Datasets:

Daniel-P-Gonzalez

/

OCD

like 0

# 1 Introduction According to the most recently updated Cochrane systematic review comprising 26 randomized controlled trial studies and 24 non-randomized studies, there is some evidence that nicotine-containing e-cigarettes (vaping) might improve the rate of 6-month smoking abstinence without causing serious adverse events when compared to nicotine replacement therapy (such as nicotine patches) among adult smokers. The effectiveness of nicotine-containing vaping over behavioral support alone is less certain, although overall the evidence appears to be favorable. As vaping might represent a promising option to help some adult smokers quit smoking, it is crucial to identify the characteristics of these smokers to guide the design of health programs that incorporate vaping as a cessation method. Real-world studies have revealed successful vaping-assisted quitting attempters to share a set of unique characteristics, including increased vaping intensity, daily use of e-cigarettes, longer vaping history, the use of a specific type of vaping device, and being male. Both qualitative and quantitative studies also point to the use of certain non-tobacco flavors and more positive overall experiences during a vaping-assisted quit attempt to be associated with potential cessation benefits. As most previous studies used regression to identify characteristics of a successful quitter who used vaping, their results are subject to limitations. In the presence of a large pool of variables, all of which are theoretical plausible predictors of smoking cessation, it is often difficult to decide which variables are to be entered into a regression equation and in what mathematical form. Additionally, as variables that describe a successful quitter are likely correlated in intricate manners, completely reflecting these relationships in a regression equation can be challenging. These issues could raise the complexity in a regression analysis and cause concerns of multicollinearity, overfitting and other statistical issues that limit the study findings. We aim to address these limitations in the current analysis using machine learning, a group of computationally-intensive and data-driven analytical methods that has gained increasing popularity in health research, including in the research of smoking cessation and behaviours of vaping. Compared to conventional regression, machine learning has strengths in mitigating the risk of model overfitting and producing highly accurate and robust prediction. To the best of our knowledge, this is the first application of machine learning in vaping-assisted smoking cessation research. Using survey data collected from a group of adults who tried vaping to quit smoking, we aim to identify and rank the importance of the top five predictors of having perceived success in vaping- assisted smoking cessation. # 2 Materials and methods ## 2.1 Study design and population This is a cross-sectional study based in Ontario, Canada. During July and August 2019, we sent an email invitation to 8,109 adult current smokers or recent quitters who had accessed two provincially based smoking cessation initiatives to respond to a survey about vaping to quit smoking, including 4,665 registrants of the Smokers’ Panel, an ongoing initiative that recruits and follows up with smokers to collect data in surveys and 3,444 participants of the Leave the Pack Behind program which provided free nicotine patches, gums and online resources to young adults. We received 1,721 replies (response rate = 21.2%). The Vape to Quit survey (see the complete survey instruments in the) was administered to individuals who identified themselves to be vaping-assisted cigarette quitters during the past 12 months, i.e., those who reported making at least one serious attempt to quit cigarette smoking (consciously trying to stay off cigarettes for good) by vaping in the past 12 months. Hence, among the 1,721 individuals, we excluded 135 people who did not consent to participate in this study, 219 people who did not smoke cigarettes at all in the past 12 months, 123 people who did not make any serious attempts to quit smoking in the past 12 months, 231 people who did not use e-cigarettes for smoking cessation in the past 12 months, and 20 people who did not give consent for us to use their data from the Vape to Quit Survey (see for a flow chart showing the study sample). These exclusions yielded 993 individuals who met eligibility criteria and thus were administered the survey. All 993 participants completed the survey and received a \$10 e-gift card while entering into a draw to win one of two \$250 Visa gift cards. In this study, we further excluded 25 people who did not confirm the use of e-cigarettes for smoking cessation in the survey, one person who did not report smoking cessation outcomes, and 78 people who had no response for at least one entire module of the survey. These exclusions yielded 889 participants in the study sample. ## 2.2 Ethics approval and informed consent This study received ethical approval by the University of Toronto Health Sciences Research Ethics Board on August 7, 2018 (#10321). Individuals provided written informed consent to participate in this study in addition to allowing us to use their responses to the Vape to Quit Survey in the analysis. Those who did not give consent were excluded from this study. ## 2.3 Outcome A binary outcome variable was created to represent the status of self-reported success in vaping-assisted smoking cessation in the past 12 months. This variable was measured using the survey question, “Over the past 12 months, would you say that vaping has helped you …?” Participants with perceived success in quitting smoking by vaping answered, “completely helped me quit smoking cigarettes.” Participants with perceived failure answered, “be not at all successful at cutting down smoking cigarettes”, “be somewhat successful at cutting down smoking cigarettes”, or “be very successful at cutting down smoking cigarettes”. ## 2.4 Candidate predictors Fifty-one characteristics representing a wide spectrum of individual level status and experiences were extracted from the survey to be candidate predictors of perceived success in vaping-assisted smoking cessation. These variables captured sociodemographic information, health status, history of cigarette smoking and quitting, preferences for vaping, side effects from using vaping to quit smoking, the use of other methods while vaping to quit smoking, history of vaping, and substance use. A Vaping Experiences Score (VES) was also used to predict self-reported vaping-assisted smoking cessation. This score was derived from results of a factor analysis on 42 vaping experiences items that we list in. For each item, participants rated how true the statement was for them on a 7-point scale with 1 being “not true at all” and 7 being “extremely true”. The factor analysis revealed six vaping experiences factors, including Relationships, Flexibility of Vaping, Side Effects, Vaping Devices, Public Reactions and Sensory Functions (see and the factor analysis done by our group). The VES was estimated as the weighted sum of factor scores over the six factors by the proportion of variance explained. To facilitate interpretation, we categorized individual’s vaping experiences into four levels, based on the four quartiles of the VES: poor (VES≤the first quartile), fair (the first quartile \<VES≤the median), good (the median\<VES≤the third quartile), and excellent (\>the third quartile). All candidate predictors were represented using a categorical variable. ## 2.5 Statistical analysis ### 2.5.1 Descriptive statistics and data imputation Characteristics of successful vaping-assisted quitters vs. unsuccessful vaping- assisted quitters were summarized and compared using the Chi-square test. Overall, less than 0.8% of data were missing, except for two variables—preferred nicotine strength used in vaping (n = 119, 13.4%) and motivation level to quit smoking (n = 80, 9.0%). We used the R package “mice” to generate five imputed copies of our data based on the algorithm of multiple imputation by chained equations, after visual inspection confirmed the assumption that data were missing at random. We used the first copy in the primary analysis and returned to the remaining four copies in a sensitivity analysis. ### 2.5.2 Variable selection Early literature recommended variable selection procedures prior to machine learning to potentially reduce the risk of model overfitting, prevent spurious associations, and to optimize model training efficiency. The utility of performing variable selection has been confirmed in more recent machine learning applications in health, including in the prediction of smoking cessation outcomes. Hence, we implemented two criteria where a predictor was selected into the model if (i) it was one of the six sociodemographic variables or (ii) it was deemed significant in a logistic regression penalized by Lasso (least absolute shrinkage and selection operator). In criterion (ii), we entered all variables into a logistic regression to be independent predictors of the odds of successful vaping-assisted smoking cessation. A loss function comprising the negative log-likelihood function and the sum of the absolute value of all coefficients was minimized, subject to an unknown tuning parameter that controlled the severity of penalty placed onto the total magnitude of coefficients. This procedure shrank the coefficient of unimportant variables to zero and thereby achieving variable selection. To identify an optimal value of the tuning parameter, a ten-fold cross-validation procedure was performed using the R package “glmnet” such that the cross-validated negative log-likelihood function was minimized. Using the optimal value of the tuning parameter, variables with a non-zero coefficient were selected into the analysis by criterion (ii). ### 2.5.3 Developing and validating a gradient boosting machine This was a predictive modeling study where we aimed to develop a classification model to predict the status of successful vaping-assisted quitters. Thus, our primary goal was to maximize the predictive power of the final model, rather than to convey any underlying relational or causal relationships as seen in an explanatory modeling study. Interested readers are referred to a previous work done by our group where multivariable logistic modeling was applied to quantify the association between the odds of successful vaping-assisted smoking cessation and various individual-level characteristics. In the present study, we used the R package “gbm” to develop a gradient boosting machine (GBM) model that classified successful vaping-assisted quitters and their unsuccessful counterparts using variables selected from the previous procedure. The GBM is an ensemble model where many weak classification tree models are converted into one single strong model to produce prediction. These tree models are developed sequentially such that each additional tree corrects the prediction error of the preceding tree, and thus continuously updating and improving the prediction. We outlined the analytical procedures below and reserved details in. A ratio of 7:3 was used to split the data randomly into a larger set used for model training (n = 623) and a smaller set reserved for model testing (n = 266). Because the number of successful vaping-assisted quitters is substantially less than that of unsuccessful quitters, with a ratio of 1:4, we applied SMOTE (synthetic minority over-sampling technique) in the R package “DMwR” to create synthetic samples of successful quitters on the basis of nearest neighbor to achieve a more balanced training set. Using data from the resampled training set, 5000 tree models were developed sequentially using a gradient descent method (assuming a learning rate of 0.01) to classify smoking cessation status, assuming each tree had at most 6 binary splits and at least 5 observations in its terminal nodes. Extensive tuning procedures were conducted using an internal ten-fold cross-validation process to prune tree models and to identify optimal values for the other parameters of GBM, in order to rule out the possibility of model overfitting while maintaining the overall performance of the GBM. We used the area under the receiver operating characteristic (ROC) curve, known as the AUC, to measure performance of the GBM on the testing set. The AUC ranges from 0.5 to 1 with higher values indicating greater discriminatory ability. A model with an AUC ≥ 0.80 is generally considered to be strong. ### 2.5.4 Ranking of variable importance We calculated a score for each predictor to represent its relative importance in the training of the GBM model using the “relative.influence” function in the “gbm” package. To calculate the score, for each variable, we computed the improvement in the Gini Index at each split in each tree, and then calculated the average improvements across all trees in the GBM. We scaled this score to be between 0%-100% such that the variable with the highest score had 100% importance. The top five predictors were identified and ranked using a bar plot. ### 2.5.5 Visualizing partial dependences We used the “plot.gbm” function in the “gbm” package to examine graphically the partial dependence of the predicted probability of reporting success in vaping- assisted smoking cessation based on the GBM model on values of the five top predictors. This procedure, recommended to be performed after the ranking of variable importance, provides visual interpretations on the approximated relationship between the top predictors and the outcome while other variables are “integrated out”. The partial dependence was estimated for each of the top five predictor at each of its value as the marginal success rate predicted by the GBM model by forcing all data to have the same value for this predictor. Visualizations were performed using the “ggplot2” package. ### 2.5.6 Sensitivity analysis We first repeated all analytical procedures on each of the four remaining imputed datasets to identify the effect of the missing values. We then assessed a parsimonious GBM model that was trained using only the top five predictors. After that, we deliberately omitted the VES variable from the model to assess its impact on the testing results. And finally, we estimated a conventional multivariable logistic model with data from the original training set by entering all variables as independent predictors of vaping-assisted cigarette cessation. Performance of the logistic model was assessed on the same testing set. Analyses were performed on R version 3.6.1 (R Foundation for Statistical Computing). # 3 Results ## 3.1 Characteristics of the study sample Of the 889 participants, 174 (19.6%) reported to have successfully quitted smoking by vaping in the past 12 months. Compared to their unsuccessful counterparts, these successful quitters were less likely to be female (49.1% vs. 58.0%) but were more likely to be employed or self-employed (82.8% vs. 75.1%). Self-reported health status did not differ between the two groups, although successful quitters were more likely to report having no formally diagnosed health conditions (38.5% vs. 29.7%). Furthermore, successful quitters had higher motivation to quit smoking in general (80.5% vs. 63.4%), but had lower likelihood of having set up a quit date in advance (33.3% vs. 45.0%). The two groups of participants showed significant differences in terms of their preference for and history of vaping. Specifically, successful quitters were more likely to use a pod system (44.5% vs. 30.9%), to use 2.1% or higher nicotine in vape (21.3% vs. 18.6%), to vape at least 10 times daily (69.0% vs. 51.6%), to vape within 15 minutes after waking up (39.2% vs. 33.9%), and having vaped 100 times (74.1% vs. 56.0%). When discussing the use of vaping for quitting smoking, successful quitters were more likely to have started vaping specifically for smoking cessation (91.4% vs. 73.8%) and to report less previously failed attempts (first time attemper: 54.6% vs. 26.6%). Regarding their latest quit attempt, successful quitters were more likely to report that they had started this attempt more than 12 months ago (24.7% vs. 9.1%). During this attempt, successful quitters were less likely to report throat irritation (54.0% vs. 62.8%) or the concurrent use of another method besides vaping to aid quitting. In terms of substance use, successful quitters were less likely to be current users of waterpipe (5.3% vs. 14.3%) or other tobacco products such as cigars (6.5% vs. 14.5%). illustrates the VES categories between the two groups of quitters. In general, successful quitters were significantly more likely to have more positive experiences as measured by the VES (p\<0.001). Notably, more than half (54.0%) of successful quitters rated their experiences to be excellent compared to just 17.9% of unsuccessful quitters. ## 3.2 Model structure and performance The Lasso logistic regression excluded 9 unimportant variables that were mostly related to preferred flavors of vaping. Entering the remaining 42 predictors into the analysis yielded a final GBM model with an AUC of 0.865, specificity of 0.750 (95% confidence interval \[CI\] 0.614 to 0.863), sensitivity of 0.851 (95% CI 0.798 to 0.902) and accuracy of 0.831 (95% CI 0.780 to 0.874) on the testing set. ## 3.3 Top five predictors of perceived success in vaping-assisted smoking cessation We identified and ranked the importance of the top five predictors of self- reported vaping-assisted smoking cessation on the basis of a score between 0%-100%. These predictors were more positive experiences measured by the VES (100%), less previously failed quit attempts by vaping (39.0%), younger age (21.9%), having vaped 100 times (16.8%) and vaping shortly after waking up (15.8%). ## 3.4 Partial dependence To enhance the interpretability of the top five predictors, we estimated the partial dependence of the predicted probability of perceived success in vaping- assisted smoking cessation on the values of these predictors. Having positive vaping experiences alone was found to be strongly associated with a higher rate of successful quitting, as the model predicted the marginal success rate of participants with poor, fair, good, and excellent vaping experiences to be 4.8%, 13.4%, 46.2%, and 77.1%. The model also suggested that, compared with participants who vaped to quit smoking for the first time, those with previous failed attempts were less likely to report success. Notably, the marginal success rate decreased consistently from 68.6% (first attempt) to 20.7% (second attempt) to 10.5% (third attempt) to just 2.5% (four or more attempt). The model found a decreasing trend of smoking cessation by age, where the youngest age group (20–29) had the highest marginal probability of quitting (49.5%), followed by participants in their 30s (30.8%), in their 40s (19.4%), and those who were 50 years old or above (5.2%). Furthermore, the marginal success rate decreased from 39.7% for participants who had vaped 100 times or more to 19.4% for those who hadn’t vaped 100 times. Finally, the marginal probability of quitting decreased by longer time between waking up and first vaping session. Specifically, smokers who started vaping within 15 minutes after waking up reported the highest marginal success rate (46.9%), which reduced to 20.8% for those who vaped in 15–60 minutes, and to 10.4% for those vaped at least an hour thereafter. ## 3.5 Sensitivity analysis Results of sensitivity analyses are presented in. Repeating procedures on each of the remaining four imputed data copies yielded very similar models with AUCs ranging from 0.855 to 0.869. Minor discrepancy emerged as in one iteration, time after waking up to vape (15.9%) replaced having vaped 100 times (14.4%) as the fourth most important predictor. A parsimonious model that was trained using only the top five predictors reached slightly reduced AUC of 0.825, sensitivity of 0.664 (95% CI 0.609 to 0.735), specificity of 0.827 (95% CI 0.704 to 0.924), and accuracy of 0.700 (95% CI 0.636 to 0.750), on the testing set, but did not change the importance ranking for the top five predictors or their importance scores. Excluding the VES variable reduced model performance on the testing set, resulting in an AUC of 0.772, sensitivity of 0.463 (95% CI 0.399 to 0.532), specificity of 0.846 (95% CI 0.724 to 0.931) and accuracy of 0.538 (95% CI 0.476 to 0.599). In this model, the top five predictors were the number of vaping- assisted quit attempts (100%), age (55.7%), having vaped 100 times (51.7%), using a pod system (45.6%) and time after waking up to vape (37.4%). Finally, a multivariable logistic model treating all variables as independent predictors of vaping-assisted cigarette cessation was estimated using the data from the original training set. The results of this analysis corroborated the importance of four of the top five predictors but ruled out the significance of time after waking up to vape. Specifically, when compared to participants with poor vaping experiences, those with fair, good or excellent experiences were associated with 2.818-fold (95% CI 1.616 to 5.075, p\<0.001), 7.593-fold (95% CI 2.475 to 33.171, p = 0.002) and 15.660-fold (95% CI 5.671 to 52.801, p\<0.001) odds of reporting a successful vaping-assisted smoking cessation. On the testing set, the logistic model reached an AUC of 0.701, sensitivity of 0.940 (95% CI 0.910–0.979), specificity of 0.248 (95% CI 0.147–0.362) and accuracy of 0.808 (95% CI 0.764–0.848). # 4 Discussion Using survey data from 889 Canadian smokers who used vaping to quit smoking in the past 12 months, we developed and validated a machine learning model that identified predictors of perceived success in vaping-assisted smoking cessation. The final model achieved high performance and suggested the VES, number of quit attempts by vaping, younger age, having vaped 100 times, and vaping shortly after waking up were the most important predictors of self-reported successful quitting. These results are robust to missing data and alternative model specification. About 20% (19.6%) of participants in our sample reported success in quitting smoking by vaping, a rate that closely resembles observations from both trial and real-world settings in Europe and the US. Thanks to a novel machine learning technique, we were able to build upon results of previous studies and more importantly, to identify and also rank the importance of, variables that were correlated with higher rate of perceived success in vaping-assisted smoking cessation. Compared with conventional logistic regression approach, we showed machine learning greatly improved model performance in the presence of a large number of variables and limited sample size. Hence, we believe that machine learning is well-suited in health outcomes research as it leverages computational power to effectively and directly identify influential variables and to quantify the strength of relationships. In our case, the ranking yielded by the machine learning analysis provides practical and actionable inputs on priority-setting in the development of interventions that maximize the effectiveness of e-cigarettes as a smoking cessation device. Consistent with early studies done by our group, we identified the VES, an index for measuring the experiences of smokers during a vaping-assisted quit attempt, as the most important predictor of self-reported smoking cessation. Notably, when holding other characteristics constant, smokers with excellent vaping experiences were associated with 16-times averaged predicted probability of quitting smoking compared to those who scored in the lowest quartile (77.1% vs. 4.8%). Furthermore, excluding the VES variable resulted in a decrease in model performance, especially in sensitivity (i.e., the ability to correctly identify true quitters). To give a more concrete example, suppose a female first-time quitting attempter in her 40s has vaped 100+ times and is currently vaping within 15 minutes after waking up (while holding other characteristics at reference level). The machine learning model predicts her perceived success rate during the current (first) quit attempt to be 2.8%, 9.1%, 12.4%, and 39.4%, corresponding to having poor, fair, good, and excellent vaping experiences. This implies improving this individual’s vaping experiences from poor to excellent may be associated with a drastic, 14-fold increase in her probability of self- reported smoking cessation. Another point worth noting is that, compared to other top predictors, the vaping experiences of smokers during a quit attempt are relatively modifiable and could be effectively improved by interventions, such as switching to a more adequate vaping device and increased monitoring for side effects. Hence, we believe the paramount importance of vaping experiences yielded by our analysis points to a feasible opportunity to prioritize strategies that enhance the experiences of smokers who rely on vaping as the main cessation aid as a way to potentially improve their likelihood of cessation. However, our results need to be interpreted with caution as they did not infer vaping experiences to be a determinant of successful smoking cessation. Hence, more in-depth research needs to explicate the specific mechanism of vaping experiences in influencing smoking cessation outcomes to support policy actions. The second and third most important predictors were the number of quit attempts by vaping and the age of smoker, although both were much less important compared to the VES. The probability of quitting decreased consistently with increasing number of quit attempts, where we found first-time quitters were more than 30-fold more likely to quit on average compared to those who had already failed 3+ times (68.6% vs. 2.5%). A similar trend was revealed in age, as the youngest smokers in their 20s were associated with about 8-times the chance of quitting smoking by vaping compared to older smokers aged 50 or above (49.5% vs. 5.2%). An interesting finding was a potential positive relationship between the likelihood of smoking cessation by vaping and indicators of vaping addiction. The model suggested that both established vapers (defined as having vaped 100 times or more) and those who vaped shortly after waking up were associated with higher success rates with vaping in quitting smoking. These observations imply that individuals with perceived success in vaping-assisted smoking cessation may have simply shifted their mode of nicotine consumption from combustible cigarettes to e-cigarettes, without eliminating the use of nicotine. Finally, the analysis confirmed the importance of previously identified predictors of smoking cessation, however they were not as important as our top five predictors. For example, daily vaping frequency was identified to be the eighth important predictor with a score of 4.4%. Specifically, those who vaped \<10 and 10+ times daily had marginal cessation rate of 26.7% and 28.8%, respectively. Furthermore, vaping device was the sixteenth predictor with a score of 1.8% where using a pod system was associated with slightly higher marginal probability of quitting than using other devices (36.8% vs. 27.5%). Our analysis has limitations. First, due to the cross-sectional nature of the survey data, we were able to identify correlates only, rather than true predictors, of perceived smoking cessation by vaping. Additionally, we were unable to determine if participants had maintained e-cigarette use after quitting smoking. These limitations should be addressed in longitudinal studies. Second, the cross-sectional data and the data-driven nature of machine learning did not permit us to draw any causal conclusions. Specifically, we were unable to attest if vaping experiences (reflected in the VES) constituted a causal factor of successful vaping-assisted smoking cessation, nor did we conduct any analysis to statistically rule out a reverse association between vaping experiences and improved smoking cessation outcomes. These limitations could be addressed by future researchers who apply a rigorous quasi-experimental method to a longitudinal dataset. Third, we did not have access to a biomedically verified smoking cessation measure and instead, relied on a self-reported quitting status. Additionally, rather than directly asking participants about their vaping-assisted quitting status (i.e., “did you quit smoking specifically using vaping for at least one week after your most recent quit attempt?”), we employed a more subjective measure by asking about their levels of perceived usefulness of using vaping to cut down cigarette smoking. Future studies can overcome this limitation by either including lab procedures to establish the abstinence outcome objectively or enhance the survey design by incorporating a direct and less subjective self-reported quitting measure. Forth, our sample size is small when compared to previous machine learning applications in health. Because the performance of machine learning models depends directly on the volume of data, the limited sample size may have impeded our ability to derive a better model. However, even with this small sample, our model reached an AUC of 0.865, indicating high discriminatory ability. Finally, the use of a convenience sample of adult smokers in Ontario, Canada may have reduced the generalizability of our findings to a broader population. Future studies should revisit this topic with larger representative samples. Our study provides several implications for future research. As we expand the current evidence on a potential link between the experiences of smokers during a vaping-assisted smoking cessation attempt and their quitting outcome, more statistical investigations are warranted to understand the direction and the strength of this association. Furthermore, although we showed our results to be largely robust to the small percentage of data (0.8%) that were absent from the dataset, future researchers that encounter more severely missed data might want to adopt a more sophisticated analytical pipeline that incorporates multiple imputations within the machine learning framework to generate robust findings. Next, we showed the gradient boosting machine outperformed the traditional logistic regression in predicting the status of vaping-assisted smoking cessation, which implies the utility of machine learning in devising predictive tools to assist smoking cessation efforts. Future studies need to explore the feasibility of deploying such machine learning-enabled tools in a real-world setting. Finally, with a surging interest in enhancing the transparency and the explainability of black-box machine learning models as well as advancing machine learning methods to gain the capability of causality testing (i.e., causal learning), future researchers might consider experimenting with these novel analytical techniques in studying the mechanism of a successful vaping-assisted smoking cessation. # 5 Conclusions Using machine learning, we identified five person-level characteristics—including vaping experiences (measured by a Vaping Experiences Score), number of quit attempts by vaping, age, having vaped 100 times and the time after waking up to vape—to be the top five predictors of the probability of perceived success in vaping-assisted smoking cessation. These results provide valuable implications for interventions aim at maximizing the effectiveness of e-cigarettes as a potential cessation device. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction The existence of adult neural stem cells in the rodent brain was reported in the 1960s, and by the early 1990s, they were successfully isolated, propagated and differentiated into neurons *in* *vitro*. Subsequent studies in the human brain also found neural stem/progenitor cells, herein referred to as adult human neural progenitor cells (AhNPCs), and like our mammalian counterparts, they were largely restricted to two neurogenic regions; the subventricular zone (SVZ) lining the lateral ventricle and the subgranular zone (SGZ) of the hippocampus. Under physiological conditions, AhNPCs of the SVZ and SGZ proliferate and migrate to the olfactory bulb and the dentate gyrus respectively to give rise to neurons and astrocytes. A number of reports also showed that AhNPCs were capable of being cultured *in vitro* from both autopsy and biopsy human brain tissue. Most of these studies utilized the neurosphere (NS) culture method, and provided evidence for self-renewal through 5-bromo-2-deoxyuridine (BrdU) incorporation studies and single AhNPC clonal expansion studies,. Furthermore, NS-derived cells differentiated into antigenically, and physiologically, identifiable neurons and glia, fulfilling the criteria for neural progenitor cells (NPCs) *in* *vitro*. More recently, techniques have been established for culturing AhNPCs as a monolayer, removing the laborious process of NS formation and allowing characterization of individual AhNPCs in culture dishes. These pioneering studies have addressed many important aspects of AhNPC cultures and have laid the foundation for future experiments. However, AhNPC culture studies are still faced with issues associated with reliable characterization and hence, experimental reproducibility. For example, most studies have found AhNPCs to only be present in the neurogenic regions, while some studies report their presence in classically non-neurogenic regions. There are also discrepancies in the literature regarding the mitotic capacity of AhNPCs *in vitro*, as some report limited (\<6 months) self-renewal, whilst others have reported successful long-term (\>12 months) cultures. Further complicating the interpretation of these previous reports is the large degree of heterogeneity observed with primary human brain cell cultures, including a fibroblast-like cell (FbC) population that also divides *in* *vitro*. It is possible that FbCs have contaminated cultures presented in previous publications, thus accounting for some of the discrepancies in the literature. It is also possible that these FbCs are, in fact, another source of NPCs in the brain. For these reasons, we studied the proliferating cell types grown from adult human brain tissue to compare their properties and determine whether FbCs and AhNPCs are two distinct populations or arise from a common source. Here, we demonstrate that the neurogenic regions of the adult human brain can give rise to two different populations of mitotic cells that have remarkable similarities during their proliferative states *in vitro*. However, AhNPCs were limited to the neurogenic regions and gave rise to physiologically active neurons, while the FbCs could also be isolated from the non-neurogenic cortical regions but failed to differentiate into mature neuronal cell types. Most importantly, we demonstrate the relatively limited proliferative ability of the AhNPCs compared to the FbCs and suggest a time frame in which AhNPC experiments should be conducted for accurate interpretation of *in vitro* experiments of adult human neurogenesis. # Results The initial isolation of cells from the adult human hippocampus (HP) and the surrounding SVZ resulted in a heterogeneous population of cells in culture. They comprised of post-mitotic astrocytes, neurons and microglia, and the mitotic AhNPCs and FbCs. Post-mitotic astrocytes were positive for glial fibrillary acidic protein (GFAP) and negative for the cell cycle marker ki67 ( A-C). Neuronal cells were βIII-tubulin positive and ki67 negative ( D-F) and microglial cells were CD45 and PU.1 positive ( G-I). The mitotic population could be divided into two groups based on their morphology and antigenicity. The first mitotic population consisted of GFAP and Nestin positive NPCs that also expressed ki67 ( A-F). These NPCs also incorporated BrdU and co-expressed PSA-NCAM and βIII-tubulin ( G-I). The second mitotic population was also Nestin and ki67 positive, but did not express GFAP ( J-L). Instead, they were immunoreactive for the collagen synthesizing enzyme prolyl-4-hydroxylase (P4H) and the extracellular matrix protein fibronectin ( M-P). Morphologically, they were generally flat and had a wide-spread cytoplasm, which together, led us to classify them as FbCs. Every region tested, including the SVZ, gave rise to neurons, microglia, astrocytes, and FbCs. However, AhNPCs could only reliably be cultured from the SVZ regions, with 10 out of the 11 cases resulting in identifiable NPCs ( K). AhNPCs were Nestin and GFAP positive ( B), proliferated as a monolayer on un-coated tissue culture flasks (semi- adhesive) and gave rise to spontaneous neurospheres (NSs; A). When differentiated, over 80% of these cases resulted in βIII-tubulin and GFAP positive cells ( K), indicating that the majority of NS forming cases show neurogenic potential. When NSs were dissociated into single cells and plated onto non-adhesive surfaces, they were capable of forming secondary and tertiary NSs ( C-D), although with diminishing efficiency. To induce neuronal differentiation, spheres were plated onto adhesive surfaces (PDL/Laminin) and medium was changed to a no or low serum containing media devoid of mitotic growth factors. Adherent NSs rapidly extended processes and cell migration soon followed ( E), and 35% of the cells expressed the immature neuronal marker βIII- tubulin ( F, 3 E). The same population was also immunoreactive for the mature neuronal marker MAP2 and the astroglial marker GFAP ( I–J), indicating that AhNPCs can give rise to at least two cell lineages of the CNS *in vitro.* Furthermore, intermediate cell types co-expressing βIII-tubulin and GFAP were observed ( G–H). The differentiated AhNPCs with neurite-bearing morphology from 5 separate cases were subjected to electrophysiological studies to confirm their functional differentiation. During the first week of differentiation, neuron-like cells exhibited a depolarized resting membrane potential (RMP) and a high membrane resistance (R_m) of −45 mV±3.2 mV and 654±210 MΩ (n = 7), respectively. When tested for active membrane properties, 2 of the 7 cells responded with a small depolarization ( D–F). Continued differentiation (3–4 weeks) resulted in RMP and R_m values sitting at near physiological levels of −65±2.1 mV and 195±50 MΩ (n = 10), respectively. Active membrane properties also matured with 3 out of the 10 cells (30%) firing an immature action potential (iAP) in response to the depolarizing stimuli (represented by A–C & G-I). These APs were accompanied by a relatively fast inactivating inward and a slow inactivating outward current. The same AP firing cells exhibited βIII-tubulin staining ( G1–4). The majority of the fiber-bearing cells elicited little or no active membrane responses to depolarizing currents or voltages ( J–L) and generally had a hyperpolarized RMP and a lower R_m. When separately grouped into iAP firing and non-firing cells, the iAP firing cells had an average RMP and R_m of −61.7±3.0 mV and 333±100 MΩ, respectively, while the non-firing cells showed glial-like properties with a RMP of −67.5±1.6 mV and R_m of 118±13 MΩ (P\<0.05). The cells that expressed both GFAP and βIII-tubulin ( J–L) did not fire an iAP but exhibited a slight inward current under large depolarizing stimuli. These results demonstrated that our AhNPC cultures gave rise to at least 2 brain cell populations *in vitro,* neurons and astrocytes. Greater than 60% of our cultures continued to proliferate for longer than 8–10 months (\>6 passages), formed NSs and expressed the NPC marker Nestin ( B). However, the ability of these cultures to generate neurons and astrocytes (their neurogenic ability) deteriorated with successive passages, reaching insignificant levels by 5–6 months (∼ 6 passages;). The morphological composition of the cultures also progressively changed, going from phase bright monolayer and sphere-dominated cultures (\<5 months) to relatively homogenous bipolar monolayer cultures with limited sphere formations (\>6 months). These long-term cultured cells showed strong resemblance to the FbCs reported previously by our laboratory, except for their bipolar morphology and the level of Nestin expression. To elucidate whether the differences in morphology and the stronger expression of Nestin was due to the NPC culture conditions, FbCs cultured in serum- containing culture conditions from the temporal cortex were cultured in our serum-free NPC proliferation media. NPC culture conditions resulted in the FbCs showing markedly different characteristics to those cultured in serum containing conditions. Immunostaining revealed that the NPC marker Nestin, although expressed at detectable levels in FbCs cultured in serum containing media, further increased during the 14 days of NPC media induction ( C & J). This increase was quantified using image analysis and q-RT-PCR and showed a 1.5 to 2 fold change in expression (P\<0.01; L). Furthermore, βIII-tubulin, which was not present at immunologically detectable levels in FbCs cultured in serum containing media ( B), increased significantly during the 2 weeks in NPC media ( I) with image analysis and q-RT-PCR data indicating a larger than 10 fold increase in expression (P\<0.01; K–L). Western blot results were consistent with the above ( M). NPC media-exposed FbCs also changed their morphology to resemble cells those from the AhNPC cultures with bipolar processes and phase-bright somas ( H). Furthermore, when the FbCs were cultured in NS forming conditions (i.e. plated onto non-adhesive surfaces at clonal densities in NPC media), a large number of cells formed primary spheres ( G) that were capable of forming secondary spheres from their dissociated cells. These results were confirmed in all of 5 trialed independent cases and demonstrated the close phenotypic resemblance between the FbCs and the cells present in our AhNPC cultures. However, like the cells found during the later passages (\>6) of our AhNPC cultures, the FbC-derived spheres were not capable of undergoing full neuronal or astrocytic differentiation, as they gave rise to cells expressing βIII- tubulin but not MAP2 or GFAP. They also failed to give rise to cells exhibiting neurophysiological properties with average RMP of the cells sitting at –6.5±6 mV and R_m 500±300 MΩ (n = 5), and failed to elicit any form of active membrane properties (data not shown). summarizes these findings, which indicate that the majority of the mitotic cells present in the later passages of a primary AhNPC cultures are likely to be the FbCs and not NPCs. To fully characterize the primary AhNPC cultures, the origin(s) of these brain- derived FbCs must be elucidated. To achieve this, various cell lineage markers were tested on the FbC cultures and compared to a fibroblast cell line derived from human lung (hLFb). The three most likely cell linages were tested; 1) mesenchymal stem cells (MSCs), 2) pericytes/blood vasculature cells, and 3) fibroblast cells. ICC and western blot analysis showed a low level of the mescenhymal/fibroblast cell marker Thy-1 in the FbCs compared to the hLFbs ( J-L), while a fibroblast-specific antigen, S100A4, was much higher in the FbCs ( G-I). A more detailed analysis of gene expression at the transcript level using q-RT-PCR (primers listed) indicated that FbCs strongly expressed genes that were associated with fibroblast cells originating from brain vasculature, such as, PDGFR-β and α-smooth muscle actin (α-SMA; M). In agreement with the ICC and western results, transcript levels for MSC-related genes were relatively lower in FbCs compared to hLFbs, while S100A4 was at least 10 fold higher ( M). Furthermore, immunohistochemical studies of adult human hippocampus sections used for our *in vitro* cultures also showed many of our FbC-markers (P4H, Vimentin, and α-SMA) staining the blood vessels, hence corroborating the likelihood that our FbCs are of blood vessel origin ( A–L). summarizes these results. # Discussion Recent developments have provided evidence for the existence of functional NPCs in the so-called neurogenic regions of the adult human brain, as previously described in rodents. By combining live cell imaging, immunocytochemistry and electrophysiological techniques, this study provided further evidence for the existence of adult NPCs in the human brain. More importantly, this study provided for the first time, evidence for the progressive changes in the cellular composition that occur during prolonged primary human NPC cultures. This area has been largely neglected, despite the fact that knowledge of the cellular composition of a heterogeneous cellular population is vital for the interpretation and reproducibility of experiments. Moreover, special emphasis was given to characterising the mitotic populations, as they are the main source of confusion regarding cell identity in primary NPC cultures. ## Primary Adult Human Brain Cell Cultures Give Rise to a Heterogeneous Population In Vitro Primary cultures isolated from the adult human brain gave rise to a heterogeneous population of brain-derived post-mitotic and mitotic cells. Initially, the post-mitotic population comprised mostly mature astrocytes and microglia, but neurons were also present in varying amounts. However, they became rapidly diluted out with subsequent passages due to their post-mitotic nature. This is in agreement with previous reports where post-mitotic cells could only be cultured for a short period of time. The mitotic cells, when cultured from the neurogenic regions (SVZ and hippocampus), comprised two distinct populations. ## Two Distinct Mitotic Populations Exist in AhNPC Cultures ### Adult human neural progenitor cells Cultures from the neurogenic regions gave rise to mitotic cells that could be classified into two types based on their morphology and immunostaining properties. The first population only arose from cultures containing the ventricular regions and the hippocampus. Morphologically, they were bipolar in shape and had phase-bright somas. These cells frequently formed spheres, incorporated BrdU and were immunoreactive for the proliferative marker ki67. The same cells co-expressed Nestin and GFAP, which are markers widely used as indicators for AhNPCs, as they are thought to arise from proliferative astroglial-like cells (type B and C cells of the SVZ and type I cells of the SGZ). Some co-expressed PSA-NCAM and βIII-tubulin, which might indicate a population that has already undergone asymmetric division to give rise to proliferative neuroblasts. *In vivo,* they represent the migrating cells predestined to give rise to granule cell neurons and interneurons in the hippocampus and the olfactory bulbs, respectively. As reported previously, spontaneous NSs arose during the earlier passages (P1– P4), and on many occasions, adhered back onto the monolayer of cells and started extending multiple neurite projections. As NS formation is a defining characteristic of neural stem cells *in vitro*, , we also tested for NS formation by sub-plating cells onto non-adhesive surfaces at every passage from monolayer cultures. Spheres frequently arose from all passages; however, in concordance with our monolayer cultures, only the spheres from earlier passages (\<P5) were capable of differentiating into cells of the CNS, and hence, only these could be classified as NSs. The NSs from earlier passages were capable of giving rise to tertiary spheres but not more. These results complement other studies that found adult NPCs *in vitro* had limited proliferative and differentiating ability. However, despite the loss of their multipotentiality and hence their NS status, their proliferative ability remained until at least passage 9 (12 months in culture; longest current culture time). These later passage cultures are likely to be predominantly comprised of FbCs and this will be discussed in further detail. Based on these observations, we highly recommend that NPC-based experiments be conducted between the 3–5 passage period. Interestingly, a couple of groups have published results in which AhNPCs were cultured for greater than 20 months (\>12 passages). These reports each used contrasting culture protocols and our protocol was more comparable with Ayuso- Sacido et al. where conventional serum-free medium supplemented with growth factors was used. Leonard et al. utilized a high serum containing, growth factor-free medium, which is widely used to culture primary microglia and astrocytes, and is commonly used by many to differentiate the NPCs. Our attempts to reproduce the high-serum culture conditions were unsuccessful as the majority of our cells developed into FbCs with very limited, if any, spontaneous sphere formation. The reasons for these differences are unclear. ### Brain-derived fibroblast-like cells (FbCs) The second mitotic cell population that co-exists with the AhNPCs, but is not only limited to cultures of the neurogenic regions is the FbC population. We have previously characterized FbCs cultured in serum containing culture conditions that were primarily designed for maintaining primary astrocytes and microglia. Here, the FbCs were highly proliferative, were morphologically large (\>50 µm) and flat, appeared phase-dark under phase-contrast microscopy, were highly immunoreactive for the collagen synthesizing enzyme P4H and the extracellular matrix protein fibronectin, and were negative for endothelial markers CD34 and Factor VIII. These fibroblast cell-like characteristics led us to classify these cells as FbCs. In our AhNPC cultures, cells with indistinguishable immunoreactive characteristics were identified and despite their morphological differences (more bipolar and phase-bright), they were classified as FbCs. In addition, the present study also found FbCs to express markers associated with NPCs, such as Nestin and Sox-2. Based on these observations, we tested: 1) whether the morphological differences were due to our serum-free NPC culture conditions, and 2) whether the FbCs were actually another form of NPCs *in vitro*. To test this, FbCs cultured in serum-containing conditions were cultured in AhNPC culture conditions for a period of 2 weeks. By removing serum and supplementing the media with B27, EGF and FGF-2, their morphology changed from a flat phase-dark to more rounded, phase-bright bipolar cells, closely resembling those FbCs seen in AhNPC cultures. As these bipolar FbCs still retained their P4H and fibronectin immunoreactivity, we concluded that the morphological differences were due to the different culture conditions. Interestingly, accompanying the morphological changes was a significant increase in the expression of the NPC marker Nestin and the early neuronal marker, βIII- tubulin. The small but significant increase in Nestin suggests that NPC conditions may be driving these FbCs further towards a precursor-like de- differentiated state. The concurrent up-regulation of βIII-tubulin came as a surprise as it represented a more neuronally committed cell line. One explanation is that the increase is due to the presence of the growth factors EGF and FGF-2, as they lead to increases in pro-neural gene expression in non- neuronal cell types. Indeed, βIII-tubulin expression was attenuated when the growth factors were removed from the media (data not shown). However, the same growth factors did not stimulate βIII-tubulin expression in the majority of AhNPCs (rather decreased it), suggesting that this is either a FbC-specific effect or that FbCs might also be trans-differentiating into cells of an early neuronal phenotype. In addition, when FbCs were cultured in non-adherent NS culture conditions, they also gave rise to clonal spheres that could be passaged into secondary spheres. This was surprising, as sphere-forming ability in our culture conditions has so far been largely associated with NSC/NPC, or cancer stem cells in others’ work. This is not the first time fibroblast cells have shown NPC-like characteristics *in vitro*, ; however, it is the first time these results have been described from cells derived from tissue of the adult human brain. Due to their brain tissue origin and the expression of NPC markers, FbCs have become an ambiguous cell type in NPC cultures, as illustrated by the fact that similar cells have been classified as both protoplasmic astrocytes and NPCs. Therefore, in order to further investigate the notion that FbCs could indeed have NPC-like properties, we compared their neurogenic capacity to AhNPCs under differentiating conditions. ## AhNPCs Differentiate into Immunologically and Physiologically Identifiable Brain Cells When AhNPC cultures of earlier passages (\<5 passages) were induced to differentiate, they gave rise to cells with distinct neuronal and astroglial characteristics. These cells were immunoreactive for neuronal βIII-tubulin, MAP2ab, and astroglial GFAP, but devoid of the proliferative marker ki67 and the NSC marker Nestin. We could not observe cells immunoreactive for oligodendrocyte marker RIP4. This was of no great surprise as others have also found oligodendrocyte differentiation to be the most challenging. As reported by numerous NPC studies, we also observed an intermediate cell type that was immunoreactive for two lineage-specific markers, βIII-tubulin and GFAP. Previous investigators have classified these cells as ‘asterons’ and described them as a trans-differentiated state of a neuron or an astrocyte. However, as we found a large percentage of asterons spreading out from the differentiating NSs and because neurons arise from GFAP positive cells *in vivo*, we suggest that asterons could be a population of semi-committed immature neurons. Differentiated NSs were studied using whole-cell patch-clamp techniques to test for functional differentiation. In agreement with previous reports, AhNPC- derived neurons gradually developed neurophysiological properties, with cells recorded at earlier time points (\<7 days) exhibiting immature passive membrane properties such as depolarized RMP and a relatively high R_m. They also showed very limited active membrane properties with the occasional cell showing a very immature ‘depolarizing hump’ in response to strong depolarizing stimuli. These ‘depolarizing humps’ are likely to be a product of outward K⁺ current activated by membrane depolarization, as blocking voltage- gated K⁺ channels abolished this signal. This K⁺ current is also observed in migrating neuroblasts of the RMS in adult rats, suggesting that it is channel activity characteristic of a developing and/or migrating neuron. By 3–4 weeks of differentiation, cells developed mature passive membrane properties with the average RMP and R_m at neurophysiological levels of −65±2.1 mV and 195±50 MΩ. Furthermore, 30% fired single immature action potentials (iAP) that were accompanied by fast-inactivating inward and slow- inactivating outward currents. These iAPs are likely to be caused by a combination of voltage-gated Ca²⁺ and not yet fully developed Na⁺ channels. The large outward currents seen in voltage-clamp are likely due to the pipette solution containing potassium gluconate instead of CsCl typically used to measure inward sodium currents. However, unlike previous reports, mature APs with typical short-lasting, overshooting waveforms could not be observed in these cultures, suggesting perhaps that the culture period of 3–4 weeks was not sufficient for complete neuronal maturation in our conditions. The possibility that iAPs were of astrocyte origin was ruled out on the basis that the astrocyte-like cells conveyed a significantly hyperpolarized RMP, lacked the ability to fire APs and labelled with GFAP, while the AP-firing cells labelled with βIII-tubulin. Therefore, this study convincingly demonstrated that the adult human brain SVZ and the dentate gyrus give rise to AhNPCs that could be propagated for up to 5 passages (∼ 6 months) without losing their bipotency to differentiate into functionally maturing neurons and astrocytes. ## FbCs Cultured in AhNPC Conditions do not Differentiate into Functional Brain Cells and therefore are not NPC by Nature Despite the large phenotypic overlap between the AhNPCs and FbCs during their proliferative states, FbCs failed to give rise to morphologically, immunologically and electrophysiologically identifiable neurons or astrocytes. Although the differentiated FbCs showed small neurite extensions and expressed βIII-tubulin, they failed to acquire more mature neuronal or astrocytic markers such as MAP2 and GFAP. Furthermore, they showed no electrophysiological resemblance to neuronally differentiated AhNPCs. This led us to believe that AhNPCs and FbCs are two phenotypically similar but functionally distinct populations of proliferative cells *in* *vitro.* ## Possible Neurovasculature Origin of the FbCs In depth characterization of the AhNPC culture system is vital in achieving a more reliable and reproducible experimental platform. Therefore, the origin(s) of FbCs were investigated based on the three most probable sources. First, mesenchymal stem/precursor cells from the blood, second, the perivascular/vascular cells in the brain parenchyma, and third, the fibroblast cells from the leptomeninges or the surrounding support tissue. The results showed that fibroblast-related genes in the FbCs, such as P4H, Vimentin and Nestin, were expressed at comparable levels to a control fibroblast cell line, reconfirming the highly fibroblast-like characteristics of these cells. In addition, our brain-derived FbCs highly expressed vasculature cell genes *PDGFRβ* and *α-SMA*, but minimally expressed mesenchymal stem cell genes, *THY1/CD90* and *CD73*. Furthermore, many of FbC markers were localized to the blood vessels in the hippocampal sections used for our AhNPC cultures, although markers such as P4H and Vimentin also labelled neuronal and astrocytic cells respectively. Overall, these data indicate a neurovasculature origin of these FbCs. In fact, a cell type strongly resembling our FbCs has been described from the adult human spinal cord. Although there are certain differences, such as their inability to form spheres under NPC culture conditions, when cultured under serum-containing conditions, these spinal cord cells behaved like our FbCs in that they proliferated as an adherent monolayer and expressed Nestin, Sox2, PDGFRβ and α-SMA. Furthermore, they also found these antigens to be localized to the blood vessels of the spinal cord. Neurovasculature contamination of brain tissue cultures is unavoidable and when taken together with the above results, it appears our brain-derived FbCs are likely to be of neurovascular origin. Therefore, investigators are urged to correctly identify the FbCs, as they can cause considerable confusion in AhNPC cultures. However, when identified correctly, they may provide another neuroectodermally primed cell line for research. ## Limited Proliferative Ability of Biopsy-obtained AhNPCs In Vitro In agreement with the majority of primary AhNPC culture studies, our results also demonstrated the limited proliferative ability and the relatively rapid loss of multipotentency seen with NPCs obtained from adult compared to those from embryonic or fetal brain tissue. This phenomenon can be explained by the fact that the number of ‘true’ neural stem cells are known to decrease during development and remain largely quiescent in adulthood until stimulated to proliferate. These ‘slow dividing’ cells proliferate at a much slower rate and give rise to the rapidly amplifying NPCs, which contribute to the majority of the dividing cells *in vivo* and the NS growth *in vitro*. In rodents, it has been reported that only the SVZ, and not the hippocampal dentate gyrus, is rich in ‘true’ NSCs. As the majority of our cultures were from the hippocampal regions that had only a limited area of the SVZ lining the inferior horn of the lateral ventricle, it is likely that our tissue contained predominantly rapidly proliferating NPCs and very little, if any, ‘true’ NSCs. Another factor to consider is the potential impact of epileptic pathology on the process of neurogenesis, as the majority of our cultured brain tissues were obtained from temporal lobectomies for the treatment of mesial temporal lobe epilepsy. At least in animal models, epileptic development can largely be divided into two phases; acute, which accounts for the first couple of months after the initial seizure generating insult, and chronic, which accounts for the period succeeding the acute phase of the disease. These distinct phases have very differing effects on neurogenesis. In general, the acute phase results in an increase in neurogenesis and neuronal differentiation. There is popular belief that this increase in neurogenesis leads to the generation of aberrant neuronal circuitry that is responsible for the generation of spontaneous recurrent seizures. However, during the chronic phase of the disease, studies generally report a decrease in neurogenesis, more specifically, in neuronal differentiation. Some studies associate this phenomenon with the deficits in hippocampal-associated learning tasks seen in animals and patients with chronic epilepsy. The majority of human studies rely on surgically obtained tissue, and since surgical intervention is usually the last mode of therapy, most cases can be regarded as chronic epilepsy. In agreement with animal studies, IHC analyses found no change or a decrease in adult hippocampal neurogenesis. As our cultures used the same tissue source as the study above, it is likely that our cultured hippocampal NPCs came from neurogenically-compromised tissue. This could account for the limited proliferative capacity and the relatively low rate of neuronal differentiation seen in not just our cultured NPCs, but also in other publications that use similar tissue sources. In summary, we have shown that neural progenitor cell cultures from the adult human brain generate a heterogeneous population of cells that gradually change in composition over several passages. We have demonstrated that there are two distinct mitotically active populations *in vitro,* the NPCs and the FbCs. Under NPC culture conditions, they showed remarkable phenotypic similarity; yet, they gave rise to very different cell types when differentiated. The NPCs gave rise to antigenically and functionally identifiable neurons and glia, while FbCs only expressed an immature neuronal marker. Therefore, it is clear that FbCs are not NPCs. However, FbCs, which might be from a neurovascular origin, do show a high degree of plasticity and a tendency towards a neuroectodermal lineage and these properties may benefit investigators by providing a readily accessible source for reprogrammable primary brain cells. As for the NPC cultures, unless new protocols are introduced that allow for the isolation and culture of ‘true’ NSCs from the adult human brain without genetic manipulation and increased cell viability, we recommend avoiding the earlier passages (\<2, due to the residual primary astrocytes and neurons) and the later passages (\>5, due to the FbCs) for conducting neurogenesis research. We believe these guidelines will allow investigators to conduct more reliable and reproducible AhNPC-based experiments and advance the field towards establishing a standardized adult human brain cell culture system. # Materials and Methods For detailed methods regarding many of the methodologies used in this study, please refer to. ## Tissue Dissociation and Culture Biopsy human brain tissue containing the anterior temporal lobe and the hippocampus were obtained from surgery for medically refractory epilepsy. All specimens were collected with written patient consent and ethical approval from the Northern X Ethics Committee (biopsy tissue) and the University of Auckland Human Participants Ethics Committee. Eleven biopsy specimens (mean age 45 years) were collected from the surgical theatre at the Auckland City Hospital and transported to our laboratory in Ca²⁺ and Mg²⁺ free ice- cold Hank’s balanced salt solution (HBSS; Gibco). Tissue containing the periventricular zone and the hippocampus was dissected and dissociated prior to being digested in HBSS containing 2.5 U/mL papain (Worthington) and 100 U/mL DNase 1 (Invitrogen) for 20 minutes at 37°C with gentle rotation, which included a gentle trituration step at 10 minutes. Enzymatic digestion was halted by the addition of NPC proliferation media; DMEM:F12 containing B27 (Invitrogen), Penicillin/Streptomycin (Gibco), GlutaMAX (Invitrogen), 40 ng/mL FGF-2 (Peprotech), 40 ng/mL EGF (Peprotech) and 2 µg/mL Heparin (Sigma). Cells were collected by centrifugation (170 *g*×10 minutes), resuspended in the NPC proliferation media and plated onto un-coated T25 culture flasks (Nunc). The following day, culture flasks were gently agitated to detach any loosely adhered cells and all the media was collected and replated onto a fresh T25 culture flask. Media was half changed every 2–3 days and cultures were serially passaged every 20–30 days. Out of the 11 specimens, 9 showed sustained growth (\>3 passages) in NPC proliferation media and were used for experimentation. Human lung fibroblasts were cultured as previously published as a monolayer on un-coated tissue culture flasks in DMEM:F12 base medium supplemented with 10% fetal bovine serum (Gibco), Penicillin/Streptomycin and GlutaMAX. ## Generation of Primary and Secondary Spheres Adherent cultures frequently formed spontaneous spheres during expansion. To further promote sphere formation, cells were plated at 50,000 cells/mL in 100 mm bacteriological grade culture dishes (Falcon) in NPC proliferation media. Primary spheres generally formed within 1 week and continued to grow in size and number. Secondary and tertiary spheres were generated by dissociating the primary spheres into smaller spheres by trituration and replating them back into new culture dishes. ## Differentiation of NPCs into Neurons and Astrocytes Whole spheres or single cells dissociated from the spheres were plated onto PDL/Laminin (Sigma) and the culture medium was changed from the NPC proliferation to the differentiation medium (DMEM:F12 containing 1% FBS, 40 ng/mL NGF and BDNF (Preprotech)). ## Immunocytochemistry Cells were fixed in 4% paraformaldehyde (PFA) for 20 minutes at room temperature and permeabilized by 3×10 minute washes in PBS containing 0.1% Triton-X (PBS-T). Antibodies were dissolved in immunobuffer comprising PBS-T with merthiolate and 1% normal goat serum. Cells were stained with primary antibodies listed in and incubated overnight at 4°C, then visualized after a 3-hour room temperature incubation with species-specific fluorescent (Alexa 488 or 594; Invitrogen) or biotin-conjugated (Sigma) secondary antibody. DAB immunoprecipitation was used to visualize biotin-conjugated secondary antibodies and all nuclei were counterstained with 20 µM Hoechst 33258 (Sigma). ## Immunohistochemistry Protocols from Waldvogel et al. were used for handling and processing of donated brain tissue. Briefly, the paraformaldehyde fixed tissue blocks were coronally sectioned into 50 µm sections and processed for immunohistochemistry as free- floating sections. In cases where antigen-retrieval was necessary, the sections were microwave-boiled in citrate buffer (pH 6.0) prior to being incubated for 3 days at 4°C with the antibodies listed in. Antibody binding was visualized using techniques described for immunocytochemistry section above and in Waldvogel et al.. ## BrdU Assay 10 µM BrdU (Roche) was added to the cultures 24 hours prior to the completion of an experiment. Cells were fixed in ice-cold methanol for 15 minutes at 4°C followed by a 45 minute incubation in 2 M HCl at 37°C. Wells were subsequently neutralized by washes in 0.1 M borate buffer pH 8.5 and PBS. Primary BrdU (1∶500; Roche) antibody was dissolved in PBS with 1% bovine serum albumin and incubated with cells overnight at 4°C. Primary antibody visualization was identical to that used for immunocytochemistry. ## Western Blot Cell culture medium was removed and the cells were washed twice in ice-cold PBS. Protein lysates were prepared and western blots were performed by protocols previously described. ## Quantitative RT-PCR Target gene expression levels were evaluated by quantitative RT-PCR using a 7900HT Fast Real Time PCR system (Applied Biosystems, Singapore). Total RNA was isolated at designated time points using RNeasy kit (Qiagen Inc.) and stored at −80°C until further use. cDNA synthesis was performed with SuperScript III first strand synthesis kit (Invitrogen) using approximately 3 µg of DNase I-treated (Promega) RNA, and subjected to q-RT-PCR using Platinum SYBR Green qPCR SuperMix-UDG with Rox kit (Invitrogen). The primers are detailed in and the relative changes were analysed according to the ΔCT method. Each PCR run included a negative RT and non-template control, as well as melting curve assays to confirm specific product amplification. The plotted data represent the mean values of at least 3 independent experiments ± standard error. ## Electrophysiology Subsets of NSs were plated onto PDL/Laminin coated coverslips (Menzel-Glaser) and whole-cell patch-clamp technique was used to examine neurophysiological properties of individual cells as previously described. Coverslips were placed in a recording chamber of an upright microscope (Olympus BX51) and perfused with artificial cerebral spinal fluid (150 mM NaCl, 4 mM KCl, 2 mM CaCl₂, 2 mM MgCl₂, 10 mM HEPES and 10 mM Glucose) at room temperature. Patch pipettes were fabricated from thick-walled borosilicate glass with an average resistance of 4–6 MΩ. Internal solution consisted of 120 mM K-Gluconate, 40 mM HEPES, 5 mM MgCl₂, 2 mM NaATP, 0.3 mM NaGTP, pH 7.2 with KOH and Alexa 594 dye (Invitrogen) for post-patch identification of the cells. During the patch-clamping procedure, cells were visualised by infrared DIC microscopy. Current and voltage-clamp recordings were made using a Multiclamp 700A amplifier (Axon Instruments). Offline analysis was performed with Axograph 8 software (Axograph Scientific). In voltage clamp, fibre-bearing cells were held at a holding potential of −60 mV. Membrane resistance (R_m), access resistance and cell capacitance were measured in voltage-clamp mode in response to a square depolarizing pulse of 20 mV at 100 Hz. In current-clamp mode, resting membrane potentials (RMPs) were recorded at zero holding current and membrane responses were examined by multiple current steps of 30 pA at 1 Hz moving the cell from −30 pA to 150 pA. Under voltage-clamp, active membrane properties were tested by applying 10 mV steps at 1 Hz, taking the membrane potential of the cell from −70 mV to 20 mV. The results are presented as a mean ± SEM for all the recorded membrane properties and passive membrane properties were statistically tested by using independent sample *t-*tests for a significance value of P\<0.05. The recorded coverslips were fixed as described in the ‘immunocytochemistry’ section above and the antigenicity of the recorded cells was visualized using fluorescent microscopy (Leica DMIRB, Germany). # Supporting Information We thank the patients and their families for their generous gift of tissue without which this work would not be possible. We also thank Lynair Roberts at the Auckland City Hospital for her administrative support and Amy Smith and Claire Lill for technical assistance. [^1]: Conceived and designed the experiments: TIHP EWM PSB RLMF MAC MD. Performed the experiments: TIHP HM. Analyzed the data: TIHP HM JMM MAC MD. Contributed reagents/materials/analysis tools: EWM PSB HHT RLMF MAC MD. Wrote the paper: TIHP MAC MD. [^2]: Dr. Teoh is a pathologist at Labtests, New Zealand, and he was the neuropathologist in charge of the authors′ epilepsy specimens used in this study. His pathological reports were part of the Auckland District Health Board’s regulation for epilepsy surgeries, and these reports were also vital in deciding which cultures will proceed in this study. As the patients fully consented the use of their tissue and their pathological reports for the authors’ research, there was no conflict of interest in this collaboration regarding data used in our studies. Furthermore, there was no commercial interest involved in sharing of the authors′ data. Therefore, this does not alter the authors’ adherence to all the PLoS ONE policies on sharing data and materials.

# Introduction Immune-related pancytopenia (IRP) is a form of bone marrow failure caused by hematopoiesis-suppressing autoantibodies. The clinical characteristics of patients with IRP include cytopenia or pancytopenia, normal or high reticulocyte (Ret) and neutrophil counts, hypoplastic or hyperplastic bone marrow, increased erythroblastic islands (EIs), and erythroid hyperplasia. Redundant autoantibodies on bone marrow mononuclear cells (BMMNCs) can be detected in the bone marrow by BMMNC–Coombs test or flow cytometry (FCM). EIs are distinct anatomical units that consist of a central macrophage surrounded by a ring of developing erythroblasts, and are considered functional units of definitive erythropoiesis. Macrophages and erythroblasts maintain the integrity of EIs through several adhesion molecules. Macrophages participate in the terminal maturation of erythroblasts by functioning as adhesion molecules, providing nutrients, transmitting proliferative and survival signals, and inducing the enucleation of erythroblasts. However, these macrophage functions have not been directly confirmed. IRP patients have more EIs in their bone marrow, which suggests that erythropoietic cells undergo hyperplasia in the bone marrow. These patients also experience varying degrees of anemia, which are worse in severe cases. In addition, more erythroblasts are phagocytosed by macrophages in the bone marrow of IRP patients with increased EIs. Immunosuppression therapy with prednisolone and human immunoglobulin decreases the autoantibodies in BMMNCs to undetectable levels. The number of EIs return to normal, and the phagocytosis of erythroblasts by macrophages ceases. Moreover, the degree of anemia is ameliorated. Information on the ineffective hematopoiesis of increased EIs in IRP patients is limited. This study aimed to elucidate the significance of increased EIs in the bone marrow of IRP patients. Thus, we investigated the erythroblast autoantibodies, biological characteristics of macrophages, and clinical characteristics of IRP patients. # Materials and Methods ## Ethics This study complied with international regulations concerning the ethical participation of volunteers. The protocol used in this study and signed informed consent forms of the participants were approved by the Ethics Committee of the General Hospital of TianJin University, China. ## Patients The participants were diagnosed with IRP according to the published criteria. Patients who suffered severe systemic infections that could not be controlled by antibiotics or those with viral infections, such as hepatitis viruses, cytomegalovirus, human parvovirus B, or Epstein–Barr virus, were excluded from the study. The following diseases were also ruled out. Myelodysplastic syndrome was excluded by a morphological assay, BMMNC culture, karyotyping, conventional cytogenesis and fluorescence in situ hybridization, as well as genetic and immunophenotyping tests. CD55 and CD59 were detected in the granulocytes and bone marrow cells of patients using FCM. A flare assay was performed to exclude patients with paroxysmal nocturnal hemoglobinuria. Patients with hemophagocytic syndrome (HPS) were excluded based on the following clinical characteristics: persistent high fever; enlarged lymph nodes, spleen, and liver; high serum lactic dehydrogenase (\>1000 U L⁻¹) and ferritin; high levels of liver transaminase; and hyperlipidemia. Chromosomal breakage was analyzed to rule out Fanconi anemia among patients manifesting clinical characteristics compatible with this diagnosis. All patient samples were obtained before therapy or after recovery and discontinuation of therapy for more than six months. ## Evaluation of Responses to Treatment Patients who participated in the treatment demonstrated increased blood count within three months of treatment. The quality of their responses was analyzed at three and six months of treatment according to the following criteria. (1) Essential cure was observed when: the hemoglobulin levels of male patients reached 120 g L⁻¹ and those of female patients reached 110 g L⁻¹, their granulocyte counts reached ≥1.5×10⁹ L⁻¹, their platelet counts reached ≥100×10⁹ L⁻¹ without suffering anemia and hemorrhage; and no relapse was observed during the one-year follow-up period. (2) Remission, defined as the absence of anemia and hemorrhage, was characterized by blood cell levels reaching the baseline (hemoglobulin count of 120 g L⁻¹ for males and 110 g L⁻¹ for females), an increase in granulocyte count of at least 3.0×10⁹ L⁻¹, and an increase in platelet count of at least 30×10⁹ L⁻¹ within one year. (3) Remarkable improvement, defined as the alleviation of anemia and hemorrhage, was characterized by an increase in hemoglobulin count of at least 30 g L⁻¹ for at least three months, and the patient no longer requiring blood transfusion. (4) No response was defined as the patient not showing any improvement after completing the treatment. ## Treatment Protocol All IRP patients were treated with immunosuppressive agents, namely, glucocorticoid alone or high-dose immunoglobulin (HDIVIG). Prednisolone was administered to all patients at a single oral dose of 0.5 to 1.0 mg kg⁻¹ body weight d⁻¹. Human immunoglobulin was infused at 0.4 g kg⁻¹ body weight d⁻¹ for 1 d to 5 d, followed by 10.0 g once a week for one to three months. The patients were prescribed supportive care, including component transfusion, liver protection, jaundice therapy, calcium supplementation, and suppression of gastric acid secretion. Smears of the initial bone marrow aspirates were examined under Wright's stain and cytochemical stains. The peripheral blood cell counts of patients were measured using an automatic hemocyte analyzer. The severity of anemia was stratified based on the hemoglobin concentrations: mild anemia, ≥90 g L⁻¹; moderate anemia, 60 g L⁻¹ to 90 g L⁻¹; severe anemia, 30 g L⁻¹ to 60 g L⁻¹; and very severe anemia, ≤30 g L⁻¹. Counts of serum bilirubin, triglycerides, ferritin, fibrinogen, and albumin were determined using routine laboratory methods. ## Immunofluorescence (IF) Assay Cells were fixed in ice-cold ethanol and washed three times with PBS for 3 min. The cells on the cover slips were blocked with 0.1% Triton X-100 and H₂O₂ for 6 min, rinsed with distilled water, and washed three times with PBS for 3 min. The cells were then blocked with rabbit serum for 10 min. The residual serum was removed, and the remaining content was incubated overnight with FITC-conjugated anti-hIgG at 4°C. The cells were rinsed three times with PBS for 3 min each, and stained with hematoxylin for 1 min. The stained cells were washed with diluted ammonia water and rinsed two times for 3 min. The prepared specimen was sealed with glycerol and examined under a fluorescence microscope. ## Macrophage Phagocytosis of Chicken Red Blood Cells (CRBCs) ### 1. Preparation of macrophages Bone marrow aspirates (5 mL) were collected in the evacuated tubes containing heparin as an anticoagulant. The BMMNCs were separated from the bone marrow aspirates using 75% Ficoll–Hypaque suspended in RPMI-1640 (Gibco-BRL, Grand Island, NY, USA) without sterile fetal bovine serum (FBS) in a 75 mL culture bottle. The BMMNCs were then incubated for 4 h at 37°C under a saturated 5% CO₂ atmosphere. The macrophages that adhered onto the bottle wall and the suspended non-adherent cells were discarded. The adherent cells were digested with 0.25% pancreatin containing 0.01% Ethylene Diamine Tetraacetie Acid (EDTA). The cells were cultured in LG-DMEM culture solution containing 10% FBS. The supernatant was removed and the cells were harvested using 0.2% trypsase solution. The BMMNCs were resuspended in RPMI-1640 and then stored for future use. ### 2. Preparation of CRBCs Venous blood (1 mL) was drawn into a sterile syringe containing Alsever's solution. The CRBCs were washed with sterile saline three times to remove the white blood cells and platelets. ### 3. Macrophage phagocytosis of CRBCs The BMMNCs were infused with 1 mL of CRBCs and incubated for 1 h (shaken every 15 min). After the cell suspension was centrifuged at 10,000 rpm for 10 min and 1000 rpm for 5 min, the supernatant was removed. Cell smears were prepared by fixing the cells with methanol/formaldehyde solution and staining them with Giemsa. A total of 100 macrophages were examined under a microscope. The phagocytic ratio and index were calculated as follows: phagocytic ratio  =  number of macrophages that phagocytosed CRBCs/total number of macrophages; phagocytic index  =  number of CRBCs phagocytosed by macrophages/total number of macrophages. ## Preparation of Bone Marrow Samples and FACS Analysis Bone marrow aspirate (2 mL) was collected in a heparinized tube and passed through a 0.2 µm membrane filter (Nuclepore) to remove cellular debris. Approximately 50 µL of bone marrow aspirate was transferred to a 5 mL tube, and incubated with monoclonal antibodies. FITC/PE/APC–conjugated GlycoA/CD15/CD34 (20 µL; Becton Dickinson) was added into the aspirate, and the mixture was incubated for 30 min at 2°C to 8°C. RBC lysis solution was used to eliminate the RBCs from the whole bone marrow aspirate. The cell suspension was centrifuged at 1500 rpm for 5 min. The resulting supernatant was discarded. PBS buffer (5 mL) was added, and the resulting mixture was centrifuged at 1500 rpm for 5 min. The supernatant was discarded, and the cells were washed two times. The cells were resuspended in 200 µL of PBS buffer and fixed with formaldehyde for FACS analysis. The cells were placed in a separate tube as the control sample, and treated with FITC/PE/APC–labeled mouse IgG1 antibodies. Approximately 50,000 events were selected in the Acquisition tab of CellQuest software to record the field of specimen tubes in FACSArial (Becton Dickinson). ## Statistical Analysis Measurement data, including the counts of Ret, EI, hemoglobulin, TBIL, DBIL, erythroid in the bone marrow, and macrophages, were presented as means ± standard deviations and analyzed using ANOVA. The differences in IgG/IgM- positive rates and treatment efficiency between groups were analyzed using chi- square test. Differences with p\<0.05 were considered significant. Statistical analysis was performed using SAS PLINK v.1.07. # Results ## Patient Characteristics A total of 81 patients diagnosed with IRP who were admitted into the Department of Hematology of Tianjin General Hospital from August 2008 to December 2011 were enrolled in this study. The ages of patients ranged from 4 years to 74 years, and the median age was 29 years. Among the 81 patients, 39 were male and 42 were female. Twenty patients (13 males and seven females) had severe aplastic anemia (SAA) but were left untreated (aged 14 years to 62 years with a median age of 33 years). Fifteen healthy volunteers (aged 11 years to 76 years) were included as controls (nine males and six females). The groups did not significantly differ in age and sex distribution. ## EIs in the Bone Marrow Aspirates of IRP Patients EIs accompanied by macrophage-phagocytosed erythroblasts were easily observed in the bone marrow aspirates of IRP patients. The IRP patients, SAA patients, and healthy controls had 7.81±3.78, 0.10±0.31, and 4.00±1.46 EIs, respectively, in every bone marrow aspirate smear (2.0 cm×1.5 cm). The IRP patients, SAA patients, and healthy volunteers significantly differed in terms of EI count (p\<0.0001;). ## Ret in the Peripheral Blood of IRP Patients The IRP patients did not significantly differ from the healthy controls in terms of the percentage of Ret in the peripheral blood (1.46±0.60 vs. 1.23±0.45) (p = 0.113;). The percentage of Ret in the SAA patients was 0.15±0.13, which was lower than those of the IRP patients and healthy controls (p\<0.0001). ## IF Detection of IgG Autoantibodies in the Bone Marrow Aspirate Smears in IRP Patients Intense green fluorescence was observed in the EIs of IRP patients, particularly between the macrophages and erythroblasts. This result indicates that the IgG autoantibodies aggregated in the EIs, specifically at the junction between macrophages and erythroblasts. The results show that 30 of the 81 IRP patients were positive for IgG autoantibodies in their EIs (positive IF of EIs). By contrast, no IgG autoantibodies (negative IF of EIs) were observed in the EIs of the SAA patients and healthy controls. ## Clinical Characteristics of IF-Positive EIs in IRP Patients The IF-positive IRP patients had significantly more EIs than the IF-negative IRP patients and healthy controls (11.13±3.20 vs. 5.86±2.53, p\<0.0001; and 11.13±3.20 vs. 4.00±1.46, p\<0.0001, respectively). The EIs of the IRP patients decreased after treatment. The number of EIs in the IRP patients with IF- positive EIs before treatment was significantly higher than that after treatment (11.13±3.20 vs. 4.33±1.86, p\<0.0001), but did not significantly differ between the treated IRP patients and healthy controls (4.33±1.86 vs. 4.00±1.46, p = 0.516). The 30 IRP patients with IF-positive EIs (22 with severe anemia and eight with moderate anemia) had a median hemoglobulin concentration of 54.87±8.0 g L⁻¹. The 51 IRP patients with IF-negative EIs (16 with severe anemia and 26 with moderate anemia) had a median hemoglobulin concentration of 63.63±8.31 g L⁻¹. The percentage of Ret in the IRP patients with IF-positive EIs was significantly higher than that in the IRP patients with IF-negative EIs (1.91±0.64 vs. 1.19±0.37, p\<0.0001). During the detection of autoantibodies on BMMNCs by FACS, the FSC/SSC gate attracted the populations of nucleated erythrocytes, stem cells, and granulocytes. The fluorescence data from events in the gate were analyzed to determine the percentages of various subpopulations (CD34⁺, CD15⁺, and GlycoA⁺). The IgG autoantibodies of the 30 IRP patients with IF-positive EIs tested positive for IgG autoantibodies (CD34 IgG, CD15 IgG, or GlycoA IgG). GlycoA IgG was detected in 29 of the 30 patients. Although 38 of the 51 IRP patients with IF-negative EIs also showed IgG autoantibodies, only two were positive for GlycoA IgG. No autoantibodies (IgG/IgM) were detected in all the control groups. The IRP patients with IF-positive EIs showed high levels of total serum bilirubin (22.34±5.92 µmol L⁻¹) and indirect bilirubin (16.03±4.40 mmol L⁻¹), which were higher than those of the IRP patients with IF- negative EIs (16.75±3.99 µmol L⁻¹ and 10.11±1.62 µmol L⁻¹, respectively). The percentage of erythroid lineage cells in the bone marrow aspirates of IRP patients with IF-positive EIs were significantly higher than that of IRP patients with IF-negative EIs (34.11±5.22 vs. 24.64±4.95, p\<0.001). This result suggests that compensatory erythroid hyperplasia occurred in IRP patients. The IRP patients with IF-positive EIs had more macrophages in their bone marrow aspirates (0.62±0.02 of IF was activated) than the IRP patients with IF-negative EIs. The IRP patients with IF-positive EIs had more activated macrophages than the IRP patients with IF-negative EIs (0.62±0.02 vs. 0.52±0.07, p = 0.004) and healthy controls (44.36±5.09 vs. 34.85±7.47, p\<0.0001). Phagocytosis of CRBCs improved the biological function of macrophages. Moreover, the phagocytic ratio and index were significantly higher among the IRP patients with IF-positive EIs than those among the IRP patients with IF-negative EIs (40.77±8.43 vs. 35.25±7.34, p = 0.004; 0.73±0.09 vs. 0.58±0.12, p\<0.0001, respectively). ## IRP Patient Responses to Treatment All patients received the glucocorticoid. Moreover, 22 cases of IRP patients with IF-positive EIs and 34 cases of IRP patients with IF-negative EIs simultaneously received HDIVIG. Three months after the initial treatment, the IRP patients with IF-positive EIs exhibited a total treatment efficiency of 56.7% positive IF of EIs (6.7% experienced essential treatment, 13.3% experienced remission, and 36.7% experienced remarkable improvement). By contrast, the IRP patients with IF-negative EIs exhibited lower treatment efficiency (p\<0.05). The IRP patients with IF-positive EIs exhibited significantly improved total efficiency (86.7% vs. 62.75%) compared with the IRP patients with IF-negative EIs. # Discussion In 2000, Shao first detected autoantibodies in the BMMNCs (bone marrow hematopoietic stem cells, nucleated erythrocytes, and granulocytes) in IRP patients using a BMMNC–Coombs test. These autoantibodies are produced by hyperfunctional B lymphocytes. IgG autoantibodies activate macrophages, thereby causing the phagocytosis of hematopoietic cells. IgM autoantibodies activate the complement system, which causes hematopoietic cell lysis. These autoantibodies bind to membrane functional antigens, such as erythropoietin receptors. These patients are diagnosed with IRP, which is also known as “positive BMMNC–Coombs test pancytopenia”. The IRP patients had more EIs in their bone marrow than the healthy controls. Whether the increase in EIs in these IRP patients are relevant to autoantibodies should be confirmed. Thus, we detected autoantibodies in EI. In the beginning of this study, laser confocal microscopy was used to observe the result of autoantibodies. The preliminary experimental results were not ideal even though FITC-conjugated goat anti-human IgG and DAPI-stained cell nuclei were observed. However, the nucleic area of each nucleated cell exhibited bright blue fluorescence, and the cell contour, nucleus structure and contour, and nucleus/plasma ratio were not evident. Thus, single nucleated cells (immature granulocytes, nucleated erythrocytes and lymphocytes, macrophages, or monocytes) were harder to identify. Furthermore, the EIs (nucleated erythrocytes at various developmental stages surrounding macrophages) and aggregated multiple nucleated cells (aggregated structure of nucleated erythrocytes, immature granulocytes, or mixed immature nucleated cells) were difficult to distinguish. Other novel techniques were used to stain the nuclei to distinguish the single nucleated cells, and verify that this technique did not affect the results of the IF assay. The cell nuclei were stained using Wright–Giemsa. However, the autofluorescence of Wright–Giemsa was so intense that it interfered with the results of the IF assay. We then stained the nuclei with hematoxylin, which obtained surprisingly good results. The green fluorescence of the FITC-labeled autoantibodies was visible, allowing us to distinguish the nucleated cells and identify the EIs, which consisted of nucleated erythrocytes at various stages that surrounded macrophages. Hematoxylin exhibited spontaneous fluorescence that did not interfere with the results of the study. Hematoxylin stained the nucleus but not the cell membrane. However, the FITC-labeled IgG autoantibodies were localized on the cell membrane of nucleated erythrocytes and/or membrane junction between the macrophage and nucleated erythrocytes. The cells were clearly distinguishable because of the significant difference in position and cell morphology. Thus, hematoxylin autofluorescence did not interfere with the results of our study. The IF micrographs showed the phagocytosis of erythroblasts by macrophages, not the overlapping structures. If overlapping structures were present, green fluorescence was observed in the cytoplasm or nucleus or exhibited a disorderly distribution. IgG autoantibodies were detected in the EIs of IRP patients, and these IgG autoantibodies primarily aggregated at the junction between macrophages and erythroblasts in the EIs. This result suggests that IgG autoantibodies functioned as adhesion factors between erythroblasts and macrophages in the EIs. We analyzed the clinical characteristics of the IRP patients with IF-positive EIs to understand the effects of EIs with IgG autoantibodies on the pathogenesis of IRP. The total number of EIs in bone marrow aspirate smears of the IRP patients with IF-positive EIs was significantly higher than those of the IRP patients with IF-negative EIs (11.13±3.20 vs. 5.86±2.53, p\<0.0001) and healthy controls (11.13±3.20 vs. 4.00±1.46, p\<0.0001). These EIs were also accompanied by macrophage-phagocytosed erythroblasts. HPS was not observed if the lymph nodes, spleen, and liver were not enlarged. Hyperlipidemia, high ferritin level, and persistently high fever were also not observed. IF detected IgG autoantibodies (CD34/CD15/GlycoA in all 30 patients and GlycoA IgG in 29/30 patients) in the erythroblasts of BMMNC-Ab of the IRP patients. Although IgG autoantibodies were also detected in the IRP patients with IF-negative EIs, these patients had less GlycoA IgG in their erythroblasts. Extravascular hemolysis was also observed in the IRP patients with IF-positive EIs, which resulted in severe anemia, increased percentage of Ret in the peripheral blood, increased number of erythroid lineage cells in the bone marrow, and increased TIBIL and IBIL levels. However, the Coombs test of the peripheral blood was negative, which excluded the possibility of autoimmune hemolytic anemia. This result indicates that erythroid hemolysis occurred in the bone marrow. We also determined the biological characteristics of macrophages in the bone marrow of IRP patients to understand the molecular interactions between erythroblasts and macrophages in the EIs. The IRP patients with IF-positive EIs had more macrophages in their bone marrow than the IRP patients with IF-negative EIs (34.11±5.22 vs. 24.64±4.95, p\<0.0001). Most of these macrophages were activated (0.62±0.02). We examined the biological function of macrophages using an in vitro CRBC co-culture. The IRP patients with IF-positive EIs exhibited a higher ratio and index of macrophages than the IRP patients with IF-negative EIs. Thus, macrophages were activated by several factors that enhanced phagocytosis. The macrophages in the bone marrow were possibly activated by IgG autoantibodies, particularly GlycoA IgG on the erythroblasts. The Fc receptors on the macrophages could bind to the Fc regions of the erythroblast IgGs. Thus, the erythroblasts with IgG autoantibodies were recruited and surrounded by activated macrophages to form a morphologic EI structure. The IgG autoantibody-mediated activation of the macrophage-phagocytosed adjacent erythroblasts caused extravascular hemolysis. These EIs with the IgG autoantibody were possibly caused by early biological processes, in which erythroblasts with the IgG autoantibody were phagocytosed by activated macrophages, not the erythropoietic niches. We also investigated the responses of the IRP patients with IF-positive EIs to HDIVIG or glucocorticoid treatments. The IRP patients with IF-positive EIs exhibited higher total treatment efficacies after three and six months of treatment than the IRP patients with IF-negative EIs. BMMNC-Ab was inhibited and removed after glucocorticoid treatment. The treatment also removed the autoantibodies in the EIs of IRP patients and decreased the number of morphologic EIs in the bone marrow smears, and the treated IRP patients did not significantly differ from the healthy controls (4.33±1.86 vs. 4.00±1.46, p\<0.0001). The treatment also removed the macrophage-phagocytosed erythroblasts, which implies that EIs with IgG autoantibodies consisted of a central activated macrophage and a ring of erythroblasts with IgG autoantibodies. The formation and integrity of the EI structure was mediated by the IgG autoantibodies. Blocking the FcR receptor on macrophages with HDIVIG or inhibiting autoantibody production through glucocorticoid treatment resulted in the following: (1) formation and maintenance of EIs that lack the essential adhesion factors (IgG and/or FcR); (2) non-activation of macrophages, which did not recruit erythroblasts; and (3) disappearance of EIs with IgG autoantibodies. These results show that the morphologic EIs in the bone marrow of IRP patients were not all erythropoietic sites. EIs with IgG autoantibodies in IRP patients were sites wherein erythroblasts were phagocytosed by macrophages. These autoantibodies could function as adhesion factors that attach erythroblasts to macrophages to form EIs. Thus, these EIs were the early biological markers of bone marrow failure in the pathogenesis of IRP. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: Z-HS RF Y-HW. Performed the experiments: Y-HW RF. Contributed reagents/materials/analysis tools: S-WD HL. Wrote the paper: Y-HW.

# Introduction Dengue viral infections have become one of the most important mosquito borne viral infections in the world and is one of the major emerging infectious diseases. In the past fifty years, its incidence has increased 30-fold with significant outbreaks occurring in five of six WHO regions. It is estimated that 2.1 million cases of dengue haemorrhagic fever (DHF)/dengue shock syndrome (DSS) occur every year resulting in 21,000 deaths. Dengue viral infections may be caused by any of the four dengue virus serotypes (DEN1–4) which are closely related. Initial infection with a particular serotype is known as primary infection, which is usually asymptomatic or results in mild disease manifestations. However, subsequent infection with other serotypes (secondary dengue infections) may lead to severe disease which manifests in the form of DHF/DSS. Currently, the pathophysiology of dengue viral infections and factors that result in severe clinical disease is poorly understood. Cross reactive memory T cells and cross reactive antibodies have been suggested to contribute to immunopathology by altering the cytokine profiles during secondary infection and are also believed to be less effective in eliminating the newly infective virus serotype. Therefore, they are thought to lead to enhanced viral replication and thus severe clinical disease. Although cross reactive T cells and cross reactive antibodies may contribute to disease pathogenesis, these mechanisms alone do not explain the immunopathological mechanisms leading to severe disease, as severe clinical disease is known to occur even during primary dengue viral infections, especially infants and in pregnant women. Furthermore, it is believed that only 0.18–1% of primary infections and 2–9% of secondary infections manifest as DHF/DSS, and that the majority of individuals who are infected with the dengue virus develop mild or asymptomatic disease. Therefore, other factors are likely to play a significant role in the disease pathogenesis. Several genetic factors have been shown to be associated with the development of DHF/DSS and some have been shown to be protective. Certain HLA- class I and class II alleles, polymorphisms in the tumor necrosis factor alpha (TNF-α), Vitamin D receptor, CTLA-4, and transforming growth factor ß (TGF-β) have been shown to be associated with development of DHF/DSS. Alleles such as HLA-A\*24, HLA-B\*53, have been shown to increase the risk of DSS whereas HLA-DRB1\*09, HLA-DR4, HLA-A\*03 and HLA-B\*18 have been found to decrease the risk of developing DSS. Certain HLA alleles such as HLA-A\*51, HLA-A\*207 have been shown to increase the risk of severe DHF only during secondary dengue (SD) infection, while certain alleles such as HLA-B44, B62, B76 and B77 appeared to offer some protection. Although Lan *et a*l have also tried to identify HLA alleles, which may increase the susceptibility for the development of DSS/severe DHF during primary dengue (PD) infection, some of the patients who had a PD infection were below the age of 1 year. Therefore, the presence of maternal antibodies in these babies could have skewed some of the results and contributed to occurrence of DHF. These studies are important but have been limited in most cases by relatively small patient groups. In this study we aimed to identify possible HLA- class I and class II alleles, which increase the risk of developing DHF/DSS during PD and SD infection in a cohort of individuals with DHF. We found that HLA-A\*24 and HLA-DRB1\*12 were associated with a significantly higher risk of developing DHF during PD infection. In addition, we also found that HLA-A\*31 and HLA-DRB1\*08 were significantly associated with a higher risk of developing DSS. These data would support the hypothesis that T cells contribute to the development of severe clinical disease during dengue viral infections. # Methods ## Study population 110 patients with clinical features suggestive of dengue infections were who were admitted to a general medical ward in a tertiary care hospital in Colombo were enrolled in the study following informed written consent. The study was approved by the Ethical Review Committee of the University of Sri Jayawardanapura and the Ethical review Committee of the University of Oxford. Serial recordings of their clinical features and laboratory investigations (platelet counts, haematocrits, white cell counts) were made until they were discharged from the hospital in order to determine the severity of dengue infection. Patients with mild dengue/dengue fever were excluded from the study, and only patients with dengue haemorrhagic fever (DHF) (those who had evidence of plasma leakage) were recruited. These patients were classified as having dengue with warning signs (moderately severe dengue) and severe dengue according to the 2009 WHO guidelines. Patients with DHF with less than or equal to a pulse pressure of 20 mmHg were classified as having shock. HLA types of the normal population was derived from a study previously done by us. As only 102 individuals (who were recruited from the Colombo district) were included in our previous study, we recruited 17 more healthy individuals who so far never had a symptomatic/clinically diagnosed dengue infection from the Colombo district. Therefore, HLA types of 119 individuals from the population were available for comparison with the dengue patients. ## Serology Dengue virus infection was confirmed by testing the serum samples which were collected after day 6 of illness with a commercial capture-IgM and IgG enzyme- linked immunosorbent assay (ELISA) (Panbio, Brisbane, Australia). The ELISA was performed and the results were interpreted according to the manufacturer's instructions. This ELISA assay has been validated as both sensitive and specific for primary and secondary dengue virus infections. Patients who only had dengue virus specific IgM were classified as having a PD infection while those who had a positive result for both IgM and IgG were classified as having a SD infection. ## HLA typing HLA class I and class II alleles were typed as previously described. DNA was extracted from whole EDTA blood samples from the patients using the QIAamp DNA blood Mini Kit (QIAGEN, UK). Extracted DNA was amplified and DNA typing was undertaken using a 144 sequence-specific primer (SSP) reactions to simultaneously detect all known HLA-A, B, C, DRB1, DRB3, DRB4, DRB5 and DQB1 specificities in an allele specific or group specific manner using the same method, reagents, PCR parameters and protocols for all loci. The PCR products were electrophoresed in 1.0% agarose gels, using xylene cyanol FF and bromophenol blue as marker dyes. The gels were run for 15 minutes at 15 V/cm in the 0.5×TBE buffer and visualized using UV illumination. A phototype was considered to be successful when the control amplifications are positive and at least one allele or group was present in each locus. Allele frequencies were estimated from the number of positive typing reactions divided by the total number of haplotypes tested. ## Statistical analysis The frequencies of HLA class I alleles and HLA class II alleles in patients with DHF and the normal population were compared to determine possible association with DHF. Graphpad Prism version 4 was used for statistical analysis. For each HLA allele, degree of association between HLA alleles and disease state was expressed as the odds ratio (OR), which is obtained from standard contingency table analysis by Haldane's modification of Woolf's method. The Fisher's exact test was used to determine the p value. The P values were further corrected by using the Bonferroni's inequality method. The corrected P value (Pc) was calculated by multiplying the p values with the number of alleles tested for each locus. Comparisons to determine possible alleles that confer susceptibility to primary dengue infections were done by comparing the allele frequencies of patients with DHF during primary infection (n = 18) and patients with DHF during secondary dengue infection (n = 92). Comparisons to determine possible alleles that confer susceptibility to occurrence of shock during DHF were done by comparing the allele frequencies of patients with DHF associated with shock (n = 29) and patients with DHF who did not progress to shock (n = 81). Only the corrected p values were used to determine if certain HLA alleles were either positively or negatively associated with primary dengue or DSS. All statistical analysis and calculation of P values were done as described above. # Results ## Clinical characteristics and severity of dengue in the study population Of the 110 patients with laboratory confirmed dengue infection, 32 (29.1%) were females. All patients had laboratory evidence of a rising haematocrit (evidence of plasma leakage) with a concurrent drop in platelet counts. 50 (45.45%) complained of abdominal pain and 49 (44.5%) had at least one bleeding manifestation. Based on the 2009 WHO diagnostic criteria, shock was defined as lowering of pulse pressure to 20 mmHg or less. Accordingly, 29 patients were classified as having shock. Of the 29 patients who developed shock, 23 (79.3%) had at least one bleeding manifestation and 5 (17.2%), had episodes of loss of consciousness. 20 (68.9%), of those who developed shock had either haematemesis or melana or significant bleeding (vaginal bleeding, epistaxsis or prolonged bleeding). Their mean platelet counts were 24.2 (SD±16.7), whereas the mean platelet counts of those who did not develop shock was 35.78 (SD±35.82). Of the DHF patients who did not develop shock 26 (32.1%) had at least one bleeding manifestation and 6 (7.4%) complained of melana and 3 (3.7%) complained of vaginal bleeding. None of the DHF patients without shock had haematemesis or severe bleeding. Based on the presence of dengue specific IgM but no IgG, 18 (16.4%) patients were found to have PD infection. The mean age of those with PD was 24.7 (SD±6.6) and those with SD was 30.1 (SD±11.4). 3 (16.6%) patients with PD and 26 (28.2%) of those with SD developed shock. Bleeding manifestations were present in 9 (50%) of those with PD and 40 (43.4%) of those with SD. However, only 3 (the 3 patients who developed shock) of those with PD had a significant bleeding manifestation (melana, haematemesis, prolong or large bleed). ## Associations of HLA class I and II alleles with DHF The HLA class I and II allele frequency in the normal population and patients with DHF is shown in. Among the HLA-A alleles, although not statistically significant after correction for multiple comparisons, A\*68 allele was associated with a reduced risk of developing DHF (Odds ratio 0.302, CI 0.1096 to 0.8341). HLA-Cw\*03 allele was also associated with a reduced risk of developing DHF (odds ratio 0.544, CI 0.2876 to 1.028), which was again not statistically significant. HLA-DRB1\*08 was strongly associated with a higher risk of developing DHF (odds ratio 10.11, CI 1.269 to 80.50). The frequency of this allele in the normal population was 0.42, whereas the frequency in patients with DHF was 4.09. However again the association was not statistically significant (Pc = 0.0935). ## Associations of HLA class I and II alleles with DHF during primary dengue infections The HLA class I and II allele frequency in the normal population, patients with DHF with PD and SD is shown in. The allele frequency of HLA-A\*24 was 44.44 in DHF with PD infection, 20.65 in patients with DHF and SD infection and 19.75 in the normal population. This allele was associated with development of DHF during PD infection (odds ratio 3.074, CI 1.454 to 6.495) which was statistically significant (Pc = 0.03). Among the HLA-B alleles, although not statistically significant, HLA-B\*15 and HLA-B\*35 were also associated with a higher risk of developing DHF during PD infection (odds ratio 2.63 and 2.56 respectively). Interestingly, the HLA-B\*51 allele was not present in the DHF with PD infection, whereas the allele frequency in DHF patients with SD infection was 8.15 and the frequency in the normal population was 7.14. HLA-Cw\*04 allele was found at a higher frequency in DHF patients with PD (27.78) than in DHF patients with SD (15.76) and the normal population (11.76) and it was associated with a higher risk of developing DHF during PD (odds ratio 2.056, CI 0.8961 to 4.716). However, again this association was not statistically significant. Among the HLA class II alleles, HLA-DRB1\*12 was found at a higher frequency in DHF patients with PD (13.89), when compared to those with SD (2.17) and normal population (3.78). This allele was found to be significantly associated with (Pc = 0.041) a higher risk of developing DHF during PD (odds ratio 7.258, CI 1.846 to 28.54). HLA-DRB1\*01 was also found at a higher frequency among DHF patients with PD (11.11) when compared to DHF patients with SD (3.8) and the normal population (5.88). This allele too was found to be associated with a higher risk of developing DHF during PD (odds ratio 3.161, CI 0.8742 to 11.43), but this association was not statistically significant (Pc = 0.511). The frequency of DQB1\*03 was lower in patients with PD (19.4%), when compared to those with SD (30.4%) and the normal population (20.6%). However, this association was not statistically significant. ## Associations of HLA class I and II alleles with development of severe dengue (dengue shock syndrome) The HLA class I and II allele frequency in the normal population, DHF who did not develop shock and DHF with shock is shown in. HLA-A\*31 allele was found to be significantly (Pc = 0.01) associated with the occurrence of shock (odds ratio 18.58, CI 2.185 to 158.0). This allele was increased 16.6 fold in DHF who developed shock when compared to those who did not develop shock. The HLA-A\*26 allele was not found in any of the patients who developed shock, whereas the frequency of this allele in the normal population was 5.46 and in the DHF patients who did not develop shock was 6.17. Therefore, HLA-A\*26 appeared to decrease the risk of developing shock (odds ratio 0.124, CI 0.007154 to 2.154). However, this association was not statistically significant (Pc = 0.463). Among the other HLA class I alleles, HLA-B\*15 and HLA-B\*51 were increased 2.5 fold in patients who developed shock when compared to those who did not. Although both alleles were associated with a higher risk of developing shock (odds ratio 2.792, CI 1.073 to 7.267), this association was not statistically significant (Pc = 0.3132). B\*07 was found at a very low frequency in patients with shock (1.72), when compared to DHF patients without shock (10.49) and the normal population (7.56). It was negatively associated with development of DSS (odds ratio 0.149, CI 0.01945 to 1.151), but this association was not statistically significant (Pc = 0.285). Among the HLA-Cw alleles, Cw\*04 was found at a higher frequency among patients with shock (20.7%), when compared to those without shock (16.7%) and the normal population (11.8%). Cw\*15 was found at a lower frequency in patients with shock (5.2%), when compared to those without shock (11.1%), suggesting that it may be protective. However, both associations with Cw\*04 and Cw\*15 were not statistically significant. HLA-DRB1\*08 was strongly associated (Pc = 0.009) with the development of shock (odds ratio 10.98, CI 2.210 to 54.56). The frequency of this allele among patients with shock was 12.07, whereas the frequency among DHF without shock was 1.23 and the frequency in the normal population was 0.42. HLA-DRB1\*04 appeared to be protective against shock (odds ratio 0.49) as it was found at a lower frequency in patients with shock (5.2%), when compared to those without shock (9.9%). However, this association was not statistically significant. There were no differences in the frequencies of the HLA-DQB1 alleles in patients with or without shock or the normal population. # Discussion In this study we have identified HLA class I and class II alleles that are associated with the development of DHF during PD infection and SD infection and also alleles that are strongly associated with the development of DSS. The allele frequencies of patients with DHF were compared with the normal population, which at the time of recruitment had never reported a symptomatic/clinically diagnosed dengue infection. However, given that 72% of children aged 12 years from the Colombo district in Sri Lanka are seropositive to dengue it is likely that the majority of individuals who consisted of our normal population would have had at least one dengue virus infection. When comparing allele frequencies of patients with DHF and the normal population, HLA-DRB1\*08 allele was associated with the higher risk of developing DHF. The frequency of this allele in the normal population was 0.42, whereas the frequency in patients with DHF was 4.09. This allele was very strongly associated with the development of DSS (odds ratio 10.98, CI 2.210 to 54.56). The frequency of this allele among patients with shock was 12.07, which is 28.7 times higher than in the normal population. HLA-A\*31 allele too was found to be significantly (Pc = 0.01) associated with the occurrence of shock (odds ratio 18.58, CI 2.185 to 158.0). This allele was increased 16.6 fold in DHF patients who developed shock when compared to those who did not develop shock. Sierra *et al* also found that HLA-A\*31 was associated with DHF, but they have not analyzed its association with DSS/severe clinical disease. However, both HLA-DRB1\*08 and HLA-A\*31 were only associated with DSS/DHF in patients with secondary dengue infection, which possibly suggests that T cell epitopes directed to this allele could be highly cross reactive. Mapping and phenotyping T cell epitopes specific for this allele could possibly reveal disease mechanisms that contribute to severe dengue infections. Although described by others, we did not find any association with HLA-B\*51 or B\*15 and severe dengue infections. However, although not statistically significant, HLA-B\*15 and B\*57 were found at a higher frequency and were positively associated with susceptibility to DSS (odds ratio 2.79). HLA-A\*24 was found to be strongly associated (Pc = 0.03) with the development of DHF during PD infection (odds ratio 3.074). The allele frequency of HLA-A\*24 was 44.44 in DHF with PD infection and 19.75 in the normal population. Several studies have shown that A\*24 was associated with severe dengue. Although not statistically significant, Lan *et al* also found that A\*24 was associated with a higher risk of developing DSS/DHF with PD when compared to SD. The other studies that reported that A\*24 was associated with a higher risk of developing DHF/DSS had not reported the association of this allele in DHF/DSS patients with PD or SD. Although we found that A\*24 was associated with DHF during PD, it was not associated with the development of DSS. In fact the allele frequency in patients with DSS was 17.24, whereas the frequency in DHF patients without shock was 27.16. Simmons et al et al also reported that HLA-A\*24 T cell epitopes in the very highly conserved dengue virus NS3 protein have been identified and are thought to be highly cross reactive. The cross reactivity of this epitope is thought to contribute to disease pathogenesis. However, the possible contribution of this allele in the development of DHF during PD infection cannot be explained by the cross reactivity between A\*24 specific dengue virus T cell epitopes. Although several studies describe the associations of HLA-class I alleles with the development DHF/DSS, only a few have also described the associations with HLA-class II alleles. HLA-DRB1\*12 allele was found to be associated (odds ratio 7.26) with the development of DHF during PD infections. This allele was 6.4 times higher in DHF patients with PD infection when compared to those with SD infection. In the overall comparison of HLA allele frequencies, although not statistically significant, HLA-A\*68 was found to be associated with a reduced risk of developing DHF (Odds ratio 0.302, CI 0.1096 to 0.8341). This allele was also not detected in DHF with PD and the allele frequency was 1.72 in patients with DSS (frequency in normal population was 7.14). Appannna *et al* also found that this allele was found at a lower frequency in patients of Indian ethnic origin with DHF. Again although not statistically significant, HLA-A\*26 and HLA-B\*07 were found at a very low frequency in patients with shock when compared to those who did not develop shock and the normal population. However, Appanna et al found that in the Malay population in Malaysia, HLA-A\*26 was associated with a higher risk of developing DHF. Appana *et al* also described that B \*13 was associated with reduced susceptibility to DSS. This allele was not detected among our DSS cohort, which probably could imply that it does reduce susceptibility to DSS. However, as the overall frequency of this allele in the Sri Lankan population was low, it was not included in the analysis. La Fleur *et al* showed that in the Mexican population, HLA-DR4 was protective against development of DHF. We too found that this allele was present at a lower frequency among patients with DSS. However this was not statistically significant. DRB1\*07 allele has also been shown to be associated with protection in the Cuban population. We found that the allele was seen at a lower frequency in patients with DHF during PD infection and that it was negatively associated with development of DHF during PD (0.47). However, this association was not statistically significant. Although Nguyen *et al* found that HLA- DRB1\*09 was negatively associated with DSS, this allele was not detected in the Sri Lankan population. This study was carried out during the largest ever dengue epidemic in Sri Lanka, which occurred during the years 2009 and early 2010. Although in this study we have investigated the association of HLA alleles with primary and secondary dengue and also in the development of shock, it would have been useful to determine if infection with a particular dengue virus serotype also predisposes to shock or DHF during PD and its association with particular HLA alleles. However, as only adult patients were recruited in the study, they only presented to hospital with at least 4 to 5 days of fever and were therefore, only recruited to our study on day 5 of illness. Therefore, they were unsuitable for testing for the infecting dengue virus serotype. In summary, in a large cohort of affected individuals, we have identified HLA class I and class II alleles that are associated with the development of DHF during PD infection and SD infection and also alleles that are strongly associated with the development of DSS. We found that HLA-A\*31 and DRB1\*08 were associated with susceptibility to DSS when infected with the dengue virus, during SD infection. A\*24 and DRB1\*12 were strongly associated with the development of DHF during PD infection. As T cell epitopes to some of these alleles have already been identified, it would be now important to investigate how epitope specific T cells could contribute to disease pathogenesis in primary and secondary dengue infections. Overall these data support a role for T cells in the pathogenesis of dengue associated disease. [^1]: Conceived and designed the experiments: GNM TR GO. Performed the experiments: GNM TR ML. Analyzed the data: GNM GO. Contributed reagents/materials/analysis tools: NF ADDS GO. Wrote the paper: GNM GO. Provided the patient samples, collected data and clinically diagnosed patients with acute dengue infection: LTR SDJ. [^2]: The authors have declared that no competing interests exist.

