counted by visual microscopic examination. Random cell invasion was determined as the total number of cells which moved in both lateral directions (A+B). Directional cell invasion was determined by subtracting B from A. All measurements were performed at least in triplicate and each experiment performed at least twice on separate days. It should be noted that cells did move in all directions around the chamber but only the lateral movement was considered in order to standardize the quantitation. Experiments were set up in the following manner: (1) a single cell chamber was created in the agarose and the cell line tested was added to the chamber (4×10⁴ cells in 10 µl media) and lateral random invasion (A + B) through the agarose layer which contained 5% FBS or other serum components was determined. (2) two chambers (3.5 mm in diameter) were created 2.5 mm apart and the lateral directional (A – B) invasion of the cells from one chamber towards the source of the test factor present in the other chamber was determined. (3) one chamber was created and the lateral random movement (A+B) of cells out of the chamber was determined by adding the test factor directly to the cells. (4) three cell chambers were created (3.5 mm in diameter), 2.5 mm apart; the two outer chambers contained untreated (UT) cells and cells pre-treated with erlotinib, PD0325901 or LY294002 respectively. The directional net cell invasion of treated cells toward EGF (50 ng/ml) present in the middle chamber was determined by subtracting the number of cells that moved out of the chamber with untreated cells. (5) three chambers were created (3.5 mm in diameter, 2.5 mm apart) and the lateral movement of cells placed into the middle chamber toward two growth factors present in the two outer chambers (competition assay) was determined. A schematic diagram is shown with the appropriate figure in the Results section to illustrate the design for each set of experiments. ## Cell Proliferation Assay The effect of growth factors (EGF, IGF-1, PDGFC and TGFβ) and EGF inhibitor (erlotinib) on pII cell proliferation was examined. Approximately 10⁴ cells were seeded into triplicate wells of 12-well plates and allowed to attach overnight. Either vehicle only or erlotinib (1–10 µM) was then added to the cells. Growth was assessed by MTT assay after 4 days of incubation. Briefly, 1 ml of MTT \[3-(4,5-dimethylthiazolyl-2)-2,5-diphenyltetrazolium bromide\] reagent (Promega) (0.5 mg/ml) was added to each well and plates incubated at 37°C for 30 min followed by the addition of 1 ml acidic isopropanol and vigorous re-suspension of the converted blue dye. Absorbance of the suspension was measured at 595 nm with background subtraction at 650 nm. The effect of the inhibitor used was compared to untreated control cells (taken as 100%). To test the effect of growth factors on proliferation, cells were plated as above and after 24 h the medium was replaced with serum-free standard DMEM (in case of using IGF-1, PDGFC or EGF) or 5% FBS containing DMEM (in case of using TGFβ) and left for another 24 h prior to the addition of IGF-1 (2–100 ng/ml), EGF (10–200 ng/ml), PDGFC (1–100 ng/ml), TGFβ (1–10 ng/ml) or vehicle. Cell growth was then determined after a further 4 days as described above. ## Western Blotting The level of total and phosphorylated Akt and ERK1/2 protein was determined in pII cells in response to IGF-1, EGF and TGFβ (1–100 ng/ml; 5–120 min) stimulation by immunoblotting. pII cells were cultured in 6 well plates with DMEM containing various additives (5% FBS, 600 µg/ml L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin and 6 ml/500 ml 100 x non-essential amino acids) until reaching 80% confluence and then serum-starved overnight before addition of growth factors (with vehicle only as control) for periods up to 120 min. Cells were then harvested by scraping, following the addition of 300 µl of lysis buffer containing 50 mM HEPES, 50 mM NaCl, 5 mM EDTA 1% Triton X, 100 µg/ml PMSF, 10 µg/ml aprotinin, and 10 µg/ml leupeptin. Protein assay was determined by the standard Bradford assay and 8 µg protein was mixed with an equal volume of 2 x SDS and heated at 90°C for 10 min. Lysates were loaded onto a 10% SDS-polyacrylamide gel and electrophoresed at 150 V for 1 h. Proteins were transferred to a nitrocellulose membrane and blocked with 2% BSA for 1 h before being incubated overnight at 4°C with either total or pAkt antibody (Ser473 from Cell Signalling, USA) (1/600 dilution), total or pERK1/2 antibody (Abcam, UK) (1/1000 dilution), or actin antibody (Cell Signalling,) (1/1000 dilution) prepared in 2% BSA. The membrane was washed and incubated with anti-HRP- conjugated secondary antibody (Cell Signaling) (1/500 dilution) for 1 h, developed with Super Signal ECL and visualized with Kodak X-ray film. ## Matrix Metalloproteinase (MMP) Activity Assay The general activity of MMP enzyme was determined using an assay kit purchased from Abcam (Cat No. ab112146) according to the manufacturer’s protocol. In brief, pII or YS1.2 (10⁴ cells) were seeded into triplicate wells of 6-well plates and allowed to attach overnight, then starved with serum free media for another 18 h. Cells were then treated with vehicle (control), IGF-1 or EGF (10–50 ng/ml) for 30 min and the MMP activity was assayed in the conditioned media; for this, 25 µl of medium was removed and added to 25 µl of 2 mM APMA working solution and incubated for 15 min at 20°C followed by addition of 50 µl of the green substrate solution supplied in the kit - a broad spectrum MMP fluorogenic peptide substrate. A kinetic measurement was then performed for the MMP activity by taking medium samples at 5 min intervals over a 1 h period after starting the reaction by using a microplate reader with a filter set of Ex/Em = 490/525 nm. In another set of experiment, pII cells were either un- treated, or pre-treated with PD0325901 or LY294002 (10 µM) for 1 h and then stimulated with EGF (50 ng/ml) for 30 min and MMP activity was measured as described above. ## Statistical Analysis Student’s two tailed unpaired t- test was used to compare means of individual groups: p≤0.05 was considered statistically significant. # Results ## Comparative Invasive Movement of Cell Lines Used in this Study As illustrated in, neither the parental MCF-7 nor the transfected (but not ER- down-regulated) YS1.2 cells were able to penetrate into the agarose. In addition, the normal breast epithelial cell line HBL100 did not show any invasive ability. MDA-231 and pII cells did enter the agarose in all directions and to a similar extent. The greatest penetration was however observed with the YS2.5 cells. ## Effect of Serum Components on Invasive Capacity of MDA-231 and pII Cells Both MDA-231 and pII cells showed absolute dependence on the presence of serum in the agarose for invasion. To investigate what minimum components in the serum may be responsible/necessary for the cell movement, the serum was substituted with an ITS solution, which is a commonly employed supplementation to many conventional synthetic nutrient media that permits substantial reduction in the serum requirement for routine maintenance of cells in culture. In the presence of ITS in the agarose, MDA-231 and pII cells both showed some degree of invasion but this was significantly less than with the optimal 5% serum condition. The addition of BSA (1.5 mg/ml) with the ITS did not further increase the degree of invasion of either cell line. To determine if the ITS effect was due to the insulin, the experiment was performed with the addition of IGF-1 to the basic serum-free medium. At 1 and 10 ng/ml IGF-1 elicited some invasive activity but significantly less than that observed with ITS. It should be noted that in all experiments, the cells inside the chamber were suspended in complete medium containing 5% serum, without which invasion was not observed. ## Effect of IGF-1, EGF, TGFβ, PDGFC, and RANTES on the Directional Invasion of pII Cells The effect of several growth factors known to affect breast cancer cells was examined. For these experiments, the agarose was prepared with ITS instead of 5% serum since some of these growth factors may themselves be serum constituents. We determined the net (directional) invasive movement of pII cells towards different concentrations of IGF-1, EGF, TGFβ, PDGFC and RANTES. pII cells were plated in one chamber and the growth factor/chemokine was added to another chamber 2.5 mm apart. The factor diffuses into the agarose layer forming a downward concentration gradient towards the cells. Net directional invasion of cells towards the factor was calculated by subtracting the number of cells which moved towards the source of the stimulus from the number of cells which moved in the opposite direction (away from the chamber containing the stimulus). The data shown in -A illustrates a concentration – dependent bell-shaped curve observed for the net invasive movement of pII cells towards IGF-1. This resembles the classic dose response curve seen for other immune cell chemotaxis (neutrophils) when using this assay. The optimal dose of IGF-1 inducing net invasion (3 fold increase over PBS) was 50 ng/ml. With TGFβ (-B), a significant increase in cell invasion was observed only at 2 µg/ml with lower concentrations showing no effect above the PBS control. EGF showed a very potent stimulatory effect even at 1 ng/ml (-C) reaching a peak at 50 ng/ml and declining thereafter in a bell shaped curve. The highest fold increase in the net invasion over PBS values was significantly greater than that observed with the other agents (300 fold increases for EGF vs. 3 fold increase for IGF-1 and TGFβ). PDGFC at similar doses did not induce pII cell directional invasion over the PBS control (-D) although these cells exhibit high expression of this growth factor. The CC chemokine RANTES (-E) showed no enhancement in net invasion of pII cells tested at the same dose range as used for the growth factors. When pII cells (placed in a central chamber) were given a choice between two growth factors (IGF-1 vs. EGF; 50 ng/ml) present in the two outer chambers (competition assay, -F), these cells preferred to move in significantly higher numbers towards EGF, which is consistent with its greater potency in inducing pII cell invasion in comparison with the other growth factors tested. Neither IGF-1 nor EGF was able to induce invasion of the ER positive parental MCF-7 cells (data not shown). ## Effect of IGF-1, EGF, TGFβ, PDGFC and RANTES on the Random Invasion of pII Cells To exclude the possibility of different rates of diffusion between the growth factors/chemokine tested toward the cell-containing chamber which might result in different response rates, in these sets of experiments, the test factor was added directly to the chamber containing the cells and random (total) invasion was determined by counting cells that had penetrated into the agarose in either lateral direction out of the well. The dose of each agent used was lower than in the directional invasive experiments, in order to maintain some degree of approximate equivalence on the rationale that in this case cells were directly in contact with the agent instead of being exposed only to the diffused component. In terms of their relative effects, (at equimolar concentrations; 10 ng/ml) the order of potency of enhancement of the random invasion of pII cells was EGF \>IGF-1\> TGFβ. PDGFC and RANTES produced no effect above the PBS control. There was no significant difference in cell growth between newly seeded control and growth factor treated cells within the 24 h period used for the invasion assays. Proliferation differences were usually seen at 2 days onwards and therefore not likely to contribute to differences observed in pII cell invasion. ## Effect of IGF-1, EGF, PDGFC and TGFβ on pII Cell Proliferation The growth rate of pII cells was significantly increased in a concentration- dependent manner after treatment with IGF-1(-A) or EGF (-B) compared with the untreated cells, with a five-fold increase at 100 ng/ml (-A and B). On the other hand, TGFβ treatment resulted in significant inhibition at 10 ng/ml (-C). PDGFC showed no effect when used at similar dose range as for the other growth factors (-D). ## Effect of Erlotinib on pII Cell Invasion and Proliferation Erlotinib treatment significantly reduced the number of invading pII cells moving towards EGF (50 ng/ml) by 75% and 90% at 1 and 10 µM respectively (-A). It was also effective in the same concentration range in reducing cell proliferation (-B). ## Effect of IGF-1, EGF and TGFβ on Activation of Akt and ERK1/2 To determine the possible signaling events downstream of growth factor activation, we examined the degree of phosphorylation of Akt and ERK1/2, two important mediators that are known to be involved in pathways leading to proliferation and invasion as well as endocrine resistance. As shown in -A, EGF treatment (10 ng/ml) significantly enhanced both Akt and ERK1/2 phosphorylation over a 30–120 min period while IGF-1 had no effect (-A). TGFβ treatment reduced Akt phosphorylation but had no effect on ERK1/2. There was no difference in total Akt or ERK1/2 levels in all of the conditions tested (-A). We subsequently tested various dose ranges and shorter stimulation times on the degree of Akt and ERK1/2 phosphorylation in response to IGF-1 and EGF treatment in pII cells. At 30 min (-B), a dose of IGF-1 at 100 ng/ml significantly enhanced the degree of Akt phosphorylation while EGF treatment enhanced Akt activity at a lower dose range (1–10 ng/ml, -B). ERK1/2 phosphorylation was not changed compared to the UT cells in response to IGF-1 stimulation even at higher doses (100 ng/ml), while EGF treatment enhanced ERK1/2 phosphorylation at 1–100 ng/ml dose range. We also measured Akt and ERK1/2 phosphorylation at shorter time points (5–15 min, -C) and found that both Akt and ERK1/2 phosphorylation was significantly enhanced in response to EGF compared to IGF-1 treatment with 10 ng/ml. ## Effect of Akt and ERK1/2 Signaling Pathways on EGF and IGF-1 Induced pII Cell Invasion To determine the significance of Akt and ERK1/2 phosphorylation in response to EGF and IGF-1 treatment, pII cells were pre-treated with the ERK1/2 inhibitor PD0325901 or the PI3K-Akt inhibitor LY294002 (100 nM-10 µM) and their random invasion in response to EGF or IGF-1 (10 ng/ml) was determined using the under agarose assay. PD0325901 treatment did not affect the degree of invasion in response to IGF-1 while it significantly reduced the degree of pII cell invasion in response to EGF at 1 and 10 µM by 64 and 80% respectively. On the other hand, LY294002 significantly reduced pII cell invasion in response to IGF-1 and EGF by 41–70%. ## Effect of IGF-1 and EGF Treatment on Matrix Metalloproteinase (MMP) Activity We studied the effect of IGF-1 and EGF treatment on the degree of extracellular (i.e. secreted) MMP activity which could explain how these growth factors can enhance the invasion of endocrine resistant breast cancer cells as indicated by the data in and. As shown in -A, IGF-1 (10–50 ng/ml) had no significant effect on MMP activity as compared with the untreated controls, but interestingly EGF at 50 ng/ml significantly enhanced MMP activity over the controls (2.2 fold increase at 60 min). Interestingly, in the ER + cells YS 1.2 neither IGF-1 nor EGF had any effect on MMP activity (-B) compared to untreated control cells, which suggests a direct link between ER knockdown, increased MMP activity, and enhanced cell invasion. In the presence of LY294002 there was a significant reduction in the degree of EGF-induced MMP activity in pII cells (38%) whereas PD0325901 produced a much smaller (14%) and non-significant effect (-C). # Discussion We studied the influence of various serum components on cell invasion and the modifying effects of erlotinib that is known to antagonize EGF action by inhibition of receptor tyrosine kinase activity. The HBL100 cells originally established from non-malignant breast cells were unable to penetrate into agarose. This was also the case for MCF-7 cells as has been shown before. MDA-231and pII cells demonstrated a similar degree of random cell invasion that was dependent upon the presence of 5% FBS in the agarose. Interestingly, YS2.5 cells showed significantly enhanced degree of invasion compared to pII cells. Both lines were generated from MCF-7 transfection by ER directed shRNA constructs but the YS2.5 display a much greater ER knockdown effect (data not shown) suggesting that the invasive capacity of the cells may be influenced by the residual level of ER in previously ER expressing cells. A recent report by Ye *et al* indicated that the degree of ER knockdown was related to increased invasion of MCF-7 cells into matrigel. They suggested a role for ER signaling in regulating E-cadherin expression and EMT through SNAIL2. The YS1.2 line, that was also derived by transfection with ER targeting shRNA but which did not exhibit reduced ER transcript levels, behaved similarly to the parental cells. This clearly demonstrates that it is the blockade of ER production that confers the ability to invade rather than transfection *per se*. Although the combination of insulin, transferrin and selenium as a culture medium supplement was able to substitute the serum to some extent; it is apparent that other serum factors are needed for invasion. We examined the effect of four growth factors on the proliferation and migration of pII cells. Directional invasion was induced by addition of varying concentrations of these factors at a distance of 2.5 mm from the cell-containing chamber, or random invasion by addition of the growth factor directly into the well containing the cells. In both scenarios, the potency of EGF was greater than either IGF-1 or TGFβ. IGF-1 has long been implicated in tumorigenesis as well as specifically in proliferation, survival and migration of tumor cells, with positive correlation to breast cancer progression. IGF-1 increases proliferation of ER positive and negative breast cancer cells. IGF-1 and EGF have previously been reported to enhance MCF-7 cell proliferation to a similar extent and this equivalent potency was also observed in our pII cells. In other studies, IGF-1 enhanced MCF-7 cell migration, and invasion with a bell-shaped dose-response curve in the same dose range used in our study, but with a different outcome. We observed that IGF-1 induced cell invasion in pII but not in MCF-7 cells (data not shown) in the under agarose assay. TGFβ plays a complex role in carcinogenesis where it has a significant inhibitory effect on the growth of cancer cells at early stages but it induces EMT and enhances cell invasion ant metastasis at late stages. It is frequently over-expressed in breast cancer and its level of expression correlates with poor prognosis. Its growth inhibitory effects have been noted in a wide variety of breast cancer cell lines – and such was also observed in our pII cells (-C). TGFβ has been reported to enhance the migration, and invasion, – of both ER- positive and negative breast cancer cells. In this study we found that TGFβ, while it did enhance pII cell invasion, it did so at a relatively high dose as compared with EGF and IGF-1, suggesting that at least its role in tumor invasion and metastasis is subsidiary to its effect as a growth inhibitory agent. EGF and its receptor have been frequently associated with the progression of several forms of cancers including those of the breast,. Although some studies have reported that EGF failed to induce proliferation of MDA-231 or MCF-7 another showed that it enhanced MCF-7 cell proliferation to a similar degree as IGF-1 consistent with our own observation on pII cells. Several reports have confirmed the role of EGF in breast cancer cell migration, and invasion. Our data suggest that EGF is the most potent enhancer of invasion among the growth factors tested. Our data also show that ER negative cells preferentially move towards a source of EGF when in competition with IGF-1 (-F). Furthermore, inhibition of EGFR activity by erlotinib profoundly inhibited pII cell invasion and proliferation. These data support the potential benefits of selectively antagonizing EGF action particularly in early endocrine resistant breast cancer to prevent invasion and metastasis. Some studies have suggested a role of the CC chemokine RANTES in inducing breast cancer cell migration, and invasion, but we found no evidence to support this. RANTES failed to induce either directional or random invasion of pII cells at dosages similar to those used in the above mentioned studies. A wide variety of growth factor induced downstream signaling molecules have been associated with breast cancer pathogenesis and shown to play vital roles in modulating processes such as cell growth, survival, EMT and invasion,. During this study we also compared the effect of imatinib, an agent that is thought to predominantly target the tyrosine kinase activity of the PDGF receptor, with suramin that has been used clinically as a broad spectrum growth factor antagonist. We found (data not shown) that imatinib was the most potent inhibitor of pII cell invasion and proliferation highlighting the importance of PDGF receptor as a potential therapeutic target to control breast cancer cell growth and metastasis. Interestingly, Li *et al* demonstrated a functional cooperation between EGFR and PDGFR during cell migration of murine fibroblasts whilst Abouantoun and MacDonald suggested transactivation of EGFR by PDGFRB could be responsible for migration and invasion of medulloblastoma cells. They found that imatinib blocked PDGF-BB activation of PDGFRB, Akt and ERK and reduced the level of PDGFRB-phospho-EGFR heterodimers. In our previously reported study using microarray display we did not observe any change in either PDGFB or PDGFRB mRNA in pII cells vs. MCF-7 but there was a significant elevation in PDGFA and PDGFD and a 50–200 fold increase in PDGFC. The latter was shown to interact with PGDFα and β receptors and has been described as a potent mitogen for mesenchymal cells. As pII cells have acquired a mesenchymal-like phenotype it may well be that imatinib is blocking the action of an over- expressed PDGFC. This drug is also clinically used to target the constitutive kinase activity of the BCR/ABL product, so it is uncertain which of these activities is responsible for our current observations. Therefore, we tested the effect of PDGFC but somewhat surprisingly, found that it had no effect on either proliferation or invasion. Moreover, PDGFC failed to show any synergistic effect on random invasion when added directly to cells together with sub-optimal concentrations of EGF, or in promotion of directional movement towards a source of EGF (data not shown). It remains to be determined if other PDGF isoforms such as A, B or D may play a role in enhancing breast cancer cell invasion. The intracellular signalling pathways that can potentially transmit the effects of growth factors have been well described. In this study we looked at two central mediators (Akt and ERK1/2) and found that both were selectively activated by EGF, suggesting their involvement in promoting EGF-induced invasion. We also found that EGF significantly enhanced general MMP activity secreted by pII but not by the ER positive YS1.2 cells and that was mediated through the Akt (and in part through EKR1/2) pathway. We also showed that EGF- induced pII cell invasion was significantly blocked by the ERK inhibitor PD0325901 and the Akt-PI3K inhibitor LY294002 which is in agreement with our data generated from the MMP and western blotting experiments. Although several reports demonstrated that IGF-1 can enhance MMP activity in both ER positive, and ER negative breast cancer cells, we did not observe any increase in general MMP activity with either pII or YS1.2 cells treated with IGF-1 at similar dose range used in these studies. Kim et al, observed enhanced MMP-9 activity mediated through JAK3/ERK pathway in EGF treated SKBR3 breast cancer cells, whereas Park et al showed enhanced MMP-1 activity through the involvement of ERK/MAPK signalling pathways. We observed a small (14%), but non- significant inhibition in general MMP activity following treatment of pII cells with the ERK inhibitor PD0325901 (-C). Our data is consistent with the findings of Lee et al who showed that EGF enhances MMP-9 activity in the ER negative MDA- MB-231 cells mediated through the Akt-PI3K signalling pathway. In summary, we have presented experimental evidence showing that both *de novo* and siRNA induced reduction of ER expression is directly associated with enhanced cell invasion. TGFβ plays a major role in inhibiting cell proliferation, whereas of the pro-proliferative factors IGF-1 and EGF, the latter plays a dominant role in cell invasion and by implication metastasis through activation of the PI3K and ERK pathways that we have previously suggested to be contributory mechanisms to endocrine resistance. This is also underlined by the effectiveness of erlotinib indicating that this drug has potential therapeutic usefulness for preventing spread of particularly endocrine resistant breast cancer. The under agarose assay utilized in this study proved to be a simple and effective means of monitoring and quantifying cell invasion by permitting direct microscopic visualization as compared with methods that rely on indirect measurements of fluorescent signals emitted from post incubation of migrating cells with dyes such as calcein. It facilitates direct comparison of cell lines as well as simultaneous exposure of cells to more than one test factor. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MAK YAL. Performed the experiments: MAK SS PMM. Analyzed the data: MAK. Contributed reagents/materials/analysis tools: YAL MAK. Wrote the paper: MAK YAL.

# Introduction Watching the mouth movements of a speaker (so called, visual speech) may help listeners to decode speech in a noisy environment, and may even alter the auditory perception of speech as in the McGurk effect. Observers can discriminate fairly reliably between silent video-clips of a speaker played as recorded (Forward mode) or time-reversed (Backward mode). It was argued that natural kinematics (recognition of biological motion) rather than linguistic competences had a role in this task. Normal and time-reversed visual speech differ kinematically in several ways, although the qualitative differences are subtle. With few exceptions (such as long vowels or fricative consonants), phono-articulatory gestures tend to be asymmetric in time. For instance, deceleration phases are longer than acceleration phases, and asymmetries are present between the opening and closing movements of the mouth. Moreover, articulatory gestures of speech obey specific constraints imposed by the motor system. Thus, the temporal inversion of these gestures often generates a sequence of unnatural movements hardly repeatable by normal people, although an experienced person can invert the temporal order of the phonemes in a sentence. In fact, the articulatory sequences that generate the phonemes during a speech are extremely complicated to perform in reverse. This is because one should reverse each phono-articulatory manoeuvre required to produce a given phoneme, as well as the specific sequence with which these manoeuvres are chained during the speech. The central nervous system is also sensitive to language familiarity in visual speech. Indeed, a familiar language can be discriminated by the analysis of the speech temporal structure (i.e., rhythm) in auditory as well as in visual modality. Temporal duration and variability of vowels and consonants differ between languages, and the timing of vowels and consonants can be visually assessed since phono-articulatory gestures generating these movements fit into different visual classes. For instance, Spanish monolingual speakers visually distinguished Spanish from Catalan, while this was not possible either for English or for Italian speakers. Discrimination ability of a familiar language persists also after a temporal reversal of visual speech stimuli. The rhythmic and global timing structures of speech visible cues (e.g., alternation of consonants and vowels, vowels duration) remain relatively unaltered after a temporal inversion of speech sequences, while semantic, lexical and phonotactic information are lost. Moreover, six-months-old infants are able to discriminate a familiar language from visual speech. These observations suggest that visual spatio-temporal cues play a more important role in identifying familiarity than linguistic competence. The brain networks involved in these processes are unknown. To our knowledge, no study so far has directly investigated the neural correlates of language discrimination of visual speech, while the few papers reporting brain sites activated by inverting the temporal order of natural visual speech measured brain activity using techniques (PET or MEG) other than fMRI. In theory, the occipito-temporal cortex (OTC) might have a specific selectivity to the spatio-temporal features of visual speech (i.e., kinematics of biological motion). Indeed, various foci in this region respond to different types of human movements and body forms. Studies comparing face movements during a speech with facial movements that cannot be construed as speech reported activations in both lateral and ventral OTC, including the temporal visual speech area (TVSA), as well as in auditory association areas of the temporal cortex in the superior temporal gyrus. However, this comparison might be affected by the presence of confounds in low-level visual features, such as differences in motion speed. A contrast immune from low-level visual confounds is the comparison of speech movements rendered normally (Forward) versus time-reversed (Backward). A recent MEG experiment showed that, during the processing of silently played lip movements, the visual cortex tracks the missing acoustic speech information when played forward as compared to backward, indicating a top-down modulatory control of auditory dorsal stream on visual areas. Also, in a PET study, the contrast Forward versus Backward engaged OTC bilaterally. However, the ability to discriminate plausible speech gestures (i.e., Forward versus Backward video clips) was localised to later stages of processing, such as the parieto-temporal cortex and motor areas in the frontal cortex. Visual speech stimuli have been shown to engage cortical motor areas involved in speech production, such as the left inferior frontal gyrus (IFG) that includes Brodmann’s Area BA 44 and BA 45 (pars opercularis and triangularis of IFG, respectively) thought to overlap with Broca’s region and the ventral premotor cortex (PMv), but also more dorsal regions of the premotor cortex (PMd). Importantly, part of the frontal areas implicated in the control of movements and speech are connected with the visual cortex. The motor theory of speech perception proposed that the activation of motor speech areas during the observation of speech might represent an implicit motor simulation of the observed gestures conducive to speech understanding. However, several authors questioned the idea that an automatic engagement of motor areas, such as IFG, during perceptual or cognitive task is evidence of a specific involvement of the motor system in perceptual or cognitive processes. The dorsal premotor cortex, rather than the Broca’s area (BA 44/45), seems to be engaged both in the execution and the observation of speech gestures. Conversely, it was found that the activity in the IFG correlates with hit-rate and response bias during speech perception tasks. Since response bias and hit rate are characteristic indexes of the decisional process, these findings might suggest that high level processes related to the generation of the response decision (e.g. whether to respond Yes or No), rather than motor simulations occurred in the IFG during visual speech. In summary, the specific role of IFG and Broca’s area in the functional architecture of speech perception remains open to debate. In the present study, we investigated the neural circuits engaged by language familiarity (Italian vs Arabic) and natural kinematics of biological motion (Forward vs Backward) of visible speech. Italian observers viewed silent video- clips of the mouth movements of Italian and Arabic actors speaking in their native language. Stimuli were rendered either in normal (Forward) mode or after a time reversal (Backward). During an fMRI session, participants were asked to identify the rendering mode (Forward or Backward). The brain regions sensitive to language familiarity and those sensitive to natural mouth movements of speech were identified through the fMRI contrasts Italian vs Arabic (main effect of language) and Forward vs Backward (main effect of rendering mode), respectively. We also computed the interaction between the two main factors by means of the contrast “language x rendering mode” \[(Italian Forward vs Italian Backward) vs (Arabic Forward vs Arabic Backward)\]. The latter contrast should identify the areas where the effects of Familiarity (Italian vs Arabic) was larger for natural (Forward) than non-natural (Backward) motion (i.e biological motion). # Methods ## Participants Forty healthy right-handed Italian volunteers took part in this study. Twenty participants (14 females, 6 males; mean age: 25 years; age range: 20–42 years) were tested in a preliminary experiment and twenty different participants (13 females, 7 males; mean age: 23 years; age range: 20–35) in the main study (i.e. fMRI and in a follow-up experiment, see later). All participants had normal or corrected to normal vision. None of them had any familiarity with the Arabic language or experience with lip-reading. Written informed consent to procedures approved by the Institutional Review Board of Fondazione Santa Lucia was obtained from each participant. Experimental protocols complied with the Declaration of Helsinki on the use of human subjects in research. ## Stimuli Ten adults (5 females, 5 males) native speakers of Arabic and 10 adults (5 females, 5 males) native speakers of Italian volunteered as actors for generating the stimuli. We chose the Arabic rather than other European languages in order to ensure that the participants would have not been exposed more than occasionally to this language before, and we verified that this was the case. The choice of the Arabic was also motivated because it differed from the Italian more than did most other European languages. Indeed, articulatory movements for the production of words in Arabic and Italian languages are different. Two of the authors (P.V. and V.M.) selected Italian speakers so that, after careful visual inspection, the general features of the lower part of the face were roughly similar to those of the Arabic speakers previously selected. None of the actors participated in the main or in the preliminary study. Each actor read four texts excerpted from newspapers in his/her native language (Arabic, A or Italian, I). The preparation of the stimuli involved three steps. First, we recorded the lower part of the face (including the upper/lower lips and the chin) of each actor with a digital camera (25 frames/s, Sony HDR-SR-8E), and stored the results as a sequence of single frames (1024 x 768 pixel, RGB TIFF format). Second, static frames were processed with Photoshop CS6 to equalize for luminance and chromatic spectrum, and cropped to the size of 972 x 694 pixels in order to display only the mouth movements. Finally, by using Virtual Dub, we transformed the sequences of frames into silent AVI video-clips lasting 14 s. The experimental stimuli consisted in the video-clips rendered either as recorded (Forward mode, F) or after reversing the frames order (Backward mode, B). The total number of available stimuli was 2 \[rendering mode\] x 2 \[language\] x 10 \[actor\] x 4 \[text\] = 160. Four examples of video stimuli, one for each category of interest (I_F, I_B, A_F, A_B), are provided as supplementary material. Additionally, we checked whether Arabic and Italian video-clips differed in motion energy. For each two consecutive frames of each video-clip, we calculated the mean of the squared differences in the red, green, and blue channels in every pixel. The motion energy was estimated as the average of these values across all pixels and frames (350 frames) of each video. We found that the motion energy was not significantly different (unpaired t-test; p = 0.7; t-value (78) = 0.38) between Italian (mean ± SD: 34.0 ± 11.8 pixel²/frame²) and Arabic videos (34.9 ± 7.8 pixel²/frame²). ## General outline of the study The study involved three successive experiments: a preliminary, purely behavioural session with the first group of 20 participants in which we estimated the ability to discriminate presentation modes (Backward/Forward); a main session with the second group of 20 participants in which this ability was estimated while measuring brain activity with fMRI; a follow-up session with the same participants of the fMRI session in which we estimated the ability to discriminate language familiarity (Italian/Arabic). ## Main task: Identification of rendering mode In both fMRI and preliminary experiments, participants were informed that the video-clips being shown could be either a faithful or a time-reversed rendering of actual speech movements. They were not informed that in half of the videos the actor’s language was Italian (I) and in remaining half was Arabic (A). The task (2-Alternative Forced Choice: 2-AFC) was to indicate whether the video was displayed as recorded (Forward) or reversed in time (Backward). Participants had to wait until the end of the stimulus before responding. Responses were entered by pressing with the right index finger one of two buttons marked “F” (Forward) and “B” (Backward), respectively. Between trials, the display was uniformly grey and participants fixated a central point (0.5° visual angle). No constraints were imposed on oculomotor behaviour during the presentation of the stimuli. Before each experimental session, participants were administered eight warm-up trials, which included at least one example for each combination of actor gender, native language, and rendering mode. The results of these trials were not analysed. ## Follow-up task: Identification of language The participants in the fMRI experiment were retested 10 (± 2) days later in a follow-up experiment outside the scanner. They viewed the same 160 silent video- clips described above. Participants were informed that the actor’s language could be either Italian or another, unspecified language, but no further information was provided. Participants were asked to wait until the end of the stimulus before responding (2-AFC). Reponses were entered by pressing with the right index finger a button marked either I (Italian) or NI (not Italian). The aim of this experiment was to gauge the accuracy with which viewers could discriminate a familiar language (Italian) from an unfamiliar one (Arabic) using only visual cues, and to test whether the time-arrow (forward vs. backward rendering mode) affects the judgment of language familiarity. We hypothesized that the visual cues used to discriminate languages (e.g., temporal variability of vowels) are different from those used for discrimination of rendering mode (e.g., the acceleration profiles of opening/closing movements of the mouth). If so, the sensitivity index (see below) should be uncorrelated between the two tasks. ## General procedure In each experiment, the total number of stimuli (160) was divided in 5 runs (32 stimuli/run) with the constraint that successive stimuli never involved the same actor. Stimuli were pseudorandomized and presented using the Presentation software (Neurobehavioural system^®). Within runs, interstimulus intervals (ISI) followed a uniform distribution (range: 2 s–4 s; mean: 3 s). The five runs were administered in a single session and were separated by brief pauses. Additionally, in the fMRI experiment, to estimate more accurately the shape of BOLD impulse response, we pseudo-randomly inter-mixed null events (N = 35, duration 8 s). Thus, the duration of each run was 10’ 50” during fMRI and 9’ 54” in the follow-up and preliminary experiments. ## fMRI experiment: Set-up Participants lay supine in the MR scanner with the head immobilized with foam cushioning and wore earplugs and headphones to suppress ambient noise. A digital projector (NEC LT158, refresh rate: 60-Hz) projected the stimuli through an inverted telephoto lens onto a semi-opaque Plexiglas screen mounted vertically inside the scanner bore, behind the participant’s head. The back-projected image was then viewed via a mirror mounted on the head coil positioned at about 4.5 cm from the eyes. The eye-to-screen equivalent distance was 66 cm, and the angular size of the projected image was 9° (width) × 6.4° (height). Responses were acquired with an MR-compatible response box (fORP, Current Designs). ## Follow-up experiment and preliminary experiment: Set-up The follow-up and preliminary experiments were performed in a quiet, dimly illuminated room. Participants sat in front a 19” LCD monitor and viewed the silent video-clips (9° x 6.4° visual angle) in a pseudo-random order at a distance of about 80 cm. Responses were entered via a high-speed button box (Empirisoft^®). ## Behavioural data analysis In the fMRI experiment and in the preliminary experiments, where the task was to identify the rendering mode (see above), responses were collated by language, rendering mode and actor. The number of responses “Forward” to Forward and Backward stimuli are indicated as N_F\|F and N_F\|B, respectively, while the number of responses “Backward” to Forward and Backward stimuli for each language are indicated as N_B\|F and N_B\|B, respectively. For each participant, the sample size was N_T = 80 for each language, thus a total of 1600 trials for each language was collected (N_T x 20 participants). For each participant, we computed a sensitivity index d’ = Z{Hit} − Z{False Alarm} and a response bias *c* = −0.5\*(Z{Hits} + Z{False Alarm}), where Z{Hit} and Z{False Alarm} are the z-scored transformed values of P{Hit} = P{F\|F} = N_F\|F/N_T, and P{False Alarm} = P{F\|B} = N_F\|B/N_T, respectively. Moreover, we calculated the probability of correct responses as P{C} = (N_F\|F + N_B\|B) / N_T. Similarly, in the follow-up experiment where the task was to identify the actor’s language (see above), responses were collated by language, rendering mode and actor. The number of responses “Italian” to Italian and Arabic stimuli are denoted as N_I\|I and N_I\|A, respectively, and N_A\|I and N_A\|A are the number of responses “Not Italian” to Italian and Arabic stimuli, respectively. For each participant, the sample size was N_T = 80 for each rendering mode, thus and a total of 1600 trials for each rendering mode was collected (N_T x 20 participants). We estimated sensitivity and response bias through d’ and *c* indexes, respectively, based on the convention that, in this case, Z{Hit} and Z{False Alarm} are the z-scored transformed values of P{Hit} = P{I\|I} = N_I\|I / N_T and P{False Alarm} = P{I\|A} = N_I\|A / N_T, respectively. Moreover, we calculated the probability of correct responses: P{C} = (N_I\|I + N_A\|A) / N_T. We considered d’ and response bias in addition to the probability of correct responses, since the latter might be inflated by response bias and lead to misleading interpretations. We expected that responses to the stimuli depended on whether participants had to discriminate between rendering mode (in the first two experiments) or languages (in the follow-up experiment). Thus, within-subject responses to the rendering-mode and language discrimination task should show different patterns, and the sensitivity index (d’) should be uncorrelated between tasks. To verify these points, we calculated the correlation coefficient of participants’ sensitivity index between the main task and the follow-up experiment task. ## fMRI data acquisitions MR images were acquired with a Siemens Magnetom Allegra 3T head-only scanning system (Siemens Medical Systems, Erlangen, Germany), equipped with a quadrature volume RF head coil. Whole brain BOLD echoplanar imaging (EPI) functional data were acquired with a 3T-optimized gradient echo pulse-sequence (TR = 2.47 s, TE = 30 ms; flip angle = 70°; FOV = 192mm, fat suppression). 38 slices of BOLD images (volumes) were acquired in ascending order (64 x 64 voxels, 3 x 3 x 2.5 mm³, distance factor: 50%; inter-slice gap = 1.25 mm; slice thickness = 2.5 mm), covering the whole brain. For each participant, a total of 1315 volumes of functional data were acquired in five consecutive runs. At the end of each run, the acquisition was paused briefly. Structural MRI data were acquired using a standard T1-weighted scanning sequence of 1 mm³ resolution (MPRAGE; TE = 2.74 ms, TR = 2500 ms, inversion time = 900 ms; flip angle = 8°; FOV = 256 × 208 × 176 mm³). ## fMRI data preprocessing Data and statistical analyses were performed using the SPM12 software (Wellcome Trust Centre for Neuroimaging, London, UK) implemented in MATLAB R2013 (The MathWorks Inc., Natick, MA) using standard procedures. After discarding the first four volumes of each run, images were corrected for head movements, realigned to the mean image, coregistered to the structural image, and normalized to Montreal Neurological Institute (MNI) space using unified segmentation, including resampling to 2 × 2 × 2 mm voxels, and spatially smoothed with a 8 mm full-width at half maximum (FWHM) isotropic Gaussian kernel. Voxel time series were processed to remove autocorrelation using a first-order autoregressive model and high-pass filtered (128-s cut-off). ## fMRI analysis Patterns of brain activations were computed using the general linear model and a Finite Impulse Response (FIR) set of base functions. Here, the FIR approach is ideal to fit brain activity, because it can identify changes of activity over time without making any assumptions about the profile of these changes. Accordingly, for each participant, the FIR estimated the level of activation in 12 successive time-bins. Each time-bin consisted of 1 TR (2.47 s), thus fitting 30 s of the fMRI data for each stimulus. We modelled 5 different event-types: Italian Forward rendering (I_F), Italian Backward rendering (I_B), Arabic Forward rendering (A_F), Arabic Backward rendering (A_B), time-locked to stimulus onset, thus obtaining 12 images (one for each time bin) for each correct trial of the 4 conditions, plus an additional event corresponding to errors trials irrespective of condition. Motion correction parameters were also included as effects of no interest. We analysed the activity related only to stimuli correctly identified (correct trials), since error trials (stimuli not correctly identified) may introduce confounding activation (i.e. contamination of the activation related to poorer performance by increased errors. However, to evaluate the effect of error trials on the fMRI activity, we did a supplementary analysis (not reported here) with all trials (correct and error trials). We found that the brain sites activated in the main fMRI analysis were also activated in the supplementary fMRI analysis, although at an uncorrected level (p-uncorr \< 0.05), thus indicating that error trials decrease the signal-to-noise ratio. At single-subject level, we estimated four effects of interest. First, we calculated the contrast representing the overall mean activity of all stimuli by averaging the estimated parameter of all conditions (\[I_F + I_B + A_F + A_B\]/4) in each bin. Subsequently, we estimated the contrasts of the three effects: (1) main effect of actor’s language (\[I_F + I_B\] vs. \[A_F + A_B\]); (2) main effect of rendering mode (\[I_F + A_F\] vs. \[I_B + A_B\]); and (3) modulatory effect of actor’s language on rendering mode (interaction: \[I_F − I_B\] vs. \[A_F − A_B\]). The resulting parameters for each contrast (corresponding to 12 images, one for each time bin) in each participant were then entered into second-level group analyses. Four separate one-way ANOVAs with 12 levels (each corresponding to one time-bin) were performed at the second (group) level. We used F-contrasts to highlight brain areas showing differential activity over the 12 time-bins, separately for each of the four ANOVAs. In particular F-contrasts subtracted the activity of the first bin (i.e. one TR at stimulus onset) from each of the other bins, thus capturing the changes of activity over-time. All analyses included appropriate corrections for non-sphericity. Statistical thresholds were set at p-FWE \< 0.05, family-wise error corrected for multiple comparisons at cluster level (hereafter, p-corr \< 0.05), using a voxel-wise threshold set at p\< 0.001. Furthermore, post-hoc t-tests on each time bin were false-discovery-rate (FDR) corrected for n multiple comparisons at p \< 0.05 across the number of bins (n = 12). ## Regions of interest In addition to the previous whole-brain analysis, we also performed an analysis based on regions of interest (ROIs). In particular, we defined regions as spheres of 8 mm radius centred on premotor areas that respond to visual speech (Premotor Ventral inferior PMvi/Broca’s xyz = −48 12 9, xyz = −51 9 9; Premotor ventral superior / premotor dorsal PMvs/PMd xyz = −39 3 54, xyz = −48 3 42; BA6 and BA 44 xyz = 48 18 18); visual motion area MT+/V5 (xyz = -42–66 2, xyz = 42–62 6), and sites in the posterior inferior temporal sulcus involved in biological motion processing (pITS xyz = -50–82 0, xyz = 48–78–4). Finally, we considered ROIs also in the fusiform face area (FFA xyz = -34–62–15, xyz = 34–62–15) and in the temporal visual speech area (TVSA xyz = -57–34 14). We applied family-wise-error small-volume-correction (FWE-SVC) to each ROI. We retained results as significant at p \< 0.05 FWE-SVC, further Bonferroni- corrected for the number of regions (n = 12). # Results ## Behavioural results ### Main task: Identification of the rendering mode (Forward or Backward) During the preliminary and fMRI experiments, observers had to indicate whether the video-clip was played in Forward or Backward mode. Observers detected the rendering mode (Forward or Backward) of video-clips with an overall probability of correct responses P{C} = 0.590 and P{C} = 0.556 for the preliminary and fMRI experiments respectively (pooled across participants and stimuli), significantly higher than chance level (two-tailed binomial test, p \< 0.001). Sensitivity (d’) for Italian (d’: 0.56 ± 0.14 and d’: 0.37 ± 0.09, mean ± s.e.m., for preliminary and fMRI experiments respectively) and Arabic (d’: 0.47 ± 0.15 and d’: 0.26 ± 0.08, respectively) was not significantly different (paired t-test; p = 0.52; t(19) = 0.66 and p = 0.16; t(19) = 1.44 for preliminary and fMRI experiments respectively). For both languages, d’ was significantly greater than 0 (one sample t-test; all p \<0.002; t(19) \> 3.25 and all p \<0.004; t(19) \> 3.32 for preliminary and fMRI experiments respectively). However, there was a significant response bias (*c* = -0.25 ± 0.07, p = 0.003, t(19) = 3.34 and *c* = -0.36 ± 0.05, p = 0.001, t(19) = 4.05 one-sample t-test, for preliminary and fMRI experiments respectively) in favour of the response “Forward” for the Italian video-clips, underlying the higher proportion of correct response in this task. By contrast, there was no response bias for the Arabic video-clips (*c* = 0.024 ± 0.06, p = 0.66, t(19) = 0.43 and *c*: -0.01 ± 0.08, p = 0.9, t(19) = 0.12, one-sample t-test, for preliminary and fMRI experiments respectively). The comparison of sensitivity (d’) and response bias (c) indexes between the fMRI and the preliminary experiments, computed for both Italian and Arabic video clips, did not show significant differences (t-test, all t(19) \< 1.28, p\>0.21). ### Follow-up experiment: Language identification All the participants in the fMRI experiment were retested in a follow-up experiment to ascertain their ability to recognize the language by means of visual-only cues. In this experiment, volunteers had to indicate if the actor’s language in the silent video clips was Italian or not. (white bars) reports for each condition the probability of correct responses (P{C} = 0.679) pooled across stimuli and participants. In all conditions, P{C} was significantly higher than chance level (two-tailed binomial test, p \< 0.001). The average d’ was not significantly different between video-clips played in forward (d’: 1.05 ± 0.12) and backward mode (d’: 0.96 ± 0.15) (paired t-test; p = 0.5; t(19) = 0.65) and in both cases d’ \> 0 (one sample t-test; all p \<0.001; t(19) \>7.02). There was a significant response bias in favour of the response “Italian” in the case of Forward video-clip (c: -0.21 ± 0.05, p \< 0.001, t(19) = 4.43, one-sample t-test). Conversely, in the case of Backward video-clips, the response bias was in favour of the response “not Italian” (c: 0.12 ± 0.034, p \< 0.01, t(19) = 2.95, one-sample t-test). ### Comparison between tasks We expected that stimuli were classified differently depending on the task, and that the two tasks had different response patterns across subjects. To verify this hypothesis, we compared the participants’ sensitivity index (d’) in the fMRI main task (rendering mode discrimination) and in the follow-up experiment (language discrimination). The analysis of d’ showed a greater sensitivity to stimuli in the follow-up experiment task compared to stimuli sensitivity in the main task (paired t-test, t(19) = 6.15, p\<0.001). An alternative possibility is that the higher d’ in the second experiment could be due to a learning process occurring after the first experiment. In this case, we should expect a correlation across participants between tasks. However, the sensitivity indexes of the two tasks were not correlated (Pearson’s r = 0.27, p = 0.23), suggesting that the two tasks rely on different processing. ## fMRI results ### Brain areas engaged by visual speech We mapped the cortical regions activated by all visual speech stimuli, irrespective of the parameters manipulated experimentally, (i.e., all stimuli vs. rest) by a differential F-test across the 12 time-bins (see). This test highlights regions having different amplitude and/or time-course of the BOLD response between conditions examined in the contrast image (in this case, all stimuli vs. rest condition). As shown in, significant effects were observed in occipital and temporal cortices, i.e. regions that are typically involved in audio-visual processing, as well as in parieto-frontal cortices, which are generally engaged by the vision of speech movements (Bernstein et al 2014), in the insula, cingulate cortex, motor and premotor areas. Activity in left motor/premotor areas was presumably related, at least in part, to the right-hand motor responses. ### Main effect of actor’s language *Whole brain analysis*. The differential F-test comparing Italian (familiar) vs. Arabic (unfamiliar) stimuli (irrespective of rendering mode) across the 12 time- bins revealed significant activations (p-corr \< 0.05, whole brain) bilaterally in the posterior fusiform gyrus (FG_a, at coordinates almost coinciding with those of FFA), extending to the inferior occipital gyrus (IOG) and right occipito-temporal sulcus (OTS, at coordinates near to those of pITS, a biological motion sensitive area), and in the precuneus (green area,). The time profiles of the estimated BOLD activity in these regions are plotted in. It is important to note that direct inspection of these activity patterns is necessary before any conclusion can be drawn, since significant differences obtained through our statistical analysis (i.e., F-test) could be due to a modulation in amplitude and/or to a time-shift. Time bins presenting a different activity level (post-hoc t-test, p-FDR corrected for multiple comparison \< 0.05 across bins) between Italian and Arabic are filled in green. FG_a showed enhanced earlier activity (see bins 1^th–3^th) for Italian stimuli independently of rendering mode, and later activity for Arabic stimuli (6^th–8^th bins, compare grey and black lines). IOG and OTS, belonging to the same cluster of FG_a, had a similar temporal profile (e.g., OTS). The precuneus showed the opposite trend, with a decreased activity in the earlier bins for Italian stimuli (see black lines in, 3^th–4^th bins) and in the later bins for Arabic stimuli (7^th–8^th bins). *ROI analysis*. A significant effect of the familiar language independently of rendering mode (p-corr \< 0.05, FWE-SVC Bonferroni) was also found in the left PMvs/PMd (green area). These regions showed increased activity for the Italian stimuli independently of rendering mode only in late bins (6^th–9^th bins). Also left FFA and right pITS showed a main effect of language, already reported in the whole brain analysis. By contrast PMvi, MT+/V5 and TVSA did not show a significant main effect of language. ### Main effect of rendering mode *Whole brain analysis*. Regions with differential responses to normal kinematics (i.e., video-clips played forward) and to implausible kinematics (i.e., video- clips played backward) were identified by the contrast Forward vs. Backward mode (main effect of rendering mode), irrespective of language (Italian or Arabic). The regions significantly sensitive to this contrast (orange regions) were found in the intraparietal sulcus (IPS) bilaterally, and in the left posterior-middle fusiform gyrus (FG_b) (p-corr \< 0.05, whole brain). show the time- course of activity in these regions. Bins in which the activity differed significantly between normal and reversed video-clips are filled in orange (post-hoc t-test, p\<0.05 FDR corrected for multiple comparisons across bins). In particular, the right IPS (1^th–2^th bins) and FG_b (1^th–2^th bins) responded more to Forward than Backward rendering mode (compare continuous and dotted lines in, respectively). Left IPS showed a similar trend in the early bins (2^th–3^th bins) at a lower statistical threshold (p-uncorr \< 0.05). Finally, the image resulting from the intersections (logical AND) between the cluster image of the left FG_a (reported above) and the cluster image of left FG_b showed that these two clusters were sharply separated (no voxel in common). *ROI analysis*. A significant effect of the rendering mode independently of language (p-corr \< 0.05, FWE-SVC Bonferroni) was also found in the PMvi, in the pars opercularis of IFG (p-corr \< 0.05, FWE-SVC Bonferroni) (see, Figs and, blue). PMvi responded more during late bins to the rendering modality (8^th–9^th bins), but selectively to Italian language, so that also the interaction calculated on the peak was significant (see also below). By contrast none of the posterior ROIs (MT+/V5, FFA, pITS, TVSA) nor PMvs/PMd showed a main effect of rendering. ### Influence of actor’s language on the rendering mode discrimination process *Whole brain analysis*. Through the contrast (\[I_F − I_B\] vs. \[A_F − A_B\]), we searched for brain sites where the response to rendering mode was affected by language. This analysis revealed significant activations (p-corr \< 0.05) in pars triangularis (BA 45) of the left inferior frontal gyrus (IFG-triang) and in lingual gyrus (LG, see blue regions). The temporal profile of BOLD responses showed that IFG-triang differentiated the Arabic video-clips played in Backward and Forward mode in the earlier bins, in particular Arabic backward stimuli strongly de-activated IFG- triang (3^th–4^th bins) (post-hoc t-test, p-FDR corrected for multiple comparison \< 0.05 across bins). In a similar way, IFG-triang differentiated Forward from Backward Italian video-clips, but at later times compared to Arabic stimuli discrimination, namely between the 8^th and 10^th bins. Moreover, Italian stimuli played backward also showed a marked negative pattern in these bins. Overall, IFG-triang responded similarly to Italian and Arabic stimuli, although the temporal patterns were shifted. Indeed, neither the difference between the maximum peaks for Italian Forward stimuli (bin 8) and Arabic Forward stimuli (bin 3) (t(19) = 0.93; p = 0.36), nor the difference between Forward and Backward condition of Italian stimuli at bin 8 and that of Arabic stimuli at bin 3 (t(19) 0.12; p \> 0.9) were significantly different (compare the differences between continuous and dotted grey lines in bin 8 and between continuous and dotted black lines in bin 3, respectively). In sum, the BOLD patterns showed that IFG-triang does not have a clear preferential response to the speech gestures most frequently performed by participants (i.e., Italian Forward stimuli). LG showed a general de-activation in all four conditions versus rest. In particular, Italian Backward and Arabic Forward stimuli involved a very similar time-course, as did Arabic video-clips played backwards and Italian video-clips played forwards. These two latter conditions were also the two most deactivating (i.e., negative BOLD patterns) conditions in this site (see 3^th–4^th bins filled in blue) (post-hoc t-test, p-FDR corrected for multiple comparison \< 0.05 across bins). *ROI analysis*. This analysis revealed a significant interaction between language and rendering mode in PMvi (IFG pars opercularis), a region that was selective for the Italian language in late bins (8^th–9^th bins) (Figs and, blue). In particular, there was a single peak of higher response to the familiar language with respect to the other three conditions, indicating a clear preferential response to the speech gestures most frequently performed by participants. By contrast, none of the posterior ROIs (MT+/V5, FFA, pITS, TVSA) nor PMvs/PMd showed an interaction between language and rendering mode. Finally, we calculated a minimum effect size of 0.15 corresponding to the lowest significant F value reported in (F(11,209) = 3.26). A partial eta-squared of 0.15 indicates a large effect. # Discussion We reported differential brain responses to visual speech kinematics, language familiarity and their interaction. Neuroimaging data showed that language familiarity and temporal rendering of silent speech video-clips modulated two distinct areas in the ventral occipito-temporal cortex. Furthermore, language familiarity modulated the left dorsal premotor cortex, while natural familiar language activated the left ventral premotor cortex in the frontal operculum. These results may indicate that phono-articulatory regions resonate in response to the visemes (visual equivalents of phonemes) of a familiar language. Since in our experiments participants generally did not decode the semantic and syntactic content of visual speech, we propose that these results are confined to the visual equivalent of the phonemic axis. Indeed, our results are in agreement with the definition of a phonological pathway more dorsal with respect to the lexical and semantic pathways, which includes IPS, the dorsal premotor region and the pars opercularis of IFG. ## Sensitivity to the time-arrow of visual speech Participants were able to discriminate above chance level visual speech gestures rendered forwards from those rendered backwards. Behavioural results were consistent across experiments. Noteworthy, the sensitivity index d’ estimated in the preliminary and in the main fMRI experiments were not statistically different. These results indicate a sensitivity of the central nervous system for temporal features (i.e., time arrow) of the visible speech, in agreement with the results obtained with a familiar language in a previous study with a similar task. Although lip-reading accuracy of hearing people is generally low and idiosyncratic, one cannot rule out a priori that observers were occasionally able to lip-read excerpts of Italian texts in the forward mode, and to use these instances as a cue for discriminating the rendering mode. However, the fact that sensitivity was not significantly different for Italian and Arabic stimuli suggests that lexical competence and speech intelligibility did not play a significant role in the task. The evidence suggests instead that better than chance performance was achieved mainly by a kinematic analysis of movements. If so, the performance reflected the ability to discriminate the motor sequences that are visually perceived as plausible from those that are perceived as implausible from the motoric point of view. This assumption can be sharpened by taking into account the response bias, which describes the position, along the decision axis, of the internal threshold for discriminating the stimuli. In our experiment, there was a response bias in favour of the response “Forward” for the Italian but not for Arabic stimuli, indicating a corresponding shift of the threshold to higher values. This invites the inference that in order to classify a movie reversed in time as ‘Backward’, it is necessary to detect more motoric incongruences in Italian than in Arabic stimuli. This inference is in keeping with the suggestion that a high threshold for detecting speech kinematic anomalies favours the stability of speech perception in environments where such anomalies in a familiar language occur due to inter-individual differences or phonetic peculiarities typical of particular social environments (regional inflexions, slang, etc.). Indeed, the participants to both the preliminary and fMRI experiments had no reason to suspect that in half of the video-clips the language being spoken was not Italian. In a follow-up experiment, participants were asked to identify the language (Italian or not Italian) spoken by the actors in the silent movies. If participants benefitted from speech intelligibility, then the sensitivity to discriminate languages should be higher for Forward than that for Backward rendered stimuli. The results of the follow-up experiment do not conform with this scenario, because the sensitivity index d’ was comparable for Forward and Backward rendered video-clips , making unlikely that speech intelligibility or lexical processes occurred in our tasks. The lack of significant correlation across participants between the sensitivity in the fMRI and follow-up experiments suggests that different processes, likely taking into account non-overlapping sets of cues, underlie language and rendering mode discrimination. Conversely, the finding that response bias during rendering mode discrimination depended on the language (and vice-versa) might indicate that, during a late stage of analysis, these signals are merged. This merging might take place in the premotor cortex, where the PMvs/PMd selected the familiar language independently of rendering, while the PMvi responded to the Italian language selectively in the natural kinematics condition. ## Ventral occipito-temporal cortex (vOTC) The main effects of language and rendering mode activated distinct regions in ventral occipito-temporal cortex. In particular, the comparison of Forward with Backward conditions showed a differential pattern of BOLD responses in the left posterior-middle fusiform gyrus (FG_b), while the posterior fusiform bilaterally (FG_a), the inferior occipital gyri and the occipito- temporal sulcus (OTS) were differentially involved with Italian versus Arabic video-clips. The fusiform sites are located posteriorly to the visual areas engaged by semantic and/or lexical processes in the vOTC, which are typically reported at y-coordinates \< -50 mm, in the anterior part of the fusiform gyrus \[see \]. Conversely, the fusiform face area (FFA), a region responding selectively to static faces, is located in the posterior region of the fusiform gyrus. The stereotaxic coordinates of FFA centre of mass (\[ – –\],) roughly correspond to those of the peak of FG_a reported here, but are posterior to those we found for the main effect of rendering mode (FG_b). Previous studies reported that multiple sites in FG respond to faces. It is likely that different foci in FG encompass distinct functional modules, as suggested by early PET studies showing that gender and face identification activated distinct regions in posterior and middle fusiform gyrus, respectively. It has been shown that the kinematics of biological movements, as well as the temporal unfolding of faces that express an emotional state engage ventral OTC. In particular, observing facial speech gestures activates FG, although it is unclear whether these activations are specific for speech because some control stimuli, such as gurning faces, activated this region more than talking faces. Indeed, the difference between speech and control stimuli may have been due to differences in low-level features, such as visual motion speed. In our study, all four experimental conditions (showing exclusively the lower portion of faces) were comparable in terms of low-level features, such as mean luminance and motion speed. By contrast, time reversal of visual speech stimuli violates motor constraints, and hence produces movements with an implausible kinematics, never occurring during real speech. Previous studies showed that coherent sequences of facial expressions engage the posterior fusiform gyrus more than a scrambled sequence. A possibility is that ventral OTC processed specific kinematic cues embedded in visible speech. Therefore, we speculate that the posterior-middle fusiform site (i.e., FG_b) was sensitive to the kinematic plausibility of speech gestures. Conversely, the more posterior site FG_a was involved mainly in processing kinematic features related to the familiarity of speech, such as the rhythm of speech that is invariant under time reversal but differs across languages. The previous observations challenge one of the most prominent models about face processing, namely the model proposing that static and dynamic face features are processed separately in ventral OTC and STS, respectively. In fact, our data suggest that ventral OTC has foci sensitive to spatio-temporal (i.e. changeable) characteristics of speech lip movements. Italian-speech stimuli evoked a higher activity peak than Arabic-speech stimuli in the more posterior site of fusiform gyrus (FG_a), whereas the sustained post-peak activity was greater for Arabic than Italian stimuli. The Arabic-speech stimuli were unfamiliar, and thus they were presumably unexpected. It is thought that a specific class of neurons (the so-called error neurons) responds selectively to unexpected or unusual stimuli. These neurons compare the sensory input with an internally generated (prediction) signal coding what is expected in a given context. In case of a mismatch between the predicted and the incoming sensory signal, error units enhance their activity. We surmise that the greater activity for Arabic than Italian stimuli in the post-peak period might be related to the activity of error units. Therefore, depending on the language (familiar or unfamiliar), this class of neurons contributed differently to the overall neural activity in FG_a, so the pattern of neural activity changes according with language. However, this mechanism is not specific to ventral OTC, but is a widespread mechanism governing several brain processes (see Friston 2010). For instance, we recently surmised that this kind of neural processing occurs also in lateral OTC when observing unfamiliar walking movements, and it might even bias balance control. ## Lingual gyrus The interaction between language and rendering mode showed significant activations in the lingual gyrus. Previous reports have already suggested that visually presented speech gestures engage this site. The novel finding reported here is that LG responds differently depending on the language. In the follow-up experiment (language discrimination), the d’ was similar between Forward and Backward stimuli, while there was a significant bias toward Arabic-speech or Italian-speech response in the Backward or Forward condition, respectively. The temporal profile of LG activity distinguishes the experimental conditions. In particular, the conditions Arabic Backward and Italian Forward were the two most deactivating conditions. Therefore, the response bias in language discrimination task could be related to a decrease of LG activity. Interestingly, in the auditory domain, LG activity has been found to depend on the familiarity of the spoken language. In the latter study, hearing a speech segment in a second language modulated LG activity differently depending on the participant’s proficiency in that language. Because LG is involved, together with fronto- parietal regions, in speech control, these findings suggest a supra-modal effect of speech language in LG, probably due to feedback from high-order centres. ## Premotor prefrontal activity reflects motor simulation of speech gestures The interaction of rendering mode and language showed a significant effect in the left IFG, comprising the pars opercularis (BA 44), which we have labelled as PMvi, and the pars triangularis (BA45). Most authors agree that the Broca’s area includes both BA 44 and BA 45 of the left hemisphere. Broca’s area was initially thought to be involved only in speech production, but current research shows that it has a more complex role possibly involving also speech comprehension. In particular, the activation of IFG during speech perception has been interpreted by some authors as the occurrence of a motor simulation of the observed movements. This idea is in accordance with the motor theory of language holding that a simulation of speech gestures in the motor regions is instrumental for speech perception and understanding. However, it is still an open issue whether the IFG activation during speech perception is related to a language-specific process, as the putative motor simulation of speech gestures, or represents a general-domain cognitive mechanism. It has also been suggested that distinct IFG foci have different roles during speech perception. Specifically, articulatory rehearsal of speech gestures would occur in pars opercularis, while pars triangularis and orbitalis could be related to cognitive-control mechanisms, such as decisional or working-memory processes. The rehearsal function of pars opercularis generalizes across different types of movement, as this region was found to respond also to observation of hand movements. The time-course of the BOLD signal that we observed in the pars opercularis of IFG fit with the predictions of the motor simulation hypothesis (Figs). According to this hypothesis, the familiar stimuli (i.e., Italian Forward) should elicit a higher level of activity than unfamiliar, implausible gestures difficult to reproduce (e.g., Italian Backward or Arabic Forward and Backward stimuli). Conversely, in the pars triangularis of IFG, Italian-speech stimuli and Arabic-speech stimuli, for which latter participants had no motoric expertise, evoked comparable responses although shifted in time (Figs). Thus, the results do not suggest a specific sensitivity for Italian-speech stimuli in the pars triangularis of IFG, but rather sensitivity for the kinematics of natural mouth movements, a kind of biological motion. Our data, limited to the case of a familiar language (Italian), are also in agreement with those reported by Paulesu et al. in which IFG activity was greater for forward than backward silent movies of a speech in a familiar language. The issue of the role of intelligibility of silent visual speech should be further investigated, as one could argue that motor simulations occur only or mainly when linguistic competences are required, as with a lexical discrimination task. However, we believe that the ability of participants to speech-read might be a confound when trying to disentangle motor from higher cognitive functions of Broca’s area. In our case, it appears that motor simulation occurs in absence of comprehension of the content of the speech. ## Summary and conclusions Previous studies focused mainly on the role of temporal auditory regions and frontal regions in processing visual speech. More recently, it has been shown that a region in the left posterior temporal cortex, the so-called temporal visual speech area (TVSA), is activated in visual phonetic discrimination, possibly integrating information coming from high-level visual areas in OTC. We did not find significant effects in TSVA, as verified through a specific ROI drawn in this region. Our data suggest that the ventral occipito-temporal cortex has a sensitivity to visual speech gestures, contrary to the view that the peculiar analysis of visual speech starts at higher cortical levels. Our results support the hypothesis that kinematic cues embedded in visible speech can be extracted through the visual pathways, outside the classical areas related to auditory speech and audio-visual integration. Finally, the selective responses of PMvs / PMd to the familiar language and of PMvi to the natural familiar language support the hypothesis that motor simulation drives premotor activity during visible speech perception. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Fragmentation and fissuring of the medial coronoid process (MCP), as well as, pathological lesions of the cartilage and the subchondral bone of the medial coronoid process are part of medial coronoid disease (MCD). MCD alongside three other clinical pictures such as the ununited anconeal process, osteochondritis dissecans (OCD) and elbow incongruity are part of elbow joint dysplasia. Fragmented medial coronoid process is the most common manifestation of elbow dysplasia (ED) and is also a common cause of thoracic limb lameness in medium to large, rapidly growing breeds of dogs with an increased prevalence in males. Most cases first present between 5 months and 12 months of age, and depending on the literature information, there is another group with an incidence of 12% showing first signs of a forelimb lameness at 6 years or older. The clinical picture often shows amongst other pathologies, the following pathological findings in the affected limb: lameness and relief posture, external rotation and pain when there is applied pressure on the medial coronoid. Another interesting finding, described by Moores et al. is that 50% of their study- population of 50 dogs showed abnormalities of the medial coronoid process without any clinical signs of lameness. Abnormalities were described as computed tomographic findings in form of fragmentation, fissures, sclerosis or hypoattenuation as well as an abnormal shape and irregular radial incisures. The aetiology of MCD is undetermined at present, but the literature agrees that it has a multifactorial origin. The implicated factors are genetic dispositions, abnormalities of the underlying subchondral bone and abnormal mechanical loading . Furthermore, other environmental factors such as exercise, nutrition, microtrauma and mineral imbalance cannot be ruled out as to be relevant. Radioulnar incongruity also seems to play an important role. Regardless of a variable sensitivity, radiographic examination is always the first choice for ED screening in practice. However, due to both superimposition of the radial head over the medial coronoid process and osteophytes, a correct assessment of the MCP is not always possible. Furthermore, a tight fit between the ulnar trochlear notch and the humeral condyle complicates the correct diagnosis. Often a suspected diagnosis of MCD can only be made based on secondary changes such as osteophytes, blurring of the cranial coronoid contour and sclerosis of the ulnar notch. Further examinations in the form of computer tomographic diagnostics are necessary to confirm the findings, which offer the advantage of more clarity, since images are not superimposed on each other and can be evaluated in different reconstructed views. Nevertheless, even by combining the two diagnostic tools radiography and CT, we do not have 100% reliable information about all bone details and of the integrity of articular cartilage. In the literature, many different methods of treating MCD are described, ranging from conservative to minimally invasive and invasive osteotomy or ostectomy. There is no official, unambiguous "protocol" for the therapy of MCD, but all options pursue a return to normal function, ameliorate pain and a slowing of the progression of osteoarthritis. Fragment removal is often recommended, but even if a fragmentation of the medial coronoid process is present, surgical removal of the fragment does not always guarantee a good outcome. This implies that the fragmentation is not necessarily the sole cause of clinical signs and it is still possible, that arthrosis will progress. It is assumed that the removal might result to a load redistribution to either the remaining portion of the medial contact area, or to the lateral elbow contact area including the radial head. Subsequently, this could potentially accelerate cartilage degeneration or cause subchondral pathology at these sites. The incongruence of the radioulnar articular surfaces may lead to an overload of the medial part of the elbow joint, and seems to play a non-negligible role in the clinical symptoms and prognosis. After the removal of the fragment, the incongruence remains and can therefore also influence the outcome. In addition, when comparing different surgical methods regarding fragment removal, the outcomes are very variable. In most of these studies no significant improvement of long-term function after fragment removal is observed. It is still ambiguous as to what affects the outcome of dogs with MCD. Many factors such as severity and duration of lameness at the time of presentation, the degree of cartilage damage and osteoarthritis and the type of lesions present in the joint, could all affect outcome and prognosis. The purpose of the current study was to evaluate if there are radiographic and CT findings which can explain the discrepancy of a unilateral forelimb lameness despite a bilateral diagnosed MCD. We hypothesised that there is a significant difference between the two diseased elbow joints on radiographs and CT images which could explain a unilateral forelimb lameness. # Materials and methods ## Inclusion and exclusion criteria Clinical records of the database of the Small Animal Teaching Hospital at the University of Veterinary Medicine Hannover were reviewed from February 2014 to April 2019 to identify dogs diagnosed with bilateral MCD. Dogs were eligible for participation in the study if they had a bilateral diagnosed MCD, where one forelimb presented no clinical signs and the other forelimb showed lameness and pain, swelling or crepitus during the orthopaedic examination of the elbow joint. The diagnosis was confirmed by radiographic and CT exams’. Inclusion required furthermore a complete documentation of radiographs, CTs as well as a subjective gait assessment. All the data was grouped together for each dog including name, date of birth, breed, sex, weight, age at diagnosis of MCD and hospital identification number. Cases with incomplete medical records were excluded. Exclusion criteria were concurrent elbow joint pathology such as ununited anconeal process, osteochondritis of the medial humeral condyle and flexor tendon enthesiopathy. ## Radiographs Radiographs of each elbow joint were taken and included a mediolateral flexed and craniocaudal view. The radiographs were scored for osteoarthritis based on the IEWG guidelines. The score was modified in such a way that only the size of the osteophytes and thus the indication of arthrosis was assessed, since otherwise many radiographs images would have been assessed directly as score 2 based on our diagnosis of MCD. Incongruence wasn’t included in our modified assessment and as the dogs with concurrent elbow joint pathology such as ununited anconeal process or OCD were already excluded, these signals were irrelevant for the evaluation. Furthermore, Trochlear notch sclerosis (TNS) which describes a radiological term of increased bone radio-opacity in the region of the ulnar trochlear notch, was quantified. The measurements were performed as described by Draffan et al and the overall TNS ratio of sclerosis to ulnar depth was then calculated. ## Computer tomography CT imaging was performed of both elbow joints from each dog with a Philips brilliance 64 slice scanner (Philips medical systems technologies LTD, Haifa Israel). Parameters varied depending on bodyweight, but most images were obtained with slice thickness of 1mm, pitch of 0.579, rotation time of 0.75 second, 120 kV and 200 mAs/slice using a bone algorithm. In order to perform the CT examination, dogs were premedicated using Acepromazin and Levomethadon, and anaesthetized using Propofol and Isoflurane in oxygen. All dogs were positioned in sternal recumbency with the front limbs extended cranially with an angle between 90° and 120° as described by Shimizu et al. To avoid interference the head of each dog was pulled back. Both elbows were scanned simultaneously. Several parameters were scored as described in, as well as, the size of the fragment was measured in cm². All measurements were made using a commercial imaging software (Easy Image, Denvis (CoSi dental GmbH, Sigmaringen Germany)). ## Statistical analysis Statistical analysis were performed using the software R version 3.6.0 (2019-04-26). Following fundamental variables were recorded for the statistical calculations: A total of 42 dogs were considered. The statistical analysis was performed using 2 data sets, which contained information of the radiographs using the modified IEWG Score and TNS. Each forelimb was regarded separately so in total 84 radiographs were available for the analysis. The other data set contained information from the computed tomographic imaging, which also included 84 scans. Furthermore, data of each dog: age, breed, sex, weight, age at diagnosis and group (non-lame limb, vs. lame limb) were considered. Descriptive statistics were generated for all variables: Metric, nearly normally distributed variables were described using the mean value (MV) and standard deviation (SD) and compared using a *t* test or Kruskal Wallis test. In contrast, skewed variables were described using the more robust median and interquartile range and checked for equal positional distribution using a nonparametric test, the Wilcoxon rank sum test or Kruskal Wallis test. Categorial variables were described using absolute (N) and relative frequencies % and compared using the χ2 independence test. A p-value of \< 0.05 was considered as significant. # Results Forty-two dogs (84 elbow joints) with bilateral diagnosed MCD met the inclusion criteria. The most common breed were crossbreeds (10 dogs), the second most common were Labrador Retrievers (9 dogs) followed by Rottweiler (4 dogs) and Airedale Terrier (3 dogs). Bodyweight ranged from 10.5 to 68.5 kg (median: 33.6kg, SD: 10.5). The gender distribution in the study-population was as follows: 21 males, 9 males neutered, 8 females and 4 female neutered dogs. Age at diagnosis ranged from 7 to 104 months (median: 36.7 months, SD: 31.3). Depending on the gait of each dog we established two groups to examine the collected data. Group I: affected non-lame limb. Group II: affected lame limb, showing a lameness degree between a slight intermittent lameness up to a continuous non weight bearing lameness. The two groups were compared concerning the radiographic and CT differences based on this classification. ## Radiographs Radiographs were evaluated regarding the modified IEWG score and the TNS ratio. The MV of the TNS ratio of the forelimbs from group I was 0.460 compared to the forelimbs from group II, which had an average value of 0.481. The median, as well as the 1st and 3rd quartile of the TNS ratio were also smaller in group I, but there was no significant difference (p = 0.072). Twenty five percent of the forelimbs from group I had a value less than 0.430 and 25% had a value greater than 0.498. Hence, 50% had a ratio between 0.430 and 0.498. In group II 25% of the forelimbs had a value less than 0.442 and 25% had a value greater than 0.530. Consequently 50% of the forelimbs had a TNS ratio between 0.442 and 0.530. In the analysis of correlation, 22 forelimbs (52.4%) of the whole study population (independent of the grouping) had the same modified IEWG Score. Regarding the relative frequency of the remaining population, 16.7% of group I had a lower modified IEWG score, while group II had in 31% a higher IEWG score. In comparison, forelimbs from group II had a modified IEWG score of 2 almost twice as often as those from group I, and almost three times as often a modified IEWG score of 3. The percentage distribution of the modified IEWG scores of 0 and 1 is as follows: group I: IEWG 0 = 57.1%, IEWG 1 = 33.3%; group II: IWEG 0 = 52.4%, IEWG 1 = 23.8%. Looking at the frequency distribution of the modified IEWG score stratified by age group (4–12 months versus (vs.) \> 71 months), no clear correlation could be found. Those two age groups were applied to this study due to several different research articles and literature in which the authors outlined that especially in these two age groups the MCD is present. ## Computer tomography To evaluate CT changes, one CT scan per limb was available for each dog, a total of 84 scans. These were evaluated according to pathology, type of fragmented coronoid process (FCP), shape of the coronoid process and dislocation. Additionally, the size of the fragment was calculated for 82 elbow joints. Independent of the clinical degree of lameness, 60 elbow joints (71.4%) showed a single fragment. 11 forelimbs (13.1%) had a fissure and in 10 cases (11.9%) multiple fragments were diagnosed. The type of FCP was a coronoid tip in 52.4% and 61% of the fragments were dislocated. The average size of the fragment was 0.159 cm² +/- 0.129. ## Pathology Of all the dogs included (84 CT scans)—independent of the grouping—54 joints (64%) showed the same pathology in the CT scans. The following pathology occurred in Group I and Group II respectively: 24 times a single fragment, two times a fissure and once a single fragment. However, the single fragment was the most common pathology: 29/42 forelimbs from group I (69%) and 31/42 forelimbs from group II (73.8%) showed this pathology. Within the group of elbows without a lameness (group I), 21.4% (9 forelimbs) were diagnosed with a fissure on CT. Eight forelimbs (19%) of group II had multiple fragments. The remaining pathologies were rare (0–4.8%) in both groups. ## Type of MCD In both groups, 42.8% had the same type of MDC. In group I the tip fragment / fissure occurred most often with 66.7%, the remaining forelimbs of this group showed a radial incisure fragment or fissure in 26.2% and 7.1% of the forelimbs had a combination of radial incisure and tip fragment / fissure. Both a radial incisure fragment or fissure, and a tip fragment or fissure were diagnosed in 38.1% of the forelimbs in group II. The remaining 23.8% had a combination of radial incisure and tip fragment or fissure. A significant correlation (p = 0.019) was found between the different types of MCD. ## Shape Each fragment can solely be classified either as flattened, round, pointed or irregular shape. Regarding the shape of the fragment, 26% of the forelimbs showed a coincident form in the two groups. The irregular shape occurred in group I in 7.1% of the forelimbs, and in 16.7% of the forelimbs of group II. In group I the pointed form is dominant with 35.7%, while in group II this is the second rarest shape with 21.4%. Regarding the flat and round form, the distributions between the two groups were as followed: group I: flat 26.2%; round 31%. Group II: flat: 33.3%, round: 28.6%. ## Dislocation Nineteen percentage of the forelimbs from group I had a dislocated fragment, compared to 59.5% of the forelimbs in group II. When considered paired, 42.9% of all elbows had a dislocation of the fragment in group II, whereas no dislocation was present in the forelimbs of group I. In 16.7% of the forelimbs there was a dislocation of the fragment in both group I and group II. The p-value \< 0.001 showed a significant difference regarding the dislocation of the fragment. One forelimb had a dislocation on the elbow which showed no lameness while the other limb showing a clinical visible lameness without a displaced fragment. ## Size of the fragment Forelimbs from group II, in regards to the median and the MV fragment sizes, had almost twice as large compared to those from group I (Group I: Median 0.09cm²; MV 0.11cm². Group II: Median 0.16cm²; MV 0.20cm²). The 1st quartile, 3rd quartile and maximum also took higher values in group II in comparison to group I (Tables and). Regarding the categories ‘fragment dislocation’ and ‘fragment size’ in relation the results of and the Boxplot. (Supporting information) show that there is a local significance (p = 0.001). A larger fragment is more likely to dislocate than a smaller fragment. # Discussion This is currently the only study which analyses the disease pattern of MCD comparatively showing two different clinical pictures in one dog. Comparing to other studies, the median age of 36.7 (SD: 31.1) months of dogs at diagnosis was above the average and the median weight of 33.6kg (SD: 10.5) was similar to other studies. Crossbreed dogs, Labrador Retrievers and male dogs were over- represented which is mirrored in other studies. The aim of the current study was to evaluate whether there are different radiographic or CT imaging findings which can explain a clinically unilateral lameness despite bilateral diagnosed MCD. The IEWG score was higher by 31% on the side of the lameness compared to the non-lame limb. However, it should be noted that both limbs had at least in 50% of cases a modified IEWG Score of 0 (57.1% vs. 52.4%). TNS values were only slightly deviated by 0.021 (0.460 group I vs. 0.481 group II), which seems to make it a worse clinical criteria to guide therapy. CT imaging might provide a better way to differentiate the two groups. A dislocated fragment, diagnosed in 59.5% of the forelimbs showing a lameness (group II), can cause lucent defects in the subchondral bone and explain pain and lameness. In this study, it could be shown that a larger fragment is more likely to dislocate than a smaller one (p = 0.001). Furthermore, there was a significant difference regarding the size of the fragment of the two groups (p = 0.001). The fragment was almost twice as large on the forelimb showing a lameness. However, a cut-off value for the size of the fragment could not be established. This study could not demonstrate an effect of the shape of the fragment, but this might be due to the measurement techniques used. It was not always simple to visualise the whole fragment properly. A tip fragment or fissure seems to occur more often in forelimbs without a clinical visible lameness. This could be explained by the smaller size of the (fissured or fractured) fragment compared to a radial incisures, causing a smaller surface of the unstable or fractured fragment facing the joint-space. Another explanation which was shown as a significant factor by Baud et al. was that the radial incisure fragment was associated to a narrower radioulnar joint space. This in turn might influence the joint mechanism negatively and cause a clinically worse lameness. It seems that this pattern was shown in the results of this study, as well shown by the p-value represented in. However, the results of group II from a stand-alone perspective showed a more or less stable distribution with regard to the type of the fragmented coronoid process, as also outlined in. Nevertheless, a study elaborated by Baud et al. supports the pattern that a fissure or fragmentation of the radial incisure is more often present in dogs with a lameness that seems also present in this study since 26 forelimbs out of group II showed a radial incisure tip fragment or fissure or a combination of both which represents almost 62% of the whole population. Since only a purely clinical lameness examination was carried out and not an objective gait analysis using plate measurement for ground reaction forces, it must be questioned whether the gait analysis performed in this study was too insensitive. It is possible that the more painful limb masked the less painful limb in the clinical gait examination, so that only unilateral lameness was diagnosed by the examining veterinarians. It has already been described in the literature that dogs showed only unilateral lameness despite bilaterally diagnosed MCD and that this disease is complex and the presentation of clinical sings can be intermittent or constant. Concerning the clinical picture of an intermittent lameness this might be comparable to OCD, which however, has a different aetiology. While the cause of OCD is a disruption of the enchondral ossification of the articular cartilage, numerous pathophysiological mechanism for MCD are postulated in the literature, finally causing a lesion of both the articular cartilage and the subchondral bone. This results in a more or less loose fragment which can cause—depending of the position–an intermittent lameness. Another limitation of the study was that arthroscopy was not performed on both sides. Arthroscopy is considered the ‘‘gold standard” technique for clinical evaluation of cartilage lesions. Arthroscopy can provide a good visualisation of the articular cartilage, consequently the assessment of the integrity of articular cartilage, but not the subchondral bone. Moreover, the detection of smaller fragments would be possible, which may be purely cartilaginous, rather than osteochondral, and therefore not detectable by CT. On the other hand, the study of Morres et al illustrates a significant correlation between the CT osteophyte score and the arthroscopic cartilage erosion score for the axial and the abaxial part of the MCP as well as the entire part of the MCP. This finding could question the indication for an arthroscopy. Moreover, an arthroscopic intervention may also cause progressive osteoarthritis and cartilage damage. A further factor which could have been a significant parameter for the aim of this study is the radioulnar incongruence. As plain radiographs are unreliable for the detection of elbow in-congruency, reconstructed CT scans out of sagittal and dorsal plane images should have been considered for a more accurate evaluation of the incongruence. As there was not always an exactly similar positioning of the dogs during the CT scans and due to the non-loaded limb nature of the CT procedure, validity of imaging might need to be questioned. Future studies should evaluate how loading and movement of the limb could affect CT imaging. There are already cadaver studies evaluating this effect in radiographs but they miss in CT. # Conclusion In summary, a decision tree for the appropriate therapy could not be determined. Though the mentioned findings are seminal and directional parameters, which could explain the discrepancy between a clinically unaffected and a lame limb despite a radiographic MCD diagnosis. Especially the evaluation of the modified IEWG score, the dislocation and the size of the fragment should not be neglected when examining this disease pattern. Although there are hints to what could explain partially the unequal clinical picture in the pathogenesis of MCD, understanding of the exact cause, especially for a better therapeutic approach is incomplete. More studies are urgently required to understand this complex disease pattern, following a problem- oriented therapy can be applied. Nonetheless each patient should be considered individually. # Supporting information The authors are grateful to Chrestos Institut for their statistical assistance. [^1]: The authors have declared that no competing interests exist.