# Introduction > *“It is wisdom to know others; it is enlightenment to know one’s > self.”* > > *Lao-Tzu, 6^th century BC* Interpersonal success rests on the mind’s remarkable ability to decode and comprehend other people’s mental states. In a challenging social landscape, mindreading is the conduit through which relationships are forged, cultivated, and sustained. As Lao-Tzu’s quote indicates and an extensive literature has demonstrated, both admirable (e.g., empathy, compassion) and unsavory (e.g., deception, manipulation) aspects of the human condition rest squarely on the ability to read minds. Indeed the utility of this skill – commonly referred to as Theory of Mind (ToM) – is most apparent when one considers the plight of individuals for whom mindreading is a decidedly problematic activity (e.g., autism spectrum disorder), a topic that has dominated psychological writings for decades –. But what of the converse situation, rather than focusing on impairments in mindreading, is it possible to identify factors (e.g., interventions) that may enhance ToM? Intriguingly, preliminary investigations have revealed that mindreading can be improved under certain circumstances. Nasal administrations of oxytocin, reading literary fiction, and extended compassion-based training have all been shown to enhance ToM. Extending this line of inquiry, here we considered whether a brief period of mindfulness-based meditation would similarly improve mindreading performance. In recent years, mindfulness interventions have been shown to remediate a range of clinical problems (e.g., depression, anxiety, stress) and to impact core aspects of social cognition (e.g., metacognition, self- referential thought, see). Emphasizing the non-judgmental appraisal of present- moment thinking, even brief episodes of mindfulness meditation exert profound effects on brain and behavior, effects that we suspect may extend to people’s mindreading skills. Several strands of research highlight important linkages between mindfulness and ToM. First, mindfulness interventions enhance executive attention, a core component of mentalizing and person understanding. Deployed effectively, executive attention enables perceivers to form multifaceted evaluations and impressions of other social agents that extend beyond rigid stereotype-based conceptions. Second, present-moment thinking facilitates the cognitive operations that map the minds of self and others. Specifically, cortical regions that support mindreading and self-referential mental activity (e.g., medial prefrontal cortex, temporal parietal junction) also play a prominent functional role during mindfulness meditation,. Overlap in these regions likely captures the influence of self-reflection during explicit mindreading. Collectively these observations suggest a straightforward prediction, mindfulness meditation may enhance ToM. To explore this possibility, participants completed two complementary ToM tasks following a mindfulness (cf. control) intervention. One task, ‘The Reading the Mind in the Eyes Test’, assessed participants’ ability to decode emotions and mental states from subtle facial cues. The other task, ‘The Cyberball Social Exclusion Game’ explored their capacity to empathize with others. Adoption of these tasks enabled us to consider the effects of mindfulness on two pivotal components of ToM, mental state attribution and empathic understanding. Importantly, participants had no prior experience with meditation and did not complete the typical 8-week mindfulness-training program. Instead, only a brief mindfulness-intervention was employed (see also). # Methods ## Participants and Design An *a priori* sample size calculation performed on G\*Power 3.0.10 (∝ = .05, *d* = 0.6, power = 80%) revealed a requirement of 72 participants. Seventy-two individuals (36 females; *M_age* = 23.8, range 18–50 years; 93% Chinese) provided written consent and took part in the research. The ethics board at the University of Hong Kong approved the manner of consent and the study (reference EA 480114). The experiment had a single factor (Condition: mindfulness or control) between-participants design. No rewards were offered for participation in the research, all participants received de-briefing and opportunity for further clarification afterwards. ## Materials and Procedure Participants were greeted by a female experimenter, randomly assigned to one of the treatment conditions (i.e., mindfulness or control), and told the research comprised an investigation (i.e., series of tasks) into people’s reactions to different types of thoughts and situations (i.e., participants were blind to the purpose of the inquiry). The experimental manipulation was then introduced. All participants were instructed to close their eyes, relax and listen to scripted audio instructions (via headphones) and that a bell would chime after 5-minutes to signal the end of this activity. Based on an established protocol, participants in the mindfulness condition were instructed to pay particular attention to the sensation of their breathing during the 5-minute period. In addition, they were told it is quite natural for the mind to be distracted and attention to wander during such a task. However, they were asked to observe these moments as fleeting states of mind and to return attention to their breathing each time a distracting thought, emotion or memory occurred. Participants in the control condition received instructions that were similar in style and length. Contrasting the mindfulness treatment, however, these individuals were told to notice each thought, emotion and memory that may arise and to be completely immersed in the experience. Following the 5-minute task, as a manipulation check, participants completed the Mindful Attention Awareness Scale (MAAS-State), a questionnaire (5-item) that probes levels of mindful-attention and awareness via a 7-point rating scale (0 = not at all; 6 = very much). On completion of the MAAS-S, participants’ mindreading skills and levels of empathic understanding were assessed using the Reading the Mind in the Eyes Test and Cyberball Social Exclusion Game, respectively. Task order was counterbalanced across the sample. ## Reading the Mind in the Eyes Test (RMET) The RMET required participants to identify the emotions/mental states expressed by eyes displaying subtle affective facial expressions. Measuring mental state inference from socio-perceptual cues, the RMET is one of the most widely used instruments in ToM research. In total, participants viewed 36 pairs of eyes (black-and-white pictures of 18 male and 18 female eyes) displaying discrete expressions (12 negative, 8 positive, 16 neutral, counterbalanced for the sex of the eyes). Their task was to report which of four presented words (e.g., hateful, jealous, arrogant, panicked) best described the emotion/mental state of the person whose eyes were displayed on each trial. ## Cyberball Social Exclusion Game Participants were asked to watch a 90-second video clip during which three players (i.e., cartoon figures) engaged in a virtual ball-tossing game (i.e., Cyberball, [cyberball.wikispaces.com](http://cyberball.wikispaces.com)). The names of the three female players were provided and participants were informed that afterwards they would have an opportunity to write a letter to Ann, one of players. In total, participants observed two rounds of Cyberball (11 ball tosses per round). Critically, during each round, Ann received the ball on only two occasions, thus was largely ignored by the other players. This paradigm has been used extensively (and successfully) in previous research to simulate social exclusion/rejection and distress (e.g.. Following the Cyberball videos, participants were given 3-minutes to write a letter to Ann, and were instructed to provide an account of their thoughts and feelings during the game. These letters were later scored for empathic content by independent raters. On completion of the tasks, participants were debriefed, thanked and dismissed. # Results Several *t*-tests were conducted in order to examine inter-group differences. The success of the experimental manipulation were confirmed, such that participants in the mindfulness condition reported a greater awareness of the present moment (i.e. state mindfulness) than their counterparts in the control condition, *t*(70) = 6.23, *p*\<.001, *d* = 1.47. As expected, participants who underwent the mindfulness induction reported higher levels of state mindfulness. Mindreading performance, as measured by the RMET, was better for participants in the mindfulness than control condition, *t*(70) = 5.62, *p*\<.001, *d* = 1.33, confirming our prediction. Additional regression analyses were undertaken to test the predictive power of mindfulness after controlling for other variables (i.e., sex and age). Using the enter method, a significant model emerged in which mindfulness best predicted mindreading scores (standardized ß = −.41, *p*\<.001 and an adjusted variance *R*² = .13, *p*\<.001). Neither age nor sex were significant predictors in this model (ß = .05, *p* = .66; ß = .05, *p* = .65; respectively). To assess the effect of mindfulness on empathic understanding, two coders who were blind to experimental condition and the purpose of the investigation scored the letters for empathic content (i.e., 6-point scale: 1 = not at all empathic; 6 = very empathic). The coders were instructed to consider the pro-social flavor of the letters, specifically the extent to which the writers had attempted to direct support, comfort and understanding to Ann (i.e., the excluded individual). Assessment of the inter-rater reliability indicated a high level of agreement in their ratings (Cronbach’s *α* = .92), scores were therefore averaged and a single measure of empathic understanding was calculated for each letter. Subsequent between-group analysis revealed that mindfulness participants expressed more empathic concern compared to participants in the control condition, *t*(70) = 2.62, *p* = .011, *d* = .62 (see and for examples of letter excerpts). Further analysis revealed that the sex of participants did not impact empathic concern, *t*(70) = 1.31, *p* = .20. In summary, significant between-group differences were found after a brief mindfulness induction. Confirming our hypotheses, participants in the mindfulness group out-performed those in control condition on both ToM tasks (i.e. mind reading and empathic expression). # Discussion As demonstrated herein, brief mindfulness meditation enhanced core components of ToM, notably mindreading and empathic understanding. Not only were participants better able to decode complex mental states from subtle facial cues following a brief period of mindfulness meditation, so too they expressed greater empathic concern when communicating with the excluded victim of a computerized ball- tossing game. Together, these findings highlight the potent effects that even short-lived mindfulness interventions can exert on basic aspects of social- cognitive functioning (i.e., mentalizing & empathizing). Questions remain, however, regarding the specific mechanism through which mindfulness meditation impacts ToM? One possibility is that elevated levels of metacognitive awareness facilitate mindreading and boost empathic responding. A commonly articulated viewpoint is that awareness of one’s own bodily and psychological states serves as an important precursor to mentalizing. That is, accurate observations of self are required for a comprehensive understanding of others. As mindfulness meditation focuses on internal (bodily and psychological) experiences, increased self-focused attention is a natural byproduct of this activity. It may therefore be the case that enhanced metacognitive awareness serves as the critical pathway through which brief periods of mindfulness improve ToM, a possibility that awaits empirical investigation. Additional questions center on the extent to which mindfulness interventions benefit all aspects of ToM. In charting people’s mindreading skills, a fundamental distinction has been drawn between the affective and cognitive sub- components of ToM. While affective ToM represents the ability to intuit what a person is feeling, cognitive ToM reflects the capacity to infer their beliefs and intentions. As the current inquiry employed tasks that consider only the affective component of ToM (i.e., RMET, empathy task), it remains unclear whether comparable improvements in mindreading would emerge on activities that tap the cognitive aspects of person understanding (e.g., false belief tasks, see also. Given however that mindfulness is acknowledged to enhance awareness of bodily states, emotions and cognitions, there is little reason to suspect that improvements in mindreading should be restricted to affective tasks. Instead, elevated metacognitive awareness may facilitate multiple strands of person understanding. While previous work has highlighted the cognitive and emotional benefits of long-term meditation training, here we demonstrated comparable effects on mindreading following only a brief (i.e., 5-minute) mindfulness intervention. The current results are far from unique, however. Elsewhere, researchers have shown that even brief mindfulness training can improve performance in a variety of domains, including: executive function, visual-spatial processing, working memory and impulse control. Of course, that brief episodes of mindfulness appear to facilitate processing much like extended training regimes raises a host of important questions pertaining to the underlying processes, mechanisms of change, strength, and duration of the respective effects. In no sense are we suggesting that brief mindfulness interventions are as effective as long-term training programs in shaping behavior. Nevertheless, emerging evidence indicates that even minimal mindfulness training is sufficient to facilitate fundamental aspects of human cognition. Elucidating the precise nature of these effects will be an important task for future research. # Conclusions The finding that brief mindfulness training improves ToM may have implications that extend beyond the laboratory. Underlying harmonious living is the ability to understand the thoughts and feelings of other social agents. Indeed, ToM is deemed by many to be one of the pinnacles of human evolution. While extant research has focused primarily on explicating the problems that emerge when mindreading goes awry, recent investigations have shifted instead to the identification of strategies (and interventions) that enhance person understanding. As the current findings demonstrate, brief mindfulness-based meditation comprises just such a tactic. At least with respect to affective ToM, mindful attention facilitates mindreading, although additional research will be required to scrutinize the hypothesized link between mindfulness and ToM. Of particular importance will be work exploring the process and consequences of brief mindfulness meditation in applied settings, such as a longitudinal study with a clinical population. The authors would like to thank the participants and reviewers (J. Smallwood and P. Williams) for helpful comments on previous version of this manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LBGT CNM. Performed the experiments: LBGT. Analyzed the data: LBGT CNM. Contributed reagents/materials/analysis tools: LBGT BCYL. Contributed to the writing of the manuscript: LBGT CNM.

# Introduction The public perception of nature has changed through human history, more recently facing two main drivers. These are pushing in opposite directions: urbanization, that heavily reduces daily contacts with plants or animals, and science, which constantly enlarges knowledge and gives new insights on sustainable living choices to preserve nature. Still no consensus exists on how to measure the importance of natural resources, the services provided by natural ecosystems, the connections of nature-enriched environments with health and quality of lifetime. The natural world is often perceived as a “surrounding” environment, to which most of us is not directly connected. Instead, our own survival strongly depends on the connections with it. Plants are of enormous importance as providers of oxygen, food, and as sources of pharmaceutical products that we need to contrast illnesses. They sustain human well-being and are food sources and shelters to animals interacting with them. Pollinators, especially bees, are also of enormous importance: they provide one of the most important ecosystem services. Pollinators are the vectors that plants use to produce new generations through the processes of fertilisation, fruit and seed formation. Currently there is a need to keep pollinators at the spotlight, since large diversity of the food (e.g. fruits and seeds) consumed by human, wild and domesticated animals, and even pets, rely on pollination to be produced. How people react to plants and nature-related concepts can however result surprising. An example is the concept of native/exotic species. Hoyle and colleagues investigated the public perception of non-native plants in gardens highlighting key factors that are actively driving acceptance or rejection of a given landscape by the public. As expected, the aesthetically pleasing appearance was one of them: beautiful flowers are accepted and planted in gardens, independently from their country of origin. However, it also turned out that potential incompatibility with native wildlife plays a role in granting acceptance of non-native species: knowledge on how plants influence the local environment may change people’s mind. Urban gardens lately received an increasing attention as repository of intrinsic values: both, for people well-being and for the ecological services they sustain. Back in 2009, Frankie and colleagues disseminated the results of a large study involving gardens in seven Californian cities: they underlined the intrinsic value of gardening as habitats for native bees. Again in 2009, Pawelek and colleagues highlighted how it was possible even to increase local pollinators in urban gardens by choosing different plants and garden designs. Almost ten years after, Burr and colleagues provided useful conservation direction for yards in the USA, melting data on insect pollinator populations and social and cultural drivers influencing people choices. Similar studies, that provide information on compatibility between ornamentals and pollinator sustainability \[i.e. ,\], perfectly match the increased apprehension about dramatic pollinator's losses that recently also reached the wide public. Botanic gardens (BG) are special places where merging exotic plant species and their relationship to local wildlife, finally guiding human perception on how they interact and what may result by hosting exotic species. BGs are special since the plants they host are not casually selected: they have been planted and catalogued according to precise criteria: an example is the status of each species (e.g. exotic, rare, endangered, with great conservation value, etc.). Moreover, they have a strong historical value that the public can appreciate. These gardens appeared in the Middle Age in convents or monasteries: they collected medicinal plants, indigenous or exotic, employed for the care of various sicknesses. In the Renaissance, they became places for the collection, cultivation and study of plants with healing properties (the first in Pisa, Italy, in 1543/44). During the eighteenth century, the period of the great explorations, BGs hosted the exotic species coming from newly discovered countries with the aim of experimenting their ornamental and economic potential. Currently, in the context of biodiversity losses, BGs have assumed a new role as repositories for the conservation of the plant biological diversity at global level. In addition, a fundamental mission of BGs is linking public direct experience to the perception of the importance of natural systems, greeting citizens as pleasure and relaxing sites while concurrently acting as “open-air museums”. We planned a transversal project, where the public perspective drove the focus of scientific research plans: 1) to address the recent findings of granting acceptance of non-native species when connected to positive compatibilities with native wildlife, we investigated exotic and native species in relation to pollinators visits. In this case, we expected *BG* to act *as a plant-pollinator network repository*: considering their urban location and the abundance of plant species with different flowering time, BGs may constantly sustain local population of pollinators. During two following years we monitored bee visits and analysed differences and similarities comparing the respective networks of exotic and native plant species, with the aim of verifying suitability of exotic species as food sources for local bees; 2) to sustain the need of deepening the connection of nature with everyday human life, we addressed to sages, renown species traditionally used for cooking and frequently planted in private and common gardens. We performed a comparative study using the *BG as an open-air laboratory*: it was a perfect location, since *Salvia* species of different geographic origin acclimatised during a long period and, notwithstanding their home range, potentially share the native pollinators assemblage. We measured flower characteristics of five *Salvia* species, recorded and identified pollinators visiting them, compared pollinators assemblages and pollinators frequencies among specie. Our aim was selecting the mostly-liked by local pollinators giving possible suggestions on the best-suited species for gardening activities; 3) with the aim of strengthen the link between scientific findings and society and promote understanding and conservation effort, *BG* was *the interactive learning promoter* of a national travelling exhibition titled “*Seduzione / repulsione*: *quello che le piante non dicono*”(*Seduction Repulsion–what plants do not say*) associated to an illustrated catalogue. The BG promoted the exhibition and the catalogue by implementing their setting-up with the involvement of local and national stakeholders, but was also actively hosting and spreading the content of the exhibition. In these deliverables, plants were presented not only for the aesthetic appeal of their flowers: they were illustrated according to the interactions they establish with other organisms, the evolutionary paths that drove them to develop given strategies, finally underlining how these strategies are not different from those also employed in human activities. # Materials and methods ## Study site This study was performed at the Ghirardi Botanic Garden (GBG) of Toscolano Maderno (Italy), on the western shore of the Lake of Garda at 86 m asl. The Ghirardi Botanic Garden (GBG) of Toscolano Maderno (Italy) granted permits to carry out field research in its premises. The town has a municipal land area of 56.73 km², of which 0.78 urbanized (urbanized surface incidence = 13.77%). In 2017, the population census reported 7969 actual residents (population density = 135.41 residents/km²) plus about 7000 transient inhabitants/tourists temporarily present in camping, hotels or other accommodation facilities during the spring and the summer. The climate is mild, generally warm and classified by the Köppen-Geiger system as continental temperate, with hot summer (Cfa). About 844 mm of precipitation falls annually; hottest months are July and August with respective mean daily maximum temperatures of 28°C and 29°C. GBG, established in 1964 as an experimental botanic station under the direction of Prof. Giordano Emilio Ghirardi, extends over a surface of about 10000 m² and, currently, the preserved plant heritage includes more than 400 *taxa* from all the regions of the world. GBG has a long history of hosting medicinal species of different origin, in relation to the favourable microclimate of the site which facilitated the acclimatization of the introduced plants. The primary purpose was the cultivation and preservation of officinal species, mainly with cardiotonic and antitumor properties. It is worth mentioning the Chinese *Camptotheca acuminata* Decne. (Cornaceae), whose seeds were sent to GBG for a presumed anticancer activity, finally documented only by recent studies. Since 2002 it is part of the non- profit network “Rete degli Orti Botanici della Lombardia” (“Network of Botanic Gardens of Lombardy”). ## GBG as a plant-pollinator network repository Plant data collection started with an accurate screening, including the earliest checklists up to the most recent contributions and reports. To match pollinator visits to plants, we recorded visits along linear transects every two weeks during two following seasons (March-September 2016 & 2017; n = 22 transects). Due to the numerous plant species and their different flowering durations, repeated walking-transects are the most suited method to detect if frequency of visits may be considered occasional (a single visit is recorded) or if some kind of constancy is observed. Along the transects, 244 plant species were present, belonging to 63 families. However, for further analyses only species showing attractiveness towards flower visitors were considered. Bee individuals were recognised at sight till the deepest possible level and recorded only once when visiting a species, even in the case of paying multiple visits at flowers simultaneously present on the plant. We also took photographic evidence of bee visits, double-checking with final bee identification list, and collected specimens. We grouped visited plant species according to their provenance: native, if originally of Mediterranean or continental Europe; exotic, if from other continents or islands. Further, they were grouped according to Pellissier and colleagues' classification of flower characteristics: wind, disk, funnel, bilabiate, tube, head, brush. These floral morphologies vary as for availability and accessibility of the floral resources, pollen and nectar; therefore, they imply different attraction potential towards diverse groups of pollinators. To depict preferences of bees in flowers with different morphologies, we constructed a network visualisation using R 2.14.0 (R Core Team, 2013), keeping exotic and native species separated. ## GBG as an open-air laboratory GBG hosts numerous species of the mint family (Lamiaceae). We selected sages, our target being five *Salvia* L. species differing by native range, flower characteristics and pollination ecology. Indeed, the genus has some peculiarities regarding pollination strategies, either carried on by bees or by birds, and compound emissions. Two species were native to Europe (*Salvia pratensis* L. and *Salvia verticillata* L.) and three exotic, of central-south American origin (*Salvia blepharophylla* Brandegee ex Epling, *Salvia greggii* A.Gray and *Salvia uliginosa* Benth.). These species were not only divergent as geographical origin: they also originally co-evolved with different pollinators: insects (mainly bees) for *S*. *uliginosa*, *S*. *pratensis and S*. *verticillata*, and birds (mainly hummingbirds) for *S*. *blepharophylla* and *S*. *greggii*. In temperate areas as the one where the GBG is located, bird- pollination is not an option. Our aim was verifying if all species equally sustain native pollinators, being rich in nectar and often planted in gardens as ornamentals or for culinary purposes. To investigate pollinators on them, we first performed a literature search on the five species, acknowledging only citations referring to observed visits. We revised the first 50 citations obtained as Google Scholar output under comparable keywords (pollinator / *Salvia* 'name of species'). In the genus *Salvia*, as in the whole family Lamiaceae, the flower is bilabiate and characterized by the typical staminal lever, a mechanism helping in a successful pollen deposition on the pollinator’s body. For testing the morphological variability at flower level, 20 randomly- selected fully-opened flowers per species were compared for what concerns the inflorescence type, the floral colour and the total length of the corolla. We measured the last parameter using a digital calliper and a stereomicroscope and evaluated three different size class: short (\<1.5 cm), medium (1.5–3.0 cm), long (\>3.0 cm). We directly recorded flower visitors on each species on sunny days, between 8:00 and 14:00 (solar hour). Patch records regarded bee observation, were repeated along the day and lasted 10 minutes (221 patch records in total). Data refers to 10 days, from May to September 2016 fortnightly. Each bee approaching the flower was recorded and classified as explained for the plant-pollinator network: each individual accounted for a single visit, notwithstanding the amount of visited flowers. ## GBG as the interactive learning promoter For this project, the GBG was the promoter, and directly involved, in the setting-up of a travelling exhibition and a printed catalogue. Information on plants and their communication media with other organisms were selected: texts, photographic material and drawings were all employed for the final editing. Multi-stakeholder’s meetings involving scientist, artists and botanic garden managers were planned in order to decide how to settle up the mobile exhibition. Selection of material to be included in the panels and how to relate it with plants in the garden was also the result of open discussions. Mobile panels were printed and exposed in the GBG greenhouse. Some plant species hosted in the garden where selected to recall the information of the panels and flagged accordingly: this way, the information read on panels was transferred to a direct experience while visiting the garden and its plant content. The exhibition was displayed across Italy thanks to the involvement of the Network of Botanic Gardens of Lombardy, and other stakeholders as local municipalities. The catalogue was reflecting the panel order, reproducing part of the same content, comprehensive of texts, photographic material and drawings. A colloquial language was mainly applied to explain context of concepts; however, scientific terminology was also employed and fully explained. Moreover, to increase empathy we highlighted those strategies adopted by men for communication efforts during artistic and social performances that resemble the ones adopted by plants. # Results ## GBG as a plant-pollinator network repository The linear transects contained 244 plant species, out of which 140 (57.4%) received at least a visit by a local pollinator. Abundance of native and exotic species along the transects was very similar: 44.3% of exotic species and 55.7% of native ones. While walking along the transects, we recorded 517 bee visits on flowers, including those made by unidentified bees. Exotic and native plant species experienced similar amounts of occasional (26 and 23, respectively) or multiple visits (36 and 55, respectively), no difference emerging between natives and exotics (Fishers's exact test: p = 0.154). The majority of visited species (61.1%) experienced more than one visit, up to a maximum of seventeen. Also combining data and looking for overall number of visits received by all exotic or all native species, no difference emerged (*t* = 1.7242, df = 130, p = 0.0871). Average number of visits would be 3.05 visit/exotic- and 3.97 visit/native plant species. Notwithstanding their origin, native and exotic plant species obviously showed convergence of flower characteristics. When considering flower morphology, among the visited species we observed that brush blossom was not represented at all, while tube and wind flower morphotypes were poorly represented: 5 records of visits in total, native and exotic combined. The other four categories (bilabiate, disk, head and funnel) were all similarly visited: 30.9% of visits to bilabiate, 20.8% of disk, 17.2% of head and 30.1% of funnel. Minor differences in trends were shown by the two groups. Among the exotic species, number of visits where in decrescent order for head, disk and funnel blossoms, while for native species highest visits were on bilabiate, funnel and disk blossoms. No preferences emerged in the number of visits recorded, when addressing flower morphotype and origin of the species, not even for the two most represented groups (for bilabiate blossom, Fishers's exact test: p = 0.1094; for funnel, Fishers's exact test: p = 0.1517). Only a few bees could not be ascribed to any of the following families: Apidae, Andrenidae, Colletidae, Megachilidae and Halictidae. The five families were differently represented, as number of visits. The highest frequency of bee visits was due to the family Apidae (56.5%), shared between the *Apis mellifera* L. (honeybee) and *Bombus* Latreille species (bumblebees). The second family in order of importance was that of Halictidae (25.2%). The remaining 18.3% was due to Andrenidae, Megachilidae and Colletidae. represents the visualisation of networks between native or exotic plants and local pollinators. The network also highlights the relative abundance and visitation rates of plants grouped accordingly to the flower morphotype they were sharing. ## GBG as an open-air laboratory Results of the literature search highlights that the most frequent pollinators observed on the five species of sages were the honeybee (*Apis mellifera*) and the bumblebees (*Bombus* spp., various species). *Xylocopa* spp. also seemed quite attracted to sages, being previously recorded on three out of the five *Salvia* species. Bees were reported even on *S*. *greggii*, considered an ornitophilous species. For *S*. *blepharophylla*, we could not find any report of direct observations, inside or outside its home range. During our observations, we also recorded the bee families previously listed in the literature (dots refer to presence on flowers). A single family was ubiquitous on all sage species: individuals of the family Halictidae (mostly *Lasioglossum* Curtis spp.) were observed on all the five documented *Salvia*. Bee assemblages on the five sages differ. The ornitophilous species were the ones attracting a less diverse group of bees. *S*. *pratensis*, a species spontaneously growing in many areas even adjacent to the GBG, was surprisingly poorly visited and even discarded by *Bombus* sp. During a total of 2210 minutes of focal observations, we recorded 883 visits paid to flowers. Many of the 10-mins slot of observations remained without any bee contact: almost the 30% of all time dedicated to observations. Considering all records combined (lower graph), *S*. *verticillata* was the species with the highest number of records: 0.43 visitors/minute, followed by *S*. *uliginosa* (0.35 visitor/minute), and *S*. *greggii* (0.28 visitor/minute). *S*. *blepharophylla* and *S*. *pratensis* had very few visitors (0.09 and 0.07 visitor/minute, respectively). Kruskal- Wallis H test confirmed the difference among sages (χ2(4) = 61.985, *p* = 0.0001). ## GBG as an interactive learning opportunity The exhibition titled (in Italian): “*Seduzione Repulsione*: *quello che le piante non dicono*” (*Seduction Repulsion–what plants do not say*) displayed ten panels: titles and content of each is reported in. The aim was explaining through clear images, drawings and text, the various media of communication employed by plants and the resultant interactions (either positive, negative, or mutualistic) with other organisms. Similarities with human daily life were underlined, reporting them in the panels as well as in the catalogue. For example, the chapter related to colours starts with a citation and the picture of a painting, “In the style of Kairouan– 1914”, of the famous artist Paul Klee. After a trip to Tunisia, the artist finally moved away from black and white works, amazed by colour variability due to intense light. The chapter continues explaining the origin and importance of the colour green in leaves for photosynthesis, and how it is used as background colour for pollinators and seed dispersers to distinguish flowers and fruits. The text gets deeper by introducing the fact that there are differences in how colours result attractive to different organisms (insects, humans, birds, mammals) and what the chemical compounds responsible for different colours are (anthocyanins, carotenoids, flavonoids), in plants and animals as well. Finally, it concludes underlining that differences in perceived colours may also be the result of differences of the surface bearing them. Therefore, as in this example, each panel/chapter of the catalogue refer to features of major importance in the natural world, but relating them with physical and emotional perceptions more often experienced by people. The catalogue was published in 2016, directly involving the GBG and the non- profit organisation “Rete degli Orti Botanici della Lombardia”. It was assembled through the contribution of A. Ronchi (texts), P. Berera (graphics), under the scientific supervision of G. Fico, while involving numerous collaborators and national stakeholders (Fondazione Cariplo, Regione Lombardia Agricoltura, Ministero dell'Istruzione dell'Università e della Ricerca). The exhibition, still available, already travelled across 11 Italian locations: Botanical Garden Pavia, 6-18/09/2015; Sala Viscontea (Bergamo, 04/10/2015-31/01/2016); Ghirardi Botanical Garden (Toscolano Maderno, Brescia, 14/05-30/06/2016); Headquarters of StelvioPark (Bormio, Sondrio, 15/07-30/09/2016); Brera Botanical Garden (Milano, 12/12/2016-14/01/2017); Castello di Desenzano (Desenzano del Garda, Brescia, 04/03-02/04/2017); Villa Pisani Bolognesi Scalabrin (Vescovana, Padova, 6-25/04/2017); Villa Litta (Lainate, Milano, 30/04-20/05/2017); Natural History Museum (Venezia, 04/11/2017-18/03/2018); Tenuta Villa Quassa (Ispra, Varese, 7-29/o4/2018); JRC, European Union Joint Research Center (Ispra, Varese, 03/05-19/07/2018). The total number of recorded visitors was 51390, in two years; Venice alone attracted more the half of them (28390), with a presence of about 5600 visitors each month and an almost constant increment in the 5 months. The catalogue is currently available at the BGs belonging to the Network of the Botanical Gardens of Lombardy and through specific requests at <segreteria@reteortibotanicilombardia.it>. The venues differed between museal institutions and botanic gardens. At botanic gardens, the exhibition encountered a more selected audience searching for rigorous scientific and botanical in-depth-knowledge, but also fascinated by living organisms and the beauty relying on flowers and green leaves. They were finding out themselves the characteristics described in the panels. Museal institutions host in general a wider public, from elderly people to families to school groups (the Network of the Botanical Gardens of Lombardy, in 2017, sum up a total of 39.049 students), ready to walk around and possibly similarly interested to various topics, from nature to history to art. Museums often advertise special exhibition to enlarge the interest of resident public, or inducing distant one to join. This function was successfully taken on by the travelling exhibition during this work. The success of the exhibition resulted in the translation, in a language rich of links with accessible experiences, the achievement of science. Moreover, the added value was that a large part of the achievements presented were resulting from activities run at the same place of exhibitions: museums, botanic gardens, universities. This was well expressed by those that visited the botanic garden with researchers actively observing and recording pollinators. The enthusiastic interest in researcher’s activities, combined with in-situ explanations, seemed to push the interest towards the connection between plants and pollinators. # Discussion ## GBG as a plant-pollinator network repository According to Hymenoptera behaviour, we may expect occasional probation in new food sources; repeated visits confirm instead appreciation of resources offered by the plant. Bee visits on exotic flowers have long been recorded, even on invasive species from very distant origin. Visitation rates may develop from occasional visits to the development of a given routine for resource collection. From the results of the present study, we can conclude that the majority of visited plants in the GBG was actively looked for by local pollinators. This indicates that, in absence of co-evolutionary processes that may have built a solid relationship between a plant species and its pollinators, there are equally attractive forces that favour the establishment of new relationships. Similarity of floral morphologies is certainly the most evident trait possibly justifying this conclusion. However, future data on resource availability may integrate the current findings. As expected, also pollinators distribution can influence records on visits. Our data pointed a greater abundance of honeybees and bumblebees, when compared to relative abundance of other bee groups. A possible explanation may be linked with honeybee distribution facilitated by men through beekeeping. For the genus *Bombus*, it usually counts on several species when in proximity of mountain areas (i.e. the surroundings of Lake of Garda), where they can find a higher number of suitable nesting sites. Generally, we observed that exotic species attracted native pollinators and, depending on species availability and matches with flower characteristics, exotics may even compensate for resources according to flower abundance of native ones. We have to keep in mind that here we did not consider negative effects of possible disruptions of native plants-native pollinators networks, or negative effects due to invasive alien species. A similar result was reported by Lowenstein and colleagues, who did not find any effect of biogeographic origin (native versus non-native) of plant species regarding pollinator presence. However, exotic and native flowering ranges at the GBG overlapped and compensated resource offer, with an overall positive effect on local bee assemblages. This is sustained by how linkages are consistent independently from the origin of a species and is also confirmed by considering the flower shape: in presence of opposite frequencies of head and bilabiate flowers, respectively between exotics and native, we still observe consistency of activity and variety of visiting pollinators. A final consideration deals with future experimental studies on pollen deposition performance and on exotic species habits that may help to interpret linkages as evidenced by Devaux and colleagues. ## GBG as an open-air laboratory Sages are largely renown as ornamentals and/or aromatic species grown in private and common gardens. Notwithstanding, when addressing to their pollinators poor scientific evidence is available. It is worth mentioning that most of the literature we could access refers to pollinator’s observation out of the home range of the target sage (see second column of). The absence of data on natural conditions turns it difficult to compare native and exotic state of pollinator networks. Global trading of these species and pollinator’s records worldwide, however, are important to formulate the potential plant-pollinator relations to be expected. For example, *S*. *greggii* is definitely attracting the wild large species of the genus *Xylocopa*, in California (USA; literature data) as well as in Italy (Europe). *S*. *pratensis*, interestingly, did not attract even *Bombus* sp., its usual pollinator. This result may eventually be linked to plant location in the garden; however, it may also reflect competition among available resources. Similarly, the large amount of observation set missing bee records could be partially due to adverse weather conditions, or to bees foraging elsewhere on more attractive plant species. Flower characteristics confirm similarities depending on the pollinators that sages were expected to be coevolved with bees for *S*. *verticillata*, *S*. *pratensis* and *S*. *uliginosa* and birds for *S*. *greggii* and *S*. *blepharophylla*. Dissimilarities in bee attraction among species need to be deeper investigated from an evolutionary point of view, since they were not explained solely by flower traits. Among the insect-pollinated ones, *S*. *verticillata* showed an outstanding number of individuals even when compared to *S*. *uliginosa*, the most alike. Similarly, there was a difference between the two ornitophilous species, with *S*. *greggii* accounting for the highest number of bees visiting its flowers. These differences may partly be accounted for flower arrangement on the plant. Sages with medium-sized flowers produces lax inflorescences, while sages with short-sized flowers exhibit dense inflorescences. Flowers are grouped in different types of inflorescences: dense panicle formed by superimposed verticillasters (*S*. *verticillata* and *S*. *uliginosa*), lax spikes with 4–6 flowers in each whorl (*S*. *pratensis*) and lax racemes with 1–2 flowers in each whorl (*S*. *greggii* and *S*. *blepharophylla*). Also, other characteristics may play a role: corolla length and colours. The corolla length varied from the short size class (\< 1.5 cm) in *S*. *verticillata* and *S*. *uliginosa* up to the medium size class (1.5–3.0 cm) in the other target species. The corolla colours ranged from the blue tones in *S*. *verticillata* (lilac-blue), *S*. *pratensis* (bluish-violet) and *S*. *uliginosa* (sky-blue with white bee-line on the upper and lower lips) up to the red ones in *S*. *greggii* (scarlet red) and *S*. *blepharophylla* (red with an orange undertone). Finally, data suggest that the choice of planting these species with the double function of delighting view and sustain local pollinators should favour *S*. *verticillata* and *S*. *greggii*, with different flower traits, colours and pollinator assemblages. ## GBG as an interactive learning opportunity The Botanic Gardens are places of knowledge, windows on the world of science from which everyone can look out. However, the exhibition contributed a technical and scientific prospective through the development of an awareness- raising process on the importance of the plant biodiversity and how it is linked to the surrounding world. Botanic gardens are very important and worldwide connected by the potential of capacity building in outreach activities and plant rescue and preservation, alive as well as for the respective genetic banks. The exhibition received an extraordinary response from the public in terms of both numbers and overall appreciation of the proposed information. The venues differed between museal institutions and botanical gardens. In both cases, the proposal of the traveling exhibition received considerable success. The major strength points are the unusual key of lecture and the language, correct and rigorous from the scientific point of view, but direct and narrative. In the exhibition, we proposed to abandon the most usual man-centred profitable/productive approach to recount the topics related to Repulsion/Attraction through a plant-centred vision. If on the one hand humans were kept out from the vision, on the other human perceptions were used to show plants as living independent organism similar to ourselves. Pollinators played a main role in this view, thanks to the recent worries related to their decrease and the tentative works to define plant lists to contribute to their survival. # Conclusions We aimed at addressing documented public concerns by translating them into research questions to be investigated in a special context, that of a botanical garden. The work arises from citizens’ apprehension described in scientific publications in the field of social sciences, tackled the alarming reports on pollinators losses and dispersion of exotic species, resumed results in outreach activities and material. The first set of data addressed to worries on exotic species planted in private or common gardens, especially on their interactions with the native fauna. Our results highlighted the establishment of similar relationships between exotic or native plants, and local pollinators. Therefore, it can be assumed that (at least some) exotic species may equally contribute to sustain local pollinator fauna thanks to the resource they provide. On this topic, a large amount of literature has offered lists of possible preferred species to be planted when gardening purposes plan intend to match pollinator sustainability \[i.e. 22,71\]. However, Garbuzov and Ratnieks pointed out a contrasting situation. On the one hand, in existing literature there is generally a low overlap of species even when addressing the same geographic region (and even other pitfalls, as poor recommendations, omitted species, lack of details), finally not providing a decisive list. However, on the other hand, these lists have a strong appeal and could turn them into a communication tool with the potential of driving further research. Our second set of data applied this concept, by choosing renown species used both, for gardening (thanks to their intense and long flowering) and culinary purposes (thanks to their chemical properties). They are very frequently mentioned in gardening articles, printed or online, and used in private as well as public gardens. Our data highlighted that different species of the same genus *Salvia* attracted a different assemblage of pollinators, and therefore a selection among species could be actively performed if their planting is also intended to sustain pollinators. Botanic gardens are present in most cities, and even in small towns. They mostly host a combination of native and exotic species, and research as well as outreach activities. These two competencies are, however, mostly kept separated. We merged them by transforming the botanic garden in the promoter of widening research activity and outputs on plant-animal interactions, by linking them with physical and emotional human experiences. The exhibition and the catalogue describe how plants interact with other organisms and how we (humans) also use similar modes of interaction: attraction means and repulsion modes coming from similar chemical or mechanical approaches. The exhibition, being a travelling one, has and still is enlarging the audience of this output; the catalogue stands as an inspiring tool. # Supporting information We are indebted to enthusiastic students and colleagues that helped in data collection: Marco Palamara Mesiano, Serena Malabusini, Davide Zanovello, Katia Valloggia and Giacomo Tassera, and to Silvana Munzi for her priceless time for discussion and comments on the first draft of the manuscript. We also are grateful to Angela Ronchi and Patrizia Berera for their invaluable contribution in the realization of the panels and the catalogue of the travelling exhibition “Seduzione Repulsione, quello che le piante non dicono”. We thank Emanuele Albini and Mauro Folli, gardeners at GBG, for the precious care of the plant collection. [^1]: The authors have declared that no competing interests exist.

# Introduction Recent surveys of pet owners have found that human attachment to their pets (animals kept within a domestic setting for personal interest, entertainment, or companionship) plays an important role in their lives as evidenced by how they refer to their pets–as friends, children or fur babies, and members of the family. Owners often attribute human qualities to their pets describing them as thinking, emotional, and creative. Young adults in distress may be more likely to turn to their dogs than to some family members. More than half of pet owners report they are closer to their pet than to their own parents and 95% of dog owners hug their companions every day. Humans have owned pets for many millennia. Evidence suggests that the first pet, the dog, was domesticated by 30,000 years ago, likely for practical purposes such as tracking and hunting. Archaeological evidence of a social bond between dogs and people, as companions, is found by 14,000 years ago suggesting that people have valued animal companions for thousands of years. Evidence of nonhuman primates (hereafter referred to as primates) as pets is similarly ancient. In an Iranian cemetery dating to 4,800 years ago, a rhesus macaque (*Macaca mulatta*) was buried with grave goods and in the same manner as the human children in the cemetery. The macaque had pathologies of the hindlimbs indicating that it was inadequately cared for (possibly kept in a cage too small to allow it much movement) and had been imported as macaques are not endemic to Iran. Approximately two-thirds of all households in the United States include pets. Much research has touted the benefits of owning a pet to one’s physical and mental health. Additionally, specially trained pets can provide service to their owners, and even save their lives (e.g. Medical Alert Dogs (MADs)). People have a strong social and emotional attachment to their pets. When asked why they own pets, reasons included having companionship, having a play partner, and the need to love and care for another creature and most pet owners believe their pets are good for them. Though dogs and cats are the most common type of pet, primates are also kept as pets. It is estimated that there are over 15,000 primates owned as pets in the United States, which is a small, but not insignificant number. This is despite the almost universal opinion of scientists and veterinarians that primates do not make good pets; primate pet ownership is detrimental to the primates themselves as well as to their human owners. Many primate pet owners do not have a sufficient understanding of their species or of how to care for them which can result in nutritional deficiencies, injuries, and behavioral disorders. Primates are naturally aggressive and injuries to their owners are not uncommon. Also, several diseases can be transmitted from primates to humans (e.g. parasites, Salmonella, rabies). Owning an exotic pet is an increasingly larger part of the wildlife trade network, it is the third largest illegal trade, and monkeys are becoming “ever more fashionable” as pets \[20, p. 2408\]. A study of the global trade in exotic pets found that primates and carnivores were the most often traded mammals, though mammals were less common than birds and reptiles. This study also found that primates were rarely traded at large physical markets. Research on pet primates has been increasing, though most studies focus on pet primates in habitat countries or countries near habitat countries that themselves have endemic primates (see, for example, Reuter et al., for Madagascar, Duarte- Quiroga & Estrada for Mexico; Ceballos-Mago et al., for Venezuela; Nijman et al., for Indonesia). Systematic studies examining the primate pet trade in the United States are rare. The goal of this research is to preliminarily investigate the pet primate trade, including which taxa are for sale in the United States, the sex and age of the primates for sale, the seller’s location, and the price to provide a baseline for future research. # Methods ## Ethical research considerations Though we were collecting publicly available information from the surface web, the research was also approved by an ethics oversight committee (Institutional Review Board, University of Utah) in 2019 as an amendment to IRB_00079146. The complete dataset is stored on an encrypted hard drive and all personal identifiers were removed from the analysis. Data were collected manually twice a month for 12 months (June 2019-June 2020) from six publicly available online exotic pet-trade websites (their identifications are not being included to avoid unintentionally raising their visibility). Prior research has suggested that data collected from the Internet can provide a reliable indicator of the exotic animal pet trade globally and can be useful in describing some aspects of the exotic animal trade. Additionally, the use of the Internet in the trade and trafficking of exotic animals has been increasing. Though primates are for sale through a variety of venues, publicly available exotic pet sale websites were selected because data are less personalized than social media, such as Facebook or Instagram, thus reducing ethical issues with data collection. For example, some Facebook groups require that you join which involves a certain level of deceit. E-commerce websites were located through a search engine (Google) using the phrases “monkey(s) for sale United States” and “primates(s) for sale United States” and we examined the first 50 results from each search for relevance. Search results that did not return e-commerce websites showing primates for sale (e.g., websites discussing the appropriateness of pet primates, forums, blogs), local classified advertising websites, such as city-specific websites, and websites that did not list specific primates for sale (e.g., websites that only included phrases such as “contact … for availability”) were excluded. From the remaining results, we selected the six most popular sites, as indicated by their ranking in the search engine results, that had the most primates for sale in our initial review. These six websites were then searched for “monkey”, “primate”, “lemur”, “loris”, “galago”, “bushbaby”, “baboon” and “ape”. We recognize that our keyword search was limited. However, more extensive searches were done for the first two months in our data collection process (e.g., broadening our keyword search for specific types of primates, specifically those listed in state regulations, such as “tarsier”, “prosimian”, “indri”, “sifaka”, “marmoset”, “tamarin”, “capuchin”, “saki”, “uakari”, “muriqui”, “guenon”, “langur”, “macaque”, “mangabey”, “mandrill”, “drill”, “gibbon”, “chimp”, “gorilla”, and “orang”), but these searches did not yield any results not initially found by our primary search terms, thus we limited the remainder of our keyword searches for efficiency. For each advertisement, we recorded the date the advertisement was posted, seller’s username, common name of the primate, sex, age, location, and price. Very few advertisements listed a genus or species and most descriptions used common names (e.g. marmoset, capuchin) so we aggregated our data by these general types. We also downloaded photos of the primates for sale. The seller’s username and photos were recorded to avoid re-counting the same advertisement if it was posted for multiple weeks or on multiple websites. Advertisements that were not clearly for a specific primate for sale (e.g. some advertisements stated “contact us to see what primates we currently have”) were not recorded in our dataset. Additionally, animals that were clearly misidentified (e.g. listed as capuchin, but photos in advertisement were clearly cercopithecine) were excluded from the analysis; two advertisements were removed for this reason. To prevent inflating the number of primates for sale, we recorded the number of primates for sale as one regardless of the number of primates in photos provided in the advertisement and despite the use of plural descriptions (e.g. monkeys) unless the advertisement specifically noted they had more than one primate for sale (e.g. if an advertisement stated three primates were for sale, then we recorded that as three primates). Advertisements that stated that there were “several” or “multiple” or “many” primates for sale were recorded as having two primates for sale. We were unable to verify the validity of these online advertisements and we are aware that pet scams exist, that they had been increasing, but that there was a decline in 2020. To reduce the impact of scams, we examined the accompanying photos as an indicator that the animal was in the possession of the seller and compared photos on different websites because posting the same advertisement can be an indicator of a scam. Also, five of our six websites have been in existence for over 10 years. Scams often offer pets for free or at a deeply discounted rate, sometimes mentioning that the animal must be sold because of a family hardship; this tactic was not present in any of the advertisements we recorded. Also, Price found that scam websites often targeted a single taxon while our websites listed a wide variety of taxa (e.g. mammals, birds, reptiles). Finally, we searched Pet Scams ([petscams.com](http://petscams.com)), which catalogs pet scam websites, and Scam Detector ([scam-detector.com](http://scam- detector.com)), an official contributor to the Federal Trade Commission which uses a complex algorithm to yield a “trust index”, for all six of our websites. None of our websites were included in the Pet Scams catalog and the Scam Detector “trust index” ranged from 55.7 to 75.8 (100 is the most reputable); the websites that scored below 70 did so because the Scam Detector algorithm determined their website was poorly designed. Sex and age were compared using a Chi-Square Goodness-of-Fit test with the null hypothesis that the number of males and females would be equal and that the number of primates for sale in different age class would be equal. Ages were binned into three categories (\< 1 year, 1–2 years, \> 2 years) because animals were often identified with descriptors (e.g. “baby”) rather than by chronological age. Primates listed as “baby” were included in the \< 1-year category while primates listed as “adult” were included in the \>2 year category, though we recognize this binning of ages does not take into account the life history of the species (i.e. different species reach adulthood at different ages). Range, mean, median, and mode were calculated for price. Spearman’s Correlation (α =.05) was used to test for an association between the median price per taxa with pooled mean adult body size (mass) and with the number of primates for sale (“supply”). # Results We identified 551 primates of 14 different taxa for sale. Marmosets were the most common taxon of primate for sale (36.7%, N = 202) followed by lemurs (21.6%, N = 119), capuchins (11.3%, N = 62), and squirrel monkeys (10.5%, N = 58. We also found one baboon and one mandrill, both male, for sale. The majority (69.1%, n = 381) were platyrrhines, 21.6% were strepsirrhines (lemurs were the only strepsirrhines for sale), and only 8.9% (N = 49) were catarrhines with green monkeys being the most common catarrhine (4.9% of total primates for sale, 55% of catarrhines for sale, N = 27). Of the 288 advertisements which identified sex of the primate, 61.8% were male (N = 178), and 38.2% were female (N = 110) (Chi-Square = 16.056, df = 1, P = 0.00006). Of the 467 advertisements listing primate’s age, there were more primates under 1 year for sale (78.6%, N = 367), compared to 1–2 years (4.7%, N = 22) and older than 2 years (16.7%, N = 78) (Chi-Square = 440.264, df = 2, P\<0.00001). Primates were found for sale in 22 states. Florida had the most primates for sale (45.7%, n = 252), followed by Tennessee (11.8%, n = 65), Texas (11.6%, n = 64), Missouri (6.7%, n = 37), and North Carolina (5.6%, n = 31). The price of the primates ranged from \$500 for a capuchin or marmoset to \$15,000 for a spider monkey with a mean of \$4,618.14 ± SD \$2931.20, a median of \$3,800, and a mode of \$3,500. Price varied widely within and between taxa. For example, prices for capuchins ranged from \$500 to \$13,500. Average price per taxa was not correlated with the size (r_s = 0.2531, p = 0.40411) or supply (r_s = -0.11279, p = 0.71371) of the primates. The two largest primates (baboon and mandrill) were near or below the sample average (baboon = \$1,500 and mandrill = \$5,000) while marmosets, the smallest primates, were just under the sample average (marmoset average = \$3,624.31). # Discussion Our finding of 551 pet primates for sale over the course of one year underestimates the true numbers because we only examined six exotic pet websites advertising primates for sale in the United States. However, primates are for sale through numerous other venues including private and commercial breeders, auctions, social media, and even pet stores. We had no *a priori* hypotheses for what types of primates would be for sale in the United States. Prior research on pet primates generally occurs in habitat countries with wild-captured animals (see, for example,) and, therefore, provided little foundational background for the U.S. market of captive-bred animals. However, several possibilities for the drivers of the primate pet trade in the United States exist. In a market-driven system, buyers (demand) are likely influencing the sellers (supply). Perceived “ease” of care and/or the depiction of primates in the media may influence the type of primate that buyers desire. Additionally, the owning of a pet primate in general (as opposed to owning a specific type of primate) may be fashionable or a symbol of status. Most profitable breeding practices may be a driving factor for the sellers. These possibilities are not mutually exclusive. ## Potential drivers of the primate pet trade in the United States ### Popularity of primate taxa in the media The desire to own exotic pets may stem from the influence of their popularity in television shows and films. Studies have shown that the sale of green iguanas increased after Jurassic Park and the sale of red-eared slider turtles increased after the release of Teenage Mutant Ninja Turtles. Similarly, the sale of clown fish increased after the release of Finding Nemo and the sale of owls increased after the Harry Potter movie franchise premiered (though see). Similar results have been reported for common pets. For example, the movie Snow Dogs was associated with an increased popularity of Siberian huskies and border collies. Several releases of, or sequels to, 101 Dalmatians were associated with trends in the increase of dalmatian sales and Jack Russel Terriers became trendy with the TV show Frasier. This trend goes back several decades as Old English Sheepdogs became popular with the release of the Shaggy DA films in the 1960s. Primates have been featured in films since the 1930s. Aldrich found that between 1990 and 2013 chimpanzee were the most common primate “actor” (42% of films). While Fleury asserted in 2013 that chimpanzees were the most common primate pet in the United States, the Chimp Care website indicates there are currently 25 chimpanzee pets in the United States and we found no chimps for sale in our study. In 2015, the Fish and Wildlife Service (FWS) Rule 80 FR 34499 classifying captive chimpanzees as “endangered” (effectively ending the private legal ownership of chimpanzees by requiring permits, granted only for scientific research,) probably drove this reduction in the number of pet chimpanzees. Though we note that it is illegal in most states to own a chimpanzee, we did find other illegal primates for sale (e.g. a mandrill was advertised for sale in Florida even though they are barred from personal possession through the Captive Wildlife Licenses and Permits Rule 68A-6.002 of the Florida Fish and Wildlife Conservation Commission and eight primates were for sale in three states, Colorado, Georgia, and New York, where ownership is banned). According to Aldrich the other primate “actors” were capuchins (33%), cercopithecines (macaques and baboons) (13%), orangutans (8%) lemurs (2%), and gorillas (2%). One capuchin, Crystal, is credited in 26 films including George of the Jungle, the Doctor Doolittle films, the Night at the Museum films, and the Hangover films. In our study, we found no gorillas or orangutans for sale, only one baboon, and nine macaques. Our results for capuchins and lemurs provide limited support for the idea that popularity in film and television may be a driver for the choice of primate for a pet. Capuchins, the third most common primates for sale, have been featured regularly in films and TV shows and, if including animated shows such as the Madagascar franchise, so have lemurs, the second most common primate for sale. Adults who were raised watching Zoboomafoo or the Madagascar franchise may see a sifaka or a ring-tailed lemur as a potential pet and not as a wild animal. In interviews with primate pet owners, some speak of how their childhood love of books and movies depicting primates spurred them to purchase primate pets in adulthood. According to Aldrich, no marmosets, the most common type of primates for sale, were featured in films between 1990 and 2013. As this study did not survey owners, it is not known whether these choices in pets were related to their depictions in film and/or on television. ### Popularity of primates in general While the main goal of this project was to uncover which types of primates were for sale, future research addressing why primates are for sale as pets in the first place is needed. While many of the advertisements we examined described the primates for sale in favorable terms, such as “cute”, “cuddly”, “friendly”, “sweet”, “adorable”, “loving”, and “tame”, most scientists and veterinarians agree that primates do not make good pets because they are aggressive, they are difficult to properly care for, and they can transmit diseases. Recent research on preferred dog breeds suggests that fashion, and not good sense, can drive the selection of specific breeds. In the United Kingdom, each of the top 50 dog breeds had at least one inherited disorder. Additionally, dog breeds are not selected based on good health, longevity or good behavior (i.e. the most popular dogs are often those with high frequencies of genetic disorders and poor behavior). Similarly, owning a pet primate may not be based on good sense and it is not uncommon for primate owners to admit that they did not realize what they were getting into when deciding to get a primate describing their pets as “risky”, “unpredictable”, and dangerous” and that caring for the primates is “labor intensive”, “totally consuming”, “expensive”, “You never know which ones will grow up and attack”, “It’s sad to watch the depression they go through if they’re not getting enough one-on-one attention”, and “If given the chance to turn back the hands of time, … ‘I wouldn’t have a pet monkey’”. Relatedly, primate ownership, like other exotic and more common pets, may be related to status. The average cost of a pet primate in our study, around \$4,000, is significantly more than the average cost of a dog or a cat. Purchasing such an expensive pet may serve to advertise the owner’s economic class and is an example of conspicuous consumption; it provides a boost to the owner’s ego. Additionally, primate pet owners, along with owners of other exotic pets, have described several reasons for wanting to own something other than a dog or cat, including that it’s “cute”, it’s “cool”, to “show off”, to “impress others”, and as a way to “get attention”. Social media could increase the desire for a pet primate. Several famous individuals have posted selfies with pet primates. For example, Justin Bieber and Chris Brown owned pet capuchins, while Kristie Alley owned pet lemurs and Michael Jackson owned a pet chimpanzee named Bubbles. Other celebrities, while not owning pet primates, have posted selfies with primates. For instance, Rihanna posted a selfie with a loris and Paris Hilton posted selfies with several different primates, including a capuchin and an orangutan. Non- traditional celebrities, such as bloggers, YouTube personalities and Instafamous, are even more influential. Kelvin Peña, otherwise known as Brother Nature, has posted selfies on Instagram with capuchins and a baboon at a Los Angeles sanctuary. He has also posted selfies on both his Instagram and Twitter with a cotton top tamarin and a ring-tailed lemur inside his home, making it likely they are his pets. Brother Nature has roughly 2.4 million followers. On TikTok, @heresyourmonkeycontent has almost a million followers and frequently posts video content of their tufted capuchin. Some primates are social media celebrities in their own right. For example, Pizzatoru, is a galago with an Instagram account and has 240,000 followers. Comments posted in response to photos of primates often express a desire to own one. Photos of celebrities with primates can be especially problematic because many consumers copy celebrities to enhance their own self-esteem. ## Size, sex, and age of primates for sale Size may be a driving factor in the choice of pet primates as smaller primates are easier to care for (e.g. smaller primates require less food, less space) and manage/handle than larger ones. They require less space and less food and may be easier to handle than larger primates. The adult size of the types of primates for sale range from 0.3 kg (marmosets) to 31.6 kg (male mandrill). In our study, 77% of the primates for sale were under 2 kg as adults and 95% were under 5 kg. A preference for small animal pets has also, but not universally, been reported for dogs. Posage et al., posit that a preference for small dogs could be because small dogs are easier to control, an idea supported by surveys of dog owners, and which would also be true for pet primates. Also, small versions of other pets, such as Munchkin cats, miniature pigs, and even miniature cows have become more popular. Despite the high percentage of smaller primates, we found listed for sale, primates do not make good pets no matter the size, as noted by veterinarians and scientists. While size is likely a driver in the choice of primate to buy and to sell, other factors are probably also at play. Smaller primates tend to reach adulthood and begin reproducing earlier and have shorter gestation periods and interbirth intervals compared to larger primates. This faster life history can increase their reproductive success; thus, they can produce more offspring over a shorter period of time. This is especially true of marmosets, the most common type of primate for sale (37% of advertisements). Marmosets regularly produce twins, with triplets also being common in captive colonies and, with post-partum estrus, they can produce two litters in a single year. In fact, marmosets and tamarins have the highest potential fecundity and fertility of any haplorrhine primate. In one study, captive *Callithrix jacchus* produced an average of 3.66 offspring per year and had an interbirth interval of 217 days. Consequently, smaller primates, especially marmosets, may be preferred by buyers because they are perceived as easier to care for than larger primates, and they may be preferred by sellers because of their higher reproductive output. Breeding practices to maximize reproductive output may also account for the sex- bias in the primates for sale. In our sample, two-thirds of the primates for sale were male. Among breeders of other animals, females make up the largest portion of the breeding establishment (see, for example,). Reproductive efficiency is not highly impacted by the number of males owned, but by the number of females; in other words, a breeder only needs a limited number of males to produce the maximum number of offspring. This could lead to breeders keeping more female offspring and selling more male offspring. The sale of infants over older animals is unsurprising as puppies and kittens are also overwhelmingly preferred as pets. Also, selling the offspring as infants reduces the amount of care and resources that must be invested by the breeder, thereby increasing profits. These profits would be compounded as the mothers, once infants have been weaned, can begin producing their next litter. Additionally, younger primates are less aggressive having not yet reached sexual maturity. ## Conservation implications The primates we found for sale are presumably from breeders and not wild-caught. It has generally been illegal to import primates since the 1975 Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES) entered in force. Also, a study of the global exotic pet trade found limited trade from primate-habitat continents (e.g. Central and South America, Africa) to North America (though see whose analysis suggests substantial export of squirrel monkeys and capuchins from Central and South America). Nonetheless, primates in captivity impact the conservation of their wild counterpart. Research has shown that when people see primates outside of their natural habitat, it can increase their desire to own a primate themselves (which can drive the extraction of primates from the wild in habitat countries), lead them to believe they are not endangered, and decreases the likelihood they will contribute to conservation; all of these will impact wild populations. Ross and colleagues report that 35% of respondents did not know chimpanzees were endangered because of their frequent presence in films and television shows. Squirrel monkeys, capuchins, and ring-tailed lemurs were also seen as acceptable pets to individuals who have seen them depicted in close contact with humans out of their natural environment. Primate pet owners may also think that by owning and/or breeding primates, they are saving them from extinction. As Reuter & LaFleur note, because of the global connectedness through social media, every primate kept as a pet is driving, either directly or indirectly, the capture of primates in the wild and is, therefore, impacting their conservation. Unfortunately, a study by Moorhouse et al., found that informing a potential owner about an exotic species’ conservation status did not impact the desire to own one. In fact, rarity of a species (i.e. those more highly endangered) may have the opposite effect and increase their attractiveness as a pet. ## Reducing the pet primate trade With the estimated high number of primates as pets in the United States and the seemingly healthy market of primates for sale as pets, this is a problem that will continue. There are several avenues for reducing the number of pet primates, including educational advertisement campaigns and legislation. One path towards reducing pet primate ownership is targeted advertising campaigns to educate those considering purchasing a primate. Moorhouse et al., found that advertisement campaigns focusing on disease transmission and legal consequences could reduce the demand for exotic pets by 39% while information on their conservation status or welfare would not reduce the demand at all. In short, potential owners of exotic pets were swayed by information on how owning an exotic pet would affect them, but not by information on how it would affect the pet primate. This study provides sound guidance for those seeking to educate the general public about the unsuitability of primates as pets. Secondly, Beetz proposes that a tax on exotic pets could limit the pet trade through simple economics. Similar sin taxes have been successful in changing behavior in other areas such as cigarette and alcohol use and the consumption of unhealthy foods and drinks (see, for example). A recent study on cigarette usage found that state taxation, federal taxation, and an anti-smoking advertising campaign all significantly reduced the purchase of cigarettes, but the taxation reduced cigarette consumption more than the advertising campaign. Since primates are so expensive to begin with, the tax would have to be high to make an impact. Additionally, for those purchasing primates as a sign of status, it is unlikely that taxation would impact their decision. Another path involves federal regulation of the pet primate trade. Currently, there are no federal regulation on the ownership of pet primates, though bills, such as s H.R.3135/S.1588 –Captive Primate Safety Act—which would ban the interstate trade and private ownership of primates, have been introduced in Congress multiple times, most recently in May 2021, but not yet passed. Instead, the regulation of the primate pet trade is handled state by state creating a patchwork of laws. Federal legislation requiring permits to own primate pets may be another route to decrease their ownership if ownership cannot be banned outright. The recent decrease in the number of pet chimpanzees following the implementation of permit requirements for all captive chimpanzees (see above) suggests this may be a viable option. # Conclusion Our study of six exotic pet trade websites uncovered over 500 primates for sale over a one year period. This despite the belief of primatologists and veterinarians that primates should not be owned as pets. They can transmit diseases to their owners and they can be aggressive, especially after reaching sexual maturity. Also, they are difficult to care for properly. Most primates were for sale in Florida and were small-bodied primates, male, and infants; this is unsurprising as it would be most profitable for the sellers focus on smaller primates with their faster life history and to sell young primates and keep female primates for breeding. While the pet primates for sale are likely captive bred, they can still have consequences for their wild counterparts. Making the ownership of pet primates illegal through federal regulation would likely reduce (but not eliminate) the number of pet primates both by making the purchase of one more difficult and by fear of legal consequences (as found by Moorhouse). Research on primate pet owners themselves, including why they decided to own a primate, what drove their decision in the type of primate to purchase, and how the primate was purchased are urgently needed because understanding the demand for pet primates is the first step in reducing the demand. # Supporting information We would like to thank E. Goodrich, M. McKnight, and three anonymous reviewers whose comments greatly improved this manuscript. [^1]: The authors have declared that no competing interests exist.

# Introduction Multiple sclerosis (MS) is a chronic disease of the central nervous system in which there are recurrent injuries both to the myelin sheaths that surround the nerve axons and also, to a lesser extent, to the axons themselves. Although it is often said that MS does not affect mortality, the available data suggest that patients with MS have a life expectancy that is decreased by about 5–10 years compared to the general population when matched for age and sex.. Only limited data are available regarding the causes of death (CODs) that account for the excessive mortality found in the MS population. In part, this is due to inherent limitations that arise from the manner in which CODs are documented on death certificates. Death certificates typically have 5 positions for categorizing a death., In Part I, the first position (the immediate COD) is meant to be the condition (eg, hepatic encephalopathy) that led directly to the death. The next 3 positions (the underlying CODs) are for those causes that led sequentially to the immediate cause (eg, chronic alcoholism leading to cirrhosis of the liver leading to the hepatic encephalopathy). The fifth position (Part II), is for significant conditions, contributing to the death but not leading directly to the underlying causes (eg, malnutrition, ascites, etc).. The use of death certificate data leads to some inaccuracies in establishing the relationship between MS and specific CODs. For example, the use of the underlying COD (as is typical for the reporting of vital statistics) is often inaccurate because MS itself is frequently listed as the sole underlying cause, and, as such, this entry provides no insight into whether a particular complication of end-stage MS (eg, an infection) may have actually led to the death rather than MS itself. Similarly, the use of the immediate COD may result in other errors. For example, many physicians list cardiac or respiratory arrest as the immediate COD in the first position on the death certificate. Nevertheless, because these entities represent the final common pathway of all deaths, they are not helpful for understanding what condition actually led directly to the death (eg, pneumonia, myocardial infarction, gunshot wound, etc). We recently reported the results from a retrospective cohort study in which we compared survival and mortality patterns in patients with MS drawn from the OptumInsight Research (OIR) database. This database contains the billing/claims data from a national commercial health insurance plan in the United States (US). We used a matched cohort design, including both patients with MS and non-MS comparators from the same source population and analyzed mortality data covering the period of 1996–2009. In this study, we were able to show that mortality rate among patients with MS was approximately twice that observed among matched controls.. In the present study, we attempted to gain insight into those CODs, which contribute to the excessive mortality seen in MS patients. To accomplish this, we developed a novel algorithm for identifying (from death certificate data) the cause most directly leading to the death. We refer to this as the “principal” COD. We compared our new principal COD algorithm with standard death certificate–based definitions using the underlying and immediate COD assignments to evaluate the extent to which this new method might provide improved COD information among individuals with a chronic illness such as MS. # Methods The project utilized a commercially available database, anonymous with respect to any individual-identifying information, as well as publically available death certificate data. Individual institutional approval was, therefore, unnecessary. ## Selection of Subjects and Determination of Mortality – OIR Claims Database Subjects were drawn from the OIR health-care claims database. The database contains billing claims information for over 39 million individuals insured through United HealthCare; there are approximately 15 million covered lives per year and 7.5 million patients with lab results. The database is geographically diverse and is representative of the commercially insured population of the US.. Patients with MS were selected for inclusion in the study if they met the following criteria: - Inclusion in the database for ≥3 months during the period from 1996–2009 – these years were selected to cover the time from the start of the modern period of treatment with MS disease modifying therapy (DMT) to the most recently available data at the time of analysis. - Age ≥18 years at the time of their first ICD9 diagnostic code of 340 (the code for MS). - At least 2 ICD9-340 diagnosis codes ≥30 days apart *or* <u>at least 1</u> ICD9 diagnostic code of 340 and ≥1 billing code for DMT, defined as any of the following drugs: interferon beta-1a, interferon beta-1b, glatiramer acetate, or natalizumab. The reason for limiting consideration to only these agents was because these are MS-specific therapies that were available during the time period covered by the analysis and, therefore, their use is likely to reflect a true MS diagnosis. Similarly, control subjects were selected for inclusion if they met the following criteria: - Inclusion in the database for ≥3 months and during the period from 1996–2009. - The same age, sex, and residence region (using the US Census categories of Northeast, Midwest, South, West) during the index year (ie, the year of entry into the insurance database) as the matched patient with MS. Up to 3 matched non-MS control subjects were selected for each patient with MS as described above. Deaths among the selected subjects were identified by linkage with the National Death Index (NDI) and the Social Security Administration Death Master File (SSA DMF). Standard algorithms, which included Social Security number, were used to determine matches. A death was considered valid if it was identified through either source. Mortality information for each subject was searched through the end of 2009. ## Determination of Cause of Death Determinations of COD were based solely from the information provided on the death certificate for those subjects whose death was identified by linkage with the NDI. This information was available for 1,451 (91.9%) of the deaths among patients with MS, and 2,127 (91.2%) of the deaths among non-MS controls. The information was not available for 128 (8.1%) deaths among patients with MS and 205 (8.8%) deaths among non-MS controls who were identified through the SSA DMF without NDI linkage. All deaths without death certificate information were categorized as being due to an unknown cause. As shown in, one of the ICD10 codes listed in Part 1 of the death certificate was assigned as the COD for each of the 3 approaches to COD assignment evaluated in this study; the immediate, the underlying, and the principal COD methods. The ICD10 codes for immediate and underlying COD were based on standard death certificate categorizations for Part 1 of the death certificate. The principal COD method used the ICD10 code at the top of Part 1 of the death certificate, with the following exceptions: ICD10 codes indicative of suicide were always considered the principal COD, regardless of its position on the death certificate; MS was considered the principal COD if the only ICD10 codes, which preceded the MS code on Part 1 of the death certificate, were those indicative of cardiac or respiratory arrest or both or if MS was the only code mentioned. Otherwise, MS was not considered the principal COD. Finally, ICD10 codes indicative of cardiac or respiratory arrest were only considered the principal COD if no other ICD10 codes (including MS) appeared on Part 1 of the death certificate. Independently of the process for assigning ICD10 codes for each of the 3 COD methods, a 4-member author panel (consisting of a physician with experience in death certificate completion, an MS neurologist, and 2 general physicians) identified 7 principal disease/injury categories of COD. These are deaths due to: 1) cardiovascular disease, 2) cancer, 3) infections (including pulmonary infections), 4) pulmonary disease (excluding pulmonary infections), 5) MS, 6) accident, and 7) suicide. The panel then reviewed all the ICD10 codes noted in any position on death certificates and assigned them into 1 of these 7 categories. If a subject had ICD10 codes that did not fit into any of the prespecified categories (eg, codes for deaths due to diabetes, liver disease, kidney disease, etc), they were categorized as death due to “other known causes.” Death certificates indicating only cardiac or respiratory arrest were classified as death due to “cardiopulmonary arrest,” and deaths without any information were classified as death due to an “unknown cause”. ## Analytical Methods Using the principal, underlying, and immediate COD methods, the overall and cause-specific mortality rate per 100,000 person-years with 95% confidence intervals (CI) were calculated for each category of death, and selected subcategories, in the MS and non-MS control populations. Differences in mortality rate per 100,000 person-years with 95% CI between patients with MS and non-MS controls were also calculated for each major category of death and selected subcategories. # Results ## Accuracy of the MS Diagnosis As discussed in the Methods, for the purposes of this study, the diagnosis of MS required either: a) ≥2 MS diagnostic codes or b) ≥1 MS diagnostic code plus a DMT prescription. The presence of a diagnostic code together with a DMT was considered to be sufficient evidence that the subject actually had MS. To determine the accuracy of the diagnosis when no DMTs were prescribed (36% of the total identified MS population), a medical chart review was conducted in 85 randomly selected such patients (thus demographically representative of the patients who did not receive DMTs). Sixty-two of these patients (73%) were judged to have definite MS –32 based on physician notes, 3 based on their use of DMTs despite not having a claim for this in the database, 26 based on both factors, and 1 based on the opinion of the neurologist reviewer. Therefore, assuming that the presence of an MS diagnostic code together with use of an MS- specific DMT accurately identified a patient as having MS, then the maximum expected false-positive rate was only 10% (0.37×0.27). Therefore, we concluded that the use of the “2-code” definition provided sufficient evidence that the subject actually had MS. ## Description of the Study Population De-identified data for 30,436 patients with MS and 90,123 controls who met the initial screening criteria for selection were provided by OIR. When matching was performed according to Kaufman et al, these procedures led to the exclusion of 34 patients with MS and 305 controls, resulting in 30,402 cases and 89,818 non-MS controls being available for our analyses. Of these, 29,411 cases (97%) had 3 matched controls; 594 (2%) had 2, and 397 (1%) had 1. Because of the matching procedures used, the distribution of demographic factors was identical for the MS and comparator cohorts. Each population was composed of 77% women and had a mean age at the index date of 44 years. The majority resided in either the Midwest (32%) or the South (44%), and almost all were born after 1920. The peak birth decades were the 1950s (31%) and 1960s (30%). The MS cohort had a somewhat shorter coverage period in the database than controls, but their post-index-date coverage was actually longer. The majority (64%) of patients with MS had ≥1 DMT prescription during their period of coverage. A total of 1579 deaths (5.2%) were observed in the MS cohort, with 2332 deaths (2.6%) in non-MS comparators. The overall mortality rate among the MS patients was approximately double that among non-MS controls (899 vs 446 per 100,000 person-years). The difference in mortality rate between the 2 cohorts was an excess of 453 deaths per 100,000 person-years in the MS population. Further details of mortality rates according to sex and region have been published elsewhere.. ### Cause-specific mortality using principal, underlying, and immediate COD As shown in, using the principal COD method, the categories with the highest mortality rate among patients with MS (other than MS itself) were cardiovascular disease (185 deaths per 100,000 person-years), infection (134 deaths per 100,000 person-years), cancer (104 deaths per 100,000 person-years), and pulmonary disease (77 deaths per 100,000 person-years). MS accounted for 133 deaths per 100,000 person-years, the third-highest mortality rate in the MS cohort. By contrast, among non-MS comparators, the categories with the highest mortality rate were: cardiovascular disease (125 deaths per 100,000 person-years), cancer (106 deaths per 100,000 person-years), infection (39 deaths per 100,000 person- years), and pulmonary disease (31 deaths per 100,000 person-years). Among patients with MS, using the underlying COD from the death certificate, the disease/injury category with the highest mortality rate was MS, accounting for 286 of the 899 deaths per 100,000 person-years. Other CODs with relatively high mortality rates included cardiovascular disease and cancer. In contrast to the principal COD, the underlying COD indicated a higher mortality rate due to MS (133 vs 286 deaths per 100,000 person-years). The mortality rate due to infection (56 deaths per 100,000 person-years) was less than half of that attributed to infection by the principal COD (134 deaths per 100,000 person- years). The most common underlying CODs in the non-MS comparator cohort were cancer and cardiovascular disease. The immediate COD indicated the highest mortality rate due to cardiorespiratory arrest, followed by cardiovascular disease, MS, and infection. Unlike the principal COD method, the immediate COD indicated the highest rate of mortality as a result of cardiorespiratory arrest (7.4 vs 148 deaths per 100,000 person- years). The mortality rate attributed to MS was similar using the principal and immediate COD approach (133 vs 128 deaths per 100,000 person years). Cardiovascular disease accounted for the highest mortality rate in the non-MS comparator population when using the immediate COD. ### Differences in cause-specific mortality rate between the MS and comparator cohorts The relative contribution of each major disease/injury category to the overall difference in mortality rate is illustrated in. Based on the principal COD algorithm, the largest contributors to the difference in mortality rate between patients with MS and non-MS comparators (other than MS) included: infections (21.0% \[95/453 deaths per 100,000 person-years\]), cardiovascular disease (13.2% \[60/453 deaths per 100,000 person-years\]), and pulmonary disease (10.2% \[46/453 deaths per 100,000 person-years\]). Notably, all 3 of these COD categories had higher mortality rates when using the principal COD approach than either the underlying or the immediate COD approaches. Together, these 3 COD categories accounted for 44.4% of the total difference in mortality (201/453 deaths per 100,000 person-years). Moreover, using then the principal COD method, only 29.4% (133/453) of the difference was attributed to MS, and cardiac/respiratory arrest accounted for almost nothing. Nevertheless, despite the fact that these specific diseases accounted for the bulk of the excessive deaths observed, it is noteworthy that for every other disease category (with the exception of cancer), there was an excessive number of deaths in the MS population compared with the non-MS cohort. Based on the underlying COD method, the differences in mortality rate between patients with MS and non-MS controls were mostly attributed to MS, which accounted for the difference in 63.1% (286/453) of the cases. Almost none of the difference was attributed to cardiac/respiratory arrest by this method. According to the immediate COD method, the differences in mortality rate between patients with MS and non-MS comparators were attributed to MS in 28.3% (128/453) of the cases and cardiac/respiratory arrest in 16.6% (75/453) of the cases. ### Subcategory analysis of principal COD Using the principal COD, subcategories of infection, pulmonary disease, and cardiovascular disease were assessed to gain greater insight into the categories that were particularly important contributors to the excessive mortality that MS populations experience. In the infection category, almost all of the excess mortality was attributed to either pulmonary infection or sepsis (90.6% \[86/95 deaths per 100,000 person-years\]). In the pulmonary disease category, more than half of the difference in excessive mortality was attributable to aspiration (58.7% \[27/46 deaths per 100,000 person-years\]). By contrast, within the category of cardiovascular disease, no single subcategory stood out as a contributor to the excess mortality. # Discussion In this study we demonstrated differences in the cause-specific mortality rate derived from the principal, the underlying, and the immediate COD reporting methods. An important goal of our study was to understand which non-MS conditions lead to the excess mortality in patients with MS. Using the principal COD method, it is apparent that infections, cardiovascular diseases, and pulmonary diseases are the most important contributors to the higher mortality rate seen among patients with MS. Moreover, analyses of subcategories of these principal COD categories indicated that certain disease states (eg, sepsis, pulmonary infection, and aspiration), which seem very likely to be intermediate steps on the pathway leading from advanced MS to death, were among the most important contributors to the excessive mortality observed among patients with MS. By contrast, and interestingly, death from suicide (which *a priori* we considered to be MS-related) was not significantly different between the MS and control populations. The main advantage of the principal COD method is that it decreases the number of cases in which cardiac/respiratory arrest or MS was the designated COD compared with the underlying and immediate COD methods. In fact, over 30% of the total mortality in the MS cohort was attributed to either MS or cardiac/respiratory arrest when using the underlying or the immediate COD approach (290 and 276 deaths per 100,000 person-years, respectively). The immediate and the principal COD methods estimated a similar mortality rate from MS itself (128 and 133 deaths per 100,000 person-years, respectively), a finding that suggests that if physicians place MS in the first position on the Part 1 of the death certificate, they almost never consider another condition to be underlying the death. Similarly, most patients (except those with rare brainstem lesions in vital locations) do not die directly from their MS. Therefore, when MS was listed as the immediate COD, it provided little information other than to indicate that MS was part of the causal path leading to death. Combined, these 2 categories accounted for 45%–64% of the total difference in mortality rate (compared with non-MS controls) when using the underlying and the immediate COD methods. By contrast, with the principal COD method, these uninformative classification categories accounted for only 30% of the difference in mortality. Indeed, even for other chronic disease states, these particular COD categories provide very little information about what actually caused the death and, thus, the principal COD method would be usefully applied to these other chronic diseases. In addition, this method will give practitioners better insight as to how the care of their patients might be improved in the future. These observations are mostly consistent with previously published experience, although there are differences in the methodologies that have been used in different studies. For example, in a report from the Danish MS Registry, the authors found that in 82% of deaths MS was listed as an underlying or contributing cause and, in 56.4% it was listed as the underlying cause. By contrast, the category of “infectious and respiratory diseases” accounted for only 4.7% of the deaths in their study. This apparent reversal in the relative importance of MS as a cause of death when comparing the Danish cohort with the results of the present study, as discussed earlier, likely reflects the non- informative nature of MS as an underlying cause and underscores the greater insight provided by our use of the “principal” COD for categorizing deaths in MS patients. Using a different methodology from ours, a retrospective cohort study from Great Britain also reported results similar to ours. That study compared mortality in MS patients with that experienced by a matched control group, and reported that MS patients had an excessive mortality due to pulmonary infections, other pulmonary disease, and cardiovascular diseases. Lastly, in the 21-year long-term follow-up of MS patients from the pivotal interferon beta-1b trial, 98.4% of the original patients were identified 21 years after the start of the randomized controlled trial (RCT), and over this interval 22% (81/366) of the patients had died. Of the adjudicated deaths 78.3% (54/69) were found to be MS-related. Patients receiving placebo during the RCT experienced a greater all- cause mortality rate compared with those patients on active therapy, and this excessive mortality was largely due to MS-related causes, especially pulmonary infections.. In our study, we used an insurance claims database to study survival patterns in MS, an approach which has potential advantages and disadvantages compared with other design strategies. Its main advantage (particularly in the US) is that, in such databases, standardized ICD10 codes are widely applied so that data is easily collated and analyzed. Moreover, in the US, death reports are obtained from 1 national source (the NDI), which has a specific format, and which can be easily linked to the claims database. Finally, because the population sizes are huge, the use of a claims database provided excellent statistical power. There are, however, potential disadvantages as well. First, in the OIR data, the diagnosis of MS was based on the person having either 2 MS diagnostic ICD9 codes or 1 such code plus a prescription for a DMT. The diagnosis was not based on a review of the medical record or an examination of the patient. Nevertheless, the presence an ICD9-340 code together with a DMT prescription (which was present in 64% of patients with MS) seems to be conclusive evidence of a patient’s physician’s belief that the patient had MS. However, in the remaining 36% (where the DMT prescription is lacking) the diagnosis may be less secure. We, therefore, specifically conducted a chart review in a cohort of 85 such patients and confirmed that the diagnosis was correct in 73%. This observation suggests that the *maximum* proportion of false positives in our study population is 10% (or less) of the total. This degree of diagnostic uncertainty would not materially affect our findings. Second, in claims data, patients were not identified at either diagnosis or onset but rather only after their enrollment in the insurance plan. Therefore, claims data do not include either the diagnosis date or the disease-onset date, so that survival can only be measured from birth rather that from those time points, which have been the principle focus of most earlier reports., – Nevertheless, both cases and controls had to survive until plan enrollment; they were age- and sex-matched so that both cases and controls survived to the same chronological age to be included. Moreover, cases and controls were also matched on their first opportunity to die and, consequently, immortal time bias cannot account for our findings. Third, because the mean age at the first MS diagnostic code was 44 years, we commenced our follow-up of the patients with MS, on average, 15 years into their illness. Consequently, our study cannot answer the question of which CODs are important for patients with MS who experience very early mortality. Nevertheless, because the vast majority of patients with MS survive substantially beyond 15 years from the onset of their disease, – this potential gap in our data is probably of little practical consequence. Fourth, the relatively brief period of enrollment in the health plans did not allow for a detailed measurement of comorbidities. While it would be desirable at some point to evaluate the impact of other diseases on CODs in patients with MS, the available data were insufficient for a rigorous assessment. And finally, because the patients with MS from the claims data were all commercially insured, they are representative of the commercially insured population and probably not of the total MS population in the US. Therefore, one must exercise caution when extrapolating the present results to other settings. In summary, the principal COD method provided an improvement over other methods of using death-certificate data to determine COD, such as the underlying or the immediate COD methods. By partially resolving the uninformative category of “death due to MS” and largely resolving the uninformative category of “death due to cardiopulmonary arrest,” this method focuses attention on the excessive deaths due to other causes ( and). Most conspicuously, these causes include infections (both pulmonary and genitourinary as well as sepsis) and aspiration pneumonias , which can be seen as expected complications in bed-bound patients. Similarly, the excessive numbers of deaths from ischemic heart disease and embolic disease may represent the consequence of immobility in advanced-stage patients and their inability to exercise. Also the greater number of deaths from accidental respiratory obstructions may reflect the common impairment of swallowing, which is known to accompany the later stages of MS. Thus, these particular complications seem likely to be the consequence of advanced stages of MS. Nevertheless, interestingly, excessive deaths were also observed in other disease categories (eg, accidental poisoning), which have a less obvious connection to advanced disease. The basis for this observation is not known. Moreover, because the principal COD method provides better insight into the actual conditions that lead to death in MS patients, it will be useful in assessing (over time) the impact of therapies (either DMTs or aggressive treatment of infections) on mortality in MS. In addition, a greater awareness of the common causes for the excessive death rate that MS patients experience (eg, fatal pulmonary infections, sepsis, and aspiration) will allow physicians to anticipate potential problems and, thereby, improve the overall care of their patients with MS. We are grateful to the patients and the investigators for their contributions to the study. Additionally, we thank Robert C. Ristuccia, PhD for assistance with preparation of the manuscript. [^1]: This project was sponsored by Bayer HealthCare Pharmaceuticals, the employer of S Reshef, and MJ Rametta. V Knappertz and D Pleimes were salaried employees of Bayer HealthCare Pharmaceuticals. M Corwin and H Golub are employed by Care-Safe, V Knappertz by Teva Pharmaceutical Industries and D Pleimes by Myelo Therapeutics GmbH. D Goodin has participated (or is currently participating) in several industry-sponsored clinical trials in multiple sclerosis; the sponsoring pharmaceutical companies for these trials have included (or do include) Ares-Serono, Merck Serono, Novartis, Berlex Laboratories, Bayer Schering HealthCare, Biogen Idec, Schering AG, and Teva Neuroscience. He has also lectured at both medical conferences and in public on various aspects of the epidemiology, diagnosis, and management of multiple sclerosis, and in many cases these talks have been sponsored directly or indirectly by one or another of the above listed companies. He has served as a temporary ad hoc consultant to several of these organizations on several occasions. M Corwin has received personal compensation for activities with Bayer HealthCare Pharmaceuticals, Ameritox, and Genova Diagnostics as a consultant. Dr. Corwin has received personal compensation for participation as a Medical Monitor and on Data and Safety Monitoring Committees for Edimer Pharmaceuticals, SynapDx, BioDelivery Sciences International, Stratatech, KV Pharmaceutical, and Meridian Medical Technologies. D Kaufman has received personal compensation for activities with Bayer HealthCare Pharmaceuticals, McNeil Consumer Healthcare, and UCB as a consultant. Dr. Kaufman has received research support from McNeil Consumer HealthCare, CSL Behring, Onyx Pharmaceuticals, Incyte Corporation, and Sanofi-Aventis. H Golub has received personal compensation for activities with Bayer HealthCare Pharmaceuticals, Ameritox, MedicaMetrix, MoMelan Technologies, and Genova Diagnostics as a consultant. Monitoring Committees for Antisense Therapeutics Limited, Sanofi-Aventis, Bayhill Pharmaceuticals, Bayer Pharmaceuticals, BioMS Pharmaceuticals, Daiichi- Sankyo, GlaxoSmithKline Pharmaceuticals, Genmab Biopharmaceuticals, Medivation, Peptimmune, PTC Therapeutics, Teva, Vivus, NHLBI, NINDS, and NMSS. He also received consulting and speaking fees and served on advisory boards for Alexion, Accentia, BaroFold, Ciba Vision, Biogen Idec, Novartis, Consortium of MS Centers (grant), Klein Buendel Incorporated, Enzo Pharmaceuticals, Somnus Pharmaceuticals, and Teva Pharmaceuticals. G Cutter has received personal compensation for participation on Data and Safety. 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]: Performed the experiments: MC DK HG. Analyzed the data: DSG MC DK HG SR MJR VK GC DP. Wrote the paper: DSG MC DK HG SR MJR VK GC DP.

# Introduction Impairments in working memory (WM) are regarded as a fundamental cognitive deficit in schizophrenia. WM refers to the short-term storage of information in the service of the active guidance of behavior. It is crucial for a broad range of cognitive operations, and WM impairments can lead to deficits in social and occupational functioning. Spatial WM deficits are present in high-risk populations, in spectrum disorders, and in unaffected relatives and therefore have been discussed as a potential endophenotypic marker for schizophrenia. Although WM deficits are well documented in schizophrenia, the specific factors that cause these deficits are not yet known. Considerable evidence suggests that the capacity of visual WM is restricted to about three or four organized chunks. Nevertheless, stable and substantial differences in WM capacity can be found across healthy individuals, which may reflect poor functioning of attentional control in the service of WM. For instance, healthy individuals with low WM capacity show reduced selectivity during WM encoding and as a consequence, may store task-irrelevant information, whereas those with high WM capacity efficiently filter out irrelevant information. The selection of relevant information to be stored in WM is associated with activitv in the prefrontal cortex and the basal ganglia and excerts control over the parietal cortex where the information is stored. Moreover, recently it has been shown that the time needed for healthy individuals to disengage from a distracting event is related to WM capacity. Although there is considerable evidence that low WM capacity in healthy participants is associated with inefficient attentional selection, it is less clear whether a failure of attentional selection of incoming information at an early stage of processing also contributes to the severe impairments in WM observed in patients with schizophrenia (PSZ). Attentional abnormalities have long been thought to be a central feature of schizophrenia, however its role in explaining their severe WM impairments has not been resolved yet. To determine whether poor WM in PSZ stems from problems of attentional selection at an early stage of processing, it is necessary to disentangle the specific types of attentional mechanisms at specific stages of processing in schizophrenia. Attentional selection is required when a subject is confronted with competing incoming stimuli and needs to restrict processing to a subset. Several studies revealed that the selection process is intact in PSZ when the relevant stimuli are salient, having a bottom-up processing advantage. Similarly, PSZ are unimpaired in their ability to select relevant among irrelevant stimuli for WM storage, when salient cues are given. However, when the selection process requires a high degree of top-down control, performance in PSZ is markedly reduced. For instance, the search time per item is significantly increased in PSZ when the target in a visual search task is embedded among highly similar distractors. Recent evidence suggests that impairments in top-down driven attentional selection occur not only on the level of perceptual processing but also on the level of WM encoding. For instance, in the presence of highly distracting stimuli, PSZ were impaired in their ability to efficiently select task-relevant items for WM encoding. In the present study, we investigated the relationship between impairments in WM and top-down control vs. bottom-up, stimulus-driven processes in attentional selection within one paradigm, i.e. the attentional capture paradigm. The basic mechanisms underlying attentional capture and its relation to WM capacity have been revealed in studies on healthy participants. However, it is largely unclear to what degree these mechanisms are impaired in PSZ and whether the potential impairment is related to the patients' reductions in WM capacity. To determine WM capacity we used a visual change detection task that has been extensively used in studies of visual WM in healthy participants,. The attentional capture paradigm that we used was developed by Fukuda and Vogel. This task allows us to test the ability to resist interference from a distractor under conditions that require either high or low demands on top-down attentional control. The first step of this task involves a visual search which requires subjects to briefly view an array of four colored Landolt Cs that are presented within placeholders for a specific target item. Subjects are asked to report the orientation of the single item that has the target color. On some trials, a task-irrelevant colored box (flanker) is briefly presented flanking one of the placeholders. The flanker could either match the color of the target item (relevant-flanker condition) or appear in a different color (irrelevant-flanker condition). In the irrelevant-flanker condition, attention would be captured automatically and involuntary depending only on the relative saliency of the physical properties of the flanker such as its sudden onset (stimulus-driven capture). As a result, visual search times are likely to be increased reflecting additional costs on processes needed to disengage from the salient stimulus feature. In the relevant-flanker condition, when the flanker appears suddenly and its color matches the color of the target, top-down driven effects in addition to stimulus-driven effects are likely to contribute to the capture effect and lead to a further increase in visual search times. According to the attentional capture hypothesis, the observer would have adopted a top-down attentional set and stimuli that match this set would capture attention (contingent capture). Thus, the contingent capture effect is thought to depend critically on the observer's intentions and goals (i.e., top-down processes), rather than the physical properties of the stimulus per se (i.e., bottom-up processes) but an alternative interpretation is also possible. To isolate top-down driven effects in the present paradigm we subtracted response accuracy in the relevant-flanker condition from response accuracy in the irrelevant-flanker condition (see Methods). Typically, in visual search tasks, PSZ show increased search times regardless of distractor type and set size. This finding may reflect a general deficit in the speed of cognitive processing rather than a specific deficit in the attention process. To account for general cognitive slowing, we implemented a staircase procedure in the present task. This allowed us to equate task difficulty in the no-flanker condition by individually establishing the presentation time of the search array needed to reach a criterion. Taking individual differences in the processing speed into account, we expected that PSZ would exhibit increased susceptibility to contingent attentional capture but not stimulus-driven attentional capture. Moreover, we predicted an inverse relationship between deficits in the contingent attentional capture task and WM capacity in PSZ. This would be taken as evidence that reduced resistance to interference from distractors at an early stage of processing results in WM encoding problems later, thus pointing to an attention-based account of the WM deficits in schizophrenia. # Methods ## Change Detection Task ### Participants Thirty outpatients with schizophrenia (*n* = 23) or schizoaffective disorder (*n* = 7) (PSZ) participated. Diagnoses were made according to Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) criteria using structured clinical interviews. The Brief Psychiatric Rating Scale (BPRS), the Scale for the Assessment of Negative Symptoms (SANS), and the Scale for the Assessment of Positive Symptoms (SAPS) were used to assess symptoms. Twenty-eight healthy control subjects (CO) without a history of DSM-IV Axis 1 disorders and no family history of psychosis were recruited from the community. CO were medication-free and screened to rule out schizotypal personality using the Schizotypal Personality Questionnaire (SPQ). PSZ and CO were matched for age (*t*₅₆ = 1.56, *P* = .13), IQ (*t*₅₆ = −1.41, *P* = .16), and handedness (*t*₅₆ = −1.18, *P* = .24). Years of education were lower in PSZ than CO (*t*₅₆ = −2.33, *P*\<.05). All subjects had normal or corrected-to-normal vision. Exclusion criteria were a history of head injury, neurological disorder or substance abuse in the six months preceding the study. All subjects gave written informed consent approved by the Vanderbilt University Institutional Review Board (IRB) and were paid. ### Stimuli, Task, and Procedure Stimuli in the change detection task were colored (red, green, blue, yellow, purple, black, and white) squares (1.2°×1.2°), presented in randomly selected positions within a centered 11.4°×11.4° region on a gray background. In each trial, participants were presented with arrays of two, four, six, or eight colored squares for 150 ms (memory array). After a retention interval of 900 ms, one colored square (test probe) was presented at the location of one of the items from the memory array. Participants made an unspeeded button press to indicate whether the color of the test probe matched or did not match the color of the original memory item in that location. Half of the trials were matches. An inter-trial interval of 1 s followed. Each of the four experimental conditions was presented equally often (40 trials per condition). Participants performed 10 practice trials followed by an experimental block of 160 trials that were presented in a randomized order. To quantify WM capacity we used an equation developed by Pashler and modified by Cowan : K = (hit rate+correct rejection rate−1)×N. This approach allows us to estimate the number of items held in memory, K, from an array size of N items, taking guessing into account. The K estimate is conceptualized as a limit in the number of discrete slots that holds a single item, which is appropriate for the change detection tasks with highly distinguishable stimuli, such as categorically different colors. The K estimate has become a standard measure of change detection performance because it corrects for response bias and allows comparisons across different array sizes, conditions, and groups. In this study, we first transformed each individual's accuracy for each array size into a K estimate. For each subject, we then calculated the mean K value across the four array sizes. The mean K values were correlated with performance in the attentional capture task (see below). ## Attentional Capture Task ### Participants The same PSZ participated as in the change detection task. Twenty-eight CO were matched for age (*t*₅₆ = 1.38, *P* = .17), IQ (*t*₅₆ = −1.33, *P* = .19), and handedness (*t*₅₆ = −0.78, *P* = .44). PSZ were less educated than CO (*t*₅₆ = −2.7, *P*\<.01). 26 CO had participated in the change detection task. All subjects gave written informed consent approved by the Vanderbilt University IRB and were paid. ### Stimuli, Task, and Procedure Stimuli were four colored Landolt Cs (1.0°×1.0°) that appeared within placeholders (1.8°×1.8°) on a black background. The placeholders were present throughout the duration of each trial. The target item was a red C. The three distractors appeared in blue, magenta, or green. In each trial a search array consisting of one target and three distractors was presented for a fixed duration that was determined for each participant using a staircase procedure (see below). Participants were instructed to identify the orientation of the single target with the target color (red). Shortly following the onset of the search array, a multi-colored pattern mask was presented at each placeholder location until a response was given. Participants indicated if the gap in the target item was on the top, right, left, or bottom, by pressing one of four arrow keys. Response accuracy was emphasized. An inter-trial interval of 2 s followed. There were three types of trials, randomly intermixed. One third of the trials were flanker-present trials, i.e. at varying intervals prior to the onset of the search array a task-irrelevant colored box (flanker; 1.0°×1.0°) was presented for 50 ms at a position that flanked the position of one of the placeholders (but never the position of the target item). In two thirds of the trials, no flanker was presented. On flanker-present trials, there were four possible stimulus onset asynchronies (SOAs) between the flanker and the search array: 50 ms, 150 ms, 250 ms, and 350 ms. In half of the flanker-present trials the flanker was drawn in the target color (relevant flanker), in the other half of the flanker-present trials the flanker was green (irrelevant flanker). Participants performed 10 practice trials followed by eight experimental blocks of 120 trials, with all conditions randomly intermixed within blocks. ### Analysis The dependent variable was response accuracy. Because response accuracy was emphasized we did not analyze reaction time. Response accuracy was first assessed as a function of group (PSZ vs. CO) and flanker condition (no flanker, irrelevant flanker, relevant flanker) across SOAs. Because performance did not differ between PSZ and CO in the irrelevant-flanker condition (see below), the stimulus-driven effect was not further assessed. In the relevant flanker condition when the flanker appeared in the target color, response accuracy reflects a combination of two potentially separable effects: attentional capture by the sudden onset of the flanker (stimulus-driven capture) and attentional capture by the target color (contingent capture). The contingent capture effect was isolated from the stimulus-driven effect by calculating the difference in accuracy between irrelevant-flanker trials and relevant-flanker trials ( = contingent capture cost) and compared between groups as a function of SOA ### Staircase Procedure Before participants performed the attentional capture task, we titrated the duration of the search array for each subject so that each subject's performance was approximately 75% correct in the no-flanker condition. In the staircase procedure, participants were initially presented with four placeholders for 500 ms. Participants were informed that a target (i.e. a red square with a gap on one side) would appear in one of the placeholders along with three distractors (i.e. differentially colored squares with a gap on one side) filling the other placeholders, and they were instructed to indicate the direction of the gap of the target with a button press. In the first trial, the target was presented for 500 ms, and thereafter, the target exposure duration was modulated from trial to trial in a following manner to obtain an individualized exposure threshold with which each individual can perform the task with approximately 75% accuracy. If a participant correctly identified the direction of the gap, the exposure duration for the next trial was shortened by 10%. On the other hand, if he/she responded incorrectly, the exposure duration for the following trial was increased by 30%. Each participant performed one practice block of 10 trials followed by four experimental blocks of 60 trials. The search-array durations for the last 20 trials in the four blocks were averaged to estimate the individual baseline search-array duration for the attentional capture task. # Results ## Change Detection Task The mean WM capacity averaged across the four array sizes was significantly lower for PSZ (*M* = 1.60, *SD* = 0.74, range: 0.14–3.14) than CO (*M* = 2.08, *SD* = 0.56, range: 1.26–3.4) (*t*₅₆ = −2.79, *P*\<.01). Lower WM capacity estimates in PSZ are consistent with past findings. WM capacity did not correlate with symptom ratings, chlorpromazine equivalent (CPE) dose, or duration of illness (all *P*-values \>.37) ## Attentional Capture Task: Staircase Procedure The individual baseline search-array durations were significantly longer for PSZ (*M* = 227.3, *SD* = 73.9, range: 85.3–382.3) than CO (*M* = 173.5, *SD* = 63.9, range: 67.3–316.5) (*t*₅₅ = 2.91, *P*\<.01). Removing one outlier from the PSZ, with an exposure time more than 3 SD higher than the group mean, did not change the result. This subject was removed from further analyses. With data collapsed across both groups (26 CO and 29 PSZ who participated in both experiments), the individual exposure time correlated negatively with the WM capacity estimate averaged across the four set sizes (*r* = −.41, *P*\<.01). Trends for a similar relationship were found when the correlations were calculated separately for each group (CO, *r* = −.36, *P* = .07; PSZ, *r* = −.29, *P* = .13). The individual exposure time did not correlate with symptom ratings, CPE values, or duration of illness (all *P*-values \>.72). ## Attentional Capture: Flanker Capture In the no-flanker condition, the mean accuracy of target identification was 76.9% for CO and 75.7% for PSZ. Thus, the staircase procedure was successful in equating task difficulty by individually establishing the presentation time of the search array. A repeated-measures ANOVA was conducted to examine the effects of flanker type (relevant, irrelevant, no flanker) and group (CO vs. PSZ) on accuracy. The analysis revealed a main effect of flanker type (*F*_2,110 = 71.97, *P*\<.001, *η2* = .57), with lower accuracy for irrelevant flankers than no flankers (CO, *t*₂₇ = 2.78, *P*\<.05; PSZ, *t*₂₈ = 4.95, *p*\<.001) and lower accuracy for relevant flankers than irrelevant flankers (CO, *t*₂₇ = 3.64, *P*\<.01; PSZ, *t*₂₈ = 6.24, *P*\<.001). Overall, performance did not differ between PSZ and CO (*F*_1,55 = 1.71, *P* = .20). However, there was a significant interaction between group and flanker type (*F*_2,110 = 3.32, *P*\<.05, *η2* = .06), indicating that the additional decrease in accuracy in the relevant flanker condition was stronger in PSZ than CO. Thus, when adjusting the individual exposure time of the search array, PSZ showed a similar stimulus-driven capture effect as CO (no group difference in the irrelevant-flanker condition, *t*₅₅ = −1.05, *P* = .30), whereas the contingent capture effect was slightly increased in PSZ vs. CO (trend for a significant group difference in the relevant-flanker condition, *t*₅₅ = −1.85, *P* = .07). A previous study in healthy participants demonstrated that individual differences in WM capacity were associated with the time needed to recover from attentional capture rather than the susceptibility to attentional capture. To test whether PSZ and CO differed in their susceptibility to attentional capture or the time needed to recover from it, the contingent capture costs (i.e. accuracy on irrelevant-flanker trials minus accuracy on relevant-flanker trials) were compared between groups as a function of SOA. In the former case we expected PSZ to show higher capture costs than CO, irrespective of the SOA. In the latter case, capture costs should decrease faster (i.e. at earlier SOAs) in CO than PSZ. To test this hypothesis a repeated-measures ANOVA was conducted with the factors SOA \[short (50-ms and 150-ms) vs. long (250-ms and 350-ms)\] and group (CO vs. PSZ). The analysis revealed a significant interaction effect (*F*_1,55 = 3.99, *P* = .05, *η2* = .07) indicating equivalent capture costs in PSZ and CO for short SOAs (*t*₅₅ = 0.06, *P* = .96), but a significant group difference for long SOAs (*t*₅₅ = 2.27, *P*\<.05). Relevant capture costs considerably decreased after the 150-ms SOA in CO, but remained high in PSZ.. We also divided all participants into high-capacity (*M* = 2.43, *SD* = 0.39) and low-capacity groups (*M* = 1.27, *SD* = 0.45) using a median split on the WM capacity and compared contingent capture costs as a function of WM capacity (low vs. high K) and SOA. This strategy followed the logic of the individual differences approach used to explore the relationship between WM and attention in healthy participants. The logic is that if an individual's WM capacity can predict capture costs one can conclude that a common factor underlies both abilities. In the high-capacity group about two thirds of the participants were CO and one third of the participants were PSZ. In the low-capacity group this distribution was reversed. Based on previous findings and our findings in PSZ we predicted a faster decrease in relevant capture costs for high vs. low capacity individuals. Specifically, we expected higher capture costs in low vs. high capacity individuals for long SOAs (SOA 250-ms and 350-ms), which was tested using independent t-tests (one-tailed). The findings revealed that relevant capture costs were significantly higher in low capacity than high capacity individuals at SOA 250-ms (*t*₅₃ = 2.41, *P*\<.01) but did not differ between groups at SOA 350-ms (*t*₅₃ = 0.81, *P* = .42). With data collapsed across both groups, we calculated the correlation between the WM capacity estimate averaged across the four array sizes and relevant capture costs for each SOA. Although there was no relationship with WM capacity at the 50-ms and the 150-ms SOA (all *P*-values \>.41), significant negative correlations emerged at SOAs 250-ms (*r* = −.33, *P*\<.05) and 350-ms (*r* = −.28, *P*\<.05). When calculated separately for each group, a trend for a similar relationship was found at SOA 250-ms in PSZ (*r* = −35, *P* = .06, all other *P*-values \>.41) but not CO (all *P*-values \>.16). Furthermore, relevant capture costs at SOA 250-ms correlated negatively with capacity estimates derived from two additional tasks, the Letter-Number-Sequencing task - reordered condition and a spatial delayed response task (DRT) with a longer encoding period (see Supporting Information S1). These findings suggest that the degree to which participants, especially PSZ, are impaired in their ability to resist interference from distractors at later SOAs is related to their WM capacity reduction. Relevant capture costs averaged across later SOAs (150–350 ms) correlated positively with SANS scores (*r* = .37, *P*\<.05; *P*-values \>.57 for SAPS and BPRS) but not at SOA 50-ms (*P*\>.66). There was no consistent relationship with CPE values (all *P*-values \>.51) and duration of illness (*r* = .41, *P*\<.05 at SOA 150 ms, all other *P*-values \>.46). # Discussion Our findings provide evidence for preserved stimulus-driven attentional capture but impaired contingent attentional capture in schizophrenia. Both groups showed equivalent capture costs when an irrelevant flanker preceded visual search, whereas capture costs were increased in PSZ compared to CO when the flanker was drawn in the target color. This new finding points to a selective impairment in the ability to resist distractor interference when top-down control, is required to disengage from the distractor. However, when the capture is purely stimulus- driven , disengagement from the distractor is unimpaired in PSZ. These results are consistent with previous reports on impaired attentional selection in schizophrenia only under conditions of top-down control. Furthermore, our results indicate that not only the process of selecting the relevant information but also the inhibition of the irrelevant information is impaired in schizophrenia. To our knowledge only stimulus-driven attentional capture has been studied in schizophrenia so far, and the existing findings are inconclusive. Consistent, with our findings, Ducato et al. did not find increased stimulus-driven attentional capture by a distractor that changed in color in PSZ. Moreover, PSZ were able to resist interference from irrelevant moving distractors to the same degree as CO when the attentional load of the central task was low. However, these investigators also found increased susceptibility to attentional capture by irrelevant moving distractors in PSZ under conditions that allowed participants to control automatic capture. Although these findings have been discussed in terms of domain-specific impairments in stimulus-driven attentional capture in schizophrenia, they are also consistent with the conclusion of a specific impairment in selective attention in schizophrenia when top-down processes are involved. PSZ needed considerably more time than CO to find the target stimulus in the no- flanker condition. Thus, it was indeed crucial to adjust for differences in the processing speed when assessing attentional capture in the PSZ. When cognitive slowing was taken into account, we found no impairments in stimulus-driven capture in the PSZ. We found evidence for increased contingent capture costs in PSZ vs. CO, however these differences appeared only at later SOAs. In CO, capture costs substantially decreased at SOA 250-ms but remained consistently high in PSZ even at the latest SOA. When dividing all participants into low- and high-capacity groups we found a similar pattern of capture costs with the main difference between groups at SOA 250-ms. Because the majority of subjects in the low-capacity group were patients it is difficult to disentangle the effects of WM capacity from those related to the disease. However, Fukuda and Vogel previously demonstrated in a student sample that recovery time from attentional capture was faster in individuals with high vs. low WM capacity. The faster recovery time observed in this study compared to our study, i.e. contingent capture costs substantially decreased already at SOA 150-ms in the high-capacity group, is most likely due to age differences in the study samples and/or the inclusion of patients in our high-capacity group. WM capacity decreases with age and indeed, WM capacity was lower in the high-capacity group of our study than the high-capacity group in the previous study. Together these findings provide evidence that recovery time rather than susceptibility to attentional capture is the critical factor that distinguishes PSZ from CO (i.e. low from high WM- capacity individuals). The findings of this study also have important implications for an attention- based account of WM deficits in schizophrenia. We reasoned that if PSZ experienced reduced resistance to interference from distractors at an early stage of processing this would result in WM encoding problems later. To test this hypothesis, we correlated WM capacity with performance in the attentional capture task. As expected, WM capacity was markedly reduced in PSZ. The capacity estimates were lower than those reported in previous studies, which might be due to the short stimulus presentation duration. The capacity estimates observed in the CO were also lower than the usual estimate of about 3–4 items. Thus, it is possible that insufficient encoding time in high WM load conditions reduced WM capacity in our participants, however it seems that this affected WM performance to a similar degree in both groups. Consistent with an attention-based account of WM deficits in schizophrenia, we found a significant negative correlation between capacity estimates and the individual exposure time of the search array, which indicates that the degree to which participants needed more time to find the target was related to their degree of WM capacity reduction. Moreover, for relevant capture costs, there was an inverse relationship with WM capacity at later SOAs that was driven in particular by PSZ. The observed correlations may reflect common processes other than those related to attentional selection, such as generalized cognitive impairments or reduced visual encoding. However, PSZ did not show a deficit in the irrelevant flanker condition and irrelevant capture costs did not correlate with WM capacity across all participants (all *P*-values \>.26). These results do not suggest a general processing deficit but indicate a specific impairment. Due to the short stimulus presentation time it is possible that reduced visual encoding contributed to reduced WM capacity in both groups. However, the target presentation time in the attentional capture task was individually determined. Thus, reduced encoding probably did not influence capture costs. Therefore, it seems that the common process that was impaired in both tasks and reflected in the correlation was not solely associated with visual encoding. Moreover, the group comparisons yielded a similar pattern of results as the continuous analysis showing that capture costs were considerably higher in PSZ than in CO, and higher in low- vs. high-capacity individuals at later SOAs. Finally, we demonstrated that relevant capture costs at SOA 250-ms also correlated negatively with capacity estimates derived from two additional tasks, the Letter-Number-Sequencing task and a spatial DRT that included a longer encoding period, speaking against an explanation in terms of impaired visual encoding. In the light of previous findings that low-capacity individuals have difficulties in filtering out irrelevant information for WM encoding, we propose that attentional dysfunctions contribute, at least partially, to WM deficits in schizophrenia. Importantly, our findings indicate that attentional selection in the service of WM is not globally impaired in schizophrenia. Rather, our data suggests that impairments in attentional selection can have detrimental effects on WM encoding specifically when top-down processes are involved. This is consistent with a previous study showing that the ability to select relevant information for WM encoding is markedly reduced in patients when the distractors have strong competitive advantage over the targets. In addition to the correlational evidence derived from the present study, manipulating the demands on attentional selection during WM encoding will be necessary to further specify the attentional processes that are impaired or spared in the context of WM in schizophrenia. Consistent with previous findings, our results suggest that the processes that are most vulnerable to top-down attentional dysfunctions are those required for WM encoding. However, the ability to resist interference from distractors is also crucial to keep the memory representation stable over time and to retrieve the information from WM. Thus, attentional dysfunctions are likely to affect WM at different stages of processing in schizophrenia. It is important to note that recovery time from attentional capture was slowest in PSZ even though their WM capacity was higher than the estimated capacity in the low-capacity group (including PSZ and a proportion of CO) who showed no difference in capture costs compared to high capacity individuals at the latest SOA. This finding suggests dissociable effects of WM capacity and schizophrenia on contingent capture costs. It seems that the mechanisms underlying low WM capacity that might be similar in PSZ and low-capacity healthy individuals cannot fully account for the increased capture costs in the PSZ and we must consider the functional impairments that are unique to the illness. We also found that relevant capture costs, rather than WM capacity, were associated with the severity of negative symptoms. It is unclear whether negative symptoms are more likely to be associated with reduced top-down control of attention when competition between different incoming stimuli needs to be resolved. Previous studies did not report a relationship with negative nor positive symptoms. Attentional impairment comprises one subscale of the SANS. Although this construct is rather unspecific in the scale, its items correlate with the other negative symptom subscales. Thus, it seems plausible that negative symptoms such as avolition or blunted affect lead to difficulties in the effortful allocation of attention and reflect a unique contribution of the illness to increased relevant capture costs. Given the various types of attention, this finding does not exclude the possibility that other types of attention are associated with positive symptoms. In this study, all PSZ except for two were medicated with the majority taking second-generation antipsychotics. To assess the influence of medication, we correlated task performance with CPE. Neither WM capacity, nor the individual exposure time of the search array, or capture costs correlated with CPE. Moreover, CO with low WM capacity were medication-free and showed relevant capture costs as well. Together these findings argue against a major effect of medication on task performance. Understanding the processes that contribute to impaired WM in schizophrenia is crucial in the search for cognitive remediation strategies. Recent studies have highlighted impairments in perceptual encoding rather than memory retention and there is some evidence that theses deficits can be improved with behavioral training. Our results add to these findings by suggesting that training of top- down attentional selection has the potential of enhancing WM encoding, thus pointing to different routes in the development of targeted behavioral therapies for WM deficits. # Supporting Information We thank Heath Nichols, Katy Thakkar, and Joel Peterman for their assistance with clinical interviews and subject recruitment, and Prof Stephan Heckers with his help in recruitment. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JSM KF EKV SP. Performed the experiments: JSM. Analyzed the data: JSM SP. Contributed reagents/materials/analysis tools: EKV SP. Wrote the paper: JSM KF EKV SP.

# Introduction In 1812, Benjamin Rush observed that individuals characterized by a psychopathy- like disorder were primarily afflicted by a ‘derangement in the moral faculties’. Recent conceptualisations of psychopathy suggest that it is a disorder characterized by callousness, egocentricity, grandiosity, often leading to profound antisocial behaviour. Psychopathic offenders understand moral and societal rules, but show a striking indifference to those rules or any consequence of breaking them. When confronted with moral choices, psychopaths show no reluctance to inflict personal harm if it helps to achieve their goals. The extraordinarily destructive effects and massive societal cost of antisocial and psychopathic behavior underscore the paramount importance of making progress in our understanding of this disorder. Factor analysis of one of the predominant conceptualizations of psychopathy, Hare’s Psychopathy Checklist-revised second edition (PCL-R) , reveal at least two dimensions of this disorder. Items that map onto Factor 1 denote interpersonal/affective traits, simplified with the term ‘emotional detachment.’ Factor 2 items are behavioural symptoms such as impulsivity or criminality simplified with the term ‘antisocial lifestyle’. Although the neurobiology of psychopathy in general is not well understood, recent data has begun to offer clues to the underlying circuitry disrupted in this disorder. Limbic structures such as the amygdala are strongly involved in emotion, the processing of emotionally laden information and the learning of conventional and moral rules. Recent work shows that psychopathic offenders have smaller amygdala volumes bilaterally, including deformations in the basolateral, lateral, cortical and central nuclei. By contrast, volumetric increases of the amygdala have also been observed in psychopathy. The amygdala is connected with cortical regions that control and modulate subcortical functioning. Connectivity between the amygdala and orbitofrontal cortex (OFC), and ventromedial PFC (vmPFC), is altered in psychopathic offenders and individuals with antisocial personality disorder. The amygdala is connected to the prefrontal cortex by a white matter tract called the uncinate fasciculus, which is disrupted in psychopathic offenders. In line, functional connectivity between the amygdala and the vmPFC is weaker in psychopathy and in boys with psychopathic traits. Furthermore, lesion studies show that damage to the vmPFC decreases the likelihood that individuals will avoid emotionally aversive outcomes. In addition, during fear conditioning psychopaths show no significant activation of the amygdala-OFC network that is recruited in healthy controls. (However, also see). Together these data imply amygdala-vmPFC dysfunction to partly underlie psychopathic behavior. While evidence for disruption in amygdalo-prefrontal circuitry is mounting in psychopathy, others have shown alterations in temporal, parietal and limbic structures. Recent functional neuroimaging work has implicated mesolimbic dopaminergic abnormalities in healthy individuals with psychopathic traits. Healthy individuals with greater psychopathic tendencies have greater activation of the ventral striatum and anterior cingulate cortex during reward anticipation. Impairments in the mesolimbic reward system were recently also found in violent offenders, suggestive of strong reward drives without consideration of potential punishment. In addition, smaller volumes of the nucleus accumbens have recently also been observed in offenders with psychopathy . Although disruption of the mesolimbic reward system has been associated with impulsive and violent behavior, this involvement has never been directly shown in psychopathic offenders, and certainly not in white matter circuitry. The major objective of our study was to examine diffusion based measures of white matter throughout the brain in psychopathy to clarify the disrupted neural circuitry in this disorder. We completed diffusion tensor imaging (DTI) in psychopathic offenders and healthy controls, using a tract based spatial statistics (TBSS) approach that provides us with analysis of diffusion-based measures in white matter that is an optimized voxel-wise approach voxels and is not limited by *a priori* assumptions regarding which circuitry is vulnerable in this disorder. Based on recent findings, we hypothesized that (i) psychopathic offenders would be characterized by disruption of white matter connecting amygdala to prefrontal cortex, and (ii) we would find disruption of white matter from nucleus accumbens to prefrontal cortex, corresponding to vulnerability of this network from recent functional imaging studies. # Methods ## Participants Psychopathic offenders (mean age ± standard deviation (SD); 33.5±7.4) were recruited from halfway houses in the Greater Toronto Area and through the Law and Mental Health Program at the Centre for Addiction and Mental Health (CAMH). Thirty-eight male offenders were interviewed of which 23 offenders met PCL-R inclusion criteria for this study. Offences included aggravated sexual assault, homicide, human trafficking and kidnap. Of these individuals, 12 psychopathic offenders reoffended within the duration of the study (October 20^th 2010–May 24^th 2011) and had either been arrested or were unlawfully at large. Therefore, 11 male psychopathic offenders were able to complete the entire study. Psychopathic offenders were included if they scored 23 or higher on the Psychopathy Checklist-Revised edition (PCL-R) (average ± SD; 28.1±3.3; range 23–34). In this community sample a relatively liberal cut-off of 23 was chosen to increase the number of subjects that could be enrolled in the study. This cut-off ensures moderate to strong psychopathic traits in criminal populations and allows for a correlational approach to psychopathy which is beneficial with relatively small sample sizes. All included offenders have been administered PCL-R interviews conducted by trained, certified forensic psychologists and psychiatrists for clinical and risk assessment purposes. Exclusion criteria included schizophrenia or any primary psychotic disorder, bipolar disorder, depressive or anxiety disorders and other personality disorders, e.g., borderline personality disorder. Psychopathic offenders were also excluded if they currently met Diagnostic Statistical Manual-IV-TR (DSM-IV-TR) criteria for substance abuse disorder, that is, no history of substance abuse or dependence for at least 6 months. All offenders were subjected to regular drug screens as part of the terms of their parole which indicated none had abused drugs at the time during which they were participants in the study. During a standard safety screening (done by SH) offenders were screened for the presence of neurological (e.g., traumatic brain injury, seizures, history of stroke), psychiatric disorders or contraindications to TMS, which were confirmed with a file review by an experienced clinician (done by TS). Right-handedness was confirmed with the Edinburgh Handedness Inventory. All psychopaths had previously been administered the Shipley Institute of Living Scale and all scored at least in the average range (corresponding to a WAIS-R score ranging from 90–110 indicating likely absence of organic brain damage. The Shipley Institute of Living Scale screens for organic brain damage and has a high correlation (r = 0.85) with the full scale WAIS-R. Healthy age-matched male controls (mean age ± SD; 32.1±6) were recruited through advertisement in and around the Centre for Addiction and Mental Health (CAMH). Psychopathology including a lifetime history of substance dependence, or substance abuse in the past 6 months in the control group was also ruled out through a structured clinical interview for DSM-IV disorders (SCID-I) *(i.e. the* Structured Clinical Interview for DSM-IV Disorders. All healthy controls also received a urine toxicology screen to rule out current substance use. ## Ethics Statement In accordance with the Declaration of Helsinki the Research Ethics Board of CAMH approved of the study and written informed consent was obtained for all participants. The rights of all participants were protected. Psychopathic offenders were explicitly told that enrollment in or drop-out of the study would by no means interfere with their treatment or parole. All included psychopathic offenders received oral and written information about the study and were screened for contra-indications. If they wanted to participate they were asked to sign a Release of Information after which a review of previous psychological assessments was conducted to retrieve the PCL-R score. Hereafter, they were called to ask if they were still interested in participating: this usually gave psychopathic offenders a period of 2–3 days to reconsider their enrollment. Upon arrival to the lab, participants were again briefly told about the procedures and were then asked to sign the Informed Consent. ## Image Acquisition DTI images were acquired using an eight-channel head coil on a 1.5 Tesla GE Echospeed system (General Electric Medical Systems, Milwaukee, WI), which permits maximum gradient amplitudes of 40 mT/m. A single shot spin echo planar sequence was used with diffusion gradients applied in 23 non-collinear directions and b = 1000 s/mm². Two b = 0 images were obtained. Whole brain coverage was obtained (no gap), oblique to the axial plane. Slice thickness was 2.6 mm and voxels were isotropic. The field of view was 330 mm and the size of the acquisition matrix was 128 × 128 mm², with echo time = 85.5 ms and repetition time = 15000 ms. To improve the signal to noise ratio, the entire sequence was repeated three times. ## Tract-based Spatial Statistics The DTI data were converted to 4D NIfTI volumes. The resulting images for all three repetitions were merged for each subject, corrected for motion and eddy current distortion and ultimately averaged using standard tools available from FSL ([www.fmrib.ox.ac.uk/fsl](http://www.fmrib.ox.ac.uk/fsl))). After brain- extraction using BET, FA images were created by fitting a tensor model at each voxel to the averaged diffusion data using DTIFit (FMRIB’s Diffusion Toolbox), implemented in FSL. Voxelwise statistical analysis of the FA data was carried out using TBSS, part of FSL. All FA images were nonlinearly registered to the target image (FMRIB58_FA) provided by the FSL software. Next, the mean FA image was created and the tracts were narrowed to generate a mean FA skeleton which represents the centres of all tracts common to the all subjects. An FA threshold of 0.2 was chosen to discard non-white matter voxels. The area surrounding the skeleton in each subject’s aligned FA map was searched perpendicular to the skeleton voxel and the locally highest FA values were projected onto the skeleton. This ensures that each subject’s skeleton remains in the group space while representing the centers of each individual’s own unique fiber tracts. The resulting individual skeletonised images were fed into voxelwise cross-subject statistics. Finally, group comparisons between subjects with psychopathy and normal controls were carried out with permutation-based analysis. This was achieved with Randomise implemented in FSL, utilizing threshold-free cluster- enhancement method (TFCE). Statistical maps were then thresholded at P\<.05 fully corrected for multiple comparisons (family-wiser error \[FWE\]-corrected). The most probable anatomical localizations for each cluster showing significant between-group differences in FA were determined with the FSL atlas tool using Talairach atlas, Harvard-Oxford cortical and subcortical structural atlases, and the JHU white matter tractography atlas (<http://www.fmrib.ox.ac.uk/fsl/data/atlas-descriptions.html>). ## Region of Interest Analysis Region of interest (ROI) analyses were conducted as post hoc tests on mean FA values extracted from significant clusters resulting from TBSS. Independent- samples t-tests were performed on each cluster to assess group differences. In psychopathic offenders, partial correlation coefficients were then computed between mean FA values extracted from significant clusters (with more than 100 voxels) and PCL-R factors, while controlling for effects of age. Statistical analyses were performed using SPSS v.17.0 for Windows (SPSS Inc., Chicago, IL). ## Probabilistic Tractography To further elucidate the disrupted neural circuitry of psychopathy, significant different voxels between groups were used to seed the tractography algorithm. To generate individualized seed masks, significant voxels resulting from TBSS were projected back onto each control subject’s native space. The use of this approach (i.e. fiber tracking in healthy controls using abnormal voxel clusters identified in a disease group) has been recently described. Probabilistic fiber tracking was conducted separately in each control subject using PROBTRACKX implemented in FMRIB’s Diffusion Toolbox (FDT). Using this algorithm, estimates of fiber orientation and their uncertainty were calculated at each voxel. This model also accounts for the possibility of crossing fibers within each voxel (bedpostX). We used the default parameters with 5000 sample pathways per each seed voxel with a curvature threshold of 0.2 (corresponding to ±80°). Pathways were also terminated after 2000 steps, using a step length of 0.5 mm. Tractography results were then transformed to the Montreal Neurological Institute (MNI) space and averaged using the FA image that required the least warping among all other images during TBSS registration as the registration target. Transformation fields from the TBSS registration stage were used to warp tractography results into the target image space. The transformed images were then averaged and linearly transformed to MNI standard space using the linear transformation matrix between the target image and the MNI standard brain. Finally, the resulting averaged image was thresholded to confine it to probabilities only greater than.5% of the maximum possible number of probabilistic pathways in a voxel (i.e. 5000 × total number of seed voxels). # Results The TBSS analysis showed five non-contiguous clusters of significantly decreased FA in white matter tracts of the psychopathic group as compared with controls (P_FWE \<0.05). These clusters involved uncinate fasciculus and inferior occipitofrontal fasciculus bilaterally, extending towards both subgenual anterior cingulate (BA 25), left amygdala, left orbitofrontal cortex (BA47) and left frontal pole (BA10). In addition, decreased FA was also noted in regions involving bilateral anterior thalamic radiations and their medial extensions pointing to subgenual anterior cingulate. Cluster one corresponded to the subcortical and limbic portions of these white matter tracts (in amygdala, thalamus, subgenual anterior cingulate) and cluster two corresponded to the frontal components of these white matter tracts (in orbitofrontal cortex and frontal pole). The right sided cluster (cluster 3) stemmed from the anterior thalamic radiation and uncinate fasciculus with extensions to thalamus, striatum, pallidum and subgenual anterior cingulate. In general, FA deficits were more pronounced in the left hemisphere. Results showed no significant regions with increased FA value in psychopathic individuals compared with the controls. In the psychopathy group, we found significant correlations between PCL-R subfactors and mean FA values extracted from two of the clusters, while controlling for age. Mean FA of the cluster 2, which mainly involved white matter in ventral regions of left frontal lobe, was negatively correlated with Factor 1 (r = –0.63, P_one-tailed = 0.04, n = 10). In addition, there was a negative correlation between mean FAs derived from the right-sided cluster (cluster 3) and Factor 2 (r = –0.68, P_one-tailed = 0.02, n = 10). However, there were no significant correlations between cluster 1 and PCL-R subscores. For one psychopathic offender only the overall psychopathy score was available. Therefore, correlational analyses were conducted in 10 psychopathic offenders. The probabilistic tractography seeded from voxels with abnormal FA encompassed the following pathways: (i) left and right ATR extending from thalamic nuclei rostrally towards frontal pole and medially to subgenual anterior cingulate and nucleus accumbens (ii) bilateral caudal continuation of the anterior thalamic radiation towards diencephalon and midbrain approaching medial ventral tegmental area/substantia nigra area (iii) left uncinate fasciculus bridging temporal pole (adjacent to the amygdala) and multiple regions in prefrontal cortex (including orbitofrontal cortex, frontal pole, ventromedial prefrontal cortex, and subgenual anterior cingulate cortex). # Discussion In this study, we show that psychopathic offenders have white matter deficits in two dissociable major networks. Namely, an amygdalo-prefrontal network (i.e., white matter tract connecting amygdala to prefrontal cortex) and a striato- thalamo-frontal network (a network connecting the nucleus accumbens, thalamus and prefrontal cortex). Importantly, correlations between the underlying factor structure of psychopathy and white matter tracts were found: we found correlations between (i) PCL-R Factor 1 and integrity of projections from amygdala to medial prefrontal and orbitofrontal cortex, and (ii) between PCL-R Factor 2 and projections from ventral tegmental area and nucleus accumbens through the anterior limb of internal capsule to thalamus and prefrontal cortex. Earlier structural and functional MRI studies have repeatedly implicated disruption of an amygdalo-prefrontal network in psychopathy. Indeed, our findings align with recent data from antisocial personality disorder, children with callous-unemotional traits, boys with conduct disorder and psychopathic offenders, all pointing to disrupted white matter within these regions. Recently, disruption of the uncinate fasciculus, which connects amygdala to vmPFC and OFC was implicated in psychopaths, supporting our findings, providing evidence for the theory that the social and emotional deficits in psychopaths reflect a deficient interaction between OFC/vmPFC and amygdala. A subsequent study that included the same 9 psychopathic individuals from the Craig et al study, plus another six (total n = 15) with antisocial personality disorder found white matter disruption in frontally-based white matter tracts connecting to the limbic system, providing further support for this notion. Furthermore, a recent study of boys with psychopathic tendencies found reduced white matter concentration in frontal cortex, raising the possibility that even mild neural alterations may lead to subclinical behaviors along a continuum. While evidence continues to accumulate for disruption in prefrontal structures and connections from prefrontal cortex to amygdala, there is substantial evidence for structural and functional disruption of the amygdala itself in psychopathy and related disorders. In youths with both conduct disorder and strong psychopathic traits, the amygdala and the OFC are shown to be significantly less active during early reinforcement learning. Blair has suggested that in psychopaths a dysfunctioning amygdala is pivotally important as it results in poor fear conditioning and passive avoidance learning, both of which are implicated in the pathogenesis of psychopathy. In fact, poor fear conditioning in childhood is significantly predictive of adult criminal behaviour suggesting a neurodevelopmental aspect of amygdala damage in psychopathy. Interestingly, a recent study suggests that functional connectivity deficits in limbic areas may already be present at early age and may prelude structural connectivity deficits. Within the limbic system, significant bilateral volume reductions in the amygdala, accompanied by surface deformations in specific amygdalar nuclei provide further evidence for disruption of this core circuitry which correlate strongly with the affective and interpersonal aspects of psychopathy. When our findings are taken together with the existing literature, it is becoming increasingly clear that structural disruption of white matter in the OFC, vmPFC and frontal pole, and white matter connections to these regions from the limbic system may explain the dysfunctional interpersonal and affective aspects of psychopathy. Although some parts of the mesolimbic reward system (e.g., the striatum) have been implicated in psychopathy, our implication of disrupted white matter in a striato-thalamo-frontal network is new. Therefore, the damage in white matter tracts connecting the nucleus accumbens and ventral tegmental area through the anterior limb of the internal capsule to the thalamus and frontal cortex highlights a key second system that may be disrupted in psychopathic offenders. Affected white matter fibers extending to the midbrain appear to parallel those of the medial forebrain bundle, which connect ventral tegmental area with the nucleus accumbens - an important part of reward circuitry. Indeed, we found that disruption in striato-thalamo-frontal circuitry was correlated with Factor 2 of the PCL-R. Other structural MRI studies have found increased volume of the striatum, albeit in different parts. Recent accounts of striatal functioning in antisocial populations have suggested a particular pattern of impairment. While normal controls recruit activity in the anterior cingulate when stimuli lose their rewarding properties, antisocial individuals retain activity in the striatum. Once these individuals are motivated to obtain a certain reward, striatal impairments may result in the inability to flexibly use contextual information to terminate certain behaviors, for instance aggressive or antisocial behavior. A recent study examined mesolimbic dopamine reward system functioning in healthy individuals with psychopathic traits. Using a combination of PET and fMRI in the same individuals, neurochemical and neurophysiological hyperreactivity of the dopaminergic reward system was correlated with impulsive- antisocial temperament, providing potential explanations for increased substance use, impulsivity (however, also see), and violence in this population. Moreover, dopaminergic networks in the VTA and striatum correlate with trait impulsivity, where higher dopamine release in the striatum is predictive of stronger individual desire for drugs. By contrast, Kiehl and colleagues found reduced functioning of the striatum in psychopathic offenders but, as stipulated by Glenn and Yang, this may have to do with the negatively valenced stimuli that were presented. It is important to note that it is unclear whether hyperreactivity of the mesolimbic reward system in healthy individuals would induce the same behavior as white matter disruption of this circuit in psychopathic offenders. That is, white matter damage in a certain network is not typically associated with hyperreactivity. Nonetheless, when our data are taken together with these recent findings, evidence suggests that disruption of the reward system circuitry may underlie a group of behaviors common to psychopathic offenders, including aggression/violence and impulsivity/substance abuse. An important limitation of our study was the relatively small sample size for a neuroimaging study. Although 23 individuals had consented to our protocol, nearly half re-offended during the relatively brief time course of the study. However, this is consistent with the nature of this population, and could be considered a reflection of the severity of psychopathic and antisocial behavior in our sample. Another important limitation is that we did not control for a past history of substance abuse in comparing psychopathic offenders to controls, and there is good evidence to suggest that effects of substances themselves can influence white matter measures. A recent study, however, shows that gray matter deficiencies in mesolimbic circuitry in psychopaths are associated with violent behaviour and psychopathy and may be relatively independent of substance abuse whereas integrity of prefrontal regions such as the OFC may be more susceptible to drug use. The drive for substance abuse itself may therefore be a component of the disorder of psychopathy, and is not easy to separate. Importantly, psychopathic offenders who completed all protocols were not using substances during the time period of the study, as they were subject to regular drug screens as a condition of their parole. One approach for future studies would be to include a third group of non-psychopathic offenders with a history of substance use as comparators. Finally, our main dependent measure of fractional anisotropy does not reveal the cellular correlates within white matter that are disrupted, as it can only indirectly index microstructural integrity of white matter. Changes in FA could be due to alteration of axonal membranes, changed axon number, disruption of the myelin sheath, or all of the above. Post-mortem investigations in this population combined with post-mortem DTI scanning could be especially valuable in elucidating these correlates, which may then be translatable to potential treatment. In summary, we found profound disruption of white matter circuitry in psychopathic offenders compared to healthy controls. Our data suggest that the neural correlates of the interpersonal affective dysfunction lies in disruption of amygdalo-prefrontal circuitry and the neural correlates of antisocial/violent behavior lie in disruption of striato-thalamo-frontal reward circuitry that is tightly linked to impulsivity and substance abuse. The extent of white matter damage attributed to neurodevelopment versus that acquired over time in psychopathy is a potentially exciting area of future investigation that deserves further exploration. We gratefully acknowledge the assistance of all persons and volunteers whose participation was essential in the successful completion of the study, in particular Neil Spiller and Eugene Hlasny. [^1]: The authors hereby declare that funding was received from AstraZeneca, Neuronetics, and Aspect Medical Inc and Pfizer. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: SH DS ZD. Performed the experiments: SH DdJ TS ZD AV. Analyzed the data: AN DF AV. Wrote the paper: SH AN AV.

# Introduction Among the archaeological deposits in the Sierra de Atapuerca (Burgos, Spain), two assemblages show evidence of cannibalism. These are the late Early Pleistocene level TD6–2 of Gran Dolina and the Bronze Age level MIR4 in Mirador Cave. Remains of hominins/humans (hereafter referred to as human remains) were recovered with obvious evidence of butchering and consumption. All of these studies concluded that human bodies were processed in the same way as animal carcasses. However, in both assemblages, a higher number of cut marks were documented on the human remains than on other animal remains. This tendency has also been observed in other cannibalized assemblages, such as Moula Guercy (France) and Brillenhöhle (southwestern Germany), in which the percentage of human bones with cut marks ranges between 40 and 60%. The cut marks, along with the anthropogenic bone breakage, can provide a remarkable amount of paleo-economic information about prehistoric human groups. Since the 1980s, a great deal of research has focused on the macroscopic and microscopic morphology of cut marks and their distribution on the skeleton and on different parts of the bones in order to infer information such as the timing of access that Pleistocene hominins had to carcasses, the handedness and, more recently, Stiner et al.suggested that through the orientation of cut marks may determine if one or several individuals were involved in the butchering. However, Egeland indicates that the orientation diversity is related to the skill and experience of the butchers. In this line of research, several proposals have been put forth as to the different variables that may influence the frequency of cut marks. Binford suggested that when the burden of extracting the tissue is lower and, therefore, more scraps of meat are attached to the bone, fewer marks should be expected. The frequency of cut marks would therefore be related, according to these observations, to the intensity of the butchering process. Milo supported a similar argument, although she added, as another variable, the difficulty in processing of carcasses. None of these assessments, however, was supported by actualistic data to substantiate these hypotheses. Egeland used an experimental approach to test the butchering process intensity and its relationship to the frequency of cut marks with quantitative data. For this purpose, he proposed studying the relationship between two simple measurements: the number of arm movements made during defleshing activities, and resulting number of individual striae. Experimental data obtained by Egeland indicates that the relationship between increased frequencies of cut marks and more intensive processing is not consistent. Previously, Lyman had suggested that the variations in cut marks frequency were related to the size of the animals butchered, with more cut marks on larger specimens. He later, however, needed to expand his explanatory models to account for the diversity of frequencies of butchered bones intra and inter-assemblages. And he finally claimed that the interpretations of cut mark frequencies may require unique contextual data in each specific assemblage. Dewbury and Russell concluded that the raw material of the stone tools used is another variable that can affect the frequency of marks. Although their results were not statistically significant, they claim that this variable should be taken into consideration. In addition, Padillashowed that, in butchery experiments, the frequencies of cut marks vary depending on the experience and skill of the butchers. Finally, Domínguez-Rodrigo and Yravedra proposed three major variables as significant factors: (1) the size of carcass and type of tool (lithic or metal); (2) bone breakage and (3) ravaging by carnivores and the consequent damages to the skeletal profiles in the original samples. According to Domínguez-Rodrigo and Yravedra, skeletal profiles ‒ regardless of their origin and cause ‒ may influence the proportion of bones bearing butchery marks. In this context, our goal is to create an experimental framework that would allow us to describe the butchery activities performed on the cannibalized human corpses recovered from level TD6–2 of Gran Dolina and level MIR-4 of the Mirador Cave. We have developed this research starting from an uniformitarian assumption. The specific objective of this experimental work is to provide actualistic observations for the interpretation of the frequency of cut marks on human remains. It is difficult to perform these experiments with human bodies. For this reason when the opportunity arose we decided that the processing of the carcass of a chimp could be helpful to our goal. We have created this actualistic framework through carrying out the butchering process (bone breakage included) on the carcass of a chimpanzee. We must point out that although there are similarities between chimpanzees and humans, we are also aware that there are considerable differences. The most important of these differences is that humans are bipedal while chimpanzees are terrestrial quadrupeds that support their forelimbs on their knuckles and not on the palms of their hands or paws like other quadrupeds. However, they can stand up and walk on their legs (bipedally) when circumstances require freeing up their hands for use. This characteristic makes the anatomy of these primates closer to hominins than that of any fully quadruped mammals, despite the clear differences in the proportions of the segments of the body. ## Archaeological context The Atapuerca complex (Burgos, Spain) contains several archaeological and palaeontological sites dating from the Early Pleistocene to the Holocene Evidence of cannibalism has been recorded in two assemblages from La Sierra de Atapuerca. ### TD6–2 assemblage The oldest evidence of the practice of cannibalism among hominins was found in level TD6–2 at the Gran Dolina site (Sierra de Atapuerca, Burgos, Spain). Several stratigraphic studies have established that level TD6 dates to the late Early Pleistocene. *H*.*antecessor* remains were recovered, along with a large number of faunal specimens and over 600 lithic artifacts. The lithic assemblage is attributed to Mode 1 technology. The hominin remains exhibit several anthropogenic modifications on their surfaces which indicate that they had been processed and consumed by other hominins After a new review of the TD6–2 assemblage, *H*. *antecessor* sample was found to contain 181 remains (155 bones and 26 teeth) from 11 individuals (MNI). Cut marks are present on all types of bones. Additionally, there is abundant evidence of anthropogenic bone breakage (percussion pits/notches or peeling). In total, 44.5% of the bones of *H*. *antecessor* exhibit some type of anthropogenic modification. Damage resulting from non-human carnivore activity has not been documented on the *H*. *antecessor* remains. ### MIR4 assemblage Mirador Cave is located on the southern slope of the Sierra de Atapuerca. In level MIR 4, a 6 m²test pit revealed an intentional accumulation of human remains, from which both bones and teeth were recovered (MIR4A). All the remains had been deposited in a small (40 x 25 x 10 cm) oval pit. Charcoal and bone dating places the chronology of level MIR4 in the Middle Bronze Age, during which time the cave was used as an animal shelter. The human remains, however, provide calibrated dates between 4400 and 4100 BP. The dates indicate that the human remains date from the Early Bronze Age, while the burial event must have occurred during the Middle or Late Bronze Age The skeletal remains of six individuals, consisting of a total of 165 specimens (148 bones and 17 isolated teeth) were recovered from MIR4A. Taphonomic analysis revealed that the individuals had undergone an intense defleshing process, and most specimens exhibited human tooth marks and evidence of boiling and intentional bone breakage. Skulls displayed a well-defined breakage pattern, with conchoidal scars around the entire perimeter of the skull and percussion marks on the temporal, parietal and occipital bones. These modifications suggest consumption as the main objective of processing the bodies. However, the remains were clearly buried intentionally in a specific way, with the skulls carefully arranged at the bottom of a pit and the remaining fragments on top of them, although the dates of some human bones and charcoal seem to indicate that the remains were buried by people unrelated to the cannibalism event several centuries afterwards. A review of the fauna from this level revealed no faunal remains inside the pit and that some human remains were dispersed across other parts of the surface and mixed with different archaeological material. Also, while the humans had only been boiled, the remains of other fauna showed evidence of being both roasted and boiled. These features, along with the dating, indicate that the remains of ungulates and human remains from the pit do not correspond to the same event and that the human remains are older than the other remains at the site. This may also alter the skeletal profile represented in the pit due to bone selection that occurred during the secondary burial. shows the percentage of human modifications on the *H*. *sapiens* and ovicaprini remains from MIR4A and MIR4. This emphasizes the high degree to which anthropogenic modifications are present, as they were found on over 60% of the hominin assemblage. Ovicaprini remains exhibit anthropogenic modifications on fewer than 30% of the assemblage. # Materials and Methods The body of the chimpanzee was made available to the IPHES team by the Fundació Mona primate recovery center (Riudellots de la Selva, Girona, Spain). The Fundació Mona is a nonprofit organization (registration 1404 in the Justice Department of the Generalitat de Catalunya) whose aim is to ensure the welfare of confiscated primates that have been detected in illegal or abuse situations. The foundation has a recovery center close to the city of Girona which is used as a home for rescued primates in order to rehabilitate and protect them. The foundation also has a research unit with an interdisciplinary team of psychologists, biologists, veterinarians and anthropologists. The chimpanzee died under veterinarian care at the Fundació Mona from natural causes at 57 years old. We will not offer images of the process described here out of respect to animal, its caregivers and the center’s generous sponsor. The chimpanzee had lost most of its dentition and had osteoarthritis in most of its bone articulations, related to its advanced age. Butchering process was recorded by photographing it and taking notes. In addition, following the advice of Nilssenthe entire process was recorded with video camera for later review, in order to expand on the notes and check that they were accurate. The images have also been used in studying the bones, and allowed us to make very reliable links between specific butchering activities and specific modifications. For skinning, defleshing and disarticulation, we used simple chert and quartzite flakes. From the Lower Pleistocene to the Holocene, chert and quartzite were the main raw materials used for tool production at the Sierra de Atapuerca sites. We chose two specific varieties of stone: Neogene chert from the Late Miocene formations around the Sierra, and quartzite from the Arenas de Utrillas facies. This chert is made up of microcrystalline grains and has moderate to good knapping properties; the quartzite is, in fact, a homogeneous and fine-grained metaquartzite. The specific features of the tools we used are presented in. Carcass butchering involved three expert butchers, who were assisted by three assistants. One expert removed the skin and butchered the right forelimb, the right hind limb, and trunk. Another butcher processed the left forelimb and the left hind limb. Finally, the third butcher processed the skull and, with an assistant, broke the bones. We have make an ANOVA test to evaluate possible differences in the frequency or the locations of the cut marks produced by the two main butchers involved in processing the chimpanzee carcass. The analytical variables used were: NISP of remains with cut marks on: Clavicles; Os Coxae; Scapulae; Humeri; Radii; Ulnae; Femurs; Tibiae; Fibulae; NISP Total; Maximum number of cut marks on one specimen; Maximum number of cut marks on one element; Total number of cut-marks; NISP of shafts with cut marks; NISP of epiphyses with cut marks. In the laboratory, bones were boiled to make it easier to remove the meat scraps, tendons and fat. The bones were then dried in the shade. A total of 314 bone specimens were collected. There were also 140 unidentifiable fragments less than 2 cm long that were recorded but not included in the analysis. The entire surface of every bone was examined, both macroscopically and microscopically (OPTHEC 120HZ), and for each anthropogenic modification (cut marks and bone breakage signals) the location, segment, portion and face were noted. We recognized three types of cut marks: slicing, scraping and chop marks. Slicing marks occur when the tool is applied with force parallel to the long axis of the tool’s edge. Scraping marks occur when the edge is applied perpendicularly to the bone surface, generating multiple unidirectional striations. Finally, chop marks are produced by applying a dynamic force (percussion) using a tool with an edge. The analysis of cut marks took into account the number of individual striae, location on the skeletal element, distribution over the surface (isolated, dispersed all over, clustered) and orientation in relation to the longitudinal axis of the bone (oblique, longitudinal, transverse). Maximum length of the cut marks was measured in millimeters. We also recorded the presence/absence and location of percussion pits, impact points, adhered flakes on the surface, conchoidal scars, hammerstone/anvil abrasions and peeling on each item studied. Peeling was described as general, classical or incipient in keeping with the suggestions of Pickering et al Finally, the experimental results were statistically compared (by correspondence analysis) with for the cut mark results for the hominin and deer (*Cervus/Dama*) remains from level TD6–2 and those for the human and *Ovis/Capra* remainsfrom level MIR4. ## Description of the butchering process The butchering process of the chimpanzee carcass consisted of making the fullest possible use of the animal by skinning, dismembering, defleshing, disarticulating it and bone breakage to extract the marrow. In this paper, the intensive butchering process is understood to mean removing the maximum amount of soft tissue suitable for consumption. We did not remove the viscera of the chimpanzee because it had undergone a necropsy, so the thoracic cavity had been opened and visceral content had been clinically removed. ### Skinning The process began with longitudinal slices from the parietals to frontal and orbital area. It then continued with oblique movements from parietal to occipital and, after that, towards the temporal bone in order to remove the right ear. The next step was to extract the lower lip, working below the chin and neck toward the left side of the mandible, removing the upper lip and finishing by extracting the left ear. Removing the skin from the trunk and limbs was easier because it required fewer slices with the tools. Longitudinal cuts were made along the inner side of the limbs, from the chest to the wrist for the arms, and from the abdomen to the ankle for the legs. The process was completed by removing the skin from the back and pelvis. During the skinning, lithic tools only came into contact with the skull and jaw bones. ### Dismembering The dismembering process consisted of several stages. First, the lower limbs were separated from the trunk, then the skull, and finally, the upper limbs were separated from the trunk. The femur and coxa were disarticulated by means of transversal cuts around the femoral head. Cranium and trunk were separated by making transversal cuts at the base of the skull to reach the articulation with the atlas, and around it. Finally, torsion was applied to assist separation. Dismembering of the upper extremities was simple, and only required cutting the flesh around the edge of the scapula. ### Defleshing The carcass was defleshed, whenever possible, along the lines of the muscular membranes, separating each of the muscles individually. However, they were sometimes too large or were difficult to differentiate, and in these cases they were cut into portions. The muscles were completely removed from the limbs, trunk and skull, and efforts were made to leave no flesh attached to the bones. The defleshing of the head began by removing flesh of the face, then of the temporal and occipital bones, and the jaw. The lower extremities were defleshed. Longitudinal cuts on the femurs made it possible to open up the muscle and reach the bone. Transverse and oblique movements were required to remove the muscle. The same process was used on the tibiae and fibulae. Defleshing of the arms began by removing meat from the scapulae with cuts on both the ventral and dorsal faces. On the humeri, longitudinal cuts were made to reach the bone shaft and continued with oblique and transverse cuts. This technique was used on the radii and ulnae, as well. The trunk was defleshed by first removing the flesh from the back and then removing the intercostal meat. Defleshing involves contact between the tool and the bones, especially in areas where there is less muscle mass, such as the skull. On the long bones, contact usually took place near the epiphysis and midshaft portions of the long bones. ### Disarticulation Disarticulation was performed after the main segments had been defleshed, except for the hands and feet. These were processed by first separating the fingers or toes and subsequently defleshing the hand or foot. Once this had been done, each metapodial bone was removed from its phalanx, and then each phalanx was separated. The limb bones were disarticulated by making transverse cuts on the bone joints, then applying torsion to the bones whenever possible. The mandible was removed from the skull during the defleshing of the head. Finally, the ribs were separated from the spine by manually bending them and were broken at an angle. ### Bone breakage Once the process of soft tissue removal was completed, we proceeded to break the skull, mandible, ribs, long bones, phalanges and metapodials. The technique used for the cranial and appendicular bones was percussion using quartzite hammers and three limestone anvils. One anvil had a flat surface and the other two had angular surfaces; when the flat anvil was used, the bones were placed on the edges. The bones were passive objects in this process, except for the skull for which the passive and active roles were combined; the active movement of the skull proved a more effective breaking technique. The metapodials and phalanges were hit on the shaft to break them and to access the marrow. The ribs were broken by simply bending them, as discussed above (section 3.2). contains a summary of the details of the breakage of each bone, and shows the regions where the bones were percussed or bent. # Results A total of 314 specimens from the chimpanzee skeleton were analyzed. Between three and 13 pieces of each long bone were recovered. 52% of the skeletal elements examined exhibited some type of modification caused, 35% with cut marks on their surfaces. These marks range in length between 1.2 mm and 50 mm. Cut marks were found alone or in groups of two to 15 striae. The latter were generally parallel or sub-parallel to one another and were oblique and/or transverse in relation to the sagittal axis of the bone. Of the cut marks, slicing marks were the most abundant. These were documented on 106 specimens. Scrape marks were recorded on eight specimens and chop marks on six. Forty-one of the skeletal elements examined exhibited signatures produced during the hammerstone-anvil bone breakage: 21 percussion pits, 20 percussion notches, and three anvil abrasions. A combination of these modifications was detected on three specimens. Peeling was documented on 37 pieces of the ribs and vertebrae. ## Cranial bones On the skull (NISP = 6, five with cut marks), the marks are related to skinning, defleshing and disarticulation. Most of the slicing marks associated with the skinning process were longitudinal (the longest of these measured 20 mm in length) and were located on the parietal, orbital, temporal and frontal parts of the skull. The sternocleidomastoid muscle was cut in order to remove the ears during skinning. There were some oblique, transversal and longitudinal cuts on the bone which were located on the parietal, frontal and zygomatic bones. These cuts are associated with the removal of the epicranius, temporalis, masseter and orbicularis oculi muscles and caused by the disarticulation and defleshing of the cranium. The mandible (NISP = 3, all with cut marks), exhibited transverse or oblique cut marks. The cuts were located on the vestibular side of the mandibular notch and on the mandibular body (M₁ zone) and are related to cutting the masseter and depressor anguli oris muscles. These marks were therefore identified as caused by the defleshing and disarticulation processes. The calvarium was broken in half lengthwise. Blows to the orbit resulted in the separation of the skull from the face and a right orbital fracture. Percussion pits and hammerstone/anvil abrasions were visible on the temporal and parietal bones. Parietal bone shows a cortical scar (23x14 mm) and an adhered flake. Left hemimandible had a transverse fracture on the chin area. ## Trunk On the vertebrae (NISP = 25, 16 cut with marks), cut marks were documented on cervical and thoracic vertebrae. The atlas exhibited transversal slicing marks on the lamina of the arch and on the transverse process and right pedicle, which were caused by cutting the levator scapulae tensor. These cut marks and the broken left pedicle are due to the disarticulation in which the chimpanzee’s head was separated from the trunk. The 5th, 6th and 7th cervical vertebrae exhibited cut marks located on the neural apophyses. These cut marks measured between 1.8 and 3.1 mm and were caused by cutting the multifidus and rotatory muscles during defleshing. All of the thoracic vertebrae, except for the 2nd, had slicing marks in clusters on their neural apophyses. The 2nd thoracic vertebra exhibited cut marks on the transverse process, related to cutting the intertransversarii muscle. All cut marks identified on the vertebrae were caused during the defleshing of the back. Peeling was visible on three lumbar vertebrae. One of them showed incipient peeling on the neural apophysis, the other one had general peeling on the articular, transverse and neural apophysis and the third one had general peeling on the neural apophysis. On the clavicles (NISP = 2, both with cut marks) the cut marks were clustered, and transversal and oblique. They were located on the coronoid tubercle, acromial and sternal extremities. Slicing marks were caused during disarticulation, by cutting the pectoralis major, deltoids and sternocleidomastoid muscles. Cut marks on the ribs (NISP = 48, 14 with cut marks) were mostly clustered and transversal and oblique. Scrape marks were visible on one rib. Almost all of the cuts (12) were located on the rib angle area and the diaphyses, but one was located on the neck of the rib. Cut marks documented on these ribs were caused by defleshing and disarticulation. These marks were made during the removal of the external intercostal, serratus anterior and longisimus dorsi muscles. Peeling was found on 32 rib specimens, 14 of them showed general peeling, 12 classical peeling and six incipient peeling. Most of the fractures were at the rib angles. Cuts marks on the coxae (NISP = 2, both with cut marks) were clustered, transversal and oblique. They were located on the lateral and medial sides and on the ischium portion. These marks appeared at the insertions of the quadratus femori and gluteus maximum muscles, indicating that the slicing marks were generated during defleshing. ## Arms and legs On the scapulae (NISP = 2, both with cut marks), cut marks were mostly clustered, and ran in several directions (transversal, oblique and longitudinal). These specimens displayed the longest (50 mm) slicing mark of the sample. The cuts were visible on the dorsal face, on the area where the supraespinatus and infraspinatus muscles attach, and on the scapular spine, where the trapezius, teres major and deltoides muscle attach. The scapular spine had slicing, chop and scrape marks. There were also cut marks on the ventral face and neck of the scapulae. These cuts were caused by cutting the teres minor and subescapularis muscles. All marks identified on scapulae are associated with defleshing and disarticulation. Cut marks on the humeri (NISP = 18, 12 with cut marks) were clustered, transversal and oblique. The slicing marks were along the bone and on the anterior, lateral, medial and posterior sides. The slicing and chop marks were caused by extracting the brachial, the medial and lateral head triceps muscles and the extensor carpo radialis longus. Cut marks were detected on the proximal epiphysis (tuberculus major) caused by cutting the supraspinatus and infraspinatus tendons. Most of cuts were caused during defleshing and the disarticulation of the humeri, the scapulae or the radii ulna. The radii (n = 19, six with cut marks) exhibited clustered, isolated and located cut marks on the ainterosseous cresta, radial tuberosity and posterior side. Cuts were generated during the removal of the flexor digitorum superficialis and biceps brachii muscles and supinator and pronator tendons, and the slicing marks were caused during defleshing. Cut marks on the ulnae (NISP = 22, eight with cut marks) were clustered and oblique and located on the shaft and all sides of the bones. Most of the cuts were slicing marks, although there were scrape marks on the medial side of the right and left ulnae. Cuts located were made during the removal of flexor digitorum profundus, extensor pollicis longus and flexor carpi ulnaris muscles. The femurs (NISP = 18, ten with cut marks) bore clustered, transversal and oblique cut marks located along the length of the bones and on all sides of the bones. On proximal epiphyses (trochanter minor and major) there are slicing and chop marks caused by cutting the illiopsoas, gluteus medius, pectioneus tendons and vastus medialis muscles. On the distal epiphyses there were slicing marks related to cutting the gastrocnemius muscle. All these marks on epiphyses were associated with disarticulation. Some of the longer cut marks (+20 mm) were located on shafts and were longitudinal. In addition to these slicing marks, clustered and transverse chop-marks (18 mm) could be observed on the shaft. These slicing marks were caused by cutting the vastus intermedius tendon during defleshing. The chop marks were caused by fracturing the femur. Cut marks on the tibiae (NISP = 23, seven cut marked) were clustered, transversal and oblique. On the proximal epiphysis and shaft, the cut marks are located where the popliteus tendon, extensor tibial anterior, soleus and plantaris muscles attach. Slicing marks were visible on the distal epiphyses (the fibular notch) and distal shaft; these were caused by cutting the extensor hallucis longus muscle and the Achilles tendon, and are associated with defleshing and disarticulation. The fibulae (n = 11, four cut marked) shows cut marks that were clustered, transversal and oblique. On the distal epiphyses and shaft there were slicing marks caused by cutting the peroneus brevismuscle. There were cuts on the proximal epiphyses (crests) caused by cutting the soleus and biceps femoris and peroneus longus muscles. The locations of the cut marks indicate that they were made during disarticulation and defleshing. All of the long bones were broken, resulting in 107 long bone specimens, of which 41(38.3%) exhibit some type of anthropogenic breakage mark, such as percussion pits, percussion notches, chop marks, adhered flakes and conchoidal scars together with hammerstone/anvil abrasions from the intensive bone breakage process. With one exception, all of the percussion marks were located on the mid shaft portions of the long bones. The exception was a humerus on which a percussion impact was located close to the proximal end. ## Hands and feet Cut marks were located on six metapodials, two metacarpals and four metatarsals. On the metacarpals (NISP = 5, two cut marked), clustered and isolated, transversal and oblique cut marks were found on the 2nd and 5th metacarpals. On the 2nd metacarpal II, slicing marks were located on the proximal epiphysis and the lateral side. On the 5th metacarpal they were located on the shaft, the distal epiphysis and the palmar side. For both, the cut marks were caused by cutting the flexor digitorum brevis. On the metatarsals (NISP = 20, four cut marked) the cuts were clustered and oblique. They were located along the bone, from the proximal to the distal epiphyses and on the medial, lateral and plantar sides. These marks were caused by cutting the flexor digitorum longus, dorsal interosseus and plantar interosseus muscles. On the plantar side of the 5th metatarsal, a scrape mark was visible that was caused bycutting the plantar interosseus muscle. On the carpals (n = 19, one cut marked), only the right pisiform bone exhibited transversal slicing marks along the length of the bone and on the anterior and posterior sides. These cuts affected the flexor carpi ulnaris, and were therefore caused during disarticulation. Among the tarsals (NISP = 9, one cut marked), only the chimpanzee’s left calcaneus shows cut marks. The cuts were clustered and transversal. Slicing marks were located on the calcaneal tuberosity and related to cutting the Achilles tendon. There were nine cut marks on the phalanges (NISP = 72, nine cut marked) , five of which were from the right hand and four from the left foot. On the hand, slicing marks were visible on the 1st, 2nd and 3rd phalanges, located on the medial side and clustered together. These cuts were caused by cutting the flexor digitorum superficialis. On the foot, the cuts were on the 1st and 2nd phalanges. These marks were oblique and are related to cutting the flexor digitorum longus and plantar interosseus muscles. Cuts documented on distal portions of the phalanges were caused during disarticulation and defleshing. Two phalanges and a 4th metatarsal had percussion pits and hammerstone/anvil abrasions, two of which were impact flakes. These marks were on the lateral and anterior side shafts. One metatarsal exhibited an adhered flake. # Discussion and Conclusions The butchering process is a set of exclusively anthropogenic activities for the purpose of obtaining nutrients and other products derived from animal carcasses. The butchering process begins with accessing the carcass of an animal, and ends with discarding the unused or exhausted remains. The primary goal of the butchering process is to prepare carcasses for transport and consumption, although other ends are also possible, including obtaining secondary materials (e.g. skin or tendons) or extracting and preparing tissues for storage and delayed consumption. The butchering process can be performed on any type of animal, including the corpses of hominins. Before and after the butchering process can have specific aspects associated to the mode of death and how the remains are finally disposed, aspects that can related with ritual treatments of the remains. While it is true that rituals may be present for both animal carcasses and human bodies, it is evident that in this work we are especially interested in the latter. However, whatever the causes and circumstances of the death of the individuals (natural, accidental or violent) and whether or not they were disposed of differently from other animals, the butchering process, whether the end goal is the consumption of humans or other taxa, has the same objectives and would therefore generally be undertaken in the same way. The process includes skinning, defleshing, dismembering, viscera removal, and bone breakage. One of the characteristics that allows us to identify cannibalism is the existence of parallels between the butchering process applied to human bodies and the carcasses of the other animals in an assemblage. In general, most descriptions of European prehistoric cannibalized assemblages describe a butchering process that is similar for humans and animals, as a result of a consistent pattern in the exploitation of meat, bone and brains of the bodies. The only possibly ritual processing described in any of these cases was that applied to the skulls in some of the assemblages that contained anatomically modern human remains. However, many of these assemblages share a common feature in that a high frequency of human remains displays anthropogenic modifications. Between 40 and 60% of the human remains found at El Mirador, Fontbrégoua (France), Gough’s Cave (Great Britain), Moula Guercy (France) and TD6–2 show anthropogenic modifications. Even for the ungulate remains in these assemblages, anthropogenic modifications have been documented on 20% of the specimens. TD6–2 and MIR 4 assemblages both clearly exhibit this characteristic although there are significant differences between them: the assemblages are from different economic systems (Early Pleistocene hunter-gatherers on the one hand, and a productive Bronze Age society on the other). Additionally, the remains were not disposed of in the same way, the remains at Mirador were in a pit, separate from other taxa and TD6–2 remains were found mixed with the bones of other animals. Lastly, the processing was different, since some were not cooked and others were. Not to mention the almost one million years between the times when the bodies were processed, eaten and disposed of. These two assemblages do, however, share the high frequency of anthropogenic modifications to the remains recovered, especially cut marks, but also bone breakage. Of the human remains from TD6–2, 44.5% exhibit some anthropogenic modification, whereas 29.7% of the medium size Cervidae remains show signs of a butchering process. In Mirador Cave, 60.1% of the human remains show signals of butchering, while only 19.7% were found on the bones of ovicaprini. How these differences in the frequencies of cut marks are interpreted could have significant consequences when attempting to interpret why cannibalism took place in these two assemblages. In our experiment, the chimpanzee was skinned, defleshed and dismembered, and finally its bones were fractured to simulate a model of making intensive use of all the tissues. The viscera had already been removed, so this part of the process could not be taken into account. The result was that 42.7% of the recovered remains exhibited some kind of human modification. This is close to that observed on the remains found at archaeological sites where cannibalism occurred, and especially close to for the remains from level TD6–2. In the experimental sample we considered fragments that in an archaeological assemblage could hardly be identified anatomically beyond the more general classifications (long and flat bones) and could not have been assigned to a specific taxonomic group with reasonable certainty. For this reason, we separated out those remains which, in our experience, could not have been identified in an archeological assemblage because of the lack of landmarks (nutrient foramens, muscle attachments, articular facets, twisted parts of bones, etc.). Removing these pieces from the sample changes the percentage of anthropogenically modified bones only slightly (to 44.2%). Therefore, our experimental results do not seem to be affected by the degree to which the remains can be identified. Any comparison must also take into consideration the fact that the taphonomic history of archaeological remains is more complex, and all of the possible effects of post-depositional modifications cannot be controlled. These include increased fracturing of the remains by diagenesis during the fossilization process. Neither the archaeological assemblage of TD6–2 or the more modern case of Mirador Cave appear to have been significantly affected by taphonomic processes after human activity. In TD6–2, dry fractures affected only 2% of the specimens. The most common post-depositional modification is manganese oxide staining, but this does not affect the bone structure, and so does not affect the degree of anatomical and taxonomic identification or the morphology of the anthropogenic signatures (cut marks or fractures). No signs of carnivore consumption on the remains of *H*. *antecessor* have been documented. In addition, the conservation of some human elements such as vertebrae and ribs, which are likely to disappear in assemblages affected by carnivores, is higher than in the case of other animals, mainly the ungulates. This leads us to conclude that the hominin bones were little disturbed after the anthropogenic activity and abandonment. The post-cannibalistic taphonomic history of the human remains from Mirador Cave has been affected by the activity of the individuals who deposited the bones in the pit, but did not participate in the consumption. However, in this case, the absence of modifications by other agents and/or processes is even more remarkable than in the case of TD6–2. Dry fracturing was only identified on one specimen (0.1%), and post-depositional modifications (trampling, manganese stains and root-etching) affect only 1.3% of the assemblage. As in the previous case, the human remains had not been modified by carnivores. Given all these characteristics, it is appropriate to compare the experimental frequency of the anthropogenic modifications to the chimpanzee’s bones with those documented in these two archaeological assemblages. Much research in the past ten years has aimed to establish the origin of the different frequencies of cut marks found in archaeological assemblages, although without much success in most cases. The effects of the raw materials used to make the tools and the number of moves made during the butchering process has been tested, but no clear relationship was established between the frequency of striae and either of these variables. A relationship has been established between the abundance of cut marks and the experience and skill of the butchers. No differences were, however, detected in either the frequency or the locations of the cut marks produced by the two main butchers involved in processing the chimpanzee carcass (ANOVA, F = 0.0614, p =. 806), although it the sample is, admittedly, small. In the case of archaeological sites, however, it is difficult to assess the number of butchers or their ability. This difficulty is partly due to the fact that the analysis of archaeological assemblages often reveals palimpsests that point to different events occurring during the same or different occupations and involving the same or different individuals, thereby rendering it impossible to objectively evaluate the number of butchers or their experience. An analysis of the distribution of the modifications to the chimpanzee’s bones revealed that the most of cut marks were located on the muscle attachments and, primarily, on the skeletal segments and elements with the largest muscles or sets of muscles. The elements with the highest number of marks were the scapulae, humeri, vertebrae and ribs; on the latter there were cut marks associated with removing the flesh from the animal’s back. Regarding the high number of cut marks on the ribs, we must take into account that no marks were caused by evisceration—a process including evisceration would yield an even higher number of marks on the ribs. The remains of *H*. *antecessor* from TD6–2 and *H*. *sapiens* from MIR4 showed cut marks in similar locations which were mostly caused by removing meat (Figs. to). In both cases, the highest number of cut marks on the long bones were located on the shaft; in the case of the ribs, the highest number were located on the proximal part of the shaft; and on the vertebrae the highest number were located on the laminae. In the latter two cases, the cut marks were also due to removal of the meat from these individuals’ backs. Some of the phalanges and metapodials from TD6–2 also exhibit cut marks. This has led to the hypothesis that the tendons in the hands and feet were possibly removed for a specific use, although Saladié et al. could not relate these marks to a specific activity. However, the processing of one hand and one foot from the chimpanzee suggests that these cut marks are due to an intensive use of the meat from these extremities. Skinning has been documented in all three cases on the bones of the skull. However, in this respect there are considerable differences in the calvariae recovered MIR4. The six documented skulls show a pattern of very definite modifications. These modifications were caused by the scalping and fracturing that resulted in what are known as skull cups. This skull breakage pattern has been found in other assemblages associated with ritual cannibalism, as in the case of Gough’s Cave, or war cannibalismThe remains from TD6–2 do not have these features, and the chimpanzee skull was not processed this way either, even though it was scalped and then fractured in order to extract the brain. This suggests that the calvariae found in Mirador Cave and in other assemblages in which this pattern is present must have been obtained intentionally. One explanation for the high frequency of cut marks on *H*. *antecessor* and *H*. *sapiens* may be that the bodies were treated differently from how animals were treated. However, the remains of the animals identified in both assemblages were processed in the same way in TD6–2 and MIR-4, except that, at Mirador Cave the human remains were cooked and burned. These similarities extend to our results from processing the chimpanzee. The only notable differences between the human and ungulate samples are related to the skinning of the carcasses. On the ungulates remains, these signs are more common on phalanges or metapodials than on skulls, suggesting that we cannot assume differential treatment. Furthermore, both sets of ungulates underwent the full butchering process, including the removal of the viscera, so these animals’ tissues were also intensively used in both assemblages. The question is, therefore, what causes the differences in the frequencies of cut marks? To try to resolve this issue, a correspondence analysis was performed to determine whether there was a relationship between the frequency of the cut marks and their anatomical distribution on the chimpanzee specimens, *H*. *antecessor* and the cervids from TD6–2, and *H*. *sapiens* and the ovicaprini from MIR 4. The most discriminating variable is found on Axis 1 (eigenvalue = 59.19), which separates the human and chimpanzee specimens (to the left of the axis) from the ungulates, which are at the right margin of the graph. This separation seems to be related to the higher number of modified axial specimens in the hominin remains. Of the two types of hominin remains studied, a stronger relationship was found between the cut marks on the chimpanzee remains and those on the *H*. *antecessor* remains. The *H*. *sapiens* results are clearly influenced by the number of cranial remains with anthropogenic modifications (Axis 2; eigenvalue = 24.22). Clustering in the cases of the cervids and ovicaprini is marked by the presence of cut marks on limb bones and their reduced presence on the vertebrae and ribs. This result seems to be related to the frequency of cut marks and the anatomical distribution in the four taxa. Obviously, we had all of the chimpanzee skeletal elements. One of the features that characterize the set of *H*. *antecessor* remains is the presence of ribs, vertebrae and phalanges, defined by Marean and Clerghorn (2003) as low survival elements as they are not typically well preserved in archaeological sites. The human remains from Mirador Cave mainly consist of calvariae and intensely processed skull fragments. These may have been used for purposes other than providing food, or for purposes in addition to providing food. Processed long bones are present are present in both assemblages. Ungulates from both assemblages are well represented by fragments of shafts of long bones; ribs and vertebrae are also present, but in low numbers, and skull fragments are poorly represented. These anatomical profiles are also visible in other assemblages in which the frequency of cut marks on the hominin remains is higher than for ungulates. In the sample from Moula Guercy (France), over 50% of the Neanderthal remains exhibit cut marks, while cut marks are found on only 22% of the deer specimens. In this assemblage 33.8% of the hominin remains are skull fragments, of which 65.2% (15 of 23) display mark If we also take into account the signals of anthropogenic fracture, 100% of the skulls were modified. In contrast, there is only one modified deer skull fragment in the assemblage. Another similar case is the Magdalenian assemblage from Brillenhöhle (southwestern Germany) in which 64% of the human remains exhibit cut marks. Orschiedt interpreted this as the result of mortuary practices. This theory is supported by the high frequency of cut marks and their distribution on components of the axial skeleton and on the phalanges. In addition, this assemblage shows obvious parallels with the Magdalenian human remains from Gough’s Cave (Great Britain), where there is clear evidence of cannibalism. This anatomical distribution of the cut marks is analogous to that in the TD6–2 assemblage, which shows a high frequency of cut marks on ribs and vertebrae, and these marks have also been documented on phalanges and metapodials. There is no doubt that these modifications are the result of processing and consumption by other hominins and not to subsequent burials. The case of Mirador Cave is slightly different because of the secondary burial of the human bones, although the modifications are clearly related to processing for consumption, regardless of whether or not there was a ritual treatment. However, the selection of these remains for burial is what made the intensively processed skulls stand out. The frequency of cut marks documented on the chimpanzee bones may be affected by the fact that we have the entire skeleton, including low survival bones, which typically show many cut marks. The remains from TD6–2 and MIR4 show evidence of specific taphonomic histories that provide insight into the episodes of cannibalism that occurred at the sites, and explain the differences between the skeletal profiles recovered in the two assemblages. At Mirador Cave, most of the human remains were found in a pit in which the remains of other animals were completely absent. The remains were clearly intentionally buried, with the skulls carefully placed at the bottom of the pit and the remaining fragments laid on top of them. In this assemblage, there is a marked presence of heavily modified skull specimens. In TD6–2, the main difference between the *H*. *antecessor* and ungulate samples lies in the fact that while the latter were clearly modified by carnivores, no hominin remains affected by scavenging carnivores have been documented (or the evidence of this is so slight as to be invisible). This has been interpreted as the result of longer occupations during episodes in which there was cannibalism. The fact that the remains of *H*. *antecessor* were not affected by secondary consumption by carnivores led to the conservation of low survival elements, mainly ribs and vertebrae, the latter of which exhibit numerous cut marks and abundant bone breakage. These data are consistent with the observations made by Domínguez-Rodrigo and Yravedra, who stated that the extent of ravaging by carnivores and bone fragmentation are the most influential variables in the total frequency of cut marks. This study compares the products of three different processes from different contexts (two archaeological and one experimental) and three different skeletal samples (a complete skeleton, one product of different events that did not include ravaging and another that has been biased by the human selection of some bones). However, in all three cases the frequency of anthropogenic modifications is high. This diversity indicates that it is not only the skeletal profile that determines the frequency of anthropogenic damage, although it definitely plays an important role. The fact that these assemblages have not been ravaged by carnivores is also an important consideration. Furthermore, our results may be affected by the differences between chimpanzee and hominin anatomy. It is clear that the higher or lower frequency of anthropogenic modifications must be judged in each particular context, as suggested by Lyman. There may be different causes that influence the frequency of cut marks: ravaging, type of tools, the size of the carcasses and the skeletal profiles that remain as a result of a variety of causes, but also a possible ritual component. The three cases analyzed here have different taphonomic histories but share the characteristic that they were scarcely disturbed by other agents or processes. Our results indicate that frequencies of anthropogenic modifications after intensive butchering of hominin corpses for food of 30% or higher are not uncommon in different contexts. The experimental butchering process that simulated the cuts made exclusively for the purpose providing food revealed numbers of modifications as high as in archaeological contexts (both where there was ritual behavior and where this was not the case). It can therefore be expected that, in cannibalized and little disturbed assemblages, the frequency of anthropogenic modifications will be high, and will usually affect over 30% of the specimens in the assemblage. We want to express our sincerest gratitude to La Fundació Mona, especially Olga Feliu, David Riba and Miquel Llorente. We also want to give special recognition to all the people—especially the volunteers—who work hard in the center to ensure the welfare of the primates housed there. We also want to express the greatest gratitude to our colleague Marina Mosquera, since she made it possible for the experiment to be carried out. We are deeply grateful to the Atapuerca research team and the participants in the fieldwork for that project. We thank the editor Michael D. Petraglia, and two reviewers for suggestions that improved the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: PS ARH RH IC AO. Performed the experiments: PS ARH RH IC AO BS MJG PM JM. Analyzed the data: PS ARH RH IC BS. Contributed reagents/materials/analysis tools: PS ARH RH IC BS. Wrote the paper: PS ARH RH IC. [^3]: Current address: IPHES, Carrer Marcel·lí Domingo s/n, Edifici W3, Campus Sescelades, 43007, Tarragona, Spain

# Introduction Microfluidics, and other engineered environments, can produce highly controlled micrometer-scale environments for the study of organismal model systems (e.g., mammalian cells). However, scientists or engineers interested in manipulating the environment of cm-scale organisms (e.g., plants) have remarkably few convenient tools at their disposal. This paucity is partly due to the demanding design requirements associated with larger scales (e.g., cost). This liability is particularly evident in the study of plants and their root systems. The development of plants in soil is an important subject of investigation. The provision of food to the global human population is under severe pressure (our supply of food is predicted to be far below demand by 2050) and depends on plant roots (97.6% of global calorie consumption is derived from plants). Roots influence a plant's yield and whether a plant will survive stresses. We know that root growth is strongly affected by its environment, soil, but our mechanistic understanding of these effects is imperfect, and strongly limited by technical challenges. Root development is a difficult process to study experimentally. (i) Plants display highly variable root systems, even when genetically identical. (ii) Roots are remarkably sensitive to a variety of stimuli (e.g., gravity, light, touch, moisture, nutrients, oxygen, temperature, trauma, electric fields). (iii) Any volume of soil is unique and impossible to replicate exactly. (iv) Its heterogeneity makes it opaque to most forms of radiation. (v) Its structural and chemical characteristics (i.e., porosity, surface chemistry, nutrient gradients, oxygen gradients, bulk composition, soil biota) cannot be independently manipulated. One approach to avoid this complexity is to characterize the growth of plants in soil-less media, e.g., hydrogels, paper, glass beads, sand. These systems are less inhomogeneous and irreproducible than soil and can be modified – usually to a limited extent – to mimic soil properties such as chemical composition, physical structure, water availability , refractive index, or mechanical strength. However, the lack of modularity, versatility, structural precision, and the very limited control over structural and chemical heterogeneities in these systems severely limits the type, complexity, and reproducibility of the experiments they can perform. Microfluidic approaches offer fascinating capabilities for the study of plant roots, but are subjected to limitations in their throughput and in the size of the plants they can host. We here demonstrate that LEGO bricks are highly convenient and versatile building blocks for building cm-scale engineered environments for plant roots. Their modularity enables the fabrication of environments with highly controlled structural and chemical heterogeneities that are suitable for convenient quantitative studies of environmental effects on plant phenotypes. # System Design A convenient experimental platform for the study of root development in controlled environments must satisfy a demanding set of design constraints. LEGO bricks, while conceived and sold as toys, satisfy these constraints. ## Modularity Modular systems can produce many structurally distinct environments from a few different components. Features can be added or removed without remanufacturing the entire experimental setup. LEGO structures are modular. The smallest bricks are 8*x*8*x*6 mm. The largest are 48*x*8*x*50 mm. The number of different structures that can be made with these units is staggering: six identical bricks can form almost a billion different structures. ## Scalability Confinement can affect the physiology of an organism. The ability to create experimental platforms of a range of sizes enables researchers to study any plant and their ensembles. LEGO structures can be easily scaled to accommodate different plant species: the smallest enclosed environment that can be produced with LEGO bricks measures 0.35 cm³ in volume, and it is theoretically possible to create LEGO structures capable of containing the largest plant species. ## Structurally precise Roots are sensitive to the physical structure of their environment. For example, the study of root thigmotropism (the response of a root to touch) requires structures that are of an exact size and shape. The molds used to produce LEGO bricks are accurate to within 5 µm, which is comparable to the diameter of a root hair and to the resolution of 3D printing (minimum layer thickness is ∼50 µm in some of the best current models). ## Capable of increasing levels of complexity A good model system allows for the controlled introduction of experimental variables. LEGO bricks can be used –as shown below – for the generation of physical barriers, air pockets, chemical gradients, and interconnecting chambers to control the growth environment of a plant. ## Simplicity Simple setups reduce the risk of operator-induced systematic errors. Differently from microfluidic approaches, the assembly of structures from LEGO bricks does not require technical training so undergraduate students can perform LEGO brick- based plant experiments from their first day in the laboratory. Simple experiments that demonstrate fundamental principles of plant growth (e.g., tropisms) or encourage experimental creativity can be conducted by school children of all ages during science education classes. ## Reproducibility Plant root experimental platforms (e.g. sand columns, rhizotrons, split-root pots) are typically made from scratch. Their reproducibility between labs or across continents cannot be guaranteed. The unique selling point of LEGO bricks is that bricks bought in separate batches are essentially identical and backward- and forward-compatible with each other. Experiments created from LEGO bricks can be accurately replicated anywhere in the world. ## Affordability The more expensive each experiment is, the fewer experiments can be conducted with finite resources. This fact is especially meaningful in developing nations and in research fields, like plant science, where throughput is an essential parameter. Individual LEGO bricks cost between \$0.10 and \$1.00 and are sold worldwide. A LEGO structure capable of growing a plant costs \$3.1 and is reusable: some LEGO bricks in our lab have been in near-constant use for two years. ## High throughput The ability to run a large number of experiments at the same time is essential for the establishment, for example, of genotype-environment-phenotype relationships. A LEGO structure like the one shown in can be assembled in less than a minute. ## Transparency Twenty eight different LEGO bricks are made from transparent polycarbonate which can be assembled into transparent structures for the real-time monitoring of plant roots over time. ## Autoclavable Tissue cultures require sterile conditions. Transparent LEGO bricks (with the exception of large base plates) are autoclavable due to their polycarbonate composition: they still fit together in the same way as they do prior to autoclaving and are still transparent after more than 50 autoclave cycles. Opaque LEGO bricks are made from acrylonitrile-butadiene-styrene block copolymer (ABS), and can be sterilized with ethanol or bleach. ## Three-dimensionality While 2D platforms offer significant advantages in terms of visualization and practicality, 3D mediums are more representative of the natural environment of roots. LEGO bricks allow for the creation of nearly arbitrary 3D structures. ## Chemical inertness Legislative standards ensure the safety to children of LEGO bricks sold in the USA and EU. These standards include maximum soluble levels of toxic or hazardous substances. ## Compatibility with existing growth environments Tools that integrate with existing experimental platforms are often the most useful. The modularity of LEGO structures enables them to integrate with laboratory protocols e.g., LEGO structures can hold gel, beads, sand, soil, 3D-printed elements, or be structurally precise elements in other setups. # Results and Discussion shows a flow diagram of the design, assembly, disassembly, and re-assembly of an experiment based on LEGO bricks. The website of the LEGO Group ([www.lego.com](http://www.lego.com)) provides a free software (LEGO Digital Designer, LDD) for the CAD-like design of structures using any available LEGO brick. The software outputs a step-by-step assembly guide and a list of the required parts. Individual bricks can be purchased through the “Pick a Brick” section of [www.lego.com](http://www.lego.com) or other outlets (e.g. local LEGO stores, EBay). Sterilization of LEGO bricks can be performed before or after assembly. The preservation of sterility requires the structure to be maintained in a sterile container during the course of an experiment. The simplest example of a plant germination and growth environment based on LEGO bricks is shown in. The LEGO bricks are assembled into a container that contains a root growth medium on which a seed is germinated and grown:, for example, shows a *Brassica rapa*, Wisconsin Fast Plant, Astroplant, *dwf1*, growing on a transparent hydrogel, Gellan gum. While gel media for root growth are very commonly used in experiments, they are not the best mimic of soil: root architectures grown in an homogeneous media will not match those of plants grown in real soil. However, gel media allows us to demonstrate three essential capabilities of LEGO-based biological environments: their ability to hold liquids, their compatibility with real-time observation and root structure analysis, and their use in generating reconfigurable environments that include controlled heterogeneities. Furthermore, LEGO environments are not limited to gel media: the environment shown in can hold other media of choice, e.g., sand, perlite, soil. Since structures built from LEGO bricks are not waterproof, their use to hold gels requires some stratagems (see Supporting Information for details and for a demonstration). The LEGO structure must be chilled in a freezer before the cool gel solution is poured in it just prior to setting. Using this approach, leakage of the gel solution was minimal. These basic environments can be easily scaled to match the dimensions of the organism under consideration and the time the organism is allowed to grow. show the use of LEGO bricks to create containers with very different dimensions (5×5×5 cm, 10×10×5 cm, and 20×20×10 cm) for the growth of Fast Plants, *Triticum polonicum* (Wheat), and *Zea mays* (Corn). The transparency and flat walls of LEGO bricks allows for good quality real time imaging of the development of the root system. shows time-lapse imaging of *Lepidium sativum* (Garden cress) roots over the course of ∼48 hrs from germination in a LEGO-based environment. The plant was chosen for its relatively fine roots (∼350 µm thickness) that would have been hard to image in a poorly transparent system. The reversible nature of the mechanical bond between the bricks provides two important capabilities: the creation of reconfigurable biological environments, and of highly controlled heterogeneities (i.e., solid obstacles, air pockets, and chemical and soil biota gradients) in an otherwise homogeneous growth medium. demonstrates a reconfigurable plant growth environment. Two Fast Plants were grown in gel in separate containers assembled on the same base plate. The LEGO brick walls separating the two containers were removed and reconfigured to make one larger container. The volume separating the two plants was then filled with more gel, fluidically connecting the two plants. demonstrates the generation of controlled heterogeneities in a homogeneous gel medium for plant growth by a simple templating strategy borrowed from the materials science “toolbox”. A gelling mixture was poured into a LEGO-based mold. LEGO-based features in the mold can be used as solid heterogeneities to study the physical interaction of plant roots with solid objects (thigmotropism). After gelation, LEGO-based molds could be removed, leaving behind precisely positioned air pockets that would serve as sources of oxygen gradients into the gel. These pockets could be then refilled with a hydrogel containing a desired chemical to generate precisely positioned one-dimensional (bottom left panel) or two- dimensional (bottom right panel) nutrient gradients. The above process can be combined to create environments with solid heterogeneities, air pockets (i.e., oxygen gradients), and chemical (e.g., nutrients, toxins, signaling molecules) gradients simultaneously. # Conclusions In summary we demonstrated that LEGO-based environments can (i) scale to the size of the organism under consideration, (ii) allow for real time monitoring of root systems in 3D, (iii) be structurally reconfigured to change the environment of an organism during its development, and (iv) generate precisely controlled heterogeneities (i.e., solid barriers, air pockets, chemical and soil biota gradients) in an otherwise homogeneous growing medium. This manuscript also proposes a broader concept: the use of reusable and mechanically interlocking building blocks for the construction of biological environments for cm-scale organisms and systems of organisms. Modular and reusable building blocks can alleviate the challenges associated with the large scales of plant science experiments, while providing new capabilities (e.g., controlled heterogeneities, reconfigurable environments) for the study of environmental effects on biosystem development. Furthermore, this concept provides materials chemists and engineers with two stimulating opportunities: (i) to creatively engage with the synthesis or development of increasingly capable cm-scale biological environments for important organisms such as plants, and (ii) to use these environments to test hypothesis concerning plants that are compatible with their skillset. Compelling opportunities lie in extending our approach to chemically synthesized bricks, LEGO-compatible 3D-printed bricks and objects, and commercial bricks from other manufacturers. Our laboratory will be introducing a set of integrated tools for the fabrication of frugal but sophisticated cm-scale environments for the study of plants and other organisms. # Supporting Information We thank Dr. Kuloth V. Shajesh for valuable discussions and William Rekemeyer for help in the laboratory. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LC. Performed the experiments: KRL TS SB AM. Analyzed the data: TS LC. Contributed reagents/materials/analysis tools: LC. Contributed to the writing of the manuscript: LC TS SB KRL. [^3]: Current address: Department of Sustainable Soils and Grassland Systems, Rothamsted Research, Harpenden, Heartfordshire, United Kingdom