# Introduction The breeding of animals using genetically engineered (GE) technology has recently become possible. This process could avoid time-consuming artificial hybridization breeding and pure breeding programs. To date, many GE animals have been produced, including fish, mice, rabbits, sheep, pigs, cows, and goats. However, the safety evaluation of GE animals and food products should be considered seriously, especially with regard to potential effects on microflora through possible horizontal gene transfer. Horizontal gene transfer, also known as lateral gene transfer, refers to the transfer of genes between different species, such as between prokaryotes and eukaryotes in a manner other than traditional reproduction. According to former reports, this phenomenon can take place between different species, including between bacteria and bacteria, between plants and bacteria, and between animals and plants. The microbial community of the gastrointestinal tract is closely associated with the host metabolism and has a complex and sensitive construction. Microflora may thus be an important intermediate by which horizontal gene transfer reaches other more advanced organisms. To date, horizontal gene transfer between GE animals and bacteria has not been reported. However, further evidence is required to investigate this issue given the significant concerns. It is important to determine whether the structure of gastrointestinal bacterial flora could be rearranged following the insertion of foreign genes into GE animals and their alteration of the host metabolism. Moreover, soil contains various types of bacteria, and the bacteria in the gastrointestinal tract can also enter the environmental soil in the form of feces produced by GE animals. Any changes in the gastrointestinal bacterial flora could thus conceivably also influence the surrounding environmental soil flora. To enhance the milk production of goats, we previously generated transgenic goats over-expressing goat growth hormone (*GH*) with beta-lactoglobulin promoter in their mammary glands by somatic cell nuclear transfer (SCNT). *GH* transgenic goats were confirmed by PCR analysis and verified the transgenic copy number and integration sites. Here, we focused on the effects of the *GH* transgenic goat on the microflora of the intestine, feces and surrounding soil. # Materials and Methods ## Ethics statement This study was approved by the Ethical Committee of Animal Experiments of the College of Veterinary Medicine at Nanjing Agricultural University. All animal care and use procedures were conducted in strict accordance with the Animal Research Committee guidelines of the College of Veterinary Medicine at Nanjing Agricultural University. All sections of this experiment adhere to the ARRIVE Guidelines for reporting animal research. A completed ARRIVE guidelines checklist is included in. ## Experimental animals and sample methods Female *GH* transgenic and non-transgenic Saanen dairy goats were raised on a farm in the Transgenic Research Center, Shanghai, China. All the goats were healthy and fed with the same fodder. During the entire experimental period in autumn, the goats were given ad libitum access to feed and water. The room temperature was maintained at 25–27°C. During housing, all animals were monitored twice daily to assess their health status. No adverse events were observed. Feces were taken from GH transgenic and non-transgenic goats, and each sample was taken when it just been defecated. Soil samples were taken from 0 m to 150 m from the GH transgenic goats’ pen with 15m in width and 30m in length. We slaughtered the goats and cut the intestine lengthwise to collect the intestinal contents from jejunum and cecum. The feces and soil samples were collected in three replications and mixed in one centrifuge tube, and the intestinal contents samples were collected only once. The samples details were performed in. All the fresh samples are stored in -70°C before analysis. ## DNA extraction and PCR detection of target DNA Microbial community DNA extraction of the fecal and intestinal samples was performed using the TIANamp Stool DNA Kit (Tiangen, China). The soil microbial community DNA extraction was performed using the EZNA Soil DNA Kit (Omega, USA). PCR amplifications of the *GH* and *neoR* gene fragments from the feces, soil and intestinal content samples were carried out, with positive and blank controls included in all procedures. The primers used were `GH-F CATCCAGAAGGAATTCATGATGGCT, GH-R AGGGTCGACCTAGAAGGCACAGCT, neoR-F CCTGTCATCTCACCTTGCTCCT` and `neoR-R ATACCTGTCCGCCTTTCTCCCT.` The PCR amplifications were carried out in 20 μl reaction volumes comprised of 10 μl of 2 × Taq Master Mix, 1 μl of each primer (10 μM), 0.2 μg of temple DNA and added ddH₂O to 20 μl. Each target gene was amplified with an initial denaturation of DNA at 94°C for 10 min, followed by 26 cycles of 30 s of denaturation at 94°C, 60 s annealing at 60°C and 60 s of elongation at 72°C, with a final elongation for 10 min at the same temperature. PCR products were visualized by electrophoresis. ## Polymerase chain reaction-denaturing gradient gel electrophoresis (PCR-DGGE) analysis PCR amplification of the variable V3 region of bacterial 16S rDNA was performed with a pair of universal primers (`338F 5’-ACTCCTACGGGAGGCAGCAG–3’` and `518R 5’-ATTACCGCGGCTGCTGG–3`) with a GC clamp of 39 bases (`CGCCCGCCGCGCGCGGCGGGCGGGGCGGGGGCACGGGGGG`) added to the 5’-terminus. PCR amplification was carried out in 50 μl reaction volumes, composed of 5 μl of 10 × PCR buffer, 4 μl of dNTP mixture, 1 μl of each primer (20 μM), 0.25 μl rTaq polymerase (5 U/μl), 2.5 ng of temple DNA and ddH₂O to 50 μl. The 16S rDNA genes were amplified with an initial denaturation of DNA at 94°C for 10 min, followed by 30 cycles of 60 s of denaturation at 94°C, 60 s annealing at 55°C, 90 s of elongation at 72°C, and a final elongation for 10 min at the same temperature. PCR products were subjected to DGGE analysis with the Dcode Universal Mutation Detection System (Bio-Rad, Hercules, USA). The PCR products were loaded onto 8% (wt/vol) polyacrylamide (37:1 acrylamide/bisacrylamide) gels in 1× TAE buffer with a denaturing gradient ranging from 30 to 60%. The gels were run at 150 V for 420 min and then silver stained. ## Cloning and sequencing The PCR products of the 16S rDNA from prominent bands were recovered with Gel Recovery Purification kits (Watson, China) and ligated into the pMD19-T vector (Takara, China). Then, these recombinant plasmids were transformed into *Escherichia coli* DH5α. Three or five clones were randomly selected for each band. These products were sequenced with an ABI Prism Big Dye Terminator Cycle Sequencing Ready Kit (Applied Biosystems, USA) with a pair of universal primers for the pMD19-T vector: `M13F (-40) GTTTCCCAGTCACGAC` and `M13R (-26) CAGGAAACAGCTATGAC.` ## Phylogenetic analysis The analysis of 16S rDNA gene sequences was performed using the database of the National Center for Biotechnology Information (NCBI) to acquire closely related sequences. Phylogenetic trees were constructed with MEGA 6.0 on the basis of 97% similarity by the bootstrapped neighbor-joining method with 1000 iterations. The similarity in the microorganism was compared by cluster analysis. Cluster analysis was performed based on Dice’s algorithm with the BioEdit 7.0, Phylip 4.0, and MEGA 4.0 software. # Results ## Detection of the transgene and marker gene The isolation of manure, soil, and intestinal contents DNA extraction were verified by electrophoresis. Neither the transgene (*GH*) nor the marker gene (*neoR*) were detected in the feces, soil or intestinal content samples. Bacterial genes were amplified with the 16S rDNA universal primers to verify the effectiveness of the DNA extraction. The results showed that the extracted DNA from all samples contained the bacterial gene. ## PCR-DGGE and cluster analysis To study the influence of the *GH* and *neoR* genes on the bacterial community structure, DGGE was performed after PCR amplification of the variable V3 region of the 16S rDNA from the microbial DNAs. Thirty-three distinct bands were found. Cluster analysis was also performed on the basis of similarity (\> 95%). The band patterns for the feces, soil and intestinal content samples showed degrees of similarity that were higher than 96%, 96.5% and 95%, respectively. ## Phylogenetic analysis Each prominent DGGE band was recovered from the gels, cloned and sequenced. We selected sequences with over 97% similarity for subsequent phylogenetic analysis. Seven groups could be found based on the phylogenetic distribution of the 16S rDNA cloning libraries: *Firmicutes*, *Bacteroidetes*, *Proteobacteria*, *Acitinobacteria*, *Chloroflexi*, *Nitrospirae* and *Acidobacteria*. Unclassified sequences were designated as “unknown”. A more detailed classification is provided in. # Discussion Most GE research to date has focused on the animal health and welfare, while the environmental assessment of GE animals is only beginning to be investigated. The intestinal flora of mature livestock should be stable. However, diverse species of bacteria colonize the GI tract and they may develop natural competence, or the ability to absorb naked DNA. It is therefore necessary to detect bacterial changes in the feces of GE livestock in any complete environmental assessment. This is the first study of transgenic goats studied on their fecal matters and environment. We collected samples from the intestinal content, feces and soil. The GE goats were raised in a farm isolated to prevent communication with external wildlife and the escape of GE goats. We minimized the risk of horizontal gene transfer in the GE goats, although it still could occur in the gut and rumen. No *GH* and *neoR* were detected in the intestinal content, feces and soil samples, suggesting that no horizontal gene transfer occurred in the course of this study. The microbial community of the gut can be affected by many factors, including animal health, age and foraging patterns. We minimized the influences of these factors in this study by selecting goats with the same rearing condition, of the same developmental stage, and foraging in the same location. PCR-DGGE and 16S rDNA sequencing were used to determine the genetic diversity of their microbial communities and to identify several uncultured microorganisms. The V3 region was selected for species identification, which can be used to distinguish bacterial species to the genus level. Cluster analysis of intestinal contents and feces on the basis of 95% similarity showed that the microflora between the transgenic goats and normal goats were similar. Similar results were also detected in the soil samples, while three soil samples (S5, S6, and S7) covered the range of the farm, and S8 was collected from the outside of the farm. Furthermore, the phyla distribution among the three generations of GE goats (F0, F1 and F2) and normal goats was in the same dominant group, which could be because the microbial flora formed after weaning, then became stable. In conclusion, we did not find any evidence of horizontal gene transfer from the *GH* transgenic goats to the gut floral of other goats or soil microorganisms. We also demonstrated that the foreign *GH* gene and *neoR* gene did not change the microbial flora in the goat intestine or the surrounding soil. # Supporting Information This work was supported by grants from the National Animal Transgenic Breeding Grant Project (No. 2013ZX08008-004 and 2014ZX08008-004) and A Project Funded by the Priority Academic Program Development of Jiangsu Higher Education Institutions. GE Genetically engineered *GH* Growth hormone SCNT Somatic cell nuclear transfer *neoR* Neomycin resistance gene [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ZB XG QZ JL WH QHY. Performed the experiments: ZB XG QZ JL WH HQY JC. Analyzed the data: ZB XG QZ ZB XG. Contributed reagents/materials/analysis tools: ZB XG QZ HQY JC. Wrote the paper: ZB QHY. Substantially contributed to the concept and design of the work: QY QHY.

# Introduction The way orthographies represent sound differs markedly across languages. English, for example, is generally thought to have a comparatively complex orthography. One promising strategy to investigate how differences across orthographies may shape the functional architecture of the reading system is to develop full-blown computational models of reading for different languages using a common framework and the same core processing components. This strategy is pursued here in the context of the Connectionist Dual Process Model of Reading Aloud (CDP), a model that was originally developed for English. The latest versions of this model (e.g., CDP++) have been shown to provide the most comprehensive account of the empirical data, outperforming all of their competitors by an order of magnitude in terms of quantitative performance (i.e., goodness of fit). Unlike English, Italian is characterized by relatively simple (i.e., transparent) relationships between orthography and phonology. It therefore provides an interesting contrast with respect to the bulk of research on the far less transparent English orthography (see e.g., for a discussion). As is often the case, simplicity at one level comes at the price of complexity at another level. Finnish is a good example of this, where grapheme-phoneme relationships are extremely simple (fully consistent) but the morphological system is highly complex. In Italian, complexity can be found at the suprasegmental (word stress) level (e.g.,). That is, words with the same syllable structure and similar spellings can have stress in different locations (syllables) and where stress goes is not always predictable from the sublexical information. For example, in the database that is used below, 77.2% of 3-syllable words have stress on the penultimate syllable (e.g., *do'mani* \[tomorrow\]), 13.2% have stress on the antepenultimate syllable (e.g., '*macchina* \[machine\]), and 9.6% have stress on the ultimate syllable (e.g., *socie'tà* \[society/group\]). Apart from Italian stress being interesting in its own right, it is interesting to compare stress assignment in English and Italian as this comparison makes it possible to investigate whether the same mechanism can be used to assign stress in languages that differ in terms of complexity. Cross-language differences between English and Italian show that not only are the linguistic “rules” of how stress is assigned quite different (cf., e.g.,), but so are other factors. These include morphology (c.f.,) and orthographic markers that help predict stress. For example, with orthographic markers, Italian uses a diacritic (*à*) to mark stress in word-final position which English does not commonly use and both languages have orthographic sequences that are correlated with certain stress patterns. The existence of different cues that may have different weights in different languages is a challenge for a model like CDP++ because it uses the same architecture and learning mechanisms in different languages. Thus, the model needs to find the “right” cues solely on the basis of the statistical spelling-to-sound properties in the training corpus. Similarly, if there are patterns in the data that are seemingly best accounted for by a rule system, such as that which has been suggested for predicting the stress of Italian verbs, CDP++ and other connectionist models must learn to approximate these patterns without the use of a rule-based mechanism. There are now a fairly large number of published studies that have investigated different aspects of word and nonword reading in Italian, many of which are specifically concerned with how stress is computed. This means that it is possible to thoroughly test computational models of Italian reading aloud. A key prediction of the CDP approach is that the relationships between spelling and sound as well as spelling and word stress can be learnt via a simple linear learning mechanism. Given the complexity of word stress in Italian, it remains a challenging question whether such a simple mechanism can correctly predict stress assignment along with a number of other effects that have been reported. We therefore constructed an Italian version of CDP++ and assessed its descriptive adequacy both qualitatively and quantitatively. In terms of the scope of data to test the model on, we focused on skilled adult reading. We selected all studies that used a simple naming or priming paradigm with literate adults, for which the authors provided the list of items used. Only those studies that used more than 12 items in each cell were chosen. We also examined the largest study that investigated the effect of stress in acquired dyslexia in Italian. Developmental studies were not used to test the model because there are a number of issues to do with development that make simulating these data beyond the scope of the current work (for a discussion of these issues and a developmental CDP++ model, see –). ## The Model: CDP++.Italian The architecture and the processing assumptions of the model are identical to those of the latest version of the English CDP++, except that rather than only using words with a maximum of two syllables, words with three syllables were also used. In line with its dual-process framework, the model includes two main routes between spelling and sound, a lexical and a sublexical route. The lexical route is identical to all versions of CDP (apart from the earliest version), excluding the database used and its properties. The new database (see below) meant that 35 phonemes used and 32 letters (including a null letter) were used. The feature level of the model contained 14 features, although the parameters were set so that feature overlap had essentially no effect on the performance of the model in any way. In terms of the other parts of the lexical route of the model, the same frequency counts were used in both the orthographic and phonological lexicons since the database we used only had one set of frequency counts. In addition, all of the words used a frequency count that was the same as those given in the database plus 2. This was done because some words have a frequency value of zero and we take log values of frequencies for some computations. This means all frequency values always end up being greater than 0. The sublexical route consists of a graphemic parser and a two-layer associative (TLA) network. The graphemic parser is designed to segment letter strings into graphemes as well as categorize the graphemes into onset, vowel, and coda categories. This categorization process allows the graphemes to be placed into the graphosyllabic template of the TLA network (i.e., its input representation), and the TLA network is then able to generate phonology from them. The two different routes converge at the phoneme output buffer, where phonemic activation is integrated, as well as at the stress output buffer, where word stress activation is integrated. At present, learning only occurs in the sublexical route when the model is in training mode. In this mode, the graphemic parser is presented with a set of letter strings generated from each word. These strings are constructed based on the idea that an *attentional window* moves over letter strings from left to right, with the model learning which grapheme is at the start of each letter string within the attentional window (i.e., a set of letters is presented and the parser produces a grapheme that can be one or more letters long as an output). Apart from just learning which grapheme is at the start of a string, the graphemic parser also learns what type of grapheme it is (onset, vowel, or coda). In running mode, this allows the graphemes to be placed in a syllable- like template (i.e., the graphosyllabic template of the TLA network) based on their categorization, since if an onset grapheme follows a vowel or a coda grapheme, it means that it must be placed into the next syllable of the template after the vowel. Finally, the parsing network has a memory for previous graphemes it has parsed. This allows some amount of context sensitivity to be learnt even when the letters in the attentional window are the same, which is important for languages like English. The parser is displayed in. In Italian, because the correspondences between spelling and sound are less complex and have less contextual sensitivity compared to English, the attentional window is only 3 letters wide (in English, we used 5). This means that with the word *maglione* \[jumper\], for example, six strings of letters would be given to the network as input patterns: *mag, agl, gli, ion, ne\*, e\*\** (note that the \* represents no letter is in the window). The teaching signal (i.e., desired output) presented to the network during learning is the first grapheme occurring in each letter string, as well as its category (for the example above: m, onset; a, vowel; gl, onset; etc.). The three categories are represented in the output by simply duplicating the set of graphemes three times, and the grapheme is put in the set that represents the category it belongs to. Once the network has learnt relationships between the letter strings and the first grapheme in the strings, the parser can break strings of any length down into graphemes as well as place them into their correct position in the graphemic buffer. Thus, in running mode, the constituent graphemes for any string of letters (regardless of whether it is a known word, a novel word, or a nonword) are generated in an entirely bottom-up fashion. The orthography-to-phonology mapping is learnt by the TLA network, which is presented with graphemes (inputs) and, during learning, phonemes and stress information (outputs). In learning mode, all of the information is presented at the same time, and the model learns simple associations between inputs and outputs using the delta rule (this is equivalent to the Rescorla-Wagner learning rule). In running mode, the graphemes are placed in the graphemic buffer of the TLA network in a position determined by the graphemic parser, and the model generates phonology and stress information based on the simple associations it has learned during training. The information from the output of this network is then used in conjunction with activation produced by the lexical network to generate a pronunciation. ## Database The lexicon of the model was constructed from all words up to 3 syllables and 8 letters long that were in the Adsett et al. database (N = 63,438). This database consists of a large number of Italian words, and morphologically simple and morphologically complex word forms are represented separately. Letters in the database with a diacritic (accent) mark were coded as an entirely separate letter, which meant there were 31 separate letters plus one for the ‘blank’ letter. In terms of phonology, stress was coded based on syllable position (i.e., 1^st, 2^nd, or 3^rd), and there were 32 phonemes in the database, of which 23 were consonants, 7 were vowels, and 2 were semi-vowels. Frequency counts were obtained by entering the words into the Google search page on the 15/8/2008 and counting the hits for each word, with an Italian language restriction. Whilst it is known that Google counts may not be perfect, the log frequencies of the counts correlated reasonably well with the log frequencies in the CoLFIS database using all items that were shared, r = .77 (N = 21279). In addition, when frequency alone was used as a predictor on the Barca et al. database of written word naming latencies, an almost identical sized correlation (r = −.24) was found with both the CoLFIS frequencies and Google counts. There are a number of very low frequency words in the Adsett et al. database that are unlikely to be known by most of the Italian speaking population, as well as a number of loan words. These words were left in the database for the sake of simplicity and generality. There is also a reasonable amount of variation in different Italian dialects, and the examples used here are taken directly from the Adsett et al. database, and thus may differ as a function of regional dialects. ## Graphosyllabic Template The basic idea of the graphosyllabic template is to allow graphemes to be put into a syllable-like structure. In learning mode, where exemplars are presented to the model and the structure is learnt, this graphemic structure is derived by trying to align graphemes with lexical phonology, although other methods could certainly be used (see for a discussion). This means that identical letter sequences can potentially be coded differently if those sequences map to different lexical phonology. To code these sequences, a number of assumptions were made about graphemes and how they are placed in the graphosyllabic template. First, in terms of the set of multi-letter graphemes, these were selected based on trying to find the minimum set that could be used to describe the Italian orthography under the assumption that single graphemes generally map onto single phonemes. Based on this strategy, 5 consonant and 9 vowel graphemes that had more than one letter were used (gl gn gh ch sc ia ie io ió iu iá ié iù ii). These graphemes could potentially occur in any place of the graphosyllabic template, and the template was organized into a CCCCVC structure for each of the three possible syllables it could contain. This structure was chosen because there are maximally 4 onset graphemes in Italian (e.g., *<u>Austria</u>* \[Austria\], which uses the onset /strj/ in the second syllable). One coda consonant was used because, excluding a small number of loan words, only a single grapheme can occur in that position. A second assumption concerned the coding of geminates. They are represented in the phonology of the database as a single coda consonant followed by a single onset consonant. In the orthography they often correspond to a sequence of two identical letters (e.g., -ss in *ca<u>ss</u>e* \[boxes\]). Accordingly, these were coded as two single letter graphemes split between the two orthographic syllables. Conversely, when the geminates corresponded to non-identical letter sequences like –gl (e.g., *ma<u>gl</u>ione* \[jumper\]), these were coded with a single grapheme that was put in the first onset slot of the second syllable which the geminate spanned. Such a distinction is consistent with the conventional splitting of end-of-line words (when the line is out of space) in Italian printed text, which is also explicitly taught to children for handwriting. That is, the geminate letters are split (e.g., *cas-se*, with *se* going to the next line), whereas two consonant letters forming a grapheme are not split (e.g., *ma-glione* is a legitimate split but *mag-lione* is not). A third assumption that was made was that the semi-vowels /j/ and /w/ were coded by a single grapheme in an onset position. Thus, it was assumed that even if a letter is nominally a vowel, it does not necessarily have to be placed in a vowel position of the template. Rather, it was assumed that a vowel letter may occur in the onset position after a consonant when it is representing a semi- vowel phoneme. Thus, for example, *partiate* \[leave\], which has the phonology /par.tja.te/, was coded as p(o).a(v).r(c).t(o).i(o).a(v).t(o).e(v) and not p(o).a(v).r(c).t(o).ia(v).t(o).e(v) (o = onset; v = vowel; c = coda). Using these vowels in onset slots of the graphemic template allows only one-grapheme- one-phoneme correspondences to be used. An alternative to using vowel graphemes in onset positions would have been to use vowel graphemes with two letters (e.g., -ia) including ones that are not necessary with the current coding scheme (e.g., -uo). Apart from having to use many more graphemes, if this strategy was used then, in some cases, a single grapheme would have needed to map to both a vowel and the semi-vowel phoneme. Using vowel letters in onset positions therefore reduces the number of graphemes a great deal and also means that single graphemes map to single phonemes in these cases. For example, the –ia grapheme may either fall in a syllable without a /j/ in the onset associated with it or it may fall in one with a /j/. With the word *partiate* (/par.tja.te/), for example, there is a /j/ after the /t/. With the current coding scheme (p.a.r.t.i.a.t.e), because –i maps to /j/ in a one-to- one fashion, there is no inconsistency. Alternatively, when words without a /j/ are used, such as *angoscia* \[anguish\] (/angɔʃʃa/), the –ia is not split, and thus there is no ambiguity either. One advantage of the present coding scheme is that it naturally accounts for context sensitivity, which reflects the fact that the pronunciation of an onset consonant is affected by the vowel that follows it. For example, the letter –g is usually pronounced /g/ when followed by an *a*, *o*, or, *u*, or /ʤ/ when followed by *e* or *i*. This means that in a word like *seguo* \[follow\] (/segwo/), currently coded as s.e.g.u.o, the –u needs to combine with the –g to activate the correct phoneme. It also needs to activate the /w/ phoneme. The –o is then left to activate the vowel. Alternatively, if the vowel was instead coded as two letters (s.e.g.uo), the –uo would need to perform all three functions – help activate the correct context sensitive onset, activate /w/, and activate a vowel. ## Creating the Training Databases The graphemic structure of the training database was created in the same way as in Perry et al., where words were first divided into contiguous consonant and vowel sequences. These were then parsed using the longest possible graphemes. To identify cases where vowel letters functioned as semi-vowels, all words where the initial parsing of the graphemes caused there to be less vowel graphemes than vowel phonemes were identified. When this occurred, the onsets of syllables were scanned for /j/ and /w/. If the vowel that came after them started with either an –i or a –u, this vowel was placed in an onset position (N = 7076). Other more minor changes included: 1. the –sc onset was split into –s and –c when it corresponded to /sk/ (e.g., <u>sc</u>appa /skappa/; N = 855). 2. –gl was split as –g and –l when it corresponded to /gl/ (e.g., <u>gl</u>oria /glɔrja/ \[glory\]; N = 836). 3. If there were less orthographic vowels than phonological ones, the string was scanned for all of the multi-letter vowels and the vowel grapheme was split if one was found (N = 813). For example *avv<u>ia</u>* /avvia/ \[start\] has the consonant-vowel sequence \[a\]\[v\]\[v\]\[ia\], and three phonological vowels. To get the number of orthographic and phonological vowels the same, the –ia was split. Thus, the final graphemes were a.v.v.i.a. 4. The onset –sch was split as –s and –ch (e.g., *<u>sch</u>erzo /sketso/* \[joke\]; N = 329). 5. The vowel sequence –iuo (e.g., *g<u>iuo</u>co* /ʤwokɔ/ \[play\], N = 6), was split as –iu and –o. This left 191 words which could not be coded, almost all of which were loan words (e.g., *delphi*). Therefore, the final training database contained 63360 words. From these words, the training database for the graphemic parser was constructed by taking each word and creating the set of 3-letter sequences that represented the letters in the attentional window with a grapheme that needed to be parsed at the start (the input patterns). These were paired with the grapheme that occurred at the start of each sequence (the output patterns). See above for an example of this. This meant that there were 417,622 training exemplars. The training database for the TLA network was constructed by simply aligning the graphemes in the graphosyllabic template of the network (the input patterns) and pairing this with the phonology of the words (the output patterns). ## Training the Graphemic Parser The graphemic parser was created in the same way as Perry et al., where a simple two-layer network with a 3 grapheme memory was trained to select graphemes from the start of strings of letters and also categorize them into onset, vowel, and coda categories. Training was also done in an identical fashion to Perry et al., where different networks were trained on the whole database as well as a number of smaller subsets of words (500, 1000, and 2000 words). The input patterns were the three letter sequences that could be derived so that a grapheme was at the start of each sequence, and the actual graphemes used were those derived from lexical phonology as described above. The output patterns were simply the grapheme and its classification (i.e., onset, vowel, coda). See above for an example of the patterns created for the word *maglione*. The input layer of the graphemic parser consisted of three main sections that contained 32 letter nodes each (i.e., 31 letters plus one “null” letter). These were designed to represent any possible sequence of letters that could occur in an attentional window that is 3 letters wide. The output layer consisted of all possible graphemes. These were repeated 3 times so any grapheme could potentially be classified into an onset, vowel, or coda category. ## Graphemic Parser Results There were far fewer errors in Italian than in English – indeed there were only 1001 (.24%) errors for the patterns that were used in training. The errors were not random, which makes it possible to look at the individual types of errors, and these appear in the Materials S1 in. Whilst the results suggest that the model is not perfect, this is at least in part because there are inconsistencies in the way graphemes are split in Italian, and the errors can help identify predictions that the model makes. For example, *esempii* \[examples\] and *capii* \[I understood\] both use an -*ii* letter sequence, but with *esempii*, the -*ii* functions as a semi-vowel and vowel, whereas with *capii* there are two vowels. This type of inconsistency causes the model to make errors with the -ii letter. This means that CDP++ predicts that people will also give a distribution of responses when confronted with the –ii pattern, since it is something which is ambiguous in Italian, and some of the responses will therefore have a different number of syllables to the other ones. Another example of this is the –ia vowel sequence, which can also be parsed into different categories (e.g., *prev<u>ia</u>* /prevja/ \[subject to\] (semi-vowel/vowel) vs. *rinv<u>ia</u>* /rinvia/ \[reject\] (vowel/vowel)). Apart from the results of the fully trained model, the models trained on a small number of exemplars also showed reasonable performance. Even when the model was trained on only 500 words, it was able to get most correspondences correct. This suggests that choosing the correct graphemes in words in Italian is fairly simple, and can be done with relatively minimal information about the entire database, which might explain in part why reading acquisition is a lot quicker in Italian than in English. ## Training the Sublexical (TLA) Network The TLA network was trained for 20 cycles using the same parameters as in Perry et al.. Phonemes and graphemes were aligned in the same way as in CDP++ (i.e., into syllables). ## CDP++.Italian: Results The items used for all of the studies simulated below were identical to those in the original studies (The exact results were not reported in a small number of the experiments. When this was the case, we estimated the results from the figures). When words were used that did not exist in the database, they were excluded from the analysis, as were nonwords that were in the database (i.e., nonwords that were actually words). A 3 standard deviation (SD) cut-off was also applied to all of the results, and these items were considered outliers, as were all words that took more than 250 cycles to produce. The number of items removed from the statistics and the reason is reported after each experiment in square brackets. All data sets were run using the same parameter set (see the Materials S2) unless otherwise stated. Due to computational constraints, we restricted the lexicon of the model. This was done by only using words in the lexicon that were identical to the one being run, except for the pseudohomophone, neighborhood, and Job et al. simulations, where we used the full lexicon. This was necessary since, with more than 60,000 items in the lexicon, it is very hard to find an optimized parameter set within a reasonable amount of time. In addition, our previous work has shown that whilst examining some properties of feedback in the model is useful, feedback has little impact on data sets that do not need it. ## Database Comparisons The first set of results we examined were those from Barca et al., a database with the reaction times for 625 nouns, 501 of which were in the model's lexicon (most of the others were 4-syllable items). We used a two-step regression analysis to predict the human reaction times. In the first step, we added the onset characteristics of the first phoneme of the words. These were taken from the database of Barca et al. In the second step, we added the naming latencies of the model (in number of cycles). The performance of the model was compared with a number of different regression analyses that used the onset characteristics of the words, log word frequency from the same database which the model used, orthographic neighborhood calculated using Levenshtein Distance, number of letters, number of syllables, and word stress. The results showed that the model plus onsets captured slightly less variance compared to the regression equation with all of the terms in it (52.3% versus 53.6%), somewhat more than onsets plus frequency (50.6%), and more than just onsets alone (46.4%). Unfortunately, as can be seen via the difference between the full regression equation and just the onsets, the amount of variance that could be captured above just simple onset characteristics was relatively small (for a similar finding in French, see) ## Words, Frequency, and Nonwords Perhaps the simplest question of all that could be asked about the model is whether it reads aloud real words more quickly than nonwords, and whether high frequency words are read aloud faster than low frequency words. Given that the Italian orthography is very regular, it is conceivably possible that at least the segmental phonology of words could be generated without lexical input, which, ignoring word stress, would predict that nonwords and words would be read aloud at a similar speed. This is not what is found, however, and it is been shown that not only are nonwords read aloud more slowly than words, but low frequency words are read aloud more slowly than high frequency ones. This suggests that lexical input is important in Italian reading. Using the same stimuli as Pagliuca et al. where both high and low frequency words were examined as well as nonwords, we examined whether the model would also show this pattern. The results showed that, just like the data, words were read faster than nonwords, *t*(85) = 14.70, *p*\<.001, and high frequency words were read faster than low frequency words, *t*(44) = 3.83, *p*\<.001 (High Frequency Words: 70.7; Low Frequency Words: 80.4; Nonwords derived from high frequency words: 120.0; Nonwords derived from low frequency words: 125.5) \[2 words not in the database, 7 nonwords in the database\]. ## Stress Regularity/Consistency Perhaps the results that are the most important in Italian reading are those to do with how reaction times are affected by stress regularity and stress consistency – that is, whether people give slower responses to words with atypical stress due to them not having a possible default stress (regularity – typically assumed to be penultimate in Italian) or due to them having a different stress pattern compared to words with similar spellings (consistency, typically measured as a friends vs. enemies ratio where friends share the same orthographic sequence and phonology but enemies only share the same orthographic sequence). The results of the model on all of the experiments reported below to do with stress regularity and consistency appear in. Colombo ran one of the seminal studies on stress effects in Italian. In her first experiment, she examined stress regularity in both high and low frequency words and found that words with irregular stress were slower to read aloud than words with regular (i.e., penultimate) stress, but that this was restricted to low frequency words. CDP++ showed the same pattern, with a main effect of Frequency, *F*(1, 101) = 75.41, p\<.001, Stress Regularity, *F*(1, 101) = 6.46, *p*\<.05, and an interaction between the two, *F*(1, 102) = 9.31, *p*\<.005. Two t-tests showed that the difference between the high frequency words was not significant, *t*\<1, but the difference between the low frequency words was, *t*(47) = 2.88, *p*\<.01. (High Frequency Regular: 80.6; High Frequency Irregular: 79.9; Low Frequency Regular: 88.7; Low Frequency Irregular: 96.7). \[9 words were not in the database, 1 outlier\]. Apart from just stress regularity, Colombo also examined whether other properties of stimuli interacted with the stress regularity (her Experiment 4). She found that stress consistency, which she defined as the extent to which the last 3 letters of a word shared the same stress pattern with other words with the same 3 letters (i.e., stress neighbors), was important. With words that were stress inconsistent (i.e., had more stress enemies than friends), RTs were slower than when they were consistent, but only when the words were also stress irregular. When the words were stress regular, no effect of stress consistency was found. CDP++ showed a relatively similar pattern, with main effects of Stress Consistency, *F*(1, 55) = 9.34, *p*\<.005, no effect of Stress Regularity, *F*(1, 55) = 1.32, *p* = .26, which was unlike the data of Colombo where a significant effect was found, and, importantly, an interaction between the two, *F*(1, 55) = 5.17, *p*\<.05, which appeared to be caused by the inconsistent words with irregular stress being especially slow (Stress Consistent/Stress Regular: 94.25; Stress Consistent/Stress Irregular: 92.6; Stress Inconsistent/Stress Regular: 97.0; Stress Inconsistent/Stress Irregular: 105.6) \[5 words were not in the database\]. Burani and Arduino (their Experiment 1) also examined the effect of stress consistency and regularity. They ran an experiment that was similar to that of Colombo where stress consistency was examined, but suggested that their stimuli were better matched than those of Colombo for a number of different reasons (see Burani and Arduino for a list of these). Interestingly, they found effects of stress consistency for both stress regular and stress irregular words. CDP++, showed a main effect of Stress Consistency, *F*(1, 42) = 8.94, p\<.01, but not Stress Regularity, *F*(1, 42) = 1.73, *p* = .20, nor an interaction between the two, *F*\<1 (Stress Consistent/Stress Regular: 92.6; Stress Consistent/Stress Irregular: 97.1; Stress Inconsistent/Stress Regular: 101.8; Stress Inconsistent/Stress Irregular: 104.8) \[15 words were not in the database, many of which were 4 syllables long\]. This pattern is therefore very similar to the one reported by Burani and Arduino. The successful simulation of both sets of results suggests that the seemingly inconsistent results between these two studies can potentially be explained once the properties of items are taken into account. Apart from stress consistency, Burani and Arduino also looked at the overall number of words that share a particular letter sequence, which they called numerosity. They found that stress irregular words with a high numerosity were named faster than stress regular words with a low numerosity. CDP++ predicted a null effect with this dataset, *t*\<1 (Regular words: 97.2; Irregular words: 97.6) \[8 words were not in the database\]. One final study looking at stress neighborhood was run by Sulpizio, Arduino, Paizi, and Burani. They examined the effect of stress neighborhood with nonwords, defined in a similar way to Colombo. With their three-syllable stimuli, they found that having a consistent stress neighborhood had a weak effect on nonwords when the neighborhood favored penultimate stress, but had a strong effect when it favoured antepenultimate stress. They suggested that rather than these results reflecting just stress consistency, as defined by the proportion of stress friends versus enemies, they were likely to occur because of a difference in the numerosity of friends versus enemies. That is, with their nonwords, those in an antepenultimate neighborhood shared a greater number of stress friends and stress enemies than those in a penultimate neighbourhood, even though the consistency ratio was similar for both types of words (for example, a word with 12 friends and 6 enemies has a higher numerosity than a word with 6 friends and 3 enemies, even though both have the same consistency). CDP++ showed a similar pattern to the data, although tended to over-predict the effect of having a consistent stress neighborhood (Experiment 1 (Penultimate vs. Antepenultimate responses), Human: Penultimate Neighborhood: 53% vs. 47%; CDP++: 73% vs. 27%; Antepenultimate Neighborhood: 25% vs. 75%; CDP++: 16% vs. 84%; Experiment 2, Human: Penultimate Neighborhood: 58% vs. 42%; CDP++: 66% vs. 34%; Antepenultimate Neighborhood: 25% vs. 75%; CDP++: 14% vs. 86%) \[Experiment 1: 2 nonwords were 4-syllables long\]. Apart from studies specifically looking at the comparison between different types of stress dominance, there are also two sets of nonwords run by Colombo and Zevin, and hence the proportion of nonwords given penultimate stress can be examined (note that Experiment 1 and 2 in their study used the same nonwords), as well as a set of nonwords used in Colombo (Experiment 5). With the nonword set used by Colombo, 69.8% of nonwords that participants gave reasonable responses to (i.e., were not errors) were given penultimate stress. CDP++ gave a result close to this, giving 61.9% of the stimuli penultimate stress \[1 4-syllable nonword\]. With the first and second experiment of Colombo and Zevin, the nonwords were deliberately chosen to be biased to give penultimate stress, and this pattern was found with CDP++ and in the real data (CDP++: 78%; Experiment 1: 76%; Experiment 2: 83%). Alternatively, in the fourth Experiment of Colombo and Zevin, a balanced set of items in terms of type of likely stress was chosen, although people produced somewhat more antepenultimate (63.2%) than penultimate responses. CDP++ produced the opposite result, favouring penultimate responses (66.3%) \[Experiment 1: 2 outliers; Experiment 4: 1 nonword in the database\]. ## Orthography-Phonology Consistency A second type of consistency that can be found in Italian relates to the orthography-phonology mapping. Job, Peressotti, and Cusinato examined this by constructing nonwords that used consonant graphemes that could only be correctly read if the following vowel was taken into account (see the final paragraph in *Graphosyllabic Template* above for the set of vowels that affect some consonants and how these are coded in CDP++). They did this by choosing words with one of these consonants in it, and then constructing two types of nonwords by changing a single vowel in them. In one case, the nonwords kept the same consonant pronunciation as the words they were derived from (the consistent nonwords; e.g., *mercoto* /merkotɔ/, which was derived from mer<u>c</u>ato /mer<u>k</u>atɔ/) whereas in the other case they did not (the inconsistent nonwords; e.g., *merceto* /mertʃetɔ/). The results of Job et al. showed that when the nonwords were mixed with words, there was an effect of consistency, with the consistent nonwords being read aloud faster than the inconsistent ones, but this did not occur when the nonwords were not mixed with words. CDP++ showed a significant result with this set of items, *t*(46) = 2.12, *p*\<.05 (148.1 vs. 164.5 cycles) \[1 outlier, 4 Errors, 14 4-syllable nonwords, 1 nonword in lexicon\]. To simulate the change of strategy as a function of list composition (i.e., no consistency effect with nonwords only), we reduced the threshold at which phonemes and stress nodes needed to be activated before naming can be finished to.5. This was done because, as noted in Perry et al., it is a reasonable way of simulating lists where only nonwords are used. This is because nonwords tend to generate less activation than words and hence people may reduce their response criterion accordingly when reading blocks of them. With this change, the model did not produce a consistency effect, *t*(49) = 1.42, p = .16 (120.1 vs. 126.1 cycles \[2 Errors, 14 4-syllable nonwords, 1 nonword in lexicon\]). Job et al. also ran an additional experiment that was the same as the other where the stimuli were run in a nonword only block, except a different set of nonwords were used. The results they found showed no significant difference between the consistent and inconsistent nonwords. CDP++ did not display a significant difference between those two groups either, even using the normal parameters, *t*\<1 (150.9 vs. 146.9 cycles) \[1 outlier\]. A more recent set of experiments that assessed whether words including complex (contextual) print-to-sound rules are named more slowly than words with no contextual rules was run by Burani, Barca, and Ellis. In contrast to Job et al., they examined simple versus complex spelling-sound patterns in words rather than nonwords. In their first experiment, they found that people were slower reading aloud the words with consonants that required a context to predict their phonology correctly. CDP++ predicted the same result, *t*(39) = 2.20, *p*\<.05 (89.4 vs. 94.4 cycles) \[15 words not in lexicon\]. In their second experiment, they added the additional factor of word frequency. Their results were somewhat mixed, presumably because with the complex rule words there was a lower density of complex letter clusters relative to Experiment 1 (See Procedure, Experiment 2), with no effect with high frequency words and the results with low frequency words being much weaker than the previous experiment. This caused the main ANOVA to fail to reach significance by items. The absolute size of the effect with the low frequency words was also smaller than the first experiment (11 ms vs. 24 ms). CDP++ predicted that there should not be a significant effect of consistency with either the high or low frequency items (both *t*'s\<1; High Frequency Consistent: 77.4; High Frequency Inconsistent: 78.0 \[2 words not in the database\]; Low Frequency Consistent: 86.8; Low Frequency Inconsistent: 87.2 \[3 words not in the database\]). ## Other Effects Apart from consistency, there are a number of different effects that have been reported in Italian. These include the effect of morphology, pseudohomophony, and orthographic neighborhood. Morphological effects are interesting because CDP++ has no explicit morphological processing layer. Thus, if CDP++ were to capture effects that are presumably due to morphological processing, it would suggest that some of these effects can be explained by factors correlated with morphemic status, such as frequency, rather than some sort of explicit morphological status or the semantics associated with particular morphemes. Burani, Marcolini, De Luca, and Zoccolotti, examined this in reading, and they found that a number of different groups, including normal adult readers, read nonwords that were composed of a root and suffix more quickly than morphologically simple nonwords. CDP++ displayed the same result, *t*(29) = 2.40, *p*\<.05 (144.5 vs. 158.5 cycles) \[1 nonword in lexicon\]. Alternatively, when a similar manipulation was examined with words, skilled readers did not show any differences. CDP++ displayed this result also, *t*\<1 (93.1 vs. 92.6 cycles) \[44 4-syllable words not in the database\], although there were a large number of items in the stimuli that it was not able to use. A very similar set of nonwords as Burani et al. was examined by Burani, Marcolini, and Stella. Similar results were also found, except that there was an exceptionally large error rate (21.7%) on the morphologically simple nonwords, which was presumably due to the different composition of the lists the nonwords were part of in the two studies. CDP++ not surprisingly gave very similar results with these nonwords compared to the previous ones, *t*(28) = 2.03, *p* = .052, and also made no errors (144.2 cycles vs. 156.7 cycles; note that one nonword that was actually an exceptionally low frequency word was treated as a nonword rather than removed due to the small number of items and thus the importance of each item in the final significance value). Whilst the difference in the error rate between the model and the real data is interesting, creating errors with CDP++ to try to simulate this aspect of this particular data set is beyond the scope of the current work. Pseudohomophone effects, where people read aloud nonwords with phonology that sounds like a word faster than nonwords where it does not, are interesting because they are generally believed to show that there is feedback from sublexical phonology in reading (see for a review). Peressotti and Colombo examined this in Italian using pseudohomophones that were orthographically often very strange (e.g., *cjfra*) and similarly matched nonwords, with the idea being that using nonwords with strange sequences of letters meant that none of the effects they found could be due to orthographic similarities between pseudohomophones and the words that they sounded like. They also compared the results to more orthographically normal non-pseudohomophonic nonwords. The results they found showed that the pseudohomophones were read aloud faster than their controls, but the difference between pseudohomophones and the nonwords with more normal spellings was not significant. Despite the strange orthographic patterns used by Peressotti and Colombo , we presented their stimuli to CDP++. Not surprisingly, the model had a high error rate, since it simply could not produce a reasonable answer for some of the nonwords, such as when a –j was used as a vowel. Initial inspection of the results showed that these errors were not distributed evenly across the groups. In the pseudohomophone and nonword control group, the model made 36 and 44 errors, compared to 13 with the orthographically normal nonwords. Because the groups were very large, however (119 items in each cell), we could still examine the RTs from the correct responses. However, rather than using between-group comparisons, as we generally do, we only used within-group comparisons. This is reasonable because stimuli triplets were matched across the groups (e.g., *ansja*, *antja*, and *antia*). The results showed that, like the human data, CDP++ predicted that the pseudohomophones would be read aloud faster than the control nonwords, *t*(57) = 2.52, *p*\<.05 (133.0 vs. 143.0 cycles) \[80 Errors, 2 outliers\]. Unlike the data, however, the pseudohomophones were slower than the orthographically more normal nonwords, *t*(58) = 2.29, *p*\<.05 (135.5 vs. 126.2 cycles) \[49 Errors, 2 outliers, 9 nonwords in database\]. Apart from just the error rates, these results should be taken with great caution because, unlike Perry et al., we allowed the network to run even if a correspondence was very poorly learnt. That is, it triggered many “dead nodes”, which we ignored (see for evidence suggesting that nonwords with very strange spellings are not processed by the normal reading system and hence reasonable to ignore). This allowed slower responses in the pseudohomophone and matched nonword control condition that would typically be excluded, and hence is likely to be responsible at least in part for the difference in reaction times between the pseudohomophones and orthographically normal nonwords. Another way that phonological feedback can be examined was done by Mulatti et al.. They examined nonwords by changing a single letter either in the first position or a latter position of a word (e.g., a word like *serpe* \[serpent\] was changed to form *berpe* and *babbo* \[dad\] was changed to form *babro*). They found that nonwords created from a first letter change were slower to read aloud than nonwords created from a latter letter change. They suggested that this was caused by serial processing of sublexical phonology and the way that activation produced at different times interacted with lexical activation. CDP++ was not able to replicate this effect, *t*\<1 (124.4 vs. 128.6 cycles). \[2 errors, 2 nonwords in lexicon of model\] Finally, neighborhood effects, where the effect of words with similar spellings to the one being read aloud is examined, are interesting because they may provide some insight into both learning and the way the lexical system functions. In Italian, Arduino and Burani reported that nonwords with many orthographic neighbors (i.e., words that differ by a single letter) were read aloud faster than nonwords with few orthographic neighbors, and that whether the nonwords had a high frequency neighbor or only low frequency ones did not appear to affect the results. When their stimuli were presented to CDP++, there was a significant main effect of whether a nonword had high frequency neighbors, *F*(1, 51) = 4.39, *p*\<.05, with the nonwords with high frequency neighbors being named more slowly than those with only low frequency neighbors, as well as a significant interaction, *F*(1, 51) = 5.81 (High Neighborhood/High Frequency Neighbors: 132.8; Low Neighborhood/High Frequency Neighbors: 123.6; High Neighborhood/Low Frequency Neighbors: 116.5; Low Neighborhood/Low Frequency Neighbors: 113.8). \[4 errors, 1 nonword in the lexicon of the model\]. Two t-tests examining the nonwords with high and low frequency neighbors separately were not significant (high: *t*(25) = 1.60, *p* = .12; low: *t*(26) = 1.86, *p* = .075). It is unclear to us exactly why the model shows the incorrect pattern with the nonwords. However, CDP++ has not been able to simulate neighborhood effects all that well in previous simulation work, and thus this may represent a problem with the model. As noted by Pagliuca and Monaghan, there might also be a lack of power because there were only 15 items in each cell. ## Acquired Dyslexia Different types of acquired dyslexia have been reported in Italian, including both surface and phonological dyslexia. One particular area of interest has been word stress. In the largest study examining this, Colombo et al. examined 22 patients with Alzheimer's Disease (AD), and classified them into 3 groups based on how advanced their cognitive decline was as measured by the Mini Mental State Examination (MMSE). They then examined their reading performance on high and low frequency words and nonwords. The words they used were divided into what they called dominant and subordinate stress, with the former defined as those with penultimate stress and the latter defined as those with ante-penultimate stress. The groups were also balanced on stress neighborhood with the dominant words having high stress consistency and the subordinate words low stress consistency. The results Colombo et al. found showed that the performance of the groups was related to their performance on the MMSE, with the group that scored the lowest also performing the worst. With that group, stress errors accounted for around 35% of all of the errors, and the percentage of correct responses was affected by both frequency and stress type (Dominant stress, High Frequency: 92.2%; Low Frequency, 83.0%; Subordinate stress: High Frequency: 79.2%; Low Frequency: 45.3%). A more complex pattern was found with nonwords, where the most severe group had an average error rate of 25%, and the more severe the AD, the more likely they were to give subordinate stress on the nonwords. At present, we only simulated the effect of word stress, since simulating the nonword results would have required more than simple parameter changes, which is beyond the scope of this paper. This simplification is reasonable because, whilst phonological deficits and nonword processing problems are often found to co-occur in AD, phonological dyslexics have been reported with no obvious phonological processing problems. Caccappolo et al., suggest that this means that sublexical and lexical mechanisms may not be directly linked. In addition, in one of the classic cases on acquired surface dyslexia in Italian, a patient was described with “virtually normal” (p. 283) performance when reading nonwords, but made many stress errors reading some types of words. To simulate the word results of Colombo et al., we simply used the same strategy for simulating surface dyslexia we have used elsewhere. We did this by increasing the frequency scaling of the lexicons to.75, with the idea that this simulates additional difficulties in lexical access. We also reduced the amount of activation going into the phoneme output buffer by changing the excitation and inhibition parameters from the phonological lexicon to the phoneme output buffer to.03 and −0.03. For the sake of simplicity, we did not reduce the level of activation going to the stress output nodes under the assumption that the generation of phonemes in AD is more difficult than the generation of stress information. Obviously, in the future, it would be possible to examine the effect of reducing activation to both phoneme and stress output nodes should the data dictate it. With the parameter changes noted, CDP++ produced results very similar to those of the severe group on overall error rates (Correct %, Dominant stress, High Frequency: 91.2%; Low Frequency: 77.1%; Subordinate stress, High Frequency: 82.9%; Low Frequency: 54.3%). The distribution of errors was also very similar, with 44% of the errors due to stress and 56% coming from other sources. These results suggest that the effects of frequency and stress dominance are inherent properties that the model is sensitive to, and that when it is parameterized such that it does not perform at near perfect accuracy, the most likely items it makes errors on are also the most likely ones that people do after cognitive decline due to AD. ## Priming All of the previous simulations relied on getting the model to produce output in a simple naming task, with each item run entirely independently of the others. However, there is also some data on stress priming in Italian, where the effect of being primed with a word that has the same or different stress to the one being named has been examined. At present we will not try to simulate all aspects of priming, as there are a number of non-trivial issues that would need to be considered to do this. These include how decay in representations should be set (i.e., the amount activation in representations reduces from one word to the next), how primes should be treated when a second word appears on top of them, and how to implement aspects of the prosodic processes not currently implemented, such as how stress is stored in the linguistic system over time. Despite the problems of modelling priming, the type of results the model would predict ignoring more intricate matters can be examined. In terms of simple priming where a prime precedes a target word, Sulpizio, Job, and Burani found that when a prime was presented for 83ms before a target with the same stress pattern, the target word was named faster than if the preceding word had a different stress pattern. They found that this occurred irrespectively of whether the target word had penultimate or antepenultimate stress. Our explanation is the same as offered by Sulpizio et al., which is that this may be explicable via the pre-activation of stress information (stress nodes in the model), which would then either reach threshold faster if the stress information is congruent or more slowly if it is incongruent. To examine this, we ran the model using the words of Sulpizio et al. with a reduced stress criterion (.58 instead of.68), which simulates the ability of the model to reach the stress threshold faster. The results showed that the size of the reaction time differences between the model with the normal and low stress criterion were relatively similar across the words with penultimate and antepenultimate stress (Penultimate Stress, Normal/Low, 99.9, 94.6; Antepenultimate Stress, Normal/Low: 106.3, 98.6; Priming effect: Penultimate: 5.3; Antepenultimate: 7.7) \[9 words not in lexicon\]. Note that due to the repeated nature of the comparison and the fact that the model is entirely deterministic, statistics are not reported since even the smallest of priming effects are almost always significant. Apart from standard priming, Colombo and Zevin examined the effect of priming across a number of trials. In particular, they used a paradigm where a set of words or nonwords with the same stress would occur before a target word and the effect of the prime words examined. The results they found suggested that the main change caused by the prime words and nonwords was a change in the dominance between lexical and sublexical processing, with people making more errors on words that did not have a dominant stress pattern when the primes caused more sublexical processing. To simulate this, we ran the words of Colombo and Zevin's first experiment, where they examined the effect of nonwords typically given penultimate stress on words with antepenultimate stress. With the normal parameter set, the model makes no stress errors. To simulate a change of dominance between lexical and sublexical routes, we increased the excitation strength of the TLA to phoneme output and stress output buffer to.12, and reduced the strength of the parameters from the phonological lexicon to the stress output buffer by.02. The results showed that this increased the error rate on the words to 14.3%, which is very similar to the experiment of Colombo and Zevin. Obviously, there are many ways we could have changed the balance between the two routes, but the results here show that a change of balance is likely to cause stress errors in the way Colombo and Zevin predicted. ## The Role of Sublexical and Lexical Process An important facet of the results that we have not explored is to what extent the lexical and sublexical parts of the model are responsible for the results. One way to examine this is to look at the performance of the model without sublexical or lexical input. This isolates the extent to which the results are simply caused by one route or the other. We first did this on the large database we first examined, removing all sublexical input. This caused the amount of explained variance to drop from 52.3% to 49.4% (note that just onsets alone account for 46.4% of the variance and the correlations without onset coding are *r* = .39 for the model with sublexical phonology and *r* = .24 for the model without sublexical phonology). We then examined all of the stress consistency studies without sublexical phonology. None of them even produced a trend towards a significant stress consistency effect, and nor was there an orthography- phonology consistency effect with Burani et al.'s words. Next, we removed lexical activation from the pseudohomophone simulations of Peressotti and Colombo, and there was no longer a pseudohomophone effect (*t*\<1). These results basically suggest two things. First, that phonology plays an important role in the quantitative performance of the model, as it does in other versions of CDP. Second, that stress consistency effects are caused by the interaction of lexical and sublexical processing and that pseudohomophone effects are caused by feedback from sublexical to lexical phonology and back again. ## Inconsistent Findings Whilst the model produced reasonable results across a broad spectrum of experiments, there were a number of results that the model produced that were qualitatively different to the human ones that were not discussed. These include (a) no significant difference between nonwords created by changing one letter at the start of a word compared to the end of a word as reported in Mulatti et al. ; (b) no stress regularity effect in Experiment 4 of Colombo ; (c) no significant difference between the high numerosity irregular and low numerosity regular words in Experiment 2 of Burani and Arduino ; and (d) no significant difference with the low frequency words with complex versus simple rules in the Experiment 2 of Burani et al.. Whilst we have no definitive explanation for why the model did not capture these, in all cases, the absolute size of the effect reported in the studies was small. In Mulatti et al. it was 15 ms, in Colombo (Experiment 4) it was 13 ms, in Burani and Arduino (Experiment 2) it was 18ms, and in Burani et al. (Experiment 2) it was 11 ms. Alternatively, the size of the effects in all of the experiments where the model did find a significant result, excluding the frequency effect reported in Pagliuca et al., was larger (Colombo (Experiment 1): 43 ms; Colombo (Experiment 4): 24 ms; Burani and Arduino (Experiment 1): 24 ms; Job et al. (Experiment 1): 21 ms; Burani et al. (Experiment 1): 24 ms; Burani et al. : 48 ms; Peressotti & Colombo : 35 ms). Given this, it suggests that it would be worthwhile investigating ways to make the model more sensitive to smaller effects in the future. Actually finding ways to increase the sensitivity of the model may be particularly challenging, especially for the nonwords of Mulatti et al.. This is because, even though one group of their nonwords differed from their basewords only on the first letter, they often shared start sequences with many other words, and hence their uniqueness to any other words based on serial position may not be as much as the examples in the title of their article might suggest (*zeading* vs. *reazing*). For example, the first nonword reported in their stimuli set, *berpe*, differs in the first letter compared to the baseword *serpe* from which it was created. However, it only differs in the 4^th letter with *berci* \[yell\] (there are in fact 102 other words that start with *ber*). This can be compared with the control *babro*, which differs in the 4^th letter compared to its baseword (*babbo*). This means that any early effects of phonological feedback generated serially would activate *berci* and *babbo* to a similar amount, and thus a positive feedback loop from these words being activated should help both nonwords similarly. The main difference then is that *babbo* is a closer neighbour to *babro* than *berci* is to *berpe* (one vs. two letters different). This means that, after the 5^th letter is parsed and activation generated, *babro* is likely to be activated more than *berci* since two phonemes would differ from the nonword compared to one. Such fine differences may be very hard for computational models such as CDP++ to capture via a lexical feedback loop. # Discussion and Conclusions The present simulations show that CDP++ did a reasonable job predicting many of the different data patterns that have been reported in the literature. The two most important effects have to do with stress and orthography-phonology regularity/consistency. Stress regularity/consistency is important for the development of a comprehensive model of reading aloud but these effects have received little attention in languages other than Italian (but see), probably because most modelling studies have focused on monosyllables (but see e.g.). Orthography-phonology regularity/consistency is important because, historically, it has been the crucial benchmark effect that challenged rule-based models in favour of connectionist models. For both of these theoretically important effects, whilst not perfect, CDP++ has captured the data remarkably well. The ability of CDP++ to simulate various aspects of stress in Italian suggests that the mapping of orthography onto stress nodes, as implemented in the Italian and English CDP++ model, is a powerful and general mechanism that does not seem to be specific or restricted to a given language. The ability of the model to simulate consistency effects in Italian provides yet another demonstration that the CDP family of models is highly sensitive to consistency, as has been shown by the English model on a number of large and exceptionally well-controlled data sets. Together, this suggests that graded consistency effects are likely to be an inherent property of the type of network and learning algorithm used, and not something that is specific to a particular orthography. It is worthwhile comparing the results of CDP++ to those of the PDP model of Pagliuca and Monaghan. Our model differs from theirs in a number of important ways. In particular, we used a lexical route under the assumption that the sublexical route cannot learn all relationships between orthography and phonology. Thus, at least when reading words, our model can perform essentially flawlessly. Alternatively, the model of Pagliuca and Monaghan was only able to read 93.7% of words correctly. It seems likely that if the network of Pagliuca and Monaghan was trained for longer or with a more powerful algorithm, better accuracy could probably be obtained. However, whether the model would still capture nonword stress consistency effects with additional training would need to be explored. When comparing the two models on the results of the experiments described above, it also becomes clear that CDP++ provides a better fit of the available empirical data than the PDP model of Pagliuca and Monaghan. CDP++ was able to correctly simulate all of the results that were correctly simulated by Pagliuca and Monaghan as well as many others that Pagliuca and Monaghan did not examine. There were also effects that were correctly simulated by CDP++ but not Pagliuca and Monaghan's model (e.g., the effect of stress consistency effect using Burani and Arduino's items). Given there are some similarities between the models, one might try to isolate why CDP++ performs better than Pagliuca and Monaghan's model. One possibility is that CDP++ uses graphemes and not letters in the input layer. However, given the simplicity of the Italian orthography, and given that Pagliuca and Monaghan organized the representations of their model into a syllable structure (as we did with CDP++) which allowed their model to generalize to nonwords very well, the effect that graphemes have over simply letters may not be especially large. Given this, the other alternative is that the relationships people learn between spelling and sound and spelling and stress are relatively simple, and are hence approximated well via a linear network. This would mean that using a 3-layer network that allows complex and more specific non-linear relationships to be learnt may allow the network to learn things that people do not (see Perry et al. for a further discussion about this in terms of the French orthography). It also means that knowing whether a PDP model trained to be almost perfect on words would behave similarly to the current model of Pagliuca and Monaghan or whether it would learn additional non-linear relationships is important. In this case, additional learning to improve the overall performance of the model on words might also cause it to over-fit the data and hence learn more complex relationships that people do not. In addition to nonword consistency and stress regularity, we also investigated pseudohomophone, morphology, and neighborhood effects. CDP++ was able to produce a pseudohomophone effect, and, like Pagliuca and Monaghan's model, it captured a morphological effect but failed to capture the full pattern of neighborhood effects in Arduino and Burani. The pseudohomophone effect is interesting as it has always been difficult for PDP models to simulate this class of effects, and has generated a reasonable amount of debate (see for a discussion). The morphological effect with nonwords confirms that both CDP++ and the model of Pagliuca and Monaghan are sensitive to morphology even though they do not have morphological processing layers. CDP++ also showed that, like the real data, the reaction times it produces with words were not affected by morphology. Finally, with the neighborhood effect, Pagliuca and Monaghan showed that their model was sensitive to this variable. However, they used a larger and currently untested stimuli set, and they also suggested that different versions of their network might be differentially sensitive to this. Obviously, a mega-study of Italian words would be useful for investigating these effects further. Apart from simulating data of normal readers, we also investigated data from acquired dyslexia. Whilst we did not try to model all of the patterns that exist, we did show that, with two very simple parameter changes, CDP++ can produce a stress dominance effect that is of a similar level to the group of patients that produced the largest effect in Colombo et al. – that is, it showed the most errors on low frequency words with subordinate stress. The model also produced an overall error rate that was very similar to that group. Whether a PDP model is able to approximate this is currently unknown, and represents an interesting challenge given that simulating surface dyslexia has historically been a problem for such models. An important added value of the present modelling enterprise is the fact that CDP++ was able to simulate seemingly discrepant findings, where conflicting results have been reported using essentially the same manipulation. One of the most disconcerting discrepancies was the one between the results reported by Colombo and Burani and Arduino with respect to stress regularity/consistency, with Burani and Arduino suggesting that the difference may be due to the items that were used. CDP++ correctly simulated both sets of results, which shows that their discrepant findings may indeed be due to the actual items selected. Finally, the model was also able to simulate quite complex findings that depended on list context manipulations (see also). For example, in Job et al.'s first two experiments, the authors found a nonword consistency effect in mixed lists of words and nonwords but not in pure lists of only nonwords. They suggested that this occurred because nonword reading can benefit from lexical feedback, and modulating the proportion of nonwords affects the extent of this lexical influence. Our suggestion, alternatively, is that modulating the proportion of nonwords affects people's response criterion (i.e., when they are willing to name the word), and this produces in the model the same pattern observed in the human data and is also consistent with other strategic manipulations that have been reported. In summary, the present work has shown that CDP++ can be easily transposed to a regular orthography with a fairly complex stress system, including mechanisms to do with grapheme parsing and learning. The model is available on-line and can be used to predict results before actually running the critical experiments. # Supporting Information The Italian CDP is available for download at <https://sites.google.com/site/conradperryshome/> and at <http://ccnl.psy.unipd.it/CDP.html>. We would like to thank Christina Burani for especially helpful comments. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CP JZ MZ. Analyzed the data: CP. Wrote the paper: CP JZ MZ. Programmed the model: CP.

# Introduction The growing number of fully sequenced genomes from both prokaryotes and deep- branching eukaryotes offers the possibility of identifying genetic transfers that may have occurred when the first eukaryotes appeared at least 1.5 billion years ago. Eukaryotic cells are chimeric entities, with both mitochondria and chloroplasts derived from endosymbiotic precursors that were distinct from the nucleocytoplasmic lineage. Additionally, many of the original mitochondrially encoded genes were gradually transferred to the eukaryotic nuclear genome *via* endosymbiotic gene transfer. It has been estimated that the α-proteobacterial ancestor of mitochondria contributed at least 630 genes to the eukaryotic nuclear genome. However, thousands of eukaryotic nuclear genes of eubacterial ancestry are not derived from an α-proteobacterial ancestor. The hydrogen hypothesis proposes that the first eukaryote was a consortium of a hydrogen- dependent archaeon and a hydrogen-producing α-proteobacterium, which was the ancestor of the mitochondrion. In this single eubacterial ancestry hypothesis, the mitochondrial ancestor would have assimilated other eubacterial genes via lateral gene transfer prior to the endosymbiotic event. The fluid chromosome model would help to explain such mixed prokaryotic sources of the ancestral mitochondrial genome. This model assumes fluid prokaryotic genomes shaped by gene losses and lateral gene transfers instead of static genomes. Thus, the expected phylogeny for a gene acquired from the mitochondrion would be common ancestry for all eukaryotes but not necessarily trace to α-proteobacteria because the ancestor of mitochondria possessed an as yet unknown collection of genes. Alternatively, it has been recently demonstrated a high frequency gene transfer system in the α-proteobacteria *Rhodobacter capsulatus* based on a virus-like gene transfer agent (GTA), which would facilitate random transfers between species. It has been suggested that the GTA system was present in the last α-proteobacteria common ancestor that could explain not only in as much so many α-proteobacterial genes came to reside in the nucleus, but also why mitochondrial and nuclear genes show such mixed phylogenetic affinities. Several hypotheses have been suggested to account for more than a single (eubacterial) endosymbiont entity. One major hypothesis is metabolic symbiosis, also known as syntrophy, with a symbiont distinct from the ancestral mitochondrion symbiont. One syntrophy case is based upon the exchange of sulfur compounds between a spirochete and an archaeal thermoplasma species. Another case is the hydrogen- driven syntrophy hypothesis, which suggests that the eukaryotic ancestor was derived from symbiosis between an ancestral δ-proteobacteria, specifically a sulfate-reducing myxococcal species, and a methanogenic archaeon followed by the eventual incorporation of an α-proteobacterium. The latter hypothesis argues for the full incorporation of a methanogenic archaebacterium within a myxococcal cytoplasm. The acquisition of mitochondria was independent and early in the existing myxococcal-archaeon consortia, The δ-proteobacteria represent one of the most diverse groups of bacteria, exhibiting a wide array of metabolic strategies, including free-living, syntrophic and pathogenic forms. This group is characterized by large variations in genome size. While species of the genus *Syntrophus* have genomes of approximately 3 Mb, the genomes of myxococcal species are among the largest in prokaryotes, with sizes of approximately 9–13 Mb, which might account for their higher biological complexity. The Myxococcales are characterized by social behavior directed toward predation and the construction of a unique multicellular structure, the asexual fruiting body. The formation of these fruiting bodies requires coordinated cellular motility and cell signaling. Thus, the Myxococcales are of special interest because they represent one of the few prokaryotic lineages that have independently acquired some degree of cellular differentiation and multicellularity. In the Myxococcales, lipids play a role in developmental aggregation, signaling and morphogenesis of the fruiting body. In addition, lipids are major contributors to myxococcal physiology as an energy reservoir, which is also the case in fungi, plants and animals, but not in most other prokaryotes. The Myxococcales have also been shown to share other similarities with eukaryotes, such as the presence of eukaryotic-like protein kinases, Ras-like G-proteins and GTPase-activating proteins that function in regulating cell polarity. Here, we employed a phylogenomic analysis to investigate a possible contribution of Myxococcales to the origin of eukaryotes. Our data revealed that several genes encoding mitochondrial proteins, mostly from the fatty acid β-oxidation pathway, were acquired from Myxococcales. # Results ## Phylogenetic analysis To identify putative genetic transfers between eukaryotes and Myxococcales, we conducted a large-scale comparative genomic analysis of 40 eukaryotic genomes, 27 archaeal genomes and 330 eubacterial genomes, including 6 myxococcal species: *Anaeromyxobacter dehalogenans*, *Haliangium ochraceum*, *Myxococcus xanthus*, *Plesiocystis pacifica*, *Sorangium cellulosum* and *Stigmatella aurantiaca*. A preliminary and conservative step consisting of BLAST (threshold E-value of e-40) homologous sequence searches and maximum likelihood (ML) phylogenetic tree reconstruction identified 93 eukaryotic proteins with a predicted myxococcal origin followed by 40 trees with α-proteobacterial, 15 with archaeal, 8 with firmicutes, 7 with cyanobacterial and 7 with chlamydial origin. The remaining trees traced to other bacterial groups but had less than four trees or were unresolved, with several prokaryotic groups preceding eukaryotes. These 93 positives were further evaluated using HMMER searches, a tool based on hidden Markov models, against additional eukaryotic genomes, and the alignments were performed again with these new taxa. From the inferred ML and Bayesian trees, we identified 15 eukaryotic proteins of obvious myxococcal origin, all with strong or moderate statistical support (see for all ML trees and for all Bayesian trees). Of note, 13 of these 15 genes of myxococcal ancestry were localized to the mitochondria in eukaryotes. We determined the organellar localization of these genes based on bibliographic references and by detecting predicted mitochondrial targeting sequences for several eukaryotes. Eight of these proteins play a role in the formation of the acyl-CoA pool; which comprises pivotal intermediates in lipid metabolism. In particular, we detected a myxococcal ancestry for two eukaryotic acyl-CoA synthetases (ACSs) and, remarkably, six fatty acid β-oxidation enzymes that degrade the acyl-CoA pool: four acyl-CoA dehydrogenases (ACDs), one electron transport flavoprotein (ETF) and one acetyl-CoA acyltransferase with thiolase activity. The ACSs identified corresponded to ACS bubblegum family members 1 and 2 (ACSBG1–ACSBG2) and ACS family member 3 (ACSF3), which activates fatty acids to form acyl-CoA, thus allowing their transport and metabolism. The ACD protein family catalyzes the oxidation of diverse acyl-CoA compounds produced during the degradation of fat and protein to enoyl-CoA. ACD subfamilies are distinguished by the metabolic pathways in which they participate and by their substrate specificity. The four eukaryotic ACDs with myxococcal ancestry participate in the β-oxidation of fatty acids, with optimal activity for acyl-CoA substrates of specific lengths: short (ACADS), medium (ACADM), long unsaturated (ACAD9) or very long (ACADVL). The other ACD subfamily identified is implicated in amino acid degradation. After removal of the amino groups from isoleucine, the remaining branched acyl-CoA is dehydrogenated by the short/branched chain acyl-CoA dehydrogenase (ACADSB). Electron transport flavoprotein A (ETFA) is, in addition to electron transport flavoprotein B (ETFB), the primary acceptor for reducing equivalents from the β-oxidation of acyl-CoA dehydrogenases. The phylogenetic trees of both electron transport flavoproteins (ETFA and ETFB) showed patchy distributions, with eukaryotes nested within different bacterial groups; however, the bacterial group first identified as preceding eukaryotes corresponded to the Myxococcales (see ETFA tree in). Remarkably, an acyl-CoA dehydrogenase gene (ACADM) is located next to the ETFA and ETFB genes in the genome of *M. xanthus*, indicating the importance of electron flow during the β-oxidation of acyl-CoA intermediates in *M. xanthus*. Moreover, the thiolase identified, acetyl-CoA acyltransferase 2 (ACAA2), performs the last thiolytic cleavage of the fatty acid β-oxidation spiral in mitochondria and it should be differentiated from the peroxisomal thiolase ACAA1 and from the thiolases ACAT1 and ACAT2 that can also act in the biosynthesis of eukaryotic ketone bodies and sterols through the condensation of two acetyl-CoA molecules to form acetoacetyl-CoA. The ACAA2 thiolase has previously been reported to have a myxococcal origin. The remaining proteins of unequivocal myxococcal ancestry belonged to different functional groups, and five of them also localized to the mitochondria. These remaining proteins are as follows: 1. three proteases: the M3 family peptidases neurolysin (NLN) and the thimet oligopeptidase (THOP1) and the cytosolic aminopeptidase NPEPL1; 2. a translation elongation factor G that is predicted to localize to mitochondria in organisms ranging from mammals to trypanosomatids and that performs GTP-dependent translocation of the ribosome during translation; 3. the Krebs cycle enzyme NAD(+)-dependent isocitrate dehydrogenase; 4. an arginil tRNA synthetase; 5. an aminotransferase, 4-aminobutyrate aminotransferase; and 6. a ceramidase, N-acylsphingosine amidohydrolase 2, that catalyzes the hydrolysis of the N-acyl linkage of ceramide, a second messenger in a variety of cellular events, to produce sphingosine. Despite the lack of a predicted mitochondrial targeting sequence, this protein has been experimentally localized to the mitochondria of mice and humans. ## Compositional analysis The position of the Myxococcales as a sister group to eukaryotes in phylogenetic trees produced in this study implies the occurrence of lateral gene transfer events from Myxococcales to eukaryotes or vice versa. We investigated this possibility further by performing a sequence compositional analysis of the 15 eukaryotic genes of myxococcal ancestry using homologs from both Myxococcales and unicellular eukaryotes. In cases of recent lateral gene transfer, it has been suggested that the nucleotide composition of the transferred gene might be more similar to that of the donor species than to that of the recipient. The average G+C content of myxococcal species ranges from 69 to 74%. To examine the possibility of a transfer from Myxococcales to eukaryotes, we calculated the average G+C content in the coding sequence (CDS) of our 15 candidate genes in the unicellular eukaryotes and myxococcal homologs and compared them to the average whole-genome G+C content for each species. We found that the 15 CDSs from unicellular eukaryotes were consistently representative of their respective genomes and were thus adapted to function in the organism in which they resided, an observation that was inconsistent with the occurrence of recent lateral gene transfer events. To further examine the possibility of a eukaryote-to-Myxococcales transfer, we used several compositional methods to measure lateral gene transfer in bacteria: (i) Karlin's method based on codon usage, (ii) the Horizontal Gene Transfer Database (HGT-DB), which relies on G+C content, codon usage, gene position and amino acid composition, and (iii) the Island Viewer server, which identifies genomic islands or clusters by integrating sequence composition and comparative genomic approaches. None of the 15 myxococcal ancestors of eukaryotic genes was predicted to have undergone lateral gene transfer from eukaryotes. This analysis only identified overlaps of 3 and 6 bp in ACADS and ACSBG, respectively, in the *M. xanthus* genome. Thus, our findings on the base composition of the 15 CDS candidates are not consistent with recent lateral gene transfers. Instead, our results suggest that ancestral genetic transfers occurred between Myxococcales and eukaryotes, and the incorporated genes might have since converged with the bulk of the genome through a process of amelioration. # Discussion ## Myxococcal origin of part of the mitochondrial β-oxidation pathway Although we used a conservative pipeline (threshold E-value of e-40) and despite few myxococcal genomes being currently available, our results clearly indicate that several eukaryotic genes have a myxococcal ancestry. Half of the genes identified encoded mitochondrial enzymes involved in acyl–CoA intermediate metabolism, primarily in fatty acid β-oxidation. Among the enzymes involved in β-oxidation with a clear myxococcal ancestry were four acyl-CoA dehydrogenases (ACDs), which catalyze the oxidation of diverse acyl-CoA compounds produced during the degradation of fat and amino acid to enoyl-CoA in a substrate- specific manner. Fatty acids can be degraded in hydrogen-driven syntrophy with symbiotic partners under anaerobic conditions. There are natural examples of fatty acid syntrophies with δ-proteobacteria. For instance, the fermenting *Syntrophaceae* family has the ability to grow on fatty acids in syntrophy with methanogens. Notably, the syntrophic oxidation of fatty acids involves hydrogen production from high potential electron donors, such as the acyl-CoA intermediates, that are oxidized by members of the ACD family. Indeed, a genomic analysis of the fatty acid degrading syntrophic bacterium *Syntrophus aciditrophicus* has suggested that the oxidation of acyl-CoA intermediates, which plays a crucial role in generating reducing equivalents, might be specific to syntrophic metabolism. The production of hydrogen from the ACD reaction is thermodynamically unfavorable and can occur only with energy input by a process known as reverse electron transfer. With regard to this mechanism, it has been suggested that ETF could transfer electrons from acyl-CoA intermediates oxidized