# Introduction Modulation of brain function via ultrasound and other applications of therapeutic ultrasound originated with the Fry brothers. Through excitation and inhibition of neuronal tissue, they were able to induce transient physiological effects without observable damage. Tyler and colleagues revitalized this concept, first showing neuron activation in a mouse brain-slice model. Next, through the use of transcranial pulsed ultrasound with a relatively large focus directed at the brains of mice, they induced observable, generally biliateral peripheral motor activity such as tail and paw flicks and whisker movements, demonstrating that ultrasonic neuromodulation (UNMOD) could stimulate entire brain circuits. These UNMOD studies – employed ultrasound emitted by readily available planar transducers as the source of stimulation. Even with acoustic waveguides, the resulting acoustic fields have lacked optimal anatomical specificity. Two recent *in-vivo* studies have used focused ultrasound (FUS), balancing relatively lower frequencies with relatively larger volumes of brain stimulation. In the first study, Yoo et al. used a pulsed, 690-kHz focused ultrasound protocol on anesthetized rabbits, showing via functional magnetic resonance imaging (fMRI), electromyography (EMG), and gross observation that one side of the brain can be stimulated, inducing motor function on the contralateral side. The second study used a 350-kHz focused ultrasound protocol to stimulate at least a cranial nerve associated with control of an eye of an anesthetized rat, with motion induced ipsilateral to the stimulation zone. This study seeks to demonstrate a new method to deploy low-frequency FUS-based UNMOD, one with significantly increased anatomical specificity yet with the potential to deploy very low ultrasound frequencies through use of vibro- acoustography techniques. Typical research-grade embodiments of vibro- acoustography use two confocal sources of high-frequency ultrasound each run at slightly different and typical high carrier frequencies to produce, at their shared focus, a source of low-frequency ultrasound at the difference frequency of the carrier waves. Here, modulated focused ultrasound (mFU) allows application of low-frequency ultrasound to small regions of the brain. To test the efficacy of mFU and the relative contributions of its constituent parts, we visually recorded evoked movements while transcranially stimulating different parts of mouse brain w under light anesthesia. The mFU protocol followed, where possible, existing published ultrasound protocols, using a difference frequency of 500 kHz. mFU induced each of unilateral and bilateral motor function that varied by location, intensity, and the inclusion or exclusion of the low- frequency temporal modulation of the high-frequency carrier wave. # Methods ## Ethics Statement All animal procedures were approved by the University of Washington Institutional Animal Care and Use Committee (IACUC). ## Animal Model and Anesthesia Male C57BL/6 mice, age 8–12 weeks, weight 22–27 g, were anesthetized with a mixture of ketamine and xylazine. Ketamine and xylazine concentrations administered were 87.5 mg/Kg and 8.75 mg/Kg, respectively, with supplemental doses administered as needed to maintain anesthesia during longer experiments. A heating pad set to 100’F maintained the body temperature of the mice while they were under anesthesia. Hair was removed from the top of each mouse’s head via shears and application of Nair® (Church and Dwight Co., Inc., Princeton, NJ, U.S.). Toe and/or tail pinches were given every 10 min to assure mice stayed reactive to such stimuli but were otherwise quiescent. Aquasonic (Parker Laboratories, Inc., Fairfield, New Jersey, U.S.) ultrasound coupling gel was placed on the skin to ensure proper transmission. Brain tissue for histological analysis was collected after perfusion with 1 mL of paraformaldehyde within 5 min of the last experimental trial. Mice not used for histological analysis were euthanized with pentobarbital, concentration 400–500 mg/Kg. ## Ultrasound Sources Two transducers were used for experiment trials: a single-frequency planar, ultrasound source and an effectively multi-frequency, focused ultrasound source. A pulsed, single-frequency UNMOD protocol using a planar piston transducer (Ultran Group, Ultran GS500-D13, State College, PA, U.S.;) was developed following King et al. and Tufail et al. : 88 bursts of 500-kHz ultrasound, each of length 200 µs, at a pulse repetition frequency of 1.5 kHz in a 1-s interval. A focused UNMOD protocol was designed to overlap the UNMOD protocol using the planar Ultran transducer. A dual element, coaxial, confocal and circular transducer and associated matching networks (H-148, Sonic Concepts, Woodinville, WA, U.S.) with a filled, central opening were used. Two Agilent Series 33220A 20-MHz function generators (Agilent Technologies, Santa Clara, CA, U.S.), controlled by a third Agilent function generator, drove two ENI brand model A150 55-dB amplifiers (Electronic Navigation Industries, Rochester, NY, U.S.) that, in turn, powered each of the two transducers within the focused transducer. A LeCroy Oscilloscope (Waverunner LT344,Teledyne LeCroy, Chestnut Ridge, NY, U.S.) monitored the voltage entering each transducer element. During the focused ultrasound for UNMOD studies each element of the transducer was driven at 2 MHz. To study the effects of mFU – the vibro-acoustography technique – one element of the focused transducer was driven at 1.75 MHz and the other at 2.25 MHz, producing a difference frequency of 500 kHz at the focus. Otherwise, the mFU parameters mimicked those of the planar transducer at 500 kHz. For both ultrasound methods, one trial consisted of ten applications of this ultrasound protocol and was completed in approximately 10 s. The length and width of the mFU transducer focus, measured at the ‘half pressure’ value during simulations in water, is 8 mm in the axial direction and 1.5 mm in the lateral direction, yielding approximately 0.015 mL. For the planar ultrasound device, the broad ‘focus’ measured greater than 40 mm in the axial direction and 12 mm in the lateral direction, yielding approximately 4.5 mL. ## Ultrasound Calibration The experimental transducers were placed in a tank filled with degassed and deionized water and the active tip of a calibrated needle hydrophone (HNR-1000, Onda Corporation, Sunnyvale, CA, U.S.) was then placed at the focus of each of the two elements of the dual-element transducer and at the point of maximum pressure from the planar transducer. To verify that the voltage into each element of the dual-element transducer produced the same pressure at a given frequency, thereby insuring that when both elements were run simultaneously each contributed equally to the pressure, peak positive pressure was measured with each element running individually. Each of the two elements produced half the peak pressure that was measured when both elements were combined; thus, we fine- tuned the voltage required by each element to produce half the peak pressure of a predetermined value. To calibrate the Ultran transducer the tip of the needle hydrophone was placed at the center of its planar face and moved axially to locate its broad, maximum peak pressure at roughly 2 cm from the face. For both experimental transducers the spatial peak temporal average intensity (I_spta) is reported. ## Ultrasound Deployment The concave side of the FUS transducer had on its distal surface a hollow, plastic cone with a large opening covered with 0.1524-mm thick latex to allow transmission of ultrasound. Between the transducer face and the latex covering, the transducer housing contained degassed and deionized water. The transducer housing was attached to a metal arm connected to a micro positioner. The positioner stage acted as a 3D-coordinate grid – allowing transducer movement through the necessary x–y, x–z, and y–z planes with sub-millimeter precision. Green laser lights attached to the transducer housing facilitated precise positioning of the transducer focus. A red light emitting diode (LED) attached to a small ruler placed near the front of the animal and within view of a video camera indicated the time of ultrasound application. Body movements and the blinking red LED were recorded with a Nikon D3200 camera (Nikon Corporation, Tokyo, Japan). A plastic, 3D-printed support positioned the mouse to allow the front paws to hang, keep the head secure, and prevent the body from rolling side-to-side. The green lasers were aligned on the surface of the skin, placing the geometric focus of the ultrasound in the same location. The height was marked on the micro-positioner, the transducer moved away from the skull, and ultrasound gel applied. Ultrasound was focused to a position 5 mm below the skin surface – the target depth chosen after imaging the mouse head with a diagnostic ultrasound device. The planar transducer was clamped directly to the micro positioner and the transducer was applied directly onto the mouse scalp coupled with ultrasound gel. ### Ultrasound Administration – Planar (Ultran) Transducer Planar ultrasound was applied in a rostral to caudal sweep along each mouse skull midline with 3 mm between stable positions A, B, and C. One ultrasound trial was administered to each of the three positions with an interval of about 5 min between trials. I_SPTA was 5.25 W/cm². During pilot studies ultrasound was applied continuously and was swept slowly from one position to another. No significant motor activity differences were observed for small movements of the transducer (1 mm scale), but rather only when traversing large portions of the mouse skull (front, middle, rear, on scales of several millimeters). Another series of trials at position B (the mid-sagittal region) were conducted with I_SPTA varied from 0.15 to 5.25 W/cm² by changing only the peak pressure and maintaining the temporal pattern. ### Ultrasound Administration – mFU at Various Intensities For a separate group of mice an anatomical position was determined for each mouse where a 10-s application of mFU via the standard protocol caused robust motor movement. Without moving the source mFU was reapplied for sets of ten more stimulations, varying the number of bursts or pulse duration to decrease the intensity, and recording the evoked movements for each combination. ### Ultrasound Administration – mFU and FUS Applied to Separate Mice The focused transducer was moved incrementally in steps of 1 mm along the top of the mouse skull through six regions each measuring 3×3 mm spanning the bregma to the lambda sutures in a manner that emphasized the parietal region. Each of these six regions was divided into a 3×3 grid to create a 54-element stimulation grid with 1-mm resolution. Each portion of this 54-region grid was stimulated with ultrasound in the same order for each mouse. FUS or mFU application was delayed by five minutes between each of the six major regions. Within a region, the trials were paused only to move the transducer to each new location between applications. ### Ultrasound Administration – mFU and FUS Applied to the Same Mice The standard protocol was amended to five stimulations per position instead of ten. The locations of the stimulations remained the same, with the exclusion of the two most rostral grids (eliminating grids 3 and 4). Within each grid the same paths were followed, except for beginning in the lower right large square (grid 6) and circling clockwise through the three remaining squares (hence to grid 5, then grid 1, then finishing at grid 4) after the stimulations were complete. We performed this study in this fashion motivated by our first results, with mFU alone or FUS alone, where as we report below stimulation of regions five and six – the most caudal regions – produced the majority of the observed induced motions. In each position mFU was applied and any motor movements noted, then the application was switched to FUS by equalizing the carrier frequencies. The trial was repeated if any motor movement with either mFU or FUS was observed. If no movement was observed, the trial advanced to the next location. ## Data Acquisition and Analysis All experimental trials were captured on video using a Nikon D3200 camera complemented by hand-written notes collected by a minimum of two lab members for each trial. This allowed incorporation into the subsequent analysis of the videos observations taken from three different perspectives. Three people reviewed the videos of each experimental trial multiple times while referring to the hand-written notes. ### Definition of a motor movement robustness scale Common movements observed over multiple trials were left and right paw raises (individually or together), paw extension, tail flicks or extension, and whisker flicks. A quantitative measure of the extent of these motions – their ‘robustness’ – was created with an ordinal scale of 0–3. A value of zero was assigned for no observed movement correlated with application of ultrasound, only movement associated with the breathing of the mice. A value of one was assigned for observed faint movements, at least a twitch, with amplitude of up to 1 mm. Paws would twitch up or down, hind legs would briefly flex, and the tail would flick, usually upwards. At this degree, generally only the tip of the tail would move. A value of two was assigned for observed moderate movements with amplitudes as high as 5 mm. Also, partial tail extension would occur, usually lasting a little over one pulse duration. A value of three was assigned for observed strong movements of 1 cm or greater in amplitude and represented the largest motions regularly observed. Less than one percent of the time larger, strained, or rare movements were observed, including limb extension and multidirectional tail movement including spinning. These may have been a sign that the light anesthesia required reapplication; in these cases additional anesthesia was administered to complete the experimental protocol. ### Qualitative measures of induced motor activity A qualitative comparison of mFU with FUS focused on those positions in a given mouse where mFU and FUS could induce motion each time they were applied. In addition to robustness, defined previously, ‘fluidity’ was defined as a measure of the sharpness or crispness of the observed movements as observed grossly, and ‘repetition’ as a measure of the consistency of each action within a trial. A binary determination was made whether mFU or FUS elicited the stronger response at the same anatomical location in the same mouse according to how appropriately they fit the categories. If mFU and FUS could not be differentiated the site was labeled as no discernable difference. # Results ## Planar Ultrasound Device Intensity Sweep shows a logarithmic fit (R² = 0.939) between the ultrasound intensity and the average degree of motor movement caused by the stimulation at that intensity. Results from three mice demonstrate that the greater the ultrasound intensity the larger the induced movement by the ultrasound. Tail movements and bilateral (only) movement of legs and whiskers were observed. ## Planar Ultrasound Device Spatial Sweep Tail movements and bilateral (only) movement of legs and whiskers were observed using the planar ultrasound device with six mice. There was an overall decrease in the number of front leg and tail movements as the planar ultrasound transducer moved from region A (caudal) to region C (rostral) of the mouse brain. Region A had the highest average success rate for front leg and tail movement, while region B had the highest average robustness. A significant difference in the robustness and success rate of front paw activity between regions A and C was also observed, as well as a difference in robustness between regions B and C. Analysis of hind leg movement showed no significant difference in success rate or level of robustness for the three regions. There was, however, a significant difference in success rate and robustness of movement for tail stimulations between regions A and C as well as B and C. ## mFU Applied with Variable Intensity In trials with three mice, the intensity of the standard ultrasound protocol was decreased by changing the number or duration of pulses. Motor responses were induced until reaching 1 W/cm² and robustness of the induced movements decreased linearly with intensity. ## mFU and FUS Applied to Separate Cohorts of Mice Evoked motor responses were sensitive to position of ultrasound delivery with a spatial resolution of ∼1 mm. Moreover, the fraction of successful stimulation events (of ten repetitions) varied considerably. Some of the five mice for each of the FUS and mFU trials showed minimal induced activity while others showed substantial induced activity under each mFU and FUS protocol. When averaged across all mice and all positions, however, there was comparable success between the ability of mFU and FUS to induce movement: mFU induced some type of motor activity in 75 out of 270 stimulations (27.78%) and FUS induced observable motor activity in 77 out of 270 stimulations (28.52%;). Both protocols also caused the same range of motor movement. Averaging over the behavioral results at each grid point over all five mice in the experiments yields low success rates. This analysis, however, obscures the fact that when ultrasound induced movement, it did so for a large percentage of the stimulations. To account for the large variance of results for movement induction, data are also reported only for trials in a given position that showed successful stimulation. The highest success rates for each of mFU and FUS were observed in positions 5 and 6. This is the most posterior location over the parietal region of the brain, and also where the most robust induced motions were observed. In contrast, the most anterior area – regions 3 and 4– showed the lowest percentage of induced motions as well as the least robust movements. ## mFU and FUS Applied to the Same Mice Trials with mFU and FUS applied to the same three mice resulted in 458 total stimulations, with 99 eliciting an observable motor response. Out of a possible 180 positions for stimulation, induced movement was observed in 37 (20.56%) of those locations with mFU, FUS, or both modalities. Of these 37 locations, 13 showed movement induced by both methods. Only mFU or only FUS stimulation induced movement in the other 24 locations, though these results varied significantly between individual mice. The distribution totals of these combinations are given in. Qualitative analysis of those 13 positions where mFU and FUS induced motion each time they were applied detected little difference between mFU and FUS in terms of robustness, fluidity, and repetition. ## Histological Analysis Hemotoxilin and eosin, and cresyl violet stained sections of brain were analyzed for damage associated with ultrasound protocols. All histology samples show unaffected brain with no interesting artifacts (data not shown). # Discussion Studies were performed with both poorly focused and very focused (plus multi- frequency) transducers to compare their ability to induce movement in response to transcutaneous/transcranial ultrasound applied to the brain. Application of ultrasound from a planar source to mouse brain induced motor movements whose amplitude scaled with peak pressure. This differs from previous results reported by King et al., where all or nothing responses were reported. This may be due to our use of ketamine/xylazine versus isoflurane. Inhalation of isoflurane inhibits the transmission of motor evoked potentials through the brainstem. Thus, centrally targeted motor stimulation may require relatively intense ultrasound stimulation to produce an observable peripheral effect. In contrast, ketamine/xylazine has been shown to have only limited effects on peripheral sensory or motor conduction. In pilot studies using isoflurane (unreported here) we observed all or nothing responses similar to those reported by King et al.. Ultrasound from the planar source produced a low-frequency (500 kHz) rapidly pulsed sequence (88 pulses at a PRF of 1.5 kHz, each pulse lasting for 200 µs) with an intensity of 5.24 W/cm², following the temporal structure of the ultrasound protocol demonstrated by Tyler et al. and intensity of King et al.. This stimulation source produced largely uniform, repeatable, and exclusively bilateral motor responses (primarily tail and front leg motion; minimal hind-leg motion) over a range of transducer locations spanning 4–6 mm in a rostral to caudal direction. Manipulation of the location of ultrasound from a planar source served as a first step towards showing anatomical specificity of UNMOD. Although the sweep was limited, it produced differential robustness of induced activity, though not type of induced activity, in a manner that corresponded with three different anatomical regions located several millimeters apart. The same low-frequency ultrasound protocol was used with the modulated (at 500 kHz) high-frequency (2 MHz) focused ultrasound made possible by vibro- acoustography. This mFU protocol generated different motor responses in mice when changing the position of ultrasound application by as little as 1 mm. Both the type of evoked movement, as well as the robustness often varied with mFU probe location. Unilateral paw movements were observed at 28% of locations tested, with the remainder of the evoked paw movements consisting of bilateral motions. In contrast, but in agreement with other studies, we observed only bilateral paw motions generated by the planar transducer. Nonetheless, these results were highly variable both within a given mouse and between mice. Kim et al. report in their studies of neuromodulation that the ultrasound focus was diameter 3.5 mm and length 6.2 mm at full width half maximum pressure (with an associated volume of approximately 0.06 mL). By comparison, the vibro- acoustography studies described here had diameter 1.2 mm and length 8 mm (with an associated volume of approximately 0.015 mL). The ultrasound carrier frequency of 2 MHz was also higher than that used previously, but is still sufficient to transmit trans-temporally through a human or primate skull. If needed, however we can in principle modify our UNMOD protocol via mFU to allow for transmission across thicker regions of the skull by making use of a lower carrier frequency of ∼1 MHz, though we have not tested this in practice. Also, King et al. and others have shown that low frequency UNMOD is effective at frequencies as low as 250 kHz. We can readily apply such low frequency UNMOD protocols within, however, a much smaller volume of brain accessible to single- frequency devices, using the vibro-acoustography paradigm. Indeed, exploration of the potential efficacy of UNMOD at frequencies in physiologically relevant bands (tens to hundreds of Hertz) is possible via our methodology. Supporting this, Greenleaf and colleagues, have deployed vibro-acoustography paradigms for imaging purposes with difference frequencies as low as 7 kHz with no intrinsic reason why they could not go lower. Exploring UNMOD at very low difference frequencies represents an important target of our next research efforts. What constituents of the mFU technique are most strongly correlated to the observed biological effects? Holding the spatial and temporal peak pressure, and the pulse repetition frequency constant while varying the pulse length and number of pulses per stimulation acts to decrease the spatial peak temporal average intensity. Over a significant range of spatial peak temporal average intensity, mFU produced comparable motor responses, though the magnitude of those responses declined linearly as intensity decreased. This linearity of movement response contrasts with the non-linear responses to varying levels of electrical stimulation delivered to the brain, likely due to underlying nonlinearities in current spread and resulting spatial summation of electrical stimulation. Linear activation of neural tissue may be a key advantage of ultrasound stimulation, as non-linear activation of the peripheral nervous system, for example, has limited the clinical utility of functional electrical stimulation. The vibro-acoustography technique reported here also introduces higher frequencies into the ultrasound stimulation protocol than have been considered in previous UNMOD studies. The anatomical specificity, robustness, fluidity, and repetition of motor activity induced by mFU versus FUS were similar. Large variances in observations for mFU stimulation were matched by those for FUS stimulation alone. This observation highlights the likely role of the radiation force found in each pulse of ultrasound (one of the constants between mFU and FUS) as a significant contributor to the observed effect, with the pulse repetition frequency of its application now meriting additional scrutiny. There were clear differences, however, in the ability of mFU versus FUS to produce a motor response when applied to the same portion of brain of the same mouse. This suggests that the pulse-associated radiation force does not represent the sole means of producing UNMOD with ultrasound. While at times mFU and FUS worked comparably well at a given location, they often did not. This difference suggests that the low-frequency component to mFU does contribute in a unique way to UNMOD, and is consistent with our direct observations that FUS alone was not always sufficient to induce motor responses. A greater understanding of this difference will require additional work, including examination of the specific anatomical targets that are receptive to mFU versus FUS. For example, while electrical stimulation of the central nervous system is known to activate axons at lower stimulus intensities then neuron cell bodies or their dendrites, the mechanism by which ultrasound activates neural tissue is currently unknown, and may depend upon the presence or absence of a low frequency component of ultrasound. Finally, both the largely bilateral movements evoked using mFU, and the large variance in induced motion both within and between mice, suggest that multiple deep brain structures are activated, with little direct stimulation of unilateral motor cortex. The regions of the brain likely stimulated during these UNMOD studies with mFU include, but are not limited to, the cerebral cortex, basal forebrain, midbrain (e.g., red nucleus and substantia nigra), hypothalamus, thalamus, hippocampus, cerebral cortex, basal forebrain, caudate striatum, and corpus callosum. All of these structures are involved in motor movement either directly or indirectly. If ultrasound stimulation preferentially activates axons at lower intensities than cell bodies (as is the case for electrical stimulation), the predominance of bilateral movements may originate from activation of large axon tracts such as the corpus callosum, which functions in part to coordinate motor activity between the two hemispheres. In addition, the red nucleus integrates information from the contralateral cerebellum and ipsilateral motor cortex, so its activation (either directly or indirectly) may result in bi-lateral movements of the upper forelimbs, although perhaps most naturally in an alternating pattern such as observed during gait, which we did not observe. The relay circuits of the thalamus may also contribute to the evoked activity, although cortical motor projections are largely lateralized with the exception of a minority of pre-frontal projections. Most probably, the sphere of activation of even focused ultrasound directly activates bi-lateral structures in the mouse brain, suggesting it is necessary to perform larger animal studies to determine the stimulation effects on individual brain areas. ## Limitations The greatest limitation to this project is the size of the mouse brain relative to the focal zone of the ultrasound sources. Even for FUS and mFU, the roughly 8-mm focal length and 1.5-mm focal width of the ultrasound’s highest intensity region is large enough to simultaneously stimulate several anatomically distinct portions of the brain. This limitation may be overcome with a larger animal model in tandem with intra-operative brain mapping. In addition, it is a worthwhile engineering task to produce ultrasound devices with a smaller focus. Several protocol parameters were not explored. These include the pulse repetition frequency of 1.5 kHz, the modulating frequency of 500 kHz, as well as a wider range of pulse lengths and number. Of particular interest would be pulse repetition and modulating frequencies that fall within the frequency band associated with natural physiological brain processes. This would also bring the project closer to direct comparison with neuromodulation by electro-stimulation. Typical frequencies measured by EEG are 8–12 Hz, “alpha” waves; 18–26 Hz, “beta” waves; and \>30 Hz, “gamma” waves. Vibro-acoustography techniques could be used to explore the potential effects of spatially compact but very low-frequency ultrasound signals on brain function. Two of the many phenomena associated with these low-frequency signals within brain are event related desynchronization (ERD) and event related synchronization (ERS). ERD refers to the somatotopically defined decreases in the low-frequency band that occur during motor movement, or decreases in the correspondence between parts of the body and specific regions of brain, while the opposite is true for ERS. Desynchronization can be observed in the idling beta activity peaking around 20 Hz, for example, when an individual processes sensorimotor information or performs a motor task. During these same activities, ERS can be observed by an increase in spectral power in the gamma frequency range. Perhaps mFU with a modulating frequency below 30 Hz could modify the normal ERD or ERS processes by increasing or suppressing the phenomena. Acute histological analysis after UNMOD trials showed damage-free brain. Repeated application of mFU and FUS yielded reproducible results, suggesting that brain function was not altered focally or acutely. In addition, the protocols used spatial peak temporal average intensities within the range that others have reported to be both efficacious and safe. Yoo et al. employed fMRI to observe alteration of brain function in deeply anesthetized rabbits via UNMOD at much lower intensities than we used. By design, they did not observe any grossly observable motor function. Because of the observations of Yoo et al., and our observations of induced motion at I_SPTA values of 1 W/cm², near the FDA limit of 0.72 W/cm², we are optimistic that the mFU embodiment of UNMOD may be deployed within FDA limits for ultrasound. These FDA limits on ultrasound intensity, thermal index, mechanical index, etc. are defined for frequencies greater than or equal to 1 MHz, however. Therefore continued attention to safety is warranted, along with use of a range of observable correlates to successful UNMOD. Examples include fMRI, or in future animal studies the use of fine-wire electromyograms (EMG) to measure potentials across different muscle groups in the legs, tail, and other anatomical structures. This would allow measurements of smaller motor excitations than are visible as gross movements. # Conclusions Transcranial ultrasound applied to the brain can transiently and nondestructively activate it using a range of parameters and devices. Previous research used pulses of low-frequency (250–700 kHz) ultrasound with spatial peak temporal average intensities (I_SPTA) of 0.1–10 W/cm², emitted from transducers that insonified large volumes of mouse brain relative to our system, and all with a single carrier frequency of ultrasound. Typical observations to date include induced motor activity timed to the delivery of ultrasound, but without the ability to vary the type of activity. This study seeks to add anatomical specificity to current neuromodulation practice through the use of focused ultrasound (FUS) by itself, or a modulated variant (mFU). ‘Modulated’ refers to adding complex low-frequency temporal modulation (500 Hz here) of the higher frequency (2 MHz), pulsed and focused waveform in the manner of vibro-acoustography. With lightly anesthetized mice as test subjects, regions of brain were stimulated with 1-mm resolution. Each of mFU and FUS alone were sufficient to induce motor activity, though not always at the same anatomical location. Induction of a variety of motor functions varied by intensity (0.1–5.0 I_SPTA), and by the inclusion or exclusion of the low-frequency temporal modulation of the high-frequency carrier wave. Responses were spatially selective, with diverse movements (both unilateral and bilateral) evoked by both ultrasound methods often at adjacent stimulation locations separated by only 1 mm. In future work we will seek to determine the relative efficacy of mFU versus FUS, to further refine the portions of the UNMOD paradigm most closely tied to its efficacy, and to study focal stimulation of central nervous system structures at the very low frequencies that arise naturally within brain. Finally, there are transcranially delivered therapeutic modalities for transiently altering brain function such as transcranial magnetic stimulation (TMS). TMS works well on shallow anatomical brain structures and within relatively large volumes of tissue. If the early promise of neuromodulation by ultrasound bears fruit, our work and that of our colleagues will point the way for a new therapeutic neuromodulatory modality, one that alters brain function in smaller volumes of tissue at greater depth than current non-invasive technologies based on existing MRI-guided ultrasound devices. We are grateful for the advice we received from Dr. Tyler during the early part of our studies, advice that insured that we could reproduce his work with planar transducers. We thank John Kucewicz of the University of Washington for performing the numerical simulations of Fig. 2. We are also grateful for the constructive comments of Brian Mogen and Mike Kasten of the University of Washington’s Bioengineering and Rehabilitation Medicine Departments, respectively. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: EM PDM JMX CJC NKC. Performed the experiments: EM JMX CJC. Analyzed the data: EM JMX CJC PDM. Contributed reagents/materials/analysis tools: PDM EM NKC. Wrote the paper: EM JMX CJC PDM CTM.

# Introduction Gastrointestinal stromal tumors (GISTs) are soft tissue sarcomas that develop primarily in the stomach (60–70%) and small intestines (20–30%), but also appear in the rectum, colon, esophagus or omentum. These tumors are quite rare, with an estimated annual incidence of 6.8 cases per million individuals in the US between 1992 and 2000, and 3300 to 6000 new US cases predicted each year, though systematic under-ascertainment of GIST cases implies the true rate is slightly higher. Data from the National Cancer Institute's Surveillance, Epidemiology and End Results (SEER) program suggest that GISTs are more common in African- Americans than Caucasians (8.9 versus 4.5 cases per 1 million individuals per year, 1992-2002) but equally common in men and women. Median age at diagnosis in the SEER population is 63 years. Unlike other gastrointestinal neoplasms, more than 90% of GISTs express the *KIT* proto-oncogene, as measured by immunohistochemical analysis of CD117, the stem cell factor receptor protein encoded by *KIT*. In approximately 75% of GISTs, this CD117 overexpression is attributable to a gain-in-function mutation in the tyrosine kinase domain of KIT. Once mutated, *KIT* may encode tyrosine kinase receptors in which the tyrosine kinase domain can be activated in the absence of stem cell factor signaling, thereby stimulating excess, unregulated proliferation of the host tumor cells. Another 10-15% of GISTs have mutations in the *PDGFRA* gene, another tyrosine kinase receptor encoding gene. Primary GIST-related *KIT* and *PDGFRA* mutations have been well characterized. Results from 3 population-based studies in Switzerland, Norway and France suggest that 50-60% of all GISTs have mutations in *KIT* exon 11, 5–10% in *KIT* exon 9, 3% in *KIT* exon 13, 1% in *KIT* exon 17, 2–5% in *PDGFRA* exon 12 and 2–6% in *PDGFRA* exon 18. The proportions observed in hospital-based or convenience samples are generally consistent with these estimates, with some variability due to inclusion criteria and small sample sizes. Most GISTs with primary *KIT* or *PDGFRA* mutations respond to treatment with imatinib mesylate (STI571, Gleevec™, Novartis Pharmaceuticals, Basel, Switzerland), an inhibitor of the KIT and PDGFRA tyrosine kinase. Imatinib is more effective in patients with mutations in *KIT* exon 11 than in patients with no tumor mutations (wild type) or exon 9 mutations. Unfortunately, roughly half of the patients who initially respond to imatinib develop drug-resistant disease after prolonged treatment. This acquired resistance may be attributable to the development of secondary *KIT* mutations in residual tumor tissue. While some *KIT* and *PDGFRA* germline mutations have been identified among families with multiple GIST cases, and a few studies have identified single nucleotide polymorphisms (SNPs) associated with soft tissue sarcoma incidence (*MDM2*), survival (*AhR*), or specific translocations common in some types of soft tissue sarcoma (*XPG/ERCC5*), no studies have looked for inherited genetic risk factors for sporadic GISTs. Though such studies are necessary to advance our understanding of disease etiology, recruitment of cases and compatible controls is limited by the disease's rarity. A population-based study with rapid case ascertainment and collection of detailed information on non-genetic risk factors would be especially arduous, as GISTs are often misclassified in reports to cancer surveillance systems. Given these constraints, we decided to investigate the role of inherited genetic polymorphisms in GIST development by conducting a case-only analysis of the association between tumor mutation type (mutations in *KIT* exon 11, *KIT* exon 9, *PDGFRA*, or wild type) and 225 variants in 39 candidate genes using tumor and blood samples collected during a phase III clinical trial of adjuvant imatinib. In previous studies, certain susceptibility loci have been linked to characteristic tumor mutations, or mutational “signatures”. These include associations between *GSTM1*-null genotype and *TP53* transversion mutations among bladder cancer patients, and certain functional polymorphisms in *XPD* and G:C→T:A *TP53* mutations among lung cancer patients. Similarly, we hypothesized that the characteristic somatic mutations in the *KIT* and *PDGFRA* genes in GIST tumors may be mutational signatures that are causally linked to specific mutagens or susceptibility loci. To address this hypothesis, we selected candidate genes previously linked to soft tissue sarcoma or to environmental risk factors for soft tissue sarcoma. We included genes related to dioxin, phenoxyherbicide, insecticide, vinyl chloride, and radiation response, as well as variants in the previously identified *AhR, MDM2*, and *ERCC5* genes. We also looked at polymorphisms in genes encoding proteins on the *AhR/ARNT* dioxin- response pathway (*CYP1A2*, *CYP1B1*, *HIF1A*, *NQO1*, and *G6PC/G6PT*), other related metabolizing pathways (*ADH1A*, *ADH1B*, *ADH1C*, *ALDH18A1*, *ALDH1A1*, *ALDH1A2*, *ALDH1A3*, *ALDH1B1*, *ALDH1L1*, *ALDH1L2*, *ALDH2*, *CYP2B6*, *CYP2C8*, *CYP2C9*, *CYP2D6*, *CYP2E1*, *CYP3A4*, *GSTM1*, *GSTT1*, *GSTP1*, *HNF4A*, *NAT2*, *NFE2L2, NOS2A*, *PTGS2/COX2*, and *SULT1A1*) – and *TP53*, a tumor suppressor and cell cycle regulation gene closely related to *MDM2*. Additionally, we selected several DNA repair genes (*ERCC2*, *RAD23B*, *XPA,* and *XPC*) in the same DNA repair pathway as *ERCC5*, as polymorphisms in these nucleotide excision repair genes can affect individual sensitivity to carcinogen-induced DNA damage. As our main objective was to conduct a preliminary assessment of these candidate genes rather than any specific variants, we conducted gene-level as well as SNP-level association analyses. # Materials and Methods ## Study population In total, 713 individuals participated in American College of Surgeons Oncology Group (ACOSOG) Z9001, a multicenter, phase III, randomized, double-blind study of adjuvant imatinib (Gleevec™) versus placebo for patients with resected, primary GISTs conducted between July 1, 2002 and April 18, 2007. Cases were eligible if they had a localized tumor of at least 3 cm that tested positive for CD117 by immunohistochemical analysis with the Dako antibody (DakoCytomation, Copenhagen, Denmark). Additional information on the Z9001 trial is published elsewhere. This genetic ancillary study includes the first 333 Z9001 participants who provided a blood sample and consented in writing to unspecified future research using their blood and tumor tissue samples. After removing individuals missing mutation data (n = 52) or more than 10% of their genotype data (n = 2), 279 participants remained. Information on each participant's race, age, sex, and tumor size, site, stage, grade and mitotic rate was available from the parent study. The study protocol was approved by the Institutional Review Boards of Memorial Sloan-Kettering Cancer Center and the University of North Carolina. All participants provided written informed consent. ## Variant selection Once we selected our candidate genes, we identified single nucleotide polymorphisms (SNPs) or deletions within those genes that potentially affected function and had minor allele frequencies (MAF) equal to or greater than 10% in the HapMap CEU population. This included nonsense, missense and splice site mutations, as well as mutations in seed microRNA regions or transcription binding sites. All selected nonsense, missense or splice site mutations were in or near coding regions (within 2000 and 500 base pairs of the 5′ and 3′ ends of the region, respectively). SNPs that did not pass the design phase (designability score \<1 or final score \<0.7) were replaced with surrogate SNPs in high linkage disequilibrium with the original candidate SNP. ## Laboratory Analysis During Z9001 enrollment, all tumor and blood specimens were banked with the ACOSOG Central Specimen Bank at Washington University School of Medicine in St. Louis, Missouri, then DNA extracted from these blood samples was sent to Memorial Sloan-Kettering Cancer Center (MSKCC) for storage at −80°C until analysis. Each sample was genotyped using the GoldenGate genotyping assay (Illumina Inc., San Diego, CA), which consisted of allele-specific extension/ligation methodology followed by universal primer polymerase chain reaction (PCR) amplification regions for the candidate SNPs. Allele-specific oligos and locus-specific oligos hybridized directly to the genomic DNA, upstream and downstream from the targeted SNP before the universal PCR reaction took place. For internal quality control purposes, twenty-seven participants underwent duplicate genotype analysis. Concordance for duplicate samples was 99.9%. SNPs were excluded if they were mono-allelic (n = 3), had a MAF less than 5% in our study samples (n = 6), showed poor clustering (n = 7), or had no individuals homozygous for the minor allele at some levels of the outcome (n = 1), leaving 208 SNPs in the final analysis. Deletions in GSTM1 and GSTT1 were detected using multiplex PCR utilizing sets of target specific and housekeeping gene specific primers. Here, individuals with no copies of the polymorphism of interest (null genotype) were differentiated from those who had one or two copies (wild type). DNA for mutation analysis was extracted from tumor tissue that was snap-frozen and then analyzed as previously described. Briefly, all cases were first tested for *KIT* exon 11 mutations via PCR analysis using Platinum TaqDNA Polymerase High Fidelity (Life Technologies, Inc., Gaithersburg, MD). Tumors without exon 11 mutations were then subjected to PCR analysis using primers for *KIT* exon 9, 13, 14 and 17 and *PDGFRA* exon 12 and 18. ## Statistical Analysis Participants were categorized dichotomously based on the presence or absence of a specific mutation type. The following outcomes were considered: i) a deletion of *KIT* exon 11 codons 557–558, ii) any other (i.e. non-codon 557-8) *KIT* exon 11 deletion, iii) a *KIT* exon 11 insertion, iv) A *KIT* exon 11 point mutation, v) a *KIT* exon 9, exon 13, exon 14, or exon 17 mutation, vi) a *PDGFRA* exon 18 or 12 mutation, and vii) no *KIT* or *PDGFRA* mutation (wild type). Although differentiation by non-exon 11 *KIT* mutations would have been preferable, the prevalence of exon 9, 13, 14 and 17 mutations was too low for independent outcome assessment. We conducted descriptive analyses of selected demographic variables and tumor characteristics, both overall and stratified by gender and race (white vs. non- white). We also compared the covariate distributions of our study population with the remaining Z9001 trial participants to look for possible indications of bias. For each variant, we calculated the race-specific MAF and Pearson χ² p-value for the association between genotype and race. We used Fisher's exact test when one or more cells had less than 5 observations. Additionally, we conducted a crude case-control analysis by comparing the genotype distributions among the white participants (n = 273) to the genotype distributions among individuals of European descent using the HapMap database. Individuals with missing mutation data were included in these descriptive analyses. The association between germline polymorphisms and somatic mutations was analyzed using logistic regression. We obtained odds ratios (ORs), 95% confidence intervals (CIs) and p-values for each SNP-mutation combination, adjusting for race, sex, and age at diagnosis. We coded genotypes as ordinal variables (0 = homozygous for the major allele, 1 = heterozygous, 2 = homozygous for the minor allele) and estimated per-allele ORs and 1 df trend tests. All p-values were corrected for multiple testing by controlling for the false discovery rate. Gene-level association tests were conducted using the sequence kernel association test (SKAT) developed by Wu et al. Here, SNPs are grouped based on prior biological knowledge, in this case occurrence in the same gene, and analyzed using a logistic kernel-machine-based multi-locus test. This method requires fewer hypothesis tests than standard techniques and improves power to detect the effect of an untyped, causal locus by incorporating data from several correlated surrogate SNPs. This method also allows for covariate adjustment, nonlinear effects, and epistasis. Briefly, this method uses a modified version of the variance component score test to assess whether the variance of subject-specific random effects differs from 0. The subject-specific model for each of *n* individuals takes the form: where y_i is the outcome for individual *i*, *x*_i1 to *x*_im are the covariate values for individual i, α₀ to α_m are the regression parameters, *z*_i1 to *z*_ip are the genotypes for individual *i* at genotypes 1 to *p*, and *h*_i = *h*(*z*_i1, *z*₁₂,…*z*_ip) = *h*(***Z*_i**) =  is a function for the subject-specific random effect defined by a positive, definite kernel function of the form K(•,•) and some γ_i, …, γ_n. Assuming **h** follows an arbitrary distribution with a mean of 0 and variance τ**K**, testing the null hypothesis H₀: h(**Z**) = 0 is equivalent to testing H₀: τ = 0, which can be accomplished using a variance-component score statistic : where logit. To obtain a p-value, we can compare *Q* to a scaled χ² distribution with scale parameter κ and degrees of freedom ν, which are modified to account for correlation between SNPs in the same SNP-set (for further explanation, see Appendix A in Wu et al). In this analysis, we opted to use a kernel that models identity-by-state (IBS), or the number of alleles shared by a pair of individuals. This kernel is the most powerful option when epistatic effects may be present. # Results Descriptive analyses are shown in. The median age for included participants was 58.0 years (range 18–85). Approximately half of the population was male (51%) and the majority were white (82%). Most tumors were located in the stomach (66%) or small intestines (31%) and were between 5 and 10 cm in diameter. 70% of evaluated tumors had exon 11 *KIT* mutations, 10% had *PDGFRA* mutations and 13% had no identified *KIT* or *PDGFRA* mutations. Non-white participants were younger, on average (53.0 years vs. 59.0 years), and more likely to have stomach tumors (74% vs. 64%) and exon 11 *KIT* mutations (84% vs. 67%). The most common exon 11 *KIT* mutation was a deletion at codons 557–558 (34%). Compared with other ACOSOG Z9001 participants, the individuals included in this genotyping substudy have similar demographic and tumor characteristics. A somewhat higher proportion of participants in this ancillary study were white (82% versus 76%), but our subpopulation had nearly identical age, gender, tumor size, mitotic rate, tumor location, and tumor mutation type distributions to the full patient pool. Genotype distributions of the 208 variants varied substantially by race , but genotype frequencies among whites in our study population were very similar to the HapMap CEU sample for the 204 SNPs available in both populations. Notable discrepancies included SNPs on several aldehyde dehydrogenase genes, *ALDH1A3*, *ALDH1A2*, *ALDH1L1* and *ALDH1L2*, and two DNA repair genes, *ERCC2* and *XPC*. The associations between each genetic variant and possible outcome are depicted in, with the strength of the association quantified by the inverse of the log of the p-value. While no SNPs were statistically significant after controlling for an FDR level of 25%, some interesting patterns emerged. Most notably, minor alleles at *CYP1B1* rs1056836 and rs2855658 were positively associated with a deletion at *KIT* exon 11 codons 557-8 (OR = 1.81, 95% CI: 1.21–2.71 and OR = 1.91, 95% CI: 1.27–2.86, respectively), while variation in another *CYP1B1* SNP, rs1800440, was positively associated with wild type tumors (OR = 2.65, 95% CI: 1.48–4.76). Having a rare variant at rs1056836 was inversely associated with wild type tumors (OR = 0.54, 95% CI: 0.32–0.92). Minor alleles in two *RAD23B* SNPs, rs7041137 and rs1805329, were more common among tumors with *KIT* exon 9, 13, or 14 mutations (OR_rs7041136 = 3.05, 95% CI: 1.52–6.12 and OR_rs1805329 = 3.24, 95% CI: 1.48–7.11) than tumors without such mutations. The rare form of a third *RAD23B* SNP, rs1805334, was also positively associated with non- exon 11 *KIT* mutations (OR = 2.45, 95% CI: 1.16–5.14). rs50872 in *ERCC2* was the strongest risk factor for *KIT* exon 11 insertion mutations (OR = 2.68, 95% CI: 1.43–5.04) and the rare variant of rs3815029 in *GSTM1* was inversely associated with non-codon 557-8 *KIT* exon 11 deletions (OR = 0.43, 95% CI: 0.25,0.75). Based on the above evidence that at least one variant in *CYP1B1*, *RAD23B*, *ERCC2*, or *GSTM1* was associated with one or more GIST mutation types at p\<0.005, we provided a detailed evaluation of the estimated effects for all of the variants in these four key genes. Effect estimates and p-values for the remaining variants were included in. This table includes results for rs4646755 in *ALDH1L1* and rs3731149 in *XPC*, the strongest risk factors for *PDGFRA* mutations and *KIT* exon 11 point mutations, respectively, both of which had p-values of 0.02. These patterns were preserved in the gene-level SKAT analysis (and), with *CYP1B1* again associated with *KIT* exon 11 codon 557-8 deletions and wild type tumors (p = 0.002 and 0.003, respectively); strong associations between *RAD23B* and KIT exon 9, 13 or 14 mutations (p = 0.002); and *GSTM1* and non-codon 557-8 *KIT* exon 11 deletions (p = 0.01). *ALDH1L2* was also strongly associated with wild type tumors (p = 0.01). As for the other three possible tumor subtypes, *ALDH2* was associated with *KIT* exon 11 insertions (p = 0.03) and the null *GSTT1* genotype was associated with *PDGFRA*-mutated tumors (p = 0.04). No genes were associated with *KIT* exon 11 point mutations (p\<0.05). Although the effect estimates were very imprecise, the associations between the rare alleles of *CYP1B1* SNPs rs1056836 and rs2855658 and *KIT* exon 11 codon 557-8 deletions were even stronger when the analysis was limited to small intestinal tumors (OR_rs1056836 = 5.18, 95% CI: 2.07, 12.95 and OR_rs2855658 = 5.17, 95% CI: 2.05, 13.03). Neither SNP was associated with the outcome in stomach GISTs. No other clear patterns emerged in site- specific subanalyses (data not shown). # Discussion In this preliminary investigation of genetic risk factors for GIST tumor subtypes we identified several genes and SNPs worthy of further investigation. This included SNPs on two xenobiotic metabolizing genes, *CYP1B1* and *GSTM1*, and two DNA repair genes, *RAD23B* and *ERCC2*. Further exploration of the relationship between GISTs and aldehyde dehydrogenase genes or other DNA repair genes (e.g. *XPC*), may also be warranted. *CYP1B1* encodes a cytochrome *P*450 enzyme that is involved with phase I metabolism of PAHs, dioxin, and other chemicals. Two of the *CYP1B1* SNPs we assessed have previously been linked to cancer. This included the rare variant at rs1056836, a missense mutation, which has been linked to increased risk of lung cancer, multiple myeloma and head and neck cancer, with a possible inverse association with pancreatic cancer. Previous evaluations of the SNP's association with breast, colorectal, endometrial and prostate cancer have produced mostly null findings. The rare allele of rs1800440, another missense mutation, was also associated with lung and head and neck cancer, with no reported association with breast or colorectal cancer. However, this SNP did exhibit an inverse association with endometrial cancer. The remaining *CYP1B1* SNP, rs2855658, is located in a seed microRNA region, but has no previously established links to cancer. Although there is little evidence of a link between cancer and the specific *RAD23B*, *ERCC2*, and *GSTM1* variants identified here, previous studies have observed associations between one or more types of cancer and other variants on these three genes. For example, SNPs in *RAD23B* have been linked to esophageal and bladder cancers and one SNP near *RAD23B* was strongly associated with breast cancer in a genome-wide association study. *ERCC2* has also been linked to bladder cancer and a large meta-analysis completed in 2006 reported statistically significant associations between *ERCC2* SNPs and skin, breast and lung cancer. Neither *RAD23B* nor *ERCC2* have been linked to any type of sarcoma. Like the seed microRNA and missense mutation SNPs in *CYP1B1* that were strongly associated with tumor mutations in the present study, some of the identified *RAD23B* and *ERCC2* SNPs also have potentially functional roles. For example, rs13181 on *ERCC2* is a missense mutation, as is rs1805329 on *RAD23B*. Additionally, *RAD23B*'s rs1805330 is a splice site mutation and rs10868 and rs1805334 are located on transcription binding sites. As previously discussed, both *RAD23B* and *ERCC2* are nucleotide excision repair genes. Polymorphisms in these and other DNA repair genes could impair an individual's DNA damage response and affect their carcinogen sensitivity. *GSTM1* is one of several genes encoding glutathione *S*-transferases, which are phase II xenobiotic metabolizing enzymes responsible for carcinogen activation or detoxification. In previous studies, *GSTM1* deletions have been linked to osteosarcoma incidence and recurrence, with a non-statistically significant positive association with soft tissue sarcoma mortality. Other studies of *GSTM1* deletions have identified positive associations between the null genotype and a variety of cancers, included oral, colorectal, cervical, and bladder. None of the association p-values were statistically significant after adjustment for multiple comparisons, whether we applied a false discovery rate correction of 25% or even 50%. While this implies that the observed associations may be due to chance, it should be noted that this was the first investigation of inherited risk factors for GISTs and our main study purpose was to identify variants worthy of further exploration. This study may also have limited generalizability. Study subjects were drawn from a predominantly white clinical trial population, and our findings may not be applicable to other racial groups or to all socioeconomic groups. As the HapMap CEU population is made up of 60 parent-child trios, it may not be an adequate comparison group for our population, especially since we were unable to adjust for unequal distributions of age, gender or other potential confounders. Outcome misclassification is also a potential concern, as tumors with *KIT* exon 11 mutations were not assessed for other *KIT* or *PDGFRA* mutations and we did not test for *PDGFRA* exon 14 mutations in any tumors. However, previous reports suggest that GISTs with 2 or more mutations are rare (\<5%), as are *PDGFRA* exon 14 mutations (\<1%),. Thus, outcome misclassification is unlikely to be a substantial source of bias. While we have only limited evidence that our outcome classification system corresponds to distinct carcinogenic processes in GISTs, linking genetic polymorphisms to tumor phenotypes is valuable for generating etiologic hypotheses. In this small, yet novel, case-only study of genetic risk factors for GIST tumor subtypes we identified several variants in *CYP1B1*, *RAD23B*, *GSTM1*, and *ERCC2* that we believe are worthy of further investigation. We hope that this exploratory analysis serves as a starting point for future research on genetic and environmental causes of these rare and understudied tumors. # Supporting Information We thank Aliaksandra Samoila and Kenneth Cheung for their technical assistance during the preparation of DNA specimens for genotyping and direct sequencing. The authors also wish to thank Michael Wu for his statistical advice. [^1]: Novartis Pharmaceuticals Corporation provided funding for this research project. RD previously served as a paid consultant to Novartis. All other authors declare that they have no conflict of interest. There are no patents or products related to this research. Novartis Pharmaceuticals' role does not alter the authors' adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: LSE IO RD. Performed the experiments: IO CRA KMO LM. Analyzed the data: KMO LSE KB LM. Wrote the paper: KMO LSE.

# Background There is an increasing trend of the use of prescription and over-the-counter (OTC) drugs globally. In the United States (US), the expenditure on prescription drugs were \$456 billion in 2017, an increase of 1.7% since 2016. The trend is expected to continue. The per capita expenditure on prescription drugs is expected to grow at 4% annually between 2018 and 2026. The major contributors to the high growth in drug expenditure include inflation, population growth, increase in per capita prescription, and per capita prescription intensity. In the US, both the proportions of prescription use and per capita prescription intensity increased between 1999 and 2012. Among 18 most commonly used prescription drugs, 11 of them have been used more commonly, including antihyperlipidemic agents, antidepressants, prescription proton-pump inhibitors, and muscle relaxants. For specific populations, the prevalence of using prescription drugs can be as high as 81% among the Medicare enrollees in the United States. In Canada, the information on the historical trends of drug use is limited. It was estimated that 41% of Canadians living in private households are currently using prescription drugs based on the estimates from the 2007 to 2011 Canadian Health Measures Survey (CHMS). The most frequently used prescription medications include lipid-lowering agents, angiotensin-converting enzyme-inhibitors, peptic-ulcer drugs, and acid-reducers between 2007 and 2009 in Canada. There are also other significant drug trends observed by researchers or the public, particularly the use of opioids. In Canada, the overall trend in medication use, such as proportions of drug use and the number of drugs used, remains unclear. In addition, there are changes made to the data collection methods in the CHMS cycle 3, conducted between 2011 and 2013. The changes to the collection methods of the CHMS included an increase in the maximum number of prescription or OTC drugs that can be identified through the Anatomical Therapeutic Chemical (ATC) classification codes or the American Hospital Formulary Service (AHFS) drug codes. In cycles 1 and 2, a maximum of 15 current drugs can be identified. In cycles 3 and 4, the limits on the maximum numbers were removed. The second change in the collection methods of the CHMS was the way the drugs were classified. In cycles 1 and 2, three categories of drugs are identified through drug codes: prescription, over-the-counter (OTC), and health products. At most 15 drug codes and five codes of new drugs can be identified for each of three categories in cycles 1 and 2. In total, at most 60 drug codes (20x3) can be documented for each CHMS interviewee in cycle 1 and 2. From cycle 3 and onwards, three drug categories were reduced to two: prescription and OTC. The restrictions on the numbers of drugs that could be reported were eliminated at the same time. The last change is about how data are stored. In the cycle 1 and 2 data sets, each survey participant has 60 drug records in the same row. From cycle 3 and on, each drug record is stored in separate rows. This requires further data linkage for the CHMS data. The changes to the data collection methods of the CHMS are considered challenges for the study of medication use in Canada using this data. The aforementioned analysis of the CHMS data by Rotermann et al. (2014) only used data from the CHMS cycles 1 and 2 implemented between 2007 and 2011 before the changes in data collection were introduced. Analysis of the CHMS medication data across four cycles and beyond require careful review of the CHMS data to account for the change in reporting. Based on available evidence, we don’t have enough information on the drug trends in Canada and are uncertain about the impact of the changes in data collection in the CHMS. We hypothesize that there is an increasing trend of prescription drug use because increasing trends have been observed in other countries and the latest data collection method does not restrict the numbers of drugs that CHMS participants can report. To test the hypothesis, this study aims to review the potential effects of the changes and verify the quality of reporting by the ATC drug categories before and after the changes made to the data collection method. If there are minimal effects caused by the changes in the data collection methods, the trends in the use and numbers of prescription and OTC drugs between 2007 and 2015 are also illustrated. # Methods The CHMS is an ongoing national survey that collects data related to health measures, biomarkers, and medication use. The details of the sampling strategies and the eligibility criteria to be sampled in the CHMS could be found in Tremblay et al. The eligibility criteria were similar across the four cycles, which were released before December 2017. In brief, those living on reserves or in Aboriginal settlements, institutional residents, and full-time members of the Canadian Forces were not eligible for being sampled in the CHMS, less than 4% of Canadians. In this study, eligible subjects were referred to as non-institutionalized Canadians. The cycles were implemented between 2007 and 2009, 2009 and 2011, 2012 and 2013, and 2014 and 2015. Based on the stratified sampling strategy, collections sites were first chosen and households were sampled from the collection sites. In these households, respondents were selected for interview. There were at least 5000 non-institutionalized Canadians sampled to draw nationally representative statistics for each cycle. The CHMS variables were cleaned according to the user guides. The values representing, “not applicable”, “not stated” and “refused” were recoded to missing values. The medication-related variables in four cycles were identified, screened, summarized and merged to the household data for analysis.\[–\] This secondary data analysis was approved by the ethics review committee at the Centre Hospitalier de l’Université de Montréal (15.115). There is no consent to participate required. ## Criteria for the assessment of the effect of the changes in medication data collection There were three criteria assumed to be important for the assessment of the effects of the changes in medication data collection. The first criterion was the total number of the types of medications ever collected from all CHMS participants. If the limits on the maximal numbers of reported drug codes were important, it was possible that the total numbers of the medication ever reported might be less in the CHMS cycles 1 and 2 because of the censoring. Second, the trends of the numbers or the use of prescription or all drugs might be influenced by the limits. For example, the limits could restrict the total numbers of medications that could be reported and some of the medications might be skipped for the lack of reporting entries. Lastly, the trends in the proportions of diabetes, thyroid, and lipid-lowering medication may reflect their respective disease prevalence. If the trends of disease-specific medication were different from those of disease prevalence, especially when there were lower proportions of disease-specific medication use in cycles 1 and 2 compared to disease prevalence, the restrictions in data collection might be the reason. ## Medication identification The total numbers of prescription and OTC drugs used in the past month were retrieved directly from the respective variables in four cycles (“medd100a” and related variables in cycles 1 and 2; “meufnum” in cycles 3 and 4). The numbers of prescription and all medications were labelled for individuals. For cycles 1 and 2, the ATC drug codes were searched from the variables that represented any prescription drug use, such as those beginning with “atc_1”. For OTC drugs and health products, the drug codes were retrieved from the variables beginning with “atc_2” or “atc_3”. The identified drug codes were categorized as prescription or OTC drugs. For cycles 3 and 4, the drug types, either prescription or OTC drugs, were first identified according to the drug type variable, “meuftype”. Then the ATC drug codes were extracted. The drugs codes were subsequently merged to individual data according to the individual identification numbers in the household data file. The grouping of ATC drug codes were based on the first of the five levels of the ATC classification, namely the first character of the ATC codes in. For example, the codes beginning with C represented therapeutic products targeting cardiovascular system. ## Disease identification In the CHMS survey, there were eight physician-diagnosed chronic conditions reported. In, three of the chronic conditions, cardiovascular disease, diabetes, and thyroid condition, were selected for several reasons. These chronic conditions were included in the surveys through cycles 1 to 4. The questions used for interview were similar from cycles 1 to 4. The medications for the conditions were previously studied with the CHMS data. For each condition, there were one to eight related variables. When individuals reported “yes’ to any of the variables, they were considered to have the conditions. ## Disease-specific ATC codes At the same time, there were three types of medications separately identified for data verification in : diabetes medication, thyroid drugs, and cardiovascular medications. The three groups were identified mainly based on the first one or two levels of the ATC codes. The three drug types were created for the comparison with disease prevalence. If there were major discrepancies in disease prevalence and proportions of medication use especially an upward increase in medication use relative to disease prevalence observed after the changes in data collection in cycle 3, there was the possibility that some of the disease-specific ATC codes in cycles 1 and 2 might not have been properly reported partly due to the restrictions on drug reporting or the changes in medication data collection. The variables for disease identification and the corresponding drug codes were listed in. ## Trend analysis The use and the numbers of prescription or OTC drugs were first described with mean values by cycles. The drug trends were then depicted with ratios with 95% confidence intervals (CIs), compared to cycle 1 (ratio = 1). The 95% CI higher or lower than one was considered significantly higher or lower than the values in cycle 1 respectively. The trends were illustrated by the four CHMS cycles. Because individuals were given weights that were designed to account for non-response and create weighted samples representative of the average population in the period of the survey, the CHMS cycles were used as the proxy for time. The binary outcome of using medication or not was described in proportions. The first levels of the ATC classification codes and the three disease-specific ATC classes were also analyzed for trends. The statistics were weighted statistics and adjusted for survey design, unless otherwise specified. The survey design required the specification of study sites, provinces, weight variables and bootstrap weights. The degrees of freedom were specified as required.\[, –\] Data processing and statistical analysis were conducted with R (v3.20) and RStudio (v0.98.113). # Results The population characteristics were listed in. There were more than 29 million of non-institutionalized Canadians represented in the four cycles and about half of them were female. In the CHMS cycles 1 and 4, there were 739 and 603 types of medication ever identified from all CHMS interviewees respectively. There were 60 variables derived to summarize the use and numbers of prescription alone or prescription and OTC drugs combined. Fifty-six of them were from the first levels of the ATC classification and four were disease-specific classes in. ## Potential impact of the changes in data collection The proportions of the use of disease-specific drugs were plotted in, with the prevalence of cardiovascular disease, diabetes, and thyroid condition. The six curves clustered in three groups in cycles 1 and 2. The highest two curves were the diagnosis and treatment of diabetes. The two curves in the middle were the treatment and diagnosis of cardiovascular disease. The lowest two curves throughout cycles 1 to 4 represented the diagnosis and treatment for thyroid conditions. The trends of the proportions or ratios of thyroid and diabetes medication use were similar to those of the disease prevalence (see for the drug codes and for statistics). In contrast, there was an abrupt increase in the use of cardiovascular drugs between cycles 2 and 3. The curve of cardiovascular drug use deviated from the prevalence of cardiovascular disease. Also, there might be a mismatch in the treatment and diagnosis for cardiovascular disease. The prevalence of cardiovascular disease remained above 57% throughout the four cycles. There was some evidence for significant changes in disease-specific drug use and cardiovascular medication use compared to cycle 1. The proportions of using cardiovascular drugs (ATC level 1: C, ATC C in short) were below 28% throughout cycles 1 to 4. The proportion of using cardiovascular drugs (ATC C) in cycle 3 was higher than cycle 1. The proportion of using thyroid drugs did not change significantly in cycle 2 (ratio = 0.74, 95% CI = 0.51 to 0.98), compared to cycle 1. ## Trends of medication use The trends of medication use were shown in and Figs to. In, the proportions of using any drug, prescription and OTC, remained above 0.88 throughout the four cycles (0.90 to 0.88 from cycles 1 to 4, ratio of cycle 4 = 1.08, 95% CI = 0.89 to 1.26). The proportions of using any prescription remained more than 0.51 throughout cycles 1 to 4 (0.53 to 0.55 from cycles 1 to 4, ratio of cycle 4 = 1.13, 95% CI = 0.89 to 1.37). In and, the numbers of prescription medications and all drugs taken were more than 3.78 and 1.5 throughout cycles 1 to 4, respectively. The numbers of all drugs in use were from 3.9 to 3.8 between cycles 1 and 4 (ratio of cycle 4 = 1.05, 95% CI = 0.86 to 1.24). The numbers of prescription medications in use were from 1.51 to 1.68 between cycle 1 and 4 (ratio of cycle 4 = 1.20, 95% CI = 0.92 to 1.48). The numbers of prescription or all drugs was not significantly different from one for cycles 2 to 4 in. ## Trends in medication use, by ATC classification In Figs and, the ratios of using any prescription and any drugs, respectively, were plotted according to the first-level categories of the ATC classification. The ratios and the proportions according to the first-level categories could be found in and Tables. The drugs of the “various” category (ATC V) could not be estimated due to insufficient sample sizes for prescription use in cycles 1 to 4 in and for the use of both prescription and OTC drugs in cycles 3 and 4 in. The proportions of using prescription for blood and blood forming organs (ATC B) use were significantly higher in cycles 3 and 4, suggesting increasing trends \[ratio of cycles 3 and 4 = 1.740 (95% CI = 1.13 to 2.35) and 1.56 (95% CI = 1.03 to 2.10) respectively\]. Hormonal prescription use, excluding sex hormone and insulins, (ATC H) in cycle 2 was significantly less prevalent (ratio = 0.72, 95% CI = 0.51 to 0.93). The proportions of using other prescription medications in the following categories throughout cycles 2 to 4 were not significantly different from those in cycle 1 in : ATC A, C, D, G, J, L, M, N, P, R, and S. In, the trends of using any drugs, including prescription and OTC drugs were shown. The use of the drug for blood and blood forming organs (ATC B) became significantly more prevalent from cycles 2 to 4 (ratio = 1.89, 2.77, and 2.50 respectively). The use of cardiovascular drugs (ATC C) in cycle 3 was significantly higher than cycle 1 (ratio of cycle 3 = 1.56, 95% CI = 1.17 to 1.94). The prevalence of use of hormonal drugs (ATC H) in cycle 2 was lower (ratio of cycle 2 = 0.75, 95% CI = 0.53 to 0.98). The use of any drugs of the other categories between cycles 2 and 4 was not statistically significantly different from that in cycle 1: ATC A, D, G, H, J, L, M, N, P, R, S, and V. ## Trends in the number of medications, by ATC classification Figs and show the ratios of the number of prescription or all drugs, respectively, throughout cycles 1 to 4. The detailed statistics were in and Tables. The numbers of prescription for blood and blood forming organs (ATC B) were higher in cycles 3 and 4 in \[ratio = 1.75 (95% CI = 1.12 to 2.38) and 1.57 (95% CI = 1.04 to 2.11) respectively\]. The numbers of hormonal prescription (ATC H) were lower in cycle 2 compared to cycle 1 (ratio of cycle 2 = 0.71, 95% CI = 0.50 to 0.91). The change in the number of the prescription medications of the other categories in use between cycles 2 and 4 was not statistically significant from cycle 1: ATC A, C, D, G, J, L, M, N, P, R, S, and V. There were significantly more prescription or OTC drugs for blood and blood forming organs (ATC B) in use in cycles 2 to 4 than cycle 1 (ratio = 1.85, 2.82, and 2.55 respectively). There were significantly more cardiovascular drugs (ATC C) in use in cycle 3 than cycle 1 (ratio of cycle 3 = 1.38, 95% CI = 1.01 to 1.75). There were significant differences in the numbers of “various” drugs (ATC V) in use in cycle 3 and 4 \[ratio = 0.35 (95% CI = -0.16 to 0.85) and 0.02 (95% CI = -0.02 to 0.06) respectively\]. The numbers of the drugs of the other categories were not significantly different from cycle 1: ATC G, J, L, M, N, P, R, and S. # Discussion To the authors’ knowledge, the restrictions on drug reporting in the CHMS cycle 1 and 2 have never been studied before, and are fundamental to the studies that aim to use the CHMS data to understand the patterns of medication use in Canada. After confirming that the changes to medication data collection after the CHMS cycle 2 may not have important or significant impact on data quality, we demonstrate the results of trend analysis of prescription or OTC drug use. There are several findings according to the trends of medication use. The first one is about the potential impact of data collection changes on the trends. The others are about the trends that we discovered based on population representative data. First, the changes made to the method of drug information collection may have minimal effects on the drug reporting, i.e. little censoring of drug codes due to the limits on the numbers of drugs reported in the CHMS cycles 1 and 2. This is related to the criteria we have established for this research objective. The numbers of all types of medication according to the ATC codes reported by all interviewees have been decreasing in contrast to the increasing trend in the use of certain prescription medications. This opposes the fact that more drugs have been approved and are now used in the market. This may suggest that the diversity of drug use is in decline. Marketing, culture, and other factors may be involved in this decline. The exact causes to this finding may need to be investigated in the future. There is a lack of prevalent upward bending in the medication use or numbers of medications between CHMS cycles 2 and 3 to support the hypothesis that changes in medication data collection lead to censoring of information in the CHMS cycles 1 and 2, except for cardiovascular drug use that abruptly increased after cycle 2. Second, only prescription medications or all drugs of one ATC level-one classification, blood and blood forming agents, have been more frequently used in cycles 2 to 4, compared to cycle 1. The use of cardiovascular drugs, prescription only or all, has been more prevalent in cycle 3, implemented between 2011 and 2013. A significant decline in the use of hormonal prescription medications, excluding sex hormones and insulin, was only observed in cycle 2. These findings are similar to the significant changes in prescription drug use in the US between 1999 and 2012, where increases in cardiovascular or anticoagulant prescription use and decrease in hormonal products have been observed. These trends may be related to the fluctuations in disease prevalence and evolving medical practices. In this study, we find that disease prevalence is an important factor. Throughout four CHMS cycles, the use of thyroid and diabetes drugs corresponds well with the disease prevalence among the CHMS participants. In the CHCMS cycles 1 and 2, cardiovascular drug use aligns with disease prevalence well. Then the changes in clinical guidelines may contribute to the rapid increase in cardiovascular drug use, such as the emphasis on the preventive role of aspirin and lipid-lowering agents. The role of aspirin (ATC code: B01AC06) has been stressed since 2009 for the primary and secondary prevention of cardiovascular events. Lipid-lowering agents, statin (ATC codes mostly beginning with C10), have been recommended for patients with type 2 diabetes regardless of their lipid profile since 2008. The changes in medical guidelines may be one of the reasons why cardiovascular drugs have been used more frequently than disease prevalence. The decrease in sex hormones in the CHMS is similar to the trend in the US that have been linked to the decrease in non-contraceptive hormone use, particularly conjugated estrogens, after the release of the Women’s’ Health Initiative Hormone Therapy Trial. Third, there are differences between disease prevalence and proportion of drug use, especially for cardiovascular conditions and related medication. There are factors discussed and tested in the literature, such as inappropriate provider- patient communication and poor disease awareness. It requires further investigation to understand the associations with these factors. The last finding is that we did not find significant increases in the overall trends in prescription or OTC medications in Canada based on a nationally representative survey from 2009 to 2015. Though increasing trends of medication use have been observed in other countries. To our knowledge, this is the first study to describe the overall trends of prescription and OTC medication use in Canada. The findings in our study show that it feasible to conduct trend analysis with the medication data in the CHMS. Several significant trends in drug use have been identified. Some of them can be partly explained by known factors, such as the changes in the clinical guidelines for cardiovascular drugs and the decrease in sex hormone use due to newly published trial results. We encourage researchers to be actively engaged with national data, such as the CHMS, to understand the underlying causes of these trends. In the future, there are still issues to be investigated, such as trends in specific drugs, drug use in certain population groups, and the exact causes of more intense use of cardiovascular drugs. ## Strengths and limitations This study has several strengths including national representativeness, systematic data collection, small proportions of missing information in drug codes, and comparison with disease prevalence. However, there are several limitations to this study. First, this is an analysis of repeated cross- sectional surveys focusing on non-institutionalized Canadians and those living outside reserves. It is likely that medication for injury and acute or severe conditions is not included for study. The drugs being used by those above the age 79 years were not surveyed and the results from the CHMS might underestimate the numbers of drugs in use. It requires caution to interpret the results and is necessary to understand the population under study. Second, the impact of the changes in the data collection methods may require further review. This study is an initial attempt to do so. Third, there are limited sample sizes in certain drug categories, especially the “various” drug category (ATC V) that may be subject to the changes in coding practice in health care (personal communication with Statistics Canada). Fourth, the age range of the CHMS participants is wide, from three to 79 years. Our analysis may not be sensitive to the drug trends that exist only in certain subpopulation, such as adolescents or the elderly. Fifth, the observation time might not be long enough to identify significant increase or decrease in drug use in most of the drug classes in this nationally representative survey, compared to the 13-year follow-up in the US. Lastly, the doses of the drugs are not documented. It may not be feasible to identify medications not distributed by pharmacies or clinics among the CHMS participants. Without information on doses and drugs that may not be obtained from health professionals or pharmacies, it may be difficult to assess the magnitude of ongoing opioid epidemic based on the CHMS data. # Conclusions There is an increasing trend in the use of blood and blood forming agents through cycles 2 to 4 and cardiovascular drugs in cycle 3. There have been restrictions on the drug reporting in the CHMS cycles 1 and 2. The removal of the limits on medication data collection might not have an important impact on the reporting of the use of prescription or OTC drugs based on several findings. First, the numbers of the types of medications identified from all participants decreased between cycle 1 and 4. Second, the rapid increase of the trends of medication use is not prevalent in 14 groups of medications stratified by the first level of the ATC classification between cycles 2 and 3, except for cardiovascular drugs. Lastly, the proportions of thyroid and diabetes medication use increase proportionally with disease prevalence between cycles 1 and 4. The impact of the limits on the numbers of medication that can be reported in the CHMS cycles 1 and 2 may be minimal. Trend analysis of medication use with the CHMS cycle 1 to 4 data is feasible. Trends of specific drugs in subpopulations and the associations with disease prevalence and prescribing practices may need to be further investigated. # Supporting information We would like to thank Danielle McGolrick for reviewing this manuscript and providing constructive comments. AHFS American Hospital Formulary Service ATC Anatomical Therapeutic Chemical BMI body mass index CHMS Canadian Health Measures Survey OTC over-the-counter [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Canadian Agency for Drugs and Technologies in Health, Ottawa, Ontario, Canada

# Introduction Scale is a fundamental concept in geography. It has a great impact on the representation, analysis, and aggregation of spatial data. Previous research shows that nearly all geographical phenomena are scale-sensitive, which further highlights the significance of scale to geographic research. Openshaw described this ‘scale-sensitive’ phenomenon as the modifiable areal unit problem (MAUP), which has two forms—the scale effect and the zoning problem. The scale effect refers to the fact that using coarser/finer analysis units will inevitably lead to different analysis results, whereas the zoning problem refers to the differences caused by the division of the study area even at the same spatial scale (e.g., dividing the study area into rectangles versus hexagons). This study focuses on investigating the scale effect. Although ‘scale’ is an ambiguous term with different semantic meanings, it is often used to refer to the size of the analysis unit or the spatial extent of the study area. In spatial analysis, the size of units directly determines the amount of details to be included in the analysis and the results generated. This process creates the scale effect. With the rapid development of information and communications technologies (ICTs), researchers in spatial science have access to a large amount of spatial data with high spatio-temporal resolutions. The significance of big geo-data on studying human mobility patterns and the socioeconomic environment has been widely recognized. When conducting these urban-oriented studies, aggregating individual-level data into areal units is unavoidable, which raises the issue of choosing an appropriate scale when analyzing big geo-data. For example, in nationwide studies, it is common to use cities as the basic research unit. However, when it comes to urban-scale studies, there is no universally adopted analysis unit. In addition, many types of big geo-data used for urban studies, such as social media check-ins, are extensive data, where the size of the analysis unit determines the amount of data to be included and affects the value of each unit, whereas for intensive data, such as temperature and elevation data, the values are independent of the size of the analysis unit. For intensive data, the mean does not change with different analysis units and the variance declines when the analysis unit gets coarser, whereas for extensive data, the mean value changes when applying different analysis units and the variance can either increase or decrease with a coarser analysis unit. Therefore, analyses relying on extensive data are more scale-sensitive and it is more important to look into the scale effect of extensive data. Previous work pointed out that the analytical results based on a single spatial scale cannot provide a complete view of the actual spatial patterns. One potential solution is to build multi-scale models by aggregating the data at various scales. However, this solution only applies to simple data handling (e.g., spatial statistics), data storage, visualization, and sharing due to its computational complexity. When discovering spatial patterns, researchers tend to arbitrarily choose one spatial scale for simplicity. For example, regular grids are mostly used for urban-oriented studies, where the sizes can vary from 200 m, 250 m, 500 m to 1,000 m. There is insufficient research on how to optimize the spatial scale in urban studies. Therefore, it is important to develop methods that can optimize the choice of spatial scale when characterizing and comparing aggregated data. Similar research was also conducted by remote sensing scientists to select appropriate spatial resolutions for image processing. The basic idea is that an appropriate spatial resolution can be determined by the spatial variation of land surface properties. Various statistical measures, such as the local variance, the variograms, the semivariance at the lag of one pixel, and the scale variance, have been applied to solve this issue. Compared with these methods, a semi-variogram, as a commonly used analysis tool in geostatistics, provides measurable information regarding the variances of spatial units at different lags. Previous work has demonstrated the feasibility of semi- variograms to quantify spatial heterogeneity and explore spatial patterns at different scales. For example, Garrigues et al. constructed a normalize difference vegetation index (NDVI) variogram to quantify the spatial heterogeneity of land cover patterns. The univariate variogram model was extended by Garrigues et al. to a multivariate variogram model that captures the spatial heterogeneity in both red and near infrared bands. Laush et al. studied the effects of spatial and spectral scales on vegetation indices for different types of land cover. Although the concepts are similar, there are fundamental differences between using semi-variograms to identify a suitable spatial resolution (i.e., spatial unit) for remote sensing imagery and using them for individual-level big geo- data. In remote sensing studies, researchers mainly focus on calculating the semi-variances between image cells to classify land use patterns and identify objects. Each image cell is considered a homogeneous unit, so it is not common to look into the variances within a certain image cell. However, individual- level big geo-data are often crowd-sourced point data, such as check-in data harvested from social media sites. Aggregating such point data to areal units inevitably leads to a loss of information. A simple example of aggregating point data is to divide a study area into grid cells, count the number of points in each cell, and use that count as the value of the cell. In this case, it is necessary to look into the variances of points within each cell (i.e., the intra-unit variance) to understand how much information was lost during aggregation. In addition, the cell values in remote sensing imagery are mostly intensive data (e.g., NDVI), meaning that the cell values are normalized and cannot be added up, whereas point-based big geo-data are extensive so the aggregated cell values can be added up. The mathematical operations are different for these two types of data, therefore, the statistical measures for intensive remote sensing data cannot be directly applied to extensive spatial data. Based on the above reasons, our study proposes an innovative strategy to identify the optimal scale for extensive spatial data. # Methods ## The semi-variogram and its key parameters As an efficient tool in geostatistics, a semi-variogram *γ*(*h*) was defined as half of the average squared difference of values between points separated at distance *h*, which is calculated as half a variogram. A set of *γ*(*h*) values can be obtained for each pairwise distance *h* as shown in. The solid line represents a fitted semi-variogram (theoretical variogram) based on scattered semi-variance values (empirical variogram). A semi-variogram is normally an increasing curve of the distance *h* since nearby locations are more likely to be more similar than locations far apart. It consists of three main parameters, the nugget (*c*₀), the sill (*c*₀ + *c*₁), and the range (*α*), reflecting different characteristics of spatial data variance. As *h* increases, the semi-variogram may reach a steady point (i.e, the sill) or increase indefinitely. In previous studies, most variograms can reach the sill within the study area, otherwise the spatial variability of the data goes beyond the predefined spatial extent. The sill (*c*₀ + *c*₁) represents the total variation of the spatial dataset being investigated. As the partial sill, the structural variance (*c*₁) reflects the intrinsic characteristic of data. The range (*α*), which is the distance where the variogram reaches the sill, representing the maximum spatial distance at which the dataset can still demonstrate spatial autocorrelation. Theoretically, a variogram should go through the origin (0,0) because when the lag distance is getting close to 0, the differences between locations separated by this distance also approach 0. However, in practice, it is common for a variogram to not go through the origin and result in a positive intersect value at the y axis. This is called the ‘nugget effect’ and the non-zero intercept is the nugget variance (*c*₀). It is can be caused by measurement errors or micro-variations that occur at a distance *h* smaller than the spatial granularity of the analysis. In other words, the nugget variance (*c*₀) either shows that there are errors during data collection, or locations separated at short distances still have substantially different values. ## Defining the nugget-sill ratio (NSR) for quantifying spatial data variance structures There are three categories of spatial variations of aggregated data based on the relation between the lag distance and the size of the analysis unit: *intra*-*unit* *variation* if the lag distance is smaller than the size of the unit; *adjacent*-*unit* *variation* if the two are equal, and *inter*-*unit* *variation* if the lag distance is larger than the unit size. Intra-unit variation measures the information loss when aggregating point data into areal units, but it cannot be calculated directly. However, it is possible to use the *γ*(*h*) when *h* is smaller than the unit size to approximate the magnitude of the intra-unit variation. To this end, the nugget variance (*c*₀) is a good indicator for estimating the intra-unit variation, because it is the lower limit of the intra-unit variation when *h* approaches 0. In addition, the nugget variance (*c*₀) is scale-related, so it can be a useful indicator to quantify the intra-unit variation of aggregated spatial data at different scales. A large nugget variance indicates more substantial information loss during the aggregation. Therefore, researchers often prefer a small nugget variance if possible. Geospatial big data, such as social media check-in data and taxi origin- destination data, are extensive and additive, meaning that the most common way to aggregate the data is using a simple sum operation. This aggregation process inevitably introduces an increasing intra-unit variation and a larger nugget variance (*c*₀) at a coarser scale. One solution to compare indicators at different scales is to make extensive data intensive (e.g., convert population to population density). Another solution is to use a normalized and scale-free indicator. In this work, we adopt the nugget-sill ratio *c*₀/(*c*₀ + *c*₁) (NSR) as the measure of spatial data variance structures. The NSR contains information about both the intra-unit variation (i.e., the nugget variance) and the inter-unit variation, as the sill (*c*₀ + *c*₁) is an approximation of the limit of the inter-variation; therefore, it is an appropriate indicator of the scale effect in this study. The NSR refers to the ratio of the micro-variance as opposed to the total variance. When point data are aggregated to areal units, attributes of all points within the same unit are represented by the aggregated attributes of the unit. This inevitably causes information loss. Previous studies used this indicator to characterize the spatial dependency between locations, i.e., a smaller NSR shows a stronger spatial dependency. It is based on the principle that the closer two locations are, the more similar attributes they have. This study uses the NSR to quantify the information loss and the scale effect when aggregating point data. ## Estimating semi-variances and calculating the NSR is used to estimate the semi-variance $\hat{\gamma}(h)$. We use regular cells as the spatial unit, and the cell size represents the scale of aggregation. In, *x*_*i* and *x*_*i* + *h* are cells separated at distance *h*; *z*(*x*_*i*) is the value of cell *x*_*i*, which is the sum of properties of all points within that cell; *N*(*h*) is the number of pairs of cells located distance *h* from each other. $$\hat{\gamma}(h) = \frac{1}{2N(h)}\sum\limits_{i = 1}^{N(h)}\left\lbrack z\left( x_{i} \right) - z\left( x_{i} + h \right) \right\rbrack^{2}$$ Considering that the semi-variance value is not statistically reliable at large distances due to the decreasing number of cell pairs *N*(*h*), we chose to only calculate the semi-variance for lag distances smaller than half of the extent of the study area. We also equally divided the x-axis into several ranges of lag distances (instead of using a specific distance) to make sure that we had enough grid pairs in each range. Therefore, is converted to the following format: $$\hat{\gamma}\left( d_{k} \right) = \frac{1}{2N(k)}\sum\limits_{l = 1}^{N(k)}\left\lbrack z\left( x_{1}^{l} \right) - z\left( x_{2}^{l} \right) \right\rbrack^{2}$$ where *k* is the index of a given distance range (i.e., a ‘distance bin’) and *d*_*k* is calculated as a representative distance of the *k*th distance range as defined in ; *N*(*k*) is the number of cell-pairs within the *k*th distance range; $|x_{1}^{l} - x_{2}^{l}\left| \in \right.\left\lbrack h_{k} - \epsilon,h_{k} + \epsilon \right\rbrack$, where *h*_*k* is the median of *k*th distance bin, *ϵ* is a distance tolerance, $x_{1}^{l}$ and $x_{2}^{l}$ represent the *l*th cell pair within the distance bin and $z\left( x_{1}^{l} \right)$, $z\left( x_{2}^{l} \right)$ are values of cells $x_{1}^{l}$ and $x_{2}^{l}$, respectively. $$d_{k} = \frac{1}{N(k)}\sum\limits_{l = 1}^{N(k)}\left| x_{1}^{l} - x_{2}^{l} \right|$$ where $|x_{1}^{l} - x_{2}^{l}|$ represents the Euclidean distance between the cell-pair ($x_{1}^{l},x_{2}^{l}$) and *d*_*k* is the mean value of pairwise distances between all cell pairs in the *k*th distance bin. According to, we can obtain a series of discrete semi-variance estimates, but to calculate the NSR, we still need to fit a continuous mathematical model to the empirical semi-variogram. These models are usually selected from a set of predefined functions, which ensures that the predicted variances are non- negative, such as the Gauss, spherical, exponential, and power models. Although the polynomial model is not included in these models, it is selected with the Gauss model in this study, as they can match the shape of empirical variograms and also follow the aforementioned ‘non-negative’ principle. The Gauss is defined: $$\gamma(h) = c_{0} + c_{1}\left( 1 - e^{- \frac{h^{2}}{d^{2}}} \right)$$ where *d* is the distance parameter; the practical range is $\sqrt{3}d$. The polynomial model is $$\gamma(h) = - \frac{c_{1}}{a^{2}}h^{2} + \frac{2c_{1}}{a}h + c_{0}$$ where *a* is the range (*α*). ## Research process shows a scale-adaptive integrated method for scale effect evaluation with three steps: (a) aggregating point data into areal units under multiple scales; (b) fitting the empirical variogram models and calculating indicators; and (c) comparing quantitative results. First, we aggregate discrete point data into cell units with different sizes. Values of discrete points within each cell are accumulated as the attribute value of that cell. For example, the number of check-ins in each cell is the sum of check-ins from all discrete points in that cell. We then estimate semi- variances for each cell pairs in the study area. Second, we fit the empirical variogram and calculate the NSR. For aggregated data at each scale, we take the same steps as follows: Before estimating semi- variance values, we need to decide the distance tolerance *ϵ*. To investigate the scale effect, we assume that the *ϵ* is only proportional to the cell size, and define that the number of cell pairs within each distance bin should be more than 30 as suggested by Huijbregts. After calculating the discrete semi- variogram estimates, we fit the Gauss and polynomial models and then calculate the NSR. Finally, we plot a diagram showing the correlation between the NSR and the scale. A small NSR value means that the intra-cell variance accounts for a small percentage of the total variance, and thus it can be considered as a guide to the optimal scale (*S*_*o*). More detailed analyses are presented in the ‘experiments with synthetic data’ section and the ‘case study’ section. In general, this method builds a bridge between scales and observed spatial data and then quantifies the scale effect by comparing a group of indicator values at different scales. # Experiments with synthetic data We designed two sets of simulation experiments to verify the feasibility of the method. The simulated data in this section are scale-free. The entire study area is a square-shaped 1,000 × 1,000 areal unit. The cell sizes were selected from 10 to 80 with a 5-unit interval. We generated simulated points from a two- dimensional normal distribution N(*μ*₁, *μ*₂, *σ*₁, *σ*₂, *ρ* = 0). Centers of the generated points were fixed to the center of the study area (*μ*₁ = *μ*₂ = 500), and *σ* (*σ*₁ = *σ*₂ = *σ*) is a variance parameter representing the spatial dispersion of the simulated data. In addition, *N* denotes the number of discrete points generated. shows the spatial distribution of the generated points and the corresponding heat maps at various scales. At a finer scale, each unit is small and more similar to each other. As the scale gets coarser, it is easier to see the spatial heterogeneity of different cells. We conducted two simulations with different *N* and *σ* to test the proposed method. Setting *σ* = 150 ensures that 99% of the generated points can be included in the study area. We set the number of points *N* as 1,000, 2,000, 3,000, 10,000, 20,000, and 30,000 to test different parameter settings. To ensure reliable results, we generated data with the same parameters 20 times and calculated their average NSR. The correlation between the NSR and the cell size are shown in. Except for the last sub-figure where *N* = 30,000, the other sub-figures all show a U-shaped curve, where the NSR first decreases and then increases when the cell size increases (i.e., when scale gets coarser). At the finest scale (cell szie = 10), the whole study space is divided into fragmented small units. Each cell only contains a small number of points. As a result, the intra-cell variation and the inter-cell variation are very close, resulting in the NSR approaching 1. It is similar to the ‘pure nugget effect’ discussed in, where the semi-variance shows similar values at all lags. The differences between cells are gradually revealed when the cell scale increases. When the scale gets too coarse, the spatial information inside each cell is highly generalized, and the variance of cell values decreases. The intra-cell variation is again getting close to the inter-cell variation, which leads to an increase in the NSR. That proves that the NSR is effective in characterizing the structure of spatial variances. We defined the optimal scale (*S*_*o*) as the cell size when the NSR reaches the first local minimum. Based on this definition, *S*_*o* = 45, 45, 30, 20, and 15 where *N* = 1,000, 2,000, 3,000, 10,000, and 20,000, respectively. However, we cannot find an *S*_*o* corresponding to when *N* = 30,000, since the scale is less than the minimum scale (10) we considered. The *S*_*o* values get smaller with an increasing *N*, which indicates the impact of data density on the scale effect. Based on the results from the previous step, we set *N* = 10,000 and explored the role of *σ* in determining the scale effect. The six graphs in demonstrate a similar U-shaped curve, which is consistent with results from the previous step, and *S*_*o* = 20, 20, 25, 25, 25, and 30 when *σ* = 140, 160, 180, 200, 220, and 240, respectively. As *σ* increases, the minimum NSR appears at a coarser scale, which implies that *σ* also has an impact on the scale effect. In other words, when aggregating data with a larger variance, each analysis unit naturally contains more information compared to when the variance of the data is smaller, thus it is preferable to use a finer analysis unit. On the contrary, when *σ* is smaller, the corresponding analysis unit should be coarser. # Case study ## Data description In addition to the simulated experiments, we selected two major cities, Beijing and Shanghai, and two medium-sized cities, Chengdu and Wuhan, to test the methodology. Because human activities are mainly concentrated in the urban area of these cities, we defined the study areas with the same dimensions (30 km by 30 km) in all cities. The study area of Beijing covers the districts within the 5^*th* ring road and the study area of Shanghai covers the central area within the outer ring road. For Chengdu, the study area covers the area within the 4^*th* ring road and the study area of Wuhan consists of the area within 3^*th* ring road and partial regions within the 4^*th* ring road. To mitigate the zoning effect, we used regular grids as the analysis unit in this case study, with cell sizes from 300 m to 2,000 m with a 100 m interval. The actual study area may be slightly smaller than 30 km × 30 km, as the number 30 km is not divisible by certain cell sizes (e.g., 700 m, 800 m, etc). We used the point of interest (POI) check-in data in Beijing, Shanghai, Chengdu, and Wuhan from Sina Weibo in 2014. Sina Weibo is the biggest microblog service in China functionally similar to Twitter. We collected our dataset using the official Weibo Application Programming Interface (API) (<https://open.weibo.com/wiki/2/place/nearby/pois>). For each POI, the record includes its place name, address, geographical coordinates, and the number of check-ins at this POI. After data filtering, we obtained 88,886, 78,864, 26,907, and 24,542 valid records for Beijing, Shanghai, Chengdu, and Wuhan, respectively. We calculated the number of check-ins within each cell at different scales. shows the heat maps of check-ins at varying spatial scales. As can be seen, the spatial distribution of check-ins demonstrates very different patterns in each city. In Beijing, the data show a polycentric pattern where there are multiple clusters of check-ins in different parts of the city. Beijing’s polycentric urban activity pattern has been discussed in many previous studies. In addition, there are more check-ins in the north than in the south. This is potentially because the northern side of Beijing is more developed with better facilities and infrastructures. Unlike Beijing, Shanghai shows a monocentric pattern where the check-ins are concentrated in the southwest of the study area (i.e., the central urban area of Shanghai). For Chengdu, cells with a higher check-in density are mainly clustered in the center and to the south of the city, and there are a few high density cells in the outer areas of the city. Check-ins in Wuhan show a morphologically polycentric pattern, which is potentially due to its complex configuration of water bodies. As shown in, the Yangtze River and the Han River divide the urban center of Wuhan into three sections. There are also a large number of lakes and other water bodies that contribute to the discontinuity of the central urban area in Wuhan. In addition, the spatial distribution of check-ins varies at different scales. At a finer scale, high-value cells scatter across the whole study area, whereas at a coarser scale, high-value cells are more clustered. For example, the heat map of Shanghai when the cell size equals 1,400 m shows a monocentric pattern; however, at other scales, the heatmaps show a polycentric pattern. ## Quantitative results of the scale effect As mentioned in the methodology, we used the NSR to quantify the scale effect of POI check-ins in the four cities. We applied several strategies to ensure the robustness of the experiment design. First, to mitigate the impact of the actual study area, at each scale, we chose ten slightly different 30 km × 30 km study areas within each city and calculated the average NSR. Second, we tested different models that can be used to fit into the estimated semi-variance values, and we chose the polynomial model due to its better fitting performance. The fitted models were evaluated based on the goodness of fit *R*², which is a commonly used indicator to measure the performance of a statistical model. illustrates the correlation between the NSR and the spatial scale in the four Chinese cities. As can be seen, all four curves are U-shaped and the maximum NSR values appear at the finest scale (i.e. 300 m). The optimal scale *S*_*o* for Beijing, Shanghai, Chengdu, and Wuhan are 600 m, 600 m, 900 m, and 700 m, respectively. Due to the differences in city sizes and the amount of data, we will compare two large cities (Beijing and Shanghai) and two medium-sized cities (Wuhan and Chengdu) separately to better interpret. As shown in, Beijing has a slightly larger NSR range than Shanghai, indicating that the spatial characteristics of Beijing are more affected by spatial scales. Both curves can be divided into two parts with 600 m as the cut-off point. The former half of both curves show a rapid decline, but the latter half indicates an increasing trend. Specifically, the NSR in Shanghai first drops rapidly and then rises slowly with slight fluctuations; while the NSR has more fluctuations in Beijing after the first local minimum. This suggests that urban configuration of Beijing is more complex than that of Shanghai. For Chengdu, the NSR values are generally lower than those in Wuhan. This shows that POI check-in data in Chengdu are more spatially dependent. Moreover, it should be noted that the NSR for Wuhan shows a more fluctuated pattern. This is probably because check-in data in Wuhan demonstrate a more dispersed pattern due to the large number of lakes and ponds in Wuhan. In addition, the *S*_*o* for Chengdu and Wuhan are coarser than for Beijing and Shanghai, which further demonstrates the impacts of data density on the scale effect discussed in the experiments using synthetic data. ## Evaluation of the optimal scale *S*_*o* To further evaluate the results, we introduced an indicator (the homogeneity within a cell, denoted by *Hom*) based on the q-statistic. Q-statistic is a statistical method for measuring the degree of spatial stratified heterogeneity and uncovering its possible determinants. Spatial stratified heterogeneity refers to the phenomenon that occurs when dividing a study area into sub- regions, the within region variance is smaller than the between region variance. A typical example of the spatial stratified heterogeneity is the differences between climate zones. The division of the sub-regions inevitably affects the degree of spatial stratified heterogeneity. In the scope of this study, a smaller *Hom* represents a greater spatial stratified heterogeneity, which further indicates a more homogeneous pattern within the sub-regions (units) and less information loss during the aggregation. It provides a feasible measure for validating whether the obtained optimal scale, *S*_*o*, corresponds to the least amount of information loss. The *Hom* indicator is defined as: $$Hom = \frac{SVI}{SVT} = \frac{\sum_{i = 1}^{n}N_{i}\sigma_{i}^{2}}{N\sigma^{2}}$$ where $\sigma_{i}^{2}$ is the variance of POI check-ins within cell *i* and *σ*² is the variance of all POI check-ins; *N*_*i* is the number of POIs within cell *i*, *N* is the number of all POIs, i.e. $N = \sum_{i = 1}^{n}N_{i}$. *SVI* and *SVT* are the sum of intra-unit variances and the total sum of variances, respectively. For given spatial point data, *SVT* is fixed while *SVI* varies with the scale. The *Hom* indicator ranges from 0 to 1. It represents the degree of intra-cell variability at different scales, which is reflected by not only the number of points within an analysis unit, but also the variance of attributes at these points. A lower *Hom* indicates that the analysis units are more homogeneous internally and the intra-unit variances are lower, and therefore we have less information loss. As the scale gets coarser, the number of POIs within each cell increases. This results in an increase in intra-cell variability. Moreover, spatial point patterns can be a result of a complex urban configuration or multiple spatial processes, thus there may be more than one local minimum value of the *Hom* indicator (i.e., the optimal scale). In this paper, we adopt the same idea as the elbow method when deciding the number of clusters and use local minimums to identify the optimal scale based on the *Hom* indicator (*S*_*hom*). shows the evaluation results for four Chinese cities. According to these results, we can conclude that for Beijing, Shanghai, Chengdu, and Wuhan, the *S*_*hom* values are 500 m, 700 m, 700 m, and 800 m, respectively. These results are slightly different from the *S*_*o* values calculated from the NSR because the *Hom* indicator only measures the intra-unit variability and does not consider the inter-unit variability. In other words, at each given scale, if we randomly switch the locations of cells, the *Hom* indicator remains constant, because the parameters in (i.e., *N*, *N*_*i*, *σ*, $\sigma_{i}^{2}$) do not change as long as the cells are still divided the same way. The city with the biggest difference between the *S*_*hom* and the *S*_*o* values is Chengdu. For other cities, the optimal scales recommended by the NSR and the *Hom* indicator are very similar. In addition, the orders of the *S*_*o* and the *S*_*hom* values for these cities are consistent. Both the *S*_*o* and the *S*_*hom* values for Beijing and Shanghai are smaller than those for Chengdu and Wuhan. This further validates the robustness of our method. # Discussion ## Influence of different fitting models Considering that fitting different models to the same semi-variogram may lead to different results, we adopted multiple models to fit the same data to investigate how the model selection may impact the scale effect. We employed the polynomial model and the Gauss model to fit the simulated data in the ‘experiments with synthetic data’ section 3 to compare the results. Because the results are similar, we only discuss one set of data (*N* = 10,000, *σ* = 150) as an example. As illustrated in, the solid and dashed lines are results fitted by the Gauss and polynomial models, respectively. It can be seen that the NSRs fitted by the Gauss model are generally higher than those fitted by the polynomial model, which is potentially due to the different characteristics of the models. Although the absolute values are different, both curves follow the same trend. In addition, the optimal scale *S*_*o* values are consistent. These results show that the choice of models does not have a substantial impact on the results in our analysis. ## Experiments with dual-centered data The patterns in the ‘experiments with synthetic data’ section are mostly single- centered point patterns where the points show one central cluster of high density values. We also designed two comparative experiments to investigate how the number of centers affects the scale effect. As shown in, we extended the single-centered simulated data to dual-centered patterns and calculated their *S*_*o*. The *S*_*o* for dual-centered patterns when *N* = 10, 000 and 20,000 are both coarser than the *S*_*o* from single- centered patterns with the same *N*, which implies that the increase of the number of centers leads to a coarser *S*_*o*. The *S*_*o* for the dual-centered pattern when *N* = 10,000 is coarser than that when *N* = 20,000. This is consistent with the conclusions from single-centered data regarding the impact of *N* on the optimized scale. ## Potential applications and limitations The MAUP is a pervasive phenomenon for both intensive and extensive geographical data. However, there is insufficient research on how to quantify the scale effect caused by the MAUP. Due to the rapid development of ICTs, more and more individual-level crowd-sourced data are generated on a daily basis, and many of these datasets are point-based. To this end, this paper defined a series of indicators (e.g., the NSR and the *Hom* indicator) to quantify how the MAUP manifests itself when aggregating point data into areal units. The methodology can be used to select the optimal scale for aggregation in various applications in geography. For example, similar to the Weibo case study, many researchers have used various types of point-based location data (e.g., social media check-in data, georeferenced mobile phone records, and taxi pick- up/drop-off data) to analyze the magnitude of human mobility in different urban districts. Research questions can range from basic summary statistics like “Which part of New York has the most Twitter check-ins during Christmas?” to a more complex one like “How should we quantify the mobility flows between urban regions based on taxi data?” In all these studies, a crucial data pre-processing step is to determine the size of the spatial unit for aggregating the point data. The proposed method provides a feasible way to quantitatively assess the influence of the scale of analysis units when applying crowd-sourced data to urban studies. The application of the proposed method is not limited to using crowd-sourced mobility data in urban geography. In demographic studies, a remaining challenge is to mitigate the MAUP caused by aggregating population data based on different spatial units. In physical geography, a similar problem exists for animal tracking data, where researchers need to determine the optimal scale for aggregating location points from tracking devices. This study takes a first step in providing a feasible solution to the aforementioned research problems. The proposed method has several limitations. First, through trial and error, we found that this method does not work well when the data is very sparsely distributed. Sparsely distributed data shows no clear spatial pattern, so it is difficult to find a suitable model to fit the estimated semi-variance values. This may limit the application of this method to sparse datasets, such as taxi pick-up data after midnight or check-in data in a remote rural area. In addition, the quality of the check-in data may be influenced by various factors such as the representativeness of social media data and the strategy of data collection, which inevitably affect the results in the case study. Social media sites like Weibo are more likely to attract users with a certain demographic profile (e.g., young people), which leads to a biased sampling of the population. For the Weibo POI data used in this study, users are allowed to check in to a POI when they are within a certain distance of the location, which naturally leads to data accuracy issues. Because this study aims to propose a methodology instead of generating empirical results, we did not directly address the data quality issues. In practice, researchers should be aware of the influence of data quality issues on the results when applying our method to their own data. # Conclusion The scale effect is an important issue in geography. With the development of ICTs, massive high-resolution geo-tagged data is available for investigating human mobility patterns and the socioeconomic environment. Spatial aggregation is necessary to investigate collective patterns from individual-level big geo- data, and this inevitably leads to the challenge of selecting an optimal scale in spatial analysis. We proposed a method to quantitatively evaluate the scale effect of extensive data, which is a common type of big geo-data. Because semi-variograms can provide rich spatial information at different lag distances, we employed the nugget-sill ratio as a quantitative measure to characterize the structure of spatial data variance at multiple scales. Two sets of simulated experiments showed that both very fine and very coarse scales lead to high NSR values, and a low NSR tends to appear at a medium scale. This observation is consistent with the structures of spatial variances. In addition, we defined the scale where the first local minimum NSR occurs as the optimal scale (*S*_*o*), the results show that as *σ* (i.e., the dispersion of spatial data) increases, the *S*_*o* value gets coarser. The conclusion is consistent with our perception that a finer analysis unit is more appropriate for data with a higher spatial heterogeneity, otherwise a coarser scale is more suitable. It demonstrates the rationality of our method in quantifying the scale effect. We also used Weibo check-in data from four Chinese cities (Beijing, Shanghai, Chengdu, and Wuhan) as a case study. The results suggest that the optimal scale *S*_*o* for these cities are 600 m, 600 m, 900 m, and 700 m, respectively. Overall, a very fine scale indicates that the analysis units are too small and there are not enough points to be aggregated in each unit; however, a scale too coarse will lead to over-generalization of the data and a substantial loss of information. Therefore, it is important to find a balance point between the level of detail and the degree of aggregation, which is the main contribution of this study. We adopted a classic geostatistical method (i.e., the semi- variogram) and provided a new perspective to quantify intra-unit variation and inter-variation at different scales. The method in this study offers a useful data processing strategy to optimize spatial scales when aggregating big geo- data in urban studies. Geospatial big data have many potential issues, such as data sparsity, data representativeness, and other data quality issues, which can lead to non- stationary results. In fact, the semi-variogram contains much more information than we explored in this study and may be useful for improving our results in the future to optimize the scale of spatial analysis. In addition, the proposed method focuses on quantifying intra-unit variation when exploring the scale effect, but in practice, different datasets and urban studies may need to adopt different indicators based on their specific needs. Future studies should focus on expanding this framework by exploring the choice of indicators for different datasets. # Supporting information We thank the anonymous reviewers and the academic editor for their constructive comments, which greatly improved the content and clarity of this paper. Ximeng Cheng and Xin Yao helped with the research design and edited an early draft of this paper. Mr. Lei Zhang helped improving the grammar and style of this work. [^1]: The authors have declared that no competing interests exist.