by ACD to membrane redox complexes, such as the complex that includes the membrane-bound iron-sulfur oxidoreductase present in *S. aciditrophicus*. Our list of putative fatty acid β-oxidation genes with a myxococcal ancestry did not include the genes encoding the central multifunctional proteins involved in β-oxidation (HADHA and HADHB). A possible explanation for this absence may lie in the fact that the human HADHA/HADHB multienzyme complex is formed by a gene fusion between 3-hydroxyacyl-CoA dehydrogenase and enoyl-CoA hydratase of HADHA and the non-covalent interaction of the thiolase activity of HADHB. This arrangement probably resulted from the combination of monospecific enzymatic functions. Thus, the gene fusion present in the human genome from which we retrieved myxococcal homologous sequences (step 1 in our pipeline) might have masked ancestral monofunctional subunits that are still functional in other eukaryote lineages (for example, in *Euglena gracilis*). Indeed, we identified a monofunctional enzyme with 3-hydroxyacyl-CoA dehydrogenase activity with myxococcal ancestry, although the statistical support for this inference was low (38% ML and 0.91 BPP). This protein is absent in Metazoa, but *Capsaspora*, Chromalveolata, *Dictyostelium*, *Naegleria* and Fungi all encode homologs of this gene. Our results contradict those of studies suggesting that the α-proteobacterial ancestor of mitochondria was the donor of multiple genes involved in acyl-CoA metabolism, including the β-oxidation pathway, to the nucleocytoplasmic lineage. This is most likely because our analysis involved a higher number of myxococcal genomes and included a relatively large taxon sampling. However, we cannot exclude the possibility that the α-proteobacterial ancestor of mitochondria was also a fatty acid degrading bacteria, with the possibility that some eukaryotic β-oxidation enzymes could have been acquired from α-proteobacterial genomes. Indeed, our results indicate that some mitochondrial β-oxidation enzymes did not descend from α-proteobacterial homologs because they were absent in the reported trees (see the ACADSB tree in and) or their topology was unrelated to the eukaryotic clade (see ACAA2, ACADM, ACADS, ACADVL-ACAD9 and ACSBG trees in and). ## Lateral gene transfer from Myxococcales to eukaryotes or fluid chromosome? Our results are compatible with the occurrence of multiple ancient lateral gene transfer events from Myxococcales to eukaryotes. We propose three possible explanations for the observation that most of the proteins with myxococcal ancestry found in this study are targeted to the mitochondria. First, myxococcal bacteria may have transferred genes to the nucleocytoplasmic lineage using direct and independent lateral gene transfers. Second, lateral transfer may have occurred *via* endosymbiotic gene transfer from a myxococcal symbiont to the nucleocytoplasmic lineage. This second proposal is compatible with the δ-proteobacteria syntrophy hypothesis, which suggests that the eukaryotic ancestor was derived from a symbiosis between an ancestral sulfate-reducing myxococcal species and a methanogenic archaeon. However, this hypothesis assumes an unrelated incorporation of the myxococcal and mitochondrial ancestors, which does not support the majority of the myxococcal ancestry proteins found in this study being targeted to the mitochondria. The δ-proteobacteria syntrophy hypothesis argues that a myxococcal endosymbiosis occurred prior and independently of the mitochondrion acquisition by an amitochondriate eukaryote. We believe that our findings are more compatible with a simultaneous origin for myxococcal and mitochondria proteins that might underlie the reason why proteins with myxococcal origin are preferably targeted to the mitochondrion. A third possibility is that ancestral lateral gene transfer events occurred between Myxococcales and the mitochondrial ancestor endosymbiont prior to the endosymbiotic event that gave rise to the mitochondrion. In this latter scenario, the myxococcal genes would have already been present in the ancestral mitochondrial genome. This option is compatible with the fluid chromosome model. Thus, myxococcal transfers to eukaryotic genomes would have originated vertically from the mitochondrial progenitor, which was not necessarily an α-proteobacterium because the first mitochondrial genome possessed an as yet unknown collection of genes. Taken together, we believe that the most parsimonious event might have been that a hydrogen-producer, a fatty acid degrading δ-proteobacteria, transferred acyl- CoA related enzymes to the hydrogen producing mitochondrial ancestor. According to the hydrogen hypothesis on the origin of eukaryotes, eukaryotes arose from a hydrogen-driven syntrophy between the hydrogen producing α-proteobacteria symbiont and a hydrogen-consuming methanogenic archaeon. In this scenario, harboring several ACD reducing equivalents producers acquired from Myxococcales could be of advantage for the hydrogen producing mitochondrial ancestor, fostering hydrogen, acetate and CO₂ production. Once the host methane production became irreversibly dependent on the symbiont hydrogen source, the methanogenic archaebacterium would have maximized its contact with the surface area of the symbiont by surrounding it, eventually engulfing the hydrogen producing symbiont. Over time, genes would have been transferred through endosymbiotic gene transfer to the eukaryotic nucleocytoplasmic lineage. If this scenario is true, we would expect that at least some mitochondrially encoded genes would trace to Myxococcales. To test this possibility, we analyzed the myxococcal ancestry of mitochondrially encoded proteins from 2209 mitochondrial genomes. Homologous sequence searches and subsequent phylogenetic reconstruction indicated that none of the mitochondrially encoded proteins had a myxococcal origin (data not shown). Taking into account this result, we cannot discard the possibility of a myxococcal endosymbiosis. The lack of myxococcal ancestry in mitochondrially encoded proteins might apparently support an independent myxococcal endosymbiotic event, but then the genes transferred to the host should be targeted not only to the mitochondrion but colonizing other subcellular compartments. As far as the myxococcal proteins identified are preferably targeted to the mitochondria, we believe that the most compatible scenario is the simultaneous origin for mitochondrial and myxococcal proteins. According to that, there are two possibilities: i) that the endosymbiotic event that gave rise to the mitochondrion also involved a myxoccocal bacterial partner that subsequently transferred a number of metabolic pathways to the host via endosymbiotic gene transfer; or ii) ancestral lateral gene transfer events took place between Myxococcales and the mitochondrial ancestor endosymbiont prior to the endosymbiotic event that gave rise to the mitochondrion. Supporting the last option, the mitochondrial genome seems to retain specific functional genes that do not descend from the myxococcal genomes. In this sense, the presence of a mitochondrial genome appears to correlate with increased respiratory capacity and ATP availability within organisms. Therefore, it has been suggested that this relationship could be the driving force behind the selective pressure to preferentially retain the genes encoding the electron transport chain in the mitochondrial genome. These genes are mostly derived from α-proteobacterial genomes; in contrast, the myxobacterially derived metabolic genes would have been transferred to the eukaryotic nucleocytoplasmic lineage. This transfer might account for the absence of mitochondrially encoded genes with myxococcal ancestry. It is worth mentioning that because Myxococcales harbors some of the largest eubacterial genomes, lineage sorting could influence our topologies. However, most of our trees have representation from a wide sampling of other eubacterial genomes. Our results indicate that the mitochondrion might be an important entry point of eubacterial genes that are different from α-proteobacterial genes. Thus, the genome of the mitochondrial ancestor should be considered a mix of different eubacterial genes, some of them still conserved in the extant α-proteobacterial and myxococcal genomes. Further, it is tempting to speculate that the large genomes of the Myxococcales could be a repository of ancestral genes from many prokaryotic lineages, including that of the mitochondrial ancestor. In conclusion, our data indicate the myxococcal origin of 15 nuclear eukaryotic genes that are not α-proteobacterial, some of which have key roles in acyl-CoA intermediate metabolism. Many other genes may also have a myxococcal origin, but the lack of a phylogenetic signal and/or our extremely conservative pipeline did not allow us to recover them. We propose that a fatty acid degrading δ-proteobacterium donated some genes to the mitochondrial ancestor prior to the endosymbiotic event, and these genes were subsequently transferred to the nucleocytoplasmic lineage via endosymbiotic gene transfer. Thus, our results support a version of the fluid chromosome model as the most plausible scenario, in which Myxococcales contributed key metabolic genes to the first eukaryotes. # Materials and Methods ## Data retrieval Proteomes encoded by 355 publicly available complete genomes of eubacteria and archaea were obtained from the National Center for Biotechnology Information (NCBI) FTP server (<ftp://ftp.ncbi.nih.gov/genomes/Bacteria>), including four complete myxococcal genomes. In addition, we included the draft assembly genomes of the Myxococcales species *P. pacifica* and *S. aurantiaca*. We sampled 40 and 49 eukaryotic genomes for the first and second analyses, respectively, with the goal of encompassing the widest possible eukaryotic diversity. Sequence analyses were performed for a range of members of the Amoebozoa (*Dictyostelium discoideum*), the Archaeplastida (*Arabidopsis thaliana*, *Chlamydomonas reinhardtii*, *Cyanidioschyzon merolae*, *Oryza sativa* and *Ostreococcus tauri*), the Chromalveolata (*Paramecium tetraurelia*, *Phaeodactylum tricornutum*, *Phytophthora sojae*, *Tetrahymena thermophila*, *Theileria parva* and *Thalassiosira pseudonana*), the Excavata (*Giardia lamblia*, *Leishmania major*, *Naegleria gruberi*, *Trichomonas vaginalis*, and *Trypanosoma brucei and cruzi*), and the Opisthokonta (*Anopheles gambiae*, *Apis mellifera*, *Aspergillus aspergillus*, *Bos taurus*, *Caenorhabditis elegans*, *Candida glabrata*, *Canis familiaris*, *Capsaspora owczarzaki*, *Ciona intestinalis*, *Cryptococcus neoformans*, *Danio rerio*, *Debaryomyces hansenii*, *Drosophila melanogaster*, *Gallus gallus*, *Homo sapiens*, *Encephalitozoon cuniculi*, *Eremothecium gossypii*, *Kluyveromyces lactis*, *Macaca mulatta*, *Monodelphis domestica*, *Monosiga brevicollis*, *Mus musculus*, *Neurospora crassa*, *Pan troglodytes*, *Rattus norvegiccus*, *Takifugu rubripes*, *Tetraodon nigroviridis*, *Saccharomyces cerevisiae*, *Schizosaccharomyces pombe*, *Xenopus tropicalis* and *Yarrowia lipolytica*). The databases used included the ENSEMBL databases for *A. gambiae*, *A. mellifera*, *B. taurus*, *C. elegans*, *C. familiaris*, *C. intestinalis*, *D. rerio*, *D. melanogaster*, *G. gallus*, *H. sapiens*, *M. mulatta*, *M. domestica*, *M. musculus*, *P. troglodytes*, *R. norvegiccus*, *S. cerevisiae*, *T. rubripes*, *T. nigroviridis* and *X. tropicalis*; the NCBI genomes for *A. fumigatus*, *C. glabrata*, *C. neoformans*, D. hansenii, *E. cuniculi*, *E. gossypii*, *K. lactis*, *L. major*, *N. crassa*, *O. sativa*, *S. pombe*, *T. thermophila*, *T. brucei*, *T. cruzi*, *T parva* and *Y. lipolytica*; the *Cyanidioschyzon merolae* Genome Project (<http://merolae.biol.s.u-tokyo.ac.jp>) for *C. merolae*; DictyBase (<http://dictybase.org>) for *D. discoideum*; GiardiaDB (<http://giardiadb.org>) for *G. lamblia*; the Integr8 Database for *A thaliana*; TrichDB (<http://trichdb.org>) for *T. vaginalis*; and the Joint Genome Institute Eukaryotic Genomics databases (<http://genome.jgi-psf.org>) for *T. pseudonana, N. gruberi, P. sojae, P. tricornutum, O. tauri, M. brevicollis, C. reinhardtii, P. tetraurelia and C. owczarzaki*. For the second analysis, we included the following genomes: *N. gruberi*, *P. sojae*, *P. tricornutum*, *O. tauri*, *M. brevicollis*, *C. reinhardtii*, *P. tetraurelia*, *C. owczarzaki* and *O. sativa*. ## Homolog sequence searches and phylogenetic reconstruction For the complete Myxococcales genomes and the *P. pacifica* proteome, homologous sequences were retrieved using blastx and blastp algorithms (E\<10⁻⁴⁰), respectively, against the human proteome. This search yielded a set of 471 human proteins that were used as seed proteins to identify homologous proteins against the ensemble of 357 bacterial and 40 eukaryotic genomes described above (E\<10⁻⁴⁰). Groups of homologous sequences for each of the 471 seed proteins were aligned using MAFFT with default parameters. Gap rich positions in the alignment were removed using trimAl v1.2, applying a gap threshold of 25% and a conservation threshold of 50%. Maximum likelihood (ML) phylogenetic trees were then reconstructed in RAxML 7.0.4 using the Whelan and Goldman (WAG) matrix of amino acid replacements and assuming a proportion of invariant positions (WAG+I). The number of bootstrapping runs was automatically determined using a newly implemented rapid bootstrap algorithm for RAxML using CIPRES-Portal 2.0. The resulting 471 ML trees were examined, yielding 93 trees with a eukaryotic cluster that putatively branched with a myxococcal clade. For each of the resulting 93 trees, we built the eukaryotic protein profile using a tool based on hidden Markov models, HMMER version 3.0 (<http://hmmer.org>). The eukaryotic proteins from these selected trees were used to repeat the searches with HMMER on a proteome dataset including 357 bacterial and archaeal proteomes and 49 eukaryotic proteomes using a cutoff range of E\<10⁻³. Groups of homologous sequences were aligned using MAFFT. Insertions and sequence characters that could not be aligned with confidence and incomplete sequences were removed. Multi-aligned sequences were manually examined. Additional phylogenetic analyses were performed using both RAxML and MrBayes analyses. One thousand replicates of rapid bootstrap analyses were performed using RAxML 7.2.6 with the general time-reversible (GTR) matrix of amino acid replacements and four gamma-distributed rates (GTR+Г) using CIPRES-Portal 2.0. Additional phylogenetic analyses were performed using a Bayesian method implemented in MrBayes with a mixed model of amino acid substitution and with gamma correction, including four discrete correction categories and a proportion of invariant sites (WAG+Г+I). MrBayes was run until a convergence diagnostic with a standard deviation of split frequencies \<0.01 was achieved or until the likelihood of the cold chain stopped increasing and began to randomly fluctuate, thus reaching a stationary state. The analyses were performed with eight chains, and trees were sampled every 1000 generations in two runs. To construct the consensus tree, the first 10% of trees were discarded. When necessary, large unresolved trees were split into smaller partitions with strong bootstrap support or differentiated long branches to facilitate the analysis. ## Mitochondrial proteome analysis Mitochondrial proteomes were retrieved from 2209 eukaryotes from <http://www.ncbi.nlm.nih.gov/genomes/GenomesGroup.cgi?opt=organelle&taxid=2759>. Approximately 30000 proteins were grouped into 53 orthologous groups, and homologous BLAST searches were performed against 357 bacterial and archaeal proteomes (E\<10⁻²⁰). Phylogenetic reconstruction revealed only 12 potential myxococcal ancestries, which were then discarded based on HMM searches (E\<10⁻³) and subsequent ML and Bayesian phylogenetic reconstructions. ## Lateral gene transfer predictions We analyzed the sequence composition of the fully sequenced myxococcal genomes of *A. dehalogenans*, *H. ochraceum*, *M. xanthus*, and *S. cellulosum* using three different methods: (i) Karlin's method based on determined codon usage and calculated using software available from the Computational Microbiology Laboratory (<http://www.cmbl.uga.edu/software.html>), (ii) the Horizontal Gene Transfer Database (HGT-DB), which relies on G+C content, codon usage, gene position and amino acid content (<http://genomes.urv.es/HGT-DB/>), and (iii) the Island Viewer server to identify genomic islands or clusters, which integrates sequence composition and comparative genomics approaches (<http://www.pathogenomics.sfu.ca/islandviewer>). In supplementary figures, the open reading frames (ORFs) were depicted using CGView. Codon usage tables and G+C content values from eukaryotes and Myxococcales were extracted from the Codon Usage Database (<http://www.kazusa.or.jp/codon/>). # Supporting Information The authors gratefully acknowledge the computer resources, technical expertise and assistance provided by the Barcelona Supercomputing Center (Centro Nacional de Supercomputación) and the Spanish National Bioinformatics Institute (Instituto Nacional de Bioinformática; INB). Parts of the simulations were performed using the freely available CIPRES-Portal 1.0 and 2.0 and Bioportal ([www.bioportal.uio.no](http://www.bioportal.uio.no)). [^1]: Conceived and designed the experiments: AS. Performed the experiments: AS. Analyzed the data: AS IR-T AP. Contributed reagents/materials/analysis tools: AS. Wrote the paper: AS IR-T AP. [^2]: The authors have declared that no competing interests exist.

# Introduction The chloroplast is an important organelle in the plant cell, and its central function is to carry out photosynthesis and carbon fixation. In general, the chloroplast (cp) genome is highly conserved among seed plants with two copies of a large inverted repeat (IR) separated by small single copy (SSC) and large single copy (LSC) regions. It usually contains 110–130 unique genes, which can be roughly divided into three large groups according to their functions: genetic system genes, photosynthesis genes and conserved open reading frames with miscellaneous functions. However, a small group of angiosperm plants appear to have escaped from this dominant pattern by evolving the capacity to gain the water, carbon and nutrients via the vascular tissue of the parasitized host’s roots or shoots. This means that these parasitic plants have reduced (or no) photosynthetic ability, and no longer need genes that encode photosynthetic proteins. With the selective constraints on their cp coding genes relaxed, gene losses occur in these parasitic plants. It is estimated that approximately 1% of all angiosperm species have resorted to a parasitic lifestyle, which has independently evolved 12 or 13 times. Compared with a rapid rise in the number of cp genomes of photosynthetic organisms available on NCBI (254 in Viridiplantae, as of December 4, 2012), there are limited data sets from parasitic plants, especially from the completely non-photosynthetic species. In higher plants, the cp genome of holoparasite *Epifagus virginiana* in the family Orobanchaceae was sequenced first, followed by four species from the holoparasitic genus *Cuscuta*, and three mycoheterotrophic plants, including *Aneura mirabilis*, *Rhizanthella gardneri*, and *Neottia nidus-avis*. However, only one species of the completely non-photosynthetic plants, *E. virginiana*, which exploit other plants via direct connections rather than by mycorrhizal fungi, has been comprehensively analyzed in its cp genome structure and composition. Orobanchaceae, as taxonomically redefined by a series of recent molecular studies, comprise around 89 genera and more than 2,000 species, making it the largest predominantly parasitic angiosperm family, the majority of which are facultative or obligate root parasites –. It contains all levels of parasitic ability ranging from nonparasitic to hemiparasitic and holoparasitic. Therefore, analyses of cp genomes of other holoparasitic species within the family Orobanchaceae could confirm the common attributes of non-photosynthetic evolution and provide point for genetic analysis of cp genome evolution. In Orobanchaceae, *Cistanche* is a worldwide genus of holoparasitic desert plants. Specifically, *C. deserticola*, commonly known as desert-broomrape and traditionally used as an important tonic in China and Japan, is distributed in Northwest China and the Mongolian People’s Republic, and is also considered to be an endangered wild species in recent years due to increased consumption by humans. *C. deserticola* is parasitized on the roots of psammophyte *Haloxylon ammodendron* (Chenopodiaceae), which mainly inhabit deserts and semi-deserts due to its high tolerance to drought and salinity. Similar to *E. virginiana*, *C. deserticola* is a completely non-photosynthetic species and usually grows underground. A number of studies about the chemical components or pharmacological effects of this species have been reported. Further analysis of its cp genome structure and composition could provide new insights on the evolution of the parasitic cp genome. Attributed to the direct connections between parasitic plants and their hosts, which allows the channelling of metabolites, such as sugars, amino acids and perhaps nucleic acids in the form of mRNA, direct haustorial contact between them usually facilitates a horizontal gene transfer (HGT) from a donor to a recipient plant. HGT, known as exchange of genes across mating barriers, has played a major role in bacterial evolution. In recent years, increasing studies have reported HGT being recognized as a significant force in the evolution of eukaryotic genomes. In plants, the evolutionarily earliest examples of HGT might be the endosymbioses that gave rise to mitochondria and chloroplasts. Since the emergence of HGT events, usually detected as incongruences in molecular phylogenetic trees, a considerable number of studies have suggested gene exchanges between hosts and parasites,. Although the HGT involving parasitic plants appears to have occurred in many parasitic lineages, the majority of reported cases of HGT have been limited to exchanges between mitochondrial genes among related species,. Cases of HGT involving cp genomes are rare. The disparity in frequency of plant-to-plant HGT between the mitochondrial and the cp genomes is considered due to an active homologous recombination system,. It is reported that a chloroplast region including *rps*2, *trn*L-F, and *rbc*L among a group of nonphotosynthetic flowering plants, *Phelipanche* and *Orobanche* species, both from the family Orobanchaceae, were detected according to the phylogenetic trees based on available data. In order to examine the effect of its non-photosynthetic life history on cp genome content, we sequenced the entire cp genome of *C. deserticola*. As a completely non-photosynthetic species from Orobanchaceae, it shows the same pattern in the process of gene loss as in chloroplasts of *E. virginiana* and other parasitic plants. We also found that *C. deserticola* has two copies of a cp gene *rpo*C2, one becoming a pseudogene, the other being horizontally acquired from the host *H. ammodendron,* according to a homology search and phylogenetic analysis. # Materials and Methods ## Genome Sequencing and Assembly The spikes of C. deserticola were collected from a plant base in Bayannur City of Inner-Mongolia area which was introduced from natural populations located in desert area of Inner Mongolia in northeastern China. The collecting permit was obtained from the owner (Jun Wei) of the plant base. The voucher specimen was deposited in the MOE Key Laboratory for Biodiversity Science and Ecological Engineering at Fudan University. For cp genome sequencing, total genomic DNA extraction was performed using the Plant Genomic DNA Kit (Tiangen Biotech Co., China), following the manufacturer’s instructions. The fragments of cp DNA were amplified by the polymerase chain reaction (PCR). In brief, due to loss and pseudogenizations in the cp genome of *C. deserticola*, PCR primers were designed using the reported PCR primers from several sources. The primers of the LSC region were designed using the reported conserved cp DNA primer pairs, which including 38 primer pairs as well as eight primer pairs flanking cpDNA microsatellites tested on 20 plant species from 13 families. Only 14 of the 38 primer pairs are useable in *C. deserticola*. Then the primers for other regions were designed according to the primers of the cp genome available in the cp genome database. Some primers were also developed from the cp genome sequences of related species (*Olea europaea* and *E. virginiana*) for specific regions. In order to amplify longer fragments, some of these primers were used combined, and some of them were designed based on the newly determined sequences of adjacent regions. By using all above primers, we covered the entire cp genome of *C. deserticola* with PCR fragments ranging in size from 500 bp to 3 kb. The overlapping regions of each pair of adjacent PCR fragments exceeded 150 bp. The amplified product was purified, and ligated into TaKaRa pMD19-T plasmids (TaKaRa BioInc, Shiga, Japan), which were then cloned into *Escherichia coli* strain DH5a. Multiple (≥6) clones were randomly selected and followed by automated sequencing using ABI 3730xl DNA Analyzer (Applied Biosystems, Foster City, CA). All fragments were sequenced 2–10 times (6-fold coverage of the *C. deserticola* cp genome on average). All these individual sequences were excluded vector, primer and low-quality reads, and then assembled using Sequencher 3.0 software (Gene Codes Corporation, USA). The inverted repeat regions (IRs) of the cpDNA were not amplified separately, but primers were designed to amplify the regions spanning the junctions of LSC/IRA, LSC/IRB, SSC/IRA and SSC/IRB. Considering two IRs cannot be distinguished by automated assembly software, we input the reads as two groups and obtained two large contigs, with each contig including one IR and its adjacent partial LSC and SSC regions. Then, the two large contigs were manually assembled into the complete circular genome sequence. ## Genome Annotation and Molecular Evolutionary Analyses Initial gene annotations were performed using the chloroplast annotation package DOGMA (<http://phylocluster.biosci.utexas.edu/dogma/>). Genes that were undetected by DOGMA, such as *psb*B, *psb*K, *trn*G-GCC, *rpo*C2, *atp*B, *acc*D, and *ycf*1, were identified by Blastn (<http://blast.ncbi.nlm.nih.gov/Blast.cgi>). The correctness of the annotation for all genes was additionally verified by a similarity search against the available plant cp genome sequences. The regions with similarity to known protein coding genes but lacking intact open reading frames (ORF) were identified as pseudogenes. tRNA genes were annotated using DOGMA and ARAGORN v1.2 (<http://130.235.46.10/ARAGORN/>), and then confirmed by ERPIN (<http://tagc.univ-mrs.fr/erpin/>). The circular gene map of the *C. deserticola* cp genome was drawn by GenomeVx followed by manual modification. An assembled and corrected sequence of *C. deserticola* cp genome was deposited in GenBank. To estimate the selection constraint on the genes remaining in *C. deserticola* cp genome, the protein-coding genes that shared between *C. deserticola*, *E. virginiana*, and related photosynthetic species *O. europaea* were chosen to calculate the ratio of the rates of nonsynonymous and synonymous changes (*dN/dS*). *Nicotiana tabacum* was also included in the analyses to calculate *dN/dS* for photosynthetic plants. Alignment was performed using ClustalW. The pairwise *dS, dN* and *dN/dS* ratios were calculated using DnaSP ver. 5. ## Isolation and Sequencing of Potential HGT Gene (*rpoC2*) In this study, we found that *C. deserticola* harbours another copy of *rpo*C2 outwith its own, which corresponds to the phylogenetic position of the host *H. ammodendron*. We propose that this copy arose via horizontal gene transfer. In order to confirm if HGT occurred across the range of C. deserticola, we sampled 50 accessions from the same plant base in Bayannur City which were introduced from five natural populations located in Alxa Left Banner, Alxa Right banner (two populations), and Urad Rear Banner in the Inner-Mongolia area, as well as Hetian in the Xinjiang area. In order to confirm if the HGT occurred across the range of *C. deserticola*, total DNA extractions from these materials and cp *rpo*C2 genes amplified by standard PCR were performed in two different labs, thus, eliminating laboratory contamination. The *rpo*C2 gene was amplified using primers f4 (5′-GATAGACATCGGTACTCCAGTGC-3′) and r6 (3′-TCATTATGGGAATGTACACGCG-5′) with the following conditions: 94°C for 2.5 min; 35 cycles each at 94°C for 1 min,55°C for 30 s, 72°C for 1 min. In *H. ammodendron*, the five clones of *rpo*C2 gene were also checked. ## Phylogenetic Analyses of the Potential HGT Gene, *rpoC2* All the copies of *rpo*C2 sequences detected in *C. deserticola* and *H. ammodendron* were used as queries for BLASTN searches against the NCBI database (E-value \<10⁻³) to identify and retrieve their homologs. On the basis of Angiosperm Phylogeny Group III, sampling for the present study focused on members of the clade Lamiales and Caryophyllales that includes two families Orobanchaceae and Chenopodiaceae. Finally, 29 sequences were sampled for the *rpo*C2 phylogenetic analyses, using *Oryza nivara* (NC_005973) as outgroups. Sequences were unambiguously aligned manually in BioEdit 7.0.4.1. Phylogenetic analyses were performed using maximum likelihood in PAUP v. 4.0b10 and Bayesian inference in MrBayes v. 3.1. The appropriate ML model of nucleotide substitution (GTR+I+G) was determined by Modeltest 3.7 according to the Akaike information criteria (AIC). Relative clade support was estimated by ML bootstrap analysis of 100 replicates of heuristic searches with settings as above. Bayesian analysis was performed with MrBayes 3.1 using same model (GTR+I+G) suggested by MrModeltest v2.2. The settings for the Metropolis-coupled Markov chain Monte Carlo process were: three runs with four chains each were run simultaneously for 1\*10⁷ generations, which were logged every 1000 generations. Convergence was considered to have been reached when the variance of split frequencies was \<0.01. The first 2500 generations were discarded as the transient burn-in period. The 50%-majority-rule consensus of trees sampled in the Bayesian phylogenetic analysis was used to construct a phylogram. # Results ## The cp Genome Structure of *C. deserticola* As expected, the cp genome of *C. deserticola* \[GenBank number: KC128846\] is greatly reduced in size (102,657 bp) and in gene content. It is a quadripartite structure typical of the majority of land plant chloroplast chromosomes with a large single copy (LSC) region of 49,130 bp separated from 8,819 bp small single copy (SSC) region by two inverted repeats (IRs), each of 22,354 bp. In angiosperms, it is the fourth completely nonphotosynthetic species and the eighth parasitic species of which complete sequences of the cp genome are now available. Among these species, the cp genome of *C. deserticola* is larger than those of other five holoparasites species (*E. virginiana, R. gardneri*, *N. nidus-avis*, *Cuscuta obtusiflora* and *C. gronovii*), but is smaller than those of other hemiparasitic *Cuscuta* species, which has more or less green color distributed throughout the stems and inflorescences. When the IR is considered only once, the cp genome of *C. deserticola* contains 60 genes, encoding 27 proteins, 4 ribosomal RNAs (rRNA) and 30 transfer RNAs (tRNA). The positions of 61 genes, including 29 unique and 16 duplicated ones in the IRs regions, were localized on the map. The cp genome of *C. deserticola* has an overall GC content of 36.8%, which is similar to *E. virginiana* (36%) but slightly lower than the photosynthetic species *Nicotiana tabacum* (37.8%). Like other land plants, GC content is unevenly distributed across the *C. deserticola* cp genome. The highest GC content is in the IRs (43%), reflecting the high GC content of rRNA genes, and the lowest is in the SSC (27.5%) region. Although *C. deserticola* has a relatively larger cp genome sequence, it also exhibited severe physiological reductions: all genes required for photosynthesis (encoding photosystem I and II components, cytochrome b6f complex, NAD(P)H dehydrogenase, photosystem assembly factors (*ycf*3, *ycf*4) and ATP synthase) suffered gene losses and pseudogenizations except for *psb*M. Additional pseudogenization is also seen in genes encoding cp-encoded RNA polymerase (*rpo*), Cytochrome c biogenesis protein (*ccs*A), and Acetyl-CoA carboxylase (*acc*D). *C. deserticola* retains many genes of the translation machinery, including 8 *rpl* genes, 11 *rps* genes, and an initiation factor, infA. Only *rpl*23 is apparently a pseudogene with nonfunctional reading frames. In *E. virginiana*, a total of five tRNAs were pseudogenized and eight tRNAs were lost. In contrast, *C. deserticola* retains almost all the rRNA and tRNA genes: two identical copies of rRNA gene clusters (16S-23S-4.5S-5S) were found in the IR regions; 30 different tRNAs, which can recognize all 61 codons present in the cp genes, were identified. Intron content of genes retained in the *C. deserticola* cp genome is conserved with other seed plants: it has 11 genes with introns, six in tRNAs and 5 in protein coding genes. Two of the 12 intron- containing genes have a single intron and two genes, *clp*P and *rps*12, have two introns. All of these belong to the group II intron, whereas *trn*L-UAA is the only group I intron. Among the tRNA genes, *trn*K-UUU has a special role, since the only RNA maturase gene (*mat*K) found on the cp genome was located in its intron. Unlike other parasitic plants, *C. deserticola* harbours the complete *trn*K-UUU gene, including its intron *mat*K gene. The examination of pairwise *dN/dS* ratio for the alignable genes shared between *C. deserticola*, *E. virginiana* and their autotrophic relatives demonstrates that most of genes are under greater constraint in fully nonphotosynthetic *E. virginiana* and *C. deserticola* than photosynthetic species *O. europaea* and *N. tabacum*. In addition, of 15 protein-coding genes shared between *E.virginiana* and *C. deserticola*, 13 genes have a higher *dN/dS* in *E.virginiana* than in *C. deserticola*. ## Horizontal Transfer of *rpo*C2 from Host to Parasitic Plants We used the general primers of *rpo*C2 to amplify the total DNA of *C. deserticola* and subsequent cloning. Contrary to our expectation, the sequences obtained resemble the genes of *H. ammodendron* but not *C. deserticola*, based on sequence similarity through BLAST and phylogenetic trees (GenBank number: KC543998). This raised the possibility that HGT may have occurred between the parasite and its host. In order to confirm this result, we ruled out that the results were due to contamination or mixing-up of templates by repeating the experiment in a different laboratory. The results of the amplification are congruent with the previous results. Then, we confirmed the presence of the *H. ammodendron* type copy in 46 accessions out of 50 samples from five *C. deserticola* populations by using the same specific primers. Four of the accessions’ lack of amplification was probably due to poor DNA quality. The transferred *rpo*C2 copy (*H. ammodendron* type) amplified from *C. deserticola* is about 1050 bp, and covers amino acid positions 98–443 (nucleotide positions 294–1329) of the *rpo*C2 gene of *O. europea*. To clarify the evolutionary characteristics of the *rpo*C2 fragment transferred from *H. ammodendron* to *C. deserticola*, we aligned the nucleotide sequences of *H. ammodendron* type *rpo*C2 amplified from both of two plants with intact open reading frames of other related species. The results indicated that the transferred *rpo*C2 fragment differed from functional copies in a few point mutations and one key nucleotide insertion (C in 927 bp), which resulted in several subsequent premature termination codons and frame shifts mutations. Because the *H. ammodendron* type *rpo*C2 was not found in the complete cp genome of *C. deserticola*, we speculated it should be transferred into the nuclear or mitochondrial genome. However, *C. deserticola*’s own *rpo*C2 copies were not detected by PCR amplification using specific primers, which make us consider that this gene was lost or turned out to be a pseudogene. Thus, we searched the finished cp genome of *C. deserticola* with *rpo*C2 homologues by the BLAST method. The results shown that *C. deserticola* also retains its own significant shortened *rpo*C2, which has turned out to be a pseudogene of only 439 bp. ## Phylogenetic Analysis of Transferred *rpo*C2 Gene The HGT result was further supported by our phylogenetic analysis. Maximum likelihood and Bayesian trees constructed using the two methods described earlier gave congruent results. The two orders Lamiales and Caryophyllales confirmed as well as supported distinct clades in the phylogenetic tree. The transferred *rpo*C2 is located in the clade Caryophyllales (host clade) but does not cluster inside Lamiales (parasitic clade), which forms a clade with a relatively strong bootstrap support. The retained *rpo*C2 (*C. deserticola* type copy) was not used in this analysis because its sequence was severely fragmented when align with other homologs. The sequence alignment and the phylogenetic distribution of the *rpo*C2 in Chenopodiaceae suggest that the horizontal gene transfer happened between the host *H. ammodendron* and parasitic plant *C. deserticola*. # Discussion ## Gene Losse in the cp Genome of *C. deserticola* Compared to more than 250 completely sequenced cp genomes of photosynthetic plants, the number of fully sequenced cp genomes of non-photosynthetic plants is very small. To date, only eight heterotrophic species, exhibiting parasitic lifestyles and having strongly reduced cp genomes, have been thoroughly investigated with respect to their cp genome sequences. In this study, we have sequenced the cp genome of *C. deserticola*, a holoparasitic species from Orobanchaceae with the expectation that comparison of cp genomic features between these two relatives will provide further insights on parasitic cp genome evolution. The overlapping PCR products have indicated the reduced circular form of the cp chromosomes in *C. deserticola*. Similar to *E. virginiana*, almost all of its photosynthetic genes have been lost or have become pseudogenes after the loss of a major metabolic function. It is different from other heterotrophic plants in many ways: it retains almost all the tRNA genes; the photosynthetic gene *psb*M remains as residues and others suffered gene pseudogenizations rather than losses as *E. virginiana*; *C. deserticola* harbours complete *trn*K-UUU gene but not its intron *mat*K gene, and so on. Some parasitic species exhibit extensive losses of tRNA genes. In *E. virginiana*, a total of 13 tRNAs were pseudogenized or lost. As in photosynthetic plants, *C. deserticola* encompasses around 30 tRNA genes in cp genomes, and it is the only one of parasitic plants which possessed a full cp tRNA set as nonparasitic plants. This suggests that the loss of the transfer RNA genes from the cp genome occurred later than those of photosynthesic genes. Most of the splicing factors are nuclear-encoded, but one maturase protein is encoded by a cp gene, *mat*K, which was located within an intron of *trn*K-UUU. The *trn*K gene is lost in all parasitic angiosperm cp genomes except for *C. deserticola* and *Neottia nidus-avis*. In the *Neottia* cp genome, the intron *mat*K is a pseudogene with strong divergence of its 5′end compared to other photosynthetic orchids. In contrast, in *C. reflexa*, *C. exaltata*, and *E. virginiana*, *matK* has been retained as a free-standing gene,. Unlike other parasitic angiosperm species, neither the *trn*K-UUU gene nor its intron *mat*K gene was missing in *C. deserticola*. It has been reported that *matK* is also needed for splicing other chloroplast group II introns in the cp genome. Thus the retaining of *matK* in *C. deserticola* is not surprising because its cp genome has retained 9 group IIa introns (including *rpl*2, *rpl*16, *rps*12, *clp*P, *trn*A-UGC, *trn*I-GAU, *trn*K-UUU, *trn*G-UCC, *trn*V-UAC). While the *trnK* gene exists in the *C. deserticola* cp genome, which was similar to photosynthetic plants, this may suggest the plant has undergone fewer losses, either due to a function of reduced level of holoparasitism, or a recent switch to this life history. The entire set of chloroplast NAD (P) H dehydrogenase consisting of 11 genes has been lost or turned into pseudogenes without exception in *C. deserticola*. What is interesting is that a loss of *ndh* genes was also present in all sequenced cp genomes of parasitic plants investigated to date, regardless of the degree of evolutionary degradation of photosynthetic capacity. It was confirmed that cp- encoded *ndh* genes were first lost in the transition to heterotrophy. It has been speculated that the condensation of the genome by loss of many non-coding regions and unimportant parts of the cp genome is an early reaction of the cp genome to the parasitic lifestyle. After calculating *dN/dS* for shared cp genes between *E. virginiana, C. deserticola* and two photosynthetic species, an obvious trend of relaxed selection was revealed in both fully nonphotosynthetic species with higher *dN/dS*. It may indicate that these genes were suffering an initial stage of pseudogenization. However, *C. deserticola* has lower *dN/dS* for more genes than *E. virginiana*, which suffered a high degree of gene loss and pseudogenization, further indicating *C. deserticola* may undergo reduced level of holoparasitism or a recent switch to this life history. The gene *psb*M, which was the only one photosynthetic gene retained in *C. deserticola*, showed a higher *dN/dS* than in photosynthetic species (*dN/dS = *0), suggesting advent of relaxed selection and initial stage of pseudogenization in this gene in *C. deserticola*. However, some unexpected high *dN/dS* were also found in *rpl*33, *rps*7 and *rpl*22 in photosynthetic species. The short length of sequences may reduce the reliability of dN/dS estimation in these genes. ## HGT from *H. ammodendron* to *C. deserticola* HGT in parasitic systems has been detected by using phylogenetic trees when a DNA sequence obtained from a parasite is placed closer to its host rather than with its closest relatives. Unexpectedly we had a windfall in the process of amplifying the cp genome sequence of *C. deserticola*. One of these sequences, *rpo*C2 gene, was present in two copies within this parasite and one of them was a homolog of their host and led to conflicting phylogenies. The most reasonable explanation for our results is that cp *rpo*C2 gene in *C. deserticola* was acquired from its host, *H. ammodendron* via HGT. In order to confirm the results and provide special opportunities for studying the evolutionary dynamics of HGT at the population level, we also collected 50 samples from five populations and successfully amplified transferred the *rpo*C2 gene from 46 accessions. In addition, the events present in most individuals spanning Xinjiang and Inner Mongolia, may suggest that the HGT of *rpo*C2 probably occurred in a *C. deserticola* common ancestor of these populations, which expanded into its present wide distribution quickly. So far, the incidence of HGT in the family Orobanchaceae is high, including one nuclear HGT event which occurred between parasitic *Striga hermonthica* (Orobanchaceae) and its host *Sorghum bicolor* (Poaceae), as well as a chloroplast region including *rps*2, *trn*L-F, and *rbc*L in a group of non- photosynthetic members (*Orobanche* and *Phelipanche*) of Orobanchaceae. Our study shows that cp *rpo*C2 has transferred from *H. ammodendron* to *C. deserticola* via HGT. However, it is impossible to presume the localization of the transferred *rpo*C2 based on the available data. We just could rule out its location in the cp genome according to our completed cp genome of *C. deserticola*. This agrees with the reports that events of foreign DNA transferred into the cp genome are rare. The possibility of disparity between plant mitochondrial and nuclear genomes vs. cp genomes in rates of HGT is that the mitochondrion and nuclear genomes contain much more non-coding DNA than compact cp genomes. As desert plants, *H. ammodendron* and *C. deserticola* have developed an extremely specialized set of morphological, biochemical and molecular traits to adapt scare nutrients and water in the soil, such as loss of leaves and the development of haustoria in *C. deserticola*. With this feeding organ, *C. deserticola* can extract water and nutrients from the parasitized host, including the nucleic acids in the form of mRNA. It is why HGT appears to be facilitated by the direct physical association between the parasite and its host in the parasitic systems.Moreover, *C. deserticola* is a typical root parasite, meaning they are usually in contact with its host through meristems. In plants, meristems are less protected than the germlines in most multicellular animals. Therefore, the genes, which transferred to the root apical meristem, could have the opportunity to be integrated in the genome and transmitted to the next generation. In our study, either the transferred *rpo*C2 or its native copy appear to be non-functional pseudogenes in *C. deserticola*. Previous work has reported plant mtDNA pseudogenes that are transcribed and edited, so this raises the possibility that some of these genes may actually be functional. The fact is that acquiring a new gene can lead to an obvious benefit to living in that particular environment. *H. ammodendron,* which is distributed across dry deserts and salt pans, has high tolerance to osmotic and salt stress. We postulate that *C. deserticola* could not only obtain the carbohydrate, minerals and water, but also the straightforward source of useful genetic information from the neighbour already adapted to that environment. In the ‘genomic era’, future work is still needed to discover more HGT events in this pair of host and parasite by next generation sequencing, especially genes in mitochondrial and nuclear genomes. The authors would like to thank Yiyao Hu and Jiaqi Wu for their helpful suggestions. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JL YZ. Performed the experiments: XL QQ TCZ. Analyzed the data: TCZ QQ XL. Contributed reagents/materials/analysis tools: ZR JZ TY MH. Wrote the paper: TCZ QQ MJCC YZ.

# Introduction The suboccipital region is one of the most complicated anatomical areas of the human body. In 1995, Hack et al. first described the relationship between the deep suboccipital muscles and the cervical spinal dura mater. They found a connective tissue bridge named the Myodural Bridge (MDB), between the rectus capitis posterior minor (RCPmi) and the dorsal cervical spinal dura mater at the atlanto-occipital interspace. The RCPmi gave off dense connective tissue that connects with the posterior atlanto-occipital membrane, and finally merged with the dorsal cervical spinal dura mater. To date, many studies have confirmed the existence of this connective tissue bridge in humans and other mammalian animals. Moreover, many studies in the last decade have shown that the MDB originate from multiple suboccipital muscles including the rectus capitis posterior major (RCPma) and the obliquus capitis inferior (OCI). In the suboccipital region, the nuchal ligament (NL) also provides a connection between the suboccipital region and the cervical dura mater. In 2014, we observed an intrinsic fascial structure called the To Be Named Ligament (TBNL). The TBNL is a dense fibrous band that originates from the lower part of the posterior border of the NL, runs anteriosuperiorly to enter the atlanto-axial interspace and merged with the posterior cervical dura mater. It thereby forms part of the MDB. Furthermore, the TBNL is formed by either arcuate fibers or radiate fibers. In this study, we also found a second termination which originated from the RCPmi and terminated at the arcuate fiber of the TBNL. Here in this research, we conducted an extensive anatomical study about the deep suboccipital region, and we found that multiple suboccipital muscles (RCPmi, RCPma and OCI) had novel terminations on the TBNL other than the traditional bony structures. We termed these novel terminations as the “second terminations” and we additionally investigated the morphological relationship between these second terminations and the variable types of the TBNL in the deep suboccipital region of humans. # Materials and methods ## Dissection of the suboccipital muscles Thirty-five donated adult head-neck specimens (23 males, 12 females) were dissected in this study. All the specimens were obtained from the Department of Anatomy, Dalian Medical University. This study was approved by the Ethics Committee of the Body and Organs Donation Center of Dalian Medical University. None of the tissue donors were from a vulnerable population and all donors or next of kin provided written informed consent that was freely given. The specimens were preserved using a formalin-alcohol mixture. A layer-by-layer dissection was performed to expose the suboccipital muscles, the NL as well as the TBNL. The second terminations of the RCPma and OCI were observed and documented, and we subsequently cut the origin of the RCPma from the inferior nuchal line to expose the RCPmi. The second termination of the RCPmi was then observed and documented. Following these procedures, the various suboccipital muscle originating second terminations and the interrelationship with neighboring structures were observed. Photographic documentations were recorded with a Canon EOS 60D camera. ## Observation of the second terminations and the TBNL The second termination was defined as bundle of muscle fibers that originated from the RCPmi, the RCPma and the OCI, and terminated at the soft tissue structures (TBNL) instead of bony structures, e.g. the posterior tubercle of atlas or spinous process of axis. We observed and counted the existence of the second terminations in all the thirty-five specimens. The TBNL had two types of fiber arrangements, the arcuate and the radiate fibers. The arcuate fibers of the TBNL originated from the lower part (below vertebra C3) of the posterior border of the NL, ran anterosuperiorly; crossed over the spinal process of the axis and continued into the atlanto-axial interspace. The radiate fibers of the TBNL arose from the upper part of the NL (above vertebra C3), ran anteriorly and straightly into the atlanto-axial interspace, and finally attached to the cervical dura mater. ## Statistics For statistical analysis, *chi-square* test was applied (SPSS 17.0, IBM, USA). A *P* value smaller than 0.05 indicated statistically significance. No specimen was excluded in the analysis. ## Ethics statement The body and organs donation Center of Dalian Medical University is a specialized office that accepts body and organs from volunteers and donors' body for scientific research and medical education. It has a strict system of accepting donated body. These bodies and organs received by this center are by the agreement of both the donors and their immediate families. The specimens used in the experiment of this article “The Second Terminations of the Suboccipital Muscles: an Assistant Pivot for the To Be Named Ligament” by Xiao- Ying Yuan et al. were all from donations of the volunteers to the donation center and were all legally collected. The donation center allowed these specimens to be used in this basic medical research. None of the tissue donors were from a vulnerable population and all donors or next of kin provided written informed consent that was freely given. This can be verified in the website below. URL: <http://home.dlmedu.edu.cn/bodc/> # Results In the suboccipital region, the paired RCPmi and RCPma were short muscles and the RCPmi lied ventrally to the RCPma. The RCPma originated from the pcciput and the RCPmi originated from the position between the inferior nuchal line of the occiput and the foramen magnum, descended to attach at the posterior tubercle of the atlas and the spinous process of axis respectively. The OCI-originated from the transverse process of atlas and descended medially to attach at the spinous process of the axis. ## Existence and incidence of the second terminations from the suboccipital muscles In our study, we found novel terminations of the suboccipital muscles, and we termed them as the second terminations. These second terminations originated from the suboccipital muscles and terminated at the TBNL, and the overall incidence of the second terminations was 34.29% (24 out of 70 sides). In these 24 sides, the RCPma-originated second terminations were found in 20 sides and had the greatest incidence rate of 28.57%; 8 sides showed the RCPmi-originated second terminations, the incidence was 11.43%; the OCI-originated second terminations were found in 6 sides, the incidence was 8.57%, which was the lowest. The suboccipital muscles gave off muscle bundles of the second terminations near the midline and these muscle bundles terminated at the TBNL, just before the TBNL entered the posterior atlanto-axial interspace. Second terminations of the RCPmi and RCPma were either composed of one or multiple muscle bundles. The fiber property of the OCI-originated second terminations was tendinous, which was more resistant to tensile forces, compared with the muscular second termination of RCPmi and RCPma. In some cases, the second terminations originated at the same side from multiple suboccipital muscles and terminated at the TBNL. The second terminations originating from both the RCPma and the OCI were observed in 4 sides; those from both the RCPmi and the RCPma were found in 6 sides. ## The relationship between the second terminations and the types of the TBNL The TBNL was a dense fibrous band that originated from the posterior funicular portion of NL, ran anteriorly through the atlanto-axial interspace and attached to the posterior aspect of the cervical dura mater. It formed part of the MDB. The TBNL was formed by either arcuate or radiate fibers according to the point of origin of the fiber tissues. The radiate type of the TBNL was found in 50 out of 70 sides with the incidence of 71.43%. Twenty of seventy sides had the arcuate type of TBNL with an incidence of 28.57%. In 20 of 70 sides with the radiate type of the TBNL, 19 sides had the second termination: 2 sides from RCPmi only, 6 sides from the RCPma only, 2 sides from the OCI only; 6 sides had the second terminations simultaneously derived from the RCPmi and RCPma, and 3 sides had second terminations from both the RCPma and the OCI. In 50 of 70 sides with the radiate type of the TBNL, only 5 sides had the second termination: 1 side from the RCPma, 4 sides from both the RCPma and OCI. There was no existence of second termination in the other 45 sides. The relationship between the second terminations of suboccipital muscles and the types of the TBNL was statistically significant (P\<0.001). Since we conducted the gross anatomy on 35 head-neck specimens, we picked up the best photographs and put them as figures in the manuscript; see Figs for the second terminations and their neighboring structures of some other head-neck specimens. # Discussion ## The existence of second terminations of the suboccipital muscles Traditionally, the origins of RCPmi and RCPma originate from the inferior nuchal line and the occiput, and the terminations of these two suboccipital muscles are at the posterior tubercle of the atlas (RCPmi) and spinous process of axis (RCPma), respectively. The OCI originates from the transverse process of atlas and descends to the medial spinous process of the axis. In this study, we found new terminations of the suboccipital muscles of RCPmi, RCPma and OCI to the TBNL and we termed these new findings as the second terminations. Based on the results of the gross anatomy, muscular or tendinous bundles separated from the suboccipital muscles (RCPmi, RCPma, OCI), ran along with the NL in the midline and finally terminated at the TBNL. The second terminations may originate from multiple suboccipital muscles, and the overall incidence of these second terminations was 34.29%. The RCPma-originated second termination showed the highest incidence (28.57%) according to our observation. None of these findings have been reported in any previous publications. ## Relationship between the second terminations and the arcuate TBNL The TBNL is first described and defined in 2014. It is a dense fibrous band, intrinsic component and enhancement of the NL that forms part of the MDB. It enters the epidural space through the atlanto-axial interspace and terminates at the posterior cervical spinal dura mater. In this study, we found a significant correlation between the appearance of the second terminations and the types of the TBNL. In the 20 sides of the arcuate TBNL, the incidence of the second terminations of the suboccipital muscles was 95% (19/20), while in 50 sides of the radiate TBNL, the incidence of second terminations was 10% (5/50). In this study, we speculatively conclude that the second terminations usually co-exist with the arcuate TBNL. The arcuate TBNL originates from the posteroinferior part of the NL, runs anterosuperiorly crossing over the spinous process of the axis, and continues ventrally into the posterior atlanto-axial interspace. The existence of the second terminations would optimize the functional performance of the arcuate TBNL. The forces of contraction of the RCPmi, RCPma and OCI may act directly through the second terminations to maintain the TBNL, which curve around the posterior part of the spinous process of the axis, thus acting as an assisting pivot for the TBNL. Additionally, the second terminations might collaborate with the arcuate TBNL to transfer forces to the MDB via the atlanto-axial interspace. This will greatly enhance the physiological functions of the MDB. To this hypothesis, analysis of the biomechanical coordination between the second terminations of the suboccipital muscles and the arcuate TBNL will be our next research target. # Supporting information The authors would thank the Department of Anatomy, Dalian Medical University for assisting with cadavers for dissections and for their cooperation. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** HJS SBY. **Data curation:** XZ. **Formal analysis:** XYY CL. **Funding acquisition:** HJS SBY NZ. **Investigation:** XYY CL JYS QQZ XZ NNM ZHF NZ. **Methodology:** HJS SBY. **Project administration:** HJS. **Resources:** HJS. **Software:** JYS XZ CL. **Supervision:** SBY. **Validation:** SBY HJS. **Visualization:** XYY CL. **Writing – original draft:** XYY CL. **Writing – review & editing:** CL SBY XYY HJS SHH OCS.

# Introduction Allogeneic hematopoietic stem cell transplantation (allo-HSCT) following a high dose conditioning regimen has been the best treatment option for many young patients with hematological disorders. The antitumor activity of this approach is based not only on high dose chemo-radiotherapy given in the conditioning regimen but also on immune-mediated graft-versus-tumor effects. These observations are the basis of the development of allo-HSCT following nonmyeloablative conditioning, in which eradication of malignant cells depends on graft-versus-tumor effects. T-cell recovery after allo-HSCT following high-dose conditioning depends on both homeostatic peripheral expansion (HPE) of donor T cells contained in the graft, and T cell neo-production from donor hematopoietic stem cells (thymo-dependent pathway). In young patients given myeloablative allo-HSCT, most circulating T cells during the first months following HSCT are the progeny of T cells infused with the grafts , while neogeneration of T cells by the thymus plays an increasing role in reconstituting the T cell pool beyond day 100 after allo- HSCT. Since HPE allow the expansion of both NK cells and non-tolerant T cells, it is generally accepted that HPE is one of the driving force of graft-versus- tumor effects. Several studies have demonstrated that IL-7 and IL-15 are the main driving forces of HPE after allo-HSCT following high-dose conditioning,. IL-7 is a γ-common chain cytokine that is secreted by stromal cells from multiple organs including thymus, bone marrow, and lymphoid organs. IL-7 is required for human T cell development since mutations in the IL-7 receptor alpha can lead to severe combined immunodeficiency. Administration of IL-7 has been shown to dramatically increase peripheral T cell numbers, primarily through augmentation of HPE. IL-15 is another γ-common chain cytokine secreted by antigen-presenting cells, bone marrow stroma, thymic epithelium, and epithelial cells in the kidney, skin, and intestines. IL-15 plays an important role in the development and function of NK cells, and of NK/T cells, and is required for optimal proliferation of CD8⁺ T cells and for homeostatic proliferation of CD8⁺ memory T cells. While high-dose conditioning regimens typically induce a profound lymphodepletion, progressive replacement of host-derived T cells by donor- derived T cells is the rule after nonmyeloablative conditioning,. This prompted us to analyze the kinetics of IL-7 and IL-15 blood levels after allo-HSCT following a nonmyeloablative conditioning with the aim of determining whether there is a rational for boosting HPE and perhaps graft-versus-tumor effects in patients with high risk disease given grafts after nonmyeloablative conditioning by administering IL-7 and/or IL-15. # Patients and Methods ## Patients and Donors Data from 70 patients transplanted between March 2007 and April 2011 at the University of Liège were included in the study. All patients were given G-CSF- mobilized peripheral blood stem cells (PBSC) after low-dose \[2 Gy (n = 60), or 4 Gy (n = 10)\] total body irradiation (TBI)-based nonmyeloablative regimen. Twenty-three nonmyeloablative recipients who were given PBSC from HLA-mismatched unrelated donors were co-transplanted with third party mesenchymal stromal cells (MSCs) as a potential way to prevent severe GVHD. Further, 3 nonmyeloablative recipients were included in a double blind randomized study assessing the impact of MSC co-transplantation on transplantation outcomes. No patient was given in- vivo T cell depletion. ## Ethics Written informed consent was obtained from each patient to undergo allo-HSCT and to collect, store and analyze blood samples for research purposes. The Ethics Committee of the University of Liège (“Comité d’Ethique Hospitalo-Facultaire Universitaire de Liège”) approved the consent form as well as the current research study protocol (protocol \#B707201112193). ## Clinical Management The clinical management has been performed as previously reported. Chimerism levels among peripheral T-cells were generally measured with PCR-based analysis of polymorphic microsatellite regions (AmpFlSTR® Identifiler®, Applied Biosystems, Lennik, Belgium). CD3 (T-cell) selection was carried out with the RosetteSep^R human T-cell enrichment kit (StemCell Technologies, Vancouver, Canada). ## Cytokines Levels EDTA-anticoagulated plasma and serum samples were obtained before conditioning and about once time per week after transplantation until day 100. Samples were aliquoted and stored at −80°C within 3 hours after collection until measurement of cytokines. Kinetic courses of IL-7 production in plasma samples were evaluated before conditioning and approximately at days 7, 14, 28, 40, 60, 80 and 100 after allo-HSCT. IL-15 serum sample levels were assessed before conditioning and approximately at days 7, 14 and 28 after allo-HSCT. IL-7 and IL-15 levels were measured by ELISAs following the manufacturer’s protocol (High sensitivity IL-7 and IL-15 quantikine, R&D Systems, Minneapolis, MN, USA). The standard curve ranges for IL7 were 0.25 to 16 pg/mL, and the minimal detectable dose was \<0.1 pg/mL. No patient had IL-7 levels below this threshold in the current study. The standard curve ranges for IL15 were 3.9 to 250 pg/mL, and the minimal detectable dose was \<2 pg/mL. Il-15 levels were between 0 and 2 pg/mL in our study in 15 patients before transplantation, in no patient on days 7 and 14, and in 1 patient on day 28. No sample dilution was performed for IL-15 assay. For IL-7 analysis, samples were diluted twice. Patient samples whose cytokine level were out of standard curve range, were re-assessed after dilution. ## Immune Recovery Immune recovery was prospectively assessed as previously described. Briefly, patients’ peripheral white blood cells were phenotyped using 4 color flow cytometry after treatment with a red blood cell lyzing solution. The following antibodies were used: CD3-ECD (Beckman Coulter, Iotest \#A07748); CD4-V450 (Becton Dickinson Horizon \#560345); CD8-FITC (Beckman Coulter Iotest \#A07756); CD56-PC7 (Beckman Coulter Iotest \#A21692); CD45RA-PE (Dako \#R7086). The percentage of positive cells was calculated relative to total nucleated cells, after subtraction of non-specific staining. Absolute counts were obtained by multiplying the percentages of positive cells by the white blood cell counts (XE-5000 hematology analyzer, Sysmex, Kobe, Japan). Absolute lymphocytes counts (ALC) were measured directly by the XE-5000 analyzer or after microscopic review of the blood smears when the automated differential was flagged. Absolute white blood cell counts were used instead of ALC when white blood cell counts were below 150 cells ×10⁹/L. ## Statistical Analyses The Mann Whitney test was used to compare counts of lymphocyte subset and cytokine levels in patients given grafts after 2 Gy or 4 Gy TBI. The Wilcoxon matched pair test was used to compare cytokines levels before and at various time points after transplantation. Generalized linear mixed models were used to analyze factors affecting immune recovery and cytokine levels after transplantation. Factors included in the models included : (1) dose of TBI (2 Gy vs 4 Gy), MSC infusion or not, number of days after allo-HSCT, number of CD3⁺ cells transplanted, donor type (related vs unrelated), patient age, and donor age for analyses examining lymphocyte counts; (2) dose of TBI (2 Gy vs 4 Gy), MSC infusion or not, grade II–IV acute GVHD the first 100 days after transplantation, number of CD3⁺ cells transplanted, donor type (related vs unrelated), patient age, and donor age, and either IL-7 or IL-15 levels on days 7–14 (median) for analyses examining lymphocyte count increments from days 14–28 (median) to days 80–100 (median); and (3) number of days after allo-HSCT, number of CD3⁺ cells transplanted, donor type (related vs unrelated), dose of TBI (2 Gy vs 4 Gy), ALC, CRP levels, donor and patient ages, and MSC infusion or not, for analyses of cytokine levels. Incidences of acute GVHD according to the cytokines levels were assessed using cumulative incidence methods. A Cox model was constructed for determining potential factors associated with the occurrence of grade II–IV acute GVHD the first 200 days after transplantation. Factors included in the model included median day 7 and day 14 IL-7 levels, median day 7 and day 14 IL-15 levels, dose of TBI (2 Gy vs 4 Gy), donor type (related vs unrelated), female donor to male recipient versus other gender combination, MSC infusion or not, patient age, and donor age. Spearman’s correlation was used to examine the relationship between parameters. Statistical analyses were carried out with Graphpad Prism (Graphpad Software, San Diego, CA) and SAS version 9.2 for Windows (SAS Institute, Cary, NC, USA). # Results ## Immune Recovery Median ALC count on day 0 was 110 (range, 10–5440) cells/µl, demonstrating the persistence of recipient T cells at the time of transplantation. While median CD8⁺ T cell levels reached the lower limit of normal values from day 60 after transplantation, median CD4⁺ T cell (including naïve CD4⁺ T cells) remained below the lower limit of normal values the first 6 months after transplantation. No significant difference of T cell subset counts were observed between 2 Gy and 4 Gy TBI regimen. Using generalized linear mixed models taking into consideration data from day 14, 28, 40, 60, 80 and 100 for each patient, counts of CD3⁺ T cells (P\<0.001), CD8⁺ T cells (P\<0.001), CD4⁺ T cells (P = 0.024), NK cells (P\<0.001) and NK/T cells (P\<0.001) increased over time but not those of naïve CD4⁺ T cells (P = 0.13). Further, high numbers of transplanted CD3⁺ T cells were associated with higher counts CD3⁺ T cells (P = 0.009), CD8⁺ T cells (P = 0.003), and CD4⁺ T cells (P = 0.0099), while high donor age was associated with lower counts of CD3⁺ T cells (P = 0.04), CD4⁺ T cells (P = 0.05), and naïve CD4⁺ T cells (P = 0.021). There was no significant association between MSC administration and lymphocyte subset counts after transplantation. ## IL-7 Plasma Levels Median IL-7 plasma levels remained below 6 pg/L throughout the first 100 days (the upper limit of normal range being 9.2 pg/mL (Quantikine© HS catalog number HS750)), although IL-7 plasma levels were significantly higher on days 7 (5.1 pg/mL, P = 0.002), 14 (5.2 pg/mL, P\<0.0001) and 28 (5.1 pg/mL, P = 0.03) (but not thereafter) than before transplantation (median value of 3.8 pg/mL). Using generalized linear mixed models, low number of transplanted CD3⁺ T cells (P = 0.001), low ALC level the day of IL-7 assessment (P\<0.0001), high donor age (P = 0.003), having received PBSC from unrelated donors (p = 0.006), and high level of CRP the day of IL-7 assessment (*P* = 0.033) were associated with high levels of IL-7. ## Il-15 Serum Levels Median IL-15 serum levels were significantly higher on days 7 (12.5 pg/mL, P\<0.001), 14 (10.5 pg/mL, P\<0.001) and 28 (6.2 pg/mL, P\<0.001) than before transplantation (median value of 2.4 pg/mL). IL-15 levels on day 7 and 14 were significantly higher in 4 Gy than 2 Gy TBI. Using generalized linear mixed models, conditioning with 4 versus 2 Gy TBI (P = 0.002), having received PBSC from unrelated donors (P = 0.001), low ALC level the day of IL-15 assessment (P\<0.001), and high level of CRP the day of IL-15 assessment (P = 0.006) were each associated with high IL-15 levels on days 7 and 14 after allo-HSCT. ## Correlation between IL-7 and IL-15 Levels and Lymphocyte Subset Counts on Days 14 or 28 after allo-HSCT Day 14 IL-7 levels inversely correlated with day 14 counts of CD3⁺ T cells (R = −0.46, P = 0.002;), CD8⁺ T cells (R = −0.41, P = 0.006), CD4⁺ T cells (R = −0.44, P = 0.004), and memory CD4⁺ T cells (R = −0.45, P = 0.003), but not with counts of naïve CD4⁺ T cells (R = −0.28, P = 0.07), NK/T cells (R = −0.04, P = 0.8) nor NK cells (R = −0.14, P = 0.4). There was a weak association between day 14 IL-7 and IL-15 levels (R = 0.27, P = 0.049). Further, day 14 IL-15 levels correlated with day 14 counts of NK cells (R = −0.32, P = 0.039;) and of NK/T cells (R = −0.32, P = 0.037), but not with those of other T cell subsets. Day 28 IL-7 levels inversely correlated with day 28 counts of CD3⁺ T cells (R = −0.47, P\<0.001;), CD8⁺ T cells (R = −0.41, P = 0.002), CD4⁺ T cells (R = −0.39, P = 0.002), naïve CD4⁺ T cells (R = −0.40, P = 0.002), and memory CD4⁺ T cells (R = −0.38, P = 0.004), but not with counts of NK/T cells (R = −0.17, P = 0.2), nor NK cells (R = −0.02, P = 0.9), nor with day-28 donor T cell chimerism levels (R = 0.0, P = 0.95). There was no significant association either between day 28 IL-7 and IL-15 levels (R = 0.07, P = 0.6). Further, day 28 IL-15 levels correlated with day 28 counts of NK cells (R = −0.32, P = 0.015;) but not with those of T cell subsets, nor with day-28 donor T cell chimerism levels (R = 0.14, P = 0.29). To further assess the potential association between early IL-7 or IL-15 levels on immune recovery, we analysed whether there was a relationship between median cytokine levels on days 7 and 14 and the difference of lymphocyte subset counts between days 80–100 (median) and days 14–28 (median). Interestingly, in multivariate analyses, early IL-7 levels did not correlate with any lymphocyte subset increment from days 14–28 to day 80–100 after transplantation, while high IL-15 levels early after transplantation correlated with a lower increment of NK cells over time (P = 0.04). ## IL-7 and IL-15 Levels did not Predict for Subsequent Acute GVHD The 180-day cumulative incidence of grade II–IV acute GVHD was 30%, a rate similar to what has been observed by other group of investigators using similar conditioning regimen. As shown in the, no statistically significant association between cytokines levels on days 7 or 14 after transplantation and occurrence of grade II–IV acute GVHD were observed. Specifically, the 180-day cumulative incidence of grade II–IV acute GVHD was 29% in patients with day 7 IL-7 levels\>median (5.1 pg/mL) versus 20% in patients with day 7 IL-7 levels ≤ median (P = 0.38). Similarly, the 180-day cumulative incidence of grade II–IV acute GVHD was 19% in patients with day 14 IL-7 levels\>median (5.2 pg/mL) versus 37% in patients with day 14 IL-7 levels ≤ median (P = 0.18). The 180-day cumulative incidence of grade II–IV acute GVHD was 24% in patients with day 7 IL-15 levels\>median (12.5 pg/mL) versus 28% in patients with day 7 IL-15 levels ≤ median (P = 0.8). Similarly, the 180-day cumulative incidence of grade II–IV acute GVHD was 25% in patients with day 14 IL-15 levels\>median (10.5 pg/mL) versus 33% in patients with day 14 IL-15 levels ≤ median (P = 0.8). Finally, in a multivariate Cox model, neither median IL-7 levels (P = 0.17 with a trend for an inverse correlation) on days 7–14 nor median IL-15 levels (P = 0.21 with a trend for a positive correlation) on days 7–14 correlated with occurrence of grade II–IV acute GVHD the first 200 days after transplantation. Similarly, the use of MSC was not associated with decreased incidence of grade II–IV acute GVHD. This could be explained by the fact that all 23 MSC recipients versus of 9 of the remaining 49 patients (18%) received PBSC from HLA-mismatched donors. None of the other factors tested (dose of TBI, donor type, female donor to male recipient versus other gender combination, patient age, and donor age) were significantly associated with the incidence of grade II–IV acute GVHD in the current study. ## IL-15 Levels did not Predict for Subsequent Relapse/Progression Given that a previous publication showed an association between high IL-15 levels and low risk of relapse/progression, we compared the cumulative incidence of relapse/progression according to IL-15 levels 14 days after transplantation in our cohort of patients. The 6-month and 1-year cumulative incidences of relapse/progression were 29% and 32%, respectively, in patients with day 14 IL-15 levels\>median (10.5 pg/mL) versus 37% and 46%, respectively, in patients with day 14 IL-15 levels ≤ median (P = 0.57). # Discussion Following allo-HSCT, eradication of residual tumor cells depends in part (in case of high-dose conditioning) or mainly (in case of nonmyeloablative conditioning) on immune-mediated graft-versus-tumor effects. Prior studies have demonstrated a close relationship between T cell reconstitution and graft- versus-tumor effects after allo-HSCT,. Given that HPE allows the expansion of potentially alloreactive T cell clones, it has been generally accepted that HPE plays a major role in graft-versus-tumor effects, but could also cause or favor acute GVHD. This prompted us to investigate the kinetics of IL-7 and IL-15 levels in a cohort of 70 patients given grafts after truly nonmyeloablative conditioning. First, patients given grafts after nonmyeloablative conditioning had only a modest (\<2 fold) increase of IL-7 levels after transplantation (contrarily to what we observed in another cohort of patients given grafts after myeloablative conditioning), that persisted up to day 21. This is probably due to the fact that nonmyeloablative patients experienced relatively mild lymphopenia (and thus continue to consume the IL-7 produced by stromal cells) as demonstrated by the persistence of median ALC counts of 110 cells/µL at the time of transplantation. Although the first T cell chimerism assessment in current patient was usually around day 28 after HSCT, a prior study analyzing data from patients given similar conditioning regimen demonstrated that a median of 50 CD3⁺ T cells of recipient origin/µL persisted on day 14 after HSCT. Further, as observed by other groups of investigators, there was a strong inverse correlation between IL-7 levels and absolute lymphocyte counts, as well as a strong inverse correlation between IL-7 levels and T cell subsets on days 14 and 28 after transplantation. Other factors associated with IL-7 levels included high CRP levels, and low numbers of transplanted T cells. Levels of IL-7 in current nonmyeloablative recipients where lower to what was observed by Thiant *et al.* in a cohort of 45 patients given grafts after fludarabine +2 Gy TBI (n = 18) or more intense but still reduced-intensity conditioning (n = 27), and where much lower than what was observed by Dean *et al.* in patients given grafts after sequential chemotherapy followed by a chemotherapy/fludarabine- based reduced-intensity conditioning. This apparent discrepancy is probably explained the fact than median ALC counts on day 0 were 110 (range, 10–5440) cells/µl in current patient versus 0 (range, 0–322) cells/µL in the Dean *et al.* study, while median counts of CD3⁺ T cells were 0 (range, 0–1900) cells/µL at the time of transplantation in Thiant *et al.* study. Il-15 levels were lower in nonmyeloablative patients conditioned with 2 Gy TBI than in those conditioned with 4 Gy TBI, demonstrating that the release of IL-15 was proportional to the intensity of the conditioning regimen. As observed by Thiant *et al.*, there was a correlation between IL-7 and IL-15 levels on day 14 (but not on day 28) after transplantation, and an inverse correlation between IL-15 levels and NK cell counts. Other factors affecting IL-15 levels included high CRP levels. Several observations demonstrate that immune recovery depended mainly on HPE the first year after nonmyeloablative conditioning regimen in current patients. Firstly, there was a strong correlation between the number of infused T cells and high counts of CD4⁺ and CD8⁺ T cells, as previously observed. Secondly, thymic function was minimal during the first 100 days after allo-HSCT given that levels of naïve CD4⁺ T cells did not significantly increase the first 100 days after transplantation despite that some naïve T cells can undergo HPE and keep their naïve phenotype. Third, there was a correlation between high donor age and low counts of CD3⁺ T cells (P = 0.04), CD4⁺ T cells (P = 0.05), and naïve CD4⁺ T cells (P = 0.021), as previously observed in patients given grafts after nonmyeloablative conditioning. Despite that, we failed to find any significant association between IL-7 and/or IL-15 levels early after transplantation and increment of T cell subset counts from days 14–28 to day 80–100, even after adjusting for potentially confounding cofactors. A number of previous studies have demonstrated that high levels of IL-7 , and/or IL-15, early after transplantation correlated with subsequent occurrence of grade II–IV acute GVHD, while others study failed to find such an association. The largest study including data from 153 consecutive allogeneic transplant recipients given grafts after high-dose conditioning and ATG observed no correlation between IL-7 levels early after transplantation and acute GVHD, while, interestingly, there was an inverse correlation between IL-15 levels early after transplantation and grade II–IV acute GVHD. Further, a recent study demonstrated that administration of IL-7 after allogeneic T cell-depleted transplantation in humans did not increase acute GVHD. In the current study, we did not observe any association between levels of IL-7 or IL-15 early after allo-HSCT and grade II–IV acute GVHD. The same was true after adjusting the analyses for potentially confounding cofactors. Differences in postgrafting immunosuppression might be the cause for these apparent discrepancies between studies. As example, it has been shown that tacrolimus (given in patients included in the current study) decreased T cell proliferation induced by IL-7, and tacrolimus levels were kept high in our patients the first weeks after transplantation (median 18.6, 16.4, 14.9 and 14.3 µg/L on days 0, 7, 14 and 21 after transplantation, respectively) probably explaining the low relatively incidence of acute GVHD observed. In summary, these data suggest that IL-7 and IL-15 levels remain relatively low after nonmyeloablative transplantation, and that IL-7 and IL-15 levels early after nonmyeloablative transplantation do not predict for acute GVHD. The authors are grateful to O. Dengis for excellent technical support. [^1]: The authors have declared that no competing interests exist. [^2]: Taking care of the patients: YB FB. Conceived and designed the experiments: MF YB FB. Performed the experiments: MdB MF MH MPM AG. Analyzed the data: MdB MF LS FB. Contributed reagents/materials/analysis tools: MdB MF YB FB. Wrote the paper: MdB FB.