# Introduction The spatial patterns of species diversity adapted to the main environmental gradient have been a hot issue in ecological and environmental sciences. Comprehending spatial patterns in biodiversity and the causal factors behind them are fundamental to developing sound conservation and management strategies, as well as to studying global climate change and predicting its biological impacts on vegetation. In mountain areas, altitude is an important gradient due to large environmental changes across a relatively short geographical range. For example, altitude drives drastic changes in abiotic factors, such as water, temperature, soil properties, and areas. In recent years, altitudinal diversity patterns have been categorized into four major patterns. The most common pattern is a unimodal curve with high richness at intermediate altitudes, which accounts for about 50% of previous studies. Another common pattern is a monotonically decreasing curve with increasing elevation, found in about 25% of previous studies. Monotonically increasing or no obvious trend with increasing elevation accounts for another 25% of previous studies. The variation in species diversity patterns can be caused by many factors, such as climate, productivity, anthropogenic influences, evolutionary history, and biotic interactions, e.g. competition and facilitation. Although there are many proposed hypotheses to explain the altitudinal patterns, the underlying mechanisms remain poorly understood. Beta diversity is important for understanding spatial patterns of species diversity. It can provide insights into the mechanisms driving community assembly and can guide the selection of conservation areas. Beta diversity is used to describe turnover in species composition across spatial and temporal scales. Many methods for calculating beta diversity have been developed. Distance-decay is the slope of the relationship between the similarity in species composition between two communities and the spatial distance separating them. This method is favored for exploring ecological mechanisms because it is sensitive to key spatial processes, such as dispersal limitation. The similarity in species composition is reduced with increased spatial distance. Different traits among plant taxa, such as body size, dispersal ability, and niche width, may have significant effects on their response to environmental gradients. The slope of the relationship between similarity and distance is steeper for plant taxa with weaker dispersal ability and narrow niches. In contrast to the long-term studies on the spatial patterns of species diversity, research on spatial patterns of phylogenetic diversity is still in its infancy. As the availability of phylogenetic information has increased in recent years, ecologists have attempted to apply phylogenetic approaches to traditional diversity analyses. Because considerable phylogenetic conservatism has been found in the niche positions of plants, species that have close phylogenetic relationships usually have similar ecological and functional characteristics. If communities are structured mainly by environmental filtering, environments could select species that have certain physiological tolerances, resulting in co-occurring species having close phylogenetic relationships. In contrast, if communities are structured mainly by interspecific competition, species occupying similar ecological niches compete more strongly with one another and cannot co-exist readily, which causes species occurring in a local community to be overdispersed in the phylogeny. Therefore, studying phylogenetic diversity and structure can reveal ecological and evolutionary processes regulating community assembly and provide novel insights into exploring the mechanisms that are driving current diversity patterns. In order to understand altitudinal patterns of species diversity and its potential causes, we used multifaceted methods to examine altitudinal patterns in two temperate mountain systems of northern China. First, we examined patterns of species richness and composition turnover along the altitudinal gradient. Then, we employed phylogenetic approaches to explore altitudinal patterns of phylogenetic diversity and phylogenetic turnover, and compared whether phylogenetic diversity and species diversity patterns are consistent. Furthermore, we examined phylogenetic structure along two elevation gradients to understand the mechanisms that are driving current diversity patterns. # Materials and Methods ## Study areas This study was conducted on Mount Tai (N36° 5′–N36° 15′, E117° 5′–E117° 24′) and Mount Lao (N36°5′–N36°19′, E120°24′–E120°42′), which are located in the western and eastern regions, respectively, of Shandong Province of northern China. Mount Tai is one of China's five most famous mountains, and was listed as a UNESCO World Cultural and Natural Heritage site in 1987. The altitude of the main peak is 1532.5 m, which is the highest in Shandong Province. Mount Lao is surrounded by the sea on two sides. It is the highest peak on the Chinese coastline at 1132.7 m. Mount Tai and Mount Lao are both nature reserves in China and play an important role in local development, such as contributing to the tourist industry and providing ecosystem services. Mount Tai features a warm-temperate continental monsoon climate. On the top of the mountain, annual mean temperature is 5.3°C and annual mean precipitation is 1124.6 mm. At the bottom of the mountain, annual mean temperature is 12.8°C and annual mean precipitation is 715.0 mm. The average frost period lasts 159 d, and the extreme minimum temperature is -27.5°C.The annual relative humidity is 63%. The climate of Mount Lao has a temperate maritime climate owing to the proximity of the ocean. Mean annual precipitation increases with elevation from 726.6 mm to 2103.8 mm. The annual relative humidity reaches 73%. The mean annual temperature is 11.9°C. The average frost period is over 186 d and the extreme minimum temperature is -21.2°C. Mount Tai and Mount Lao are rich in vegetation with temperate deciduous broad-leaved forests and temperate coniferous forests. ## Field sampling Vegetation studies were carried out on Mount Tai and Mount Lao in August and September 2012 and 2013, with the permission of the scenic spot management committee of Mount Tai and Mount Lao. In order to systematically sample, the altitudinal gradient was divided into 100 m-wide elevation belts. Two or three study plots (20m × 30m) were set up in each elevation belt. The sampling sites were chosen selectively to include mature forests which were representative of typical vegetation types in each altitudinal belt and to be away from slope crests and clough, large rocky outcrops, and stream gullies. A total of 63 plots were sampled on Mount Tai and Mount Lao. All trees in each plot were identified to species and their diameter at breast height (DBH) was recorded if DBH≥5 cm. All shrubs and herbs appearing in each plot also were also identified to species. Each plot was located by using GPS (Garmin 621SC, Garmin Corp). Elevation, slope, aspect, and human disturbance for each plot were also recorded. Total basal area of trees per plot was calculated to indicate canopy density. The assessment of human disturbance was based on the distance from plots to surrounding roads, hiking trails, and infrastructure, the extent of trampling by tourists, and the amount of trash. The intensity of human disturbances was classified into four grades: 1 = no obvious disturbance, 2 = weak disturbance, 3 = medium disturbance, 4 = heavy disturbance. ## Data analysis In this study, species diversity was defined as the total number of species recorded in each plot. We calculated total richness, tree richness, shrub richness, and herb richness. Beta diversity was determined by the similarity of species composition between two plots and calculated using the Jaccard similarity index: $$\beta_{j} = (b + c)/(a + b + c)$$ where a is the numbers of species occurring in both sampling units, and b and c are the number of species unique to one sampling unit and to the other sampling unit, respectively. The higher the Jaccard similarity index, the larger the difference in species composition between two plots. β_j was calculated separately for trees, shrubs, herbs, and total species. We used the Qian and Jin‘s (2015) PhytoPhylo megaphylogeny as a backbone to construct phylogenies of angiosperm plants for Mount Tai and Mount Lao, respectively ( and Figs). We applied S.PhyloMaker package of R software to generate phylogenies based on Scenarios 3. Phylogenetic diversity was quantified using Faith’s index, which is defined as total branch length among all taxa in a plot. Phylogenetic beta diversity was measured by using the nearest taxon method. For each species in one sampling unit, the nearest phylogenetic neighbor in another sampling unit was found, and the mean of each sampling unit was calculated. The mean nearest phylogenetic distance between each pair of plots within each mountain was calculated. We measured the community phylogenetic structure using Net Relatedness Index (NRI). First, mean phylogenetic distance (MPD) between all possible pairs of taxa in each plot was calculated, and then MPD was compared to randomly generated null communities (MPDnull). Finally, standardization by the standard deviation of phylogenetic distances in the null communities was performed. The formula for NRI is given by: $$NRI = - 1 \times \frac{(MPDsample - MPDnull)}{sd(MPDnull)}$$ Positive values of NRI show that co-existing species in a plot are more closely related phylogenetically than expected by chance (phylogenetic clustering). Negative values of NRI show that species appearing in a plot are less phylogenetically related than expected by chance (phylogenetic overdispersion). In order to detect how species diversity changes along the elevation gradient, linear and quadratic regression analyses were performed. Species richness or phylogenetic diversity was used as the response variable, while altitude was used as the explanatory variable. Distance-decay was used to indicate the spatial patterns of beta diversity. We calculated Jaccard’s similarity index and phylogenetic beta diversity between each pair of plots within each mountain, and then regressed them on altitudinal divergence. In order to identify the main environmental factors affecting plant diversity, Pearson correlation analysis was performed to assess the relationship between alpha diversity (species richness and phylogenetic diversity) and environmental variables, with the altitude, slope, aspect, disturbance, and total basal area as explanatory factors. Phylogenetic diversity were calculated using Phylocom 4.2. Regression and correlation analyses were conducted using R Software, version 3.2.2. # Results ## Altitudinal patterns of species richness and phylogenetic diversity The total species richness on Mount Tai and Mount Lao has a consistent altitudinal pattern. With increasing elevation, the total number of species recorded in each plot decreased linearly. A unimodal pattern was detected for trees in both mountain systems. Trees richness increased and then decreased along altitudinal gradients. Maximum species richness occurred at about 1000 m on Mount Tai and 600m on Mount Lao. Shrubs have a monotonic decline from low to high elevation on Mount Tai. However, we did not found clear patterns for shrubs on Mount Lao. Herb richness exhibited a monotonically decreasing pattern on Mount Tai and inverted hump-shaped pattern on Mount Lao. There was no clear trend for total phylogenetic diversity along the elevation gradient in the two mountain systems. For tree species, we also did not found clear patterns along the elevation gradient. A significant negative relationship between the phylogenetic diversity of shrubs and elevation was detected for species on Mount Tai; however, no obvious relationship was detected for species on Mount Lao. For herb species, the monotonically decreasing pattern was only found for species on Mount Lao; no clear pattern was found for species on Mount Tai. Pearson correlation analysis showed that elevation, human disturbance, and total basal area play an important role in regulating species richness and phylogenetic diversity. Moreover, human disturbance and total basal area had greater impacts on shrubs and herbs than on trees. ## Altitudinal patterns of species turnover and phylogenetic distance The similarity in species composition between two plant communities decreased as altitudinal divergence increased. The regression slope between Jaccard similarity index and altitudinal divergence was not uniform among different plant groups. On Mount Tai, the regression slope of trees, shrubs, and herbs was 0.000237, 0.000195, and 0.000221, respectively. On Mount Lao, the regression slope of trees, shrubs, and herbs was 0.000388, 0.000231, and 0.000170, respectively. The phylogenetic distance consistently increased with increased altitudinal divergence in our study, although this trend was not significant for herbs on Mount Lao. On Mount Tai, the regression slope between phylogenetic distances and altitudinal divergence was 0.06975, 0.06938, and 0.05419 for trees, shrubs, and herbs, respectively. On Mount Lao, the regression slope between phylogenetic distances and altitudinal divergence was 0.12758, 0.08851, and 0.01016 for trees, shrubs and herbs, respectively. ## Variation in phylogenetic community structure along elevation gradients The phylogenetic structure of plant communities exhibited an inverted hump- shaped pattern on Mount Tai. However, no clear pattern was detected from low to high elevation on Mount Lao. # Discussion The patterns of total species richness in response to elevation gradientson Mount Tai and Mount Lao are consistent, with a monotonically decreasing pattern. However, the species richness for the plant groups presented have different altitudinal patterns. In this study, species richness of trees had unimodal altitudinal patterns across the elevation gradients in the two mountain systems. Climate is a conspicuous factor regulating species distribution and richness in many areas. McCain put forward a climatic model and proposed that hump-shaped diversity patterns might be attributed to optimal climate conditions (e.g. precipitation and temperature) at the middle elevations. High-altitude areas have lower temperature, which limits plant growth, while low-altitude areas lack adequate rainfall and cannot meet moisture requirements for plant growth. An optimal range of temperature and precipitation occurred in the middle elevations of our study sites. Favorable climatic conditions provide the most available primary productivity in the mountain ecosystem and support the survival of more species. In addition, the harsh climate conditions, e.g., strong wind, intense solar radiation, and low fertility soil, may prevent the appearance of some species at high altitude, while interspecific competition may eliminate other species and drive down biological diversity in mild climate conditions at low altitude. Altitudinal patterns in shrubs richness and herbs richness varied. Three patterns were found in our research: a monotonically decreasing pattern, an inverted hump-shaped pattern, and no significant pattern. This finding may be due to shrubs and herbs in the understory layer being more sensitive to changes in the microenvironment, such as disturbance and canopy cover. Schmitt et al. compared the plant species composition of two forest fragments in the mountainous highlands of the Great Rift Valley, and found that the two forests shared many of same woody plant species, but with obvious differences in herb layers. They proposed that herbs were affected not only by changes in moisture caused by the elevation gradient, but also by local changes in moisture caused by other factors such as insolation or small streams. Mount Tai and Mount Lao are important nature reserves and tourist scenic spots. Tourist activities are the major disturbance sources and low elevations suffer more anthropogenic impacts on Mount Tai and Mount Lao. The areas with moderate impacts may have greater environmental heterogeneity, which offers species greater opportunities for establishment and growth. Tourist activities such as camping, bushwalking and plant picking, are more likely to influence plant diversity in the understory of mountain forests. The trampling of tourists will reduce the cover or vigor of resident species, which can lead to an increase in resource availability in the understory. More available resources could provide more ecological niches for colonization by new species. In addition, forest canopy density could influence light intensity, moisture, and temperature in the understory. Forest communities with higher basal areas tend to have higher stand density and lack sufficient sunshine in the understory, which lead to lower species diversity. Therefore, the current diversity pattern of shrubs and herbs along the elevation gradient may be shaped by the interaction of altitude, anthropogenic disturbances, and canopy density. The similarity of species composition decreased with increasing geographical distances for all plant groups in this study. The rate of species turnover with altitudinal divergence is faster for trees than for shrubs and herbs in the two mountain systems. Species with greater dispersal were less susceptible to the limitations of dispersal distance; they could colonize more broadly and have lower species turnover rates. Shrubs and herbs in the understory layer were easily disturbed by human activities. Human activities provide new mechanisms for the spread of understory plants through seed dispersal, increasing the probability of understory plants expanding their current distribution. Comparing the rate of species turnover between pteridophytes and spermatophytes, Hong found that pteridophytes, whose propagules are spores dispersed by wind, have a lower distance-decay rate than spermatophytes, which are dispersed by seeds or fruits and are less vagile than pteridophytes. These results show that dispersal limitation plays an important role in structuring the spatial distribution of plants. It is generally believed that species diversity is a good surrogate for phylogenetic diversity. Bryant (2008) explored taxon richness and phylogenetic diversity of soil bacteria and plants along an elevation gradient, and found that the patterns of taxon richness reflected those of phylogenetic diversity for both soil bacteria and plants. However, in our study, the altitudinal patterns of species richness do not completely mirror phylogenetic diversity patterns. There is no monotonically decreasing trend for total phylogenetic diversity from low to high elevations (Figs). Besides contemporary climate, the controlling factors for phylogenetic diversity may include geological history. In the future, phylogenetic diversity pattern studies combining with environmental data and geological history should be conducted. In addition, conservation strategies based simply on taxon richness are not sufficient. Increasing plant communities’ phylogenetic diversity can enhance their ability to cope with climate change, through an increase in genetic diversity. The selection of conservation areas should consider preserving high taxon richness, while maximizing phylogenetic diversity to improve diversification potential in the future. Phylogenetic beta diversity and species beta diversity had a consistent altitudinal pattern (Figs). Phylogenetic beta diversity demonstrates how phylogenetic relatedness varies across environmental and geographical gradients. Phylogenetic relatedness declined as altitudinal divergence increased in our research. Phylogenetic beta diversity and the decay rate of phylogenetic similarity were both lower in herbs, which indicated herbaceous plants have close phylogenetic relationships among plant communities, have wider niche breadth, and are more widely distributed than woody plants. Phylogenetic structure can reveal the prevailing ecological processes influencing the observed patterns of species diversity. Based on the assumption of phylogenetic niche conservatism, environmental filtering leads to coexisting species having similar physiological tolerances, resulting in phylogenetic clustering. Interspecific competition leads to a community with distantly related species, resulting in phylogenetic overdispersion. Qian et al. (2014) analyzed phylogenetic structure of angiosperm assemblages along elevation gradients in Changbaishan, China, found that angiosperm assemblages tended to become more phylogenetically clustered at higher elevations. Graham et al. (2009), Pellissier et al. (2013) and Machac et al. (2011) studied phylogenetic structure of hummingbird communities, butterfly communities and ant communities, also found that communities were phylogenetically clustered at higher elevations. In contrast, tree communities were found to have a tendency to be more phylogenetically overdispersed as elevation increased in Malesian mountain forests. In our study, the phylogenetic structure of plant communities on Mount Lao did not display significant patterns along the elevation gradient, but plant communities on Mount Tai exhibited an inverted hump-shaped pattern. This result indicates that environmental filtering is the main driver of plant community assembly at high and low elevations on Mount Tai. High-altitude areas usually have lower temperature, while low-altitude areas usually lack adequate moisture. After filtering by harsh climate conditions, species surviving in high and low elevation areas are from few lineages that have evolved the ability to tolerance these habitats. Favorable climatic conditions usually occurred in the middle elevation areas and support the survival of more species. Therefore, inter- specific competition may be the main driver of plant community assembly in the middle elevations. Mount Lao is surrounded by the sea on two sides. The weather is tempered by the ocean, and this region has milder climate than Mount Tai. Therefore, the role of environmental filtering was not significant at both ends of the elevation gradient on Mount Lao. The emergence of the different phylogenetic structure patterns above is probably caused by differences in sampling scale, spatial distribution, taxonomic groups, or local climate conditions among different studies. Therefore, research into the phylogenetic structure patterns are insufficient and should be expanded to include different taxa and different regions in the future. In addition, a deeper understanding of the mechanisms driving current diversity patterns can be achieved, when sufficient environmental and functional traits data are available for these ecosystems. # Conclusions In this study, we identified the prevailing forces structuring the species distribution of plants across an elevation gradient on two mountains by using phylogenetic and traditional diversity analyses in combination. Phylogenetic structure analyses revealed that environmental filtering is the main driver of plant community assembly at high and low elevations on Mount Tai. However, the role of environmental filtering was not significant at both ends of the elevation gradient on Mount Lao where the climate is milder. At low elevation, anthropogenic disturbances contributed to the increase of plant diversity, especially for shrubs and herbs in understory layers, which are more sensitive to changes in microenvironment. Species beta diversity and phylogenetic beta diversity consistently increased as altitudinal divergence increased. However, the altitudinal patterns of species diversity did not completely mirror phylogenetic diversity patterns. Therefore, conservation strategies simply based on taxon richness are insufficient. The selection of protected areas should consider preserving high phylogenetic diversity to improve diversification potential in the face of global climate change. # Supporting Information We are grateful to Penglin Lu, Yihai Chen, Da Yin, and local experts for their help during our field investigation, and thank Yunfei Cai for assistance in species identifications. Thanks to the academic editor and reviewers for their helpful suggestions and professional editors at Editage for their English editing. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: WZ ND RW. Performed the experiments: WZ DH ND. Analyzed the data: WZ ND. Contributed reagents/materials/analysis tools: RW WZ DH. Wrote the paper: WZ JL ND.

# Introduction Cross-linking of high affinity IgE receptors (FcεRI) on mast cells is known to play an important role in allergic and hypersensitive diseases. Fukuoka et al showed that activation of human mast cells via FcεRI results in the secretion of tryptase, which generates sufficient amounts of C3a from C3 to cause mast cell degranulation. They proposed that C3a-induced mast cell activation may play an important role in mediating allergic diseases. Indeed, Shafer et al., recently demonstrated that IgE-mediated passive cutaneous anaphylaxis resulted in local increase in C3a levels and that subsequent activation of C3aR on mast cells contributed to allergic skin response. Not surprisingly, we have shown that C3a causes degranulation and chemokine generation in human mast cells and in transfected RBL-2H3 cells. However, the mechanisms involved in the regulation of C3aR signaling in mast cells remain poorly defined. A large number of multi-domain scaffolding proteins, including PDZ (<u>P</u>SD-95/<u>D</u>lg/<u>Z</u>o1) domain containing proteins, associate with GPCRs. There are three general classes of PDZ domains; class I domain, which recognize the carboxyl terminal motif S-T-X-Φ, (where “Φ” indicates hydrophobic amino acid and “X” indicates any amino acid), class II domain which recognize carboxyl terminal motif Φ-X-Φ and class III domains, which recognize D/E-X-Φ as their preferred carboxyl terminal motif. Na⁺/H⁺ exchanger regulatory factor-1 and 2 (NHERF1, EBP-50, *SLC9A3R1* and NHERF2, TKA-1, *SLC9A3R1*) are two class I PDZ domain adapter proteins that bind to several GPCRs to regulate their signaling. Both NHERF1 and NHERF2 have been shown to bind to parathyroid hormone receptor (PTHR) and β2-adrenergic receptor (β2-AR) via their PDZ binding domains to modulate receptor desensitization and internalization. NHERF proteins can also regulate GPCR signaling in receptor- independent manner via their association with downstream signaling proteins. For example, NHERF1 and NHERF2 associate with phospholipase C-β to regulate GPCR activation. Moreover, NHERF1 blocks PTH-induced ERK1/2 phosphorylation downstream of PKA via a receptor-independent but Akt-dependent pathway. C3aR is one of the few GPCRs expressed in human mast cells that possess a class I PDZ motif. We therefore postulated that NHERF1 or NHERF2 could associate with C3aR to modulate signaling in mast cells. We have recently utilized lentivirus short hairpin (sh)RNA to stably knockdown the expression of receptors, protein kinases and adapter molecules in human mast cell lines, HMC-1, LAD2 and primary CD34⁺-derived human mast cells to study receptor signaling. We used a similar approach for the current study and used both human mast cells lines (HMC-1 and LAD2) as well primary human CD34⁺-derived mast cells and show that although NHERF1 and NHERF2 do not interact with C3aR, they play a critical role in modulating C3a-induced degranulation and chemokine production. # Results ## NHERF1 and NHERF2 are expressed in human mast cells but they do not interact with C3aR NHERF proteins are highly expressed in epithelial cells of the gastrointestinal tract and airways. Since the goal of the present study was to determine the role of NHERFs on regulating C3aR signaling in mast cells, we wanted to first test if these proteins are expressed in human mast cells. We therefore performed reverse transcription PCR on cDNAs obtained from an immature (HMC-1), mature (LAD2), and primary (CD34⁺-derived) human mast cells using primers specific for human NHERF1 and NHERF2. shows that mRNAs for both NHERF1 and NHERF2 were detected in all three types of human mast cells. To determine if NHERFs physically associate with C3aR, we co-transfected HA-tagged C3aR and Flag-tagged NHERF1 or NHERF2 in HEK 293 cells and performed co-immunoprecipitation experiments. Surprisingly, NHERF1 or NHERF2 failed to interact with C3aR even when the cells were stimulated with C3a (lane 3, upper panel). It is quite possible that the lack of C3aR-NHERF interaction reflects the experimental condition used for the present study. It has previously been shown that NHERF1 and NHERF2 strongly associated with β2-adrenergic receptor (β2-AR). We therefore co-expressed HA-tagged β2-AR and Flag-tagged NHERF1 or NHERF2 and performed co- immunoprecipitation study under the exact same condition as was done for C3aR- NHERF interaction. As shown in (lane 1, upper panel), β2-AR co- immunoprecipitated with NHERF1 and NHERF2. This clearly demonstrates that despite the presence of class I PDZ motif at the carboxyl terminus of C3aR, it does not associate with NHERF1 or NHERF2. ## Silencing the expression of NHERF1 and NHERF2 in human mast cells NHERF proteins regulate signaling in a variety of cells via their interaction with receptors and downstream signaling proteins,. This suggests that these proteins could regulate C3aR signaling and mediator release in human mast cells. To test these possibilities, we used the Mission shRNA lentivirus system to stably knockdown the expression of NHERF1 and NHERF2 in human mast cells. Although HMC-1, LAD2 and primary CD34⁺ cells endogenously expressed C3aR, we used HMC-1 cells for these initial studies mainly because these cells can be cultured in large quantities and are more amenable to genetic manipulation. Cells were separately transduced with lentivirus containing 5 different shRNA constructs targeting different regions of NHERF1 and NHERF2. For control, we used a scrambled shRNA construct. After transduction and selection with puromycin, quantitative real-time PCR and Western blotting were performed to determine the extent of NHERF knockdown. As shown in, clone 2 (TRCN0000043736) for NHERF1 and clone 1 (TRCN0000043707) for NHERF2 showed \>90% decrease in mRNA. Consistent with this observation, Western blotting analyses showed complete absence of these proteins in knockdown cells. We therefore used these HMC-1 knockdown cells for subsequent experiments. ## NHERF1 and NHERF2 do not regulate C3aR desensitization, internalization, ERK1/2 phosphorylation, Akt phosphorylation or chemotaxis Agonist-induced phosphorylation of C3aR results in β-arrestin-2-mediated receptor desensitization and internalization. NHERF1 associates with PTHR, blocks β-arrestin-2 binding to inhibit desensitization and internalization. By contrast, NHERF1 promotes β-arrestin-2 interaction with the chemokine receptor CCR5 to enhance receptor internalization. These findings suggest that NHERF proteins could regulate C3aR desensitization and internalization in mast cells. We have recently used intracellular Ca²⁺ mobilization assay in HMC-1 cells to determine the roles of GRKs and β-arrestins on C3aR desensitization. We therefore used this assay to determine the roles of NHERF1 and NHERF2 on C3aR desensitization. As shown in, C3a caused a rapid increase in Ca²⁺ mobilization in shRNA control cells which decayed rapidly to baseline levels in ∼200 sec. The kinetics of this response was not altered in NHERF1 or NHERF2-silenced HMC-1 cells , indicating that these adapter proteins are not involved in C3aR desensitization. To test further the roles of NHERF1/NHEFR2 on C3aR desensitization, shRNA control or knockdown cells were stimulated with C3a, washed twice before re-exposure to the same concentration of C3a. shRNA control cells demonstrated at least a 80–90% reduction in intracellular Ca²⁺ mobilization indicating that the receptors were desensitized. A similar response was also observed in NHERF1 and NHERF2 knockdown cells. These studies clearly demonstrate that NHERFs do not regulate C3aR desensitization in HMC-1 cells. To investigate the role of NHERFs on agonist-induced C3aR internalization, shRNA control, NHERF1 and NHERF2 knockdown cells were exposed to buffer or C3a (100 nM) for different time intervals (20–300 sec) and loss of cell surface receptors was determined by flow cytometry. In shRNA control cells, C3a caused a robust internalization of its receptors with up to 70% of the receptors internalized within 5 min. However, consistent with our finding that NHERFs does not associate with C3aR, NHERF1 and NHERF2-silenced cells did not demonstrate a significant difference in receptor internalization as compared to control cells at any of the time points tested. These findings demonstrate that unlike the situation with PTHR and CCR5, NHERF proteins do not regulate C3aR desensitization or internalization. NHERF proteins have been shown to associate with a number of signaling proteins to regulate Akt and ERK phosphorylation independently of their interaction with PDZ motif containing GPCRs. Thus, NHERF1 associates with phosphatase and tensin homolog (PTEN), a major regulator of phosphatidylinositol 3-kinases (PI3K) pathway, to enhance Akt phosphorylation. By contrast, NHERF1 blocks PTH-induced ERK1/2 phosphorylation downstream of PKA via a receptor-independent but Akt- dependent pathway. NHERF1 and NHERF2 also associate with phospholipase C-β to modulate ERK1/2 phosphorylation. We therefore expected NHERF proteins to modulate C3a-induced ERK1/2 and or Akt phosphorylation in mast cells. We found that in shRNA control cells, C3a caused a transient ERK1/2 phosphorylation that peaked between 0–5 min and returned to basal levels thereafter. However, silencing NHERF1 or NHERF2 expression had no effect on C3a-induced ERK1/2 phosphorylation. Similar results were also observed for C3a-induced Akt phosphorylation in HMC-1 cells. In neutrophils, the chemokine receptor CCR2 forms a signaling complex with NHERF and phospholipase Cβ2 and that disruption of this complex leads to decreased neutrophil migration. Given that C3a is a potent mast cells chemoattractant, we examined the role of NHERFs in regulating C3a-induced mast cell chemotaxis. C3a (10 nM) caused chemotaxis of HMC-1 cells in shRNA cells and there was no significant difference in this response in NHERF1 or NHERF2-silenced cells. ## NHERF1 and NHERF2 promote C3a-induced mast cell degranulation, NF-κB activation and chemokine CCL4 generation Our next goal was to determine if NHERF1 and NHERF2 regulate C3a-induced mast cell degranulation. HMC-1 is an immature mast cell line that shows little or no degranulation response to C3a. LAD2 is a mature mast cell line that expresses C3aR and responds to C3a for degranulation. We therefore silenced the expression of NHERF1 and NHERF2 in LAD2 cells. As in HMC-1 cells, lentiviral shRNA induced ∼90% knockdown of the NHERF1 and NHERF2 in LAD2 mast cells. Surprisingly, knockdown of NHERF1 and NHERF2 resulted in a significantly decreased degranulation response to varying doses of C3a. This effect was specific for C3a as silencing the expression of NHERFs had little or no effect on antigen-IgE- mediated response. We have recently identified a novel GPCR, MrgX2 that is expressed in human mast cells. Cortistatin-14 (CST), a ligand for MrgX2, induced substantial degranulation in shRNA control LAD2 cells and this response was significantly inhibited in NHERF1 and NHERF2 knockdown cells. These data suggest that C3aR and MrgX2 share a common pathway for inducing degranulation, which is different from IgE-mediated response. We have previously shown that both primary cultures of human CD34⁺-derived mast cells and the LAD2 mast cell line express C3aR and respond to C3a for degranulation. To confirm the biological relevance of studies with LAD2 mast cells, we silenced the expression of NHERF1 and NHERF2 in CD34⁺-derived primary human mast cells. Unlike the situation in LAD2 cells, we were able to silence the expression of NHERF1 and NHERF2 in CD34⁺-derived primary human mast cells by ∼50%. More importantly, even this knockdown efficiency was sufficient to cause a significant reduction in C3a-induced degranulation. We next sought to determine the roles of these PDZ domain proteins on C3a-induced NF-κB luciferase activity in mast cells. Although HMC-1 cells, LAD2 mast cells and CD34⁺-derived primary mast cells express C3aR and respond to C3a for Ca²⁺ mobilization, C3a does not induce NF-κB activation in HMC-1 cells unless the cells are transfected with exogenous receptor. Therefore, we transfected C3aR along with NF-κB luciferase constructs in shRNA control and NHERF-silenced cells and performed NF-κB luciferase and chemokine release assays. C3a induced NF-κB luciferase activity in shRNA control HMC-1 cells. Interestingly, there was a significant reduction in NF-κB luciferase activity in NHERF knockdown cells when compared to shRNA control cells. Given that NF-κB plays an important role in the generation of proinflammatory cytokines, we tested the effects of NHERF1 and NHERF2 knockdown on C3a-induced chemokine CCL4 production. We found that C3a-induced CCL4 production was significantly decreased in NHERF knockdown cells. # Discussion The roles of NHERF proteins in regulating receptor desensitization, internalization and signaling have been studied extensively for GPCRs such as β2-AR, PTHR and opioid receptors. Most of these studies have utilized NHERF-null renal tubule cells or transfected rat osteosarcoma ROS 17/2.8 (ROS), opossum kidney (OK), PS120, and HEK 293 cells,. For the present study, we utilized lentivirus shRNA to stably knockdown the expression of NHERF1 and NHERF2 in human mast cell lines; HMC-1 cells, LAD2 and primary CD34⁺-derived human mast cells that endogenously express C3aR. Using this approach, we have uncovered novel roles of NHERF1 and NHERF2 on C3a-induced degranulation, NF-κB activation and chemokine generation in the absence receptor desensitization, internalization, ERK1/2 and Akt phosphorylation. NHERF1 and NHERF2 are class I PDZ domain proteins that associate with the carboxyl terminal of several GPCRs. Class I PDZ proteins interact with ligands terminating in the sequence \[S/T\]- X-Φ, where X is promiscuous and Φ is a hydrophobic residue, generally Leu but also Ile, Val, or Met. Our results indicate that although C3aR possess a class I PDZ protein binding motif in its carboxyl terminus (S-T-T-V), NHERF proteins do not interact with the receptor. It has been proposed that in addition to the class I PDZ motif, NHERFs seems to additionally favor \[D/E\] at position -3. Hall *et al.*, have reported that the sequence D-T-R-L is favored for binding by NHERFs. However, NHERFs also bind to N-K-P-V and S-V-G-L sequences in opioid and CCR5 receptors. Thus, the consensus binding motif for NHERFs seems to vary and may require additional binding sites upstream or downstream of the PDZ binding site. Accordingly, it has been proposed that selectivity for PDZ protein-receptor interaction may be influenced by upstream amino acids at positions 5 and/or 7. Thus, the inability of NHERFs to associate with the C3aR probably reflects the lack of these upstream structural determinants required for NHERF1/2 binding. In addition to GPCRs, NHERF proteins associate with a variety of downstream signaling proteins to regulate receptor internalization, ERK1/2 phosphorylation, Akt phosphorylation. However, the findings in the present study that NHERF1 and NHERF2 have no effect on these responses but promote mast cell degranulation and chemokine production clearly demonstrate that these proteins have distinct functions in human mast cells. The mechanism by which NHERF1 promotes mast cell degranulation remains unknown. NHERF1 contains 31 Ser and 9 Thr residues, which make up roughly 12% of the molecule. Hall *et al.*, showed that GRK6 interacts with PDZ motif of NHERF1 to promote its phosphorylation at Ser289. In our recently published work, we have shown that, as for NHERF1, silencing the expression of GRK6 has no effect on C3a-induced Ca²⁺ mobilization but blocks degranulation. Given that GRK6 associates with and phosphorylates NHERF1 (at Ser289), it is possible that this modification provides a mechanism for C3a-induced mast cell degranulation. Interestingly, PTH causes protein-kinase C (PKC)-mediated phosphorylation of NHERF1 at Ser77 (PDZ1 domain) and this modification regulates PTH action,. PKC also phosphorylates NHERF1 at Ser162 (PDZ2 domain) and Ser337/338 (C-terminus). It is noteworthy that C3a-induced mast cell degranulation requires PKC activation and PKCα binds to NHERF1 through PDZ domain. It is therefore possible that C3aR, PKCα and NHERF1 form a signaling complex to promote phosphorylation of downstream signaling proteins to support degranulation. An interesting finding of the present study was that both NHERF1 and NHERF2 are involved in promoting mast cells degranulation. It is noteworthy that NHERF1 dimerizes with NHERF2. This suggests that formation of a complex between NHERF1 and NHERF2 may promote C3a-induced mast cell degranulation and that silencing the expression of either protein leads to inhibition of the response. These possibilities will be the subject of future investigations. Another interesting finding of the present study was that knockdown of NHERF1 and NHERF2 decreased C3a-induced NF-κB activation and chemokine production in human mast cells. However, the molecular mechanism via which NHERF proteins promote chemokine gene expression is unknown. Na⁺/H⁺ exchanger 3 (NHE3) is a member of a group of integral transmembrane proteins that is regulated by NHERF. It has been proposed that in monocytes and macrophages, NHEs are rapidly activated by inflammatory stimulus such as IL-1, TNF-α and LPS which leads to the production of a variety of cytokines. In conjunction with this hypothesis, Németh *et al.*, have demonstrated that inhibition of NHEs suppressed IL-12, MIP-1α and MIP-2 production by LPS- activated macrophages. In a subsequent study they showed that this inhibition of NHEs suppressed both the inhibitory IκB degradation and NFκB-DNA binding activity leading to a decreased activation of NF-κB and a substantial inhibition in cytokine production. Since NHERFs regulate NHE3 it is tempting to propose that C3a-induced NF-κB activation and cytokine production occurs via a NHE3-dependent pathway and silencing of NHERFs abrogates this pathway. Another possibility is that lack of NHERFs disrupts the cytoskeletal association of actin with the ERM proteins and NHE3 and this change in the actin microfilament organization mediates reduced NF-κB activation and cytokine generation. Whether this or other mechanism(s) participate on the effect of NHERFs on C3a-induced responses in mast cells remains to be determined. In summary, the current study demonstrates that NHERF1 and NHERF2 promote degranulation and chemokine production, the two critical responses of human mast cells following activation via C3a. Thus, we provide the first demonstration that the PDZ proteins, NHERF1 and NHERF2 may act as novel regulators of C3aR signaling in human mast cells. Since C3a has been reported to potentiate IgE- mediated allergic responses *in vivo*, targeting these adapter proteins in human mast cells may serve as a valuable therapeutic for allergic disorders. # Materials and Methods ## Materials Mission shRNA bacterial glycerol stocks for NHERF1 and NHERF2 were purchased from Sigma Life Sciences (St. Louis, MO). Indo-1 AM was from Molecular Probes (Eugene, OR). All tissue culture reagents were purchased from Invitrogen (Gaithersburg, MD). Anti-human C3aR was obtained from Santa Cruz Biotechnology (Santa Cruz, CA), PE-labeled donkey anti-mouse IgG was purchased from eBioscience (San Diego, CA). All recombinant human cytokines were purchased from Peprotech (Rocky Hill, NJ). Rabbit anti-ERK1/2, anti-phospho-ERK1/2 and anti- phospho Akt antibodies were purchased from Cell Signaling (Beverly, MA). SuperSignal® West Femto Maximum Sensitivity Substrate and HRP labeled Goat anti- rabbit IgG were from Thermo Scientific (Rockford, IL). Purified C3a was obtained from Complement Tech, Inc. (Tyler, TX). CCL4 ELISA kit was purchased from R&D Systems (Minneapolis, MN). Amaxa cell transfection kits and reagents were purchased from Lonza (Gaithersburg, MD). Anti Flag antibody (M2) and anti HA (HA-7) agarose beads were purchased from Sigma-Aldrich (St. Louis, MO). Cortistatin-14 (CST) was obtained from American Peptide (Vista, California). ## Mast cell culture HMC-1 cells were cultured in Iscove's modified Dulbecco's medium (IMDM) supplemented with 10% FCS, L-glutamine (2 mM), penicillin (100 IU/mL) and streptomycin (100 µg/mL). LAD2 cells were maintained in complete StemPro-34 medium supplemented with 100 ng/mL rhSCF. To obtain primary mast cells, human CD34⁺ progenitors (Hutchinson Cancer Research Center, Seattle, WA) were cultured in StemPro-34 medium supplemented with L-glutamine (2 mM), penicillin (100 IU/ml), streptomycin (100 µg/ml), rhSCF (100 ng/ml), rhIL-6 (100 ng/ml) and rhIL-3 (30 ng/ml) (first week only). Hemidepletions were performed weekly with media containing rhSCF (100 ng/ml) and rhIL-6 (100 ng/ml) (32, 56). Cells were used for experiments after 7–10 weeks in culture. ## Co-immunoprecipitation HEK293 cells transiently transfected with 4 µg of cDNA encoding either the HA- tagged C3aR or HA-tagged β2-AR along with 2 µg of Flag-tagged NHERF were lysed, and clarified by centrifugation. HA-C3aR was immunoprecipitated by addition of anti-HA agarose beads (Sigma) with agitation at 4°C for 4 h. Anti-HA immunocomplexes were washed twice in RIPA lysis buffer and once with ice-cold PBS. Whole-cell lysates and HA affinity matrix immunocomplexes were transferred to nitrocellulose membranes for immunodetection of Flag-NHERF and the C3aR. The C3aR was detected with a 1∶1,000 dilution of mouse anti-HA (12CA5) antibody (Boehringer Mannheim) and Flag-tagged NHERF was detected with a 1∶1,000 dilution of mouse anti-Flag monoclonal antibody (M2). The secondary antibody used in both the cases was a 1∶4,000 dilution of a HRP-conjugated anti-mouse polyclonal antibody (Santa Cruz Biotechnology). ## Lentivirus and stable transduction of shRNAs in mast cells NHERF1 and NHERF2 targeted shRNAs in lentiviral plasmids were purchased from Sigma-Aldrich. The clone that gave the highest knockdown efficiency (TRCN0000043736 for NHERF1 and TRCN0000043707 for NHERFF2) was used. A scrambled control non-target vector (SHC002) which does not bind to any known human mRNAs was also purchased from Sigma-Aldrich. Lentivirus generation was performed according to the manufacture's manual. Cell transduction was conducted by mixing 1.5 ml of viral supernatant with 3.5 ml of HMC-1 or LAD2 (5×10⁶ cells) or CD34⁺-mast cells (3×10⁶ cells). Eight hours post-infection, medium was changed to virus-free complete medium, and antibiotic (puromycin, 2 µg/ml, Sigma) selection was initiated 16 h later. Cells were analyzed for NHERF1 or NHERF2 knockdown and used for subsequent assays 4 days following initiation of puromycin selection. ## Reverse transcription PCR and quantitative PCR Total RNA was extracted from mast cells using TRIZOL, treated with DNase I and reverse transcribed to cDNA using first strand cDNA synthesis kit (GE). The primers used for human NHERF1 were: forward 5′ TACAGAAGGAGAACAGTCGTGAAGC and reverse 5′ GCCAGGGAGATGTTGAAGTCTAGG. The primers used for human NHERF2 were: forward 5′ CCGACAAGGACACTGAGGATGG and reverse 5′ CGCTTGTTGACTCGCATGGC. For quantitative PCR, gene expression was analyzed using real time PCR with Taqman® Fast Universal PCR Master Mix on a Taqman 7500 Fast Real-Time PCR System (Applied Biosystems, Foster City, CA). Taqman probes for hGAPDH and hNHERFs were used for real time PCR to analyze the knockdown efficiency. The amplification conditions were as follows: initial denaturation at 95°C for 20 sec, followed by 40 cycles of amplification: 95°C for 3 sec, 60°C for 30 sec. Analysis was performed according to ΔΔ-Ct method. ## C3a Receptor desensitization Receptor desensitization assay based on Ca²⁺ mobilization was determined as described previously. Briefly, 1×10⁶ HMC-1 cells were washed twice with buffer (119 mM NaCl, 5 mM KCl, 25 mM HEPES, 5.6 mM Glucose, 0.4 mM MgCl₂, 1 mM CaCl₂) containing 1 mg/ml BSA and incubated with 1 µM of Indo-1 for 30 min in dark. Cells were then washed and resuspended in 1.5 ml of the same buffer and time course of Ca²⁺ mobilization (0–5 min) was determined using Hitachi F-2500 Fluoro spectrophotometer (San Jose, CA) with an excitation wavelength of 355 nm and an emission wavelength of 410 nm. Cells were removed from the cuvette, washed twice and Ca²⁺ mobilization to a subsequent exposure of C3a (100 nM) was determined. ## Receptor Internalization shRNA control and NHERF knockdown HMC-1 cells (2.5×10⁵) were stimulated with or without C3a (100 nM) at 37°C. Cells were washed, resuspended in 50 µl of ice-cold FACS buffer (PBS containing 2% FBS), and stained with C3aR antibody or isotype control (2 µl) on ice for 1 h. After washing twice with ice- cold FACS buffer, cells were labeled with Phycoerythrin (PE)-labeled donkey anti-mouse (1.5 µl) secondary antibody and incubated on ice for 1 h. Cells were washed twice and fixed in 300 µl of 2% formaldehyde. The samples were acquired and analyzed on a BD LSR II flow cytometer (BD Biosciences). ## Western blotting Control and NHERF knockdown HMC-1 cells or LAD2 cells (1×10⁶/ml) or CD34⁺-derived mast cells (0.5×10⁶/ml) were washed twice in ice-cold PBS and lysed with RIPA buffer (150 mM NaCl, 1.0% NP-40, 0.5% Sodium- deoxycholate, 0.10% SDS, 50 mM Tris \[pH 8.0\], 5 mM EDTA, 10 mM NaF, 10 mM Na- pyrophosphate and protease inhibitor cocktail). Protein bands were separated on 10% SDS PAGE gels and immunoblotted onto nitrocellulose membranes and probed with anti-NHERF1 (Santa Cruz Biotechnolgy, Santa Cruz, CA) or anti-NHERF2 (Sigma-Aldrich) antibodies. Following incubation with the secondary anti-rabbit- HRP antibody, membranes were developed using the SuperSignal® West Femto Maximum Sensitivity chemiluminiscent substrate. For ERK/Akt phosphorylation, cells were serum starved overnight, washed twice, resuspended in serum free IMDM medium at a concentration of 1×10⁶/ml and stimulated C3a (100 nM) for different time points. Three-fold volume of ice- cold PBS containing 1 mM sodium orthovanadate was added to stop the reaction. Total cell lysate was prepared with RIPA buffer and subsequently analyzed by Western blotting using rabbit polyclonal antibodies for phospho-p44/42 MAPK (pERK1/2) phospho-Akt (pAkt) and p44/42 MAPK (ERK1/2). ## Chemotaxis assay C3a (10 nM, 30 µl) or buffer was added to the lower wells of a 96-well chemotaxis chamber (8 µm pore size; NeuroProbe, Gaithersburg, MD). HMC-1 cells (0.5×10⁶) were added on top of the membrane of 96-well chemotaxis chamber. After 3 h incubation at 37°C and 5% CO₂, chemotaxed cells were collected from the lower chambers. Triplicate wells were pooled and the cells were resuspended in thirty microliters of complete IMDM. The chemotaxed cells were counted under a hemocytometer slide and the results are expressed as absolute number of cells that had chemotaxed. ## Degranulation Assay LAD2 cells or CD34⁺-derived human mast cells (1×10⁴) were seeded into 96-well plates in a total volume of 50 µl of buffer containing 1 mg/ml BSA and exposed to different concentrations of C3a C3a or CST (100 nM). For some experiments, LAD2 cells were pretreated with NP-specific human IgE (AbD serotec, 1 µg/ml) for 16 h prior to stimulation with NP-BSA (Biosearch Technologies, 100 ng/mL). For total β-hexosaminidase release, control cells were lysed in 50 µl of 0.1% Triton X-100. Aliquots (20 µl) of supernatants or cell lysates were incubated with 20 µl of 1 mM p-nitrophenyl-N-acetyl-β-D-glucosamine for 1.5 h at 37°C. The reaction was stopped by adding 250 µl of a 0.1 M Na₂CO₃/0.1 M NaHCO₃ buffer and absorbance measured at 405 nm. ## NF-κB luciferase reporter activity and CCL4 chemokine release assay shRNA control and NHERF knockdown HMC-1 cells (3×10⁶) were co- transfected with NF-κB luciferase reporter gene construct (pNF-kB-LUC and p-Renilla Stratagene, Santaclara, CA) (in a 10∶1 ratio) along with HA-tagged C3aR using Amaxa nucleofector device and Amaxa kit V as per manufacturer's protocol. Six hour post-transfection, medium was replaced with IMDM containing 10% FBS. Following 18 h of incubation, cells were stimulated with C3a (100 nM for 6 h). Cells were then harvested, washed in ice-cold PBS and finally lysed in Promega passive lysis buffer (Dual Luciferase assay kit; Promega, Madison, WI). NF-κB luciferase activity was measured using Turner biosystem 20/20 Luminometer (Promega, Madison, WI). Results expressed have been normalized to Renilla luciferase activity. Chemokine release assay was performed as previously described. CCL4 chemokine levels in the supernatants were quantified by sandwich ELISA according to the manufacturer's protocol. ## Data analysis The results are expressed as ± S.E.M for the values obtained from at least three independent experiments. GraphPad Prism software (Graph Pad, Version 5.0 San Diego, CA) was used to analyze data for statistical significance. The statistical significance was determined by unpaired two-tailed *t* test, and two-way ANOVA with Bonferroni's post test. A p value \<0.05 was deemed significant. We are grateful to Dr. Joseph Butterfield (Mayo Clinic, Rochester, MN) for supplying us with HMC-1 cells. We also thank Drs. Arnold Kirshenbaum and Dean Metcalfe (NIAID/NIH) for providing LAD2 mast cells and the FACS core facility of the School of Dental Medicine, University of Pennsylvania for flow cytometry acquisition and analysis. We thank Dr. Randy Hall (Emory University School of Medicine) for providing cDNAs for NHERF1 and NHERF2. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HS HA. Performed the experiments: HS KG. Analyzed the data: HS KG. Contributed reagents/materials/analysis tools: HS KG. Wrote the paper: HS HA.

# Introduction Beta-blocker (BB) treatment decreases the mortality rate in patients with acute coronary syndrome (ACS). Therefore, its use is currently recommended as a class I-A indication in clinical practice guidelines. Several strategies have been developed to encourage the prescription of BBs early after the diagnosis of ACS. However, it has been reported that nearly 22% of eligible patients do not receive the medication. All patients with ACS are considered eligible regardless of the concomitant administration of fibrinolytics or primary angioplasty; being in Killip class I or II; and the absence of bradycardia (FC \< 60 bpm), severe hypertension, and advanced atrioventricular block (AVB). However, the nonprescription of BBs might be due to intolerance to the medication rather than to nonadherence to evidence-based therapy. The literature presents little information about the frequency that patients hospitalized with ACS and in Killip I or II classification do not receive BBs because of their intolerance to the drug, and how this effects mortality. Moreover, little is known about the clinical predictors of nontolerance to BBs during hospitalization for ACS. The present study aims to evaluate the characteristics associated with nontolerance to BBs in patients with ACS, and to identify its impact in the 6-month all-cause mortality. # Methods The procedures were approved by The Research Ethics Committee of the Botucatu School of Medicine (FMB, UNESP; OF213/2004-CEP) and were conducted in accordance with the Declaration of Helsinki. All participants had ability to consent and signed an informed consent form. This is a prospective and longitudinal observational study comprising 377 consecutive patients older than 18 years and admitted to the hospital with unstable angina (UA), non-ST-segment elevation myocardial infarction (NSTEMI), or ST-segment elevation myocardial infarction (STEMI) \[–\] diagnosed up to 48 h after the onset of symptoms. The patients were admitted to the intensive care unit of the Emergency Hospital and in the coronary unit of the Botucatu School of Medicine University Hospital from March 1, 2003 to December 31, 2006. Exclusion criteria were Killip class III or IV at admission, heart rate \<60 bpm, systolic blood pressure (SBP) \<100mmHg, and PR interval \>0.24s, second or third AVB and history of asthma or severe obstructive pulmonary disease. After the ACS diagnosis, the patients underwent clinical and laboratory evaluations, according to the standardized protocols implemented in the intensive care unit at Botucatu Medical School for patients with ACS. All patients were administered with acetylsalicylic acid as well as submitted to mechanical or chemical reperfusion when indicated. Patients who fulfilled the inclusion criteria were treated with BBs. The medication used in all cases was metoprolol, following the international guidelines. Briefely, Intravenous metoprolol tartrate was given in 5 mg increments by slow intravenous administration (5 mg over one to two minutes), repeated every five minutes for a total initial dose of 15 mg (three doses). Patients who tolerate this regimen then received oral therapy beginning 15 min after the last intravenous dose (25 to 50 mg every six hours for 48 hours of metoprolol tartrate) followed by a maintenance dose of 100 mg twice daily. Patients who do not receive a beta blocker during the first 24 hours because of early contraindications were reevaluated for beta blocker candidacy for subsequent therapy. Oral metoprolol tartrate 25 to 50 mg every 6 to 12 hours, titrating upward as needed. The doses were tittered up to the recommended full-dose. They also received angiotensin- converting enzyme inhibitors (ACEIs), unless contraindicated. The contraindications to ACEIs include arterial hypotension and severe renal dysfunction (serum creatinine level \>2.5 mg/dL in men or \>2.0 mg/dL in women). After the onset of treatment, patients were defined as nontolerant to BBs if they developed bradycardia, hypotension, AVB, or severe and symptomatic ventricular dysfunction to any dose of BB. ## Data collection Gender, age, ST elevation, comorbidities (arterial hypertension, diabetes mellitus, dyslipidemia, overweight/obesity, renal failure), tobacco use, previous ACS, and pharmacological treatment at admission were recorded, as well as blood pressure, heart rate, Killip classification, and blood glucose level at admission. The culprit artery was determined with coronary angiography during the in-hospital period. ## Outcome measures In the first analysis, the primary outcome was the intolerance to BB during the hospitalization period and the 6-month mortality thereafter. ## Statistical analysis Data management and analysis were performed using SYSTAT 12.0 (SYSTAT Software Inc. 2007). Summary data are expressed as either mean (SD) or proportions. Univariate regression analysis was performed in order to assess which variables were independently associated with either BB treatment intolerance or 6-month all-cause mortality. Those variables that showed a statistically significant association with each outcome were introduced into a multivariate model. The analysis of BB intolerance included the variables age, gender, SBP, Killip class, serum creatinine, and use of ACEIs and statins. For the 6-month mortality analysis, age, history of ACS, intolerance to BBs, and the presence of atherosclerotic plaque in the left anterior descending (LAD) coronary artery were considered. The level of significance was set at 5%. # Results During the follow-up period, 7 patients lost contact with the service and were excluded from the analysis. Therefore, 370 patients (224 men and 146 women; p\< 0.001) were consecutively included. The demographic and clinical characteristics of the studied population are presented in. Male patients were younger than female ones (59 ± 12 years and ± 12.5 years, respectively; p \< 0.001). Acute events diagnosed at hospital admission were UA in 33.7% (n= 124), NSTEMI in 32.6% (n= 122), and STEMI in 33.7% (n= 124). Statistically significant differences were found between male and female patients in hypertension (p\< 0.001), diabetes mellitus (p 0.001), smoking (p\< 0.001), and dyslipidemia (p\< 0.001). At the initial evaluation, 85.7% of men (n= 192) were in Killip I class, and 14.3% received the classification Killip II. Among the females, 72.6% (n= 160) were in Killip I and 27.4% (n= 40) were in Killip II (p= 0.002). Plasma creatinine was higher among males (1.39 ± 1.34 g/dL vs. 1.14 ± 0.76 g/dL; p= 0.024). Of the whole population, 76% received BBs (48% males \[n= 172\] and 28% females \[n= 99\]). Among the female patients, 30% did not tolerate any dose of BBs, whereas only 20% among the males showed intolerance (p = 0.042). Coronary angiography was performed in 348 patients. An atherosclerotic plaque causing an obstruction \>60% in the LAD coronary artery, right coronary artery, and/or circumflex coronary artery was observed in 243 (69.8%), 180 (51.7%), and 160 (46%) patients, respectively. The univariate logistic regression for the outcome “BB intolerance” indicated that the Killip II classification more than doubled the risk for non-tolerance to this therapy as compared to Killip I class. In addition, for each unit increase in plasma creatinine, the risk of BB intolerance increased by 30%. The nonuse of ACEIs and statins increased the risk of intolerance to BB use.    The results of the multivariate logistic regression analysis for the outcome “BB intolerance” including age, gender, systolic blood pressure serum, creatinine, non-use of ACEI, and statin are presented in. Killip class II remained associated with intolerance to BBs. The effects of the nonuse of ACEIs and statins persisted, as in the univariate regression analysis.  Death at 6 months was associated with age, medical history of arterial hypertension, Killip II at admission, increased serum creatinine, and the presence of a \>60% lesion in the LAD coronary artery. In addition, in-hospital intolerance to BB increased the risk of death during follow up by 4.5 times.   Multivariate logistic regression indicated that female gender, medical history of arterial hypertension, Killip II class, increased serum creatinine, atherosclerotic plaque in the LAD coronary artery, and nonuse of BBs and/or ACEIs were independently associated with death at 6 months. In the adjusted model including clinical variables and angiographic findings, we have found that intolerance to BBs and anterior descending coronary artery disease were the most important predictors of all-cause mortality. # Discussion The present study aimed to identify variables associated with intolerance to BB use in hospitalized ACS patients. We here show that some clinical parameters easily obtained in the emergency department on admission can predict such intolerance. In addition, intolerance to treatment with BB increases the risk of death at 6 months in these patients. Our study found that 23% of the ACS patients eligible to did not tolerate BB treatment. This result is similar to those described by other researchers worldwide. However, the other studies suggested that patients who were not prescribed such drugs were in fact potential candidates for the treatment. For example, a Korean study has recently reported that the use of statins was limited to the equivalent amount of such medications that can be reimbursed from the local health system; in other words, in addition to the lack of prescriptions, the use of these drugs seems to be limited by economic factors. The results of the present study suggest that the nonprescription of the drugs is due to the intolerance to their use rather than to the nonobservance of guidelines. The patients who did not tolerate BB use were those in Killip II class, with intolerance to ACEI, and were not under statin treatment. The univariate regression analysis showed that beyond the Killip II classification, impaired renal function is also a predictor for the nonuse of BBs. Several studies have analyzed the association between impaired renal function and the treatment of ACS. Most of those studies have demonstrated that patients with ACS who present with some degree of impaired renal function do not receive the optimal recommended therapeutic approach. There may be several causes for that beyond the hemodynamic conditions at the emergency department, such as the presence of comorbidities that can limit the use of appropriate drugs. In our population, we did not identify factors that would explain the intolerance to the recommended medication for ACS in patients with impaired renal function. The nonuse of ACEIs and statins was associated with BB intolerance. Concerning ACEI, arterial hypotension would explain this association. However, the association of the lack of statin prescription with the nonuse of BB was intriguing. Patients who were not treated with statins had double the risk of nontolerance to BB. As our results do not allow us to explain such association, we discuss some considerations about the subject as follows. The use of statins is recommended for patients with ACS and also as an adjuvant drug in the primary prevention of the disorder. The beneficial effects of statins for patients with stable or unstable ACS have been attributed to their rheological effects and antithrombotic activity while stabilizing atherosclerotic plaques. Improvement of endothelial function, reduction of circulating C-reactive protein, and reduction of thrombogenicity have also been reported as beneficial effects of statin use. Furthermore, some reports have suggested an enhanced benefit of statins when used in association with BBs, mainly before a cardiovascular or noncardiovascular surgery. A gender-related difference was noticed in the treatment. This is in agreement with the reports of others stating that elderly female patients receive less aggressive therapies when hospitalized with ACS, maybe because aggressive prescriptions are discouraged in this specific patient group. It is known that immediate reperfusion therapy and the introduction of adjuvant medications such as BBs, ACEIs, and statins are highly relevant for the better management of patients with ACS. In addition, the more aggressive use of BBs, ACEIs, and statins for secondary prevention contributes to the reduction of morbidity and mortality in ACS patients. ACS patients presenting an unstable hemodynamic condition at admission or a history of previous ACS events, kidney failure, or overt heart failure received less adjuvant pharmacological approach, in accordance with other reports. However, our results suggest that the lack of prescription of the optimal treatment is due to the patient’s medical condition rather than to the nonobservation of the guidelines. The overall 6-month mortality was associated with age, arterial hypertension, Killip II, elevated serum creatinine, atherosclerotic plaque in the LAD coronary, and the nonuse of BB. Notably, the lack of BB use during hospitalization more than quadrupled the risk of death, regardless of the ACS type. Altogether, the results of the present study suggest that the parameters associated with BB intolerance during hospitalization should be aggressively managed in order to allow the use of the drug. Therefore, we emphasize the need for an individualized therapeutic strategy that would allow enough hemodynamic stability to make BB prescription possible in ineligible patients. We thank the Hélio Rubens de Carvalho Nunes by statistical analysis of the study data*.* [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: ALC. Analyzed the data: LMDS BBM. Wrote the manuscript: LMDS BBM. Principal investigator: LMDS. Research associate: ALG ALAF ABF EF. Researcher reviewer: BBM. Author: LMDS. Co- authors: ALG ALAF ABF EF. Author reviewer: BBM.

# Introduction Bacillary dysentery, caused by Shigella bacteria, is a bacterial infection of intestines which may result in severe diarrhea. The infection is spread from person to person via oral-feces, food or drinking water. Epidemics are frequently occurred in overcrowded populations with poor sanitation and most cases occur in summer and autumn. Bacillary dysentery is still a public health problem in China, especially in Guangxi Province. The incidence of bacillary dysentery has been reported to exhibit distinct regional differences in China, with the highest incidence of 132.37 per 100,000 in Beijing, and lowest incidence being reported in Fujian and Shanghai Provinces in 2011. In different regions, the main influence factors of bacterial dysentery were different. In this paper, the socio-economic factors of bacillary dysentery were analyzed based on regional differences in Guangxi Province. Though the morbidity and the mortality of bacillary dysentery have decreased considerably in Guangxi since the 1990s, a considerable burden still exists, particularly among the children and the older people with low economic status. Chinese Government has developed a strategic plan for the national surveillance of bacillary dysentery, but the risk of bacterial dysentery has increased in recent years, and it has attracted wide attention of scholars in China and abroad. In the past decade, the relationship between the bacillary dysentery and the meteorological factors has been reported in many studies,, The seasonality of bacillary dysentery incidence indicated that meteorological factors might play an important role in its epidemiology. But no significant progress has been made in terms of the model investigating the socio-economic factors of bacillary dysentery. In Guangxi Province, the socio-economic factors, such as public health and the level of per capita income, played an important role in bacillary dysentery's epidemiology, control and prevention. For time series data, multiple regression model is a good choice. While for regional differences, the spatial correlation analysis has its unique advantage. Spatial correlation have been widely used in health studies, but there is few study about spatial variations of social-economic impacts on bacillary dysentery in Guangxi. In this paper, global spatial autocorrelation was employed, which indicated whether bacillary dysentery and socio-economic factors were randomly located over the study area, or followed some spatial pattern, thereby indicating some underlying process. Through the spatial correlation analysis on bacillary dysentery and its socio-economic factors in Guangxi Province, the main risk factors of bacillary dysentery can be found, and the corresponding control measures were also put forward in this paper. # Data and Methodology ## Data sources The epidemiological case data, population data and socioeconomic data were obtained from Guangxi Province, China. Guangxi Province is in the Pearl River basin of southern China, from 20° 54′ to 26° 23′ N and 104° 29′ to 112° 03′ E. Guangxi has a subtropical climate, and usually has heavy rain in the summer. The summers are generally long and hot. Average annual temperature is 17–23°C, and average annual precipitation is 1250–1750 mm. Guangxi has extremely complex geological features, with significant differences between regions. Bacillary dysentery is a legally mandated notifiable disease in China. All clinical and hospital doctors are required to report cases of bacillary dysentery to local Centers for Diseases Prevention and Control. In this study, both clinical and laboratory diagnosed cases were collected. Bacillary dysentery is a very common disease in China, then it is not difficult for doctors to make a correct clinical diagnosis. Therefore, in different regions of Guangxi Province, it is believed that the disease data quality is reliable. Monthly- notified cases of bacillary dysentery were provided by Guangxi Center for Diseases Prevention and Control. This retrospective study was approved by Guangxi Center for Disease Prevention and Control. The written informed consent was not obtained. The patient records information has been anonymized and de- identified prior to analysis. The study period covered 2 years (2009–2010). Demographic data for Guangxi were collected from local government reports. The socio-economic data were collected from Guangxi Statistical Yearbook, Guangxi Health Statistics Yearbook and Guangxi Civil Affairs' Statistical Yearbook. The data were analyzed at a county-level. ## Methodology The spatial autocorrelation concept is based on the first geography law introduced by Tobler. “Everything is related to everything else, but nearest things is more related than distant things”. Global indicators of spatial autocorrelation measure if and how much the dataset is autocorrelated throughout the study region. In the field of infectious diseases, spatial autocorrelation measures the degree of dependency among the incidence of infectious disease in different areas, considering their similarities and their distance relationships at the same time. In this paper, the relationship between bacillary dysentery and related socio- economic factors were measured based on spatial autocorrelation at county level in Guangxi Province. In general, the first step is to establish the corresponding weight matrix; followed by visual analysis, to obtain the basic distribution of the attribute; the last step is the global correlation analysis, including single variable space autocorrelation analysis and multivariate spatial autocorrelation analysis. In this paper, only the multivariate spatial autocorrelation analysis was implemented, since the research objective was to explore the main socio-economic factors of bacillary dysentery. The spatial multivariate autocorrelation was based on the software of ArcGIS10.2 and GeoDA 0.9.5i. The variables of socio-economic driving forces were proportion of primary industry, proportion of secondary industry, proportion of tertiary industry, per capital GDP, per capital government revenue, sex ratio of male and female, percentage of illiterate population in total population aged 15 and above, rural population proportion, rate of younger than 5 year old children in total population, popularization rate of tap water in rural area, access rate to the sanitation toilets in rural, number of hospitals per thousand persons, number of beds in hospitals per thousand persons, medical and technical personnel per thousand persons. One of the principal global indicators of autocorrelation is the Moran's index I, defined in formula as follows, Where is the total pixel number, and are intensities in points and (with ≠) respectively, is the average value, and is an element of the weight matrix. ∈ \[−1; 1\]; if ∈ \[−1; 0\] there is a negative autocorrelation; if ∈ \[0; 1\] there is a positive autocorrelation. Theoretically, if converges to 0, there is null autocorrelation, in most of the cases, instead of 0 the value used to affirm the presence of null autocorrelation is given in equation:where is the number of events in the whole distribution The credibility test of Moran's I is calculated as Moran's I statistics have been widely used in social and health studies –. A few applications have been presented from epidemic in recent years –. Most epidemic phenomena displayed geographic patchiness, and it was found at all spatial scales – from micrometers to continental and ocean-wide scales showed that the Moran's I of epidemic incidence data was scale-dependent. It was found that the occurrence of spatial autocorrelation of epidemic was highly dependent on the aggregation level of the natural and socio-economic data. This paper presented a study of Moran's I on bacillary dysentery as one of the few early attempts. Moreover, this paper used Moran's I statistics to explain some remaining questions from previous publications about bacillary dysentery, and the method of Monte Carlo simulation was used to test the confidence level of Moran's I. # Results and Discussion ## Spatial distributions of socio-economic factors Prior to the analysis of global spatial autocorrelation, it was necessary to understand the regional distribution of each variable. The areas with high incidence of bacillary dysentery were distributed in the northwest of Guangxi, Nanning city and its surrounding districts, Hezhou city and its surrounding districts. The 14 socio-economic attributes were distributed as follows: The areas with high percentage of Children younger than 5 year old in total population were mainly located in the southeast and northwest parts of Guangxi; those with high percentage of illiterate population in total population aged 15 and above in the northwest, and parts of the southeast and northeast of Guangxi; and those with high sex ratio of male and female in the south and east of Guangxi; the rural population more dispersedly with no significant aggregation; the areas with high percentage of primary industry in the south and part of the north in Guangxi; the areas with high percentage of secondary industry in the east and parts of northwest; and the areas with high percentage of tertiary industry in the northwest. The areas with large number of hospitals per thousand persons were located in parts of northern Guangxi. The number of beds in hospitals per thousand persons, and medical and technical personnel per thousand persons are dispersed; the areas with high per capital GDP were mainly located in the southwest and northeast; those with high per capital government revenue in parts of west and northeast of Guangxi; and those with high popularization rate of tap water in rural area and access rate to the sanitation toilets in rural in the south of Guangxi. ## Spatial correlation of bacillary dysentery and the socio-economic factors The spatial correlation between bacillary dysentery and the socio-economic factors were executed. The proportion of primary industry and the rate of bacillary dysentery incidence showed a positive correlation at 94.2% confidence level, a higher degree of confidence. The proportion of secondary industry and the rate of bacillary dysentery incidence showed a negative correlation at 90.8% confidence level, a higher degree of confidence. The spatial correlation between the proportion of tertiary industry and bacillary dysentery incidence was not significant. The reasons might be that in Guangxi the agricultural machinery production level was not high, agricultural labors relied on input, while the low-level sanitary conditions of farmers' working environment caused susceptibility to bacillary dysentery and workers' working environment was better than farmers'. The tertiary industry included a variety of career types without susceptibility to bacillary dysentery. Per capital GDP, per capital government revenue and the rates of bacillary dysentery incidence showed a negative correlation at about 70% confidence level, a general degree of confidence. Per capital GDP and per capital government revenue are representative of the level of economic development. Thus, the negative correlation means good economy is helpful for the improvement of health conditions and then reduction of the sensitivity to bacillary dysentery. Sex ratio of male and female, percentage of illiterate population in total population aged 15 and above, rural population proportion and the rate of bacillary dysentery incidence showed a positive correlation at about 30% confidence level, a poor degree of confidence. Thus, the sex ratio of male and female, percentage of illiterate population in total population aged 15 and above, rural population proportion might increase the bacillary dysentery incidence. For these variables, the regional differences were not obvious, and this may be the reason why the confidence degree was poor. Rate of children younger than 5 year old in total population and the rate of bacillary dysentery incidence showed a negative correlation at 73.6% confidence level, a general degree of confidence. Therefore, a higher percentage of children younger than 5 year old was related to a higher incidence of bacillary dysentery, which was because children had a low immunity and was more susceptible to infection of bacillary dysentery. Popularization rate of tap water in rural area, access rate to the sanitation toilets in rural and the rate of bacillary dysentery incidence showed a negative correlation at 99.8% confidence level, a very high degree of confidence. The reason was that water pollution and poor sanitation were the main causes of outbreak and prevalence of bacillary dysentery, so the improvement of drinking water and sanitary conditions will be contributed to decrease the incidence of bacillary dysentery. Number of beds in hospitals per thousand persons, medical and technical personnel per thousand persons and the rate of bacillary dysentery incidence showed a negative correlation at over 60% confidence level, a general degree of confidence. This was because number of beds in hospitals per thousand persons, medical and technical personnel per thousand persons represented medical conditions, and better medical conditions were helpful for the control of spreading of the epidemic. Number of hospitals per thousand persons and the rate of bacillary dysentery incidence showed a positive correlation at 68.7% confidence level, a general degree of confidence. The more hospitals per thousand persons were often accompanied with low-technology smaller hospitals, so the correlation was positive. ## Comprehensive analysis of the socio-economic factors of bacillary dysentery In summary, the 14 socio-economic factors can affect the incidence of bacillary dysentery. The socio-economic factors can be divided into four aspects: economic development, health development, medical development and human own condition. Economic development includes proportion of primary industry, proportion of secondary industry, proportion of tertiary industry, per capital GDP and per capital government revenue; health development includes popularization rate of tap water and access rate to the sanitation toilets in rural areas; medical development includes number of hospitals per thousand persons, number of beds in hospitals per thousand persons, medical and technical personnel per thousand persons; and human own condition includes sex ratio of male and female, percentage of illiterate population in total aged 15 and above, rural population proportion, rate of children younger than 5 year old in total population. According to the former analysis, both health development and economic development play very important roles in the incidence of bacillary dysentery. Therefore, it is really necessary to increase the popularization rate of tap water in rural area, the access rate to the sanitation toilets in rural area and the level of economic development. In addition, the four aspects of socio-economic factors are not isolated from each other, but interacted with each other. For example, the development of economy can improve the level of both health development and medical development. The development of health condition and medical condition can also improve the human environment which can attract more investment and then promote the development of economy. The economic development can enhance the education level, and reduce the proportion of agricultural population, and thereby reduce the incidence of bacillary dysentery. Therefore, it is necessary to improve all the four aspects. Only the progress of the four aspects can achieve more effects in prevention and control of bacterial dysentery. # Conclusions Through the spatial correlation analysis, the following conclusions can be drawn: Popularization rate of tap water in rural area, access rate to the sanitation toilets in rural and the rate of bacillary dysentery incidence showed a significantly negative correlation. The two factors were the most important risk factors for bacillary dysentery. The proportion of primary industry and the rate of bacillary dysentery incidence showed a positive correlation. The proportion of secondary industry and the rate of bacillary dysentery incidence showed a negative correlation, which both factors played important roles in the epidemiology. The socio-economic factors can be divided into four aspects: economic development, health development, medical development and human own condition. The four aspects are not isolated from each other, but interacted with each other. Therefore, it is necessary to improve all the aspects. Only the progress of the four aspects can achieve more effects in prevention and control of bacterial dysentery. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CN HL LY. Performed the experiments: CN GZ LZ. Analyzed the data: CN HL LY. Contributed reagents/materials/analysis tools: GZ LZ. Wrote the paper: CN HL LY.

# Introduction Extracorporeal membrane oxygenation (ECMO) is an artificial means of providing oxygenation and carbon dioxide (CO₂) elimination in patients who have refractory acute respiratory or hemodynamic failure. It temporarily supports heart and lung function, which may provide time for injured lungs to recover and enable the treatment of underlying diseases. Neurological complications such as both hemorrhagic or ischemic stroke are frequent (10% to 20% according to a series). Several mechanisms can play a role in the onset of a brain injury. First, pre-ECMO hemodynamic disturbances (sepsis, acidosis, lactate level) can alter cerebral circulation. Second, cerebral blood flow (CBF) can decrease after carotid canulationcanulation. Third, jugular canulation can induce high venous pressures in the superior vena cava, which can impact cerebral venous drainage. Finally, these intracranial pressure changes alter cerebral autoregulation after initiating ECMO and thus induce a decrease in CBF. Currently, certain techniques provide insight into cerebral perfusion during ECMO. Near-infrared spectroscopy (NIRS) measures cerebral tissue oxygenation saturation (StcO₂). StcO₂ depends on oxygen blood transportation as well as CBF. During pediatric cardiac surgery, StcO₂ desaturation has been found to occur in 70% of patients and has been inversely correlated with morbidity and mortality. During the canulationcanulation procedure, low StcO₂ has been related to a decrease in cerebral perfusion. Zulueta et al. have shown that intraoperative high cerebral StcO₂ desaturations during cardiac surgery were correlated with macrocirculatory changes involving a low cardiac index. Their results confirmed the impact of systemic hemodynamic dysfunction on cerebral circulation. NIRS monitoring during cardiac surgery allowed for the detection of cerebral ischemia when a StcO₂ decrease was associated with acute hypotension. Current evidence indicates that the systemic events during ECMO can induce the onset of cerebral ischemia and neurodevelopmental outcomes that are detected by a decrease in StcO₂. Continuous StcO₂ monitoring can measure cerebral hemodynamic changes if no hypoxic events or rapid hemoglobin changes occur. Nevertheless, no study has assessed the continuous monitoring of StcO₂ during prolonged ECMO in neonates and young infants. We hypothesized that low cerebral tissue saturation (StcO₂) during prolonged pediatric ECMO can be implicated in brain ischemic lesions and thus be associated with the rate of survival. The main goals of this study were to evaluate the prognostic role of StcO₂ on survival and its association with the occurrence of brain injuries. # Patients and methods ## Study design This observational study was conducted from September 2012 to March 2014 at the Pediatric Intensive Care Unit (PICU) of the Trousseau University Hospital in Paris. This PICU is also a neonatal and pediatric ECMO center and has 10 pediatric and 8 neonatal beds. This study included all neonates and young infants under 3 months of age assisted by ECMO for refractory respiratory and/or circulatory failures according to the following criteria: refractory hypoxemia with an oxygenation index (OI) \> 40 during more than 6 hours or an OI \> 30 for more than 12 hours; an alveolar-arterial oxygen difference (AaDO₂) \> 620 mm Hg; a PaO₂ \< 40 mmHg for more than 6 hours; refractory hypercapnia with respiratory acidosis with a pH \< 7.1; refractory septic shock; and refractory cardiogenic shock. No cardiac surgery patients were included in this study. In veno-venous (V-V) ECMO, canulation was surgically performed at the right internal jugular vein and/or femoral vein. In veno-arterial (V-A) ECMO, venous canulation was performed at the right internal jugular vein and arterial canulation at the right common carotid artery. Pediatric circuits with a centrifugal pump (Rotaflow console, Maquet, Hirrlingen, Germany) or non- occlusive roller pump (A100 console, Rhône-Poulenc, Paris, France) were used. Membrane oxygenators used were either Quadrox-i pediatric (Maquet, Germany) or Hilite 800 LT (Medos, Stolberg, Germany). The continuous cerebral tissue oxygen saturation was recorded with an Invos® device (Covidien medical, Medtronics, Minneapolis, USA). Two sensors were placed on both temporo-parietal regions. StcO₂ recording began within the first 6 hours after initiation of ECMO and ended at the discontinuation of ECMO support. Data were gathered at the end of each recording. The INVOS Analytics Tool software was used to analyze all the data; the following main parameters were measured: mean StcO₂, the duration of oxygen desaturation below the threshold of 20% from baseline and the duration of oxygen desaturation below the StcO₂ value of 50%. The mean StcO₂ value typically fluctuated between 60%-70% in normal conditions. The calculation of the oxygenation index \[(MAP x FiO₂)/PaO₂\] was performed by measuring the PaO₂ of arterial blood gas before placement on ECMO. If an arterial blood gas was not available, we used a specific scale for estimating the PaO₂ with the pulse arterial oxygen saturation according to the method described by Severinghaus. Magnetic resonance imaging (MRI) was performed after the discontinuation of ECMO and withdrawal of cannulas in all surviving patients. Brain injuries were defined as ischemic or hemorrhagic lesions based on the extent and topography of the affected regions. The French Research Ethics Committee (Comité de Protection des Personnes Ile-de- France) approved this study without any further requirements owing to the lack of interventional procedures. Parents were informed of the protocol in progress and were able to present their objection to the collection of any data regarding their infants. Consents to participate are not applicable because it is an non- invasive observational study. ## Statistical analysis The continuous variables were analyzed using a non-parametric Wilcoxon test, whereas categorical data were analyzed by Fisher's exact test. Statistical significance was defined as p\<0.05. Statistical analyses were performed with R programming software. # Study objectives The primary objective of the study was to assess the association between StcO₂ and survival during pediatric prolonged ECMO. There were two secondary objectives: first, to compare the StcO₂ in brain-injured or deceased patients versus survivors without injury; and second, to compare V-V ECMO patients with VA-ECMO patients. # Results ## Description of the population Thirty-four patients were enrolled in our study. The mean ± SD age was 24 ± 36 days, the mean birth weight was 3480 ± 815 g and the mean gestational age was 39 ± 2 weeks. The etiologies were diverse: diaphragmatic hernia (12 patients), meconium aspiration syndrome (7), severe bronchiolitis (4), refractory septic shock (4), opportunistic infections (4), and malignant pertussis (3). Before ECMO, the disease severity parameters were OI = 39 ± 16, AaDO₂ = 520 ± 178 mmHg, PaO₂/FiO₂ = 79 ± 50, pH = 7.2 ± 0.16, and PaCO₂ = 50 ± 16 mmHg. Thirty-two patients (94%) had been treated with vasopressors, and 10 patients (29%) had been treated with inotropic drugs. Seventeen of the patients were on V-V ECMO, and 17 infants were on V-A ECMO. The mean duration of ECMO was 10±7 days. No differences in demographic characteristics were observed between survivors and non-survivors. However, other significant differences were observed. The pre-ECMO mean OI was higher in survivors than in non-survivors (51 ± 15 vs 30 ± 18, p = 0.02), and the pre-ECMO PaCO₂ was higher in survivors than in non-survivors (57 ± 18 mmHg vs 43 ± 9 mmHg, p = 0.02). There were more patients who received V-A ECMO among non-survivors than survivors (71% vs 29%, p = 0.04). Half of the patients were discharged alive from the intensive care unit. Seven patients of the 17 survivors (41%) presented brain lesions on cerebral MRIs. All injuries were due to ischemic stroke. Ischemic lesions were located in the left hemisphere (4 patients), in the right hemisphere (1) and in both hemispheres (2). The mean temporo-parietal StcO₂ in survivors and non-survivors were 69 ± 8% and 54 ± 14% (p = 0.0008), respectively, in the right hemisphere and 67 ± 7% and 56 ± 16% (p = 0.03), respectively, in the left hemisphere. The mean StcO₂ was significantly reduced in the right hemisphere (56 ± 14%) among brain-injured and deceased patients compared to survivors without brain injury (70 ± 8%, p = 0.002). Nevertheless, no significant difference was observed in the left hemisphere (59 ± 7% vs 66±6%, p = 0.14). The duration of oxygen desaturation (time spent below 50% StcO₂ and time spent below 20% of the mean StcO₂) was significantly longer in non-survivors than survivors. The duration of oxygen desaturation was also significantly increased in non-survivors and brain-injured patients compared to survivors without brain injury. This significant difference was observed in both hemispheres. The mean StcO₂ and duration of oxygen desaturation were not different in patients with V-V ECMO compared to V-A ECMO. # Discussion This is the first study exploring continuous cerebral tissue oxygen saturation during pediatric prolonged ECMO and its association with morbidity and mortality. First, we found that the mean StcO₂ decreased in non- surviving newborns and young infants during prolonged ECMO. Second, the mean StcO₂ and the duration of desaturation remained normal in patients without brain injury. Although the high ECMO mortality was due to the severity of the underlying diseases (ARDS, septic shock), the morbidity was primarily due ischemic or hemorrhagic stroke. In our ECMO experience, the ischemic brain lesions occurred in 20% of newborns. Rollins et al. reported up to 62% of cases that had abnormal MRIs after neonatal ECMO, whereas ultrasound imaging was abnormal in only 24% of the cases. Therefore, although MRI improves the screening of cerebral injuries, earlier detection should lead to improved care. Neurological injuries during ECMO were closely linked to cerebral perfusion. The decrease of StcO₂ below 50% during extracorporeal support was correlated with an increased rate of cerebral injuries. In other studies, StcO₂ monitoring during surgery detected alterations in cerebral hemodynamics. In this study, the duration of oxygen desaturation was longer in brain-injured and deceased pediatric patients. It should be noted, however, that the mean duration of desaturation was only 20 minutes during the entire ECMO period. This duration of desaturation was very short compared to the total ECMO support (a mean of 10 days). It appears unlikely that these desaturating episodes significantly affected the mean StcO₂. Therefore, the decrease of the mean StcO₂ could be more related to less intense but more prolonged episodes of cerebral hypoxia and/or ischemia during ECMO. Our results suggest that continuous StcO₂ monitoring may be an important approach to predict both mortality and cerebral ischemia. The ischemic stroke events during ECMO were located in the left hemisphere in 70% of the cases. The reasons are likely multifactorial. First, cannula insertions into the internal carotid artery and/or in the right jugular vein may directly impact cerebral perfusion. Jugular canulation rapidly changes the pressures within veins due to decreases in venous drainage. Second, the definitive ligation of the jugular vein prolongs high pressure levels in the right cerebral circulation. Similarly, carotid canulation abolishes right anterograde circulation. Thus, the collateral blood flow between the two hemispheres is profoundly modified after canulations and so may be the major prognostic factor of cerebral ischemia during ECMO. Within this context, the cerebral tissue oxygenation desaturation may become an interesting clinical marker of hypoperfusion. Our study has some limitations. First, our cohort size was small. Second, only newborns and young infants were included, which limits extrapolations to older children. Third, this was an observational non-interventional study without a control group. Fourth, we did not monitor continuous arterial saturation during the study period. Thus, we cannot exclude that cerebral tissue desaturation was partly due to arterial desaturation and not only to a decrease in CBF. Concerning pre-ECMO parameters, PaCO₂ was higher before ECMO in surviving patients. PaCO₂ also has a strong vasodilator effect on cerebral arteries. Hypercapnia may exert an effect on cerebral perfusion during ECMO, increasing cerebral blood flow and subsequently StcO₂. Arituk *et al*. have shown in 126 patients that cerebral StcO₂ decreased to 10% when the mean PaCO₂ decreased from 38 to 30 mmHg without a corresponding alteration of arterial oxygen, lactate, mean arterial pressure or blood flow on the extracorporeal support. The authors concluded that there was a vasoconstriction effect on cerebral arteries. Conversely, StcO₂ increased from 53% to 63% when PaCO₂ increased from 31 to 41 mmHg during cardiopulmonary bypass. Thus, it is likely that, in the survivor group, an improvement in cerebral perfusion due to hypercapnia-induced arterial vasodilation may have reduced the ischemic process during ECMO. # Conclusion This is the first observational study assessing the prognostic value of cerebral tissue oxygen saturation on survival during pediatric prolonged ECMO. The study also evaluated the relationship between desaturations measured by continuous StcO₂ monitoring and brain injuries. The mean StcO₂ was lower in non-survivors and the duration of cerebral tissue oxygen desaturation was higher in non-survivors or in brain-injured newborns and young infants. Future studies should explore the hemodynamic effect of PaCO₂ on the cerebral collateral arteries during ECMO. Additionally, some neuroprotective strategies should be developed and assessed using target StcO₂ values. The authors wish to thank Covidien for their loan of an NIRS monitor and sensors during the study period. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** MPC PLL SR JR. **Data curation:** MPC PLL SR JR. **Formal analysis:** AF MPC PLL. **Investigation:** MPC PLL SR JR JG IG AD MD SJ. **Methodology:** MPC PLL SR JR RG RC. **Project administration:** MPC PLL. **Resources:** FC CS. **Software:** MPC PLL. **Supervision:** MPC PLL SR JR DM RC. **Validation:** MPC PLL. **Visualization:** MPC PLL. **Writing – original draft:** MPC PLL SR JR DM RC. **Writing – review & editing:** MPC PLL SR JR DM RC.

# Introduction Coral bleaching has become a major threat to coral reef ecosystems worldwide. Bleaching occurs when stress to the coral-algal symbiosis causes corals to expel their endosymbiotic algae (zooxanthellae) and, if prolonged or particularly severe, may result in partial or complete coral mortality. While many sources of stress have caused corals to bleach, “mass” coral bleaching (at scales of 100 km or more) has only occurred when anomalously warm ocean temperatures, typically coupled with high subsurface light levels, exceeded corals' physiological tolerances. This was observed during recent major El Niño-Southern Oscillation events (*e.g.*, 1982–83, 1997–98, and 2002) and verified by laboratory experiments. These bleaching events caused coral death at numerous sites around the world, with impacts on reef habitats, structures, and biodiversity that lasted a decade or more. From June to October 2005, a warm-water anomaly developed across the tropical Atlantic Ocean and greater Caribbean Sea region. Satellite-based sea surface temperature (SST) observations from the U.S. National Oceanic and Atmospheric Administration (NOAA) detected a large region of warming ocean temperatures that reached a maximum anomaly of +1.2°C vs. the long-term mean when averaged across all Caribbean reef sites. Elevated temperatures persisted for many weeks and helped fuel the most active Atlantic hurricane season on record and the most severe and extensive mass coral bleaching event observed in the Caribbean. NOAA's Coral Reef Watch (CRW) developed and maintains a suite of operational satellite sea surface temperature (SST)-based products that provide coral bleaching nowcasts and alerts. HotSpots are positive SST anomalies beyond coral's tolerance level that reflect instantaneous thermal stress and Degree Heating Weeks (DHWs) providing a a measure of sustained thermal stress during a 12-week period. In 2005, NOAA warned coral reef managers and scientists of anomalously warm conditions as they developed and spread across the greater Caribbean region. The maps of sustained thermal stress indicated levels that could cause mass coral bleaching and significant mortality, and guided both the timing and location of researchers' field observations. As a result, collaborators from 22 countries undertook the most comprehensive documentation of basin-scale bleaching to date. # Results NOAA measured sustained thermal stress in 2005 that exceeded 16°C-weeks in some regions, far greater than the thresholds that have usually been associated with the onset of mass coral bleaching (DHW  = 4°C-weeks) and mortality (DHW  = 8°C-weeks). As the event developed, water temperatures rose across the basin to levels well above normal (i.e., long-term average condition) and remained above normal for more than 7 months, resulting in especially severe thermal stress at the northern end of the Lesser Antilles. Analysis of retrospective satellite data showed that the sustained thermal stress in the Caribbean during 2005 was more intense than any of the previous 20 years. The timeline for the geographic spread of the 2005 Caribbean thermal stress was decomposed into seven major phases as identified in in late-May (i), thermal stress was observed off South America; by mid-June (ii), the Caribbean coast from Colombia to Nicaragua experienced elevated temperatures. In July (iii), the western Caribbean warm anomalies persisted from Panama to Nicaragua and the extreme western Atlantic east of the Lesser Antilles began to warm. Through August (iv), reefs in the Gulf of Mexico, Florida, the Bahamas, and the Lesser Antilles experienced high levels of stress, while low-level stress was present across most of the Caribbean. In September (v), the center of warming progressed along Cuba, Hispaniola, and Puerto Rico to the Leeward and Windward Islands while low-level stress persisted throughout the Caribbean. By October (vi), thermal stress subsided in the Gulf of Mexico; however, warm anomalies intensified in the Windward Islands and expanded into the southern Caribbean. As the region of maximum warming moved southward during November (vii), waters around the northern Antilles cooled; low-level heat stress affected the northern coast of South America until it mostly dissipated around the end of December 2005. After initial reports of bleaching in Colombia in June, CRW distributed alerts via the Internet as the thermal stress spread and intensified. Teams (represented by the many co-authors on this paper) deployed throughout the region to monitor the bleaching event as it developed, and subsequently to monitor coral mortality. Coral bleaching, other disease conditions, and mortality extended across the entire Caribbean – bleaching was especially intense along the Antilles, and was observed in most Caribbean coral species in depths to 40 m. Over 3600 field surveys were recorded from 28 jurisdictions (*i.e.*, states, territories) in 22 countries. After quality control, data from 2575 field surveys were used in the bleaching analysis and 1077 were used in mortality studies. Surveys were grouped by 0.5-degree pixel at twice-weekly time intervals to allow satellite data and field surveys to be analyzed at comparable scales. Several species and sites were reported to bleach for the first time, including: the first known bleaching at Saba; the first documented mass bleaching of the Flower Garden Banks, including at least partial bleaching of all *Millepora alcicornis* and *Montastraea cavernosa* colonies; and the first reported mass bleaching of *Acropora palmata* in Virgin Islands National Park (VINP), a species listed as threatened under the US Endangered Species Act (ESA) since 2006. Surveys conducted from the peak of thermal stress through January 2007 were analyzed to assess coral mortality. Detailed and repeated monitoring revealed that a combination of bleaching and other disease outbreaks killed coral colonies stressed by high temperatures. Some researchers identified continued mortality as late as October 2007, beyond which it was difficult to attribute further mortalities to this bleaching event with sufficient certainty. In parts of the Caribbean, temperatures remained anomalously high during the boreal winter-spring and into mid-2006, although remaining below the bleaching threshold. Many corals remained bleached, and disease and mortality continued through much of 2006. Mortality exceeded 50% in several locations and made this the worst case of thermal stress-related mortality documented in the Caribbean to date, and one of the worst cases globally. The pattern of high thermal stress followed by subsequent mortality across much of the Caribbean was consistent with the pattern seen since the 1980s and 1990s in the Florida Keys, where outbreaks of other diseases have frequently been seen in years that followed thermal stress and bleaching. In the Florida Keys in 2005, bleaching was less severe than in the Caribbean proper. However, increased temperatures were quickly followed by a loss of resistance to pathogenic disease and an increased abundance of microbial pathogens in *A. palmata*, perhaps explaining the high incidence of disease following the thermal stress by either contagious or opportunistic pathogens. A longitudinal study of cohorts of corals in this region also revealed that more extensively bleached corals were more susceptible to disease outbreaks. In VINP, video surveys of permanent transects revealed that mortality occurred in colonies due to bleaching, and in colonies that showed disease symptoms either during bleaching, after recovery from bleaching, or even without visible bleaching. Frequent monitoring of *A. palmata* also revealed that bleached corals suffered greater disease-associated mortality than unbleached colonies, indicating that disease severity was dependent on host susceptibility. In Barbados, corals remained bleached for 8 months or longer before dying, ; even a year after temperatures dropped below bleaching thresholds, some corals remained bleached or pale at many sites, particularly within the important reef-builders of the *Montastraea annularis* species complex, which are now under consideration for ESA protection. Fortunately, thermal stress was lower off Venezuela (including Los Roques, Aruba, Bonaire, and Curaçao) and bleaching, disease, and mortality were limited with no long-term community decline. Comparison of satellite data with field surveys demonstrated a strong coherence between thermal stress and widespread bleaching and mortality. However, significant variability was seen in the severity of coral bleaching among reefs within each 0.5-degree satellite pixel, presumably due to variations in local conditions (*e.g.*, hydrodynamics, light, community composition). Consistent with CRW's previously established bleaching levels, significant coral bleaching began near 4°C-weeks (Alert Level 1), with widespread mass bleaching and significant mortality occurring above 8°C-weeks (Alert Level 2). However, bleaching also occurred at sites experiencing maximum stress levels below 4°C-weeks, indicating that either the 4°C-weeks threshold may have been conservative or the 0.5-degree spatial resolution failed to detect localized high temperatures. Bleaching has been reported to depend on numerous local factors, including light level, temperature variability, and past thermal stress history. These could have influenced bleaching variability within and among reefs in each 0.5-degree pixel as well. Coral mortality in 2005 was highest in jurisdictions in the northern and central Lesser Antilles where stress exceeded 10°C-weeks. In the areas where thermal stress levels were less than 8°C-weeks, significant mortality was rare (2 of 143 surveys, \<1.5%;). Above this threshold, significant mortality was observed in 31% of events. It was likely that local conditions at scales finer than those detected by satellite observations increased or decreased the effect of the thermal stress within and among reefs at the sub-pixel scale (*e.g.*, coral community structure, small-scale hydrodynamics, past bleaching; the analysis of which were beyond the scope of this study). Despite local variability, thermal stress values exceeding approximately 8°C-weeks successfully predicted significant mortality. Thermal stress of this magnitude should be weighed carefully by reef managers. In 2005, little mortality was seen below 8°C-weeks of thermal stress while above it there was an ecologically important 1-in-3 risk of mortality. The slow rate of recovery seen in Caribbean reefs, suggests that such high levels of mortality may determine the fate of coral reef ecosystems in this region for decades to come. # Discussion Unlike many past Caribbean bleaching years, strong tropical climate forcing was only a minor driver of Caribbean SSTs in 2005. In their analysis of temperature anomalies across the tropical North Atlantic in 2005, Trenberth and Shea indicated that half of the warming (0.45°C of the 0.9°C anomaly *vs*. a 1901–1970 baseline) was attributable to monotonic climate change, while only 0.2°C was attributable to the weak 2004–05 El Niño, and even less to the Atlantic Multi-decadal Oscillation (\<0.1°C). Despite the lack of strong tropical forcing, 2005 fell among the warmest years on record. NOAA’s Extended Reconstructed SST product , showed that average ocean temperatures during the July-October period for the Caribbean exceeded temperatures seen at any time during the prior 150 years. Anticipated future warming of ocean waters is expected to increase the likelihood of future Caribbean bleaching events. High ocean temperature also contributed to the record 2005 hurricane season that damaged coral reefs in Jamaica, Cuba, the Yucatan, Flower Garden Banks, and the Florida Keys as well as causing major damage to communities and loss of human life. Hurricanes have been observed to cause mechanical damage to coral reefs, including damaging coral tissue and dislodging colonies, weakening corals in ways than could slow recovery following bleaching, and contributing to long-term ecosystem decline. However, hurricanes that pass within several hundred kilometers of coral reefs have been shown to cool anomalously warm SSTs below bleaching thresholds, and were probably significant in reducing thermal stress and preventing more severe bleaching in the Florida Keys in 2005. The absence of such cooling by tropical cyclones in the Leeward Islands most likely contributed to the extreme warming, bleaching, and mortality seen there. The major hurricanes that cooled waters around the Florida Keys in 2005 (**D**ennis, **E**mily, **K**atrina, **R**ita, **W**ilma) were strong enough to reduce the Caribbean-average HotSpots. Many Caribbean reefs have changed dramatically since the early 20^th century as a result of a wide array of human disturbances. It is unlikely that natural climate variability was the cause of declines in Caribbean reefs during recent decades, as coral reef community composition had remained remarkably stable for the prior 220,000 years. While bleaching is far from the only cause of reef decline in the Caribbean, the repeated coral bleaching events since the 1980s have been strongly attributed to anthropogenic climate change. The mass bleaching and mortality from the 2005 warming further disturbed Caribbean ecosystems that were already under assault. Coral bleaching is expected to be an even greater threat to coral reefs in the future,. Mass coral bleaching from thermal stress, followed by outbreaks of contagious or opportunistic diseases, have become a threat common to coral reefs globally. Bleaching and mortality such as that seen in the Caribbean in 2005 will undoubtedly have long-term consequences for Caribbean coral reefs, as these corals have shown very slow rates of recovery to mortality from mass bleaching. This means that any future bleaching is likely to add to the damage caused in 2005, just as the 2005 event continued the decline of reefs that have suffered past mortality from bleaching, disease, and local stressors. As this paper went to press in 2010, major bleaching was again striking reefs in the Caribbean, in some places worse than in 2005. Major bleaching events have returned to the Caribbean every five years or less, and with growing intensity. With no real sign of recovery after bleaching in Caribbean reefs, these repeated events are likely to have caused reef decline that will extend beyond our lifetimes. The data presented here will aid researchers and resource managers as they develop actions to protect reefs against the thermal stress anticipated in coming decades, especially as new studies identify ways in which reductions of other sources of stress can increase reef resilience to climate change. As global ocean temperatures continue to rise, policy makers will need to address anthropogenic climate change, and managers will have to take concerted efforts to enhance the resilience of coral reefs for us to have hope of preventing dramatic losses of valuable coral reef resources. # Materials and Methods NOAA Coral Reef Watch (CRW) thermal stress products used in this study were based on nighttime-only Advanced Very High Resolution Radiometer (AVHRR) sea surface temperature (SST) data from sensors aboard operational NOAA Polar- Orbiting Environmental Satellites (POES), produced in near-real-time at 0.5-degree (50-km) spatial resolution. SST anomalies compared the measured temperature with the expected value at that time of year for each pixel. HotSpots were computed as positive anomalies above the mean temperature of the climatologically warmest month at each satellite data pixel, based on the NOAA operational climatology from years 1985–1990 and 1993. Degree Heating Weeks (DHWs) for any given time accumulated HotSpot values ≥1°C over the preceding 12-week period. The satellite-derived quantities calculated for this paper at each reef pixel surveyed included: the date of first issuance of Bleaching Watch alert (HotSpot \>0°C); the value of maximum DHW (°C-weeks) experienced during the event; and the date when temperatures dropped below stressful levels (HotSpot  = 0). The DHW map included values in coastal regions that were masked as land in the operational CRW products. For the purpose of this figure only, the coastal values were inferred using kriging, a common statistical technique. However, all data used for the subsequent analyses and comparison with field data were retrieved from NOAA operational products. Spatial averaging of satellite metrics was performed using the original operational data from the greater Caribbean pixels containing, or nearest to, coral reef locations within the region \[100W-55W, 5N-35N\]. CRW operational products were first made available on 12-Sep-2000. The 22-year time series of annual maximum DHW was produced from a retrospective suite of products that emulated the CRW near-real-time operational product for the period 1985–2006 using data from the Pathfinder Version 5.0 SST dataset. Spatial averaging was undertaken using the same pixels used for the operational data. Field surveys of coral bleaching and mortality included at least the following quantitative data: 1) measures of coral bleaching as coral cover bleached (%), number of coral colonies bleached (*n*) and total number of colonies surveyed (*N*), or both; and/or 2) measures of coral mortality as coral cover dead (%), number of coral colonies dead (*n*) and total number of colonies surveyed (*N*), or both; 3) average observation depth (m); 4) observation date; and 5) observation location, including latitude, longitude, and reef site name. Data were quality controlled to exclude observations that met any of the following criteria: 1) bleaching observations taken before the onset of thermal stress (first issuance of Bleaching Watch alert); 2) bleaching observations taken after subsidence of thermal stress, defined as the 90^th day following the date of the last No Stress alert in 2005; and 3) mortality observations taken before the maximum DHW value occurred in 2005. Multiple observations (quadrats or transects) taken at any reef site on the same date and depth (±5 m) were combined into a single survey of either means of percent cover data or proportion of the number of colonies surveyed. The 2575 bleaching surveys used in this analysis spanned the period 3-Jun-2005 through 13-Feb-2006. The 1077 mortality surveys used to estimate mortality associated with the thermal stress event were conducted during 25-Jul-2005 to 20-Jan-2007. In some cases, multiple surveys from within 0.5-degree pixels were conducted on multiple dates during the period between the onset of thermal stress and 90 days after thermal stress subsided. However, these were almost always either new surveys at different sites or different, random sets of observations within transects. There were insufficient cases of repeated surveys of the same transect to analyze how bleaching changed through time either during the warming or cooling phases of the event. However, a few resurveyed sites did show some degree of recovery after the peak of bleaching. Reports detailing change through time at individual sites have been published and continue to be published elsewhere. As the multiple researchers who took part in this paper used a variety of methods, the work presented here was a meta-analysis of surveys conducted by numerous research institutions during the 2005 bleaching event. The techniques used were all highly comparable, well-accepted field methods. The authors assumed that differences among techniques were randomly distributed with respect to thermal stress. Past comparisons among coral reef survey methods have demonstrated that while there are some biases among methods, most provide comparable results when comparing among similar types of observation such as percent coral cover or disturbance. It is important to note that the percentage of colonies bleached was often higher than the percentage of cover bleached because (1) small colonies bleached more often than large colonies; and/or (2) both partially- and wholly-bleached colonies were counted as bleached in some survey methodologies. However, a statistical comparison of the linear regressions of percent cover bleached and percent colonies bleached vs. thermal stress found no significant differences between the slopes of the two parameters (cover  = 3.91±0.89 vs. colonies  = 3.43±0.70, expressed as slope ±95% confidence interval). This supported the assumption that the different observation methods provided comparable results for this meta-analysis. Also, because any visible bleaching probably indicated a loss of most of the zooxanthellae originally present, it was appropriate to include any degree of bleaching, from pale and partially bleached to fully bleached colonies, as an indicator of significant stress in the corals. The same applies to partial and complete mortality as either indicated a thermal stress response resulting in mortality due to bleaching or various other diseases. Therefore, partial and complete bleaching and partial and complete mortality of corals were combined as observations of bleaching and mortality, respectively. Mortality data included only corals that expert observers determined had recently died; however, the actual cause of mortality typically was not identifiable. An analysis of reefs in the region showed that 4% recent mortality normally existed as a background level during surveys in years lacking any major disturbance. This 4% background level of mortality was then considered in establishing the level of mortality considered significant in. As was expected for an accumulated variable such as mortality, total percent coral mortality did rise slowly with time after the thermal stress. The observers did not feel that they could accurately separate mortality due to the 2005 bleaching event from other causes beyond 20-Jan-2007, thus determining the end date of the data used. Finally, the data density and non-random distribution of data submissions did not permit the standardization of mortality as a function of time since observation. Operational satellite products from the co-located (or next-nearest) satellite pixel were compared with all field observations. A linear regression was used to compare mean coral bleaching (combined cover and colonies datasets) with thermal stress (observed DHW at the time of the survey). For surveys that occurred after the peak of thermal stress, the observed DHW may have declined from the maximum thermal stress experienced at that location. This could have resulted in a level of bleaching greater than that expected from the observed DHW against which it was compared. Each data point represented the average of all surveys for a given 0.5-degree pixel conducted during the twice-weekly time period (temporal resolution) of the satellite data, plotted against the DHW value observed for that pixel and time period. The relationship between observed DHW and percent coral bleached was highly significant (*slope*  = 3.41, *intercept*  = 26.94, *DF*  = 359, *p*\<0.0001, *r²* = 0.24). Given the variability of monitoring techniques employed, sampling errors within each technique, and local factors at individual reef sites (*e.g.*, shading, ponding), the explicative power of the satellite metric (*r²* = 0.24 for percent coral bleached) supported the predictive relationship between the thermal stress monitored by CRW satellite products and the observed bleaching during this event. However, it was clear that inclusion of other information, including higher spatial resolution SST-based products, may further refine bleaching predictions in the future. For consistency, mortality data were considered only for observations after the peak of the thermal stress event (*i.e.*, the maximum DHW) within a pixel and were analyzed against the maximum thermal stress. For this study, the threshold for significant mortality was defined where the observed value was twice the regional baseline mortality; *i.e.*, 8%. The nature of this analysis was very broad, combining field datasets across time, space, and survey methodology. No attempt was made to separate mortality induced by bleaching from that resulting by other diseases as both were related to thermal stress. The results showed strong predictive power. However, thermal stress was far from a perfect predictor of mortality as local variability in the response of corals at and within individual reef sites likely played a critical role due to differences in circulation, shading, past thermal stress, and other factors that may have conferred local resilience. Hurricanes extract heat from the upper ocean and induce vertical mixing. Both mechanisms have been shown to reduce the high temperatures of surface waters that cause coral bleaching. While 2005 was a record hurricane season, none passed near the Lesser Antilles where some of the highest bleaching and mortality were observed. This can be seen in hurricane tracks acquired from the National Hurricane Center ([www.nhc.noaa.gov](http://www.nhc.noaa.gov)). Surface temperatures in this region remained above climatological values throughout the May-December period, with no respite from thermal stress. The NOAA Extended Reconstructed SST data, used in were averaged across reef- containing pixels (2-degree resolution) within the region \[91W-55W, 5N-35N\] and are presented as anomalies relative to the 1901-2000 mean. # Supporting Information We thank the many researchers who contributed their data to NOAA Coral Reef Watch, ReefBase, and Coral-List to document this event. For each author, there are many more individuals in the various laboratories who were critical to the success of this work. The manuscript contents are solely the opinions of the authors and do not constitute a statement of policy, decision, or position on behalf of NOAA or the U.S. Government. Contribution 1053 of INVEMAR. [^1]: Conceived and designed the experiments: CME JAM SFH TBS GL TRLC WS. Performed the experiments: TBS LAF BB EB CB CB MB AB LBW AC BC MC MJCC OD EdlG GDP DD DLGA DSG RG SG HMG JH EAHD EH CFGJ RJJ EJD LK DK PK JCL DL JM CM JPM KM JM WJM EMM EM CAOT HAO DPT NQ SR ARR SR JFS JAS GPS BS SCCS EV SMW CW EW EHW KWR YbY. Analyzed the data: CME JAM SFH GL KBR. Wrote the paper: CME JAM SFH TBS. [^2]: The authors have declared that no competing interests exist.