# Introduction Polycyclic aromatic hydrocarbons (PAHs) are the major component of coke oven emissions produced during incomplete combustion of natural or synthetic fuels. Recent studies, including those from our lab, found that increased exposure to PAHs was associated with reduced cardiac autonomic function, assessed by heart rate variability (HRV), which is considered one of the main pathophysiologic pathways for air pollution-mediated adverse cardiac events. However, the mechanisms through which the cardiac autonomic system responds to PAHs exposure have not been sufficiently understood. Systemic inflammation may be one of the potential mechanisms linking air pollution including PAHs to cardiovascular autonomic decline. It has been reported that elevated levels of inflammation biomarkers, especially that of C-reactive protein (CRP) and interleukin-6 (IL-6), are inversely associated with various HRV indices in apparently normal adults – and in patients with coronary heart disease (CHD). However, to our knowledge, no studies have evaluated the relationship between inflammation and HRV in occupational population exposed to different levels of PAHs. Heat shock protein 70 (Hsp70), one of the main HSP family members, has a dual role as chaperone and cytokine by inducing pro-inflammatory cytokines such as IL-6 production. A number of studies have shown that circulating Hsp70 play a crucial role in the pathogenesis and/or prognosis of cardiovascular diseases. We previously found that exposure to PAHs, as reflected by urinary 1-hydroxypyrene (1-OHP), resulted in a dose-dependent increase in levels of plasma Hsp70, suggesting that high Hsp70 levels may serve as a danger marker among workers exposure to PAHs. We thus hypothesized that circulating Hsp70 may be associated with cardiac autonomic function in response to occupational PAHs exposure. To test the above hypothesis, we examined the association of urinary monohydroxylated PAHs (OH-PAHs) with plasma IL-6 and Hsp70, and the effects of IL-6 and Hsp70 on HRV in 800 workers exposed to different levels of PAHs. # Materials and Methods ## Ethics Statement The study was approved by the Ethics and Human Subject Committee of Tongji Medical College and informed written consent was obtained from each subject. ## Study Subjects The cross-sectional survey was conducted throughout the workday (Monday to Friday) from Oct to Nov 2009 in Wuhan (Hubei, China). Eight hundred healthy workers who had worked for at least one year at different sites, including at the top, side, bottom, or adjacent workplaces of coke ovens, and at the offices, in the same coke oven plant, were included in this study. Subjects who had suffered from cardiovascular diseases, serious medical conditions affecting HRV or chronic inflammation in the previous 3 months were excluded from the study. In the morning, workers visited the occupational health examination center. 5 mL of venous blood and 20 mL of urine samples were collected at the beginning (pre- shift) of workshift from the workers who were on the day shift, and the end (post-shift) of workshift from those who were on the night shift. All blood and urine samples were divided into aliquots and frozen at −80°C until laboratory analysis. Questionnaires, physical examination and Holter monitoring followed to be completed. Information on demographic variables, occupational history, medications, history of disease, and lifestyle habits including smoking, passive smoking, alcohol consumption, physical activity, and diet were collected using standardized occupational questionnaires. ## Determination of Urinary PAHs Metabolites We measured 12 urinary PAHs metabolites \[pyrene metabolite: 1-hydroxypyrene (1-OHP); naphthalene metabolites: 1-hydroxynaphthalene (1-OHNa), 2-OHNa; fluorene metabolites: 2-hydroxyfluorene (2-OHFlu), 9-OHFlu; phenanthrene metabolites: 1-hydroxyphenanthrene (1-OHPh), 2-OHPh, 3-OHPh, 4-OHPh, 9-OHPh; chrysene metabolite: 6-hydroxychrysene (6-OHChr); and benzo\[a\]pyrene metabolite: 3-hydroxybenzo\[a\]pyrene (3-OHBaP)\] by gas chromatography-mass spectrometry (GC/MS, Agilent 6890N+5975B, Agilent Technologies Inc., Santa Clara, CA, USA) as previously described. Briefly, each 3.0 mL urine sample was extracted three times to elevate the detection rate. The set of the standard curve was operated about every 100 samples and re-run about 10% of the total samples were used as quality control. The identification and quantification of urinary PAHs metabolites were based on retention time, mass-to-charge ratio, and peak area using a linear regression curve obtained from separate internal standard solutions. The limits of detection (LOD) for the urinary PAHs metabolites were in the range 0.1–1.4 μg/L; default values were replaced with 50% of the LOD. Valid urinary PAHs metabolite concentrations were calibrated by levels of urinary creatinine and calculated as μg/mmol creatinine. Because 6-OHChr and 3-OHBaP were below the limits of quantification, we only analyzed 10 metabolites of PAHs. ## Detection of IL-6 and Hsp70 in Plasma IL-6 and Hsp70 levels were measured in EDTA-plasma using IL-6 and Hsp70 enzyme- linked immunosorbent assay (ELISA) kits (NeoBioscience Technology Company, China and Stressgen Bioreagents Company, Victoria, BC, Canada). The Hsp70 assay only detects inducible Hsp70 and does not detect other Hsp70 family members such as constitutive Hsp70, Grp78, DnaK (*Escherichia coli*), or Hsp71 (*Mycobacterium tuberculosis*). The inter-assay and intra-assay coefficients of variation of the assays were \<10%. ## Measurements of HRV The measurement of HRV was conducted between 8∶00 am and 11∶00 am and measurement methods have been described previously. After at least 5-min rest, each participant was seated comfortably on a chair and was fitted with a 3-channel digital Holter monitor (Lifecard CF; Del Mar Reynolds Medical, Inc., Irvine, USA) with a 1024 samples/second sampling rate for 10 minutes. We cleaned the participant’s skin with an alcohol wipe and abraded slightly to keep good lead contacts, and placed the separate electrode in the position according to the instruction and technical manual of Lifecard CF. All of the HRV indices were calculated on 5-min epoch in the entire recording. Only heart rates between 40 and 100 beats per minute were submitted to analyses. We selected five consecutive minutes of ECG reading in the statistical analysis without atria and ventricular premature beats and flutter. The HRV spectrum was computed with a fast Fourier transform method and analyzed in both time and frequency domains. The measured time domain parameters include: a) SDNN (standard deviation of NN intervals in milliseconds), is an estimation of total HRV power; b) RMSSD (the root mean of square of successive differences between adjacent normal NN intervals in milliseconds), reflects the activities of parasympathetic nervous system. The frequency-domain variables include: a) low-frequency (LF msec²; 0.04–0.15Hz) may represent the combination of both parasympathetic and sympathetic activity of heart rate; b) high-frequency (HF msec²; 0.15–0.4Hz) shows the actions of the parasympathetic modulation of heart rate; c) total power (TP msec²; approximately≤0.4Hz) is a mixture of total variability of heart rate. ## Statistical Analyses We assessed the normality of all variables with the one-sample K-S test. Normal distributions of HRV indices and values of all urinary PAHs metabolites and plasma IL-6 and Hsp70 were obtained by natural logarithmic transformation. Because there is no well-established cutoff point for defining high versus low total OH-PAHs exposure, while 33^rd and 67^th percentiles or the interquartile range (75^th vs. 25^th percentile) was used as cutoffs in previous studies, we dichotomized total OH-PAHs at the 33th percentile as either low or high. Plasma IL-6 and Hsp70 were considered as categorical variables at different cut-off points. Baseline characteristics between urinary PAHs metabolites categories were compared using *t*-tests for continuous variables and chi-square tests for categorical variables. The trend for urinary PAHs metabolites levels in the four environmental exposure groups were tested by using simple linear regression. Pearson partial correlation coefficients (r) were computed to examine the associations among ln-transformed individual PAHs metabolites and ΣOH-PAHs. The trend for IL-6 and Hsp70 among quartiles of PAHs metabolites and the trend for HRV among quartiles of the IL-6 and Hsp70 were tested by multivariate linear regression models. Moreover, effects of IL-6 and Hsp70 on HRV stratified by PAHs metabolites were estimated using general linear models. Finally, to analyze potential interactions between IL-6 and PAHs metabolites as well as Hsp70 and PAHs metabolites, interaction product terms were included in the model. Potential confounders, including age, sex, length of work, smoking status, alcohol use, BMI, physical activity, working sites (top-oven, side and bottom-oven, adjunct-oven, and at the offices), and workshift (pre-shift and post-shift) for acute exposure – over the course of a day and weekday (Monday to Friday) for intermediate-term exposure – over the course of a week were used in the analysis. All statistical tests were derived from two-sided analyses and *P*\<0.05 was considered as statistical significance. All analyses were conducted using SPSS (version 12.0). # Results ## Characteristics of Study Population in Relation to Urinary Total OH-PAHs Levels General characteristics of the workers classified by urinary total OH-PAHs levels are presented in. The distribution according to age, length of work, current smoker, current alcohol drinker, BMI and physical activity in the two groups was similar (all *P*\>0.05), but differed in sex and workshift (pre- shift) (*P*\<0.001). The median level of IL-6 was lower in the low total OH-PAHs group compared with the high total OH-PAHs group (1.86 pg/ml vs. 2.22 pg/ml), whereas Hsp70 level was similar. Averages of the five HRV indices were lower among those with the high total OH-PAHs levels, although SDNN and TP were not statistically significant. ## Urinary Levels of PAHs Metabolites at Different Working Sites As our previous report, we defined office workers as the control group, and workers at adjunct-oven, and at side and bottom-oven, and top-oven as low, intermediate and high exposure groups, respectively. The concentrations of ΣOH- PAHs and nine urinary PAHs metabolites with the exception of 4-OHPh were significantly elevated with increasing environmental PAHs exposure levels (all *P*_trend\<0.01). ## Correlations of Urinary PAHs Metabolites Almost all individual PAHs metabolites were significantly correlated with each other, except 3-OHPh and 9-OHFlu as well as 2-OHPh, 3-OHPh and 4-OHPh. The correlation coefficients between individual PAHs metabolites and ΣOH-PAHs ranged from 0.262 to 0.851. The strongest correlations between individual PAHs metabolites and ΣOH-PAHs was 1-OHNa (r = 0.851), followed by 1-OHP and 2-OHNa (r = 0.836 and 0.772), the lowest correlations was 4-OHPh (r = 0.262). ## Effects of PAHs Metabolites on IL-6 and Hsp70 as Well as HRV After controlling for potential confounders included age, sex, length of work, smoking status, alcohol use, BMI, physical activity, working sites, workshift and weekday, increasing quartiles of four urinary OH-PAHs (1-OHNa, 1-OHPh, 9-OHPh and 1-OHP) were significantly associated, in a dose-responsive manner, with increased IL-6 (*P*_trend = 0.004, 0.008, 0.0001 and 0.001, respectively). However, no significant dose-response relationship between PAHs metabolites and Hsp70 was found after multivariate adjustment. In addition, increasing quartiles of 2-OHNa and 1-OHPh were significantly associated with decreased HF (*P*_trend = 0.028 and 0.012, respectively), but no significant association between other metabolites and HRV indices was observed. ## Dose-dependent Decrease in HRV Associated with Increased IL-6 As shown in, after adjustment for covariates, increasing quartiles of IL-6 levels were significantly associated with decreases in TP and LF (*P*_trend = 0.014 and 0.006). Compared to the lowest quartile, individuals with the highest quartile of IL-6 were significantly associated with 2.8% and 4.8% reduction in TP and LF. ## Association of IL-6 with HRV by PAHs Metabolites Given that HRV had significant association with IL-6 levels, we next examined whether the relationship of IL-6 with HRV differed in low- versus high-PAHs metabolites. As shown in, after adjustment for confounders, we found that elevated IL-6 levels were significantly associated with decreases in TP and LF in the high-PAHs metabolites groups (ΣOH-PAHs, 1-OHP and 2-OHPh, all *P*_trend\<0.05), but not in the low-PAHs metabolites groups. Significant interactions between IL-6 and ΣOH-PAHs on TP (*P*_interaction = 0.017), and between IL-6 and 1-OHP, 2-OHPh and ΣOH- PAHs on LF (*P*_interaction = 0.027, 0.044 and 0.001, respectively) were also observed. ## Association of Hsp70 with HRV by PAHs Metabolites No significant associations between Hsp70 and HRV were found in total population after adjusting for confounders (data not shown). When we examined the relationship of Hsp70 and HRV stratified by PAHs metabolites, we found that there was a significant increase in SDNN, TP and LF associated with Hsp70 when multiple PAHs metabolites were low (including 1-OHP, 1-OHNa, 2-OHPh, 9-OHPh and ΣOH-PAHs, all *P*_trend *\<*0.05), but not when they were high. Tests for interaction between Hsp70 and multiple PAHs metabolites were significant on SDNN, TP and LF (all *P*_interaction\<0.05). # Discussion Increasing epidemiological evidences have indicated significant relationship between inflammatory markers, most commonly CRP and IL-6, and various HRV indices in CHD patients, type I diabetes, and in apparently normal adults.However, until now, the association between inflammatory markers and HRV in workers exposed to different levels of PAHs has been less clear. We hypothesized that inflammation and stress response might play an important role in cardiac autonomic dysfunction induced by PAHs and potential dose-response relationships among them. In the present study, we used IL-6 as a marker to test this hypothesis, because that IL-6 is a good indicator of cytokine cascade activation, accurately reflects the inflammatory status, and is stable as its half-life is longer than other proinflammatory cytokines. We found a significant dose-dependent relationship between four urinary OH-PAHs and IL-6; and IL-6 was significantly associated, in a dose-responsive manner, with decreased HRV, suggesting that certain OH-PAHs rather than others may be a significant contributor to the autonomic dysfunction through an inflammation mechanism. What is new and intriguing about our results is there were significant interactions between some PAHs metabolites and IL-6 on HRV. The link between IL-6 and HRV was only significant in workers with high-PAHs metabolites, but not in those with low-PAHs metabolites, indicating that high-PAHs exposure amplified the association of inflammation and HRV. PAHs have been documented as one of the most important components of coke oven emission. Measurement of the urinary OH-PAHs metabolites, as internal markers, is an important way of assessing exposure to PAHs, since it takes into account all absorption routes of exogenous compounds. Due to relatively short half- lives of PAHs, the information provided by biomonitoring of urinary OH-PAH is limited to recent exposure. Most potential health effects are likely associated with exposure over time, thus considering the exposure time is essential for data interpretation, especially for chemicals with short half-lives. Li et al evaluated the variability of nine urinary OH-PAHs metabolites over time in non- occupationally exposed subjects and found that the within-day variance exceeded the between-day variance, and the between-subject variance out-weighted the between-day variance for certain metabolites. However, our results did not substantially change even after adjusting for indicator variables of exposure time such as workshift for potential acute exposure (over the course of a day) and weekday intermediate-term exposure (over the course of a week), which strengthened the robustness of the observed association. Studies have shown that exposure to high levels PAHs was positively associated with the prevalence, mortality or morbidity of cardiovascular disease, and one mechanism by which PAHs may play a role in the development of cardiovascular disease is through inflammation. Thus, high-PAHs exposure has been speculated to contribute to increasing systemic inflammation and subsequent reduced autonomic function, which might explain the significant relationship between IL-6 and HRV in high-PAHs exposure. The hypothesis is supported by studies that have identified that elevated PAHs levels were associated with increased IL-6 or CRP in non-occupational exposure population, and systemic inflammatory activity was associated with a decrease in HRV. However, previous studies have only focused on one aspect of the relationships among PAHs, inflammatory biomarkers or HRV; no studies so far have examined all of them together, thus we cannot compare directly with that of other published studies. Luttmann-Gibson et al showed that systemic inflammation modified susceptibility to air pollution-associated autonomic dysfunction in elderly subjects, but the cross-sectional nature of our study precludes any inference about the causal direction of the association between inflammation and HRV related to PAHs exposure. Together with the results of prior studies, these findings make it clear that significant decreased HRV is indeed associated with inflammation among workers with high-PAHs exposure, implicating that pollution can lead to autonomic dysfunction through inflammation. Intracellular Hsp70 has been found to provide a spectrum of protection against any of a variety of stresses, preventing damage of cells and tissues, while plasma extracellular Hsp70 (eHsp70) has been shown to have protective or deleterious effects in cardiovascular events,. Some studies have shown that increased Hsp70 level was associated with low risk of cardiovascular diseases. In contrast, several studies including some from our lab suggested that eHsp70 might possibly play a harmful role in the pathogenesis and progression of cardiovascular diseases. Although autonomic dysfunction have been associated with and may even precede various forms of cardiovascular diseases, no study has reported the effect of plasma Hsp70 on HRV. In this study, we did not find significant association between Hsp70 and various HRV indices in total population. However, we observed significant and consistent interactions between multiple PAHs metabolites and Hsp70 on HRV. There was a dose-response increase in SDNN, LF and TP in relation to elevated Hsp70 levels when PAHs metabolites was low but not when it was high, suggesting that increased plasma Hsp70 may have played a protective role in autonomic function when workers were exposed to relatively low levels of PAHs. Our previous studies showed that exposure to coke oven emission resulted in a dose-dependent increase in levels of plasma Hsp70, but we did not explore whether there was a reverse relationship between PAHs levels and Hsp70 when workers were exposed to low coke oven emission as the relatively small sample size (n = 101). We speculate that circulating Hsp70 levels can be in the range which has been shown to exert protective effect against light damage when exposure to low-PAHs levels, and subsequent increase of HRV. Several limitations of the present study merit consideration. First, the cross- sectional design is not able to prove the existence of a causal relationship whether inflammation leads to decreased HRV or vice versa. Second, monitoring of the 5-min HRV represents only short-term cardiac autonomic nerve modulation and the definitive clinical significance should be elucidated. Third, we did not adjust for multiple testing because of a high correlation both in various HRV indices and in PAHs metabolites. Over adjustment for multiple comparisons may increase the type II error and render truly important associations insignificant. Lastly, we could not rule out residual and unmeasured confounders, although we adjusted for a wide range of confounding factors. # Conclusions In this study, we observed the dose-response relationships between urinary OH- PAHs and IL-6 as well as IL-6 and HRV indices. In particular, a prominent effect of IL-6 on HRV was found in the high-PAHs metabolites groups, whereas high Hsp70 levels may have a protective effect on HRV in low-PAHs metabolites groups. However, further researches are needed to confirm these findings in prospective studies and elucidate the possible mechanisms of these associations. # Supporting Information We are particularly grateful to all volunteers for participating in the present study and staffs in the Institute of Industrial Health, Wuhan Iron & Steel Corporation. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JY RZ X. Zhang. Performed the experiments: JY RZ XH YF LY X. Zhu QD. Analyzed the data: JY RZ X. Zhang. Contributed reagents/materials/analysis tools: TW X. Zhang. Wrote the paper: JY RZ X. Zhang. Revised the manuscript: JY RZ TW X. Zhang.

# Introduction Cellulose and chitin, both β-1,4-linked linear saccharides composed of glucose (Glc) or N-acetylglucosamine (GlcNAc), respectively, are the two most abundant biomasses distributed in the plant and animal kingdoms, respectively. The biodegradation of these saccharides proceeds via the same path, endo-enzymes first degrade higher degree polymerized saccharides into oligosaccharides and then exo-enzymes degrade oligosaccharides into monosaccharides. Cellulase (EC 3.2.1.4) and chitinase (EC 3.2.1.14) are the endo-splitting enzymes, and β-glucosidase (EC 3.2.1.21) and β-N-acetyl-D-hexosaminidase (EC 3.2.1.52) are the exo-splitting enzymes required for cellulose and chitin degradation, respectively. The similarity between these two biomolecules hints at a convergent evolution between the degradation enzymes from plants and chitin- containing animals. The recent crystal structural information of exo-splitting enzymes provides evidence of this linkage. Rice (*Oryza sativa*) β-glucosidase BGlu1 (Os3BGlu7), which exhibits high activity toward cellooligosaccharides, belongs to glycosyl hydrolase family 1 according to the CAZy database. The catalysis proceeds via a double displacement mechanism by which two acidic residues act as nucleophile and acid/base catalyst, respectively. The crystal structure of BGlu1 revealed that BGlu1 possesses a classic (β/α)₈-barrel catalytic domain constituting a substrate binding pocket with subsites for binding both the leaving Glc (-1 subsite) and the other cellooligosaccharide residues (+1 subsite, +2 subsite and so on). The amino acid residues constituting the subsites for binding the cellooligosaccharide residues, in particular the +1 subsite, are thus determinants for both substrate affinity and specificity. Three residues, Trp³⁵⁸, Ile¹⁷⁹ and Tyr¹³¹, are found to be crucial for the +1 Glc binding. Trp³⁵⁸ stacks with the +1 Glc through its indolyl group, Ile¹⁷⁹ forms a hydrophobic interaction with the +1 Glc through its isopropyl group, and Tyr¹³¹ forms a hydrogen bond with the C3-OH of the +1 Glc through its phenolic hydroxyl group. In this way, the +1 and -1 sugars are stabilized and exist around a 90° dihedral angle adjacent to each other, yielding a conformation where the glycosidic bond between them becomes more susceptible to attack by the catalytic residues. The structural alignment reveals that the active pocket architecture of OfHex1 and BGlu1 is very similar although their overall topology as well as amino acid sequence shares very little similarity. Insect OfHex1 from the Asian corn borer (*Ostrinia furnacalis*) is a β-N-acetyl-D-hexosaminidase specialized for chitin degradation during molting and metamorphosis. This enzyme attracts much attention because of its potential as a species-specific target for developing eco-friendly pesticides. OfHex1 shows high activity toward β-1,4-linked chitooligosaccharides but is not able to hydrolyse β-1,2-linked GlcNAc from *N*-glycans, thus exhibiting very high substrate specificity. According to the CAZy database, β-N-acetyl-D-hexosaminidase belongs to family 20, and uses a substrate-assisted double displacement mechanism by which the 2-acetamido group in the substrate GlcNAc, instead of the acidic residue in a β-glucosidase, acts as the nucleophile. OfHex1 also contains a -1 subsite and +1 subsite for sugar moiety binding, and the three corresponding amino acid residues found in BGlu1 for binding the +1 GlcNAc are found in OfHex1(Val³²⁷, Glu³²⁸ and Trp⁴⁹⁰). We thus hypothesized that both cellulolytic and chitinolytic enzymes may possess a similar mechanism to increase their affinities and specificities toward physiological substrates. In this study, a comparative structural investigation between BGlu1 and OfHex1 was performed. Although the two enzymes exhibit no similarity in either their overall structure or amino acid composition of their active pockets, crucial conserved residues were discovered by structure-based sequence alignment among cellulolytic and chitinolytic enzymes from different species. The significance of these residues was investigated by site-directed mutagenesis, biochemical characterization and crystal structure analysis. This research may provide a basis for the development of pesticides that are suitable for use in cases where plant protection is a priority. # Materials and Methods ## Structure Comparison and Multiple Sequence Alignment Structure comparison was performed by PyMOL (Schrödinger LLC, Portland, OR) and structural figures were also prepared by PyMOL. Multiple sequence alignments were performed using PROMALS3D and the alignment figure was prepared with ESPript. ## Preparation of Enzymes Mutations of OfHex1 (V327G, E328Q and E328A) were made by In-Fusion™ Advantage PCR Cloning Kit (TaKaRa) using the following primers: V327G (forward primer, 5′-GGTGAGCCCCCATGCGGTCAGCTC-3′; reverse primer, 5′-CGCATGGGGGCTCACCGCAGTATGATTTCC-3′); E328Q (forward primer: 5′- CAGCCCCCATGCGGTCAGCTCA -3′, reverse primer: 5′- ACCGCATGGGGGCTGCACGCAG -3′); E328A (forward primer: 5′-GCGCCCCCATGCGGTCAGCTCA-3′, reverse primer: 5′-ACCGCATGGGGGCGCCACGCAG-3′); W490A (forward primer: 5′-GCTTGTTCTCCTTACATCGGATGGCAG-3′, reverse primer: 5′-GATGTAAGGAGAACAAGCGTTGTTACCAGC-3′). Full-length PCR products were cloned into the expression vector pPIC9 (Invitrogen). Then the expression vector plasmids were linearized by *Pme*I (New England Biolabs) and transformed into *Pichia Pastoris* GS115 cells by electroporation. After growing on a RDB plate at 30°C for 48 h, the positive clones were selected by PCR. The positive clones were cultured in BMMY broth at 30°C for 144 h, and methanol (1% of the total volume) was added every 24 h. Wild-type and mutant OfHex1 were purified from the culture supernatant by ammonium sulfate precipitation (65% saturation), affinity chromatography on a HisTrap™ HP column (5 ml, GE Healthcare) followed by anion exchange chromatography on a Mono Q™ 5/50 GL column (1 ml, GE Healthcare). The purity of the mutants was analyzed by SDS- PAGE. The β-N-acetyl-D-hexosaminidase from *Streptomyces plicatus* (SpHex) was purchased from New England Biolabs. The β-N-acetyl-D-hexosaminidase from *Serratia marcescens* (SmChb) was expressed and purified according to the reported method. ## Enzymatic Assay The enzymatic properties of wild-type and mutant forms of OfHex1, SmChb and SpHex were determined using GlcNAcβ1,4GlcNAc \[(GlcNAc)₂, Sigma\] and GlcNAcβ1,2Man (Dextra Laboratories). For the substrate (GlcNAc)₂, the reaction mixtures contained 0.04, 0.08, 0.12, 0.16 and 0.2 mM of substrate and an appropriate amount of enzyme in 50 µl of Britton-Robinson’s wide range buffer \[pH 7.0 for OfHex1, mutants of OfHex1(V327G, E328A, E328Q and W490A) and SmChb, pH 4.0 for SpHex\]. For GlcNAcβ1,2Man, the reaction mixtures contained 0.03, 0.05, 0.1, 0.2 and 0.4 mM of substrate and an appropriate amount of enzyme in 50 µl of Britton-Robinson’s wide range buffer at the same pHs mentioned above. After incubation at 25°C for a specific period, 10 µl of the reaction was immediately analysed using a TSKgel amide-80 column (4.6×250 mm, Tosoh) at 25°C. The reaction velocity was quantified by comparing the peak area of the product GlcNAc to a standard curve of GlcNAc at known concentrations. In both cases, the substrate consumption was limited to less than 10%. The *K*_m and *k*_cat values were also calculated by linear regression of the data using Lineweaver-Burk plots. TMG-chitotriomycin was kindly provided by Professor Biao Yu (Institute of Organic Chemistry, Chinese Academy of Science). The inhibitory kinetics of TMG- chitotriomycin for the wild-type and mutant forms of OfHex1, SmChb and SpHex were studied using 4MU-β-GlcNAc (4-methylumbelliferone-*N*-acetyl-β-D- glucosaminide, Sigma) as a substrate. The *K*_i values were calculated by linear regression of the data using Dixon plots. ## Crystallization and Data Collection Mutant OfHex1 (E328A) was incubated with excessive TMG-chitotriomycin (5-fold the amount of protein), and then concentrated to ∼7 mg/ml. Vapour diffusion crystallization experiments were set up at 4°C by mixing 1 µl of protein and 1 µl of mother liquor consisting of 100 mM HEPES (pH 7.4), 200 mM MgCl₂, and 33% PEG400. Diffraction data of mutant OfHex1 (E328A)-TMG- chitotriomycin complex was collected using an in-house Rigaku Micromax-007 HF (Rigaku Raxis IV++ Image Plate, wavelength 1.5418 Å, at 180 K), and processed using the Crystal Clear software package. ## Determination and Refinement of Structures The structure of mutant OfHex1 (E328A)-TMG-chitotriomycin complex was solved by molecular replacement with Molrep using the structure of OfHex1-TMG- chitotriomycin complex (PDB accession number: 3NSN) as the search model. There was one monomer in the asymmetric unit for the structure. Structure refinement was achieved by Refmac5 and CNS. Model building was performed in Coot. The quality of the final model was checked by PROCHECK). All structural figures were prepared by PyMOL. # Results ## Structure Comparisons between BGlu1-cellotetraose and OfHex1-TMG-Chitotriomycin Structural comparisons between BGlu1-cellotetraose (PDB accession number: 3F5J) and OfHex1-TMG-chitotriomycin (PDB accession number: 3NSN) were performed to reveal structural differences and similarities. BGlu1 is a monomeric enzyme with only one catalytic domain while OfHex1 is a dimeric enzyme with an N-terminal zincin-like domain and a C-terminal catalytic domain in each subunit. Although both BGlu1 and OfHex1 possess non-classical catalytic (β/α)₈-barrels, the overall topology of the two enzymes’ catalytic domain is quite different as supported by the fact that the C^α atoms of the catalytic domain of BGlu1 (residues 1 to 476) superimposed on those of OfHex1 (residues 207 to 594) gives a RMSD value of 3.8 Å (based on the overlap of 275 C^α atoms) by a Dali pairwise comparison. In addition, BGlu1 has four distinguished loops (Loop A, residues 25–65; Loop B, residues 177–206; Loop C, residues 314–363; Loop D, residues 387–403) while OfHex1, in accordance with BGlu1, has two corresponding loops including *L*_314–335 (residues 314–335) and *L*_478–496 (residues 478–496). By using the pair-fitting alignment using the atoms in the ring of the -1 sugar (5 carbon atoms and 1 oxygen atom) as fitting pairs, the spatial arrangements of the residues comprising the -1 subsites of BGlu1 and OfHex1 are different. Except for the His¹³⁰/Asn¹⁷⁵/Glu¹⁷⁶ in BGlu1 and His³⁰³/Asp³⁶⁷/Glu³⁶⁸ in OfHex1 constituting the catalytic residue clusters, only two pairs of residues were found to be spatially conserved and with similar functions. Trp⁴³³ in BGlu1 and Trp⁵²⁴ in OfHex1 are located in the bottoms of the -1 subsites and stack with the non-reducing end sugar rings while Gln²⁹ in BGlu1 and Glu⁵²⁶ in OfHex1 make hydrogen bonds with C4-OH/C6-OH and C4-OH of -1 sugars, respectively. However, a striking similarity in the architectures was discovered in the +1 sugar binding sites. Pair-fitting alignment was performed by using the atoms in the ring of the +1 sugar (5 carbon atoms and 1 oxygen atom) as fitting pairs. The spatial locations of interactions for binding the +1 sugar are conserved between the two enzymes. It is interesting to note that the residues involved in the interactions are located in distinguished loops. Tyr¹³¹, Ile¹⁷⁹ and Trp³⁵⁸ in BGlu1 are located in the loop between the β3 strand and the α3 helix, Loop B (between the β4 strand and the α4 helix) and Loop C (between the β6 strand and the α6 helix), respectively. Accordingly, both Val³²⁷ and Glu³²⁸ in OfHex1 are located in *L*_314–335 while Trp⁴⁹⁰ is located in *L*_478–496. Trp³⁵⁸, which stacks with the +1 sugar from the down side in BGlu1, corresponds to Trp⁴⁹⁰ in OfHex1. Ile¹⁷⁹, which forms a hydrophobic interaction with the +1 sugar in BGlu1, corresponds to Val³²⁷ in OfHex1. Moreover, Tyr¹³¹, which forms a hydrogen bond with the C3-OH of the +1 sugar via a water molecule in BGlu1, corresponds to Glu³²⁸ in OfHex1 which directly interacts with the C3-OH of the +1 sugar. The carboxyl oxygen atom of Glu³²⁸ in OfHex1 is positioned in the same location as a water molecule in BGlu1 which intermediates hydrogen bonding formation between the +1 sugar and BGlu1. One may notice that BGlu1 Glu⁴⁴⁰ also makes a water-mediated hydrogen bond with the C3-OH of the +1 sugar, and is almost close enough to make a direct hydrogen bond (3.4 Å in the 3F5K structure), although it also H-bonds to O4 and O6 of the -1 subsite glycosyl residue. However, OfHex1 does not contain such a residue in the active site. Thus, the spatial conservation of the residues Val³²⁷/Glu³²⁸/Trp⁴⁹⁰ in OfHex1 and Ile¹⁷⁹/Tyr¹³¹/Trp³⁵⁸ in BGlu1 is believed to be functionally important. ## Conservation of Residues in the +1 Subsite among Chitinolytic Enzymes Structure-based multiple sequence alignments were performed to determine whether the Val³²⁷, Glu³²⁸ and Trp⁴⁹⁰ residues in OfHex1 are spatially conserved at the +1 subsite among all chitinolytic β-N- acetyl-D-hexosaminidases. The sequences selected for analysis were taken from representative species including: 1) a choanoflagellate (*Monosiga brevicollis*), a close living relative of animal ancestors ; 2) a branchiopod (*Daphnia pulex*), a close relative of the ancestors of higher crustaceans and insects ; 3) a crustacean (*Fenneropenaeus chinensis*); 4) an insect (*Drosophila melanogaster*); 5) a fungus (*Candida albicans*) and 6) a plant (*Arabidopsis thaliana*). OfHex1 (PDB accession number: 3NSN) and two bacterial β-N-acetyl-D-hexosaminidases, SmChb from *S. marcescens* (PDB accession number: 1QBB) and SpHex from *S. plicatus* (PDB accession number: 1HP5) , were selected as input structures. The results of the amino acid sequence alignment indicated that, although the length of the loop varies, Trp⁴⁹⁰ in the loop (*L*_478–496 in OfHex1) is highly conserved in all of the selected chitinolytic β-N-acetyl-D- hexosaminidases. This demonstrated that the stacking interaction between the +1 sugar and the conserved Trp is vital for the function of the chitinolytic β-N- acetyl-D-hexosaminidases. Compared to the loop *L*_478–496, the loop *L*_314–335 varies in both length and amino acid composition, a function of the biodiversity of the species chosen for the sequence alignment. The residues in the *L*_314–335, namely Val³²⁷ and Glu³²⁸ in OfHex1, are also conserved in most β-N-acetyl-D- hexosaminidases. Although the enzymes from *S. marcescens* and *D. pulex* possess a Gln instead of Glu³²⁸, this may be considered a conservative substitution in terms of the similar possibilities for hydrogen bond formation. Val³²⁷ was replaced by a smaller residue (Ala or Ser) in the β-N- acetyl-D-hexosaminidases from plant (*A. thaliana*) and the original species (*M. brevicollis* and *D. pulex*), suggesting the species-specific conservation of Val³²⁷ in bacteria, fungi, shrimps and insects. Moreover, the absence of Val³²⁷ in enzymes from lower species (*M. brevicollis* and *D. pulex*) implies the existence of a separate evolutionary branch of chitinolytic β-N-acetyl-D-hexosaminidases. The structural comparison revealed that the direction of the loop (*L*_314–335 of OfHex1 and SmChb) is clockwise while that of SpHex is counter-clockwise. Although the primary structures of *L*_314–335s and *L*_478–496s in these enzymes are totally different, both Val³²⁷ and Trp⁴⁹⁰ in OfHex1 can be found in SpHex and SmChb (Val²⁷⁶ and Trp⁴⁰⁸ in SpHex, Val⁴⁹³ and Trp⁶⁸⁵ in SmChb). For Glu³²⁸ in OfHex1, SmChb possesses Gln⁴⁹⁴ but SpHex does not possess any comparable residues in this position. Glu³²⁸ in OfHex1 functions through a hydrogen bond with the C3-OH of the +1 sugar. Gln⁴⁹⁴ in SmChb, however, functions the same way as the Tyr¹³¹ in BGlu1, that is, by interacting with the C3-OH of the +1 sugar via a water-mediated hydrogen-bond. ## Significance of the Residues Conserved in the +1 Subsite To study the role of Val³²⁷, Glu³²⁸ and Trp⁴⁹⁰ during substrate binding and catalysis, the enzymatic properties of both the wild-type and mutants of OfHex1 were compared by using β1,4-conjugated (GlcNAc)₂ and β1,2-conjugated GlcNAcβ1,2Man. Using (GlcNAc)₂ as the substrate, the catalytic efficiency after site-directed mutagenesis was determined. The *k*_cat values of the mutants, V327G, E328Q, E328A and W490A were found to be lower than the wild- type, ranging from 12% (E328Q) to 26% (V327G). The *K*_m values of V327G, E328Q, E328A and W490A mutants were 1-, 2.5-, 8.6- and 12.7-fold higher, respectively, than that of the wild-type. Based on the *k*_cat/*K*_m values, the catalytic efficiencies of the mutants for (GlcNAc)₂ were much lower (from 1.3- to 14.7-fold) than the wild-type. Furthermore, using 4MU-β-GlcNAc as the substrate, the *K*_i values of TMG-chitotriomycin for the wild-type and mutants V327G, E328Q, E328A and W490A were determined. The *K*_i values for TMG-chitotriomycin toward the mutants, V327G, E328Q, E328A and W490A, were increased by 1.1-, 1.6-, 42- and 2277-fold, respectively, compared to the wild- type. It is worthy to note that the mutation of Glu³²⁸ to Ala instead of Gln seriously impaired the ability of OfHex1 to bind the substrate. These results demonstrated that Glu³²⁸ and Trp⁴⁹⁰ but not Val³²⁷ are vital for substrate binding. To better understand the function of Val³²⁷, additional substrates were studied. As OfHex1 has evolved to degrade β1,4-glycosidic bonds in linear chitooligosaccharides, it is surprising to observe that the mutation of Val³²⁷ to Gly results in OfHex1 having the remarkable capability to hydrolyze the substrate GlcNAcβ1,2Man. The catalytic efficiency of V327G will be further explored in the discussion section. ## Crystal Structure Analysis of E328A Complexed with TMG-chitotriomycin To reveal the structural basis behind the biochemical data, crystal structural information is needed. Since we have obtained the structure of OfHex1 V327G mutant previously, in this study, the crystallization of E328A in complex with TMG-chitotriomycin was performed and the complex structure was resolved to 2.5 Å. The statistics of data collection and structure refinement are summarized in. The coordinates were deposited in the Protein Data Bank with accession number 3VTR. The structure showed well defined unbiased map for TMG-chitotriomycin (except GlcNAcIII at the reducing end) and side chains of Asp³⁶⁷ and Glu³⁶⁸. Structure alignment of the OfHex1-TMG-chitotriomycin complex with the E328A-TMG- chitotriomycin complex was performed using PyMOL. According to the alignment, the mutation of Glu³²⁸ to Ala resulted in obvious conformation changes. The catalytic residues, Glu³⁶⁸ and Asp³⁶⁷ are rotated about 180° and 90° compared to those in the OfHex1-TMG-chitotriomycin complex but with similar conformations as those in the *apo*-structure of OfHex1. The rotation of Glu³⁶⁸ led to the collapse of the two important hydrogen bond networks (Asp²⁴⁹-His³⁰³-Glu³⁶⁸ and Glu³²⁸-H₂O (I)-Glu³⁶⁸). In addition, the mutation of Glu³²⁸ to Ala increased the volume of the active pocket from 646.8 ų to 758.3 ų as measured with CASTp, suggesting an important role of Glu³²⁸ in constructing the active pocket. The conformation of TMG-chitotriomycin, in particular the GlcNAcII moiety, was found to be very different from that observed in the wild-type. Although the sugar rings of GlcNAcII in both complexes are in the ¹*C*₄ conformation and are superimposable, the C5- hydroxymethyl groups of GlcNAcII orient to different directions. In addition, the conformation of GlcNAcIII in OfHex1 is ^O*S*₂, but the electron density was not clear enough to identify its conformation in OfHex1 (E328A). However, there is no obvious minus peaks found around GlcNAcIII with the current conformation (^3,O*B*) in the difference map. Therefore, this unclearness of the electron density map can be considered as the result of the flexibility of GlcNAcIII. Furthermore, mutation of Glu³²⁸ to Ala leads to changes in polar interactions between the wild type and the mutant when complexed with TMG- chitotriomycin. The most obvious change is the absence of short polar interaction between Glu³²⁸ and the C3-OH of GlcNAcI of TMG- chitotriomycin. Interestingly, a water molecule was found to occupy the position of the side chain of Glu³²⁸. Several other changes in intermolecular polar interactions were also observed in the E328A-TMG-chitotriomycin complex when compared to the wild type. Firstly, the distance between the nitrogen atom of the 2-acetamido group (GlcNAcI) and the oxygen atom of the carbonyl group (main chain of Val³²⁷) increased from 2.8 Å to 3.2 Å. The interaction between the nitrogen atom of the 2-acetamido group (GlcNAcII) and the oxygen atom of the carbonyl group (main chain of Trp⁴⁹⁰) is missing, but instead, C6-OH of GlcNAcII forms a hydrogen bond with the main chain of Val³²⁷. # Discussion Enzymes capable of degrading cellulose and chitin are undoubtedly significant because these saccharides are the two most abundant forms of biomass in nature. Since cellulose and chitin are similar in both their chemical structure and catabolic procedures, it can be deduced that the associated hydrolases would share some commonalities, for example, their substrate binding mechanism or glycosidic bond preference. A structural comparison may provide new clues to uncover the evolutionary divergence of these enzymes between plants and animals as well as provide supporting information for the use of their associated biomass in industrial or medical applications. However, to date, little work has been performed with regard to comparing cellulolytic enzymes with chitinolytic enzymes. Herein, we investigated the structures of a plant cellulolytic enzyme (BGlu1) and an insect chitinolytic enzyme (OfHex1) in an attempt to find their similarities and differences. A structural comparison revealed that both GH1 BGlu1 and GH20 OfHex1 possess three residues essential for binding the +1 sugar of substrates –. Site-directed mutagenesis, enzyme kinetics as well as crystal structure determination of OfHex1 revealed that Glu³²⁸ and Val³²⁷ are both essential residues for determining glycosidic bond specificity and the +1 sugar binding, respectively. ## BGlu1Tyr¹³¹ vs. OfHex1Glu³²⁸ Both Tyr¹³¹ (BGlu1) and Glu³²⁸ (OfHex1) residues form hydrogen bonds with the C3-OH of the +1 sugar. Tyr¹³¹ (BGlu1), the residue conserved in most plant cellulolytic β-glucosidases , interacts with the +1 sugar in BGlu1 via a water molecule. Accordingly, OfHex1 Glu³²⁸, the residue conserved or conservatively replaced (e.g., by Gln) in chitinolytic β-N-acetyl-D-hexosaminidases, directly interacts with the +1 sugar in OfHex1. Like BGlu1 Tyr¹³¹, the conservatively replaced residue, Gln⁴⁹⁴, in SmChb also interacts with the +1 sugar via a water molecule. Site-directed mutagenesis indicated the change of Glu³²⁸ to Ala impaired OfHex1’s affinity toward the substrate, (GlcNAc)₂, by 8-fold and the inhibitory activity of TMG-chitotriomycin by 42-fold ( &). However, the mutation of Glu³²⁸ to Gln only caused a slight impairment in binding ( &). Thus, we deduced that the impairment is most likely caused by the absence of a strong polar interaction (2.67 Å) between Glu³²⁸ and GlcNAcI. Moreover, using (GlcNAc)₂ as the substrate, the *K*_m values of SmChb (with Gln⁴⁹⁴ corresponding to Glu³²⁸ of OfHex1) is in accordance with that of E328Q, and the *K*_m values of SpHex (no residue at the corresponding spatial location) is in accordance with that of E328A. These results further confirmed the involvement of Glu³²⁸ in substrate binding of OfHex1. Taken all together, since all the chitinolytic enzymes including SmChb, SpHex and OfHex1 contain the conserved Val and Trp residues at their +1 sugar binding sites, the difference in their affinities for (GlcNAc)₂ is most likely because of the presence of Glu³²⁸ (in OfHex1). The appearance of a conserved Glu at the +1 sugar binding sites might be the result of positive evolution. ## BGlu1 Ile¹⁷⁹ vs. OfHex1 Val³²⁷ Both Ile¹⁷⁹ (BGlu1) and Val³²⁷ (OfHex1) function at the +1 subsites by sandwiching the +1 sugar together with the residue Trp³⁵⁸ (BGlu1) and Trp⁴⁹⁰ (OfHex1), respectively. Sequence alignment indicated that the BGlu1 Ile¹⁷⁹ is replaced by Val in the plant cellulolytic β-glucosidase BGQ60 from *Hordeum vulgare*. Accordingly, OfHex1 Val³²⁷ was conserved in most chitinolytic β-N-acetyl-D- hexosaminidases, although there are several exceptions from *A. thaliana*, *M. brevicollis* and *D. pulex*, the corresponding residue of which is replaced by Ser or Ala. As indicated by the *K*_m values for cellobiose, BGQ60 has much higher binding affinity at +1 subsite than BGlu1. But the mutation of Ile¹⁷⁹ to Val does not improve BGlu1’s affinity for cellobiose, indicating Ile¹⁷⁹ is not essential for the cellobiose binding. Accordingly, mutation of Val³²⁷ to Gly did not impair OfHex1’s affinity for the substrate \[(GlcNAc)₂\] and the inhibitor (TMG-chitotriomycin) ( &). These results lead us to believe that OfHex1 Val³²⁷ may not be so crucial for substrate binding. However, a dramatic conformational change in Val³²⁷ is observed after OfHex1 binds TMG-chitotriomycin, namely the isopropyl group of Val³²⁷ moves from a vertical position to a parallel position around the indolyl plane of Trp⁴⁹⁰. As a result, the entrance size of the active pocket is enlarged from 7.25 Å to 8.26 Å. Thus, we assumed that this conformation change may suggest that Val³²⁷ plays some other essential role. Based on the *k*_cat/*K*_m values, although V327G hydrolyses GlcNAcβ1,2Man 3623-fold slower than (GlcNAc)₂, it is still 9% faster than StrH. StrH is a bacterial β-N-acetyl-D-hexosaminidase from *Streptococcus pneumonia* that is believed to be able to specifically hydrolyse β1,2-linked GlcNAc substrates. According to the structural comparison of OfHex1 wild-type, V327G and StrH (GH20A-GlcNAcβ1,2Man complex), OfHex1 has a narrow active site pocket which constrains the +1 GlcNAc sugar via hydrophobic and polar interactions with Val³²⁷, Glu³²⁸ and Trp⁴⁹⁰. However, StrH (GH20A) has a shallow active pocket in which the +1 Man is almost solvent exposed. Furthermore, as shown in, the conformations of both the +1 mannose in the StrH complex and the +1 GlcNAc in the OfHex1 complex are ¹*C*₄ chairs, but the two superimposed sugar ring is separated by a 35.4° dihedral angle. Thus, if GlcNAcβ1,2Man binds to the active pocket of OfHex1, steric hindrance might be encountered between the C3-OH and C6-OH of the +1 Man and the isopropyl group of Val³²⁷. This steric hindrance is removed by the mutation of Val³²⁷ to Gly. Thus, V327G is conferred with the ability to catalyse the hydrolysis of GlcNAcβ1,2Man. On the basis of the above findings, Val³²⁷ was speculated to have a role in the β1,4 glycosidic bond specificity of chitinolytic β-N-acetyl-D- hexosaminidases. In summary, two conserved residues in chitinolytic β-N-acetyl-D-hexosaminidases with totally different functions were discovered by comparative structural analysis between the insect chitinolytic β-N-acetyl-D-hexosaminidase OfHex1 and the plant cellulolytic β-glucosidase BGlu1. OfHex1Glu³²⁸ is vital for the binding of chitooligosaccharides while OfHex1Val³²⁷ is responsible for the glycosidic bond specificity. Together with the previously determined role of Trp⁴⁹⁰ in binding chitooligosaccharides, the functions of all the three conserved residues comprising the +1 subsite of chitinolytic β-N-acetyl-D-hexosaminidases have been revealed. # Supporting Information The authors thank Professor Biao Yu (Institute of Organic Chemistry, Chinese Academy of Science) for providing TMG-chitotriomycin. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: TL QY XS. Performed the experiments: TL LC WC LL. Analyzed the data: TL YZ QY WZ JZ. Wrote the paper: TL QY.