# Introduction Crohn’s disease (CD) is a type of Inflammatory bowel disease (IBD), a chronic inflammatory condition that is the result of an inappropriate immune response towards elements of the intestinal microbiota. Due to chronic inflammation and exposure to pro-carcinogenic microbes, CD patients are at a greater risk of developing colorectal cancer (CRC). CD patients harbor microbiomes that differ from non-CD individuals in both composition and function, with reduced diversity and an increase of <u>a</u>dherent/<u>i</u>nvasive <u>*E*</u>. <u>*c*</u>*oli* (AIEC). AIEC are closely associated with the intestinal mucosa, termed the mucosal niche, and induce inflammation in murine models of IBD, including the well-established, inflammation-susceptibe *Il10*^*-/-* mouse model. AIEC strains lack genes associated with virulence in diarrheagenic *E*. *coli* and are defined through the *in vitro* ability to adhere to and invade epithelial cells and survive in macrophages. However, as AIEC are genetically diverse, there is no universal genomic feature that distinguishes the pathotype, limiting our ability to detect these pro-inflammatory strains from among patient microbiota. *Escherichia coli* make up a diverse species that is known to variously have beneficial, commensal, or pathogenic impact in the mammalian gut. However, most commonly used methods for characterizing the composition of gut microbiota from fecal or tissue samples, notably sequencing of selected variable regions of the 16S rRNA operon, do not differentiate among these closely-related but functionally divergent types. Emerging single-molecule (third-generation/long- read) sequencing technologies including Pacific Biosciences (“PacBio”) and Oxford Nanopore Technologies (“nanopore”) produce reads dozens to hundreds of kilobases in length, often without prior amplification. Recent work has demonstrated the utility of long reads for sequencing the entire 16S gene or entire rRNA operon, with corresponding increase in taxonomic resolution by capturing more variable sequence, including all nine variable regions of 16S, the internal transcribed spacers (ITS), and 23S. Traditional metataxonomic analysis using next-generation sequencing of one or two adjacent hypervariable regions of the 16S operon are unable to discriminate among closely-related species and genera, much less capture strain-level variation. We propose a full- length rRNA sequencing approach–enabled by longer reads produced by nanopore sequencing–that is capable of discriminating among closely-related strains of *E*. *coli*. To support the extension of these approaches to characterize strain-level variation of tissue-associated (mucosa-associated) microbiota contributing to IBD and in models of experimental colitis in mice, with a particular focus on mechanistic studies of AIEC, we produced complete genome assemblies for eight AIEC and non-AIEC *E*. *coli*, describe the genomic variation among these strains, particularly within the rRNA operon, and demonstrate the accurate identification of these strains in mixed *in vitro* and *in vivo* microbiota. ## Methods ## DNA extraction and nanopore sequencing We extracted ultra-high-molecular-weight genomic DNA from liquid *E*. *coli* cultures using a modified phenol:chloroform protocol. Briefly, 1–1.5 ml of stationary-phase liquid culture was pelleted and resuspended in TE/NaCl/Triton buffer. Cells were lysed by addition of SDS to a final concentration of 2% and Proteinase K to a final concentration of 300 mg/ml and incubated at 55°C for 30 minutes. DNA was purified twice by addition of 1x volume phenol:chloroform and phase separation. DNA was precipitated by addition of 0.1x volume sodium acetate and 2x volume isopropanol. Ultra-high-molecular weight DNA was “hooked” out with a melted glass capillary and washed in 70% ethanol, dried, and resuspended in nuclease-free water (NFW) by incubation overnight at 4°C. Multiplexed sequencing libraries were prepared for nanopore sequencing using Oxford Nanopore’s Ligation Sequencing Kit (LSK109) and Barcoding Expansion (NBD104) per the manufacturer’s recommended protocol with the following modifications. Instead of SPRI bead cleanup after each ligation reaction, we added 1 volume of “clumping buffer” (9% PEG 8000, 1 M NaCl, 10 mM Tris), incubated at room temperature for 30 mins, pelleted by centrifugation at 13K rpm, washed twice with 70% ethanol, then allowed to dry and resuspended in NFW by incubating at 4°C overnight. This “bead-free” cleanup both avoids shearing and loss of high-molecular-weight DNA and selects for larger fragments in the PEG precipitations. Multiplexed samples were sequenced on R9.4.1 flow cells on either MinION or GridION with real-time basecalling using Guppy v3.2.2 in high-accuracy mode. The distribution of read lengths is shown in. ## Illumina sequencing Strains NC101 and HM670 were sequenced by the Microbial Whole Genome Sequencing Center ([www.migscenter.com](http://www.migscenter.com)) on the NextSeq 2000 platform, yielding paired-end reads (2x151bp). Strains denoted “CU” were Illumina whole-genome sequenced by Enterome (Paris, Ile-de-France, France) on the MiSeq platform, yielding paired-end reads (2x300bp). ## Genome assembly Basecalled reads for each sample were assembled using Miniasm v0.2, followed by four rounds of polishing with Racon v1.3.1, and Medaka v0.10.0. Each assembly consisted of one contig representing the full-length genome. Assembled genomes were normalized to start at the origin of replication and re-polished across the previous breakpoint. Assemblies were further polished with Illumina sequencing data using Pilon v1.22–1. Annotation was performed using the NCBI Prokaryotic Genome Annotation Pipeline \[PGAP release 2020-02-06.build4373; \]. Predicted genes were subsequently assigned putative function by aligning to the RefSeq non-redundant protein database with Diamond v0.9.22 allowing for frame-shift sensitive alignment (diamond blastp—more-sensitive—frameshift 15). Sequence data and assembled genomes have been made publicly available under NCBI BioProject ID PRJNA759208. ## Serotype, MLST, virulence, and antibiotic resistance genes Polished assemblies were aligned against EcOH database for serotyping, multi- locus sequence typing (MLST), and VirulenceFinder databases, and NCBI antimicrobial resistance database using Minimap2 with parameter ‘-cx asm5’ to allow for indel-sensitive alignment appropriate for hybrid-assembled genomes. For serotype and MLST, the single best-aligning type gene (minimum edit distance) is given as the predicted type. Predicted serotypes, MLSTs, AMR, and virulence genes were further validated using corresponding web-based tools provided by the Center for Genomic Epidemiology (, <http://www.genomicepidemiology.org/services/>), which showed complete agreement. ## rRNA operon analysis All copies of the rRNA operon in each assembly were identified by aligning to the *E*. *coli* K12 reference rRNA using Minimap2. Variants relative to the K12 reference were identified for each assembled rRNA copy across all assemblies based on the alignments. Total polymorphism distance was computed between every pair of alignments across each region of the rRNA operon (16S, ITS, 23S). ## Bacterial growth conditions Bacteria were grown in Luria broth (LB) at 37°C and 250 rpm unless otherwise indicated. ## Animal care Germ-free mice were reared in the National Gnotobiotic Rodent Resource Center at UNC Chapel Hill. All animal experiments and procedures were approved by UNC’s Institutional Animal Care and Use Committee (IACUC). ## Murine stool samples To generate an in vivo gut microbial community containing the seven *E*. *coli* strains of interest and endogenous mouse microbiota, we inoculated mice as follows. Three (2m/1f) interleukin-10-deficient (*Il10*^*-/-*) mice (129Sv/Ev background) were reared germ-free to adulthood (8–10 weeks) and colonzied by oral gavage with an even mixture of a total of 10⁷ CFU clinical *E*. *coli* strains isolated from the intestinal tissue of Crohn’s disease and non-Crohn’s disease patients. All strains have been classified as AIEC and non-AIEC using standard *in vitro* assays to evaluate adhesion/invasion to Caco2 epithelial cells and uptake/survival in J774 macrophages. To validate our results by tracking individual strains with PCR, each strain was marked with an antibiotic resistance cassette and molecular barcode inserted into a neutral chromosomal region using Tn7 transposon insertion. During experimentation, these strains were referenced by their blinded laboratory designation and barcode: JA0018/A1, JA0019/A3, JA0022/B6, JA0036/C5, JA0044/D2, JA0048/D5, and JA0091/C2. The identity of these strains is described in. Five strains were originally isolated by KW Simpson at Cornell University as the strains CU39ES-1, CU532-9, CU568-3, CU37RT-2, CU42ET-1; HM670 was isolated previously and gifted by Barry Campbell from Liverpool University; and LF82 is the prototypic AIEC strain gifted to KW Simpson by Arlette Darfeuille-Michaud. After colonization, mice were maintained in specific pathogen free (SPF) housing, where they acquired a simplified mouse microbiota. After 2 weeks of colonization, we gave kanamycin water *ad libetum* for 2 weeks to suppress this microbiota and ensure that all strains could persist to some extent. Mice were sacrificed 6 weeks later (by CO₂ asphyxiation), for a total of 10 weeks colonization. A stool sample was removed from the lumen of the distal colon for our analysis. DNA was extracted and purified as described in. ## In vivo colonization and persistence studies To demonstrate that the seven human-derived strains and murine *E*. *coli* NC101 could colonize germ-free *Il10*^*-/-* mice throughout the gastrointestinal tract, we singly housed 1 male and 1 female mouse per *E*. *coli* strain, for a total of 16 cages. Mice were gavaged with 10⁸ CFU of a single strain and moved to SPF housing. After 1 week, each mouse received a fecal transplant (prepared anaerobically from a pool of 7 C57BL/6 WT mice) via gavage to provide niche competition. Stool samples were collected almost daily and CFUs were quantified on kanamycin plates to monitor fecal *E*. *coli* colonization. Mice were sacrificed by CO₂ asphyxiation after 5 weeks, and the following tissues were harvested and CFUs quantified on kanamycin plates to measure viable colonizing bacteria: ileal content, cecal content, colon content, colon tissue and colon mucus layer. *E*. *coli* CFUs were also quantified from stomach content, duodenal content, and jejunal content, but only 1–3 samples of each tissue had detectable bacterial growth of more than 10² CFUs/10 mg. To demonstrate that consistent colonizing strain, CU42ET-1-D5, and inconsistent colonizing strain, HM670-C2, colonized germ-free *Il10*^*-/-* mice in a similar manner across a larger cohort, we housed 7 (3M/4F) or 8 (4M/4F) mice, respectively, in two cages for each cohort. Mice were gavaged with 10⁸ CFU of a single strain and moved to SPF housing. After 1 week, each mouse received the same fecal transplant as used previously (prepared anaerobically from a pool of 7 C57BL/6 WT mice that were *Helicobacter spp*. free) via gavage to provide niche competition. After 3 weeks of colonization (2 weeks post-FMT), we administered kanamycin water *ad libetum* for 2 weeks to suppress this microbiota and ensure that all strains could persist to some extent. Mice were sacrificed 6 weeks later (by CO₂ asphyxiation), for a total of 10 weeks colonization. The following tissues were harvested and CFUs quantified on kanamycin plates to measure viable colonizing bacteria: colon content, colon tissue and colon mucus layer. ## In vitro co-culture assays to measure pro-inflammatory cytokine production To assess the inflammatory capacity of the clinical strains, J774 macrophages grown in DMEM complete media (supplemented with 10% heat-inactivated FBS and 100 U/ml penicillin-100 mg/ml streptomycin) were transferred to 12-well plates at 5x10⁵ cells/well. Cells were grown overnight at 5% CO₂ at 37°C, and washed twice with PBS before adding fresh culture media without antibiotics. Cells were infected at an MOI of 1 for 4 hours before removing the bacteria, washing with PBS, and adding fresh media with 100 μg/mL gentamycin. Cells were grown another 20 hours at which point the supernatants were collected for ELISA and the cells were collected for qPCR analysis. ## ELISA Supernatants from stimulated J774 macrophages (above) were analyzed for cytokine IL12p40 production through ELISA following manufacturer’s protocol (BD Bioscience Opt EIA, catalog \#555165). They were assayed in 4 independent experiments with 2–3 technical replicate wells. ## Quantification of cytokine expression by qPCR Stimulated J774 macrophages were washed with PBS and transferred into 1 mL Trizol (Invitrogen) and RNA was extracted following the manufacturer’s protocol. Isolated RNA was subjected to DnaseI treatment (Invitrogen) prior to cDNA synthesis. cDNA synthesis was completed using qScript cDNA SuperMix (Quantabio). qPCR amplification was performed in triplicate with SYBR green qPCR chemistry (Bioline) using primers for *Tnfa* (F-5’-`ACCTCACACTCAGATCATCTTCTC`-3’, R-5’-`TGAGATCCATGCCGTTGG`-3’), *Il1b* (F-5’-`ACAGAATATCAACCAACAAGTGATATTCTC`-3’, R-5’-`GATTCTTTCCTTTGAGGCCCA`-3’), *Ifnγ* (F-5’-`CGGCACAGTCATTGAAAGCC`-3’, R-5’-`TGTCACCATCCTTTTGCCAGT`-3’), *Csf2* (F-5’-`ATGCCTGTCACGTTGAATGAAG`-3’, R-5’-`GCGGGTCTGCACACATGTTA`-3’), *Il12b(p40)* (F-5’-`CGCAAGAAAGAAAAGATGAAGGAG`-3’, R-5’-`TTGCATTGGACTTCGGTAGATG`-3’), and *Gapdh* (F-5’-`GGTGAAGGTCGGAGTCAACGGA`-3’, R-5’-`GAGGGATCTCGCTCCTGGAAGA`-3’) on a QuantStudio 6 Real-Time PCR System. *C*_*t* values were normalized to *Gapdh* to generate Δ*C*_*t* values, and fold changes were calculated by ΔΔ*C*_*t* to the Δ*C*_*t* of unstimulated controls. These were assayed in 3 independent experiments in technical triplicates. ## Quantification of strains by qPCR Stool DNA (6 ng each) from three mice was subjected to qPCR with primer pairs targeting each barcode to quantify relative amounts of each barcoded strain from within a complex population. Amplification was performed in duplicate using SYBR green qPCR chemistry (Bioline) using a universal barcode forward primer (F-5’-`GCTTGGTTAGAATGGGTAACTAGTTTGCAG`-3’) and barcode-specific reverse primers for A1 (R-5’-`TTCCCGAGCGCACCACAAA`-3’), A3 (R-5’-`ACACATACATCTCGCACGCAAACG`-3’), B6 (R-5’-`AAACCAACATCTCCCTCGCCC`-3’), C2 (R-5’-`GGTGATGTTTGGGCGTGGTAGAA`-3’), C5 (R-5’-`ATAACACTCCCGCCCACGAGAA`-3’), D2 (R-5’-`TTCGAACTCGACCGCCAACCAAA`-3’), and D5 (R-5’-`CCACTCAATCACGCAACACCC`-3’) with *E*. *coli* 16S primers (F-5’-`ATTGACGTTACCCGCAGAAGA`-3’, R-5’-`GGGATTTCACATCCGACTTGA`-3’) on a QuantStudio 6 Real-Time PCR System. *C*_*t* values were normalized to *E*. *coli* 16S rRNA to generate Δ*C*_*t* values, and fold changes were calculated by ΔΔ*C*_*t* to the Δ*C*_*t* of the mouse inoculum. ## Mock tissue-associated microbiome Aliquots of purified genomic DNA isolated from independent strain cultures were mixed at equal abundance by weight. HMW DNA was isolated from a frozen aliquot of 10⁷ HEK293 cells using the Circulomics Nanobind kit (<https://www.circulomics.com/nanobind>) per the manufacturer’s recommended protocol. Human DNA was mixed with microbial mixture at a ratio of 99.5: 0.5% to represent a realistic tissue-associated microbiota composition. ## Full-length rRNA amplicon sequencing For stool and *in vitro* host/microbiota samples, primers for proximal 16S (27F: `AGRGTTTGATYHTGGCTCAG`) and distal 23S (2241R: `ACCRCCCCAGTHAAACT`) were used to amplify the full-length rRNA operon (4,500bp). Starting with \~5 ng expected microbial DNA, 25 ul PCR reactions were prepared with LongAmp Taq 2x Master Mix (New England Biolabs, Inc; M0287L) and 0.4 uM of each primer. We ran 20 cycles consisting of denaturation at 94°C for 10s, annealing at 51°C for 30s, and extension at 65°C for 225s, followed by final extension at 65°C for 10min. Typical yield is 700ng of full-length amplicons, which were verified by agarose gel electrophoresis and fluorescent quantification (Qubit). Amplicon libraries were prepared for nanopore sequencing using the Oxford Nanopore Ligation Sequencing Kit (LSK109) per the manufacturer’s protocol and sequenced on R9.4.1 flow cell on GridION with real-time basecalling using Guppy v3.2.2 in high- accuracy mode. # Results ## Genome assembly and annotation To establish a baseline for variation among known enteric *E*. *coli*, we produced finished genomes for eight AIEC and non-AIEC strains (NC101, CU39ES-1-A1, CU532-9-A3, CU568-3-C5, CU37RT-2-D2, CU42ET-1-D5, HM670-C2, and LF82-B6). These genomes were sequenced, assembled, polished, and annotated as described in the Methods. Sequence data and assembled genomes have been made publicly available under NCBI BioProject ID PRJNA759208. Each genome was assembled into a single complete, circular contig representing the entire genome. describes the serotype (O and H), multi-locus sequence type (MLST), and virulence and antibiotic genes found in each. illustrates these eight genomes, annotated genes, and approximate homology between them. ## Strain-level variation As illustrated in, we observe significant genome-wide structural variation while broadly conserving GC content and annotated gene density within homologous blocks, even among these closely-related strains with similar adherent-invasive (AI) or non-AI phenotype. We computed the pairwise hamming distance between sequences across each strain and each copy of the rRNA operon for the 16S, ITS, 23S, and entire rRNA region. We evaluated several regions independently for informative variation among these strains *in silico* by aligning our whole-genome sequence to either 1) the V1-V2 hypervariable region of 16S, 2) V3-V4 hypervariable region, 3) the entire 16S sequence containing all nine hypervariable regions and conserved spacers, 4) the entire rRNA operon, including 16S, ITS, and 23S, and 5) the entire genome. Very long nanopore reads allow us to use this as a proxy for PCR amplification of the various rRNA amplicon analyses since these reads average 10–20 Kbp, and thereby most often cover the entire region of interest. In general, and not unexpectedly, the larger the region used for classification, the greater the accuracy in detecting the correct strain. Hypervariable regions V1-V2 (27F-338R) and V3-V4 (343F-806R) produces poor results (averaging 35% and 27% accuracy, respectively), little better than chance since there is little or no discriminating variation within that region. The entire 16S gene performs better (70%). Notably, there is a huge amount of variation in the accuracy across strains owing to their relative dissimilarity in a particular region. Full-length rRNA achieves an average 87.3% accuracy at the strain level, followed closely by all reads aligned against the entire genome (i.e. shotgun metagenomics) at 91%. Additional benefits of whole-metagenome sequencing include the observation of genes that may imply the functional capacity of the community without explicitly characterizing its taxonomic structure, and the ability to assemble the metagenome into an approximation of its component microbial genomes. These are important, tangible benefits, but are beyond the scope of this paper. There is a relatively small improvement in classification accuracy using the entire genomes compared to the full-length rRNA operon. Like commonly used 16S primers that target subsets of the hypervariable regions, the primer pair used for the full operon are well-conserved across bacteria, but presents a much larger sequence with which to accurately classify sequences. This coupled with cost-effective multiplexing and long-read sequencing, despite much higher error rates than standard Illumina sequencing, should make this a viable approach for characterizing complex microbiome samples. ## In vivo colonization studies We performed *in vivo* colonization studies using a well-established IBD mouse model. We first determined if these human-derived strains and murine-derived NC101 could colonize and persist in mice in the presence of a competing microbiota. We colonized two singly-housed, adult, germ-free, *Il10*^*-/-* mice each with a single *E*. *coli* strain, then after one week provided niche competition by gavaging with a murine fecal transplant as outlined in. After five weeks total, gastrointestinal tissues, including ileal content, cecal content, colon content, colon tissue, and colon mucus layer, were harvested and viable *E*. *coli* were quantified by serial plating. reveals that almost all *E*. *coli* strains could colonize the lumen and intestinal tissue with complex community competition, even without kanamycin to suppress the competing microbiota. Two strains of *E*. *coli* colonized inconsistently across both mice, with high levels of *E*. *coli* in one mouse and low or undetected levels of *E*. *coli* in the other. To verify if this would persist across a larger cohort, consistent colonizing strain, CU37RT-2-D5, and inconsistent colonizing strain, HM670-C2, were given by gavage to seven or eight germ-free, *Il10*^*-/-* mice, respectively. After one week we provided niche competition by gavaging with a murine fecal transplant, and then after three weeks we provided kanamycin water *ad libitum* for two weeks to supress the competing microbiota as outlined in. After ten weeks total, gastrointestinal tissues, including colon content, colon tissue, and colon mucus layer, were harvested and viable *E*. *coli* were quantified by serial plating. reveals that consistent colonizing strain CU37RT-2-D5 colonized all seven mice to high levels, while inconsistent colonizing strain HM670-C2 colonized five of eight mice at the lumen and intestinal tissue. Future studies are needed to validate these results and draw strong conclusions, including colonization in inflamed and non-inflamed environments. To validate the pro-inflammatory potential of these strains, each strain was co- cultured with J774 macrophages and inflammatory colitis-inducing cytokine transcription and secretion was measured. The results in demonstrate that all strains induced inflammatory cytokine production. Inducing inflammatory cytokine production is an indicator of inflammatory potential, but it is not a known predictor of tissue inflammation or colonization *in vivo*. Further effort is needed to compare *in vitro* cytokine production with *in vivo* inflammation elicited by colonizing AIEC isolates. ## Full-length rRNA metataxonomics To further evaluate the utility of performing strain-level metataxonomics using full-length rRNA sequencing on an Oxford Nanopore device, we prepared and sequenced one stool sample from *in vivo Il10*^*-/-* colonization studies depicted in. *Il10*^*-/-* mice, a well-established model of microbially driven intestinal inflammation, were colonized with an equally- proportioned pool of the seven *E*. *coli* clinical isolates, and then housed in an SPF facility where they become colonized over time with a natural mouse microbiome. After ten weeks, the mice were sacrificed and stool samples were collected for nanopore sequencing and quantification of each strain by targeted qPCR. We extracted DNA from the resulting stool samples. Full-length rRNA amplification and sequencing using universal 16S 27F and 23S 2241R primers produced \>600,000 reads with a median length of 4,149 bp. We assigned these reads to individual rRNA copies by aligning to our database of known sequences in the assembled genomes (40 copies, five in each of eight strains). illustrates the distribution of alignment identity (number of matching nucleotides) over the full-length rRNA (4,245bp). Sequences derived from the stool sample show a bimodal distribution with peaks at 3.9Kbp (92% identity) and 2.9Kbp (68% identity), representing the expected divergence and sequencing error matching *E*. *coli* and non-*Escherichia* genera, respectively. The proportion of reads assigned to each strain is shown in. There is a skewed representation of strains in the stool sample due to natural variation in their ability to colonize and thrive and due to interstrain competition between isolates in the *Il10*^*-/-* mouse gut. We confirmed the relative abundance of each strain in the stool sample through qPCR of the molecular barcode inserted in the neutral Tn7 site of each *E*. *coli* genome. The qPCR showed a similar result, as there was skewing among the *E*. *coli* isolates in the mouse gut as shown in with the CU568-3-C5 strain most abundant. Both consistent and inconsistent colonizing strains showed lower abundance in this stool sample as there were multiple *E*. *coli* isolates providing competition upon initial infection, and this data is a measure of relative abundance across *E*. *coli* strains and not total load with the stool. Additionally, the inconsistent colonizing strain still colonized in 5 of 8 mice. We additionally prepared a mock mixture consisting of all eight strains at even abundance mixed 0.5: 99.5% with human DNA (HEK293) to simulate a realistic tissue-associated microbiota sample. Full-length rRNA sequences were successfully amplified and sequenced as done with the stool sample. We observed relatively uniform representation of the eight strains in the mock mixture, demonstrating our ability to amplify and properly classify closely-related rRNA sequences in a host-tissue-like context, as illustrated in. # Discussion Alterations in the composition and function of gut microbial communities are implicated in both IBD and CRC, with CD patient microbiomes harboring an increase in mucosa-associated Enterobacteriaceae, including AIEC. After excluding genes associated with diarrheagenic *E*. *coli*, AIEC are defined by their *in vitro* adherent-invasive behavior. There is a clear need to extend our knowledge of AIEC to pathoadaptive determinants of colonization and virulence *in vivo*. Recent work highlights mechanisms AIEC have evolved to more effectively colonize the host mucosa including hypermotility and metabolic changes and T4SS-linked biofilm formation. These efforts involved either long *in vivo* host adaptation experiments or genome-wide screens with a limited number of *E*. *coli* strains. Other molecular mechanisms including the type 1 pili-CEACAM6 interaction that promotes adhesion of AIEC to intestinal epithelial cells do not offer a genetic marker to distinguish AIEC. Competitive colonization with pools of known AIEC and non-AIEC strains as a parallel strategy could help elucidate traits and genes important for colonization, but requires methods to discriminate between closely-related microbes. To relate specific traits of different AIEC to pre-clinical models and patients, we need to be able to distinguish individual strains from within a complex community. As AIEC are mucosa-associated, the ability to evaluate colonization of inflamed or cancerous lesions and normal mucosa would be optimal. The commonly used metataxonomic approach of sequencing one or two amplified hypervariable regions of the 16S gene is unable to distinguish among closely- related, but functionally diverse, species and strains. Consistent with previous work, we demonstrate that sequencing of longer regions of the 16S, ITS, and 23S operons, enabled by modern long-read sequencing technologies, dramatically increases taxonomic specificity. We illustrated the increase in sensitivity and specificity of strain detection using full-length rRNA amplification and whole- metagenome shotgun sequencing in *in silico* models, mock mixtures, and *in vivo* colonized murine microbiome. This focused study highlights specifically the challenges in resolving closely- related but functionally divergent AIEC and non-adherent/invasive *E*. *coli* strains within mixed microbial populations. We demonstrated accurate identification and discrimination of these strains in complex *in vitro* and *in vivo* models relevant to colitis. Our novel nanopore sequencing approach illustrates an efficient and effective method for discrimination of known pathogenic or specifically colonized microbiota, including tissue-associated bacteria, to enable investigation of spatiotemporal dynamics of microbial colonization without relying on genetic engineering, molecular barcoding, or *in vivo* fluorescence markers. While these results suggest a method for sensitive identification of known strains, this approach does not suggest a general taxonomic analysis protocol for microbiome studies since the genomes of the strains of interest must be known *a priori* in order to properly identify them. Full-length rRNA sequencing following PCR amplification permits accurate strain identification even when the microbial abundance—for example in tissue-associated microbial samples—is very low. We demonstrated high concordance among relative abundance between sequenced full-length rRNA sequences and qPCR results targeting the inserted molecular barcodes, suggesting that abundances are reliably preserved through amplification and nanopore sequencing. However, even higher accuracy and explicit identification of microbial genic content can be captured by shotgun metagenome sequencing when there exists enough material to extract high-quality microbial DNA. Future work includes optimization of the full-length rRNA protocol for tissue samples with adherent microbes. Our analysis of *in vitro* mixtures indicate this should be feasible with microbial:host DNA content as low as 0.5%—consistent with expected abundances in many human mucosa, including the gut—but we have so far failed to amplify full-length rRNA from resected mouse gut. Additionally, given the extraordinary utility of shotgun metagenomic sequence, it is worthwhile to explore the several existing methods for selective extraction and sequencing of microbial DNA to expand the availability of non- PCR-based approaches in the context of tissue-associated microbial communities. # Supporting information We acknowledge Lacey Lopez, Christopher Broberg, Adrienne Franks, and Cassandra Barlogio for experimental assistance. [^1]: The authors have declared that no competing interests exist.

# Introduction Erectile dysfunction (ED) is defined as the inability to achieve or maintain an erection satisfactory for the completion of sexual activity. ED is a prevalent health condition, which is estimated to affect 17.7–18.4% men aged ≥ 20 years in the USA and 32.2% in Europe, may be episodic or chronic, and can be associated with other sexual dysfunctions including reduced sexual desire and premature ejaculation. Common risk factors for ED encompass age, obesity, diabetes, hyperlipidemia, lower urinary tract symptoms, hypertension, low physical activity, smoke, psychiatric conditions and psychological factors. Opioids may increase sexual desire in the short term, but their long term use is known to negatively impact on sexual function and to lead to erectile dysfunction (ED). Various studies explored ED in men on opioid replacement therapy (ORT) and reported a 21–52% prevalence, with a peak of approximately 80% in a longitudinal study on a small group of patients in China. Most studies were on small samples of men on ORT, but a meta-analysis suggested that factors associated with sexual dysfunction include age, familial status, medical comorbidity, psychiatric illness, testosterone levels, opioid dosage, duration of treatment, and other current substance use disorders. The majority of studies on ED were conducted on men under methadone maintenance treatment, and a meta-analysis indicated a 52% pooled prevalence for sexual dysfunction among methadone users. Buprenorphine was introduced more recently as an alternative ORT. European data indicate that methadone is prescribed in approximately 63% of patients, while buprenorphine is prescribed in 37% of them, although these figures differ greatly from country to country. There is agreement that substitution treatment, either methadone or buprenorphine, may improve ED in opioid users. Pharmacodynamics of methadone and buprenorphine differ, in that the former is a μ-opioid receptor agonist, while the latter is a mixed agonist-antagonist opioid that acts on the μ-opioid receptor with low intrinsic activity and high affinity, and on the κ-opioid receptor with no intrinsic activity but high affinity. Human and animal data indicate that opioids may act at different sites in the hypothalamic–pituitary axis, and lead to opioid associated androgen deficiency (OPIAD), an endocrine dysfunction that includes increase of prolactin and decrease of follicle stimulating hormone, luteinizing hormone, testosterone, estradiol, and oxytocin. OPIAD results in hypogonadism, and is reported to occur after a few weeks of opioid intake and when exceeding a daily oral morphine milligram equivalent dosage (OMMED) of 100 mg. Animal data suggest lower risk of OPIAD with buprenorphine than methadone probably through the offset of hypothalamic–pituitary axis inhibition related to κ-opioid receptor antagonist activity. Data from humans are discordant, in that some studies reported more frequent sexual and ED to methadone, while other ones found no significant differences between the two maintenance treatments. Sexual functioning is an important determinant of quality of life (QoL), but the effect of ED on QoL has been seldom explored in men under ORT. The aims of the present cross-sectional multicentre study are to explore the prevalence and severity of ED in patients under methadone or buprenorphine maintenance treatment, the factors related to ED, and the impact of ED on patient’s QoL. To this aim, we recruited a large sample of patients, collected data on a number of demographic, clinical, and psychopathological factors, and explored their role as predictors of ED and QOL through univariate and multivariate analysis. # Materials and methods Patients were recruited from November 1^st, 2014 to January 31^st, 2015 in twenty-two National Health Service Drug Addiction Units that belong to the *Gruppo InterSERT di Collaborazione Scientifica* (GICS), a scientific collaborative network dealing with drug-related problems and located in Italy. The inclusion criteria were: a) male sex, b) age 18 years or older, c) history of opioid use, d) having been on oral methadone or oral buprenorphine treatment for at least 6 months, e) absence of any relevant comorbidity that might have influenced erection and/or QoL, including diabetes mellitus, peripheral vascular disease, neurological diseases that may be associated with ED (stroke, Parkinson’s disease, multiple sclerosis, spinal cord injury), severe obesity (i.e., body mass index \> 35), drugs that might interfere with sexual function including hormonal treatment for prostate cancer or benign prostatic hypertrophy, urological conditions or previous surgical procedures that might cause ED, cancer, severe heart, lung, liver or kidney disease. Written informed consent was obtained prior to inclusion in the study, which was conducted in accordance with the declaration of Helsinki and approved by the Ethics Committee of the Verona University Hospital. No benefit was provided for participation in the study that was voluntary and confidential. Demographic (age; education: grade school, high school, university; employment: unemployed, employed; marital status: single or divorced, engaged or married; parenthood: no, yes), and clinical variables (type of opioid: methadone, buprenorphine; daily OMMED: mg; smoke: yes, no; cigarettes: cigarettes/day; psychoactive drugs: yes/no; type of psychoactive drug: benzodiazepine, antidepressants, antipsychotics, more than one) were derived from clinical records. OMMED was calculated using standard dosage conversion calculations, i.e. a conversion factor of 4.7 for methadone and of 37.5 for oral buprenorphine. Psychopathological symptoms were measured with the Symptom Check List-90-R (SCL-90-R), a widely-used 90-item scale that includes a number of different subscales exploring the severity of respondents’ symptoms over the previous seven days. Each SCL-90-R item is rated on a 5-point Likert scale ranging from ‘not at all’ (0) to ‘extremely’ (4). We derived nine subscales, i.e., somatization (SOM), obsessive-compulsive (OC), interpersonal sensitivity (IS), depression (DEP), anxiety (ANX), hostility (HOS), phobic anxiety (PHO), paranoid ideation (PAR), and psychoticism (PSY), and the global severity index (GSI) from the Italian Version of the SCL-90-R. The GSI, which was obtained by adding the scores of all 90 items and dividing by 90, is considered the single best indicator of the current level or depth of an individual's disorder, in that it combines information concerning the number of symptoms reported with the intensity of perceived distress. Erectile function was explored with the abridged five-item version of the International Index of Erectile Function (IIEF-5), a self-administered questionnaire that explores quality of erectile function and sexual intercourse confidence and satisfaction. Each IIEF-5 item is rated on a 5-point Likert scale, ranging from 1 to 5, with higher scores indicating better erectile function and satisfaction. The severity of ED (range: 5–25) was graded as none (IIEF-5 score: 22–25), mild (17–21), mild to moderate (12–16), moderate (8–11), and severe (5–7), and the absence of sexual intercourses was also recorded. QoL was scored with the 12-Items General Health Questionnaire (GHQ-12), one of the most widely used screening tool, which measures changes in psychological health and is composed of 12 questions on mood states over the previous two weeks: lost sleep, feelings of being under strain, could not concentrate, felt unable to play a useful role, could not face problems, could not make decisions, could not overcome difficulties, felt unhappy, did not enjoy day-to-day activities, felt depressed, lost confidence, and felt worthless. GHQ-12 items were scored on a 4-point scale, ranging from 0 to 3, with higher values indicating more severe psychological distress, and the overall score (range: 0–36) was calculated. ## Statistical analysis All tests were carried with the IBM SPSS version 20.0 statistical package. The normality of variable distribution was analyzed with the Skewness-Kurtosis test. The Pearson’s χ² test and Bonferroni’s corrected post-hoc with the non-parametric Mann-Whitney U test were applied to categorical variables. For continuous variables, the one-way ANOVA and post-hoc with Bonferroni’s correction were used in case of normal distribution, while the Kruskal Wallis H test and Bonferroni’s corrected Mann-Whitney U tests were applied if the distribution was not normal. The homogeneity of variances was analyzed with the Levene’s test, and the data were logarithmically transformed before submitting them to ANOVA if the variances were inhomogeneous. Correlations were explored with the Spearman’s ρ correlation coefficient. Multivariate analysis was used to explore the influence of the demographic and clinical covariates on ED and QoL measures. Linear regression model analysis was applied for IIEF-5 and GHQ-12 scores (continuous dependent variables). The risk of having ED was also explored with logistic regression model analysis (binary dependent variable: any ED, no ED), and the results were expressed as odd ratios and 95% confidence intervals (CI). The goodness of fit of the logistic regression model was assessed using the Hosmer and Lemeshow test. *P* \< 0.05 (two-tailed) was taken as the significance threshold for all the tests. # Results We recruited 1000 consecutive male patients. Demographic and clinical characteristics of those (N = 797, 79.7%), who answered the IIEF-5 questionnaire, are reported in Tables and, respectively and the full dataset is attached as supporting information. In our cohort, 598 patients (75.0%) were taking oral methadone (average daily dosage: 58.0 ± 47.3 mg, median: 50, range: 2–350), and 199 patients (25.0%) were on oral buprenorphine (average daily dosage: 8.3 ± 6.7 mg, median: 6, range: 0.5–40). According to IIEF-5 score, 96 patients (12.0%) were sexually inactive, 22 (2.8%) had severe ED, 49 (6.1%) had moderate ED, 86 (10.8%) had mild to moderate ED, 136 (17.1%) had mild ED, and 408 (51.2%) reported no ED. In sexually active patients, average IIEF-5 score was 20.4 ± 5.4 (median: 23, range: 5–25). Only 19 patients (2.4% of the whole sample, 6.4% of sexually active patients with any ED) were under treatment for ED that was prescribed by the Drug Addiction Unit physicians in 4 of them. None of them was treated with androgen replacement therapy. Patients were grouped into sexually inactive ones (N = 96), those with any ED (i.e., mild to severe, IIEF-5 score \< 22, N = 293), and those with no ED (i.e., IIEF-5 score ≥ 22, N = 408). Age significantly differed according to the group, in that it was larger in sexually inactive patients (42.9 ± 11.3 years) than those with any ED (39.4 ± 9.5) and those with no ED (37.2 ± 9.5; *p* \< 0.001), and all post-hoc comparisons were significant between the three groups. Employment status was significantly different between groups (*p* = 0.034), and post-hoc showed a significantly larger number of employed patients (58.1%) in the no ED group than the sexually inactive one (44.8%;). Smoke was significant between groups (*p* = 0.01), and post-hoc comparison showed a significantly larger number of smokers in the any ED group (15.0%) in comparison to the other groups. The number of cigarettes/day was significantly different across groups (*p* = 0.004), in that it was larger in the sexually inactive group (17.1 ± 8.5) than in the any ED group (13.9 ± 8.6;). The use of psychoactive drugs significantly differed across groups (*p* \< 0.001), and post-hoc showed that it was smaller in the no ED group (30.4%) in comparison to the other ones. The type of psychoactive drugs also differed across groups (*p* = 0.001), in that post-hoc showed significant difference between the no ED group and the other ones. Among psychoactive drugs, the use of benzodiazepines (*p* = 0.001), neuroleptics (*p* = 0.034), and more than a single active principle (*p* = 0.019) significantly differed between groups. Daily OMMED was significant between groups (*p* = 0.032), and post-hoc showed that it was smaller in the no ED group (270.1 ± 237.5) than in the sexually inactive one (310.2 ± 229.0;). GHQ-12 score significantly differed across groups (*p* \< 0.001), in that it was smaller (i.e., better QoL) in the no ED group than the other ones. The other demographical and clinical variables were not significant between groups. SCL-90-R GSI was significantly different across groups (*p* \< 0.001), and all post-hoc comparisons were significant. All SCL-90-R subscales, except HOS (n.s.) were significantly different (SOM: *p* = 0.005, OC: *p* \< 0.001, IS: *p* \< 0.001, DEP: *p* \< 0.001, ANX: *p* \< 0.001, PHO: *p* \< 0.001, PAR: *p* = 0.013, PSY: *p* \< 0.001). Post-hoc showed significant difference in sexually inactive vs. no ED groups comparison for all subscales except HOS, in any ED vs. no ED groups comparison for OC, IS, DEP, ANX, PHO and PSY subscales, and in sexually inactive vs. any ED groups comparison for IS and DEP. Age (ρ = -0.18, *p* \< 0.001) and daily OMMED (ρ = -0.08, *p* = 0.025) were significantly and negatively correlated with the IIEF-5 score, while no significant correlation was found between the number of cigarettes/day and IIEF-5 score. A significant negative correlation was found between IIEF-5 score and GHQ-12 score (ρ = -0.18, *p* \< 0.001). For sexually active patients (N = 701), covariates, which turned out significant in previous between-groups analyses (i.e., age, employment, smoke, cigarettes/day, psychoactive drug, psychoactive drug type, daily OMMED, all SCL-90-R subscales except HOS) were entered into multivariate analysis. Multivariate logistic regression model showed that psychoactive drugs and some SCL-90-R subscales (i.e., DEP, ANX, PHO, PSY) had significant direct effect, and the number of cigarettes/day had significant inverse effect on the risk of developing any ED, while the other covariates were not significant. Linear regression model showed that employment, smoke, and some SCL-90-R subscales (i.e., DEP, ANX, PAR, PSY) had significant positive effect, while age, psychoactive drugs, and daily OMMED had significant negative effect on IIEF-5 score. Multivariate linear regression model was applied to explore the influence of demographic (age, education, marital status, parenthood), clinical variables (type of opioid, daily OMMED, smoke, cigarettes, psychoactive drugs, type of psychoactive drug), SCL-90-R GSI, and IIEF-5 score on GHQ-12 scores in sexually active patients (N = 701). Age, education, daily OMMED, IIEF-5 score, and SCL-90-R GSI significantly worsened, while employment significantly ameliorated QoL according to GHQ-12 score. # Discussion To the best of our knowledge, this is the largest report on ED in men on methadone or buprenorphine maintenance treatment. The main findings, which will be discussed below, were as follows: a) nearly half of patients in our sample were sexually inactive or reported some degree of ED; b) some demographic and clinical variables, as well as most of psychopathological symptoms measured with the SCL-90-R, significantly differed according to the presence or absence of ED; c) multivariate regression model indicated that age, employment, smoke, psychoactive drugs, daily OMMED and some SCL-90-R subscales significantly influenced the risk and severity of ED; d) QoL was worse in patients with ED and significantly correlated with ED severity; e) age, education, employment, daily OMMED, IIEF-5 score, and SCL-90-R GSI significantly influenced QoL in the multivariate analysis. The 36.8% prevalence of any ED in our sample is in line with previous studies. Among our patients, 12% were sexually inactive, but we have no information whether this was caused by the severity of ED, or by other factors. Despite the severity of ED was reported to range from mild to moderate in many patients, its impact was relevant, given the relatively young age of our sample (mean age: 38.6 years). Moreover, the possibility that some patients reported less severe ED symptoms because they felt uncomfortable or embarrassed should be considered. Age was significantly higher in patients with ED and significantly correlated with ED severity (i.e., the higher the age, the more severe the ED) in multivariate analysis. This finding is in keeping with those in opioid users, and in the general population. Being employed was found to be associated with reduced severity of ED in our study. This factor was either not consistently associated with ED, or seldom explored in previous studies. Since a direct cause-effect relationship between employment and sexual dysfunction is difficult to hypothesize, we may speculate that psychological and social consequences of being employed might have reduced the severity of ED in our patients. At variance from previous reports, marital status did not turn out to be associated with ED in the present study. Differences between the samples or cross-cultural differences might have contributed to this finding. The effect of smoke appears to be complex in our group of patients. Smoking was more frequent in patients with ED, and was a significant predictor of ED severity in multivariate analysis (i.e., more severe ED in smokers). The number of daily cigarettes turned out to be significantly higher in sexually inactive patients than those with ED, and a significant predictor of any ED in logistic multivariate analysis (i.e., the smaller the number of cigarettes/day, the higher the risk of ED). While smoke stands amongst the main factors associated with ED in general population, the role of this factor was unclear in previous studies on men using opioids. Smoke, which is common in people using opioids, may contribute to ED through increase in oxidative stress and inflammatory markers, that bring about vascular changes. Having been the assessment of this factor based on self-report in our study, a recall (i.e., patients did not precisely remember the number of daily cigarettes) or response bias (i.e., patients reported a smaller number of daily cigarettes) might have contributed to the apparently contrasting finding that the number of cigarettes/day was inversely associated with the risk of ED. Moreover, smokers often vary their smoking habits over the years, and the use of smoking pack-years instead of daily cigarettes would have represented a better measure of lifetime tobacco exposure in our patients. The use of other psychoactive drug was less common in patients with no ED than those with any ED or who were sexually inactive, and this factor was significant predictor for the risk of any ED and its severity. Benzodiazepines and neuroleptics were found to be more frequent in patients with ED than those with no ED, and the use of more than one class of psychoactive drugs was more frequent in sexually inactive patients than those with no ED. This finding, which is in accordance with previous studies, is not surprising, since psychoactive drugs may have negative effects on sexual function through various mechanisms. Moreover, the coexisting psychopathology might contribute to ED through psychological mechanisms. Of interest, antidepressants, which were reported to be associated with sexual dysfunction in previous reports, did not differ across groups in our sample. We did not find significant difference between methadone and buprenorphine treatment. Previous studies reported contrasting results on this point, with a meta-analysis indicating that methadone has 4.01 odds ratio (95% CI: 1.52–10.55) for sexual dysfunction in comparison to buprenorphine. Reasons for previous contrasting data might include the small number of patients, and/or the absence of multivariate analysis to correct for potential confounders. Daily OMMED significantly differed between groups, being larger in sexually inactive patients, intermediate in patients who reported any ED, and smaller in those with no ED. Moreover, daily OMMED was significantly correlated with the IIEF-5 score (i.e., the larger the dosage, the more severe the ED), and the correlation survived multivariate analysis. Data from previous studies are conflicting on this point, as some of them found significant correlation between opioid dosage and sexual dysfunction, while other ones reported negative findings. It is tempting to speculate that higher opioid dosage could have resulted in more marked effects on hypothalamic–pituitary axis that in turn might have been responsible for ED in our patients. Nearly all SCL-90-R subscales were significantly different across groups, being more severe in sexually inactive patients than any ED and no ED groups. For some subscales, a significant difference was found when comparing patients with ED to those without ED. Five SCL-90-R subscales (i.e., DEP, ANX, PHO, PAR, PSY) were significant predictor of either the risk of ED or its severity in multivariate analysis. These data, which are in accordance with previous studies in the general population with no medical problems, and in patients receiving opioid maintenance treatment, underscore the importance of comorbid psychopathology in patients with ED. The association might be bidirectional, because psychological factors may result in ED that, in turn, may worsen psychopathology. Future longitudinal studies would be important to better explore the directionality of the cause-effect relationship between psychological comorbidity and ED. Despite the recommendation that androgen replacement therapy should be considered in males diagnosed with OPIAD, none of our patients were on this treatment. A small minority of them (i.e., 6.4% of sexually active patients with any ED) received treatment for ED. These figures appear to differ from those in patients under long-term opioids for back pain, 19% of whom were reported to receive prescription for ED or androgen replacement therapy, especially in case of high dosage and longer duration of opioid treatment. Reasons for this negative finding in our sample might include concern on the possible cardiovascular and hepatic side effects, the risk of prostate cancer, or the lack of awareness for this therapy by the treating physicians and/or patients. QoL was significantly worse in patients who were sexually inactive or who reported ED in comparison to those with no ED. A significant correlation was found between IIEF-5 and GHQ-12 scores, and multivariate analysis showed that ED score was a predictor of QoL impairment (i.e., the worse the ED, the worse the QoL). The other factors that influenced QoL severity were age, education, employment, daily OMMED, and SCL-90-R GSI, most of which were found to significantly influence ED. It is not surprising that ED may impact on QoL through impairment of physical and psychological functioning, but surprisingly this outcome was seldom explored in men under ORT. The first and most important limitation of the present study is the lack of data on testosterone or hypogonadism, but previous studies reported no correlation between testosterone levels and ED severity, especially in younger men, probably because serum levels do not reflect biologically available testosterone, a more accurate marker of hypogonadism. The second limitation is that, because of the large sample in the present study, the assessment of ED was subjective and patients did not undergo objective measures, which could have helped better separate physical and psychological determinants of ED. The third limitation is possible response bias, since some patients might have concealed or reported a smaller severity of their symptoms, because they felt uncomfortable or embarrassed when answering questions on ED. This hypothesis is strengthened by the relatively large number of patients (nearly 20%), who preferred not to answer the IIEF-5 questionnaire and were not included in the study. Other limitations include the cross-sectional nature of the study, and the lack of information on coexisting viral infections, such as viral hepatitis or HIV. Hepatitis C virus infection, a common comorbid condition in opioid users, can negatively affect sex hormone levels, but its role in ED is still contradictory. Our data showed that nearly half of our patients under ORT reported ED or were sexually inactive, ED severity was significantly correlated to QoL, and cigarette smoking and use of psychotropic drugs was predictive of ED. Sexuality is still a neglected issue in people with substance use disorders, but sexual and ED may have important effects on emotional and social well-being, including self-esteem and sense of self-efficacy, and may affect outcome by improving adherence to treatment. Exploring ED complaints in men using opioids under ORT is a critical component of care, because self-report on this problem still occurs in a minority of patients, and IIEF-5 represents a simple and quick self- assessment tool. Counselling patients about the importance of smoking cessation is paramount especially in this population. Future studies should explore whether reducing ED symptoms in men under ORT might improve QoL, and reduce relapse into opioid use. # Supporting information Members of the Gruppo InterSERT di Collaborazione Scientifica (GICS) in alphabetical order: Arzillo C., Benigna L., Bersani N., Bersani P., Biasin C., Bossi C., Bottazzo A., Bove A., Caccamo E., Cancian S., Cantanchin F., Cantiero D., Canzian G., Cargnelutti D., Casalboni D., Casarini R., Cibin M., Civitelli P., Cozzi T., De Cecco L., Del Zotto R., Dersini F., Duranti I., Fadelli M., Favero E., Fontana N., Franceschini A., Gaiga M., Gardiolo M., Gentile N., Gervino D., Ghezzo N., Giacomin MA., Kashanpour H., Mazzo M., Meneghello D., Mihalcea C., Milan E., Montresor M., Moratti E., Pani A., Pavani V., Pellachin P., Peroni F., Piazza M., Pieri M., Prosa D., Residori M., Righetti P., Ripoli MA., Riscica P., Rizza C., Rizzetto V., Rossi A., Rovea A., Ruffato A., Ruzziconi C., Sabbion R., Santo E., Scarzella M., Sembianti N., Simonetto P., Smacchia C., Stellato M., Stimolo C., Suardi L., Vaiana A., Zavan V., Zerbetto E., Zerman M. [^1]: The authors have declared that no competing interests exist. [^2]: ¶ Membership of the Gruppo InterSERT di Collaborazione Scientifica (GICS) is provided in the Acknowledgments.

# Introduction Polycystic kidney disease (PKD) is the most frequent hereditary cause of chronic kidney disease.. Peritoneal dialysis (PD) is often avoided for PKD patients when it comes to the choice of renal replacement therapy. Indeed, enlarged kidneys and liver may reduce the peritoneal extensibility, leading to increased intraperitoneal pressure. Higher prevalence of abdominal wall hernia, leaks, and diverticulitis-related peritonitis have been reported with PD administration in PKD patients. These complications may directly impact on PD technique survival and on patients’ outcomes. However, only a small number of studies have been designed to analyze the incidence of these events and their relation with PD technical survival or global patient survival. The purpose of this review and meta-analysis was to investigate the outcome of patients with PKD treated by PD. # Methods ## Study design This systematic review with meta-analysis was conducted according to a prespecified protocol and was reported using Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. ## Search strategy and selection criteria A bibliographic search was performed from the inception to December 31^st 2017 in the following databases: Pubmed, Embase, Google scholar and Scopus. We also screened references of included articles to identify other potential studies. The search strategy, on the article title, was as follows: "polycystic kidney"\[title\] OR "polycystic kidney disease"\[title\] OR ADPKD\[title\] or "autosomic dominant polycystic kidney disease"\[title\]) AND "peritoneal dialysis"\[title\]. One author (VD) performed the full search strategy and removed duplicates. ## Study selection and data extraction After eliminating duplicates, two authors (VD, MS) independently reviewed the titles and abstracts of all articles. Disagreements were resolved by consensus. Agreement between the two authors was assessed using the Kappa coefficient. After agreement, the full text of all articles designated for inclusion was obtained. Two authors (VD, MS) checked to ensure that all articles met the criteria for inclusion in this analysis, and then independently extracted the data into a standardized form. Extracted data were: study design, country, number of subject included, percentage of male, age, comorbidity (Charlson index, diabetes mellitus and hypertension), percentage of patients treated by automated peritoneal dialysis, transfer to hemodialysis, access to kidney transplantation, dialysis adequacy, hemoglobinemia, albuminemia, overall survival, PD technique survival defined as permanent cessation of PD therapy due to PD related complications, and occurrence of peritonitis or abdominal hernia. Study authors were contacted to obtain missing data. Studies were included if they presented at least two of the following parameters: overall survival, PD technique survival, incidence of peritonitis, and incidence of abdominal hernia. Studies were excluded if they presented any one or more of the following criteria: case report, case series, abstracts, commenters or letter to the editor, language other than French or English. ## Risk-of-Bias assessment The quality of included studies was assessed independently by two researchers (VD, MS) using the Newcastle-Ottawa scale (NOS) for cohort studies. The NOS consists of three quality parameters, namely selection, comparability, and outcome assessment. The NOS assigns a maximum of four points for selection, two points for comparability and three points for outcome. NOS scores of ≥7 were considered as high quality studies and 5–6 as moderate quality. Disagreement was resolved by joint review of the manuscript to reach consensus. Publication bias was assessed using funnel plots and the Egger’s regression test if there were up to 10 eligible studies included in the meta-analysis. ## Statistical analysis The primary outcome was overall survival. Secondary outcomes were: 1/ PD technique survival defined as permanent cessation of PD therapy due to PD related complications, (considering any other outcome as censored data), 2/ percentage of peritonitis and 3/ frequency of abdominal hernias. Extracted data were presented as number and percentage for qualitative variables, and as mean and standard deviation (or median and range) for quantitative variables. Heterogeneity between studies was assessed using the Cochran Q statistic and I² test. A random effects model was used independently of the existence or absence of heterogeneity between the results of the studies because results of studies with different design and patients’ characteristics were pooled. For time to event outcomes, when hazard ratios (HR) were not specified, they were estimated according to the information presented in the paper. PKD and non PKD patients were compared through random effects models weighted by the inverse-variance method to estimate pooled HR and odds ratios (OR) with 95% confidence intervals (CIs). Sensitivity analyses were performed. All analyses were performed using R version 3.1.2 (R Foundation for Statistical Computing, Vienna, Austria). # Results ## Study characteristics Using the search strategy, we identified 9 eligible studies presenting at least two outcome of interest. The agreement in selection of studies between the reviewers was excellent (κ = 1). All the studies included in the meta-analysis were considered as high quality studies (details showed). Publication bias, using funnel plot and the Egger’s regression test, was not assessed as the number of studies included in this meta-analysis was less than 10 studies. General characteristics of the studies included in the meta-analysis are presented in. A total of 7,197 patients were included for analysis, including 882 PKD patients. Seven of the 9 studies were retrospective. In two of the 9 studies, PKD patients were significantly younger than non PKD. Clinical characteristics of patients included are summarized in. Patients did not differ between the PKD and non PKD groups in terms of hypertension, access to kidney transplantation and transfer to hemodialysis. In one of the 9 studies, Charlson index was found to be higher in non PKD patients. Most of the patients were treated by continuous ambulatory peritoneal dialysis while only Yang et al. reported a higher prevalence of automated peritoneal dialysis in the PKD group. Biological characteristics of patients included are shown in. There was no difference in serum albumin level and dialysis adequacy in term of total weekly Kt/V urea between PKD and non PKD patients. In two studies, hemoglobin level was higher in PKD patients. ## Outcomes ### Overall survival There were 5 studies that reported hazard ratios for PKD and non PKD patient survival. 6,378 patients from Europe and China were included in the meta- analysis. PKD status was associated with a better global survival with a Hazard Ratio of 0.70 \[95% CI, 0.54–0.92\]. ### PD technique survival Seven studies reported hazard ratios for PD technique survival in both groups, which included a total of 7,046 patients from Europe, China and Taiwan. There was no difference between PKD and non PKD groups in terms of PD technique survival (HR = 0.98 \[95% CI, 0.83–1.16\]) as shown in. ### Infectious complications Seven studies reported odds ratios for the incidence of infectious peritonitis episodes in PKD and non PKD patients treated by PD. 6,767 patients from Europe, China, Taiwan and Turkey were included in the meta-analysis. There was no statistical difference in occurrence of infectious peritonitis between PKD and non PKD patients (OR = 0.86 \[95% CI, 0.66–1.12\]) as shown in. ### Abdominal wall hernias There were 7 studies that reported odds ratios for the incidence of abdominal wall hernias in PKD and non PKD patients treated by PD, including a total of 2,923 patients from Europe, China, Taiwan and Turkey. PKD patients were found to have an increased risk of abdominal hernia with an Odds Ratio of 2.28 \[95% CI, 1.26–4.12\] as shown in. ### Sensitivity analysis Because the large weight of the study by Lobbedez et al. among the included studies might affect the entire analysis, we performed sensitivity analysis to check the stability of the previous results without including the study by Lobbedez et al. (except for abdominal wall hernias because this study was not included in analysis for this outcome). Results are summarized on. Hazard ratios and Odds ratios showed a great stability even in the absence of the study by Lobbedez et al. The only observed change in the absence of the study by Lobbedez et al. was the absence of statistical relevance in the analysis of overall survival. # Discussion To our knowledge this is the first meta-analysis designed to study the outcome of PKD patients on PD. We found that compared to non PKD patients, PKD patients on PD had i) better global survival, ii) no difference in PD technique survival, iii) no difference in peritonitis rate but increased risk of abdominal hernias. Our results should reassure nephrologists managing PKD patients regarding the choice of PD as renal replacement therapy. We found that PKD patients had better survival compared with non PKD patients. This difference could be explained by lower comorbidities reported in PKD patients: younger age, lower Charlson index and higher hemoglobinemia. On the other hand, patients with this hereditary disease are usually diagnosed earlier and monitored with a better nephrology care and a planned access to PD in optimized conditions. Only one study reports this factor but it could certainly explain at least in part the good clinical outcome for PD group in other studies.In addition, PKD patients have less comorbidities and may therefore have better outcome and a relatively lower risk of death or transfer to hemodialysis because of a higher competing risk of being transplanted. Lastly, in one study, diabetes mellitus was found to be the only predictor of all-cause mortality independently from PKD status. Furthermore, the meta-analysis strategy allowed us to include 7,046 patients for the analysis of PD technique survival. PD technique survival did not differ according to PKD status. Moreover, there were no statistical difference between PKD and non PKD patients in terms of access to kidney transplantation or transfer to hemodialysis. Koc et al. reported that causes of death and transfer to hemodialysis were not different between PKD and non PKD patients (p = 0.35 and 0.36, respectively). In the study of Kumar et al, in multivariate analysis, hypoalbuminemia at initiation of PD was found to be the main risk factor for PD therapy cessation independently from PKD status. Previous reports have suggested an increased incidence of mechanical complications in PKD patients treated with PD, such as abdominal leak or hernia. In these studies, abdominal hernias were not associated with an increased risk of PD discontinuation. In the present meta-analysis, we confirm that episodes of abdominal hernia appeared statistically more often in PKD patients treated with PD. However, abdominal wall complications have been found to be more frequent in PKD patients at all stages of kidney disease including before end stage renal disease. As a consequence, it is likely that hernias may not be directly related to increased intraperitoneal pressure but may be related to collagen defects and thus be observed in PKD patients treated with other renal replacement therapies. Increased risk of abdominal hernias was not associated with a decrease in technique survival. This suggests that for these patients, the therapeutic intervention such as reducing intraabdominal pressure (reduced volume infused of peritoneal solution or using automated peritoneal dialysis) is successful for maintaining them on PD therapy. In addition, studies included in this meta- analysis did not show any difference in term of fluid leak episode incidence. Some reports indicate that PKD may be associated with an increased risk of diverticulitis, leading to greater risk of peritoneal infection in patients treated with PD. In the present meta-analysis, including 6,767 patients, we found that PKD was not associated with a higher risk of occurrence of peritoneal infection. Furthermore, incidence of peritonitis episode requiring catheter removal did not differ between groups in Kumar et al. and Xie et al. (25% in PKD group vs 21% in non PKD group and 6,7% in PKD group vs 3,3% in non PKD group respectively). Staphylococcus *spp* was the main causative micro-organism in both groups while gram-negative organisms incidence did not differ between PKD and non PKD patients. PD treatment offers numerous advantages, however only 7 to 10% of end-stage renal disease patients are treated with PD. This reflects, at least in part, the clinicians’ fear of technical failure. PKD patients are probably more concerned by this issue, because of the common misconception that these patients will develop infectious and mechanical complications if treated with PD, due to increased intraperitoneal pressure and a higher incidence of diverticulitis. Our results suggest that PD is a safe renal replacement modality for PKD patients. To the best of our knowledge, there is no available study designed to assess the impact of different PD modalities (continuous ambulatory peritoneal dialysis or automated peritoneal dialysis) on PKD patients’ outcomes. First limitation of our study is related to the small number of publications on this subject. Additionally, all the studies included in the meta-analysis are retrospective or registry based studies. Indeed, there is no randomized clinical trial available on this subject. Therefore, a potential selection bias may limit the relevance of our conclusions for the entire PKD population. However, we collected all the study available (ie 9) and a total of 7,197 patients across the 9 studies (n = 2,923 to 7,046 patients for each outcome). This was a sufficient prerequisite to perform a meta-analysis. Another limitation is related to the relative importance of the study by Lobbedez et al. among the included studies due to the large number of patients included (n = 4162) accounting for almost half of the total population and Furthermore the design of the Lobbedez study, was different it was the only one including survival analysis after exclusion of the diabetic patients. However our results show strong stability when analyzed with or without Lobbedez’s study. In conclusion, this meta-analysis showed that PKD patients treated with PD seem to have an increased survival and an increased rate of abdominal hernia, without any impact on PD technique survival. There was no statistical difference in peritonitis rate between PKD and no PKD patients. Therefore, our data suggest that peritoneal dialysis is a safe modality to treat end stage renal disease of PKD patients and should be offered to these patients. However, properly designed controlled studies are needed to determine whether all PKD patients are eligible for PD or whether some specific criteria should be determined. A particular attention should be given to the impact of total kidney volume and liver size in the feasibility of PD. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction The posterior temporal region of the non-human primate brain (areas MT/MST), and its human homologue, known as area V5 or human MT complex (hMT⁺), are responsive to visual motion. Electrical stimulation of this region in non-human primates can influence motion direction discriminations, suggesting that its activity is critically linked to perceptual decisions. Although fundamental to our current understanding of motion perception, studies in non-human primates cannot ascertain conscious perceptual experiences during these direct alterations of neural activity. To determine whether a brain region is causally linked to a perceptual experience, one must modulate its neural activity. Causal necessity can be established by inactivation (e.g. lesion) of the brain region and observing a perceptual deficit, whereas causal sufficiency is established by modulating its activity (e.g. by electrical stimulation) and observing a corresponding change in the perceptual experience. Non-invasive methods such as functional magnetic resonance imaging (fMRI) have provided evidence in the human brain of relationships between hMT⁺ responses and subjective visual motion perception (for review, see). However, correlational techniques like fMRI and electroencephalography (EEG) cannot establish a causal relationship between hMT⁺ activity and conscious motion perception. Non-human primate lesion studies first demonstrated the necessary role of MT in motion discrimination judgments. Subsequent reports addressed the necessity of human MT⁺ in the conscious experience of visual motion. For instance, visual motion blindness (akinetopsia) was reported in a few patients with extensive stroke in the posterior temporal region. Deficits in motion processing have since been reported in healthy controls during transcranial magnetic stimulation (TMS) of posterior temporal cortex, in one patient with epilepsy during electrical stimulation of the anatomical area around hMT⁺, including superior, middle and inferior temporal and angular gyrus, and in a few patients with variable amounts of brain damage in the vicinity of the anatomical locus of hMT⁺. In contrast to these findings of disruption of motion perception, reports of positive percepts caused by functional alteration of hMT⁺ are missing. Although some studies in humans have elicited “motion percepts” by electrical stimulation in various regions of the brain, the precise anatomical location of these stimulation sites and their spatial relationships to hMT⁺ remain uncertain. Penfield first reported illusory motion caused by electrical brain stimulation (EBS) of the posterior temporal region in some cases of intraoperative monitoring. Plant and colleagues reported a patient who saw a moving colorless “fog”, without moving objects, during seizure auras as well as during electrical stimulation of epileptic tissue. Lee and colleagues reviewed the evidence of visual illusions caused by electrical stimulation of human visual cortex and suggested that the experience of “visual movement” can be elicited at many sites across cortex. We note, however, that the definition of visual movement was not specified. In a study of one patient implanted with intracranial electrodes, Matsumoto and colleagues were the first to relate evoked potentials from magnetoencephalography (MEG) during a visual motion task with a patient’s reported illusions of objects moving in depth during electrical stimulation of the posterior superior temporal sulcus. These previous findings of positive percepts must be interpreted with caution due to several caveats. The cortical tissue causing illusory percepts could have been diseased (epileptogenic), and the presence or absence of epileptic after- discharges (triggered by EBS) was not reported. In addition, the precise location of the stimulation was not adequately established by neuroimaging methods. Indeed, a more recent study failed to produce a visual motion percept by electrical stimulation at the border of fMRI-defined hMT⁺ in one patient , leaving open the question of whether electrical stimulation of hMT⁺ is sufficient to induce visual perceptions. The question of the spatial relationship between effective sites of induction of visual illusions by EBS and the site of visual stimulus-induced activity recorded by fMRI and electrocorticography (ECoG) remains unexplored. Moreover, the relationship between fMRI and ECoG signals during motion perception has not been characterized but has the potential to provide a bridge between human fMRI measures and electrophysiological recordings in animals. Combining three methods of neuroscientific inquiry (i.e. fMRI, ECoG, and EBS) in the same conscious human subjects allowed us to address the critical link between fMRI and electrophysiological correlates of motion perception and the role of hMT⁺ in the conscious perception of motion. # Results ## Co-localization and pattern of responses to motion as measured by BOLD fMRI and ECoG In three subjects, functional imaging using fMRI independently revealed higher levels of blood-oxygenation-level-dependent (BOLD) responses bilaterally in the posterior inferior temporal sulcus when viewing moving, compared to static, visual stimuli. This area of increased BOLD activation in response to moving visual stimuli was labeled area hMT⁺ in each individual separately (see for details). Intracranial electrophysiological recordings (ECoG) in subject B revealed a marked spatial overlap between the BOLD response and electrophysiological activity during the same task. During blocks of moving images, there was a significant increase in power specific to the theta (4–7 Hz) and high-gamma (50–120 Hz) bands only in the electrode directly overlapping with the fMRI- defined area hMT⁺. This electrophysiological signature is consistent with previous reports of the relationship between electrophysiological and BOLD measures. Note that no ECoG recordings were performed in subject A. The temporal profile of the power in the theta and high-gamma frequency bands shows several noteworthy findings. The profile of high-gamma and theta responses is very distinct after the first second of motion stimulus presentation. At the onset of the motion stimulus, the relative power of high-gamma band activity increases up to ∼190% of the power during the static stimulus and sustains an elevated power (∼120% of power during static stimulus) for the entire four seconds of the motion stimulus. The relative theta power is modulated at the frequency of the stimulus, with peaks in the theta power occurring approximately at the mid-point between transitions from inward to outward movement of the concentric circles. Interestingly, the peaks in theta power reach a higher level for outward than for inward motion, suggesting similarities in the response properties of our recorded theta modulation in the human brain to non-human primate neuronal tuning in MST, which shows a higher proportion of cells responsive to expansion than contraction. Changes in the high-gamma and theta frequencies occur in individual four-second trials, only for the electrode overlapping with the area of significant BOLD modulation. For each electrode, we plotted the mean relative power for both the theta and high-gamma bands over each individual four-second trial. These plots illustrate the clear separation of responses to motion and static stimuli in the mean relative power of the high-gamma and theta bands, only in electrode II (middle row). For electrode II, 91% of motion trials show a response above the mean power of the high-gamma band (i.e. above y = 1), and 90% of static trials show a response below the mean relative power of the high-gamma band. Theta band responses during individual trials are similarly consistent (89% of motion trials above mean theta power and 94% of static trials below mean theta power). For both high-gamma and theta bands, the mean of the distribution during the motion condition is larger than the mean of the static condition (p\<0.001, t-test; for all 3 experiments and each band). Electrodes that did not overlap with areas of significant BOLD modulation failed to show significant electrophysiological activation in response to the same motion stimulus in ECoG recordings. In all other analyzed electrodes (N = 14), none of which overlapped fMRI-defined area hMT⁺, the distributions of responses to motion and static trials are not well-separated (p\>0.15, t-test, for all 3 experiments and each band), with approximately equal numbers of points of each condition falling above and below the mean relative power (see Electrodes I and III in as examples). In subject C, intracranial electrodes were situated near the border of, but not within, fMRI-defined area hMT⁺. In this subject, we did not find any significant task-induced theta or high- gamma band activity in any intracranial electrodes, congruent with the idea that the electrophysiological and BOLD signals agree spatially. ## Electrical brain stimulation in hMT+ causes illusory visual motion As part of routine brain mapping procedures conducted for clinical purposes, electrical stimulation was performed in all three patients. During this process, a weak and focal electrical current was delivered to the brain area located between two electrodes (i.e. bipolar stimulation) while subjects were lying comfortably in the hospital bed with their eyes open. Patients were generally unaware of the timing of electrical stimulus delivery, which also included interspersed sham stimulations. Subjects were asked to describe in detail all changes in perception or subjective experience during electrical stimulation. We define illusory visual motion percepts as any change in conscious visual perception that (a) involves the translocation of one or more parts of the visual environment across visual space and (b) is directly elicited by the electrical stimulation. Reproducible, vivid, illusory visual motion percepts occurred when electrical charge was delivered through electrodes that were localized within the hub of fMRI activity corresponding to hMT⁺ in subjects A and B. The qualitative experience of the percepts was stereotyped within each individual regardless of stimulation intensity (1-12 mA) or duration (3–6 sec). The conscious illusory experiences in subjects A and B were similar but not identical. Electrical stimulation of right hMT⁺ in subject A caused displacement and transposition of the entire visual field to the left (i.e. optical allesthesia). This reported illusory percept of the visual field “jumping” to the left was spontaneously generated and present even with eyes fully deviated in left lateral gaze. In subject B, electrical stimulation of left hMT⁺ caused an illusory percept of objects moving in the contralateral (right) upper visual field as if they were “vibrating” (subject’s word). For example, while looking at the experimenter’s face, the subject reported that “the top right corner of the face is vibrating”. The effect was limited to the subject’s upper right visual field quadrant. Interestingly, when the subject’s eyes were closed and he was asked to imagine an object he had just seen, the imagined object was reported as “vibrating” during electrical stimulation and not during sham stimulation. The intensity of the illusory experiences in both subjects was not subtle. The subjects volunteered their descriptions readily and seemed to be completely captivated by the intensity of the experience. Importantly, subjects successfully kept fixation during electrical stimulation trials. Exhaustive direct inspection of video recording obtained during electrical stimulation did not reveal any macroscopic eye movements in either subject during electrical stimulation (see ; note the consideration of imperceptibly small eye movements). The total number of electrical stimulation trials at each site and the number of times a motion percept was elicited at that site is shown in. Across all subjects, stimulation directly over hMT⁺ (III–IV in Subject A, I–II in Subject B, see for locations) elicited illusory motion percepts in 92% (24 of 26) of trials. Electrical stimulation at sites directly neighboring hMT⁺ (II–III and IV–V in Subject A, II–III in Subject B) elicited the illusory motion in 39% (7 of 18) of trials, but these positive trials only occurred at the highest stimulation amplitudes tested. At these neighboring locations, one of the two bipolar electrodes was overlapping hMT⁺. In contrast, stimulation at all other cortical locations, where neither stimulating electrode overlapped hMT⁺, elicited illusory motion 0% (0 of over 100 trials) of the time. Sham stimulation trials, which were interspersed between hMT⁺ stimulation trials and did not involve current delivery, also did not elicit any illusory motion (0 of 6 trials). Illusory motion was not elicited by stimulation at any electrode sites in subject C, who had electrodes positioned adjacent to, but not within, fMRI- defined area hMT⁺. Together with the results from subjects A and B, these negative findings in subject C further suggest a high degree of spatial congruence between fMRI and electrophysiological responses to motion and conscious perceptions of motion elicited by electrical stimulation. Although not the focus of the current report, other perceptual illusions (such as an urge to move the contralateral hand, or tingling in the contralateral side of the body) occurred at some other electrode sites. None of these percepts were related to visual motion perception. # Discussion We report that electrical stimulation of functionally defined cortical area hMT⁺ causes reproducible illusions of visual motion. This illusory visual motion was only elicited when the site of electrical stimulation was precisely overlapping with the area of fMRI activation, defined independently in each subject in response to visual motion stimuli. Moreover, the electrophysiological activity recorded by ECoG during the same task was clearly limited to the electrode overlapping the area of fMRI activation. We interpret these results in the context of previous human and non-human primate studies that have shown the causal necessity of area MT for motion perception. Our results show, for the first time, that altering neural activity in hMT⁺ by electrical charge delivery is sufficient for producing complex positive illusions of visual motion. They also provide converging evidence from three different methodologies (i.e. fMRI, ECoG, and EBS) that allows inferences regarding the electrophysiological basis of the fMRI signal and the relevance of fMRI and ECoG correlates of a perceptual task to human conscious perception. ## Necessity and sufficiency- a causal link between the activity of the hMT+ network and subjective visual motion perception A substantial body of previous research has provided strong evidence to support correlations between hMT⁺ activity and visual motion perception. Causal links between MT activity and motion direction discrimination judgments have been demonstrated in the non-human primate , but studies in non-human primates have limited ability to address the subjective perceptual experience produced by experimental alterations of MT neuronal activity. The loss of cortical tissue surrounding the anatomical location of MT/hMT⁺ has been shown to produce loss of motion sensitivity in non-human primates and akinetopsia (motion blindness) in humans, both negative symptoms. Similarly, transcranial magnetic stimulation (TMS) to hMT⁺ can lead to transient loss of motion sensitivity. These prior studies provide evidence for the necessity of MT/hMT⁺ in motion perception. While disruption of function (negative effect supporting necessity) can occur following lesions or TMS, positive percepts (supporting sufficiency) can only be achieved by altering, rather than stopping, the activity of a critical network. Reports of positive percepts of motion are much more rare and have not been linked to human area MT⁺ as defined by BOLD fMRI. The placement of intracranial electrodes in the human brain is a unique opportunity to observe the effects on conscious perceptual experience during alterations of neural activity. Reproducible and consistently elicited conscious motion percepts caused by electrical charge delivery to hMT⁺, as reported here, satisfy conditions of sufficiency. That is, altering the neural activity within hMT⁺, and the network it is connected with, is sufficient for producing vivid subjective motion percepts. In one previous study, electrical stimulation at the border of hMT⁺ failed to elicit any percept (similar to our finding in subject C). Blanke et al also failed to produce positive illusory visual percepts during electrical stimulation of the temporo-parietal region in a single patient, but it is noted that the posterior extension of their electrode grid only covered the anterior portions of the junction between the inferior temporal sulcus (ITS) and the ascending limb of the ITS, where area hMT⁺ is generally thought to be located. Also, no fMRI or ECoG measures of visual motion perception were obtained. In addition to exact location, precise electrical stimulation parameters may be crucial in determining whether positive or negative perceptual phenomena occur. The lack of positive perceptual phenomena in these previous studies are in line with our own null result in subject C, and can be explained by our observations in subjects A and B that the positive phenomenon of illusory visual motion is elicited only if the site of EBS is co-localized precisely with the brain site that shows positive functional response (identified by fMRI or ECoG) during visual motion perception. This need for functional localization is clear when considering the individual variability of the location of hMT⁺ with respect to anatomical landmarks. ## Visual imagery is affected by electrical stimulation of hMT+ In our experiment, we asked subject B to close his eyes and imagine a recently viewed object “in his mind’s eye” while electrical charge was delivered to hMT⁺. Interestingly, the subject reported the same visual motion illusion (“vibrating”, or oscillatory left-right motion of the imagined object) caused by electrical stimulation. In contrast, sham stimulation during imagery trials elicited no positive reports by the subject (i.e. he did not see any change in the mental image; see). Therefore, the percept produced by electrical stimulation of hMT⁺ affects a mental image similarly to a real visual image. This finding lends support to the hypothesis that mental imagery may be an emulation of perception and that the neurons that code for a mental image may be the same as, or overlap with, those used in visual perception. ## Propagating electrical charge within selective anatomical networks The spatial spread of electrical charge is an important consideration for interpreting results from EBS experiments. Although little is known about the effect of electrical stimulation of the cerebral cortex in the human brain, the emerging evidence from cortical micro-stimulation (micro-EBS) and deep brain stimulation (DBS) in mammalian brains strongly suggests that the electrical charge delivery is more likely to recruit neural fibers whereas the activity of neurons in the stimulated area is either unchanged, blocked through depolarization blockade, or only altered in a sparse and distributed set of neurons. Reliable recruitment of neural fibers will lead to propagation of electrical activity along the afferent or efferent fibers and will reach the brain regions that are connected with the stimulated area of the brain. Given that each region of the brain has selective neuroanatomical connectivity with cortical and subcortical structures, the propagation of electrical activity will only affect the activity of a selective neuroanatomical network. Thus it may be difficult to compare the functional effect of EBS, as used in brain mapping procedure, to the effect of TMS, micro-stimulation, DBS, or structural lesioning. In other words, during brain mapping, a volley of 50 Hz signals may cause depolarization blockade (i.e. impairment of function) in the actual target of electrical stimulation but, at the same time, the volley of 50 Hz electrical signals recruits a selective neuroanatomical network in the gamma band frequency. In our experiments, it is possible that hMT⁺ may have been blocked by the depolarization blockade, but in conjunction with the recruitment of its selective neuroanatomical network (such as V1), the manipulation seems to be sufficient to lead to a subjective experience of visual motion. Given that the network is recruited artificially with 50 Hz signals, the resulting subjective experience is an illusion of visual motion when there is no real motion in the visual field (i.e. a positive phenomenon). It is interesting to note that back-propagation of signals from hMT⁺ to V1 is thought to be necessary for visual awareness of motion percepts. Whether the effect of EBS is excitatory or inhibitory depends on stimulation frequency, and stimulation frequency at 50 Hz, as in our study, is more likely to be inhibitory. Inhibitory effects on connected brain areas may be as relevant as the excitatory effect of electrical stimulation for causing positive illusory phenomena. It is likely that the inhibitory effect of EBS on the areas connected to MT, such as visual areas V1 to hV4 and parts of parietal cortex, which are involved in maintaining the stability of the visual world, may result in instability of visual images and hence the illusion of motion. Because the EBS in our study was performed in a purely clinical setting for clinical diagnosis, which does not easily accommodate research stimuli/procedures, we were unable to test the ability of subjects to perceive normal visual motion during electrical charge delivery to area hMT⁺. However, given the magnitude of the illusory percept caused by the EBS, it is more than likely that the subjects would have failed to perceive normal visual motion during the procedure. Therefore, our finding of positive illusory percept is not in conflict with the previous findings of impairment in visual motion perception during electrical stimulation of hMT⁺. ## Mechanistic interpretations of different perceptual experiences The precise perceptual experiences reported by the two subjects differed and would be difficult to predict *a priori*. Nevertheless, previous literature suggests that both types of percepts are supported by hMT⁺ activity. Subject A’s percept is qualitatively similar to the phenomenon of “apparent motion”. This phenomenon describes the perception of jumping motion between two sequentially blinking stationary stimuli separated in space. In humans, hMT⁺ activity, and perhaps feedback from hMT⁺ to early visual cortex, correlates with the perception of apparent motion. As for subject B’s percept, there is also evidence that MT in monkeys and humans is required for perceiving lateral oscillatory motion. We further propose that the percepts in subjects A and B may both be related to the role of hMT⁺ and its selective neuroanatomical network in supporting the stability of the visual world during normal vision. Specifically, the reported illusion in subject A is reminiscent of descriptions of a shifting visual world after retrobulbar paralysis of the eye muscles. The similarity of these descriptions, along with the proposed roles of MT and parietal regions during saccadic eye movements, suggests that electrical alteration of activity in hMT⁺ in subject A may have caused alteration of activity in its anatomical network (i.e. synthetic and erroneous signal from hMT⁺ to its connected parietal areas). These synthetic signals could be interpreted by the receiving areas as a corollary discharge for an eye movement that did not, in fact, take place. A corollary discharge would be expected to result in a shifting visual world in preparation for an eye movement. The experience of visual jitter in subject B may be related to MT’s normal active role in suppressing movement of the visual world due to microsaccadic eye movements. Introducing spurious signals through electrical stimulation of the set of neurons underlying these computations would conceivably alter the relationship between MT signals and ongoing microsaccadic eye movements, leading to perceptions of microsaccades in a restricted region of the visual field. (The effect would be spatially localized because hMT⁺ is organized retinotopically—see). Currently, we cannot distinguish between such an indirect effect and the possibility that the alternating electrical current from EBS is directly interpreted as alternating left-right motion in this subject. However, we can exclude the possibility that EBS directly caused eye movements that explain the percept because the percept was limited to one quadrant of the visual field, while an induced, microscopic nystagmus would be equally salient in all parts of the visual field. We note that all subjects were able to keep visual fixation during electrical stimulation trials, although we cannot exclude the possibility that electrical stimulation caused imperceptible eye movements. However, such small eye movements would be unlikely to explain the large visual motion percepts experienced by subject A, or the spatially localized percepts (within a visual field quadrant) experienced by subject B. Even if small eye movements were to explain the reported percepts, it is interesting that they would have occurred only with electrical stimulation of hMT⁺. The differences in the percepts reported by the two subjects might be attributed to the involvement of different sub-regions of hMT⁺, MT and MST, each of which could have their own network connectivity. While our current methods did not allow us to specifically address whether different sub-regions were stimulated in each subject, future studies can incorporate stimuli intended to differentiate between MT and MST to test this hypothesis. Finally, since electrode grids were implanted in the right hemisphere of subject A and the left hemisphere of subject B, the differences in reported perceptions may also be due to a left-right hemispheric functional asymmetry in the affected networks. ## Linking fMRI, ECoG, and EBS ECoG recordings are a field potential aggregated from approximately 5x10⁵ neurons underlying each electrode, similar to the number of neurons in an fMRI voxel (10⁵ neurons/mm³×∼5–30 mm³ voxel size). This similar spatial resolution to fMRI, in conjunction with the similarity in the signal type to the local field potential (LFP), puts ECoG recordings in a unique position to link fMRI BOLD findings in humans to LFP responses in non-human primates. The ECoG response to the same motion stimulus as used for fMRI was limited to the theta and high-gamma bands, suggesting that these particular frequency bands correlate with the hMT⁺ BOLD signal response. Future studies can test the generality of these findings in more subjects. We note the strong similarity of our ECoG recordings from hMT⁺ to LFP recordings from area MT in the non-human primate using microwire electrodes ( in ). In both cases, there is increased power in the high-gamma band (∼50–120 Hz) at the onset of the stimulus. In our recordings, using a long four-second stimulus, the strong high-gamma band response decreases somewhat after approximately 500 ms. The theta power is sustained at a high level throughout the stimulus, although it is also temporally modulated by the stimulus. Such differential dynamics of signals across frequency ranges will be an interesting point of study in the future. Combining the multiple methodologies of fMRI, ECoG, and EBS provides an especially powerful set of interrelated findings to help understand specific functions of cortical areas. ## Epileptic brains Although our results were obtained in patients with epilepsy, we believe the results are unlikely to be explained by pathological factors. As noted, area hMT⁺ was void of any epileptiform activity in all three patients, and data from any electrodes showing epileptic activity were excluded in our electrophysiological analysis. The positive illusory percepts were also recorded without the presence of any after-discharges. Our study included only three subjects, but it should be noted that the posterior regions of the brain are rarely implanted with electrodes and thus intracranial recordings from hMT⁺ are uncommon. Restrictions due to the clinical setting of this research provided other challenges as well. We were not able to perform ECoG recording from the hMT⁺ electrodes in Subject A because the EBS procedure was performed shortly before surgery and we could not delay the surgery in order to obtain those recordings. ## Conclusions Taken together, our findings are consistent with studies in non-human primates suggesting a crucial role of area MT and its interconnected network in conscious motion perception. We demonstrate that electrical stimulation of area hMT⁺, as defined by fMRI and verified by electrophysiological responses in individual subjects, elicits a conscious experience of visual motion in awake human subjects. In the context of previous research, our results show that the hMT⁺ network circuitry is both necessary and sufficient for producing conscious motion percepts. The spatial agreement of fMRI and electrophysiological measures allows inferences about the link between these stimulus-evoked signals and their ultimate relation to conscious visual perception when the activity of the same part of the brain is electrically modulated. # Materials and Methods ## Ethics Statement Our study was approved by the Stanford University IRB Office for Protection of Human Research Subjects. All subjects signed informed consent for participation in our research study. ## Subjects Our subjects were three patients (1 male, 2 female) undergoing epilepsy surgery for intractable epilepsy. Our study did not cause additional risk to any participants, and the intracranial procedures were conducted entirely for clinical reasons to localize the source of epileptic discharges. Our diagnostic studies revealed no pathological activity in hMT⁺. Patient A was diagnosed with multifocal epilepsy originating from frontal and posteromedial regions, whereas patients B and C were diagnosed with epileptic foci in the medial (but not lateral) parieto-occipital region, after resection of which, both subjects, to date, remain seizure free. ## Functional Magnetic Resonance Imaging (fMRI) Localizer sessions were aimed at identifying motion-responsive areas. The stimulus consisted of a set of concentric dark gray circles on a gray background. The stimulus alternated between static and moving in blocks of 16 sec. During motion blocks, the circles expanded and contracted at a rate of 0.5 Hz (i.e. one full expansion and contraction every two seconds). Each run (n = 2) lasted 208 secs (192 secs in subject B) and included 6 blocks of motion and 7 blocks of static stimuli (6 static in subject B). Subjects fixated on a white dot in the center of the screen and pressed a button anytime the fixation dot randomly flashed red. All subjects performed this independent task at near 100% accuracy, indicating stable fixation. Functional magnetic resonance images were acquired on a 3T GE MRI scanner and an 8-channel volume head coil using a spiral-trajectory pulse sequence with the following parameters: one shot, TR = 2000 ms, TE = 30 ms, flip angle = 77°, FOV = 220 mm, voxel size = 1.72×1.72×2 mm³ in subjects A and C, 3×3×2.5 mm³ in subject B. Twenty-one oblique slices covering occipital and temporo-parieto- occipital cortex were prescribed approximately along the AC-PC plane. We analyzed fMRI data using the freely available, open-source mrVista software package (<http://vistalab.stanford.edu/software/>). The acquired BOLD signal from each voxel was first divided by its mean in order to compute a time series of percent modulation. High-pass temporal filtering was used to deduct baseline drifts from the time series. Small motion artifacts within and across scans were corrected using an affine transformation of each temporal volume in a data session to the first volume of the first scan. The data were analyzed on a voxel-by-voxel basis using a general linear model (GLM) that modeled the BOLD signal using a two regressors (motion and static), with an additional DC regressor for each run to account for shifts in baseline. Statistical maps were computed as voxel-wise t-tests between the motion and static conditions. Area hMT⁺ was defined by the contrast motion \> static at a statistical threshold equivalent to a false discovery rate of 5% (q = 0.05) in each individual subject. The resulting statistical contrast maps were interpolated to the T1-weighted volume anatomy and restricted to gray matter layers. These maps are projected onto a cortical surface mesh (consisting of the surface along the gray-white boundary) for visualization. In subject A, the fMRI hMT⁺ localizer was performed post-surgically, while subjects B and C participated in the same localizer session before electrode implantation. ## Electrode Localization We used MS08R-IP10X-000 strips and IG64C-SP10X-0TB grids made by AdTech Medical Instrument Corporation (<http://www.adtechmedical.com>) for recording and stimulation in our subjects. These electrodes have the following parameters: 4 mm flat diameter contacts with 2.3 mm diameter of exposed recording area (4.15 mm²) and inter-electrode distance of 1 cm. Post-surgical computed tomography (CT) images indicating the location of electrodes were aligned to preoperative T1-weighted structural MRI images using a mutual-information algorithm, implemented in SPM5 (<http://www.fil.ion.ucl.ac.uk/spm>). The electrodes were easily identified in the CT scans and their locations were manually marked. These images were visualized using ITKGray, a segmentation tool based on ITKSnap. The resulting images were manually aligned to 3D mesh renderings of the T1 anatomical images produced using mrVista, on which the fMRI activation is displayed, thereby conserving the electrode to T1 anatomical image alignment. This procedure allowed us to construct 3D visualization of electrode locations relative to each patient’s cortical anatomy within a few millimeters (\<∼3 mm) in error. The accuracy of estimated electrode sites was also validated by digital photos, obtained intraoperatively. ## Electrophysiological Recording and Analysis After implantation of the electrodes and post-surgical stabilization, the hMT⁺-localizer task was administered to the patients for ECoG recordings (patients B and C only). This task was identical to the one described for fMRI, except that blocks were 4 seconds in length instead of 16 because of the increased temporal resolution of ECoG over fMRI. There were 22 blocks of motion and 23 blocks of static, giving a run time of 180 s. Subject B completed three runs, and subject C completed two runs. We recorded signals at 3051.8 Hz through a 128-channel recording system made by Tucker Davies Technologies (<http://www.tdt.com/>). Off-line, we applied a notch filter at 60 Hz and harmonics to remove power line noise. We removed channels with epileptic activity, as determined by the patient’s neurologist. To visualize electrophysiological responses, we created event-related spectral perturbation (ERSP) maps based on the normalized power of electrophysiological activity during each condition. A Hilbert transform was applied to each of the 42 bandpass filtered time series to obtain instantaneous amplitude and power. Using the Hilbert-transformed time series, time-frequency analysis was performed for event-related data. We logged the onset and duration of each trial via photodiode event markers for each experimental condition time locked with the ECoG recording. Event markers were used to align and average power at each frequency band over repeated trials for each condition to create ERSP maps. The ERSP was scaled by the total mean power at each frequency in order to compensate for the skewed distribution of power values over frequencies and the result was converted to decibel units. In order to test the significance of changes in ERSP, we compared each ERSP frequency-time point with a constructed “null” ERSP. We first generated a surrogate data set by transforming the original instantaneous power time series into the Fourier domain and adding random phases, resulting in a surrogate of instantaneous power that has randomized phase but preserved amplitude. Therefore, the first and second order moments of the surrogate remained unchanged but its local temporal structure was removed. A “null” ERSP was then constructed from the surrogate data with the same number of trials (randomly selected) as the condition of interest. We constructed a set of “null” ERSPs by iterating the surrogate procedure 470 times (e.g. for the presented ERSP in, we generated 47 surrogate data sets and for each set, we shuffled the surrogate events 10 times). We expect that the distribution of the “null” ERSP at each frequency-time point approaches a Gaussian distribution with sufficient iterations (law of large numbers). We tested the Gaussianity of the constructed distribution by monitoring kurtosis. We kept the absolute value of the distribution kurtosis below 0.5 (the kurtosis of Gaussian is zero) by increasing the number of iterations of the surrogate procedure. Following this procedure, we used a normal distribution to fit the “null” ERSP at a given frequency for one cycle period in order to estimate its mean and standard deviation. We shifted and scaled the ERSP at each frequency-time point relative to the obtained mean and standard deviation (Z-score). We then converted the normalized ERSP (Z-scores) to p-values. Finally, we used a false discovery rate method to correct for multiple comparisons and to set a significance threshold level for each subject, electrode and condition. Tests of significance for increases or decreases in the ERSP map were performed separately. The parameters for presented ERSPs are: (q = 0.1; p-values for the increase and decrease are 0.02 and 0.001, respectively). ## Electrical Brain Stimulation (EBS) Electrical stimulation was performed as part of routine clinical procedure of brain mapping to determine areas of hyperexcitability whose stimulation causes the patient’s typical behavioral seizures, and to determine the function of each brain region before making a decision about the extent of epilepsy surgery. Electrical charges used (50 µC/cm²/pulse) in each patient were within the safety parameters and appreciably less than the ones used in older classical studies by Penfield and colleagues (∼700 µC/cm²/pulse). Stimulation was performed using the following parameters: Square wave currents from 1 to 12 mA at 50 Hz and with a pulse width of 200 µs. The impedance of these electrodes is measured to be approximately 400 Ω at 1 kHz. Subjects were comfortably lying in their hospital bed during bipolar electrical stimulation, with their eyes open (except where noted) and fixated on an object in the room. Eye movements were monitored by video recordings. Care was taken not to influence subjects’ reports of perceptions by asking open-ended questions (“Did you hear, see, or feel anything strange?”) and by including the same questions during sham stimulation trials. # Supporting Information We thank the patients for participation in these studies. We thank R. Dougherty, R. Fisher, K. Grill-Spector, W. Newsome, A. Wagner, and B. Wandell for insightful comments throughout the study, and the staff at Stanford Epilepsy Monitoring Unit for assistance and guidance during intracranial recordings. Thanks also to Sharmin Nasrullah for support in implementing electrode localization techniques. [^1]: Conceived and designed the experiments: AMR KSW NW JP. Performed the experiments: AMR JP. Analyzed the data: AMR MD JC AS. Contributed reagents/materials/analysis tools: AMR MD JC. Wrote the paper: AMR JP. Edited the manuscript: AMR MD KSW NW JC AS JP. [^2]: The authors have declared that no competing interests exist.

# Introduction The use of Next Generation Sequencing (NGS) for the analysis of complex microbial communities has increased dramatically in recent years. Reasons for this include a continual decrease in cost and an ever greater appreciation of the ability of NGS to more comprehensively characterise microbial communities than traditional culture based methods. NGS has been advantageous in determining the role of the microbiome in disorders like Inflammatory Bowel Disease, diabetes, and obesity, or environmental communities like wetland soils and oceans. There are many methodological choices to be made when conducting a sequence- based microbiome study. These decisions have led to the introduction of a variety of technical variables that affect the compositional signal to various degrees, potentially limiting the ability to investigate the main hypothesis or to compare results relating to communities that are similar but which have been investigated using different methods. Factors such as sampling methods, DNA extraction protocol, amplification, purification and quantification along with sequencing depth can significantly impact results. For instance, using different purification and quantification methods can lead to a five-fold difference in sequence counts while a one-step versus two-step PCR method can led to significant differences in alpha and beta diversity between replicates. The majority of microbiome studies have relied on 16S rRNA gene amplicon sequencing. There are nine different variable regions within the prokaryotes ubiquitous 16S rRNA gene (V1-V9), each flanked by highly conserved stretches of DNA suitable for primer binding. Depending on sequencing technology and chemistry it is possible to sequence a number of adjacent variable 16S rRNA gene regions. However, none of the currently available technologies offer full-length gene sequencing at sufficient depth to allow for multiplexing larger numbers of samples on the same run. Unfortunately no standard approach exists for selecting the most appropriate primer pair suitable for all taxa and type of samples, and the decision is often made based on anecdotal evidence and/or advice from the published literature,. One of the first considerations before embarking on a microbiota project is to select a sequencing technology. Traditionally, the most common options are Roche 454 GS-FLX, the Illumina MiSeq (lower output, longer reads) and HiSeq (higher output, shorter reads) and the Ion PGM, each offering a series of advantages and disadvantages (see <http://www.molecularecologist.com/next-gen-fieldguide-2014/> for a guide). Both the Illumina and Ion instruments utilise a sequencing by synthesis approach where Illumina use DNA templates immobilised on glass slides and optical detection of fluorescently-labelled nucleotides, whereas templates for the Ion Platforms are immobilised in wells on a semi-conductor chip followed by electrical detection of released hydrogen ions. The Illumina and Ion technologies have been compared for amplicon sequencing using various sampling environments, variable regions of the 16S rRNA gene and analysis pipelines. In one case, when stringent quality filtering and lower sequence similarity cut-off when clustering operational taxonomic units (OTUs) were applied on V4 reads sequenced, negligible differences in alpha and beta diversities were observed within and between soil samples when comparing the MiSeq and the PGM. This concordance was further supported when comparing MiSeq and PGM derived microbiota composition as determined by sequencing V1-V2 amplicons generated using a 20-species mock community and human-derived samples. In the latter case it should be noted that, some significant differences were attributed to the PGM failing to produce full-length reads for certain organisms. Furthermore, while not comparing amplicon sequencing and using relatively early versions of sequencing chemistry on an isolated *E*. *coli* species, Loman and colleagues found MiSeq to have lower error rates and longer reads than the PGM, which on the other hand had the fastest turn-around-time. Comparative studies were also conducted to assess the initial potential of the MiSeq to replace the Roche 454 GS-FLX, while also evaluating the effect of the variable region studied. Kozich and co-authors established a dual-index barcoding approach suitable for variable MiSeq read lengths and amplicon regions, in particular V3-V4, V4 and V4-V5 regions. In terms of read quality, MiSeq was either comparable or better than the GS-FLX Titanium, and the V3-V4 better than the V4-V5 region. Another study compared amplicon sequences of seven tandem variable regions produced by the GS-FLX Titanium and Illumina GAII (predecessor of HiSeq) and showed the V3-V4 and V4-V5 primer combinations performed worst and best in terms of classification accuracy, irrespective of the technology used. It is clear that the choice of primers can have a major effect on the outcome, which was also further substantiated by Tremblay and co- authors, as the V6-V8 or V7-V8 regions returned taxonomic composition from a synthetic community that differed to higher degree than what the V4 region did. With the ever increasing number of technological variables that have the potential to have non-trivial effects on microbiota composition analysis, it is critically important to maintain a consistent methodology within studies and when comparing studies, or to have evidence that any inconsistencies that exist do not bias results. A more expensive alternative to 16S rRNA gene amplicon sequencing is shotgun metagenomic sequencing, which bypasses gene-specific amplification and potentially sequences all fragmented DNA, including that from other microorganisms and viruses, in a community. While providing much more information, including encoded functions of the microbiota, the vast amount of sequence data obtained however leads to a new set of challenges in terms of data processing, storage and analysis. For instance, the Illumina HiSeq 2500 platform can yield over 1,000,000,000,000 bp (1 Tbp) of raw sequence data, which may increase several-fold during downstream processing and analysis. Shotgun sequencing is also possible using both the Illumina MiSeq and Ion PGM albeit with less throughput compared to HiSeq. Some non-metagenomic studies have evaluated these platforms and demonstrated comparable results when used to detect blood pathogens, diagnose dementia, and detect gene variants across four microbial genomes. In the current study we investigated the impact of various amplicon primer combinations and sequencing technologies on the analysis of complex microbial communities. More specifically we compared amplicon and shotgun data generated by Illumina MiSeq, HiSeq and Ion PGM through the use of six human stool samples using two primer sets covering two different 16S rRNA gene regions (V1-V2 and V4-V5). We also assessed the depth requirements for analysing stool shotgun datasets, and thus if the MiSeq and/or PGM represent suitable alternatives to the HiSeq. # Materials and Methods ## 16S rRNA gene amplicon sequencing Stool samples were collected from six elderly individuals and stored at -80°C during the ELDERMET project, approved by the Cork Clinical Research Ethics Committee of the Cork Teaching Hospitals (CREC), which granted full approval on the 19th February 2008 (Ref: ECM 3 (a) 01/04/08). Formal written consent was obtained at the time of recruitment, on the basis of an Information Sheet/Safety Statement, following an ethics protocol that was approved by CREC in compliance with pertaining local, national and European ethics legislation and guidelines to best practice. DNA was extracted from stool samples using previously described methods, together with a modified Qiagen DNA extraction procedure. Briefly, DNA was extracted using a QIAamp DNA stool Kit with the addition of an initial bead beating step. Microbial DNA from stool samples was used as template for PCR, which contained 25μl Biomix Red (MyBio, Kilkenny, Ireland), 1 μl forward primer (Sigma Aldrich, Dublin, Ireland) (10pmol), 1 μl reverse primer (Sigma Aldrich) (10pmol), template DNA and PCR grade water (MyBio), to a final reaction volume of 50μl. Conditions were optimised so that only 1 band of the correct sizes was obtained and all PCR were completed in triplicate (see for primers and further details). Triplicate PCR products were pooled and cleaned using AMPure magnetic bead purification system (1:1.8 DNA:AMPure ratio) (Beckman Coulter, UK). Cleaned samples were quantified using Picogreen Quant-iT quantification and the Nanodrop 3300 (Fisher Scientific, Dublin, Ireland). Samples were subsequently pooled in an equimolar concentration of 10pM and prepared for MiSeq sequencing using standard Illumina protocols. Libraries were mixed with Illumina generated PhiX (20% of 12.5pM) control libraries and were denatured using freshly prepared NaOH and sequenced using a V3 600-cycle kit. For the PGM, libraries were pooled at a concentration of 10pM and sequenced according to Ion PGM protocols. ## Metagenomic shotgun sequencing For Illumina MiSeq shotgun sequencing, samples were initially tagmented, whereby the Nextera Transposome with sequencing adaptors combines to template DNA resulting in fragmentation of the DNA and the addition of adaptors using the Nextera XT kit from Illumina. A limited 12-cycle PCR was completed during which time sequencing adaptors and indexing primers were added to the DNA. Amplicon samples were then normalized and pooled, followed by sequencing on the MiSeq platform using Illumina protocols for a 2 x 300 cycle run, with an insert size of 400 bases. Shotgun libraries for Ion PGM were generated according to instructions from the ‘Ion Xpress^™ Plus gDNA Fragment Library Preparation’ User guide (Publication number MAN0007044). Libraries were sheared, size selected and individually barcoded using the Ion Xpress Barcode Adapters. Following library quantification and equimolar pooling, the Ion OneTouch^™ 2 system was used to prepare template positive ion sphere particles containing the clonally amplified DNA libraries using the ION PGM^™ Template OT2 400 Kit, allowing up to 400 bp single-end reads. Enrichment of the template positive ISPs was performed using the Ion OneTouch^™ ES and an enrichment percentage of 18% was obtained, which was within the range recommended in the ION PGM^™ Template OT2 400 Kit guide (Publication number MAN0007218). Sequencing was performed on the Ion PGM using an Ion 318v2 chip and the Ion PGM Sequencing 400 kit (guide number MAN0007242). Shotgun Illumina HiSeq sequencing reads were obtained from the published ELDERMET dataset. The paired-end read lengths were 2 x 90 bp with an insert size of 300 bases. DNA was extracted from samples using the same method as used above. ## Bioinformatic analysis MiSeq reads were merged and filtered using *join_paired_ends*.*py* in QIIME version 1.8 using the *fastq-join*.*py* tool, whereas the single-end PGM reads were not. Demultiplexing of both MiSeq and PGM reads was carried out using *split_libraries*.*py* also on QIIME with default parameters retaining only reads matching the main length distributed per primer and with an average quality score of Q25 or above. The differences in quality filtering lengths is due to reverse primers being present in the MiSeq reads. Chimeric sequences were removed via USEARCH version 7.0.1090 using the *uchime_ref*.*py* command along with the ChimeraSlayer GOLD database. OTUs were clustered using the QIIME script *pick_closed_reference_otus*.*py* and the RDP database version 11.4. The Mothur implementation of the RDP classifier was used to assign taxonomy from phylum to genus with a bootstrap cut-off of 80%. Any sequences with less than 80% bootstrap values were assigned as unclassified at that particular rank. Species counts for amplicon data were generated using SPINGO with default parameters. All three shotgun datasets reads were aligned to the human genome version 20 (hg20) to filter out human-derived sequences using Bowtie2 version 2.2.3. Illumina HiSeq and MiSeq reads were subsequently quality filtered and trimmed using Trimmomatic version 0.32 and only allowing a quality PHRED cut-off score of at least Q22 across a sliding window of 20 bp. Reads with a minimum length of 30 bp were also removed. Only PGM reads with a quality score of greater than Q15 and longer than 30bp were retained for downstream analysis. All metagenome assemblies were performed using IDBA_UD version 4.1.2 and MetaVelvet version 1.2.02. Phylogenetic binning was achieved using MetaPhlAn version 2, Kraken version 0.10.5-beta and GOTTCHA version 0.7.5. MetaPhlAn2 classifies sequences via clade-specific marker genes, Kraken uses exact alignment of *k*-mers and a lowest common ancestor approach, while GOTTCHA maps reads to non-redundant signature databases to classify at multiple taxonomic levels. Genes were predicted using MetaGeneMark version 3.26. Metaphor was used to predict core and unique genes with thresholds set to 30% amino acid identity across an alignment covering 50% of both sequence lengths. The core and unique genes were then mapped against the EGGNOG database version 4 using BLAST to create functional profiles for each of the samples and datasets retrieving the top hit with an E-value of 1e-5. ## Statistics All statistical analysis was performed in R version 3.1.3. In each of the heatplots, Spearman correlations, along with Ward D2 clustering, were performed on the relative abundance at genus level of each sample. As the data was largely non-parametric, Spearman correlations were chosen to prevent breaking the statistical assumptions of Pearson correlations. A Mann-whitney test was used to analyse differences in the taxa between clusters. Where necessary, the P-values were corrected for multiple testing using Benjamini and Hochberg. A P-value of \<0.05 was considered significant. # Results ## Microbiota composition The data generated reflected the different outputs of the three platforms. For the amplicon datasets the PGM produced 57,720 (mean) ± 9,841 (SD) V1-V2, and 33,454 ± 10,488 V4-V5 reads per sample, respectively, while the MiSeq produced 181,758 ± 108,343 V1-V2, and 102,824 ± 22,154 V4-V5 reads per sample, respectively. For the shotgun datasets there was also a marked difference between the three sequencing technologies, with 26,590,475 ± 51,650 HiSeq, 1,352,748 ± 458,483 MiSeq and 962,226 ± 170,251 PGM reads were generated per sample, respectively. We performed hierarchical clustering analysis on the microbiota composition of all six stool samples in order to assess the effect of the amplification primer combination (where relevant), sequencing strategy (16S rRNA gene or shotgun), sequencing technology and type along with metagenomic read classifier. shows a heat-plot with hierarchical clustering of the proportional taxonomic abundances at the genus level, with only genera in a minimum of 20% of the datasets included. All shotgun datasets fell into one large cluster with three distinct sub-clusters, labelled 2, 3 and 4. It is worth highlighting that although the shotgun samples clustered together, there were major discrepancies between the taxonomic profiles (sub-clusters) dictated by the metagenomic classifier used with one exception, sample 6 sequenced on the PGM and classified by GOTTCHA, which clustered with the MetaPhlAn2 sample 6 datasets. In the MetaPhlAn2 cluster (cluster 4), the datasets grouped by sample in each case, which is preferable as it suggests the technical variation is less than the inter-individual variation. For all six samples, the HiSeq and MiSeq datasets clustered together while the PGM sample was located to the side of the sub-cluster. For the GOTTCHA classifier, datasets grouped by sequencer more than by sample. Here there were no case where all three shotgun technologies clustered together by sample. For the third shotgun classifier, Kraken (cluster 2), five of the six samples clustered by sample with the exception of the MiSeq dataset for sample 2. Unlike MetaPhlAn2, the PGM formed sample-wise sub-clusters with HiSeq or MiSeq, with the two Illumina technologies not forming any sub-clusters. Out of a total of 163 genera, 23 were statistically significant between cluster 3 (GOTTCHA) and 4 (MetaPhlAn2) in where the most significant genera included *Ruminococcus* (increased in cluster 3; P-value = 9.88 x 10⁻⁰⁵), *Blautia* (increased in cluster 3; P-value = 1.30 x 10⁻⁰⁵) and *Campylobacter* (increased in cluster 3; P-value = 9.30 x 10⁻⁰⁶). When comparing Kraken, cluster 2, to the other two shotgun classifiers (cluster 3 and 4) there were 52 statistically significant different genera. These included *Buchnera*, *Cellulomonas* and *Cellvibrio*, all increased in the Kraken dataset each with an adjusted P-value of 1.82 x 10⁻¹¹. Of the 15 most significantly different genera, all but one were absent from the GOTTCHA and MetaPhlAn2 clusters, thereby indicating possible false positives detected by Kraken. The three aforementioned taxa are also not predominant colonisers of the human gut thus reinforcing the possibility of inaccuracies in Kraken assignments. See for a full list of taxonomy comparisons. For the amplicon datasets, sample-wise clustering was less prevalent than for the metagenomic datasets. MiSeq V1V2 amplicons were contained in a distinctive sub-cluster, contained within the cluster labelled 1 in , clearly separated from the rest of the amplicon datasets. A second sub-cluster contained all the sample 3 and 6 amplicon datasets, with the exception of the V4V5 Miseq dataset and the aforementioned V1-V2 MiSeq dataset. The third sub-cluster contained the majority of the V4V5 MiSeq samples (4 of 6) along with two V4V5 PGM samples. In this case the amplicons clustered by 16S rRNA gene primer combination, as opposed to by sample or by technology. The final sub-cluster contained the majority of the V1V2 PGM datasets (4 of 6) along with 3 of the 4 sample 5 datasets (V1V2 MiSeq being the missing dataset). Investigating the differences between cluster 1 (amplicon data) and clusters 2–4 (shotgun data), uncovered 91 genera to be statistically significant, therefore showing the large differences between amplicon and shotgun classification methods of reads. The full list of taxonomy comparisons are found in. As for bacterial taxa that were the most abundant across all of the datasets, there were some families that differentiate the six subjects regardless of methodology used: For example, *Porphyromonadaceae* genera were consistently high in Sample 6 datasets compared to the other samples, and so were genera belonging to the *Prevotellaceae* family in Sample 3, irrespective of primer combination or sequencing technology. For samples 1 and 5 the shotgun-based methods appeared more sensitive with respect to detecting *Enterobacteriaceae* genera within the *Proteobacteria* phylum compared to the amplicon-based approaches, which could be attributed to the difficulty of discriminating such taxa at 16S rRNA gene level. also highlights the number of unique species in each dataset, as identified by MetaPhlAn2 for shotgun data and SPINGO for amplicon data. Note that these were species that could be confidently classified as such, and should not be mistaken as number of unique OTUs. The highest numbers of unique species among all shotgun methods were detected in the HiSeq datasets, comparable to those resulting from the analysis of amplicons. The success of the HiSeq with respect to shotgun sequencing is not surprising given the greater sequencing depth it can provide resulting in detection of rarer species. The lowest number of unique species overall was detected in the MiSeq shotgun datasets, which is not due to total number of reads as PGM had fewer of these. For the amplicon datasets, the highest number of unique species was detected with the PGM datasets for five of the six datasets. Although the species counts for the pooled PGM amplicons was higher when compared to the MiSeq amplicons, the difference was not statistically significant (P-value = 0.24). However, when comparing particular primer combinations, the difference in the V1-V2 species counts between the two technologies was significant at the 10% level (P-value = 0.093). We further analysed the effect of varying sequencing depth on the number of unique species detected for each amplicon run. The highest numbers of species were detected at each read depth by the V1V2 amplicon on the PGM, while the lowest was the V1V2 on the Illumina MiSeq. All primer datasets reached saturation in the number of new species detected, other than the V4V5 primer on the PGM which was limited by the number of reads for some samples. However, despite this, more unique species were detected with this primer/technology combination than both MiSeq datasets, which had vastly more reads. ## Shotgun sequencing depth To investigate which technology was most suitable for shotgun sequencing, we performed random subsampling of reads to determine occurrences at even sequencing depths, in recognition of the fact that the HiSeq coverage was substantially higher than the coverage for MiSeq and PGM. shows the median N50 values across each of the six samples per technology, including three replicates (random sub-samplings) for each sample. At the lowest sequencing depth selected (150,000 reads) the assembly using the MiSeq data had the highest N50 (minimum contig length above which 50% of all reads are assembled into), possibly due the longer read lengths. However, as more reads were added, the HiSeq data began to outperform the assembly from both the MiSeq and the PGM technologies. The MiSeq and PGM datasets became limited by read number and their N50 value plateaued at 1.7 million and 950,000 reads, respectively. Due to the large number of HiSeq reads, the N50 peaked at 10 million reads after a large increase at 1.7 million reads. Two of the six HiSeq datasets (Samples 1 and 5) had a very large N50 at 600,000 reads. In order to ensure that the results were not affected by the assembler selected, the datasets were also assembled using both Velvet and MetaVelvet. Interestingly, the same two samples for the HiSeq datasets had an elevated N50 for both Velvet and MetaVelvet, however at 1.3 million and 950,000 reads respectively. Furthermore, unique species detection was also performed on the sub-sampled shotgun sequencing-derived reads. At low sample depths the HiSeq, MiSeq and PGM datasets were comparable with few differences in the number of species detected. At 950,000 reads, the PGM data reached the read limit, but was still similar to the other technologies in terms of number of species. However, at 1.7 million reads, the HiSeq species counts continued to increase while the MiSeq counts level off. This could possibly be due to the fact that the longer MiSeq read lengths result in more accurate species assignments relative to HiSeq, leading to earlier plateauing. In the overall graph ( insert) the HiSeq counts continued to increase without levelling off completely even at the 25 million read point. ## Encoded functions From within the categories of shotgun datasets, the core and unique genes were predicted using Metaphor. This was carried out on 600,000 reads per dataset in order to allow for comparative results at equal sequencing depth. For the core genes all three technologies gave broadly the same results, however the HiSeq data had the most poorly characterised genes out of the three datasets, along with the lowest number of genes with a “Metabolism” function and the highest with no function. Surprisingly, this technology did not predict any core genes for the categories, “Energy Production and Conversion” or “Inorganic Transport and Metabolism”, whereas both of these categories were present in the core gene profiles of the MiSeq and PGM datasets. The MiSeq datasets predicted the highest number of genes within the “Metabolism” category, while the PGM data predicted the highest for “Information Storage and Processing”, whilst also being the only technology to predict core genes in the category “Cell Motility”. The number of genes predicted by MetaGeneMark are listed in. At a read depth of 600,000 sequences, the MiSeq datasets predicted the most genes for each of the 6 samples while the HiSeq datasets gave the lowest gene number of 5 of the 6 samples. This is a possible reason why this technology gives the most detailed core and unique gene profile. # Discussion The NGS technologies Illumina MiSeq, HiSeq and Ion PGM have shown significant promise in delivering cost-effective, high-resolution insights into microbiomes from various environments. However, due to a multitude of technical variables, careful comparisons are required to provide recommendations for suitable methodological approaches. In response to this, we compared the taxonomic composition of six stool samples using two different primer combinations covering two 16S rRNA gene variable regions. We then compared these results with those of shotgun sequencing using Illumina and Ion technologies. Following either OTU clustering of amplicon reads or taxonomic classification by binning of shotgun reads, all at genus level, we compared microbiota composition of the different datasets. Even though the gut microbiota is generally regarded as individual specific, it was apparent that some amplicon datasets clustered according to technology and/or primer set, rather than by subject. In particular, microbiota composition from all V1V2 MiSeq and four of the six V4V5 MiSeq datasets grouped together in separate sub-clusters. The V1V2 and V4V5 PGM datasets clustered by sample opposed to technology in 3 of the 6 samples (samples 1, 3 and 6) while the V4V5 MiSeq data clustered with V4V5 PGM data per sample in 2 of the 6 samples (samples 5 and 6). To ensure that the differences in classifications between shotgun and amplicon sequencing were not simply due to a particular shotgun classification method, we compared the compositional clustering with three classifiers of shotgun reads, MetaPhlAn2, GOTTCHA and Kraken. The shotgun datasets grouped together in a sub- cluster separated from the amplicon datasets, which might be expected as these methods are independent of amplification bias and 16S rRNA gene copy number differences. With MetaPhlAn2, all Illumina HiSeq and MiSeq datasets were consistently closer to each other than to the PGM shotgun sequences. This is seen to a smaller degree with GOTTCHA, where three of the six samples sub- clustered the Illumina technologies, but not at all for Kraken assemblies. In terms of clustering by sample over method, MetaPhlAn2 gave the most optimal results with all datasets clustered by sample groups, closely followed by Kraken where this occurred for 5 of the 6 samples in separate sub-clusters. GOTTCHA failed to cluster any dataset by samples, indicating its higher sensitivity for technological artefacts between sequencing methods. However, it must be noted that measuring accuracy based on individual sample clustering is not always a reflection of performance, as GOTTCHA datasets clustered more closely to MetaPhlAn2 and although sample clustering is observed when using Kraken, many of the taxonomic assignments may be false positives as previously mentioned. Unsurprisingly, Illumina HiSeq shotgun sequences translated to the highest number of species, compared to the other two shotgun datasets, which were more than an order of magnitude smaller. Sub-sampling that simulated lower HiSeq coverage revealed, however, that even equal number of reads could result in more observed species for HiSeq. As this technology produces shorter reads compared to MiSeq and PGM it is possible that the number of species is artificially inflated as a result of higher sequence variation created from incorrect alignment to the reference marker genes. While not directly comparable with species observed through shotgun sequencing, V1-V2 amplicons, which are expected to be more variable than V4-V5 amplicons, sequenced by PGM resulted in the highest species counts. Despite having the largest number of reads per sample, the V1-V2 region on the MiSeq had at each subsampling point the lowest number of unique species identified. This could be due to the questionable reliability for this primer combination in relation to unexpected clustering and failure to detect expected genera. Curiously, Salipante *et al*., found that sequencing using the same V1-V2 primers on the PGM led to higher error rates when compared to the MiSeq, particularly for a mock community of 20 organisms where deviating abundances of single strains have much greater effect on the overall community composition than in a high-diversity sample. Other reasons for the different results in Salipante et al. study could be attributed to discrepancies in amplification (one-step PCR reaction) and taxonomic assignment (older RDP-classifier version and BLAST). The benefits to using metagenomic shotgun over amplicon sequencing are clear in terms of increased information content and reduced biases related to amplification and gene copy numbers. However, it is currently not established what sequencing depth is required for the different technologies; this is a more pertinent issue for shotgun than for amplicon sequencing, due to its much higher cost per sample. We therefore assembled the randomly sub-sampled shotgun datasets and compared the common N50 metric across the three sequencing technologies. As expected, the MiSeq technology, with its non-overlapping 300 bp paired-end reads, had marginally higher N50 values than HiSeq and PGM. An N50 peak occurred at 10 million reads for the HiSeq data suggesting that this was the optimal point for sequencing depth for stool samples and 100 bp paired-end reads with 300 bp insert size. There was no peak observed for the PGM or the MiSeq in the available coverage range, which may suggest that the coverage may not be sufficient to reach an optimal level of assembly. Somewhat surprisingly, for two of the six samples there were drastically elevated N50 values at 600,000 HiSeq reads, irrespective of which random sub-sampling set. Such early N50 peaks were also observed using two other assemblers, albeit for a different number of reads, and has previously been reported when assembling sub-samples of an isolated bacterium. In that case, the authors reasoned that this could be due to chimeric reads, duplications or sequencing errors, and recommended that assembled contigs should be incrementally assembled in sub-sections before a final merge. We also suggest that for our data, this read depth may be where the majority of high abundant species are assembled and as more rare taxa are added the assembly becomes less efficient. In terms of functional categorisation of assembled shotgun sequences, we found the MiSeq and PGM datasets to largely contain equal proportions of predicted core genes from the assembled contigs. For the HiSeq assemblies there were, however, substantially fewer core genes involved in “Metabolism” and more genes with unknown function. This may be attributable to the fewer number of predicted complete genes, which is plausible for this shorter-read technology. To summarise, this is, to our knowledge, the first reported study comparing both amplicon and shotgun sequencing for Illumina and Ion technologies. Although shotgun sequencing did not suffer from the same degree of technology-dependent bias seen with the amplicon sequencing, there were some major distinct differences between phylogenetic binning software, with MetaPhlAn2 producing the most favourable results. GOTTCHA failed to cluster any datasets by sample, however sub-clustered with MetaPhlAn2, while Kraken clustered separately from the other two binners and also appeared to produce a high number of false positive taxonomic assignments. The variation of microbiota composition between the majority of gut samples proved to be lesser than between the compared sequencing technologies and variable 16S rRNA gene regions. In particular, the V1-V2 MiSeq showed poor performance, while the V4-V5 region was marginally more reliable on both platforms. There is evidence that the MiSeq and PGM offer valuable information when used for shotgun sequencing, however, in order to detect the majority of species in samples and to perform a high quality assembly, deeper sequencing is required. Species assignment is also dependent on read length, which is shorter for the HiSeq. We subsequently showed that there may be no assembly-related benefit in sequencing greater than 10 million HiSeq reads per stool sample. Nevertheless, as the cost of shotgun sequencing is lower on the HiSeq instrument compared to MiSeq or PGM, this platform may still be preferable even though MiSeq produces longer reads and somewhat better assemblies at low sequencing depth. Caution should however be applied with regards to taxonomic binning, and comparisons such as those described in this study must be carried out to prevent methodological biases eclipsing the true biological picture. Hence, we advise laboratories with particular interests in certain microbes to optimise their protocols to accurately detect these taxa using different techniques. # Supporting Information The authors wish to thank Dr. Fiona Crispie and Ms. Vicki Murray for their extensive help with the sequencing in this study. This publication has emanated from research supported in part by a research grant from Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2273 and 11/PI/1137 and by FP7 funded CFMATTERS (Cystic Fibrosis Microbiome-determined Antibiotic Therapy Trial in Exacerbations: Results Stratified, Grant Agreement no. 603038). [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: PC MC AC FF. Performed the experiments: AC FF. Analyzed the data: AC FF. Contributed reagents/materials/analysis tools: PC MC. Wrote the paper: AC FF AOD CS RS PC MC.

# Introduction Depression, one type of non-communicable diseases, has triggered great public health concerns in both developed and developing countries due to its high prevalence and heavy burden of diseases: it is one of the most prevalent disorders, affects nearly 350 million people, and accounts for 12.7% of all- cause mortality around the globe. Moreover, it leads to other health problems such as type 2 diabetes, cardiovascular diseases, and suicide. These secondary comorbidities further affirm the tremendous burden that depression accounts for. China, home for nearly 18% of the global population, encounters a great threat posed by depression: with nearly 17% of the global disease burden of psychiatric-related illness, of which 30% can be attributed to depression. The figure is even more threatening in working age adults: 43% of mental disorder related disease adjusted life years (DALYs). Considering both the pattern that depressive symptomology increases as one ages and the current trend of population ageing in China, the threat will possibly be enlarged. Conventional interventions such as cognitive behavioral therapy and antidepressants have been recognized as effective means for depression treatment, nevertheless, many are confined to a small scale of population due to their high demands for professional resources. Similar to other low and middle income countries (LMICs), China also suffers from a profound gap between need and supply in mental health services. This suggests that the above-mentioned approaches can hardly be an effective approach to solve the large-scale problem facing China nor other LMICs, and therefore, novel approaches with greater accessibility are urgently needed. In recent years, emerging evidence started to regard Physical Activity (PA) as an important solution, considering its low demand for professional resources and contribution to one’s physical health (such as health function and sleep duration) and emotional social support), which are closely related to one’s depressive symptoms. In this case, various organizations have published recommendations on PA. For instance, WHO suggests older adults taking at least 75 minutes of Vigorous Physical Activity (VPA) or at least 150 minutes of Moderate Physical Activity (MPA) or an equivalent combination of moderate-to- vigorous PA (MVPA) each week; The American Heart Association further encourages the elderly to spend at least 30 minutes per day on 5 days of the week on MPA or at least 20 minutes on 3 days of the week on VPA. However, the exact role of PA on depression awaits statistical validation. Among the limited number of studies that attempted to investigate the protective role of PA on depression, findings turned out to be largely heterogeneous. Some studies generally concluded with PA’s protective effect on depression; other studies have further investigated PA from various dimensions, including intensity, duration and volume, and resulted in more complex findings: some indicated that individuals who engaged in VPA/MPA or PA with higher frequency had lower odds of depression; nevertheless, some suggested that LPA rather than VPA/MPA, lower frequency (1–2 times/week), or PA with smaller volume than recommended was protective. Moreover, some indicated that the association differed significantly with gender. One study concluded with no statistically significant associations between depression and PA of any kind, and one even found that people who engaged in high volume of PA were more likely to suffer from depression. The lack of consensus regarding this issue on the one hand indicates the association between PA and depression depends largely on the specific cultural and demographic contexts, and on the other hand, suggests that PA is a multidimensional construct that consists of intensity, frequency, duration and volume, and the relationship cannot be fully comprehended by a single dimensional analysis. Moreover, Chinese citizens are greatly influenced by Eastern culture and present distinctive patterns of PA: East Asians tend to take fewer physical activities and at lower intensity than those in Western countries. Thus, whether the PA recommendations that have been designed and subject to Western population fit Eastern context and how PA is associated with better mental health in middle- and older-aged Chinese residents still awaits further investigation. To the best of our knowledge, no study has examined the correlation between depression and PA in China by taking into consideration all four key dimensions of PA (intensity, frequency, duration and volume). To this regard, we conducted this study on middle- and older-aged Chinese with two objectives: 1) to understand the patterns of PA; 2) to measure the associations between depression and PA from various aspects including intensity, frequency, duration and volume. # Materials and methods ## Study sample We extracted our data source from the China Health and Retirement Longitudinal Study (CHARLS 2015) Wave 4 conducted in 2015. CHARLS is a nationwide survey program covering 450 villages and 150 counties in 28 provinces with an aim to provide comprehensive and quality data on the demographic background, family characteristics, health status, work, and retirement status of the mid-aged and older residents in China. The data were collected using a four-stage, stratified, cluster sampling method. Detailed sampling technique was documented in Zhao et al.’s study. The CHALRS research team has obtained ethical approval from the institutional review board (IRB) at Peking University. Amongst 20,967 surveyed community residents, 9,118 participants were selected for this cross- sectional study according to the following criteria: 1) aged 45 and older, 2) gave information on whether or not they had conducted any type of PA (VPA/MPA/LPA), 3) were assigned a case weight to balance the sample with the source population. ## Variables ### Outcome variables Depression: a 10-item Center for Epidemiology Studies Depression Scale (CES-D 10) was used to screen depression. The answers for CES-D 10 are on a four-scale metrics coding from 0 to 3. The total score ranges from 0 to 30, with higher scores indicating more depressive symptoms. The CES-D 10 has been used in previous studies and showed good reliability (Cronbach’s alpha = 0.815). Several studies have reported a cut-off point of 12 with good validity to identify clinically significant depression. In this case, participants with a CES-D 10 score of 12 or above were regarded as in risk of depression. ### Main explanatory variables CHARLS classified PA into 3 levels according to intensity: 1) VPA: activities that require hard/high intensity physical effort and make one breathe much harder than normal (e.g. heavy lifting, digging, aerobics, fast bicycling, cycling with a heavy load, etc.); 2) MPA: activities that can result in breathing somewhat harder than normal (e.g. carrying light loads, bicycling at a regular pace, mopping the floor); 3) LPA: walking, including walking at work, travelling from place to place, and walking for recreation. Participants were asked: 1) did they conduct VPA/MPA/LPA for at least 10 minutes continuously for a usual week; 2) if yes, how many days they normally took VPA/MPA/LPA in a week; and 3) how much time (\<30min, \<2h, \<4h, ≥4h) they spent on PA each time. Frequency: answers for VPA/MPA/LPA ranged from 0-7days/week, and were categorized into never (0 days)/ 1–2 days/week/ 3-5days/week/ 6-7days/week. Duration: this study took into consideration that some participants did not took any PA, and therefore, categorized this variable into 5 scales, including no PA/\<30 min per time / 30–119 min per time / 120–239 min per time / ≥240 min per time. Volume: According to the WHO international guideline for PA, at least 75min/week VPA, or 150min/week MPA, or 150min/week MVPA was recommended for the elderly. In addition, WHO has also warned that VPA/MPA over 300min/week may be harmful to the senior population. In this case, we calculated the total length of VPA/MPA/LPA by: 1) estimating the daily duration of VPA/ MPA/ LPA using the average value of each time; 2) using the following formula: $$Volume\ of\ VPA/MPA/LPA = Frequency\ of\ VPA/MPA/LPA\ *\ Duration\ of\ VPA/MPA/LPA$$ Regarding volume of VPA, responses were classified into 4 groups: no VPA/ \<75 min/week \<300 min/week/ ≤300 min/week; whereas regarding total length of MPA, responses were classified into 4 groups: no MPA/ \<150 min/week/ \<300 min/week/ ≥300 min/week. Due to the lack of international guidelines for LPA, we adjusted total length of LPA by 5 levels: no LPA/ 1^st quartile (≤105 min/week)/ 2^nd quartile (≤525 min/week)/ 3^rd quartile (≤1260 min/week)/ 4^th quartile (\>1260 min/week). In terms of MVPA, The score was computed by multiplying the length of VPA and MPA by an assigned metabolic equivalent value (MET): MPA = 4 MET, VPA = 7.5 MET. Since the lower and higher cut-off points for the sufficient MVPA are 150 min/week MPA and 300 min/week VPA, which equal to 600 MET and 2250 MET, respectively. Therefore, the volume of MVPA was then coded as no MVPA/ not sufficient (\<600 MET)/ sufficient (600–2249 MET)/ overlong (≥2250 MET). ### Control variables This study included socio-demographic variables, health behaviors and health status related variables for adjustment. Socio-demographic variables included: gender (male, female), age (45–54, 55–64, 65–74, ≥75), residency (urban, rural), education (illiterate, primary school and lower, middle school, high school and higher), marital status (married/cohabitating, single), living near children (no/ yes), total annual household income (1^st quartile (≤\$232/y), 2^nd quartile (≤\$1279/y), 3^rd quartile (≤\$5338/y), 4^th quartile (\>\$5338/y)). Health behavior-related variables included: alcohol intake (no, yes), and smoking (no, yes). Health status-related variables included: BMI (underweight (\<18.5 kg/m²), normal (18.5–24.99 kg/m²), overweight (25–29.99 kg/m²), obese (≥30 kg/m²)), diabetes (no, yes), and hypertension (no, yes). ## Data analysis Frequencies and percentages were calculated for descriptive data. Associations between levels of PA and outcome variables were assessed by binary logistic regression with the potentially confounding variables being adjusted. Adjusted Odd Ratios (ORs) and their 95% Confidence Intervals (95% CIs) were presented as measures of effect. Case weights (provided by the CHARLS study) were used to adjust for the stratified sampling method and non-response patterns. Data were analyzed using R Version 3.5.1. # Results ## Characteristics of the participants Basic characteristics of the whole sample population, as well as of those who were with or without risk of depression are shown in. Of the 9,118 participants, a greater proportion were female, in the age group of 45–74 years, rural residents, with a primary school or lower level of education, partnered and living near children. The majority of respondents did not consume alcohol or smoke; have not been diagnosed as diabetes or hypertension; and were either in normal weight or overweight. According to CES-D 10 scale, over one fourth of the participants were in risk of depression. Generally speaking, participants with depression were more likely to be female, aged 65 and older, in rural areas, with a primary school or lower level of education, with no partner, with lower family income (first two quartiles), non-drinker, non-smoker, with diabetes or hypertension, and in normal weight or underweight. ## Frequency of PA in men and women Generally speaking, the proportion of respondents taking PA increased as the PA intensity decreased in both men and women (31.30% for VPA, 55.04% for MPA and 81.14% for LPA). A phenomenon that the majority of participants either took no PA or had frequent PA (6-7days/week) was observed in all three levels PA. Despite that in LPA, the distribution differed between men and women in VPA and MPA: A larger proportion of men taking part in VPA whereas a significantly larger proportion of women involved in MPA, especially in frequent MPA. ## Duration of PA in men and women Amongst those who took VPA, the majority of men and women spent more than 4 hours per day. Different patterns between men and women were observed in MPA duration: amongst those who took MPA, a significantly larger proportion of women spent 30–119 minutes on MPA each time, whereas the distribution was more evenly in men. Women and men shared similar patterns regarding LPA duration. ## Volume of PA in men and women summaries the total length of time the participants spent on VPA, MPA, LPA and MVPA each week. Similar to, the distribution was heavily bipolarized in VPA, MPA and total MVPA: there was a large proportion of participants who did not take any PA, whereas at the same time, a considerably large proportion spent much more time on PA than they were recommended to. A significantly larger proportion of men spent over 300 min/week on VPA whereas a significantly larger proportion of women spent over 300 min/week on MPA. In total, over 58% men and women took sufficient amount of MVPA as they were recommended to. ## Relationship between PA and depression in men and women outlines the associations between different dimensions (intensity, frequency, duration and volume) of PA and risk of depression in the whole sample population, and compares the differences between men and women. Regarding frequency, the participants who spent 3 days or more on VPA had significantly higher risk of depression (OR = 1.57, 95% CI: 1.18, 2.08 for 3–5 days/week and OR = 1.24, 95% CI: 0.99, 1.54 for 6–7 days/week, respectively) than those who did not take VPA. The association was statistically significant in men but not in women. However, spending 1–2 days/week on MPA was associated with lower risk of depression in women (OR = 0.58, 95% CI: 0.36, 0.91), whereas spending 6–7 days/week was associated with 1.45 times higher likelihood (95% CI: 1.05, 2.00) of depression in men. No significant association was observed between any level of LPA frequency and depression in the whole sample population or subgroups. Regarding duration, spending over 4 hours on VPA each time was correlated with greater odds of depression (OR = 1.39, 95% CI: 1.12, 1.74). The association was strong in men (OR = 1.65, 95%CI: 1.18, 2.32) but not statistically significant in women (OR = 1.24, 95% CI: 0.91, 1.64). Spending less than 30 min/day on MPA was associated with better mental health status in women, however, men who took 30–119 min/day MPA had 1.57 times higher risk (95% CI: 1.06, 2.33) of depression compared to those who reported not being involved in MPA. No significant association between any level of LPA duration and risk of depression in the whole sample population or subgroups was observed. Regarding volume, spending more than 300 min/week on VPA was associated with greater odds of depression (OR = 1.31, 95% CI: 1.08, 1.59) in the whole sample and in men (OR = 1.65, 95% CI: 1.22, 2.24). Similar patterns were observed in MVPA. Taking 150–299 min/week MPA was associated with smaller odds of depression in the whole sample population (OR = 0.65, 95% CI: 0.41, 1.02) and in women (OR = 0.49, 95% CI: 0.28, 0.87). Nevertheless, men taking 300 min/week or more on MPA was associated with higher likelihood (OR = 1.33, 95% CI: 0.98, 1.81) of depression. # Discussion To the best of our knowledge, the study is the first one to examine the risk of depression in relation to the intensity, frequency, duration, and volume of PA in community-dwelling Chinese adults using a nationwide sample, and also, one of the very few studies to depict the patterns of PA in middle- and older-aged Chinese adults. ## Prevalence of depression In summary, the findings revealed high prevalence of risk of depression in middle- and older- aged Chinese adults (25.17%) and an upward trend (from 21.82% to 27.46%) with increasing age level. This finding is in agreement with those from recent studies, indicating that 23.6% to 27.0% of the elderly Chinese confront depression, while the risk of depression rises as age increases. The figure is significantly higher than that was estimated two decades ago (3.86%). Moreover, it is much higher than the average level of LMICs (1%) and even some developed countries. This, on the one hand, indicates the tremendous threat facing China in terms of the rising disease burden of depression; and on the other hand, reiterates the urgent need to identify approaches with broad accessibility to address this issue in a context where professional resources are scarce. Meanwhile, the present study found gender difference on the prevalence of depressive risk: the prevalence in women was significantly greater than that in men. Similar findings can be retrieved from other studies against Chinese older adults. This indicates that older women are in higher need for effective approaches to prevent depression. ## Patterns of PA Generally speaking, around three fifths of the participants reached the threshold outlined in the WHO guidelines. This figure is similar to those reported in other studies subject to older Chinese adults or elderly in LMICs, and supports the statements Wen et al. and Chang et al. made that East Asians tend to engage in PA at lower intensity. Significant differences in the frequency, duration and volume of PA were observed between different levels of PA intensity and different genders. Generally speaking, men were more active in VPA and women were more active in MPA. Significantly uneven distribution of frequency and volume was observed in VPA and MPA. These findings are consistent with those extracted from previous studies. The reasons may be attributed to the purpose of PA: most of the elderly who take MVPA are for domestic or occupational purposes, therefore, tend to have a high frequency and volume of PA. The patterns of duration can serve as indirect evidence for the above speculation: the majority of those who engaged in VPA spent more than 4 hours each time, which often occurs when one engages in a job that requires heavy labor work. In contrast, the distribution of MPA duration was more dispersed between “30-119min per time”, “120-239min per time” and “\>240min per time”, whereas a relatively larger proportion of respondents reported 30–119 minutes each time. Taking into consideration that the prevalence of MPA was significantly higher in women than that in men, we speculate that respondents who engaged in MPA were mainly for domestic purpose. ## Associations between PA and depression Generally speaking, risk of depression was correlated with the intensity, frequency, duration and volume of PA, whereas the direction and type of correlation mainly depended on the intensity and gender. Regarding VPA, compared with those who took no VPA, those who had higher frequency (3–5 days/week or 6–7 days/week), longer duration (≥240 min/day) and greater length (≥300 min/week) had greater odds of suffering from depression. However, the statistically significant correlation was only found in male but not in women. This may be due to the purpose of VPA: on the one hand, in traditional Chinese culture, men are the ones who take major responsibility to support family; on the other hand, the higher frequency, longer duration and larger volume of heavy-labor work may indicate lower household income and heavier financial burden. This, therefore, may lead to men suffering from higher risk of depression. Regarding adverse effect, similar findings have been documented in some studies, and the finding complies with WHO guidelines that indicate the potential harm MVPA may pose on one’s health when the total length exceeds 300 min/week. Strategies such as encouraging lower frequency and smaller amount but not large volume of VPA, or promoting taking PA with happiness may be worth consideration. Regarding MPA, engaging in sufficient but less than 300min/week MPA was associated with lower risk of depression for the whole sample population, whereas different magnitudes were observed between women and men. Compared with those who did not have MPA, women who had smaller frequency (1–2 days/week), shorter duration (\<30 min/day) and moderate length (150–299 min/week) of MPA were in lower risk of depression. However, this inverse associated was not observed in men. Moreover, engaging in higher frequency (6–7 days/week) and greater length (≥300 min/week) of MPA was positively correlated with higher likelihood of depression in men. Similar findings can be retrieved from several recent studies that found PA, especially MPA, more effective in terms of reducing risk of depression in women than men. Some even reported that only women, but not men, benefited from taking part in PA. The potential reason for this gender difference may be attributed to women’s more developed connectedness and their better ability to benefit from the social aspects of PA than men. Our speculation can be underpinned by one study, in which PA was associated with depression but the associated become statistically insignificant when social network was considered. These findings reveal the necessity to consider the gender difference when developing intervention strategies. In addition, we found that the threshold for a statistic significance was lower than that was recommended in some western guidelines. This indicates that the condition varies between different populations and intervention strategy design should be rooted in the specific context. For instance, in light of large proportion of no MPA and small proportion 1–2 day/week MPA in women, interventions could be designed to promote moderate volume of regular MPA. Some experience could be extracted from British policy that encourages MPA such as housework, gardening and home maintenance tasks. Regarding LPA, no statistically significant association was observed between risk of depression and LPA of any kind. This finding can be supported by previous literature that found no significant association between walking and depression and it can be further underpinned by WHO guidelines in which recommendations have only been given to VPA and MPA but not to LPA. However, we noted that some studies reported benefits LPA brought to one’s risk of depression. The difference in outcome may be attributed to two reasons: first, walking was part of everyday life to most of the Chinese residents and its contribution to one’s physical health is limited compared to moderate or intense PA. Second, as one study further divided LPA into high-light PA and low- light PA, and found high-light PA but not low-light PA was related to better well-being, the insignificant associated found in this study may be due to the unavailability to further divide LPA. ## Strength and limitations To the best of our knowledge, this is the first study to explore the association between risk of depression and PA from multiple dimensions including intensity, frequency, duration and volume in middle- and older-aged Chinese using a nationwide dataset. Additionally, this study is one of the few studies that further investigate the correlation by taking gender difference and four key dimensions of PA into consideration. Finally, as Rothon et al indicated, one of the weaknesses of the prior studies is the failure to control for other clinically confounding factors. Health behaviors (such as smoking and drinking) and health status (diabetes, hypertension and BMI) related variables were considered and adjusted in this study. The findings should be interpreted with caution due to the following limitations. First, although we have observed some associations between physical activity and depression, we are not able to establish a causal relationship between these two variables using a cross-sectional study. Identification of causal relationships will have to await future research using a longitudinal design. Second, PA was self-reported, which may result in recall bias and underestimation of PA. Also, different from the International Physical Activity Questionnaire and Global Physical Activity Questionnaire that ask the exact length of time one spends on PA (numbers of minutes/hours), the questionnaire used in CHARLS only classified the duration into four categories (\<30min, \<2h, \<4h, ≥4h), which may result in lower validity in calculating the total volume of PA. Further studies may benefit from assessing PA using objective measures (such as accelerometers). Finally, the present study was carried out subject to community-dwelling residents, and therefore, may not be generalizable to the institutionalized individuals. # Conclusions In conclusion, the findings revealed a considerable high prevalence of risk of depression in middle- and older-aged Chinese, especially in women, which calls for attention and more cost-effective approaches to address this issue. This study found that the risk of depression was correlated with the intensity, frequency, duration, and volume of PA, whereas the direction and magnitude mainly depended on the intensity and gender. To be specific, after controlling one’s demographic, health behaviors and physical health status, men with higher frequency, longer duration and greater length of VPA were associated with higher odds of suffering from depression than those who did not engage in VPA. No such association was found in women. Compared with not taking any MPA, smaller frequency, shorter duration and moderate volume of MPA were associated with lower risk of depression in women. However, similar associations have not been found in men. These findings not only serve to further understand the relationships between PA and depression in middle- and old-aged adults in China, but also remind the need to consider China’s cultural context when designing interventions according to international guidelines. The latter can help to inform policymakers to consider these variations to design target strategies with higher effectiveness. The authors thank the CHARLS research team and all respondents for their contribution. WHO World Health Organization CHALRS China Health and Retirement Longitudinal Survey PA physical activity VPA vigorous physical activity MPA moderate physical activity LPA light physical activity MVPA moderate-to-vigorous physical activity LMICs low- and middle- income countries CES-D 10 10-item Center for Epidemiologic Studies MET metabolic equivalent of task YLD years lived with disability DALY disease adjusted life years [^1]: The authors have declared that no competing interests exist.

# Introduction **Hepatitis B (HB)** is a serious blood born infection that affects the liver and caused by hepatitis B virus (HBV). It is infectious and the most common cause of chronic hepatitis, liver cirrhosis and hepato-cellular carcinoma. Hepatitis B is a very important public health problem affecting almost 10% of the world population. According to 2009 WHO report, about 2 billon people are affected with HB worldwide, more than 350 million suffered from chronic lifelong infection and, more than one million of individuals die because of cirrhosis and liver cancer every year. HBV is contagious and easy to be transmitted from one infected individual to another by blood to blood contact, mother to child, unprotected sexual intercourse, sharing of eating utensils and other barber shop and beauty salon equipment. The main transmission routes include prenatal infection, skin and mucous membrane infections caused by contaminated blood or body, sexual contacts and injection drug abuser. In addition, tattooing, ear piercing, acupuncture, dialysis, and even using a syringe can be the source of infection. In volunteer blood donors, the prevalence of anti HB reflection ranges from 5–10%. But the prevalence is higher in lower socio-economic status, older age group and those persons exposed to blood products. Prevalence of infection, modes of transmission and human behavior conspire to geographically different epidemiologic patterns of HB infection. The practice of modern medicine have contributed a lot in the increase of the case and spreading of blood born diseases like Human immune deficiency virus and HBV due to lapse in the sterilization technique of instruments and improper hospital waste management as 10 to 20% health care waste is regarded hazardous. Prevention against any disease is proportional to knowledge, attitude and practice (KAP) of the population and reflection of the importance that is paid to health related issue by the society. Health care workers should familiarize themselves with “universal precautions”, which is defined by Center for Disease Control, as a set of precautions designed to prevent transmission of Human immunodeficiency virus (HIV), HBV, and other blood-borne pathogens when providing first aid or health care. Under universal precautions, blood and certain body fluids of all patients are considered potentially infectious for HIV, HBV and other blood borne pathogens. A safe and effective vaccine against HBV is available since 20 years and is effective in preventing infection and the serious consequence of hepatitis including liver cancer and cirrhosis when given before or after exposure. In Ethiopia, the Expanded program for immunization program (EPI) delivers eight vaccines to protect children against the following serious childhood illnesses: tuberculosis, poliomyelitis, diphtheria, pertussis, tetanus, *Haemophilus influenza-B* (Hib) infections, hepatitis B disease and measles. The hepatitis B vaccines are new vaccines introduced into the EPI programme in Ethiopia in 2007. Since 1990 new HB infections among children and adolescents have dropped by more than 95% and by 75% in other age group. Vaccination gives long term protection from HB infection, possibly life-long. KAP surveys are representative of a specific population to collect information on what is known, believed and done about a particular topic, and are the most often used study tool in health- seeking behavior research. Knowledge is usually assessed to see how far community knowledge corresponds to biomedical concepts. Practices in KAP surveys usually inquire about preventive measures or different health care options. Normally, hypothetical questions are asked, so it permits statements about actual practices, rather, it yields information on people’s behaviors or on what they know should be done. Medical and health science students, being part of the health care delivery system, are exposed to the same size of risk as other health care workers when they come in contact with patients and contaminated instruments. They are the first level of contact between patients and medical care. They are expected to undertake activities related to patient care with the beginning of their clinical years. As on date, very few studies have been conducted to find out the knowledge and practice of medical students about Hepatitis B in Ethiopia. This study assessed the knowledge and practice of students towards HB transmission and prevention among medical and health students in Haramaya University. # Materials and Methods ## Study Area and Period This cross- sectional study was conducted from February 1–15, 2013 in College of medical and health science, Haramaya University, largest and oldest educational institution in the country. The University is running 50 degree, 50 masters and 13 PhD programme. The University has 12 colleges (including the health science and medical collage that has 8 departments). It has one teaching referral hospital, one health center and one clinic. The college of medical and health science is located in Harar town 25 kilometer away from the main campus and 500 kilometer away from Addis Ababa. All Medical and Health Science students of Haramaya University who were on clinical attachment included in this study. The students who were not on clinical attachment, those who have mental illness, seriously ill and blind were excluded from the study. ## Sample Size Determination The sample size was calculated by using Epi Info 3.5 software package based on single population proportion formula (estimated prevalence rate of 50% konwledge, margin of error 5% and a 95% confidence interval). The calculated sample size was 384. Since the source population is less than 10000, we used correction formula and adding 10% for none response, the resulting minimum sample size was 322. ## Sampling Method Study participants were selected using systematic sampling technique. First we stratified students based on year of study and department. Second, the sample size was distributed proportionally to each year of study based on the student population they have. Finally using students list respondents were selected by systematic random sampling. In case of absenteeism the next number was included in the study. ## Variables Knowledge and practice of students towards transmission and prevention of Hepatitis B virus (HBV) were considered as dependent variables and the independent variables were age, sex, and residential area, and marital status, year of study and departments of the study population. ## Data Collection Process A self-administered structured questionnaire was used to collect information about the socio-demographic characteristics of respondents, knowledge towards transmission and prevention method of hepatitis B virus and practice towards prevention HBV. The English version of the questionnaire was used to collect the information from the respondents. The questionnaire was pretested on 5% of pre- clinical students in the college other than the study participants and necessary changes were implemented. Training was given for data collectors and the overall data collection activities were supervised by principal investigator. The participants were anonymously responded to the items on the questionnaire. ## Data Analysis Data was checked for completeness and consistency. Coded data were entered and cleaned using Epi Data software and analyzed using SPSS version 20.0 (SPSS, Chicago, IL, USA). Descriptive statistics were conducted using frequencies and proportions. Bivariate and multivariate analyses were carried out using logistic regression to examine the relationship between the outcome variable (mean knowledge and practice) and selected socio-demographic factors. Adjusted and unadjusted odds ratios (OR) and their 95% confidence intervals (CIs) were used as indicators of the strength of association. A p-value of 0.05 or less was used as cut off level for statistical significance. ## Ethical Consideration The research protocol was approved by the medical and Health Sciences College Research Ethics review Committee of the Haramaya University. Then informed written consent was obtained from each study participants. Individual participant consent was waived/by the Ethics Committee. That means the committee save grads not a rule to be obeyed. Written informed consent was obtained from the immediate caretaker, or next of kin, prior to inclusion, on behalf of children participating in the study. Moreover confidentiality assured for all the information provided and personal identifiers were not included on questionnaire. # Results A total of 322 students belonging to 2^nd, 3^rd, 4^th, 5^th, and 6^th year from six different departments were approached for the study and all of them were participated in the study making response rate of 100%. Majority of the students (91%) were from the age group 20–24 and 232 (72%) of the respondents were male. Majority, 307 (95.3%) of the study participants were single in marital status and more than half; (55.6%) of the respondents came from urban area ****. ## Knowledge towards Hepatitis B Transmission and Prevention Knowledge was assessed by questions focusing on sign and symptoms, transmission, treatment and prevention. Each response was scored as ‘yes’ or ‘no’. The scoring range of the questionnaire was 16 (largest) to 0 (smallest). A cut off level of \<12 was considered as poor whereas \> = 12 was considered as adequate knowledge about HB. Knowledge scores for individuals were calculated and summed up to give the total knowledge score. Out of the 322 participants, 141 (43.8%) were within the poor knowledge range whereas 181 (56.2%) showed adequate knowledge about HB. Poor knowledge was apparent in responses to questions relating to transmission (question 7–10). Nearly half; 156 (48.4%) of the study participants were not know that HB has post exposure prophylaxis and the mean knowledge score for the entire study cohort was 11.52±2.37 ****. ## Assessment of Practices towards Hepatitis B Prevention Practices towards HB prevention were assessed by asking five questions as shown in. Each question was labeled with good or poor practice. A score of 1 was given to good while 0 was given to bad practice with a score range of largest of 5 to a smallest of 0. The scale classified practice as good with score \>3 and poor ≤3. Majority of the respondents, 276 (85.7%) never screened for HB and 279 (86.6%) stated a negative immunized status against HB. It was interesting to know that nearly one third 102 (31.7%) of the respondents never asked for screening of blood and blood products before transfusion, and 53 (16.5%) of the respondents never asked for a new syringe when required. Majority, 245 (76.1%) of the respondents were never participated in any education program on HB. The mean score for HB related practices was 2.04±1.15 revealing poor practices among the study participants. Out of 322 participants, 43 (13.4%) were vaccinated against HBV. In the vaccinated group, 15 (4.7%) completed all 3 doses of their vaccination schedule and remaining 28 (8.7%) students were incompletely vaccinated. Reasons for not getting vaccinated were lack of information in 67 (20.8%) students, no need was felt by 9 (2.8%) students, 15 (4.7%) had fear of injection and 45 (14%) said they ignorance ****. ## Association of Demographic Characteristics and Mean Knowledge and Practice Scores The association of demographic characteristics and mean Knowledge and practice scores using multivariate logistic regression is presented in. Among the demographic variables, study department and sex of the respondents were significantly associated with both mean knowledge and practice scores. The difference of mean knowledge and mean practice score across study department of the students were found to be significant at 95% confidence interval (P = 0.007) and (P = 0.001). Being psychiatry and medical laboratory students is statistically associated with poor knowledge towards transmission and prevention of HB and poor practice towards prevention of HB as compared to medical students. Even though year of study is not statistically associated with poor mean knowledge score towards transmission and prevention of HB (P = 0.17), being second year students or starting clinical attachment in early year of study (2^nd) is positively and significantly associated with poor mean knowledge of HB (OR, 11.7 and P = 0.042). Multivariate logistic regression showed that study department significantly associated with poor practice towards prevention of HB (P = **0.001)**. There is no variation towards poor practice in prevention of HB across year of study ****. ## Correlation between Mean Knowledge and Practice of the Students towards HB Correlations were interpreted using the following criteria: 0–0.25 = weak correlation, 0.25–0.5 = fair correlation,0.5–0.75 = good correlation and greater than 0.75 = excellent correlation. This study showed significant positive but weak linear correlations between knowledge and practice (r = 0.173, p = 0.002). The result reaffirms the relationship between knowledge and practice with infection control measures even though the correlation is weak in this study. The positive correlations between knowledge-practice in this study reaffirm the relationship between knowledge and practice with infection control measures. It is concluded that adequate knowledge can lead in good practices. # Discussion The current study sought to evaluate knowledge and practice towards HB among medical and health science students who started clinical attachment. Results of the study showed poor knowledge and practice towards HB. The mean knowledge score was *11.52±2.37* indicating low level of knowledge towards HB among the study cohort. Scientific knowledge about HBV transmission is essential for medical students. They can take proper protection during their clinical posting as HBV is 50 times easier to transmit than HIV. The knowledge about transmission of Hepatitis B through sexual route (65.5%) by used needles and syringes (71.7%) by blood transfusion (89.8%) was high. But through vertical transmission (55.9%) and contaminated water/food prepared by person suffering with these infections (22%) was low among overall medical and health science students. Among study participants, 62.4% of them know that HB can cause liver cancer and affect other organs other than liver, again a major sign of concern. These results are in line with the findings from studies reported from B.J.Medical College, Ahmadabad, and Gujarat, India; where majority of the medical students had correct knowledge on mode of transmission. In the present study, 13.4% of the students received one or more doses of hepatitis B vaccine and among which 4.7% of the students were fully vaccinated against Hepatitis B. This was lower than the vaccination status of 87.8% study done at Muhammad Medical College Mirpurkhas, 29.3% reported among medical students of B.J. Medical College, 35% reported in civil hospital of 60 laboratory technicians, 88.1% in **study done at two national/regional congresses and two university hospitals in Iran (6)** **,** 63% reported from India among medical students and 42% reported among medical students of Lahore. The most frequent reason for not getting vaccinated in the present study was inaccessibility 166 (51.6%), lack of motivation 67 (20.8%) and followed by ignorance 45 (14%) or fear of injection 15 (4.7%). The finding is consistent with a study result from medical college of Mirpurkhas, Pakistan. These are serious issues and baseless reasons and need to be improved by education. This study also showed that 89.4% of the respondents have poor practice towards prevention of HB. The positive correlations between knowledge and practice in this study re affirm the relationship between knowledge and practice with infection control measures. It is concluded that adequate knowledge can lead to good practices. The findings are in line with the results presented by the study in India and Pakistan. # Conclusion and Recommendation The present study concludes that there is poor knowledge among the medical and health science students entering into the profession practice about the hazards of Hepatitis B, its mode of transmission and prevention. Moreover, majority (95.3%) the students were not fully vaccinated against Hepatitis B and 48.4% of the students were no aware about the availability of post exposure prophylaxis for HB, which made them more vulnerable to the disease in their professional life and 75.8% of the students were not participated in health education program related to Hepatitis-B. Study department is the strong independent correlates for poor knowledge of the students towards transmission, prevention of HB and practice in preventing HB. Since medical and health science students are at increased risk of acquiring needle stick injury, and exposed to blood and blood products in their professional practice, medical and health science students should be vaccinated upon entry into the medical college. Medical and Health Science Colleges should have Occupational or Student Health Departments that must take responsibility for HBV testing, vaccination, monitoring vaccine response and providing post- exposure prophylaxis. It is also recommended that a policy be implemented for complete vaccination and giving training on infection prevention for all medical and health science students before they start clinical attachment (professional practice) in all medical and health science colleges in the country despite of study department (profession). It’s also advisable to make sure vaccine availability and accessibility. The authors would like to thank all data collectors and study participants involved in the study. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YMM KTK. Performed the experiments: YMM. Analyzed the data: YMM KTK. Contributed reagents/materials/analysis tools: YMM KTK. Wrote the paper: YMM KTK.