# Introduction The concept of synthesizing simple gene units to realize a desired function or to reproduce a known function is new in biological systems. After confirmation of the unit's desired functional behavior, a large assembly of such units can be organized to perform complex biological functions. This is like engineering small integrated chips to build a computer to derive a targeted function. In efforts towards engineering biological functions, a *repressilator* was first demonstrated as a synthetic genetic clock expressed in *Escherchia coli* producing oscillations in expressed protein concentrations. A mathematical model based on standard chemical kinetics was also proposed that predicted the observed oscillations. The dynamics of coupled synthetic genetic networks (SGNs) was also investigated – theoretically to understand the generation of synchronous rhythm in an assembly of *repressilator*s via quorum sensing type interaction. Quorum sensing is a form of exchanging information that a bacterial colony uses to develop a common rhythm. This quorum sensing type of indirect coupling is set-up between the SGN cells through diffusion of auto-inducing small molecules in a common medium. When the feedback via auto-inducing agents inside a SGN cell is reinforcing, the coupled dynamics show a state of in-phase synchrony, whereas when the feedback is repulsive, then the coupled dynamics show various possible states –: in-phase and anti-phase synchrony, inhomogeneous limit cycles, inhomogeneous steady states, and homogeneous steady states. Recently, in a biological experiment, evidence of in-phase synchronized quorum of genetic clock units was found. However, more complex features, as chaos, antiphase, and multistability in synchronous rhythm of coupled genetic clocks are yet to be observed experimentally. Mathematical models are always a very useful tool to predict complex behaviors of dynamical systems using numerical simulations. Experimental verification of rich multistability requires an accurate knowledge of the model parameters which is often very challenging in biological experiments. An alternative experimental approach using electronic analogs of the SGN was undertaken – to confirm the numerical results and to search for possible coupled dynamics. Although it is difficult to simulate the biological experiment exactly in a circuit, the advantage of an electronic SGN is accessibility of system parameters and their controllability that allows a systematic exploration of known and predictable dynamics. Earlier – electronic circuit analogs of SGN displayed oscillations with 120° phase shifts between the oscillating variables, qualitatively in agreement with the genetic *repressilator*, however, the multistability of coupled SGNs was missing since access to and control of the system parameters was lacking. In this paper, an electronic analog of the SGN is specifically designed to derive more accurate kinetic parameters of the *repressilator*. The goal is to control the parameters and thereby to realize desired sets of various kinetic parameters used in the simulations of the mathematical model. As a result, the circuit shows agreement between the measurements and the numerical predictions. The building block for the electronic *repressilator* is a circuit model for a single negative regulatory gene. This circuit shall be useful in a variety of other SGN investigations in addition to the *repressilator*. Designing electronic circuits of SGN also has the purpose of reverse engineering where knowledge of the biological experiments can be utilized for new technology and applications. # Methods ## Genetic Network Repressilator The structure of the *repressilator* consists of three repressive genes connected in a loop, with each gene producing repressor to the subsequent gene. The genes (*i* = 1,2,3) each produce their own mRNA, which translate the repressor protein. Gene 1′s repressor inhibits transcription of gene 2′s mRNA, 2′s repressor inhibits 3′s mRNA, and 3′s repressor inhibits 1′s mRNA. Taking into account standard chemical kinetics for production, degradation and inhibition, a dynamical system model of 6 first order differential equations was used for the mRNA and the protein concentrations. An electronic analog of the *repressilator* was also proposed earlier, where they used three voltages as the variables, thus reducing the above model to a set of three differential equations. However, these circuits did not simulate the kinetic parameters of the *repressilator* model. A reduced three variable model is also used here but special care is taken to retain the parameters for the chemical kinetics as in the original model. The reduced genetic network (RGN) *repressilator* is defined as,where *i* = 1, 2, 3 for three genes, and the loop is closed by the condition *x*₀ = *x*₃. The product of the *i^th* gene is *x*_i, and its production is inhibited by *x_i-*₁. The parameter *α* accounts for the maximum transcription rate in the absence of an inhibitor, *β* is the decay rate of protein degradation and *n* is the Hill coefficient for inhibition. The Hill function is commonly used to account for sigmoidal binding kinetics. In this RGN model, there is no distinction between the mRNA and the transcribed repressor protein. We find that by this reduction, the fundamental features of the SGN model are not affected. ## Description of the RGN Circuit The building block for the RGN *repressilator* circuit is the circuit for a single negative regulatory gene shown in. The desired dynamics are given by (1), where *x_i* corresponds to the output of the circuit and *x_i*_-1 is the input. The goal is to use a circuit which accounts for the kinetics of gene inhibition, production of repressor, and degradation of repressor. The gene's product is analogous to the charge coming from the collector of the transistor and the rate of production is the transistor current. The concentration of product *x_i* is proportional to the voltage *V_i* across the capacitor, and the rate of decay is the current through *R_C*. The kinetics of gene inhibition is determined by the circuitry that couples the input voltage *V_i*_-1 (the inhibitor) to the voltage at the base of the transistor. ### Circuit Analysis Here the circuit parameters are determined that correspond to particular values of kinetic parameters *α*, *β_i*, and *n* in (1). The dynamical equation for voltage *V_i* iswhere *I_t*(*V_i*_-1) is the transistor's collector current and *V_i*_-1 is the variable input voltage. The equations are expressed in dimensionless form using dimensionless time *t*/*τ* where the time-scale is chosen by *τ*  =  *R_CC*₀. The capacitor value is then *C*_i  =  *C*₀/*β_i* and (2) becomeswhere the dot denotes time derivative in dimensionless time *τ*. In order for the circuit to model the gene kinetics it is desired that *I_t*(*V_i*_-1) approximates the Hill function,where *x_i*_-1 =  *V_i*_-1/*V_th* and *V_th* represents an equilibrium constant for binding of repressor *x_i*_-1 to the gene's DNA. *I*_max is the maximum current through the transistor corresponding to gene transcription in the absence of an inhibitor. The production is half-maximal, *I_t*  =  *I*_max/2, when *V_i*_-1  =  *V_th*. Dividing both sides of (3) by *V_th*, the *α* and *β_i* are given by *α*  =  *I*_max*R_C*/*V_th* and *β_i*  =  *C*₀/*C_i*. Note that *I*_max is not due to saturation of the transistor since the voltage *I*_max*R_C* is chosen so that the emitter-collector voltage across the transistor does not reach zero. Instead *I*_max is due to saturation of the op-amp as discussed below. Next we determine how the Hill coefficient *n* relates to the circuit parameters. The Hill function behavior consists of a transition of *V_i* from one value to another one as *V_i*_-1 increases, with the transition occurring in the region around *V_i*_-1  =  *V_th* as shown in. In addition the slope in the transition region is not constant. The circuit using the two op- amps U₁ and U₂ approximates this behavior by saturating the op-amp output and by using different gains in the transition region, a larger gain for *V_i*_-1 \< *V_th* and smaller for *V_i*_-1 \> *V_th*. Op-amp U₁ is configured as a subtraction amplifier with output *G*₁(*V*_i-1 − *V*_th)  =  *G*₁Δ*V* where *G*₁ is negative. Op-amp U₂ has different gains *G*₊₂ and *G*_-2 depending on the sign of Δ*V*. Taking saturation of the outputs into account gives the voltage at the output of U₂, where *V_±sat* are the saturation levels. The next step is to consider how *G*Δ*V* controls the transistor current. The voltage drop across *R*_b1 iswhere the forward bias voltage drop across the diode is 0.6 V and *V_CC* is the supply voltage, +5 V. The diode in series with *R*_b1 compensates for the transistor's emitter- base voltage drop, so that the voltage across *R*_b1 is approximately the same as the voltage across *R*_E. The current *I_t*(*V*_i-1) through *R*_E and the transistor is therefore (6) divided by *R*_E. The maximum current *I*_max occurs when the output of U₂ is saturated at *G*Δ*V*  =  *V_-sat*, giving For comparison with the Hill function it is useful to express *I_t* in terms of *I_max* and the normalized input voltage *x*_i-1,where Δ*x_i*_-1  =  (*x*_i-1−1). In order to approximate the Hill function the overall gain *G*₁*G*_-2 for *V*_i-1\<*V_th* is chosen such that slope *dI_t*/*dx*_i-1 of the transistor current matches the slope of the Hill function at *x*_i-1 =  1 (*V*_i-1 = *V*_th). At *x*_i-1 = 1, the output *G*Δ*V* is not saturated, so *G*  =  *G*₁*G*_-2. Equating the slopes gives the condition relating Hill coefficient *n* to the overall gain *G*₁*G*_-2, Choosing the gain *G*₊₂ (for *V*_i-1\>*V_th*) to be less than *G*_-2 improves the transistor current's approximation to the Hill function. We find that choosing *G*₊₂ ≅0.3*G*_-2 works well for a range of parameter values *α*, *β_i*, and *n*. shows the Hill function (red dashed) and the predicted (green solid) and measured (blue dots) transistor current for *n* = 3.75. ### Model Parameters, Circuit Parameters, and Design Considerations Given a circuit it is useful to be able to easily determine the corresponding model parameters. From the previous section *α*, *β_i*, and *n*, are expressed in terms of circuit parameters by As an example, in the gain for op-amp U₁ is *G*₁ = −6.8 and gain for op-amp U₂ is *G*_-2 =  −22 for non-saturated overall gain *G*₁*G*_-2 = 150. For *V*_i-1 \> *V*_th the gain *G*₊₂ is approximately −6.9. For *V*_th = 50 mV, *C*₀ = 1 µf, *I*_max = 3 mA, supply *V_CC* = 5 V, and LF412 op-amp saturation *V_−sat* = −3.5 V, the resulting model parameters are *α* = 60, *β_i* =  1, and *n* = 3.8. It is also useful to be able to determine circuit parameters that achieve a desired set of model parameters. Starting with (10), *I*_max*R_C* must be far enough below the supply *V*_CC so that the emitter-collector voltage of the transistor never reaches zero. For *V*_CC = 5 V, *I*_max*R_C* is chosen to be around 3 volts and the emitter-collector voltage never gets less than about 1volt so that the transistor never goes into saturation. The choice of *I*_max has some freedom. For *I*_max =  3 mA, this determines *R_C* = 1 kΩ and *R_E* = 330 Ω in order to get voltage drops of 1, 1, and 3 volts across *R_E*, the transistor, and *R_C*, respectively, when *I*_t  =  *I*_max. The remaining free circuit parameter to adjust for a desired value of *α* is *V*_th. Rearranging (10) gives For example, for *I*_max*R_C* = 3 V, a value of *α*  = 100 is obtained using *V*_th ≅30 mV. The capacitor value *C*_i to use for a desired *β_i* is easily given by (11) as The third circuit parameter is the overall gain *G*₁G_-2 which is determined by *n* and *α*. Rearranging (12) gives For *V_CC* = 5 V and *V_−sat* = −3.5 V, then *V_CC* – 0.6− *V_−sat* = 7.9 V, and with *I*_max*R_C* = 3 V, then *G*₁*G*_-2≈2*nα*/3. For example, if the desired parameter values are *α*  = 100 and *n* = 4, then the required overall gain is *G*₁*G*_-2≈267 which can be split as desired between *G*₁ and *G*_-2. Thus, the circuit parameters that are adjusted to obtain a desired set of model parameters are *V_th*, *C*_i, and *G*₁*G*_-2. A careful selection of the op-amp is important for good approximation of the Hill function by the transistor current. The op-amp must be able to recover satisfactorily from saturation of its output. The circuit in is tested with *V*_th set to zero and with *C_i* = 0. Results for the LF412 are shown in. The input voltage *V*_i-1 (blue line) is a triangle wave from a signal generator and the measured outputs are *G*Δ*V* (red line) from the output of U₂ and the final output voltage *V*_i (green line). The red curve shows that the LF412 saturates at *V_+sat*  =  +4.5 V and at *V_-sat*  =  −3.5 V when using a ±5 V supply, and that the circuit makes the transition from high gain to low gain when *V*_i-1 goes from negative to positive corresponding to *x_i*_-1 surpassing one. The green line shows the expected inhibitory response with respect to input *V*_i-1, and the maximum value of *I*_max*R*_C = 3 V. ### *Repressilator* Circuit The electronic *repressilator* circuit consists of three negative regulatory gene circuits connected in an inhibitory loop as shown in. The triangle symbol contains the 2 op-amps, transistor, and circuitry which determine parameters *α* and *n*. # Results and Discussion The model parameters *α*, *β_i*, and *n* are now ably determined by circuit parameters. Our results of circuit measurements and numerical predictions are shown for three cases: (1) identical genes, *β*-ratio  = 1∶1∶1; (2) gene *i* = 1 with faster decay, *β*-ratio  = 3∶1∶1; (3) gene *i*  = 1 with slower decay, *β*-ratio  = 0.3∶1∶1. Gene products are the normalized voltages *x*₁ (blue), *x*₂ (red), and *x*₃ (green). Circuit measurements are solid lines, numerical predictions are dashed lines. The circuit parameters *V*_th and overall gain *G*₁*G*_-2 were varied using (13) and (15) in order to set *α* and *n*. *V_th* varied from 30 mV to 120 mV, and *G*₁*G*_-2 from 73 to 220 corresponding to *α* = 25 to 100 and *n* = 3.0 to 6.6. The figures show results for (*α, n*)  =  (50, 6.6) and (100, 3.3). Other sets of parameters produced results with similar agreement of measurements and numerical predictions. The time constant was *τ*  =  *R_CC*₀  =  (1 kΩ)(1 μf)  = 1 ms. It follows that *C*_i =  (1 µf)/*β_i*. shows the *repressilator* dynamics for *β*-ratio = 1∶1∶1 where *C*_i = 1 µf for each capacitor. The three state variables have the same shape, but with 120° phase shift. Numerical predictions are in close agreement with the measurements. For the dynamics in gene *i* = 1 has its capacitor reduced to 0.33 µf, so it has *β*₁ = 3. Thus the *β* -ratio for the genes in the *repressilator* circuit is 3∶1∶1. Increasing the gene product's decay rate causes larger oscillations for the product, reduced oscillations for the gene's inhibitor, and an increased oscillation frequency for the *repressilator*. For the dynamics in electronic gene *i* = 1 has its capacitor increased to 3.3 μf, so it has *β*₁ = 0.3. Thus the *β*-ratio for the genes in the *repressilator* circuit is 0.3∶1∶1. Gene *i* = 1 now has reduced oscillations, its inhibitor gene *i* = 3 has increased oscillations, and the *repressilator* frequency has decreased. The circuit presented here as an electronic analog of a synthetic genetic network known as the *repressilator* shows good agreement between experimental measurements and numerical predictions. The circuit includes control of parameters for the Hill function which is used to model the kinetics of gene expression and inhibition in the cyclic 3-gene network. Previous electronic analogs of the *repressilator* – did not concentrate on the kinetics and control of the parameters, and thereby did not capture many complex dynamical features. With the ability to control the model parameters, this circuit will be useful for investigations of multistability of coupled *repressilators* as well as for other SGN dynamics. The authors thank Siddhartha Roy, Indian Institute of Chemical Biology, for valuable comments and suggestions. [^1]: Conceived and designed the experiments: EHH EV JK SKD. Performed the experiments: EHH. Analyzed the data: EHH SKD. Wrote the paper: EHH SKD. [^2]: The authors have declared that no competing interests exist.

# Introduction Worldwide, bronchiolitis remains one of the most common reasons for hospitalisation of children. Over 3 million children are diagnosed with bronchiolitis annually. The incidence of bronchiolitis is higher in some populations, including Alaskan Native and Indigenous Northern Territory (NT) infants. In the latter group, hospitalisation rates for bronchiolitis are higher (352 vs. 62.6 per 1000) and infections are more severe than non-Indigenous children. Bronchiolitis is a clinical syndrome that is diagnosed in children up to 24 months of age. The most common infecting virus, respiratory syncytial virus (RSV) occurs in 50–80% of cases, although an increasing number of viruses (e.g. human rhinoviruses (HRV), coronaviruses, bocavirus), including multiple infections are being identified. Current recommended therapies in hospitalised children are limited to oxygen (O₂), fluids and hypertonic saline nebulisation. Antibiotics are rarely advocated in the management of bronchiolitis unless the illness is very severe or when a secondary bacterial infection is suspected. However, semi- synthetic macrolides (e.g. azithromycin, clarithromycin) which have immuno- modulatory, and/or anti-microbial properties and *in-vitro* anti-viral effects, which may be beneficial in children with bronchiolitis and high nasopharyngeal carriage rates of bacteria. Three randomised placebo-controlled trials (RCTs) have been published and these RCTs used different doses and duration of macrolides to treat hospitalised bronchiolitis. These studies also differed in affluence of settings which may reflect differences in the frequency and severity of acute respiratory infections in these populations. Thus not surprisingly, results from the existing RCTs differed in the effect on reducing length of hospitalisation and O₂ requirement. A Turkish trial reported improved clinical outcomes. In comparison a European and a Brazilian trials showed no improvement. Bacterial infections in children with RSV positive acute lower respiratory infections range from 3.5% to 31%. The higher rate is more likely in less- affluent settings and/or with those with more severe disease. Viral-bacterial co-infections are more likely when the upper airways are densely colonised with bacteria or during repeated infections. In the NT, children have early acquisition of bacteria in the nasal space. This is more likely to be similar to Turkey where high rates of pneumonia and bronchiectasis are also reported. Thus, we conducted a RCT on children hospitalised with bronchiolitis. Our primary objective was to determine whether a single large dose of azithromycin (compared to placebo) reduced length of stay (LOS) and duration of O₂ requirement in children hospitalised with moderate to severe bronchiolitis. We also determined the influence of azithromycin on the incidence of respiratory readmissions and presence of bacteria and viruses in the nasopharynx. # Methods ## Study design Our double-blinded, placebo-controlled, RCT was conducted at the Royal Darwin Hospital between June 2008 and December 2011 and The Townsville Hospital between October 2010 and December 2011. ## Trial registration The trial was registered with the Australian and New Zealand Clinical Trials Register. Clinical trials number: ACTRN12608000150347. ## Ethics statement The study was approved by the Human Research Ethics Committee of the Northern Territory Department of Health and Menzies School of Health Research (HREC 07/60) and The Townsville Health Service District Human Research Ethics Committee (HREC/10/QTHS/9). Individual written informed consent was obtained from children's parents or legal guardian. ## Participants Children were enrolled if they were ≤18 months, admitted with a clinical diagnosis of bronchiolitis (according to standardised hospital protocols; ≤18 months, with cough and coryza, wheezing +/− crackles, respiratory distress with both tachypnoea (respiratory rate \>50 beats/min) and retractions), required supplemental O₂ and consented within 24 hrs of hospitalisation. Children were excluded if they had: severe disease (admitted to intensive care unit); chronic lung disease, congenital heart disease, contraindications to macrolide use (e.g. liver dysfunction, hypersensitivity), diarrhoea (\>2 stools of watery consistency more than normal pattern), received macrolides (in last 7-days), or clinical and radiological features consistent with a primary diagnosis of pneumonia at time of randomisation. ## Protocol and interventions used across both sites Study staff visited the paediatric wards twice daily to assess newly admitted children. After consent, children were randomised to receive a single large dose of oral liquid azithromycin (30 mg/kg) or placebo suspension (equal volume). The placebo suspension was made up of confectioner's Sugar, Hydroxypropyl Cellulose, Xanthan Gum, Syloid 244, Sodium Phosphate Tribasic, Imitation Vanilla Creamy Flavour, Black Cherry Flavour, Quinine Sulphate (ground Quinate 300 mg Tablets). Children were managed by the paediatric team of each hospital according to the same clinical protocol for bronchiolitis (e.g. criteria for commencement and weaning of O₂) that was standardised \>6 months before commencement of the trial. Children were allowed to receive concurrent medications specified by the attending physician, except macrolide antibiotics. The protocol for this trial and supporting CONSORT checklist are available as supporting information; see and. ## Randomisation, allocation and blinding Randomisation was stratified by age (≤6 or \>6 months), ethnicity (Indigenous or non-Indigenous) and site (Darwin or Townsville). Randomisation was by computer generated permuted blocks and treatment allocation concealed by opaque stickers. Upon enrolment, children were assigned the next treatment on the appropriate stratified list. Neither the study team (researchers, hospital staff) nor parents were aware of the assigned treatment group until the end of the trial. The placebo medication was manufactured by the Institute of Drug Technology Australia Limited (Melbourne, Victoria). It had a similar smell and taste to active azithromycin. Azithromycin (Pfizer, Australia) was repackaged by IDT. Both medications were prepared as powder in identical opaque bottles and sealed with an aluminium foil. ## Clinical assessment and outcome measures Standardised data collection forms were used to record demographic, medical history and clinical data from each child (table-1). This included O₂ requirement and level, respiratory rate, temperature and heart rate. Other therapies (intravenous fluids, antibiotics) and routine investigations (full blood count and chest x-ray) were also documented. Children enrolled in the study were assessed twice daily by the attending doctor (blinded and not an investigator) for clinical signs inconsistent with bronchiolitis and associated with known azithromycin side effects. Outcome measures were collected every 12 hours until the study endpoint was reached. The primary endpoints were: LOS for respiratory illness and duration of O₂ requirement. LOS was defined as time from admission to time for ‘ready for discharge’ (Sp0₂ consistently \>94% in air for \>16 hrs) and feeding adequately. ‘Ready for discharge’ differed from LOS, as discharge from hospital in our setting is often delayed due to other social factors, especially in children from remote Indigenous communities. Other outcomes were (i) any respiratory related readmissions within 6 months of discharge and (ii) identification of respiratory viruses and bacterial pathogens. Adverse events were monitored by study staff every 12 hours until discharge. Respiratory readmissions were collected from the medical charts; as there these children had no access to any other hospitals in the region, this is a reliable outcome. ## Specimen collection and process A nasopharyngeal swab (NPS) was taken prior to administration of study medication and 48 hrs later (or at discharge). NPS were placed in skim milk tryptone glucose glycerol broth media and were transferred on ice stored at −80°C in accordance to published guidelines. Assessment for viruses and atypical bacteria were described previously –. Nucleic acids were extracted from 0.2 ml of each NPS using the High Pure Viral Nucleic Acid kit (Roche Diagnostics, Australia), according to the manufacturer's instructions. Polymerase Chain Reaction (PCR) methods were used to detect RSV (A and B), adenovirus, parainfluenza (1, 2, 3), influenzavirus (A and B), HRV and enterovirus, coronaviruses, bocavirus type 1, human metapneumovirus (hMPV), KI (KIPyV) and WU (WUPyV) polyomaviruses, *Chlamydophila pneumoniae* and *Mycoplasma pneumoniae*. For bacterial analysis, NPS were thawed and 10 µL aliquots cultured overnight on selective media at 37°C in 5% CO₂; identification of *Streptococcus pneumoniae*, *Haemophilus influenzae*, *Moraxella catarrhalis* and *Staphylococcus aureus* used established techniques as previously described. ## Statistical methods We formally compared baseline characteristics of Indigenous vs. non-Indigenous children, with appropriate statistical tests. We did not formally do this between treatment groups in accordance with current CONSORT recommendations (available <http://www.consort-statement.org/consort- statement/13-19-results/item15_baseline-data/>. Accessed 28^th June 2013). Our pre-specified analysis plan, stated that non parametric methods be used if data were not normally distributed. Data were presented as medians and interquartile range (IQR) for LOS and O_2. Differences between groups were tested using the Mann-Whitney test. A 95% confidence interval (CI) was obtained for the difference in medians between treatment groups. Subgroup analysis was performed by ethnicity (Indigenous vs. non-Indigenous) and by age (≤6 and \>6 months). Differences in proportions were tested with Fisher's exact test. We looked at time to readmission within 6 months of hospital discharge using Kaplan-Meier survival plots. ## Sample size We calculated that a total sample of 92 children (equal numbers of Indigenous and non-Indigenous children recruited) would provide 90% power to detect a difference in the mean LOS of 24 hrs between each treatment for each ethnic group at the 5% significance level assuming the standard deviation was 24 hrs in each group. This study was underpowered to detect differences in rates of readmission between treatment groups. # Results We recruited 97 children and data from 96 children were analysed. The major reason why 450 children did not meet the inclusion criteria was they did not require supplemental O₂ or were admitted over the weekend. During recruitment, 21 children admitted into intensive care were excluded; 17 were Indigenous. One participant was excluded from the analysis of primary outcomes; they had received a macrolide in the previous 7 days (this child was randomised to placebo). This child was included in the analysis of secondary outcomes. Of the 96 remaining children, demographic and clinical characteristics were similar between the treatment groups (Table-1). No children received steroids during hospitalisation. Of the cohort, 10 children were previously hospitalised for a respiratory episode, all were Indigenous; 3 in azithromycin group and 7 in placebo group. When data were grouped by ethnicity, a higher proportion of Indigenous children lived in remote areas (n = 45, 74%; p = \<0.001), were exposed to cigarette smoke during pregnancy (n = 37, 61%; p = \<0.001), or in their household (n = 44, 72%; p = \<0.001). They were more likely to have coexisting co- morbidities (i.e. skin infections (n = 17, 28%; p = 0.02) or secondary pneumonia (n = 13, 21%; p = 0.01). More Indigenous children (n = 34, 56%, p = \<0.001) received antibiotics prior to hospitalisation. The antibiotics given were ceftriaxone (n = 14, 41%), procaine penicillin (n = 7, 21%) and amoxicillin (n = 5, 15%). In hospital, additional antibiotics were more often prescribed in Indigenous children (n = 52, 85%, p = \<0.001) (Table-1). LOS was similar in both treatment groups. The median LOS in the azithromycin group was 54 hrs, compared to 58 hrs in the placebo group (difference between groups of 4 hrs, 95%CI −8, 13, p = 0.6). The median time on O₂ in the azithromycin group was 35 hrs, compared to 42 hrs in the placebo group (i.e. reduction of 7 hrs 95%CI −9, 13, p = 0.7). No child required admission into intensive care and there were no adverse or serious adverse events. All children contributed to readmission data. There was no significant difference in the number of respiratory readmissions within 6 months (10 per group, OR = 0.9, 95%CI 0.3, 2, p = 0.8) or time to readmission (logrank p = 0.9) between treatment groups. 70% of children readmitted, were reported to have a wheeze-associated illness. Indigenous children (n = 61) had longer LOS; median 59 hrs compared to 51 hrs in non-Indigenous children (n = 35) (difference of 8 hrs, 95%CI -25, 1.5, p = 0.07). This was similar with duration of O₂; 43 hrs in Indigenous children and 35 hrs in non-Indigenous children (difference of 8 hrs 95%CI -22, 1.4, p = 0.08). A higher proportion of Indigenous children were readmitted for a respiratory illness (n = 16 (26%) compared to non-Indigenous children (n = 4 (11%)), difference 15% (95%CI 0, 30%) p = 0.05. Indigenous children were more likely to be re-hospitalised earlier (Indigenous n = 16, non-Indigenous n = 4, OR = 2.8, 95% CI 0.9, 8.8), logrank p = 0.08. There was no evidence that the difference in either LOS or O₂ between treatment groups varied according to ethnicity or age (table-2). ## Viral and bacteria data All but one child had a baseline NPS. NPS could not be obtained on all participants at 48 hrs due to discharge occurring during evenings or weekends. One participant's family withdrew consent for NPS. At baseline, viruses were not detected in 23 (24%) participants. One or more virus was detected in 54 (56%) children. Two or more viruses were detected in 19 (20%) of children. RSV was the most common (n = 48, 50%), followed by HRV (n = 16, 17%), hMPV (n = 5, 5%) and coronavirus (n = 5, 5%). depicts the frequency of virus detection at baseline and 48hrs. There was no reduction in the mean number of viruses detected per child from baseline to 48hrs; azithromycin 1.0 to 0.8 (95% CI −0.2, 0.6, n = 34), placebo 0.9 to 1.0 (95% CI −0.3, 0.2, n = 37). Table-3 summarises NPS bacteria detected at baseline and 48 hrs. A reduction in the mean number of respiratory bacteria was detected per child; in the azithromycin group from 1.2 to 0.5 bacteria (difference 0.7 95%CI 0.25, 1.1, p = 0.01), and zero change in the placebo group; 1.3 to 1.3 bacteria. Compared to baseline NPS, children who received azithromycin alone (n = 14) (i.e. received no additional antibiotics in hospital) were less likely to have *S. pneumoniae* (3/14, vs. 0/11), *M. catarrhalis* (5/14 vs. 1/11), *H. influenzae* (6/14 vs. 1/11), and *Staphylococcus aureus* (1/14 vs.0/11) at 48 hrs. 3/14 children did not have NPS at 48 hrs. # Discussion We found that a large single dose of azithromycin (compared to placebo), did not have large clinical effects on LOS, length of O₂ requirement or readmission within 6 months of discharge. Azithromycin reduced the proportion of children with respiratory bacteria in their NPS but had no significant effect on viral detection by PCR. Of the 3 published RCTs on macrolides to treat hospitalised children with bronchiolitis, – only one reported improved clinical effects i.e. reduced LOS, duration of O₂ requirement and lower readmission rates. Our findings are similar to the other two trials, showing that a single dose azithromycin does not shorten LOS and O₂ requirement. However, methodological differences among trials need to be considered. One of the larger trials included only children with RSV-confirmed bronchiolitis, thus limiting generalisation to bronchiolitis caused by other pathogens. Other differences included: age, ethnic populations, concurrent use of antibiotics, treatment practices and macrolide type, dose and duration. The immunomodulatory difference between azithromycin and clarithromycin may also account for the differential results between Tahan et al's study with ours and the other 2 RCTs. For example, azithromycin increased the production of IL-10 whereas clarithromycin inhibited the production of IL-6 by dendritic cells in animal work. Tahan and colleagues used a daily dose for 3-weeks of clarithromycin, but ours like Kneyber et al, (7-day daily dose) used a short course of azithromycin. However, Tahan et al's study had very small numbers (n = 21) and high attrition. Thus, it remains unknown if a longer course of azithromycin may be effective in reducing readmission rates. Of the published RCTs on macrolides for bronchiolitis in children, our patient profile is most like that of the Turkish study. However, unlike the Turkish RCT, we did not find a beneficial effect of azithromycin on clinical outcomes. Possibly contributory reasons include the very high concomitant use of antibiotics in our group; different treatment regime used the density of bacterial carriage and secondary co-morbidities. The common use of antibiotics in children with bronchiolitis in our setting relates to the high rates of concomitant infections among children. A similar treatment practice occurs in Alaskan children. 56% of Indigenous children in our trial received antibiotics before admission and 87% during hospitalisation. While the use of antibiotics is common practice in such settings, its effectiveness and possibly increased adverse events remains unknown. Ideally concurrent antibiotics should have been disallowed in our study but it was not possible to alter clinicians' practice and our *a priori* protocol allowed the concurrent use of antibiotics other than macrolides. We used a single large dose of azithromycin, which is equivalent to one week of treatment for several reasons. In our setting, early (as early as 2-weeks of age) nasopharyngeal colonisation of respiratory pathogens occurs in Indigenous children. Azithromycin potentially has a beneficial microbiological effect on these pathogens. Azithromycin also has the benefit of a long half life and tissue penetration requiring less frequent dosing, compared to other antibiotics. This is important in our setting where adherence to treatment regimes can be challenging. While our trial did not find significant differences between treatment groups for clinical outcomes, our study had some novel data. Firstly, none of the published RCTs on macrolides for children with bronchiolitis report data on the impact on viral detection or bacterial carriage. As viruses were identified by PCR, it is not surprising no difference in viral detection were found at 48 hrs (although azithromycin may have some anti-viral effect). Future research should look at the impact of azithromycin on viral load/copies. While our numbers were small, we showed a significant difference in the mean number of respiratory bacteria per child; from 1.2 to 0.5 bacteria (difference 0.7 95%CI 0.25, 1.1, p = 0.01) in the azithromycin group. This is important in our setting as NPS carriage of respiratory pathogens (e.g. *S. pneumoniae, H. influenzae, M. catarrhalis*) are among the highest reported globally (over 80%), compared to non-Indigenous children (50%). Secondly, our study provides a clinical picture of hospitalised cases of bronchiolitis in different geographical and ethnic groups in Northern Australia, where acute respiratory infections may be one precursor of high rates of chronic respiratory illness, This is the first data published, showing NPS detection of viruses and bacteria from Indigenous children hospitalised with bronchiolitis. Wheezing and bacterial infections in young children have been shown to be associated in one prospective cohort study. Thus our data may have implications in settings, where acute and chronic respiratory diseases are particularly prevalent and more severe, including Alaska and New Zealand. We found that Indigenous children exhibited longer LOS, O₂ requirement and earlier time to hospital readmission than non-Indigenous children. The most likely reason why the latter aspects are different from our previous study is because we excluded children managed in intensive care. In our cohort, Indigenous children were more likely to be readmitted for another respiratory illness. This is not surprising as Indigenous children in the NT are 5 times more likely to be hospitalised for pneumonia and influenza than non- Indigenous children. Whether readmission is related to the insult from bronchiolitis can only be postulated. Recurrent hospitalisation for respiratory illness is an independent risk factor for developing bronchiectasis and/or respiratory dysfunction in adulthood. In our region, bronchiectasis affects 1 in every 68 Indigenous children. In the follow up of our cohort, 6/61 (10%) Indigenous children (3 in azithromycin group, 3 in placebo group) have subsequently been diagnosed with bronchiectasis and an additional 4 children are awaiting chest scans. The prevalence of readmission for a respiratory illness within 6-months in our trial was 21%; 70% had a wheezing illness. This was similar to the Turkish trial at 24% (53% were wheezing). The two most common viruses found in our cohort, RSV and HRV have been implicated for ongoing wheezing. New Zealand data have also recently described high prevalence (70%) of on-going intermittent wheeze 12-months post hospitalisation with acute lower respiratory infections. In addition, wheezing and persistent cough can also be problematic post acute bronchiolitis. Our trial (and the other published RCTs of macrolides for acute bronchiolitis) did not assess this, a known clinical research gap. Despite providing new data, there are other several limitations to our study, in addition to the concurrent use of antibiotics. Having older children increased the risk of including asthma prone children. We also did not limit to the first bronchiolitis admission. Removing the children with recurrent disease in a secondary analysis made no difference to study outcomes. As our study was limited to a single dose, it remains uncertain if any macrolides, or a longer macrolide treatment course, is beneficial in high risk children who do not receive any other antibiotics. Only having two sites, may also affect the generalisability of the results. # Conclusion In children hospitalised with moderate to severe bronchiolitis and requiring supplemental O₂, we found that a large single dose of azithromycin (compared to placebo) did not have any clinical benefit to reduce LOS, duration of O₂ requirement or readmission rates within 6 months of hospital discharge. Azithromycin reduced the proportion of bacterial carriage, but had no significant effect on reducing proportion of viruses. Further research is required to determine whether earlier administration and longer duration of azithromycin is beneficial to improve the clinical and microbiological outcomes of acute bronchiolitis, associated co-morbidities and prevent ongoing respiratory morbidity in this population. # Supporting Information We thank Lesley Versteegh, Clare Wilson, Susan Pizzutto, Kobi Schutz, Nerida Jacobsen and Louise Axford-Haines for assisting in data collection. We thank Kim Hare, Vanya Hampton, Jane Gaydon and Rebecca Rockett for processing the viral and bacterial samples. We thank Linda Ward for statistical support. We also thank the medical and nursing staff for their ongoing support and identifying the children for the study. We are grateful for all the children and families who participated in the study. [^1]: The authors declared that no competing interests exist. [^2]: Conceived and designed the experiments: AC PM CM. Performed the experiments: GBM CM AW. Analyzed the data: GBM MC. Contributed reagents/materials/analysis tools: GBM MC AC PM IM TS CM AW. Wrote the paper: GBM AC PM MC IM TS CM AW. Set up and coordinated the study, managed the project, performed the data analysis and drafted the manuscript: GBM. Conceptualised the study, was responsible for obtaining the grants, interpreted the data and edited the manuscript: AC. Assisted in study design and data interpretation and contributed to the manuscript: PM. Responsible for standardising the management of bronchiolitis in their units, recruiting participants and edited the manuscript: CM AW. Assisted in the viral components of the study and the manuscript: TS IM. Assisted in the data analysis and edited the manuscript: MC. Read and approved the final manuscript: GBM PM MC CM AW TS IM AC.