# Introduction The composition of feeds used in salmon aquaculture has undergone significant changes over the last decades. The rapid growth in the aquaculture industry, but stable production of the major protein resource, fishmeal (FM), has led to increasing use of alternatives. Alternative protein sources are required to contribute to a well-balanced diet and to support optimal fish growth performance, health and disease resistance. Currently, proteins derived from insects, terrestrial animal co-products and microbial ingredients are considered to be valuable alternatives. In commercial diets, plant-derived proteins have already replaced two third of the proteins of marine origin. Plant ingredients are the most attractive protein sources due to their low cost of production, high protein content and availability. Inclusion of plant ingredients in salmonid feeds can, however, result in reduced growth performance and feed utilization, and health issues due to anti-nutritional factors (ANF). ANFs in plant-based diets have been associated with detrimental effects on the intestine of several salmonid species, and soybean meal-induced enteritis (SBMIE) is a well described condition in Atlantic salmon (*Salmo salar* L.). Currently, in commercial salmon diets, the refined soy product with a reduced level of ANF, soy protein concentrate, is the primary protein source of plant origin and has not been shown to cause pathological changes in the distal intestine (DI) of salmonids after short-term dietary exposure. However, a certain degree of inflammation and ectopic epithelial cells have been observed in the DI of salmonids when fed diets based on FM and soy protein concentrate over a longer period of time. Microbial ingredients have proven to be high-quality protein sources with the ability to mitigate the negative effects of plant-derived proteins. Microbial ingredients such as yeast and bacteria contain bioactive compounds with immune- modulating properties that improve the changes caused by SBM. Moreover, bacterial meal has been shown to prevent the development of SBMIE in a dose- dependent manner. Mannan oligosaccharides, compounds found in yeast cell walls, have been used as a prebiotic and shown to protect the intestinal mucosa and reduce inflammation and leukocyte infiltration. However, the degree of bioactivity of these compounds depends on the microbial origin as well as the fermentation conditions and downstream processing of the microbial product before incorporation into the salmon diet. Assessing the bioactivity of novel dietary microbial ingredients has traditionally involved the evaluation of local intestinal tissue responses, such as changes in morphology, gene expression or microbiome. Local intestinal responses can induce systemic responses that could contain new biomarkers for health and disease. The innate immune system may respond to local inflammatory or immunomodulatory stimuli in the intestine of fish and in turn elicit changes systemically. The release of cytokines into the circulation stimulates hepatocytes to produce proteins and release them into the circulation to regain homeostasis. Plasma proteomic analysis is a post-genomic tool that allows the investigation of complex biological systems involved in physiology and pathology. Plasma proteome profiles in response to certain inflammatory or immunomodulatory stimuli can be useful diagnostic tools for fish diseases and indicators of fish health. In this study, an inactive dry yeast strain of *Candida utilis* (*C*. *utilis*) was used as an alternative protein source with functional properties in FM and SBM based diets. SBM was used as a dietary challenge to induce SBMIE. Increasing levels of *C*. *utilis* were included in the diets to evaluate the immunomodulating properties of the yeast, in particular, the ability to prevent and counteract SBMIE. We combine histomorphological evaluation, immunohistochemistry, morphometry and gene expression of the DI with plasma proteome analysis. By combining these methods, we aim to identify local intestinal tissue responses and changes in plasma protein profiles in Atlantic salmon resulting from dietary treatments. Our results show that inclusion of *C*. *utilis* as an alternative protein source does not alter the local morphology and immune cell population in the DI or induce major changes in the plasma proteins of Atlantic salmon. # Materials and methods ## Experimental ingredient and diet preparation The inactive dry *C*. *utilis* yeast corresponded to a commercial product called Lakes States^® Type B produced by LALLEMAND SAS (Blagnac, France). shows the proximate composition of the test inactive dry yeast. The crude protein content was determined by multiplying nitrogen content by a conversion factor of 6.25. All diets used in this study were formulated to meet or exceeded the nutrient requirements of Atlantic salmon, produced by extrusion technology at the Center for Feed Technology (FôrTek) at the Norwegian University of Life Sciences (Aas, Norway), and stored at -20°C prior to feeding. The extruded pellets were dried to \~ 6% moisture content before vacuum coating with fish oil. The diets consisted of a FM-based control diet (FM group) and the following six experimental diets; a diet containing 200 g/kg *C*. *utilis* (FM200CU group), and five diets containing 200 g/kg SBM together with 0 (SBM group), 25, 50, 100 or 200 g/kg *C*. *utilis* (SBM25CU, SBM50CU, SBM100CU and SBM200CU groups, respectively). ## Fish husbandry and feeding trial The experiment was conducted according to laws and regulations for experiments on live animals in EU (Directive 2010/637EU) and Norway (FOR-2015-06-18-761). Vaccinated salmon (Aquavac PD7, MSD Animal Health, Bergen, Norway) were acquired from Sørsmolt AS (Sannidal, Norway) and maintained according to the guidelines established by the Norwegian Animal Research Authority at the Research Station Solbergstrand of Norwegian Institute of Water Research (Drøbak, Norway). Fish were acclimated to seawater, housed in 300 L tanks supplied with ultraviolet light treated seawater (8 °C; 34.5 g/L NaCl) in a 7–8 L per min flow-through system, and fed with a commercial marine-based compound feed not containing soybean-derived products (3-mm pellet; Polarfeed AS, Europharma, Leknes, Norway) under continuous light during a 4-month period prior to conducting the feeding trial. The water temperature, dissolved oxygen and pH level were measured and recorded daily. At the beginning of the experiment, 360 fish (average initial body weight of 526 g) were randomly assigned to 18 tanks (20 fish/tank) and acclimated to the FM based control diet for two weeks prior to feeding experimental diets. Feeding was approximately 20% in excess twice daily using automatic belt feeders based on a daily estimate of fish biomass and uneaten feed per tank, which was collected from the tank outlet after each feeding period. Following the acclimation period, each experimental diet was randomly allocated to the fish tanks (two tanks/diet) and fed for 30 days (period 1) as described above. After 30 days, the feeding strategy was changed, and new diets were fed for 7 days (period 2). As a control, one fish group received FM through the experiment. To assess whether *C*. *utilis* were able to counteract enteritis induced by SBM, four fish groups received SBM diets combined with different inclusions levels of CU (i.e. SBM25CU, SBM50CU, SBM100CU, SBM200CU) in period 1. One fish group received FM200CU in period 1 to evaluate if *C*. *utilis* in combination with FM alone would affect the DI. In period 2, the ability of *C*. *utilis* to prevent SBMIE was assessed as the diet was changed to SBM in this group. Finally, three fish groups were fed SBM diet to induce SBMIE in period 1, and in period 2 the diets were changed to either FM, FM200CU and SBM200CU to evaluate if these diets were able to reverse SBMIE. The feeding strategy is illustrated in. ## Sample collection At each sampling point (0, 7, 30 and 37 days), 8 fish per diet (4 fish per tank) were randomly selected and anesthetized by immersion in 60 mg/l of tricaine methanesulfonate (MS-222, Sigma-Aldrich, MO, USA) and subsequently euthanized by a sharp blow to the head. Before dissecting fish, fish weights and lengths were recorded, and blood samples were taken from the *vena caudalis* (tail vein) using a heparinized syringe and centrifuged (1300–2000 x g for 10 min) to isolate blood plasma, which was aliquoted and stored at –80°C until proteomic analysis was performed. DI tissue was sampled and kept in RNAlater^® (Merck KGaA, Darmstadt, Germany) at 4°C for 24 h, and then at –80°C, until RNA extraction. DI tissue samples were also collected and preserved for subsequent histology, morphometry and immunohistochemistry. ## Histology, immunohistochemistry and morphometric measurements ### Histology Approximately 1 cm segment of the DI was open longitudinally and the intestinal content was carefully removed. The tissue was fixed in 10% formalin for 48 h at room temperature and further processed according to routine histological procedures. Briefly, tissue was embedded in paraffin with an orientation to ensure longitudinal sectioning. Sections (2 μm) of paraffin-embedded DI tissue were mounted on glass slides (Menzel Gläser, Thermo Fisher Scientific, Braunschweig, Germany) and processed for staining with hematoxylin and eosin. A blinded, semi-quantitative histological scoring of the DI was performed using the criteria described in detail by Baeverfjord and Krogdahl. Briefly, the criteria were: 1) shortening of both the simple and complex intestinal mucosal folds, 2) appearance of the enterocytes including supranuclear vacuolization, cellular heights and presence of intraepithelial lymphocytes (IELs), 3) widening of the central lamina propria of the simple and complex folds by connective tissue and 4) infiltration of leucocytes in the lamina propria. Each criterion was given a score ranging from 0 to 2, and half scores were included (i.e. 0, 0.5, 1, 1.5, 2). Score 0 indicated normal morphology and score 2 represented marked changes. Score 0.5 was regarded as changes within the normal range. ### Immunohistochemistry Histological sections of DI from the fish sampled at day 30 (8 fish/diet), prepared as described above, were subjected for immunohistochemical analysis, and the following diet groups were included: FM, FM200CU, SBM, SBM25CU and SBM200CU. CD3ε and CD8α positive T lymphocytes were identified in DI tissue sections by immunohistochemistry using a monoclonal anti-CD3ɛ antibody (dilution 1:600) and a monoclonal anti-CD8α antibody (kindly supplied by Karsten Skjødt, dilution 1:50), respectively. This analysis was performed as follows: formalin- fixed, paraffin-embedded DI sections (4 μm) were mounted on poly-L-lysine-coated glass slides (Superfrost Plus, Thermo Fisher Scientific, Braunschweig, Germany) and left to dry at 37°C. The slides were incubated at 58°C for 30 min before deparaffinized in xylene and rehydrated in graded alcohols to distilled water. Antigen retrieval was done by using hydrated autoclaving at 121°C for 15 min in 0.01 M citrate buffer, pH6. Endogenous peroxidase was inhibited with 0.05% phenylhydrazine (0.05%; Sigma-Aldrich, MO, USA) in phosphate buffered saline, preheated to 37°C, for 40 min. The sections were stored in PBS overnight at 4°C and then incubated with normal goat serum (dilution 1:50; Sigma-Aldrich) in 5% bovine serum albumin /0.05 M tris-buffered saline to avoid non-specific binding for 20 min. The blocking solution was tapped off without washing, and the sections were incubated with primary antibody diluted in 1% bovine serum albumin/Tris-buffered saline for 1 h. Control sections were incubated with only 1% bovine serum albumin. The sections were incubated in the kit polymer-HRP anti-mouse (Dako En Vision+ System-HRP, Dako, Glostrup, Denmark), as a secondary antibody, for 30 min. Peroxidase activity was detected with 3,3’-diaminobenzidine following the kit procedure. The sections were counterstained with hematoxylin for 30 s followed by washing in distilled water before mounting with Aquatex (Novoglas Labortechnik Langenbrinck, Bern, Germany). Unless otherwise stated, the sections were washed three times with phosphate buffered saline for 5 min between each step. All incubations took place in a humid chamber at room temperature. ### Morphometric measurements and calculation of immune cell density Morphometric measurements and calculation of the density of immune cells were performed from immunohistochemically labelled sections mentioned above. ImageJ software, version v1.51r, was used to perform the measurements and calculations. Images were captured with a Zeiss Axiocam 506 color camera connected to a light microscope (Zeiss Axio Imager M2, Carl Zeiss, Germany) at a 10 X magnification. The measurement scale was set to 2.26 pixels/μm in ImageJ and the measurements were converted from μm to mm. Counting of immunohistochemically labelled cells was performed using the multi-point tool. The freehand selection and segmented line selection were used to measure fold area and height, respectively. The fold height was measured from stratum compactum to the tip of the epithelium lining the fold. The fold area was measured from stratum compactum, including the middle of the fold base on each side, and the whole simple fold. The immunohistochemically labelled cells were counted within the measured area of the simple fold. The density of CD3ɛ- or CD8α-labelled cells was calculated as follows: Cell density = (no. of labelled cells)/area. Simple folds were subjected to all the measurements mentioned above and the folds were selected as the first appropriate simple fold located to the left of a complex fold. An appropriate fold was defined as a fold that appeared long, not bent and had an intact epithelium that was attached to the basement membrane all the way to the tip of the fold. Between 2–6 measurements were collected from each individual with a total of at least 30 measurements from each group for each measurement. A mean for each individual was calculated based on the measurements. ## RNA isolation, cDNA synthesis, quantitative PCR A small piece of DI tissue (approximately 0.5 cm) from FM, FM200CU, SBM and SBM200CU diet groups (8 fish/diet) at day 30 were subject to gene expression analysis. Total RNA was extracted and purified using RNeasy^® 96 kit (Qiagen, Valencia, USA) and QIAcube^® HT system (Qiagen), according to the manufacturer’s protocol. After the first washing step, on-column DNase treatment was performed using PureLink TM DNase kit (Thermo Fisher Scientific, Waltham, Massachusetts). RNA concentration and quality were measured using NanoDrop TM 8000 spectrophotometer (Thermo Fisher Scientific). Purified total RNA was stored at −80°C until further analysis. Prior to cDNA synthesis, all samples were normalized to 400 ng/μL. cDNA synthesis was performed using AffinityScript QPCR cDNA Synthesis kit following the manufacturer’s guidelines (Agilent Technologies, Santa Clara, CA, USA). The total reaction volume was 10 μL using 5 μL of Mastermix, 1.5 μL random hexamer primers, 0.5 μL AffinityScript RT/ RNase Block enzyme mixture and 3 μL DNase treated RNA. The resulting cDNA was stored at −80°C before use. All quantitative PCR (qPCR) reactions were performed in duplicates and conducted in 96 well plates on LightCycler^® 480 system (Roche Diagnostics, Mannheim, Germany). Each reaction consisted of a total amount of 12 μL divided into 6 μL LightCycler 480 SYBR Green I Master (Roche Diagnostics), 2 μL primers (5μM), and 4 μL cDNA. The qPCR conditions were 95°C for 5 min before a total of 45 cycles of 95°C for 5 seconds, 60°C for 15 seconds, and 72°C for 15 seconds. To confirm amplification specificity, each PCR product was subject to melting curve analysis (95°C 5 s; 65°C 60 s; 97°C continuously). Primers tested are listed in. Glyceraldehyde-3-phosphate dehydrogenase (GAPDH) and hypoxanthine phosphoribosyltransferase I (HPRTI) were chosen as reference for normalization since these genes did not show significant differential expression between the diets and have previously been described as suitable reference genes in the DI of salmon. The crossing point (Cp) values were determined using the maximum second derivative method on the basis of the LightCycler^® 480 software release 1.5.1.62 (Roche Diagnostics). The geometric mean of the CP- values for GDPH and HPRTI was used as an index. The qPCR relative expression of mRNA was calculated using the ΔΔCt method. ## In-solution digestion and protein sequence analysis by LC-MS/MS Proteomic analysis was performed, according to methods described by Lagos *et al*. 2017, using four biological replicates per treatment of plasma taken at the end of the first feeding period (30 days) (FM, SBM, SBM200CU and FM200CU). In brief, frozen plasma (−80°C) was thawed and diluted to 40 μg of total protein in PBS, and the pH was adjusted to 8 by adding ammonium bicarbonate (Sigma-Aldrich, Darmstadt, Germany). Subsequently, the proteins were digested with 10 μg trypsin (Promega, sequencing grade) overnight at 37°C. The digestion was stopped by adding 5 μL 50% formic acid and the generated peptides were purified using a ZipTip C18 (Millipore, Billerica, MA, USA) according to the manufacturer’s instructions, and dried using a Speed Vac concentrator (Concentrator Plus, Eppendorf, Hamburg, Germany). The tryptic peptides were dissolved in 10 μL 0.1% formic acid/2% acetonitrile and 5 μL analyzed using an Ultimate 3000 RSLCnano- UHPLC system connected to a Q Exactive mass spectrometer (Thermo Fisher Scientific, Bremen, Germany) equipped with a nanoelectrospray ion source. For liquid chromatography separation, an Acclaim PepMap 100 column (C18, 2 μm beads, 100 Å, 75 μm inner diameter, 50 cm length) (Dionex, Sunnyvale CA, USA) was used. The mass spectrometer was operated in the data-dependent mode to automatically switch between MS and MS/MS acquisition. Survey full scan MS spectra (from m/z 400 to 2,000) were acquired with the resolution R = 70,000 at m/z 200, after accumulation to a target of 1e5. The maximum allowed ion accumulation times were 60 ms. The proteomic analysis was performed by the Proteomic core facility of the University of Oslo. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier PXD012051. ## Data analysis ### Histology, morphometric measurements, T-cell density and gene expression Non-parametric data from the histological evaluation were analyzed by Kruskal- Wallis followed by post hoc Dunn’s test with a comparison of mean rank. Shapiro- Wilk normality test was used to test the normal distribution of the data from morphometric analyses and T-cell density and further analyzed by one-way ANOVA followed by Tukey’s multiple comparisons test. Morphometric analyses and T-cell density analyses were performed at the individual level using the mean of measurements of between 2–6 simple folds per fish. Results of qPCR (means ± standard deviations) were analyzed using One-way ANOVA with Dunnett\`s multiple comparison test (*a* \< 0.0001). These analyses were performed in GraphPad Prism, version 7.00 and 8.0.1 (GraphPad Software Inc., San Diego, CA, USA). ### Proteomic data analysis The resulting proteomic data, MS raw files, were analyzed using MaxQuant and Perseus version 1.6.0.7 based on MS1 intensity quantification. Then identifications were filtered to achieve a protein false discovery rate (FDR) of 1%. Peptide identification was determined using fragment mass tolerance for the MS1 6 ppm and the fragment mass tolerance for the MS2 20 ppm. MS/MS spectra were searched against the salmon proteome available in September 2019 (<https://www.uniprot.org/proteomes/query=taxonomy:82390>), and reverse database searches were used in the estimation of FDR. The analysis was restricted to proteins reproducibly identified in at least two of the four replicates per diet, making the minimum number of peptides used to identify each protein an average value of 2. Row-wise normalization was applied to provide Gaussian-like distributions for adjusting the differences among protein data. Protein raw data were transferred to log normalization; missing value imputation was used to replace the not identified proteins on the quantitative analysis and then performed on autoscaled data (mean-centered and divided by the standard deviation of each variable). A diagnostic plot was utilized to represent normalization procedures for normal distribution assessments. Volcano plot analysis and data modeling were performed in R, using R package MetaboAnalystR. UniprotKB database was used for the functional annotation of the proteins. # Results ## Growth and general health All groups of fish accepted their allocated diets and no significant differences were found in feed intake nor growth rate among dietary treatment. The average initial weight was 526 g and the average final weight was 667 g on day 37, considering that the weight was measured as bulk, this indicates that in general fish gained weight during the experimental period. The general health of the fish in this experiment was good, but two fish died during the experimental period, one from SBM50CU group and one from SBM100CU group, for unknown reasons. ## Histology All fish sampled at day 0 showed normal DI morphology. Briefly, simple folds were long and slender with a thin lamina propria, whereas the complex folds were tall with a narrow lamina propria and a partial central core of smooth muscle. Intestinal epithelial cells were tall with the nucleus located basally, large vacuoles located apically and many IELs. Goblet cells were scattered among the epithelial cells towards the apex, and there was a higher presence of goblet cells at the apex of the complex folds. The lamina propria adjacent to the stratum compactum was thin and numerous eosinophilic granule cells were present. At day 7, FM and FM200CU groups showed normal DI morphology as described above. In general, all fish groups fed diets containing SBM, including SBM diets with *C*. *utilis* inclusion, displayed changes in the DI morphology consistent with SBMIE (described in detail below). Nevertheless, in SBM25CU and SBM50CU, there was variation within the groups ranging from individuals showing no changes to other individuals with moderate changes in DI morphology. At day 30, no morphological changes were seen in the DI of FM and FM200CU groups. However, a moderate SBMIE was present in the DI of SBM, SBM100CU and SBM200CU groups. Both simple and complex folds were shorter with a widening of the lamina propria within the folds and adjacent to the stratum compactum. A fusion of the simple folds was frequently observed. There was an increased presence of connective tissue in the lamina propria and the increased infiltration of leucocytes consisted mainly of eosinophilic granule cells and to a lesser extent of lymphocytes. The intestinal epithelial cells showed a reduction in height with nucleus displaced in a more apical position and small supranuclear vacuoles. In SBM25CU and SBM50CU groups, there was still a variation within the groups as seen on day 7. At day 37, fish previously fed SBM, either alone or in combination with *C*. *utilis*, had normal DI morphology after being fed FM for 7 days. Similarly, most of the fish that had diets changed from SBM to FM200CU had DI morphology regarded as normal, but there were some fish in these groups that had mild enteritis. Changing diet from FM200CU to SBM induced in general a mild SBMIE, whereas a shift from SBM to SBM200CU maintained a moderate SBMIE. The FM control group was normal, as described above. No tank effects were observed at any of the sampling points. In, only fish from the tanks that were subjected to immunohistochemical analysis at day 30 are presented, i.e., two SBM-groups are omitted from the figure. Fish with diet shift from SBM combined with *C*. *utilis* to FM at day 37 had normal DI morphology but are excluded from as this group did not differ from the SBM group. ## Immunohistochemistry, morphometry, density of immune cells At day 30, CD3ɛ positive cells showed an abundant presence at the base of the epithelium and extending along the entire length of simple folds of FM and FM200CU groups. Only a few CD3ɛ positive cells were observed in the lamina propria adjacent to the stratum compactum, and were rarely present in the lamina propria of the simple folds. A weak diffuse labelling was observed in the smooth muscle, which was interpreted as background labelling. The negative controls were blank. The density of CD3ɛ positive cells in the simple folds of groups fed diets containing SBM was increased compared with the density in FM group. IEL’s that showed labelling for CD3ɛ were located as individual cells at the base of the epithelium, but occasionally clusters of CD3ɛ-labelled IEL’s were observed. CD3ɛ positive cells were more frequent in the lamina propria adjacent to the stratum compactum of fish fed SBM compared with fish fed FM, but there were only a few CD3ɛ-positive cells present in the lamina propria of the simple folds. CD8α-labelled IEL’s showed the same distribution as the CD3ɛ-labelled IEL’s being located basally between the intestinal epithelial cells in all diets. In general, the presence of CD8α-labelled IEL’s was lower than the presence of CD3ɛ-labelled IEL’s. Morphometric measurements of simple folds, both length and area, revealed no significant differences between FM and FM200CU groups. The simple folds in the DI of fish fed SBM, SBM25CU and SBM200CU were significantly shorter and had a considerably smaller area than the simple folds in fish fed FM200CU and FM. There were no statistically significant differences between the simple folds of fish fed diets containing SBM either alone or combined with *C*. *utilis*. The density of CD3ε and CD8α-labelled cells in fish of FM and FM200CU groups was significantly lower compared with the density in fish from groups fed diets containing SBM. There was a statistically significant difference between the density of CD8α-labelled cells in fish of SBM group and the density CD8α-labelled cells in fish from the SBM25CU (p = 0.0465). ## Gene expression Among the tested genes, only mRNA expression of Aquaporin 8 (*aqp8*) was significantly down-regulated in the SBM and SBM200CU groups compared with the FM control group. There was no significant difference between FM200CU and the FM control group. The transcription levels of superoxide dismutase 1 (*sod1*), glutathione S-transferase alpha 3 (*gsta3*), annexin A1 (*anxa*), and catalase (*cat*) were not different among dietary groups (p \> 0.05). ## Plasma proteome We performed proteomic analysis on plasma sampled at day 30 from four individual fish from treatment FM, SBM, SBM200CU and FM200CU. In total, 286 salmon proteins were identified, and after filtering for proteins present in at least two of the four individuals per diet, 158 proteins were selected for further analyses. This criterion was used due to the variability among fish. A Venn diagram was built up with the 158 proteins showing the overlapping of proteins detected in the four dietary treatments. There were 126 plasma proteins shared between the four groups. Moreover, each dietary treatment presented unique proteins except for diet SBM200CU (D6). The five unique proteins present in D1 (FM) were creatine kinase, lymphocyte cytosolic protein 2, histone H3, kininogen-1 and electron transfer flavoprotein alpha polypeptide. While in D2 (SBM), we detected three unique proteins: fatty acid-binding protein intestinal, transketolase and 14-3-3 beta/alph-1 protein. We did not detect unique proteins in D6 (SBM200CU), whereas in D7 (FM200CU) we found five unique proteins: GMP/IMP nucleotidase, hemoglobin subunit alpha, beta-globin, flavin reductase and L-lactate dehydrogenase. In order to study the relative expression of proteins, we generated volcano plots comparing each dietary treatment to the control. We observed that SBM induced the differential expression of nine proteins compared to the control diet (D1:FM), while diets containing yeast (200CU) show twelve and ten proteins respectively (SBM200CU, FM200CU) compared to the control diet. # Discussion Research on the effects of nutrition on fish health and disease has mainly focused on intestinal local immune responses rather than evaluating overall health impact. Therefore, the present study used an integrated approach to achieve a better understanding of the effect of feeding inactive dry *C*. *utilis* yeast, SBM and increasing levels of *C*. *utilis* yeast to Atlantic salmon in presence of SBMIE. Herein, we discuss how both local changes in the DI, including morphology, immune cell profile and gene expression, and systemic changes in the plasma proteome could reflect challenges posed by dietary treatments. A previous study has shown that there were no significant negative effects on feed intake, specific growth rate or feed conversion ratio when up to 30% *C*. *utilis* was included in the diet for salmon, but DI morphology was not assessed. A study where parr where fed 200 g/kg *C*. *utilis* with FM showed that this diet combination did not alter the DI morphology when compared with FM control diet. The present study is the first to demonstrate that FM200CU diet maintains a DI morphology similar to the FM based control diet in sea-water adapted farmed salmon. Furthermore, FM200CU presented a similar T-cell population profile in the DI compared with the control FM group, indicating no stimulation of the local T-cell population when SBMIE was not present. No significant alterations in mRNA expression further support the assumption that FM200CU does not impose any local effect in a normal DI. On the other hand, it is widely known that SBM induces inflammatory changes in the DI morphology, and the local immune response of the DI to SBM has been described as a T-cell mediated inflammatory response. At d 30, the SBM, SBM25CU and SBM200CU groups had increased CD3ε and CD8α cell populations in the DI, which is consistent with this finding. The CD3ε- and CD8α-lymphocytes were mainly confined to the basal part of the DI epithelium with only a few CD3ε-labelled cells scattered in the lamina propria adjacent to stratum compactum. However, Bakke McKellep *et al*. reported that lamina propria adjacent to stratum compactum and stroma of complex folds were rich in CD3ε-labelled cells in DI presenting with SBMIE. The CD3ε- and CD8α related observations might imply that the SBM used in these studies differed in immunostimulatory properties. Interaction between *C*. *utilis* and SBM has been described by Grammes *et al*. where feeding 200 g/kg *C*. *utilis* together with SBM (200 g/ kg) prevented SBMIE development in salmon. In our study, the severity of SBMIE in fish fed the highest inclusion levels of *C*. *utilis* (i.e. SBM100CU and SBM200CU) in combination with SBM was similar to that of fish fed the SBM diet. Adding lower levels of *C*. *utilis* to SBM diets (i.e. SBM25CU and SBM50CU) resulted in a large variation within the groups, ranging from normal morphology to moderate SBMIE. However, there was a decreased presence of CD8α-cell population in the SBM25CU group compared with the SBM group indicating that *C*. *utilis* has an immunomodulating effect locally in the DI when included at a low level in the SBM diet. Thus, partial prevention of SBMIE occurred only with lower inclusion levels of *C*. *utilis* which might differ with the previous work. This inconsistency could be due to differences in the degree of bioactivity of the *C*. *utilis* yeast, or in the ANF content of the SBM, and/or the experimental conditions. For example, Øverland and Skrede have speculated that the inconsistent effect of yeast on host immunity can be attributed to yeast strain, fermentation conditions, and downstream processing when manufactured. Furthermore, Miadoková *et al*. concluded in their study that the biological activity of glucomannan isolated from *C*. *utilis* is dependent on the combined application with other biologically active compounds. Also, the host itself can be the reason for this inconsistency as it has been demonstrated that the severity of SBMIE can differ between strains of rainbow trout. Additionally, the composition of the diet can induce variation in both the mucosa-associated and digesta-associated microbiota in Atlantic salmon. The mRNA expression profile of *aqp8* gene in the DI in the diet groups (FM, FM200CU, SBM, SBM200CU) indicates an association of *aqp8* with the resulting DI morphological and immune cell responses to nutritional challenges observed in this study. Our results confirm previous findings showing suppression of *aqp8* gene expression in intestinal inflammatory processes in salmon such as SBMIE. The relation between *aqp8* expression and inflammation is consistent with cell- based studies showing that reduced *aqp8* expression is linked to increased oxidative cell stress damage and implying that *aqp8* is a key player in the maintenance of redox cellular status. Intestinal inflammation in salmon resulting from SBM has been shown to have wider systemic effects in plasma including increased insulin levels and reduced plasma bile salt and cholesterol levels. In the present study, we found three proteins uniquely expressed in the SBM group; fatty acid-binding protein intestinal, transketolase and 14-3-3 protein beta. Salmonids fed SBM often present with reduced lipid digestibility, which can explain the presence of fatty acid-binding proteins in the plasma, due to tissue damage in the distal intestine. The shift in diet at d 30 of the experiment showed that the developed enteritis observed in the DI of fish fed either SBM alone or in combination with *C*. *utilis* was resolved after feeding FM diet for 7 d. It is also important to point out that the degree of SBMIE was reduced from moderate to mild after 7 d feeding FM200CU diet to those fish fed SBM diet for 30 d. It can be suggested that all fish would have returned to a normal state if the experiment had lasted longer. Additionally, feeding FM200CU for 30 days prior to a diet shift to SBM did not prevent development of enteritis in the DI. The present study revealed that dietary challenge induced the differential expression of specific proteins of the proteome but did not demonstrate a direct link to the local changes in the DI. This might have several explanations, such as sampling point, high variability among individuals and difficulty to measure individual feed intake. It should be noted that there is a lack of ideal biomarkers in gastrointestinal diseases in humans, such as inflammatory bowel diseases (IBDs), which further elucidates the difficulties of finding changes in the plasma proteome that can be directly linked to local inflammation or immunomodulation of the intestine. Nevertheless, there were specific proteins that were significantly expressed in our study that are of interest. We found that in the SBM group, there is an increased expression of periostin when compared with FM group. Periostin is a matricellular protein, belonging to the fasciclin family, that are also defined as an extracellular matrix protein that binds to cell-surface receptors. This protein is an emerging biomarker reflecting type 2 inflammation in allergic diseases in humans, and it has also been suggested that periostin mediates intestinal inflammation in mice and promote tumorigenesis in humans. In both SBM and SBM200CU diet the expression of kininogen-1 is decreased. The plasma kallikrein-kinin system (KKS) has been shown to be activated in patients suffering with active IBDs where the high- molecular-weight kininogen and prekallikrein were significantly decreased in plasma assuming an increased KKS activation/consumption. Lipocalin is produced in the intestine by intestinal epithelial cells and is upregulated in presence of intestinal inflammation, and therefore is an attractive a biomarker for IBDs in humans. In our study, lipocalin was significantly decreased in plasma of the SBM200CU group. Additionally, L-lactate dehydrogenase was found to be differentially expressed in all dietary treatments, which would suggest that further experiments investigating the activity of this protein could be useful to elucidate the mechanism of action of yeast in salmon. L-lactate dehydrogenase, which catalyzes the conversion of lactate to pyruvate, consequently its activity increases under periods of muscular damage. Therefore, some authors suggest that the levels of lactate dehydrogenase in blood can be used as a diagnostic tool to predict cardiomyopathy syndrome or skeletal muscle inflammation on salmon, however the plasma levels do not correlate consistently with histological scores. Differences between the plasma protein profiles of fish receiving SBM diets (SBM and SBM200CU) and FM diet were expected, as systemic changes may accompany a local tissue inflammation, thus plasma protein profiling could be a useful tool to determine the association between systemic response and outcomes of nutritional local challenges. However, the overall changes in the plasma proteome may be due to specific components in the diets (e.g. immunomodulating compounds of *C*. *utilis*) rather than local responses towards inflammation. Although obvious differences between SBM and SBM200CU were not manifested at morphological levels, we show that the SBM200CU group had a significant differential expression of complement C1q-like protein 4 and complement component C7 precursor in the plasma when compared with control group (shown). Complement factors are part of the innate immune system that enhance the ability of antibodies and phagocytic cells to clear microbes and damaged cells and promote inflammation. C7 is a precursor protein that, together with four other proteins, forms the membrane attack complex (MAC) of complement. Complement C1q-like protein 4 has a C1q-like domain, similar to domains in C1 complex, and is assumed to take part in controlling aspects of inflammation, adaptive immunity and energy homeostasis. Non-specific cellular and non-cellular immune responses have been reported when microbial ingredients have been included in fish diets, and increase in complement activation in serum has been demonstrated in sea bass, jian carp and rainbow trout fed diets containing β-glucans. Also, increased activities of complement in plasma of Atlantic salmon receiving an intraperitoneal injection of glucans from *Saccharomyces cerevisiae* have been demonstrated. Interestingly, complement proteins were not differentially expressed in SBM or FM200CU groups, which suggests that the combination of SBM and yeast might trigger a different immune response than yeast alone. Additionally, heat shock protein 90, which is involved in the response to stress, was differentially expressed in the diets containing yeast (SBM200CU and FM200CU), but not in the group fed SBM alone. Currently, the effect of heat shock proteins on fish health is not clear, however our results showed that the inclusion of yeast in diets reduces its expression, which might be due to the immunomodulatory components present in the membrane of yeast. This study combined traditional methods of assessing local tissue responses with plasma proteomic analysis in order to achieve a better understanding of the effects of nutrition on fish health and disease, and to identify systemic protein profiles in response to the dietary treatments. The inclusion of 200 g/kg of *C*. *utilis* yeast to a FM based diet (i.e. FM200CU) as a novel protein source with potential functional properties produced similar DI morphology, immune cell population and gene expression profiles, compared with a FM based control. Feeding the SBM diet induced SBMIE while feeding lower inclusion levels of *C*. *utilis* in combination with SBM (i.e. SBM25CU and SBM50CU) reduced the severity of SBMIE. Interestingly, higher inclusion levels of *C*. *utilis* yeast (i.e. SBM100CU and SBM200CU) did not show any significant protection against SBMIE, which is in contrast to a previous study that has shown protection. Changes in the plasma proteomics between groups with different dietary challenges were observed, however, their potential for the identification biomarkers for health or disease is yet to be elucidated. The absence of any clear alterations of expression of proteins within specific biological pathways would indicate that the effects due to diets were mainly restricted to local tissue responses. In conclusion, our results suggest that C. utilis does not alter intestinal morphology or induce major changes in plasma proteome, and thus could be a high- quality alternative protein source with potential functional properties in diets for Atlantic salmon. # Supporting information The authors thank Aleksandra Bodura Göksu for helping with immunohistochemistry techniques and to Ricardo Tavares Benicio for helping with feed manufacture, feeding the fish and sampling. 10.1371/journal.pone.0218360.r001 Decision Letter 0 Angers Annie Academic Editor 2019 Annie Angers 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. 9 Jul 2019 PONE-D-19-15096 Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses PLOS ONE Dear Dr. Øverland, 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. Most importantly, you should pay attention to the statistical analysis of the proteomic data, as carefully explained by Reviewer 3. In addition, both Reviewers 1 and 2 express concerns about the discussion of the manuscript, which should be carefully revised.  We would appreciate receiving your revised manuscript by Aug 23 2019 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, Annie Angers, Ph.D. 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\. We note that you have included the phrase “data not shown” in your manuscript. Unfortunately, this does not meet our data sharing requirements. PLOS does not permit references to inaccessible data. We require that authors provide all relevant data within the paper, Supporting Information files, or in an acceptable, public repository. Please add a citation to support this phrase or upload the data that corresponds with these findings to a stable repository (such as Figshare or Dryad) and provide and URLs, DOIs, or accession numbers that may be used to access these data. Or, if the data are not a core part of the research being presented in your study, we ask that you remove the phrase that refers to these data. 3\. Please amend the manuscript submission data (via Edit Submission) to include author Leidy Lagos. \[Note: HTML markup is below. Please do not edit.\] 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: Partly Reviewer \#2: Yes Reviewer \#3: Partly \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes Reviewer \#2: I Don't Know 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: The present work studies how the addition of C. utilis in soybean meal based diets impacts or reverts the changes induced when compared to a control fishmeal based diet. The results show that low levels of yeast reduced the severity of the enteritis caused by soybean meal based diets and C. utilis is proposed as an alternative protein source in functional diets for Atlantic salmon. In my opinion, this is a very attractive and complete study that combines different techniques in order to evaluate Atlantic salmon intestinal health. However, I have some comments and concerns that I believe should be addressed before considering this manuscript for publication. Major concerns: The experimental design seems carefully planned and complex, with many groups and variables. However, not all groups were used for all analyses and there is no explanation why. For example, it is stated that low doses partially prevent SBMIE, but these groups were not included in the gene expression or proteomic analyses. In addition, an explanation of the different effect that low and high levels of C. utilis have on inflammation should be included. The discussion seems to focus on randomly selected proteins. There is the description of the function of some of these proteins but I fail to see the point and relevance in this particular study. In my opinion, the whole discussion should be revised to follow a particular relevance point regarding this research, animal model and application. Other comments: L71: This sentence looks weird. Maybe the authors meant: “…an inactive dry Candida utilis strain was used…” or “… an inactive dry yeast strain of Candida utilis… “. L108: Vaccinated against? When mentioning average body weight values the error should be included. For instance, L116. L123: The feeding strategy was changed to what? In general, the experimental design is not clearly explained and even though there is a nice explanatory figure the text should be more clear. L200: The amount of tissue used for RNA isolation should be stated. L215: The concentration of the primers should be stated. Also, in the supplementary figure of the primer sequences the symbol for GDPH is GADPH, please unify the nomenclature. L253: I think a reference to the salmon proteome should appear here. L269: Why a series of PCA? Were several analyses performed? If so, they should be described. Also, verify the use of analyses instead of analysis for the plural. L289: The average initial and final weight are mentioned (again I’m missing the errors), but there is no comparison among groups. Were there differences? Maybe a table should be included with the group values? L303: “described in detail below” (remove the “ed”) L293: section Histology: A panel of pictures showing the most significant changes (like the ones described in L308-312) would be nice. Also, how do the authors know exactly which cell types are present (L311-312) without specific markers? Some of these cells are difficult to identify in intestinal sections. L323-324: Why where those groups excluded from the figure? L366-367: The difference mentioned here is not represented in the figure, in fact, regarding the figure statistics these groups were not different in this parameter. L387: The sudden change in nomenclature does not simplify the reading of the data, it actually complicates it. The nomenclature should be maintained along the manuscript (including the figures). In addition, why were the groups with low yeast dose excluded from this analysis? This should be explained in detail, particularly because the work concludes that the low doses are the ones having more positive effects but there are no results about them in this part of the study. L403 and Fig6C: This heatmap seems to show differential expression values. If this is correlation I don’t understand against what each molecule is correlated. Please revise these results and explanations. L409: “Analyses were carried out first with….” Why first? Was there a second analysis? L445: These inconsistencies should be further explained. L473: Delete DI. L473-475: The statements about the role of aqp8 in homoeostasis should be further supported. L489-492: I do not see the point of mentioning these particular roles of MHCI here. Maybe the authors are trying to make a point I do not understand. L494: One study or several? If one, change for “A previous study”. Otherwise use plural. L509: Downregulation in which tissue? Figure 3: Scale bars should be added to the pictures. Figure 5: The fold change is relative to what? Reviewer \#2: Review of PONE-D-19-15096 ” Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses”. This manuscript describes the effects of using a yeast strain as a functional protein source to boost gut health in fish at risk of intestinal enteritis when fed soy bean-based meal. The study tackles a highly relevant problem and is generally very well written and clear. Overall, this study will provide an important contribution to the field and fit well into the PLoS one journal. Below I provide some general and more specific comments to help the authors improve the manuscript for publication. General comments I want to applaud the authors for performing such a holistic analysis opposed to using just one single method. This is very much needed and sets a great standard for future studies. However, what I miss a little is an attempt to better link all the data sets together in a biological way. Could simply be through discussing results, but in a way to e.g. link the genes tested for expression responses with the proteins analyses all the way to the tissue histological data. Maybe add a conceptual figure with the entire path from DNA -\> RNA -\> proteins -\> tissues and the highlight where you have generated data and results. As is, the different data types are treated quite independently and in a somewhat random order. For example, the discussion on lines 467-469 represents a good example of how gene expression data is integrated with insights from histology and immune response data. Contrary, on lines 484-488 is an example where the protein plasma data is mainly discussed in isolation, whereas it would benefit from a stronger focus on the underling biological interactions with e.g. gene expression, histology and immune response data. Specific comments L. 108: What was the fish vaccinated against? Was it a standard vaccination program or inly for a specific disease? Pease specify. L. 158: Which and how many fish were used for this immunohistochemistry analysis? Was it the same 8 fish as above or new ones? Please specify. L. 224: Do you mean “… as index” instead of “… an index”? L. 229 + 231 + throughout: Be consistent whether or not you add a space between the degree symbol and “C” when listing temperature values. L. 253: Please provide a reference for the salmon proteome reference database. L. 262-263: Can you clarify this description as it sounds like all individuals are already part of "a cluster" when you begin. So, when does the model know to end the analysis? Is it e.g. trying to minimise the number of clusters based on some preset value? This remains unclear as currently written. L. 275: Add space before “\[31\]”. L. 388-389: It is not clear to me how this "simplifies" except using fewer letters. The figures have plenty of space to write out the FM and SBM names respectively, and these are much more informative than having to cross read between the text here and the figure to know what D1-D4 represents. I would prefer to just keep the FM and SBM based names in the figures as well. Same for Fig. 7. L. 438-444: Could this maybe also be due to individual responses among fish, and that these responses are partly controlled by the different genotypes among individuals? For example, if host genotype selects for different gut-microbiota communities, one could speculate that such different gut environments will differ in their ability to respond to the provided C. utilis strains? I agree this is very speculative, but the explanation do warrant mention as opposed to only focusing on how external parameters could explain the observed variation in gut responses; which seems to be the standard way of interpreting these types of gut data. L. 503: re-write to "proteins with significant over expression in... ". The proteins themselves are not significant, it is their expression level that is. Figure 2: Please add labels for all Y axes. Figure 3: Please clarify the different colors in the figure images. Are CD8 positive cells the purple colored ones? Or are both CD3 and CD8 colored brown? If so, what are the purple stains? Figure 4: Discuss whether the low sample size (n=7) may cause type-2 errors for the non-significant comparisons? Figure 5: Add both a and b letters between the two diet categories in the plots where there are significant differences. Also, same concern as for figure 4 about regarding lack of statistical power with the relatively low sample sizes. Was any prior power analyses performed to inform the used sample sizes? There are some differences in e.g. plot C that could potentially be significant with more statistical power, which could for example be obtained through an increased sample size. Reviewer \#3: General Impressions: In this manuscript, the authors performed a feeding trial on Atlantic salmon in which fish were fed fish meal or soybean meal-based diets that were supplemented with varying amounts of Canadis utilis yeast. Changes in intestinal morphology and the plasma proteome of the fish were monitored at several points in the study in order to monitor both intestinal and systemic effects of the different diets. Overall, the experimental design appears valid and the data collection and analysis of the histological data is appropriate. However, in the eyes of this reviewer, the analysis of the plasma proteomics data could be improved to allow for more meaningful interpretation of the data. The current methodology does not provide significant evidence that the observed changes in the plasma proteome are related to the differences in diet. However, after the authors address the following comments, the manuscript would likely be suitable for publication. Major Comments: • The authors relied primarily on Spearman’s correlation coefficients to identify proteins of interest from the plasma proteomics data. Therefore, the authors are essentially looking for cases where protein abundances do or do not correlate across the different diets. However, since there is a clear control in the experiment (fish meal-based diet), it seems that comparing all protein abundance to that control would be more informative. Then, emphasis would be placed on proteins that have ‘abnormal’ expression relative to the fishmeal fed fish. This would also allow for volcano plots to be generated and included in the manuscript, which would aid reader understanding. • Further description of how the quantitative values used in the proteomics analysis is needed. It appears that precursor abundance (from the MS1 scans) were used, but that is not explicitly stated. If that is the case, was peak intensity or peak area used? Also, was the ‘match between runs’ feature enabled in MaxQuant? • A table showing all of the proteins identified in the plasma analyses and their quantitative values should be provided in the supplement. Ideally, fold change values should also be included. It’s possible that such a table was uploaded to PRIDE, but the PXD number provided does not bring up any results when searched. This is likely because the dataset has not been publicly released but the login info should be provided to reviewers. • It is concerning that the fishmeal-based and soybean-based diets don’t cluster in either the heatmap or the PCA plots. It seems likely that such large differences in the feed would causes systemic changes in the blood which would lead to high correlation between the D1 and D7 samples, but that was not observed. This suggests that either the analysis method was flawed, or that individual variation was too large in order to adequately detect differences from the feed. • Although this reviewer is not well versed in the VIP method used here, it seems that this technique is likely to put too much emphasis on proteins that are correlated purely by chance. Further justification of this technique should be included, or the analyses should be removed. • It would be helpful if the reviewers included a few representative images of the changes in morphology observed between the groups that are summarized in Fig. 2. This could be included as an additional supplemental figure. Minor Comments: • L252: The fragment ion tolerance used here is much too large for orbitrap data. A tolerance of 0.02 Da would be more appropriate. • L253: Where was the salmon proteome database derived from? Was it a public repository such as uniprot? If so, when was it downloaded? • L316: No data is shown in this figure where SBM + C. utilis was fed first, followed by FM. This sentence says otherwise. • Fig. 4 caption – L371: What is meant hear by ‘data are expressed as mean � SD for each individual’? Isn’t each point just representative of the mean for each fish? This should be clarified. • L407-408: It doesn’t make sense to say that proteins ‘belong’ to a specific domain. Please revise. • Fig. 7 caption – L424: What do the colors in panel C represent? Also, why are proteins with VIP scores \< 2 included here? • L473: Remove ‘DI.’ • L476: Change considering to considered. \*\*\*\*\*\*\*\*\*\* 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 \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] 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.0218360.r002 Author response to Decision Letter 0 23 Aug 2019 Reviewer \#1: The present work studies how the addition of C. utilis in soybean meal based diets impacts or reverts the changes induced when compared to a control fishmeal based diet. The results show that low levels of yeast reduced the severity of the enteritis caused by soybean meal based diets and C. utilis is proposed as an alternative protein source in functional diets for Atlantic salmon. In my opinion, this is a very attractive and complete study that combines different techniques in order to evaluate Atlantic salmon intestinal health. However, I have some comments and concerns that I believe should be addressed before considering this manuscript for publication. Major concerns: The experimental design seems carefully planned and complex, with many groups and variables. However, not all groups were used for all analyses and there is no explanation why. For example, it is stated that low doses partially prevent SBMIE, but these groups were not included in the gene expression or proteomic analyses. In addition, an explanation of the different effect that low and high levels of C. utilis have on inflammation should be included. The discussion seems to focus on randomly selected proteins. There is the description of the function of some of these proteins but I fail to see the point and relevance in this particular study. In my opinion, the whole discussion should be revised to follow a particular relevance point regarding this research, animal model and application. Unfortunately, we could not analyze all samples in all analyses. Based on earlier experience with microbial ingredients, we assumed that the highest concentratiosn of CU would have the most effect when added to SBM based diet. Thus e.g., in the proteomic analyses, only FM (D1), SBM (D2), SBM200CU (D6) and FM200CU (D7) groups were subjected to analyses. In other analyses, we have included higher number of samples/dietary treatments. We have revised and re-arranged large parts of the discussion. Other comments: L71: This sentence looks weird. Maybe the authors meant: “…an inactive dry Candida utilis strain was used…” or “… an inactive dry yeast strain of Candida utilis… “. Changed to: “… an inactive dry yeast strain of Candida utilis…” L108: Vaccinated against? Farmed Atlantic salmon in Norway are routinely vaccinated as pre-smolts against a number of diseases. In this case, the Aquavac PD7 was used (contains vaccine against bacterial furunculosis, vibriosis, cold water vibriosis, winter ulcer) and viral (infectious pancreatic necrosis (IPN), pancreas disease, infectious salmon anaemia (ISA)) diseases. We have added this, thus the sentence now reads: “Vaccinated salmon (Aquavac PD7, MSD Animal Health, Bergen, Norway) were acquired from…”. When mentioning average body weight values the error should be included. For instance, L116. To avoid too much stress for the fishes at the start of the experiment, it is common to register bulk weigh of all fish per tank. Thus, in this experiment, we have only the bulk starting weight of the 20 individuals per tank. We can therefore only give an average initial body weight (526 g) withour the error. L123: The feeding strategy was changed to what? In general, the experimental design is not clearly explained and even though there is a nice explanatory figure the text should be more clear. A paragraph was added in order to make the feeding strategy more clear: Following the acclimation period, each experimental diet was randomly allocated to the fish tanks (two tanks/diet) and fed for 30 days (period 1) as described above. After 30 days, the feeding strategy were changed and new diets were fed for 7 days (period 2). As a control, one fish group received FM throughout the experiment. To assess whether C. utilis were able to counteract enteritis induced by SBM, four fish groups received SBM diets combined with different inclusions levels of CU (i.e. SBM25CU, SBM50CU, SBM100CU, SBM200CU) in period 1. One fish group received FM200CU in period 1 to evaluate if C. utilis in combination with FM alone would have an effect on the DI. In period 2, the ability of C. utilis to prevent SBMIE was assessed as the diet was changed to SBM in this group. Finally, three fish groups were fed SBM diet to induce SBMIE in period 1, and in period 2 the diets were changed to either FM, FM200CU and SBM200CU to evaluate if these diets were able to reverse SBMIE. The feeding strategy is illustrated in Fig. 1. Subsequently L132-135 were changed to: At each sampling point (0, 7, 30 and 37 days), 8 fish per diet (4 fish per tank) were randomly selected and anaesthetized by immersion in 60 mg/l of tricaine methanesulfonate (MS-222, Sigma-Aldrich, MO, USA) and subsequently euthanized by a sharp blow to the head. L200: The amount of tissue used for RNA isolation should be stated. The sentence have been changed to: A small piece of DI tissue (approximately 0.5 cm) from FM, FM200CU, SBM and SBM200CU diet groups (8 fish/diet) at day 30 were subject to gene expression analysis. L215: The concentration of the primers should be stated. Also, in the supplementary figure of the primer sequences the symbol for GDPH is GADPH, please unify the nomenclature. The primer concentration has been added. GDPH had been changed to GAPDH L253: I think a reference to the salmon proteome should appear here. The salmon proteome was added to the main text (<https://www.uniprot.org/proteomes/?query=taxonomy:8030>) L269: Why a series of PCA? Were several analyses performed? If so, they should be described. Also, verify the use of analyses instead of analysis for the plural. The description of the analysis was edited for a better understanding. “An initial PCA was used as an unsupervised method to find the directions of maximum covariance among FM, SBM, SBM200CU and FM200CU was performed using the prcomp package in R to see the distribution of the proteins present in all the groups”. L289: The average initial and final weight are mentioned (again I’m missing the errors), but there is no comparison among groups. Were there differences? Maybe a table should be included with the group values? This was not a growth experiment. But to be sure that fish in all dietary treatments were eating, and growing, the feed intake was registered. Fish weights were only registered as bulk weight per tank at the start of the experiment, and as individual body weights for the fishes that were sampled in weeks 0, 1, 4 (period 1) and 5 (period 2). To avoid unnecessary disturbance, the fish left in the tank at the different time points were not taken out for weighing. Thus at the end of the experiment, there were very few fish left per tank. We feel that growth responses should therefore not be presented in a table. However, there were no significant differences due to dietary treatment when it comes to both feed intake nor growth. Thus we have added “nor growth rate” to make it more clear to the reader. The new sentence reads: “All groups of fish accepted their allocated diets and no significant differences were found in feed intake nor growth rate among dietary treatment”. L303: “described in detail below” (remove the “ed”) The authors don’t think it will be correct to remove “ed” in this context and want to leave it as it is. L293: section Histology: A panel of pictures showing the most significant changes (like the ones described in L308-312) would be nice. Also, how do the authors know exactly which cell types are present (L311-312) without specific markers? Some of these cells are difficult to identify in intestinal sections. A panel of pictures (S2 Fig.) is included showing the overall differences between the groups but also variation within the SBM25 group. Yes, it is indeed difficult to identify specific leucocytes in intestinal sections but not impossible. There was not made an effort to make a semi- quantitative assessment of the various cell types in the different diets, thus only the leucocytes that were in abundancy will be mentioned. Therefore, L309-312 will be reduced to: “There was an increased presence of connective tissue in the lamina propria and the increased infiltration of leucocytes consisted mainly of eosinophilic granule cells and to a lesser extent of lymphocytes.” L323-324: Why where those groups excluded from the figure? L323-324: these two SBM groups were similar with the SBM group presented in the figure and therefore omitted from the figure to generate equal groups. The individuals from both of these groups are now included making the SBM group n = 23. L366-367: The difference mentioned here is not represented in the figure, in fact, regarding the figure statistics these groups were not different in this parameter. The data for morphometric measurements and CD-densities were analyzed in JMP, but the figure was made in GraphPad. Running wrong statistical analysis in GraphPad resulted in wrong statistical differences expressed as “a”, “b” and “c” in the figure, and consequently, the wrong figure was submitted. The significant differences were presented in the result section and discussed in the text according to the statistics that were originally run in JMP. To simplify the text, the data set has been rerun in GraphPad and the correct figure has been submitted in the revised version. L240-247 have been updated to: “Non-parametric data from the histological evaluation were analyzed by Kruskal- Wallis followed by post hoc Dunn’s test with a comparison of mean rank. Shapiro- Wilk normality test was used to test the normal distribution of the data from morphometric analyses and T-cell density, and further were analyzed by one-way ANOVA followed by Tukey’s multiple comparisons test. Morphometric analyses and T-cell density analyses were performed at the individual level using the mean of measurements of between 2-6 simple folds per fish. Results of qPCR (means ± standard deviations) were analyzed using One-way ANOVA with Dunnett\`s multiple comparison test (a \< 0.0001). These analyses were performed in GraphPad Prism, version 7.00 and 8.0.1 (GraphPad Software Inc., San Diego, CA, USA). L387: The sudden change in nomenclature does not simplify the reading of the data, it actually complicates it. The nomenclature should be maintained along the manuscript (including the figures). In addition, why were the groups with low yeast dose excluded from this analysis? This should be explained in detail, particularly because the work concludes that the low doses are the ones having more positive effects but there are no results about them in this part of the study. Besides the explanation included in the main text, we included a extra legend on Fig 6 and 7, referring to the different groups and its nomenclature. The aim of the proteomic analysis was to identify potential biomarkers or protein express uniquely on the SBM induced enteritis groups. Therefore, we choose the extreme groups. Nevertheless, we did not observe unique proteins which can be used as markers. Then we focus on the proteins that were differently expressed, but present in all the groups. We agree with the Reviewers at it could be very interesting to run proteomic analysis considering the low dose, it will be considered in our future trials. L403 and Fig6C: This heatmap seems to show differential expression values. If this is correlation I don’t understand against what each molecule is correlated. Please revise these results and explanations. The heatmap was generated based on the pairwise comparisons across the four time points. As per your suggestion, to be simplified and easy for the audience to follow, we have re-analyzed and produced three heatmaps, with each diet treatment to the control (FM). The old heatmap figure is replaced with new heatmaps (Fig 6C). L409: “Analyses were carried out first with….” Why first? Was there a second analysis? The description of the analysis was edited as follow: “PCA analysis was carried out first with all selected proteins across dietary groups, then PLS-DA multivariate analyses were performed to detect the proteins responsible for the differentiation between dietary groups (Fig 7D).” L445: These inconsistencies should be further explained. These inconsistencies have further been explained with: Furthermore, Miadoková et al concluded in their study that the biological activity of glucomannan isolated form C. utilis dependent on the combined application with other biologically active compounds \[56\]. Also, the host itself can be the reason for this inconsistency. It has been demonstrated that the severity of SBMIE can differ between strains of rainbow trout \[57\], and that diet can induce variation in both the mucosa-associated and digesta- associated microbiota in Atlantic salmon \[58, 59\]. L473: Delete DI. The discussion has been rewritten, thus this is not relevant in the new discussion. L473-475: The statements about the role of aqp8 in homoeostasis should be further supported. L473-475 has been removed as our data does not support this statement. L489-492: I do not see the point of mentioning these particular roles of MHCI here. Maybe the authors are trying to make a point I do not understand. This paragraph has been rewritten in the discussion. L494: One study or several? If one, change for “A previous study”. Otherwise use plural. This line has been rewritten as part of the rearrangement of the discussion. L509: Downregulation in which tissue? In the study performed by Grammes et al, a downregulation of LYME mRNA was observed in DI. Our study indicate that in DI inflammation there are high plasma levels of LYME. Figure 3: Scale bars should be added to the pictures. Scale bars are added to the pictures Figure 5: The fold change is relative to what? The relative fold changes are calculated in relation to the FM group. Reviewer \#2: Review of PONE-D-19-15096 ” Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses”. This manuscript describes the effects of using a yeast strain as a functional protein source to boost gut health in fish at risk of intestinal enteritis when fed soy bean-based meal. The study tackles a highly relevant problem and is generally very well written and clear. Overall, this study will provide an important contribution to the field and fit well into the PLoS one journal. Below I provide some general and more specific comments to help the authors improve the manuscript for publication. General comments I want to applaud the authors for performing such a holistic analysis opposed to using just one single method. This is very much needed and sets a great standard for future studies. However, what I miss a little is an attempt to better link all the data sets together in a biological way. Could simply be through discussing results, but in a way to e.g. link the genes tested for expression responses with the proteins analyses all the way to the tissue histological data. Maybe add a conceptual figure with the entire path from DNA -\> RNA -\> proteins -\> tissues and the highlight where you have generated data and results. As is, the different data types are treated quite independently and in a somewhat random order. An attempt to link the data sets together in a biological way have been made in the rewritten discussion. For example, the discussion on lines 467-469 represents a good example of how gene expression data is integrated with insights from histology and immune response data. Contrary, on lines 484-488 is an example where the protein plasma data is mainly discussed in isolation, whereas it would benefit from a stronger focus on the underling biological interactions with e.g. gene expression, histology and immune response data. We agree with the Reviewers. Therefore the discussion have been rewritten. Specific comments L. 108: What was the fish vaccinated against? Was it a standard vaccination program or inly for a specific disease? Pease specify. Farmed Atlantic salmon in Norway are routinely vaccinated as pre-smolts against a number of diseases. In this case, the Aquavac PD7 was used by Sørsmolt AS. We have added this information, thus the sentence now reads: “Vaccinated salmon (Aquavac PD7, MSD Animal Health, Bergen, Norway) were acquired from…”. L. 158: Which and how many fish were used for this immunohistochemistry analysis? Was it the same 8 fish as above or new ones? Please specify. The same 8 fish from each diet group were used for histology, immunohistochemistry and qPCR. L158-159 has been changed in order to make this clearer: Histological sections of DI from the fish sampled at day 30 (8 fish/diet), prepared as described above, was subjected for immunohistochemical analysis, and the following diet groups were included: FM, FM200CU, SBM, SBM25CU and SBM200CU. L. 224: Do you mean “… as index” instead of “… an index”? L224 changed to “… an index” L. 229 + 231 + throughout: Be consistent whether or not you add a space between the degree symbol and “C” when listing temperature values. When listing temperature values, all have been changed to °C with no space between degree symbol and “C”. L. 253: Please provide a reference for the salmon proteome reference database. The salmon proteome as been included in material and method L. 262-263: Can you clarify this description as it sounds like all individuals are already part of "a cluster" when you begin. So, when does the model know to end the analysis? Is it e.g. trying to minimise the number of clusters based on some preset value? This remains unclear as currently written. We agree and the sentence have been modifies as follow: “The clustering results are presented in the form of a heatmap, with levels of protein expression across the three dietary groups (SBM (D2), SBM200CU (D6) and FM200CU (D7) compared to the control FM (D1). Hierarchical clustering was performed with the hclust function in R package stat. UniprotKB database was used for functional annotation of the proteins.” L. 275: Add space before “\[31\]”. Space has been added before “\[31\]” L. 388-389: It is not clear to me how this "simplifies" except using fewer letters. The figures have plenty of space to write out the FM and SBM names respectively, and these are much more informative than having to cross read between the text here and the figure to know what D1-D4 represents. I would prefer to just keep the FM and SBM based names in the figures as well. Same for Fig. 7. Besides the explanation included in the main text, we included an extra table on Fig 6 and 7, referring to the different groups and its nomenclature. L. 438-444: Could this maybe also be due to individual responses among fish, and that these responses are partly controlled by the different genotypes among individuals? For example, if host genotype selects for different gut-microbiota communities, one could speculate that such different gut environments will differ in their ability to respond to the provided C. utilis strains? I agree this is very speculative, but the explanation do warrant mention as opposed to only focusing on how external parameters could explain the observed variation in gut responses; which seems to be the standard way of interpreting these types of gut data. This is a good point which have already been briefly mentioned/discussed in the following sentences in the manuscript: “Also, the host itself can be the reason for this inconsistency. It has been demonstrated that the severity of SBMIE can differ between strains of rainbow trout (Venold 2012), and that diet can induce variation in both the mucosa- associated and digesta-associated microbiota in Atlantic salmon (Gajardo 2016, 2017)”. However, as we do not have the actual genotypes of these salmon (they were purchased from Sørsmolt AS), a further discussion of this point would be just speculation. L. 503: re-write to "proteins with significant over expression in... ". The proteins themselves are not significant, it is their expression level that is. This line has been rewritten as part of the rearrangement of the discussion. Figure 2: Please add labels for all Y axes. The label “Scale” is added for all Y axes. Figure 3: Please clarify the different colors in the figure images. Are CD8 positive cells the purple colored ones? Or are both CD3 and CD8 colored brown? If so, what are the purple stains? Color development of both CD3 and CD8 was performed by 3,3’-diaminobenzidine, therefore both CD3 and CD8 positive cells are labeled with brown color. As the detection of these cells were run in two different assays where specific antibody was applied, we can count positive CD3-labelled cells when antibody towards CD3 was applied and we can count positive CD8-labeled cells when antibody towards CD8 was applied. In both cases, the slide was counterstained with hematoxylin that gives a purple color in order to visualize to the structure of the rest of the intestinal mucosa. Figure 4: Discuss whether the low sample size (n=7) may cause type-2 errors for the non-significant comparisons? The sample size was in fact n=8, except for the SBM group, but we agree that this is indeed a low sample size. We would of course have preferred to have a higher number of fish, but the experimental set-up, number of available tanks and limit of biomass in each tank did not allow for a higher number. A larger sample size, e.g., by one more tank per dietary treatment (i.e., 3 tanks instead of 2) could have strengthened our results. However, since we have measured/analyzed many different parameters in the same individual fish, and these analyses do support each other, we still feel confident with these results. Figure 5: Add both a and b letters between the two diet categories in the plots where there are significant differences. Letter a and b were added between the two diets where there are a significant differences Also, same concern as for figure 4 about regarding lack of statistical power with the relatively low sample sizes. Was any prior power analyses performed to inform the used sample sizes? There are some differences in e.g. plot C that could potentially be significant with more statistical power, which could for example be obtained through an increased sample size. Unfortunately, the wrong figure was submitted. The correct figure is submitted in the revised version, and here, the differences are significant. Reviewer \#3: General Impressions: In this manuscript, the authors performed a feeding trial on Atlantic salmon in which fish were fed fish meal or soybean meal-based diets that were supplemented with varying amounts of Canadis utilis yeast. Changes in intestinal morphology and the plasma proteome of the fish were monitored at several points in the study in order to monitor both intestinal and systemic effects of the different diets. Overall, the experimental design appears valid and the data collection and analysis of the histological data is appropriate. However, in the eyes of this reviewer, the analysis of the plasma proteomics data could be improved to allow for more meaningful interpretation of the data. The current methodology does not provide significant evidence that the observed changes in the plasma proteome are related to the differences in diet. However, after the authors address the following comments, the manuscript would likely be suitable for publication. Major Comments: The authors relied primarily on Spearman’s correlation coefficients to identify proteins of interest from the plasma proteomics data. Therefore, the authors are essentially looking for cases where protein abundances do or do not correlate across the different diets. However, since there is a clear control in the experiment (fish meal-based diet), it seems that comparing all protein abundance to that control would be more informative. Then, emphasis would be placed on proteins that have ‘abnormal’ expression relative to the fishmeal fed fish. This would also allow for volcano plots to be generated and included in the manuscript, which would aid reader understanding. Yes, we re-analyzed the data. According to your suggestions, we have generarted the volcano plots and heat maps, by comparing each diet treatment to the control (FM). The old heatmap figure is replaced with new heatmaps (Fig 6C) and new volcano plots were added (Fig 7A, 7B, 7C). • Further description of how the quantitative values used in the proteomics analysis is needed. It appears that precursor abundance (from the MS1 scans) were used, but that is not explicitly stated. If that is the case, was peak intensity or peak area used? Also, was the ‘match between runs’ feature enabled in MaxQuant? We performed the anlysis in Mascot and MaxQuant, however, we used the LFQ on MS2 level (Spectral count) and not on MS1 (Peak areas) as in MaxQuant to perform the analysis. The MS raw file were analyzed in both Mascot and Scaffold to get higher confidence in the protein identified. • A table showing all of the proteins identified in the plasma analyses and their quantitative values should be provided in the supplement. Ideally, fold change values should also be included. It’s possible that such a table was uploaded to PRIDE, but the PXD number provided does not bring up any results when searched. This is likely because the dataset has not been publicly released but the login info should be provided to reviewers. Yes, an excel sheet containing all the proteins identified with their quantitative values and fold change values are submitted as a supplementary table (S3 Table). • It is concerning that the fishmeal-based and soybean-based diets don’t cluster in either the heatmap or the PCA plots. It seems likely that such large differences in the feed would causes systemic changes in the blood which would lead to high correlation between the D1 and D7 samples, but that was not observed. This suggests that either the analysis method was flawed, or that individual variation was too large in order to adequately detect differences from the feed. It is important to mention that we did not observe severe enteritis on the SBM fed fish, which might have influenced the results on plasma proteome. On the other hand, as mentioned by Reviewer 3, we observed a very big variation between individual. This is a persistent concern in feeding trial with fish, since there is not an available method to measure individual feed intake as in mammals. Therefore we use each tank as unit and calculate feed intake per tank. Although we perform histology to verify the degree of enteritis of each fish used in the analysis, we are not sure how much of the antinutrient (SBM) each fish ingested. Our main aim goal was to identify proteins uniquely expressed in each group, but it seems at SBM enteritis is restricted to a local inflammation (distal intestine) instead of a systemic inflammation. • Although this reviewer is not well versed in the VIP method used here, it seems that this technique is likely to put too much emphasis on proteins that are correlated purely by chance. Further justification of this technique should be included, or the analyses should be removed. In the initial PCA plot (Fig 7D), there is significant overlap of dietary groups, which leads to difficulty in discerning which proteins are responsible for the changes seen between the dietary groups. To address this problem, we used a known approach (Eriksson L, Umetrics AB (2006) Multi- and Megavariate Data Analysis, Part 1, Basic Principles and Applications: Umetrics AB.) using Partial Least Squares – Discriminant Analysis (PLS-DA) Variable Importance Projection (VIP) scores to highlight the significant changes between these groups. These VIP scores are normally used for variable selection as heat maps combined with group difference proteins to highlight the significant differences in dietary groups. We don’t think, this VIP proteins are correlated purely by chance, as in the discussion, we have highlighted about the role of these proteins on immune system and how they are related with SBM enteritis. • It would be helpful if the reviewers included a few representative images of the changes in morphology observed between the groups that are summarized in Fig. 2. This could be included as an additional supplemental figure. We added Fig. 2S, where morphologic changes are presented due to the diets. Minor Comments: • L252: The fragment ion tolerance used here is much too large for orbitrap data. A tolerance of 0.02 Da would be more appropriate. We appreciate the comment of Reviewer 3, the MS analysis was performed at Proteomic Core Facility of University of Oslo. But the recommendation will be taken into account in our future analysis. • L253: Where was the salmon proteome database derived from? Was it a public repository such as uniprot? If so, when was it downloaded? The reference salmon proteome was download from a public repository UniProt, as mention in the main text. The date of downloaded was December 2017. • L316: No data is shown in this figure where SBM + C. utilis was fed first, followed by FM. This sentence says otherwise. All the groups are included in the new figure. • Fig. 4 caption – L371: What is meant hear by ‘data are expressed as mean � SD for each individual’? Isn’t each point just representative of the mean for each fish? This should be clarified. Data are expressed as the mean for all individuals � SD. • L407-408: It doesn’t make sense to say that proteins ‘belong’ to a specific domain. Please revise. Yes, we revised the sentence and changed to “most of the proteins have signal protein domain”. • Fig. 7 caption – L424: What do the colors in panel C represent? Also, why are proteins with VIP scores \< 2 included here? The colors in the panel C represents the protein abundance at 4 dietary treatments. Red represents higher and dark green represents lower abundance. This sentence is added to the figure legend. We have replaced the figure with new one showing the proteins only with VIP \>2. • L473: Remove ‘DI.’ This line has been rewritten as part of the rearrangement of the discussion. • L476: Change considering to considered. This line has been rewritten as part of the rearrangement of the discussion. 10.1371/journal.pone.0218360.r003 Decision Letter 1 Angers Annie Academic Editor 2019 Annie Angers 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 Sep 2019 PONE-D-19-15096R1 Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses PLOS ONE Dear Dr. Øverland, 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. While all three reviewers consider the general response to comments satisfactory and that the manuscript is significantly improved, Reviewer 3 has underlined new concerns regarding the proteomic data analysis. Based on the possibility that the conclusion reached in the manuscript may be based on invalid interpretation of the results. The reviewer kindly suggests many viable options to overcome this conundrum, and I am therefore confident that the matter can be resolved. Specific comments by both reviewers 2 and 3 should also be addressed prior to publication. We would appreciate receiving your revised manuscript by Oct 28 2019 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, Annie Angers, Ph.D. Academic Editor PLOS ONE \[Note: HTML markup is below. Please do not edit.\] 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: (No Response) 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 \#2: Yes Reviewer \#3: Partly \*\*\*\*\*\*\*\*\*\* 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: Yes 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: (No Response) Reviewer \#2: I have now been through this revised version of the manuscript. I am generally happy with all the authors responses to my original review. So I am happy to hereby recommend the manuscript for publication. But I do have a few minor comments to help with some final improvements: 1\) Regarding the Y-axes labels on figure 2, it may be more appropriate to label them "Change" rather than "Scale" that seems not very informative. 2\) I suggest that the authors either delete or re-write the very kast sentence of the discussion: "Further research is needed to evaluate the impact of yeast strain and the fermentation and/or down-stream processing conditions of the yeast on functional properties in relation to gastro intestinal health and systemic responses." As is, this is not very informative. Consider adding some more concrete suggestions as to how such "Further research" can be designed to better help move this entire research field and add more values to readers of the paper. Reviewer \#3: The authors addressed a number of concerns raised by the reviewers and the manuscript has been significantly improved. Unfortunately, this reviewer still has significant concerns about the analysis of the proteomics data and the conclusions that are drawn from this data. In the original manuscript, it was not made clear that the authors were utilizing spectral counting to obtain quantitative information from the proteomics data. This should have been stated in the manuscript, and needs to be added. The spectral counting method has been phased out in the field due to its low reproducibility and accuracy. When using a high resolution, accurate mass instrument such as the Q Exactive that was used here, MS1 intensity-based quantitation provides far superior quantitative values than does spectral counting. The reasons for this are well described in the literature, but are primarily the result of the somewhat stochastic nature of the data collection when data-dependent acquisition strategies are employed (as was done in this study). The fact that an MS2 mass tolerance was used for searching that was more than an order of magnitude higher than it should have been (0.8 Da rather than 0.02 Da) further reduces the quantitative accuracy of the spectral counting strategy used here. For the reasons described above, the quantitative values presented should be treated with caution. The proteins shown to be differentially expressed in the volcano plots (Fig. 7) are likely valid, as this analysis incorporates both the calculated fold change, as well as a p-value, which incorporates other variables such as inter-sample variability. Unfortunately, many of the proteins that the authors focus on in the discussion (such as HMP, LYME, and HIIN) were not shown to be differentially expressed by this analysis, they simply didn't correlate well to the control. This alone is not adequate evidence of a change in protein abundance. With these concerns in mind, this reviewer suggests the authors choose one of 3 options: 1\. Reanalyze the data using MS1 intensity-based quantitation (as can be completed in MaxQuant or other software packages such as Scaffold Q+, Proteome Discoverer, etc.). 2\. Focus the discussion only on proteins that showed differential abundance in the volcano plot analyses (Fig. 7) 3\. Remove the proteomics data entirely and focusing on the gene expression, morphological, and other data presented in the manuscript. Any of the above approaches would be valid and would make the publication suitable for publication in the opinion of this reviewer. Specific Comments: L262: The date the database was downloaded should also be included. L403: Don't say 'expressed' here. The proteins were likely expressed in the samples, they just couldn't be detected by the mass spectrometer. Figure 6: The Spearman's correlations done here do not provide sufficient evidence to say a protein is changing abundance. Panel C should be removed from the manuscript as well as all discussion of the results presented in this panel. \*\*\*\*\*\*\*\*\*\* 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 \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] 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.0218360.r004 Author response to Decision Letter 1 28 Oct 2019 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: (No Response) Reviewer \#2: I have now been through this revised version of the manuscript. I am generally happy with all the authors responses to my original review. So I am happy to hereby recommend the manuscript for publication. But I do have a few minor comments to help with some final improvements: 1\) Regarding the Y-axes labels on figure 2, it may be more appropriate to label them "Change" rather than "Scale" that seems not very informative. The label on the Y-axis has been changed from "scale" to "change" according to the reviewer’s request. 2\) I suggest that the authors either delete or re-write the very kast sentence of the discussion: "Further research is needed to evaluate the impact of yeast strain and the fermentation and/or down-stream processing conditions of the yeast on functional properties in relation to gastro intestinal health and systemic responses." As is, this is not very informative. Consider adding some more concrete suggestions as to how such "Further research" can be designed to better help move this entire research field and add more values to readers of the paper. The last sentence has been deleted as the discussion has been changed. Reviewer \#3: The authors addressed a number of concerns raised by the reviewers and the manuscript has been significantly improved. Unfortunately, this reviewer still has significant concerns about the analysis of the proteomics data and the conclusions that are drawn from this data. In the original manuscript, it was not made clear that the authors were utilizing spectral counting to obtain quantitative information from the proteomics data. This should have been stated in the manuscript, and needs to be added. The spectral counting method has been phased out in the field due to its low reproducibility and accuracy. When using a high resolution, accurate mass instrument such as the Q Exactive that was used here, MS1 intensity-based quantitation provides far superior quantitative values than does spectral counting. The reasons for this are well described in the literature, but are primarily the result of the somewhat stochastic nature of the data collection when data-dependent acquisition strategies are employed (as was done in this study). The fact that an MS2 mass tolerance was used for searching that was more than an order of magnitude higher than it should have been (0.8 Da rather than 0.02 Da) further reduces the quantitative accuracy of the spectral counting strategy used here. For the reasons described above, the quantitative values presented should be treated with caution. The proteins shown to be differentially expressed in the volcano plots (Fig. 7) are likely valid, as this analysis incorporates both the calculated fold change, as well as a p-value, which incorporates other variables such as inter-sample variability. Unfortunately, many of the proteins that the authors focus on in the discussion (such as HMP, LYME, and HIIN) were not shown to be differentially expressed by this analysis, they simply didn't correlate well to the control. This alone is not adequate evidence of a change in protein abundance. With these concerns in mind, this reviewer suggests the authors choose one of 3 options: 1\. Reanalyze the data using MS1 intensity-based quantitation (as can be completed in MaxQuant or other software packages such as Scaffold Q+, Proteome Discoverer, etc.). 2\. Focus the discussion only on proteins that showed differential abundance in the volcano plot analyses (Fig. 7) 3\. Remove the proteomics data entirely and focusing on the gene expression, morphological, and other data presented in the manuscript. Any of the above approaches would be valid and would make the publication suitable for publication in the opinion of this reviewer. We are grateful for the specificity and professional comments of reviewer 3. We decided to take option number 1, which means at we reanalyzed the entire dataset using MaxQuant and Perseus. The LFQ intensities were exported to an excel sheet. The log-transformed LFQ values were analyzed using the pipeline described in material and methods. We created volcano plots, comparing the protein pattern of each dietary treatment with the control diet. This comparison resulted in 9, 12 and 10 significantly expressed protein when the FM was compared with SBM, SBM200CU and FM200CU, respectively. Although option 1 was more complicated, we are very grateful for the comments, which helped us to increase the reliability of our results. Specific Comments: L262: The date the database was downloaded should also be included. We used the latest salmon proteome available and the date is included in the manuscript L403: Don't say 'expressed' here. The proteins were likely expressed in the samples, they just couldn't be detected by the mass spectrometer. The sentence was changed as marked in the manuscript. Figure 6: The Spearman's correlations done here do not provide sufficient evidence to say a protein is changing abundance. Panel C should be removed from the manuscript as well as all discussion of the results presented in this panel. In accordance with the reviewer’s advice, Panel C in Figure 6 has been removed from the manuscript and the discussion revised accordingly. 10.1371/journal.pone.0218360.r005 Decision Letter 2 Angers Annie Academic Editor 2019 Annie Angers 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 Nov 2019 PONE-D-19-15096R2 Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses PLOS ONE Dear Dr. Øverland, 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. The manuscript is much improved and almost ready for publication. However, the final concerns raised by Reviewer 3 still need to be addressed, and minor adjustments to the text are also needed. We would appreciate receiving your revised manuscript by Dec 26 2019 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, Annie Angers, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (if provided): \- Please address the possible data discrepancies between table S3 and table 2 referred to by Reviewer 3; \- Make sure that proteins that were not found in a replicate are not given a value of zero but treated as missing in your analysis (or clearly state that you did not use 0 in your analysis) \- Make the minor adjustments to text suggested by Reviewer 3. \[Note: HTML markup is below. Please do not edit.\] 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 \#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 \#3: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? 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 \#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 \#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 \#3: General Impressions: The authors have clearly addressed the concerns raised in the previous review, and the reanalysis of the proteomics data has dramatically improved the quality of the manuscript. I greatly appreciate the authors hard work and dedication to providing the highest quality data possible. I have noticed one discrepancy in the data analysis that should be addressed in the final version of the manuscript (described below), but I feel that the manuscript is otherwise ready for publication. I’ve also included a few very minor comments that I feel will improve the readability of the publication. Major Comments: The authors state that only protein that were identified in at least two of the four replicates were used for quantitative analysis. However, when looking at Table S3 and comparing that to the results presented in Table 2, there seems to be some discrepancies. For example, C0HAL2 (elongation factor 1-alpha) is reported as having a very significant fold-change between the FM (D1) and SBM (D2) in Table 2, yet no quantitative values for this protein are reported in the D2 group in Table S3. Why is this? Please ensure that your filtering is working as intended. Additionally, proteins that were not identified in a replicate should not be assigned a quantitative value of 0. Instead, these values should be treated as missing (null), or missing value imputation should be used to provide baseline quantitative values. It’s difficult to determine how such values were treated in the analysis, and that should be stated outright. As an aside for future experiments, use of the ‘match between runs’ feature in MaxQuant should dramatically decrease the number of missing values present in the dataset. Minor Comments: Line 241: It seems as if part of this sentence was accidentally removed as there is no mention of the addition of trypsin. Line 243: Another sentence or two very briefly describing the column and MS parameters that were used would be useful here. Line 261: Was a fragment mass tolerance of 0.8 used in the reprocessed data as well or was this a carryover from the previous version of the manuscript? As mentioned previously, a much tighter tolerance (\~0.02 Da) would be more appropriate. Lines 276-277: Saying that the average weights increased does not mean that every fish gained weight. Consider rephrasing. Table 2: It would be nice if you included the p-value for each protein here as well. Figure 6: As a note, the ‘presence’ of a protein in one sample and not in another is much more likely a result of the stochastic nature of the MS data collection rather than a protein actually being present in the sample or not. Therefore, comparisons such as the one shown here are often not very informative. \*\*\*\*\*\*\*\*\*\* 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 \#3: No \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] 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.0218360.r006 Author response to Decision Letter 2 28 Nov 2019 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 \#3: General Impressions: The authors have clearly addressed the concerns raised in the previous review, and the reanalysis of the proteomics data has dramatically improved the quality of the manuscript. I greatly appreciate the authors hard work and dedication to providing the highest quality data possible. I have noticed one discrepancy in the data analysis that should be addressed in the final version of the manuscript (described below), but I feel that the manuscript is otherwise ready for publication. I’ve also included a few very minor comments that I feel will improve the readability of the publication. Major Comments: The authors state that only protein that were identified in at least two of the four replicates were used for quantitative analysis. However, when looking at Table S3 and comparing that to the results presented in Table 2, there seems to be some discrepancies. For example, C0HAL2 (elongation factor 1-alpha) is reported as having a very significant fold-change between the FM (D1) and SBM (D2) in Table 2, yet no quantitative values for this protein are reported in the D2 group in Table S3. Why is this? Please ensure that your filtering is working as intended. --\> We follow the criteria “at least two peptides in at least half of the replicates” per group. It means to asseverate that a protein is present in one or all the groups, should be at least in two samples. As an example, protein C0HAL2 is detected in all replicated of diet 1, in one of the replicates on diet 6 and in three out of the four replicates on diet 7. It was not detected on diet 2, therefore the significant difference between diet 1 and diet 2. From the total 158 proteins, 126 proteins were present in all dietary treatment, however the other 32 proteins were present in either one, two or three of the treatment, not in all, as stated in the manuscript. Additionally, proteins that were not identified in a replicate should not be assigned a quantitative value of 0. Instead, these values should be treated as missing (null), or missing value imputation should be used to provide baseline quantitative values. It’s difficult to determine how such values were treated in the analysis, and that should be stated outright. --\> The function missing value imputation was used for the quantitative analysis. The imputed value is shown in Table S3 and it is mentioned in the manuscript as follow: “Protein raw data were transferred to log normalization; missing value imputation was used to replace the not identified proteins on the quantitative analysis and then performed on autoscaled data (mean-centered and divided by the standard deviation of each variable) \[29\].” As an aside for future experiments, use of the ‘match between runs’ feature in MaxQuant should dramatically decrease the number of missing values present in the dataset. --\> We are very grateful for the suggestions, we have been learning a lot in this process so we are confident that in future experiments we will take into practice the knowledge acquired. Minor Comments: Line 241: It seems as if part of this sentence was accidentally removed as there is no mention of the addition of trypsin. --\> Yes, it was an editing mistake. The sentence has been rewritten as follow: “Subsequently, the proteins were digested with 10 μg trypsin (Promega, sequencing grade) overnight at 37°C. The digestion was stopped by adding 5 μL 50% formic acid and the generated peptides were purified using a ZipTip C18 (Millipore, Billerica, MA, USA) according to the manufacturer’s instructions, and dried using a Speed Vac concentrator (Concentrator Plus, Eppendorf, Hamburg, Germany)” Line 243: Another sentence or two very briefly describing the column and MS parameters that were used would be useful here. --\> The information required was added into the text “The tryptic peptides were dissolved in 10 µL 0.1% formic acid/2% acetonitrile and 5 µL analyzed using an Ultimate 3000 RSLCnano-UHPLC system connected to a Q Exactive mass spectrometer (Thermo Fisher Scientific, Bremen, Germany) equipped with a nanoelectrospray ion source. For liquid chromatography separation, an Acclaim PepMap 100 column (C18, 2 µm beads, 100 Å, 75 μm inner diameter, 50 cm length) (Dionex, Sunnyvale CA, USA) was used. The mass spectrometer was operated in the data-dependent mode to automatically switch between MS and MS/MS acquisition. Survey full scan MS spectra (from m/z 400 to 2,000) were acquired with the resolution R = 70,000 at m/z 200, after accumulation to a target of 1e5. The maximum allowed ion accumulation times were 60 ms” Line 261: Was a fragment mass tolerance of 0.8 used in the reprocessed data as well or was this a carryover from the previous version of the manuscript? As mentioned previously, a much tighter tolerance (\~0.02 Da) would be more appropriate. --\> This corresponded to the previous analysis, the new analysis was performed using: “The fragment mass tolerance for the MS1 was 6 ppm and the fragment mass tolerance for the MS2 was 20 ppm”. This information was added into the text Lines 276-277: Saying that the average weights increased does not mean that every fish gained weight. Consider rephrasing. --\> The sentence was rephrased as follow: “The average initial weight was 526 g and the average final 277 weight was 667 g on day 37, considering that the weight was measured as bulk, this indicates that in general fish gained weight during the experimental period” Table 2: It would be nice if you included the p-value for each protein here as well. --\> We have included the p-value in Table 2 as required by Reviewer 3. Figure 6: As a note, the ‘presence’ of a protein in one sample and not in another is much more likely a result of the stochastic nature of the MS data collection rather than a protein actually being present in the sample or not. Therefore, comparisons such as the one shown here are often not very informative. --\> We are grateful for the comments and it will be considered in our future experiments. 10.1371/journal.pone.0218360.r007 Decision Letter 3 Angers Annie Academic Editor 2019 Annie Angers 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. 9 Dec 2019 Candida utilis yeast as a functional protein source for Atlantic salmon (Salmo salar L.): Local intestinal tissue and plasma proteome responses PONE-D-19-15096R3 Dear Dr. Øverland, 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, Annie Angers, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Thank you for the extensive revisions to the manuscript. I am sorry that the processed took so long, but I feel that the improved version was worth the efforts. Reviewers' comments: 10.1371/journal.pone.0218360.r008 Acceptance letter Angers Annie Academic Editor 2019 Annie Angers 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. 12 Dec 2019 PONE-D-19-15096R3 *Candida utilis* yeast as a functional protein source for Atlantic salmon (*Salmo salar* L.): Local intestinal tissue and plasma proteome responses Dear Dr. Øverland: 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. Annie Angers Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist. [^2]: Current address: Cargill Aqua Nutrition North Sea, Bergen, Norway

# Introduction Bullying is a relevant problem widespread in schools. Bullying involves a power imbalance between a target and his or her perpetrator(s) and tends to be repeated over time. Reports indicate that the bullying perpetration rates range between 6.4% and 11%, and victimization rates range between 21% and 48.5%. The consequences for victims can be dramatic because peer victimization is associated with numerous mental health problems and a considerable decline in quality of life. In recent years, bullying has spread to social networks. In this context, cyberbullying has been described as willful and repeated harm inflicted \[onto a victim\] through the use of computers, cell phones, and other electronic devices. Cyberbullying rates are also high, oscillating between 17% and 24% for perpetration and between 24% and 26.6% for victimization. The perpetration and victimization of peer aggressions tend to overlap. Thus, the percentages of pure victims and pure perpetrators were found to be relatively lower than those known as bully-victims and cyberbully-victims. An attempt was made to explain the existence of this double role through mechanisms of reaction to aggression. Those children and adolescents who are targets of aggression by other youths might seek revenge and act aggressively against their assailants. Similarly, when a child or adolescent bullies others, she/he can—as a result—become the target of bullying and cyberbullying behaviors. In support of these reciprocal mechanisms, some studies have found longitudinal evidence of reciprocity between victimization and perpetration; however, other studies have found only evidence for victimization as a predictor of perpetration. For example, in a longitudinal study, traditional bullying victimization predicted traditional bullying perpetration and cybervictimization predicted cyberbullying perpetration six months later, but perpetration did not predict the increase in victimization. Similarly, in a longitudinal study of traditional bullying, victimization increased the likelihood of involvement in bullying perpetration. In another study, cybervictimization predicted cyber perpetration in girls but not in boys, and perpetration did not predict victimization in girls or boys. In a longitudinal study of three waves, cyber victimization predicted an increase in cyber perpetration through the mediation of cyber witnessing. Other studies have found evidence for the opposite relationship. For example, in a large sample in South Korea, adolescents who bullied others were highly likely to be bullied by others in the following year. Finally, another study did not find bidirectional longitudinal associations over time between cyber victimization and perpetration. The severe consequences of bullying and cyberbullying have encouraged the development of numerous preventive intervention programs in the school setting (for a review see). The results of these interventions have been mixed, especially regarding cyberbullying. For example, in a review by Evans and collaborators, they found that only 11 of the 22 reviewed studies (50%) displayed significant effects. Furthermore, several review studies have found that the effectiveness of interventions may be moderated by the degree of development of children and adolescents (for example,). Although the findings are contradictory, some reviews concluded that the effects of interventions decrease in older youth. For instance, through a hierarchical meta-analysis, found that traditional antibullying interventions were effective from early childhood to early adolescence and that there was a decline to a null effect in the case of interventions with middle adolescents (Grade 8 and higher). The same finding has been observed in investigations of cyberbullying prevention interventions. For instance, in an examination of the effects of the KiVa antibullying program on cyberbullying and cybervictimization, the effect of KiVa on cyberbullying was moderated by age; thus, the reduced odds of endorsing higher frequencies of cyberbullying in the posttest in the experimental condition were observed only when students’ age was below the sample mean. A suggested explanation for the aforementioned findings is that whereas younger youth may be willing to accept the authority of adults and follow curriculum activities, older youth may be especially motivated to be respected and maintain their autonomy with regard to adults. For this reason, older adolescents may be especially sensitive to attempts by adults to modify their behavior through interventions and react in an opposite manner than expected. Notably, some interventions with individuals in middle and late adolescence have produced an opposite result of the intended effect. A new approach to interventions, termed “wise interventions,” is attracting interest (for a review see). Wise interventions emphasize the subjective creation of meanings, how people interpret themselves and social situations, and, in doing so, can effectively change behavior in a recursive manner over time. A proliferation of wise interventions has addressed several social and personal problems. Many of these interventions have shocking results because they tend to be very brief in duration and produce lasting changes in people's behavior. Wise interventions do not address a lack of capacity or risk behaviors directly and incorporate three basic motives that drive the search for meaning: (a) precision (people want their interpretations to be accurate); (b) self- integrity (people want to think about themselves as adequate, moral, and competent); (c) belonging (people want to feel accepted and included by others, belong to social groups, and contribute positively to the lives of others). Wise interventions allow people to adopt new belief or behavior themselves rather than follow a directive from someone else (for example,). In this manner, the design of wise interventions allows them to be perceived as respectful regarding the autonomy and status of students so that students realize that they make their own decisions. The results obtained with this type of intervention in adolescent behaviors are enormously promising. For example, Yeager and his collaborators designed a short, universal wise intervention designed to change the implicit beliefs about the personality of adolescents (Incremental Theory of Personality Intervention; ITPI). Specifically, the intervention focused on teaching that personality can change. The key elements of the ITPI were as follows: (a) students have an active role to facilitate the deeper processing of the message; (b) the intervention is not presented to the students to change their behavior so that the students do not feel manipulated, and this reduces their resistance to the intervention; and (c) the intervention has long-term effects due to the recursive processes that influence the effects that accumulate over time. The ITPI has been shown to reduce the symptoms of anxiety and depression among adolescents (for example,) and bullying behavior. The results obtained show that their effectiveness is moderated by the level of development of adolescents. For example, in a sample of adolescents, the effect of the intervention on the reduction of cyberbullying perpetration was greater when the level of testosterone, an indicator of pubertal maturation, was lower. Researchers have not examined whether bullying and cyberbullying prevention interventions affect the aforementioned reciprocity between victimization and perpetration. As a result of the intervention, adolescents could learn to not react aggressively when they are the target of rejection and aggression by peers, breaking the vicious circle of violence. This learning could occur because of interventions based on incremental theories of personality. The ITPI teaches that when young people behave aggressively or reject peers they often do so because they have personal problems or do not know how to behave differently and that aggressions do not last forever because people can change. Therefore, this type of intervention could affect adolescents’ reaction when they are victims of aggression, reducing the likelihood of reacting aggressively. Notably, Yeager and colleagues found that following an ambiguous provocation scenario, the ITPI reduced hostile attributions, vengeful desires, and aggressive retaliation. ## This study Two studies have found that an ITPI reduced aggressive behavior in adolescents. However, additional knowledge is required to understand the mechanisms through which ITPI acts. Although findings have been inconsistent, research has suggested high reciprocity between victimization and perpetration. Therefore, in this study, we proposed that the ITPI intervention could reduce the predictive reciprocal associations between victimization and perpetration. This mechanism could partially explain the effectiveness of the intervention. Although in this study we examined victimization and perpetration as outcomes, we expected that the effects of the intervention would be stronger for perpetration because it refers to the behavior of the adolescent receiving the intervention, whereas victimization depends mainly on others’ behavior. Similarly, based on the same reason, we expected that the intervention would reduce the predictive path from victimization to perpetration rather than the path from perpetration to victimization. Because reviews have indicated that the level of development of the participants can moderate the effects of interventions, we included age as a moderator in the analyses. We also controlled for the overall level of perpetration and victimization in the classrooms because the classroom context can influence the behavior of adolescents. Finally, sex was added to the analyses because researchers have found sex differences in the prevalence rates of bullying and cyberbullying victimization and perpetration (for example,). # Materials and methods ## Participants This study is part of a project conducted in Bizkaia (Basque Country, Spain). Twenty high schools were contacted and 10 agreed to participate. Eligible schools had general student populations, and the schools’ principals agreed to the randomization and delivery of the assigned intervention. Classrooms were selected in each school in order to obtain participants from different grades. The total sample was 1329 adolescents from 33 classrooms. A subsample of 535 adolescents participated in a previous study to assess the effectiveness of the ITPI in bullying and cyberbullying behavior. The total sample was used to test the hypotheses of the current study. Of the total sample, 462 participants were excluded (exclusion criteria are indicated); thus, the final analytic sample was 858 high school students (52% boys) aged 12 to 17 at pretest (*M* = 14.56, *SD* = 0.97). Their parents’ socioeconomic status was determined by following the recommendations of the Working Group of the Spanish Society of Epidemiology and the Spanish Society of Family and Community Medicine, who advise researchers to consider parents’ most recent job. This criterion led to the following distribution of socioeconomic class: 15.8% low, 20.8% low-medium, 20.2% medium, 23.8% high-medium, and 19.4% high. The distribution by grade was 280 in 8^th grade (*M*_age = 13.56, *SD* = 0.52), 363 in 9^th grade (*M*_age = 14.60, *SD* = 0.47), and 215 in 10^th grade (*M*_age = 15.80, *SD* = 0.47). A double-blind randomized controlled trial was performed with two parallel groups. Randomization was performed at the individual level within each classroom on the day of the intervention, and after the pretest, 452 participants were assigned to the experimental condition and 406 to the control condition. Randomization was performed by block and divided by sex, and no differences were observed between the control and the experimental group in the pretest in sex (p =.416) or age (p =.725). The notebooks with the task for each intervention were inside an envelope to ensure blinding of the assignment and condition. The participants completed the intervention individually, and the researchers collected the notebooks inside the envelopes. displays the participants’ flow chart. No statistically significant differences were found in pretest measures of bullying perpetration, cyberbullying perpetration, and cyberbullying victimization among adolescents who completed all the study measures and those who did not. However, adolescents who did not complete the three study measures were mostly boys (27.8% of boys and 19.7% of girls), χ² (1, *N* = 858) = 7.81, *p* \<.01; of older ages (14.79 for noncompleters versus mean age of 14.49 for completers), *t*(856) = 3.86, *p* \<.001; and with higher mean scores on bullying victimization (*M* = 3.30 for noncompleters versus 2.61 for completers), *t*(846) = 2.12, *p* \<.05. ## Procedure Students completed the intervention and assessment in their classrooms during school hours. The session was administered by a trained psychology research assistant who led students to work quietly and privately. Informed consent was required both from parents and adolescents. An overview of the procedure is presented in. One week before the interventions began, a survey was administered during school hours (pretest), and a similar posttest was administered six months and 12 months afterward (clinical trials ID: NCT03481699). As the majority of the adolescent population in the Basque Country is bilingual, the questionnaires were presented in both Spanish and Basque so that each adolescent could choose the language in which to respond. The Ethics Committee of University of Deusto approved this study. ## Intervention In this study we delivered an experimental intervention modeled on the intervention developed by Yeager and collaborators. The experimental and control interventions were adapted to local cultural specifications and translated into Spanish and Basque. The experimental intervention was divided in three parts and its total duration was 50–60 minutes. In the first part, participants read a scientific article that provides information about individuals´ potential to change. They read conclusions of neurological and behavioral studies that demonstrate that the thoughts and feelings in the brain control the behavior and that under correct circumstances the pathways in the brain can change. Next, participants wrote three sentences based on the science explaining that people can change. In the second part, participants read extracts written by other students that had participated in the study and written their own conclusions. The purpose of using of the testimonials obtained from other participants was to add credibility to the incremental theory of personality. In the final part, participants described a time when they felt isolated, rejected, or disappointed by another person at school. Next, they imagined that the same event happened to another student and write one to three paragraphs describing what they could do or say to help the other student to understand that people can change and the things happening to him/her might also change. The control intervention was developed to be performed in parallel with the experimental intervention and also has three main parts (see Supplementary Information, Annex 1 by Yeager and collaborators). However, the control intervention contained scientific information and information on the human brain. In the first part, the participants read a scientific article about the brain, including brain localization and the role of different brain areas in supporting cognitive functioning. In the second part, the participants read stories written by other adolescents explaining how they became accustomed to the sensory and physical environment (e.g., the building, sounds, smells) of high school. In the last part, participants wrote about how and why students adapt to the physical environment at high school. ## Instruments The Revised Peer Experiences Questionnaire (RPEQ;) was used to assess participants´ bullying perpetration and victimization experiences. The bullying perpetration and victimization scales comprised nine parallel items that asked participants to rate how often they have perpetrated or they have received an aggressive behavior in the last six months on a 5-point scale ranging from 1 (*never*) to 5 (*a few times a week*). Sample items are “I threatened to hurt or beat a peer up” (perpetration scale) and “A kid gossiped about me so that others would not like me” (victimization scale). Researchers that have examined the RPEQ found adequate psychometric properties in studies involving adolescents. Test–retest reliability and internal consistency were adequate. The version of the RPEQ used in this study had been translated into Spanish and Basque by means of back-translation procedures and validated in other studies with good psychometric properties. Cronbach coefficients were.81.87, and.90 at T1, T2, and T3 for perpetration, respectively; and.82.87, and.87 at T1, T2, and T3 for victimization, respectively. Cyberbullying victimization and perpetration were assessed using the Cyberbullying Questionnaire (CBQ) to measure perpetration and victimization in the past six months. Each scale includes nine items, for example, “sending pictures of an acquaintance that could be humiliating” (perpetration) and “receiving threatening or insulting messages from other people” (victimization). The following response scale was used: 0 (*never*), 1 (*1 or 2 times*), 2 (*3 or 4 times*), and 3 (*5 or more times*). The CBQ has shown adequate factorial and convergent validity, in addition to an acceptable internal consistency. In this study, Cronbach coefficients were.79.89, and.91 at T1, T2, and T3 for perpetration, respectively; and.74.84, and.89 at T1, T2, and T3 for victimization, respectively. ## Overview of the statistical approach The pattern of missingness was examined. In the general sample, for self- reported measures, Little’s MCAR test was statistically significant, χ²(235) = 420, *p* \<.001. Thus, we used full information maximum likelihood (FIML), a recommended method to manage missing values when they are not distributed randomly. FIML estimates parameters by using all the available data, including cases without data. We used hierarchical linear modeling 7–3 with robust standard errors. We estimated separate models for bullying and cyberbullying victimization and perpetration. Each model was accompanied by Level 1, 2, and 3 equations. For instance, for cyberbullying perpetration, for Level 1, regression equations modeled variation in the repeated measures as a function of time (i.e., the three waves of data) and cyberbullying victimization. Because cyberbullying was measured at each time point, it was modeled as a time-varying covariate to investigate whether cyberbullying perpetration was predicted by changes in cyberbullying victimization over time. Time was coded as 0, 1, or 2. For Level 2, the equations modeled individual differences in the Level 1 parameters (i.e., intercepts and slopes) as a function of between-subject variables. Level 2 predictors of the intercept included condition (0 = control, 1 = experimental) and age. Level 2 predictors of the slope included the same predictors and interaction terms between the condition and the other variables. The inclusion of these parameters at Level 2 allowed us to test the effects of condition and age on both the intercept and the change in the outcome variables over time. Sex (1 = female, 0 = male) was also included as a predictor of the intercept to control for sex differences in cyberbullying perpetration. We included random effects for intercept and time at Level 2, allowing for variability between individuals in the initial levels and changes over time. Where the random effects were not significant, they were removed from the final models. We tested models that included sex and the Sex x Time, Sex x Condition, and Sex x Condition x Time interaction terms. Sex was a significant predictor of the intercept. However, because sex did not moderate the effect of the intervention on victimization nor on perpetration, we report the most parsimonious models without these components. Finally, at Level 3, we included the classrooms and their average level of cyberbullying perpetration as a predictor of the intercept and the time slope. Thus, the level of cyberbullying in each classroom focused on the grand mean. Random effects for intercept and slope were included, allowing variation between classrooms. A similar model was used to predict bullying perpetration. In this case, bullying victimization was used as a predictor, instead of cyberbullying victimization, and the average level of bullying perpetration in the classroom was used at Level 3. We also tested parallel models in which the outcome was cyberbullying versus bullying victimization and predictors were cyberbullying versus bullying perpetration. For these models, the average level of victimization in the classroom was used at Level 3. The effect sizes were calculated to compare the differences between groups in changes in the outcome variables from baseline to follow-up using the estimated marginal means obtained in the mixed models. Where the differences between groups in changes were statistically significant, positive values of Cohen’s *d* indicated greater decreases in outcomes (e.g., greater reduction in bullying perpetration) for adolescents in the ITPI group compared with the adolescents in the control condition. Negative values of Cohen’s *d* indicate greater decreases in the control group. Data is available at <https://osf.io/mh92w/> (DOI [10.17605/OSF.IO/MH92W](https://doi.org/10.17605/OSF.IO/MH92W)). # Results ## Descriptive statistics and preliminary analyses shows the correlation coefficients for the study variables. All the variables were significantly associated (*p* \<.001). The highest correlations were between perpetration and victimization at each time for traditional bullying and cyberbullying. depicts the descriptive statistics and comparisons between the intervention and control conditions. Several *t* tests indicated no significant differences between the control and ITPI groups in any study variable at baseline (*p* \>.05). There were not significant differences in any variable depending on socioeconomic level, except for cyberbullying perpetration at the one-year follow-up, in which adolescents of medium socio-economic class scored lower than adolescents of low-medium (*p* =.007) and high-medium (*p* =.005) socioeconomic classes. ## Effects of the ITPI presents the results of the mixed models for bullying and cyberbullying perpetration as outcomes, and presents the random effects. The average level of bullying/cyberbullying perpetration in the classroom was significantly associated with the intercept; thus, the classroom context is a predictor of the initial individual level. Sex was significantly associated with initial level of perpetration with girls scoring lower than boys in both bullying and cyberbullying perpetration. There were no differences in initial levels of perpetration between the experimental and control groups. Although the slope for time was not significant, it indicated an increasing tendency over time for both bullying and cyberbullying perpetration. For bullying but not for cyberbullying, the level of perpetration in the classroom predicted a lower increasing tendency over time, probably because classrooms with high levels of aggressive behavior have less space to increase aggressive behavior. The negative significant time x age interaction indicates that bullying and cyberbullying increase less as age increases. Moreover, as hypothesized, the time x condition x age interaction was statistically significant, indicating that the effect of the intervention on perpetration levels is moderated by age. displays the trajectories of both bullying and cyberbullying perpetration for adolescents within the experimental and control conditions of participants under and above the median age (Median = 14.48). For cyberbullying, among younger participants, those in the experimental group display a slightly decreasing trend, whereas those in the control group display an increasing tendency. Among older participants, the trajectories are very similar in the experimental and control groups. For bullying, a similar pattern is observed among the younger participants. In the case of older participants, the experimental group displays an increasing tendency from the pretest to the six-month follow-up and a slight tendency to decrease from the six-month to the one-year follow-up. The control group, by contrast, displays a stably decreasing trend from the pretest to the one-year follow-up. presents the means for bullying and cyberbullying perpetration over time estimated in the mixed models. The effect sizes comparing mean change scores from baseline to 6-month follow-up, and from baseline to the 12-month follow-up, were small. The results also indicate that victimization was a strong predictor of perpetration for bullying and cyberbullying. Because victimization was included as a Level 1 variable, these results indicate that perpetration increases when victimization increases. More important, the intervention reduced the predictive association between victimization and perpetration. displays this association for adolescents in the experimental and control groups. As observed, the association is more intense in the control group than in the experimental group. Age did not moderate the role of victimization on the prediction of perpetration or the effect of the intervention on the association between victimization and perpetration. presents the results for bullying and cyberbullying victimization as outcomes. Average level of victimization in the classroom was a strong predictor of initial level of individual victimization for bullying and cyberbullying. The changes in both bullying and cyberbullying perpetration were strong predictors of both bullying and cyberbullying victimization over time, respectively. However, no effect of age and condition were observed for victimization. # Discussion Peer aggression is a highly prevalent problem among adolescents. In addition, victimization and perpetration tend to overlap; that is, many adolescents react with violence when subjects of aggression because they are motivated by the desire to take revenge. The objective of this study was to examine whether an intervention aimed at teaching an incremental theory of personality, that is, that people can change, reduced the reciprocity between victimization and perpetration of bullying and cyberbullying. We expected that the effects of the intervention would be stronger for perpetration than victimization because perpetration involves the behavior of the adolescent receiving the intervention, whereas victimization depends mainly on others’ behavior. In consistency with our hypothesis, the ITPI had no effect on the trajectory of victimization over time or on the predictive relationship between perpetration and victimization, neither for bullying nor for cyberbullying. This result is not surprising because the victimization does not depend mainly on the behavior of the recipients of the intervention but on the behavior of other people, including people outside of the study. The results suggest that the ITPI acts on the behaviors of the adolescents who receive them. We also predicted that victimization would act as an antecedent of perpetration and that the intervention would reduce the predictive path from victimization to perpetration. As we expected, and consistent with other studies, victimization was a strong predictor of perpetration over time for bullying and cyberbullying. By contrast, perpetration was not a predictor of victimization. Consistent in both forms of aggressive behavior, the intervention reduced the intensity of the association between victimization and perpetration. This effect was not moderated by the age or sex of the participants. Namely, the findings indicate that the probability of perpetrating bullying/cyberbullying as a consequence of victimization was lower among adolescents who received the ITPI than in adolescents in the control group. This finding is critical because it suggests that the ITPI can contribute to the subjective creation of meanings such that when adolescents experience rejection and aggression from others, they could interpret these experiences in a more benign manner, which inhibit retaliatory reactions. The ITPI teaches through stories of other young people that often adolescents who behave rudely to others or commit acts of aggression do so because they have personal problems or do not know how to behave differently. The ITPI also teaches that people can change. Therefore, when adolescents who have received the ITPI are the target of aggressions, they could interpret that the other person’s actions may be caused some difficulty and that the attacks will likely not last forever; thus, the desire for revenge and the probability of reacting aggressively would be reduced. This is consistent with the results obtained by Yeager and collaborators, that is, after the ITPI the adolescents reacted to provocation with a more positive attitude and less desire for revenge. Although the main objective of the study was not to test the effects of the ITPI on the trajectory of bullying and cyberbullying, the results show that its effectiveness is moderated by age. Specifically, in terms of cyberbullying, among the youngest adolescents (\<14.48), those who received the ITPI showed a slight tendency to reduce aggressive behavior that contrasts with the increasing trend in the control group. Among the oldest ones, the trajectories were similar in the two groups. This result largely coincides with that obtained with the KiVa program by Williford and collaborators. In the case of bullying, the result among the youngest is very similar to that obtained for cyberbullying. However, among the oldest participants, there is a paradoxical effect of a greater reduction in the perpetration of bullying in the control group than in the experimental group. This result is reminiscent of the findings obtained with other interventions in which results opposite to those expected have been observed (for a review see). Traditional bullying behavior takes place face to face, and it is possible that among older adolescents, who are more motivated to maintain their social status, even a subtle intervention, such as that used in this study, may provoke opposite reactions. Finally, male adolescents scored higher than female adolescents on perpetration of bullying and cyberbullying and on cyber victimization. However, sex did not moderate the effects of the ITPI in the outcome trajectories or in the predictive association between victimization and perpetration. Thus, we dropped sex from the predictive models to increase their parsimony, and sex was maintained only as a predictor of the baseline level. This study has limitations that provide opportunities for further research. One limitation is the exclusive use of self-reports, which can contribute to increasing the association between victimization and perpetration. A second limitation is the loss of participants over time, especially in the one-year follow-up. Some adolescents had changed schools and many others could not be present in the classroom on the days when the data was collected. The oldest boys with higher scores on bullying victimization were the participants with the highest likelihood of not completing all the study measures; thus, the results may not be generalizable to this particular group of participants. Finally, the study is based exclusively on quantitative data; thus, conducting qualitative studies to examine how adolescents create meanings after the intervention would be valuable. Thus, for example, a topic for further research could be an evaluation of how adolescents interpret victimization behaviors and if they do so in a more benign manner than the adolescents in the control group. Despite these limitations, the study has strengths that allow for valuable conclusions. An RCT design was used with a large sample of adolescents, The study included a follow-up of six months and one year. In addition, contributions were made to the field of preventive bullying and cyberbullying intervention by examining how an intervention alters the reactions of adolescents to victimization. ## Conclusions Findings of this study indicate that a brief intervention teaching that people can change reduces the reciprocity between victimization and perpetration. This mechanism could explain the beneficial effects of the ITPI and of other interventions. The intervention is within the so-called Wise Interventions, and its brevity makes it a low-cost tool easily implemented in educational contexts. The results also indicate that early implementation of the intervention is critical because its effectiveness was greater among the youngest participants. This research was supported by a grant from the Spanish Government, (Ref. PSI2015-68426-R) and from the Basque Country (Ref. IT982–16 and Ref. PI_2016_1\_0023). [^1]: The authors have declared that no competing interests exist.

# Introduction Depression is one of the most frequent psychiatric disorders in patients with type 2 diabetes, with a lifetime prevalence of 24%–29% and a point prevalence of 10%-15%. Depression in type 2 diabetes is significantly associated with poor glycaemic control, chronic complications, increased mortality, reduced physical and mental functioning, higher health costs and decreased adherence to diet and hypoglycaemic medications. A recent epidemiological study demonstrated that individuals with diabetes and comorbid depression have a worse health-related quality of life and higher health service usage. Collaborative care models for depression in diabetes demonstrated to significantly improve outcomes for depression and glycaemic control. Depression may be underdiagnosed in diabetes. The limitations in diagnosing depression in diabetes include the lack of specific diagnostic criteria and the overlap of symptoms. Egede and Ellis noted that poor management of diabetes is a barrier for early recognition of depressive symptoms, such as fatigue, changes in weight and appetite, sleep disturbances, and motor retardation. Furthermore, increased appetite is a cardinal diagnostic criterion for both major depression and diabetes, anhedonia may result from changes in lifestyle due to diabetes, loss of interest may be secondary to limitations to pursue usual free-time activities due to diabetes complications (e.g., visual problems, neuropathy), and loss of libido and reduced self-esteem may result from sexual dysfunction. Campayo and co-workers suggested that depression in diabetic patients may be qualitatively different from depression observed among individuals without diabetes who have ‘primary’ (i.e. no organic cause) depression, but no supporting data were provided. A recent large prospective study showed that individuals with current depressive and/or anxiety disorders fad a 10-fold increased odds of diabetes incidence at two years, and the incidence was highest for those with comorbid depression and anxiety. A valid and timely diagnosis of depression is of clinical relevance especially given that detection and treatment of depression improves glycaemic control. Another potential confounder for a valid diagnosis of depression in diabetes is the potential comorbidity with anxiety. This is substantial in the general population, with 57% of individuals with major depression also meeting diagnostic criteria for an anxiety disorder. Generalized anxiety disorder (GAD) is more frequent in diabetes than in age-comparable healthy controls, and is associated with poor health behaviours. In the general population, comorbid depression and anxiety is highly prevalent. Several studies examined the comorbidity between depression and anxiety in diabetes, but they were limited by important methodological shortcomings, such as a diagnosis of diabetes based only on self-report. The aim of the present study was to examine the syndromal pattern of depression and anxiety in type 2 diabetes using a data-driven approach, as well as to examine the overlap with anxiety. We used baseline data from community-based participants of the Fremantle Diabetes Study Phase II (FDS2), and used latent class analysis (LCA) to examine the homogeneity of depressive symptoms and the comorbidity between depression and anxiety. LCA assesses the symptom profile of individual patients and produces classes of patients in terms of their pattern of symptoms. Given that the classes thus identified do not overlap, individual patients can belong to only one group. Based on our experience with LCA in medical conditions, we expected to find three classes characterized by high, moderate and low probabilities for symptoms of depression, as well as a high comorbidity between depression and anxiety. # Methods ## Patients The FDS2 is a longitudinal observational study of known diabetes conducted in a postcode-defined geographical area surrounding the city of Fremantle (Western Australia). Details of FDS2 recruitment procedures, sample characteristics including classification of diabetes type and non-recruited patients have been published elsewhere. Briefly, between 2008 and 2011, 4,667 diabetic patients were identified in the local population of 153,000 (crude diabetes prevalence 3%), of whom 1,668 (36%) were recruited to FDS2, in addition to 64 surviving FDS phase I participants who had moved out of the study area (1,732 participants altogether). ## Ethics Statement The Human Research Ethics Committee of the Southern Metropolitan Area Health Service specifically approved the study, and all subjects gave written informed consent. ## Clinical and psychiatric assessment Each participant underwent a baseline assessment that included a comprehensive clinical questionnaire, physical examination and fasting biochemical tests. Diabetes type was assessed from treatment history, BMI, age at diagnosis, and nature of first presentation and/or self-identification. Non-insulin treated patients and those ≥60 years of age at diagnosis were usually considered to have type 2 diabetes, as were patients \<60 years of age at diagnosis and taking insulin at the time of the study entry but whose first treatment was not insulin. In these cases, case records were consulted for evidence of ketonemia, as well as islet cell antibody, GAD antibodies, serum insulin and C-peptide levels, if available. Complications were ascertained using standard criteria. All participants were assessed with the PHQ-9, a brief and accurate assessment of depression that has been validated in patients with diabetes. The PHQ-9 is a dual-purpose instrument that can establish depressive disorder diagnoses as well as ratings of depressive symptom severity. Patients with type 2 diabetes have significant physical complaints that may artificially increase PHQ-9 scores on items that are not necessary symptoms of major depression. Using a PHQ-9 cut point rather than the diagnostic algorithm may increase the frequency of false- positive cases of depression. Therefore, as suggested by the PHQ-9 authors, we used the instrument's diagnostic algorithm for making diagnoses of major and minor depression. Major depression was diagnosed when the first or second symptoms (i.e. depressed mood or loss of interest/anhedonia) were present more than half the days, and at least 5 of the 9 symptoms were present during the same period. Minor depression was diagnosed whenever 2 to 4 symptoms (including depressed mood or loss of interest/anhedonia) were present for more than half the days. Lifetime history of depression was assessed with the Brief Lifetime Depression Scale (BLDS). This instrument was developed and validated by our group for use in diabetes. The BLDS was modelled on the PHQ-9 and included similar items, general format and language, but the participants were asked whether they had ever had a period during their lives lasting for two weeks or more when they experienced any of the nine symptoms of depression listed in the DSM-IV/5. For the study of anxiety in the FDS2, a specific questionnaire (the Generalized Anxiety Disorder Scale (GADS)) was devised by one of the authors (SES) with experience in developing psychiatric instruments to rate the severity of anxiety. The GADS includes 9 items that correspond to all 9 DSM-IV criteria for GAD. Based on the PHQ-9, GADS items were rated as “not at all present”, “present several days”, “present more than half of the days”, and “present nearly every day”. As with the PHQ-9, a symptom was rated positive whenever the patient rated “more than half of the days” or “nearly every day”. To make the instrument comparable to the DSM-IV criteria for GAD, the time criterion was the past 6 months. The validity of the GADS was examined by having a psychiatrist (SES) assess a convenience sample of 23 patients with the Structured Clinical Interview for the DSM-IV-Research Version (SCID)-Anxiety Disorders Module, blind to the GADS results. Kappa statistics showed a high diagnostic concordance for DSM-IV GAD diagnosis (kappa  = 0.88). Six of the 23 patients received a diagnosis of GAD based on the SCID, as compared to 4 based on the GADS (McNemar's test, *P* = 0.50). Test-retest reliability was assessed in 25 patients, who completed the GADS twice, with an interval of (mean ± SD) 9±6 days. The intra-class correlation (ICC) was high (ICC  = 0.88). The presence of cognitive deficits was based on the following algorithm (Bruce et al., 2008): patients scoring ≤26 on the MMSE were further assessed with the Clinical Dementia Rating (CDR) (Hughes et al., 1982) to establish the presence of dementia. This instrument generates ratings ranging from normal cognition (CDR 0) to severity levels of dementia (CDR 1–3) and an intermediate category defines very mild cognitive impairment or questionable dementia (CDR 0.5). ## Statistical Analysis To test for group differences, generalised linear modelling (GLM) was used. LCA (cluster procedure) was used to determine the latent structure of patients with diabetes and to categorize individuals based on their responses to the PHQ-9 and GADS. LCA was calculated using the software LatentGold version 4.0.4. LCA assumes that a ‘class’ explains the association among a set of symptoms. LCA computes latent class probabilities or prevalence, and conditional probabilities. The basic assumption of LCA is that a (presumably homogeneous) population of individuals is a mixture of distinct, but internally homogeneous subgroups. LCA assesses the symptom profile of individual patients and produces classes of patients as suggested by their pattern of symptoms. Since the clusters do not overlap, individual patients can belong to only one group. For this analysis, we included the nine DSM-IV symptoms for major depression and the four DSM-IV symptoms for generalized anxiety disorder that do not overlap with major depression as latent class indicators. Models with one to five classes were estimated. Each patient was assigned to the latent class to which the largest posterior probability was calculated. The best-fitting model was chosen based on having the lowest values for the following goodness of fit indices: Bayes Information Criterion (BIC), Approximate Weight of Evidence (AWE) (which takes into account classification performance), and the Corrected Akaike Information Criterion (CAIC). The level of contribution of each indicator to the final latent class structure is assessed by the information content statistic (*R²*), which can be interpreted similarly to the commonalities in traditional factor analysis. For the clinical interpretation, the mean level and frequency of responses for each psychiatric symptom within each class were examined. After determining the best class solution, we used latent class membership as a between-subject variable in a one-way analysis of variance (ANOVA) followed by post-hoc comparisons using Tukey's HSD, and *X²* analyses to examine demographic and clinical differences among the latent classes. The computer package IBM SPSS Statistics 20 (IBM Corporation, Somers, NY, USA) was used to compare baseline characteristics across the four LCA classes. Data are presented as proportions, mean ± SD, geometric mean (SD range), or, in the case of variables which did not conform to a normal or log- normal distribution, median \[inter-quartile range, IQR\]. For independent samples, two-way comparisons for proportions were by Fisher's exact test, for normally distributed variables by Student's *t*-test, and for non-normally distributed variables by Mann-Whitney U-test. Multiple comparisons for proportions were by Fisher's exact test or Chi-squared test, for normally distributed variables by one-way ANOVA, and for non-normally distributed variables by Kruskal-Wallis H-test, adjusted using the Bonferroni correction. A two-tailed significance level of *P*\<0.05 was used throughout. # Results ## Demographic and clinical findings FDS2 recruited 1,551 participants (90%) with clinically defined type 2 diabetes. Complete data for our study was obtained from 1,337 patients (86%). When compared with the 214 patients with missing data, the latter were more likely to be female (46% vs. 57%, *P* = 0.005), and had worse glycaemic control (HbA_1c (median \[IQR\]) 6.8% median \[IQR\] \[6.2%–7.6%\] vs. 7.0% \[6.4%–8.5%\], respectively, *P* = 0.001). On the other hand, there were no significant between-group differences in age (mean years ± SD) (65.8±11.1 vs. 64.9±14.4 years, respectively, *P* = 0.41), diabetes duration (8.8 years \[2.8–15.6\] vs. 10.0 years \[3.0–17.8\], respectively, *P* = 0.12), or cognition (8.0% vs. 7.1% cognitively impaired (P = 1.00). ## Latent class analysis Items selected for the LCA included all 9 items from the PHQ-9, and the items of worrying/anxiety, feeling restless, feeling tense, and feeling irritable from the GADS (the remaining items were not included due to overlap with PHQ-9 items). The modelling started with probing the presence of conventional latent classes (i.e., cluster modelling), running cluster models of up to 5 latent clusters. Inspection of conditional probabilities showed a reduction in BIC and CAIC values from the 2- to the 5-cluster model, but at the cost of increasing number of parameters (NPar). From the 4-class model onwards, decreased accuracy was unacceptably high (i.e., \>10%) and the AWE index has started to ascend. The 4-cluster model showed a pattern of probabilities extending from one extreme (low probabilities for all symptoms) to the other (high probabilities for all symptoms), with the exception of two intermediate classes showing a mixed pattern of probabilities. This finding suggests that a single discrete factor (anxious depression) with 4 ordered levels or classes (e.g., no anxious depression, subclinical anxiety, minor anxious depression and major anxious depression) provides the best explanatory model. Thus, an inter-factorial effect and orderliness of latent classes was obtained, as opposed to random and/or pseudonominal composition of latent classes. Class-specific endorsement profiles for depression and anxiety symptoms are shown in. Latent class 4 included 104 patients (8%) and showed high partial conditional probabilities (\>0.60) for all the anxiety and depression symptoms, except for psychomotor changes (\>0.40) and suicidal ideation (\>0.30). This class was labelled “major anxious depression”. Latent class 1 included 439 patients (33%) and showed low partial conditional probabilities (\<0.10) for all symptoms of depression and anxiety. This class was labelled “no anxious depression”. Latent class 3 included 293 patients (22%) with partial conditional probabilities intermediate between classes 1 and 4, and was labelled “minor anxious depression”. Finally, latent class 2 included 501 patients (38%) and showed higher conditional probabilities than class 4 on the anxiety items, sleep changes and loss of energy. This class was labelled “subclinical anxiety”. We used final class membership as a between-subject variable to evaluate differences in depressive and anxiety symptoms as a function of latent class membership. The ANOVAs evaluating mean symptom severity differences on each item as a function of latent class membership were all significant at the *P*\<0.001 level. Follow-up pairwise comparisons using Tukey's HSD tests demonstrated the following differences: For GAD worry the ANOVA was significant, and all four groups were significantly different to each other (Tukey HSD *P* = 0.0001). Similar results were obtained for GADS restlessness, GADS tension, GADS irritability, loss of interest, sad mood, sleep changes, loss of energy, appetite changes, worthlessness/guilt, and poor concentration. For the items of psychomotor changes and suicide ideation there were significant differences on post-hoc tests between all classes except between classes 1 and 2. There were significant demographic and clinical differences between classes. The *major anxious depression class* was significantly younger than the other classes, whilst the *minor anxious depression class* was significantly younger than the subclinical anxiety class. Depression (major or minor) was significantly more frequent in the *major anxious depression class* as compared to the other classes, and in the *minor anxious depression class* as compared to the remaining two classes. Depression before the onset of diabetes was significantly higher in the *major anxious depression class* as compared to the other classes, and in the *minor anxious depression class* as compared to the remaining two classes. Patients in the *major anxious depression class* had a significantly higher frequency of GAD as compared to the other classes, whilst patients in the *minor anxious depression class* had a significantly higher frequency of GAD than the remaining two classes. Depression (minor or major) and/or GAD were present in 89% of patients in the *major anxious depression class*, 29% in the *minor anxious depression class*, 2% in the *subclinical anxiety class*, and none in the *no anxious depression class*. Finally, the *major anxious depression class* showed a significantly younger age at onset of diabetes, higher HbA_1c levels, higher serum triglycerides levels and a higher frequency of insulin use than the other three classes. This class also showed a higher BMI than the *subclinical anxiety* and the *no anxious depression classes*. When the class with major anxious-depression was compared with patients meeting DSM-IV criteria for major depression, there were no differences on demographic, diabetes-related or psychiatric variables. # Discussion To our knowledge, this is the first study to examine the validity of diagnostic criteria for depression in diabetes using LCA in a large community sample of patients with type 2 diabetes, and there were several important findings. First, LCA of depressive and anxiety symptoms from our sample produced evidence for a 4-class solution: a *major depression-anxiety class*, a *minor depression- anxiety class*, a *subsyndromal anxiety class*, and a class with *no depression or anxiety*. Second, all nine DSM-IV diagnostic criteria for major depression identified a class of patients with a high frequency of major depression, and further changes to these criteria are not necessary. Third, all symptoms of anxiety had similar high probabilities as symptoms of depression for the *major depression-anxiety class*, suggesting that anxiety symptoms should be added to the DSM-IV/5 diagnostic criteria for major depression to identify a class with more severe psychopathology. Therefore, our findings do not support separating anxiety from depressive symptoms for patients with type 2 diabetes. Fourth, the *major depression-anxiety class* had worse diabetes-related parameters than the class without anxiety or depression, suggesting that anxious depression is a marker of more malignant diabetes. Finally, we also identified a class that included anxiety and depression symptoms of a lesser severity, which we termed *minor depression-anxiety*. About one third of this group had clinically relevant anxiety and/or depression, but there were no demographic or clinical differences when compared to the class without anxiety or depression. These findings should be considered in the light of some limitations of our study. First, given the large sample of individuals to be assessed and the comprehensive quantity of clinical assessments, we were unable to interview patients with more appropriate structured psychiatric interviews. Nevertheless, we assessed depression using the PHQ-9, a valid and reliable instrument to diagnose depression based on DSM-IV criteria and to measure the severity of depression. Second, anxiety symptoms were assessed with the GADS, a new instrument specifically developed and validated for the present study. As with the PHQ-9, the GADS allows a diagnosis of GAD based on DSM-IV criteria and measures the severity of anxiety. We did not use the Generalized Anxiety Disorder 7-item scale (GAD-7) given that this instrument was not available when our study was designed. However, we believe that the GADS is an improvement over the GAD-7 given that the GADS assesses all the DSM-IV/5 criteria for GAD over the past 6 months, whilst the GAD-7 assesses a partial list of symptoms during the past 2 weeks. This was a cross-sectional study, and the clinical progression of the classes identified in this study will need to be examined in our ongoing biennial follow-up study. Finally, we did not assess for the category of ‘diabetes distress’ in our study. The syndrome of major depression was criticized for not differentiating an expected emotional reaction to a significant stressor from a pathological reaction, whereas ‘diabetes distress’ may better capture the worries and fears of a chronic illness. However, the construct of ‘diabetes distress’ is in need of further validation, and it is not included into the major psychiatric nomenclatures. Our study provides empirical support for a class that includes the symptoms of major depression in diabetes, and demonstrates this class to be significantly related to worse diabetes outcomes. While depression is recognized as one of the most frequent psychiatric disorders in type 2 diabetes, questions remain regarding the validity of using DSM-IV/5 diagnostic criteria when the symptoms of depression and diabetes can overlap. Given the impact of depression on the management and prognosis of diabetes, it is critical to validate diagnostic criteria for depression for use in clinical practice. Using LCA, we found for the first time that all nine- DSM-IV criteria for major depression identify a class of patients with a high chance of having major depression. Moreover, this group also showed significantly worse clinical parameters, such as higher fasting plasma glucose, higher HbA1c, greater use of insulin, higher BMI, increased central adiposity and increased serum triglycerides than the class without anxious depression, which provides partial validation of the classes obtained. The classes with high probability of depressive symptoms also included a high probability of anxiety symptoms (worrying, restlessness, tension and irritability) and had a high frequency of GAD. This finding demonstrates the high comorbidity of depression and anxiety in diabetes, and suggests that a new diagnostic category of “major anxious depression” should be used in type 2 diabetes. In the general population, this comorbidity is associated with a poor prognosis, recurrent depression, worse treatment outcomes of depression, more frequent suicide attempts, greater symptom severity, diminished social supports, and more past-year distressing life events. Starr and Davila proposed a model of depression-anxiety comorbidity in which anxiety symptoms lead to depressive symptoms via maladaptive anxiety response styles. Another relevant finding from our study was a class with intermediate symptoms of anxiety and depression, which we labelled *minor depression-anxiety*. This class included a relatively large proportion of our sample (22%), and future studies may determine whether this class may evolve into a more severe depression with significant impact upon diabetes variables. Cyranowsky et al recently demonstrated that anxiety commonly emerges alongside subthreshold depressive symptoms. The finding that 29% of this class had depression and/or anxiety suggests that this class is clinically relevant. It has been demonstrated that subclinical depression has a significant negative impact on diabetes control and the relationship between depressive symptoms and poor self- care appears to be linear. A third class, which included 38% of the sample, had minor symptoms of anxiety but no depression. Only 2% of this group had clinical depression and/or anxiety, but had a significantly higher lifetime prevalence of GAD and depression as compared to the class with no anxiety or depression. The clinical relevance of this class should be further examined in longitudinal studies. In conclusion, we demonstrated that the DSM-IV criteria for major depression may be used unmodified to diagnose depression in diabetes. We also found a high comorbidity between depression and anxiety, suggesting that symptoms of anxiety may be included into a new set of diagnostic criteria for major depression or mixed anxiety-depression in type 2 diabetes. Future longitudinal studies will validate this construct and examine the clinical relevance of minor anxious depression and subclinical anxiety in type 2 diabetes. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: SES WAD MD. Performed the experiments: SES WAD VC. Analyzed the data: MD WAD. Contributed reagents/materials/analysis tools: WAD MD. Wrote the paper: SES DB TMED.

# Introduction Sepsis is a lethal syndrome of physiologic, pathologic, and biochemical abnormalities induced by infection, which is one of the major global public health concerns. Although recently there has been extensive researches demonstrating the mechanism and treatment of sepsis, sepsis is still the principle cause of death in intensive care patients worldwide. There are a variety of scores for the diagnosis of sepsis that can be evaluated, but there is still a lack of valuable indicators for the study on prognostic factors. Anemia is one of the risk factors for death resulting from sepsis and septic shock, and hematocrit (HCT) is the percentage of red blood cells in the volume of the whole blood, which is one of the critical biomarkers for the diagnosis of anemia. HCT can be utilized as a critical prognostic biomarker in several cancers, including lung cancer, renal carcinoma, and epithelial ovarian cancer. However, little is known about the prognostic value of HCT for patients with sepsis in surgical intensive care units (ICUs). At present, clinical researches have mainly focused on the relationship of anemia indicators such as the red blood cell distribution width (RDW) and platelets with the prognosis of sepsis. In spite of limited researches on the impacts of HCT on the prognosis of patients with sepsis, a few studies are limited to the evaluation of anemia before sepsis surgery. Accordingly, the HCT level of first admission patients was employed in our research to dissect out whether it influenced the prognosis of sepsis, thus helping doctors assess the condition in a timely manner and provide a basis for subsequent prognostic measures. # Method ## Data source All data used in the study was extracted from the MIMIC-IV (v1.0) database, which is the hospitalization information of patients admitted to the Higher Medical Center in Boston, Massachusetts, the USA from 2008 to 2019 to conduct a retrospective study. This database is a relational database containing the actual hospitalizations of hospitalized patients in an advanced academic medical center in Boston, Massachusetts, USA. The MIMIC-IV database is based on the success of MIMIC-III. It integrates improvements to the deficiencies of the MIMIC-III database, including laboratory measurements, medications, recorded vital signs, SOFA score, SAPS II score, etc. ## Study population and data extraction PgAdmin4 has used run structure query language (SQL) and extracted data from the MIMIC IV database. Patients are identified by ICD-9 codes and extracted from the database. The extracted variables included age, gender, admission type, comorbidities, laboratory parameters, severity score, time of entering and leaving the hospital, time of entering and leaving the ICU, and date of death. Laboratory parameters include hematocrit, albumin, lactate, bilirubin, hemoglobin, red blood cell distribution width, platelet, serum creatinine, potassium, etc. All laboratory data were extracted from the data generated within the first 24 hours after the patient entered the ICU (i.e., the baseline value). Data in this retrospective analysis is accurate medical data, accessed free of charge. The project has been approved by the Institutional Review Board of the Massachusetts Institute of Technology and BIDMC and has received an informed consent exemption. The author has the right to use and download the database through the Protecting Human Research Participants exam. ## Patient population Adult patients (age ≥18 years) admitted to ICU (internal medicine ICU, surgical ICU) who had the exact hematocrit index and corresponding ID within 24 hours after admission to the ICU. Patients who lack corresponding data will be excluded. Sepsis criteria: This study uses patients diagnosed with sepsis in the database and meets sepsis-3 to define the sepsis cohort. ## Statistical analysis Analyses were performed using Stata 16. Baseline characteristics of all patients were stratified according to the quartiles of hematocrit (HCT) and divided into three groups: low hct level (male≤42%, Female≤37%), regular hct level (male 42%\~49%, female 37%\~44%), and high hct level (male≥49%, female≥44%). Continuous variables were presented as median, and categorical variables were presented as a number. Nonparametric Wilcoxon tests or Kruskal-Wallis tests were used to compare continuous variables between different groups. Categorical variables between the groups were compared using the X2 test and Fisher exact test. A P value \< 0.05 between the two groups was considered a significant difference. A Multivariate Cox proportional hazard model was constructed to determine the independent effects of three groups on 30-day mortality. Variables with P \<0.05 in univariate analysis were further included in the multivariate Cox proportional hazard model. We compared the survival rates using log-rank tests and presented the results as Kaplan-Meier curves. A Kaplan-Meier survival curve was constructed to compare the 30-day mortality of the three groups of hct levels. P\<0.05 indicates that the difference is statistically significant. # Result ## Population and baseline characteristics A total of 2057 patients with sepsis 3.0 in the MIMICIV database were included which have complete data of hematocrit level within 24 hours after ICU admission. All baseline characteristics are summarized in. Differences in age, laboratory parameters, comorbidities, and scores between the two groups were statistically significant. HCT level was divided into three groups. Variables with missing data are relatively common in the MIMIC IV database. The number of deaths within 30 days after admission to the ICU was 911. Compared with the 30-day survival group, the death group was older, and the levels of hematocrit, albumin, platelets, and hemoglobin were lower than those in the survival group (P\<0.05). # Association of Hematocrit level with 30-day outcomes The 30-day all-cause mortality rate of patients was 44.2%, As shown in, compared with the survival group, the proportion of patients with low hct levels in the 30-day mortality group was significantly increased, and the proportion of the regular group and the high hct level group gradually decreased. The difference was statistically significant (P\<0.05). ## Hematocrit is an independent prognostic predictor in sepsis patients Survival analysis was conducted to explore the impact of hct level on 30d mortality. Notably, from the previous analysis, we know that patients in the lower hct level had worse survival rates. Basic demographics and laboratory parameters for the prediction of 30-day mortality were investigated using a univariate Cox analysis regression model. Variables include age, low hct levels, albumin, platelets, bilirubin, red blood cell distribution width, lactate, Scr, SBP, DBP MAP, SAPS II score, hypertension, heart failure, and mechanical ventilation are all statistically significant (p\<0.05). Adjust the univariate analysis for the potential confounding factors associated with 30-day mortality in patients with sepsis, And then, the 30-day mortality was assessed with a multivariable Cox proportional regression model. According to the results, the low hct level remained an independent prognostic factor for sepsis (P\<0,05). Compared with the regular hct level, the 30-day mortality risk of the low hct level is increased, and the difference is statistically significant \[HR = 1.589 .95%CI1.099–2.297, P\<0.05\], although the risk of death in the high hct level also increased, the difference was not statistically significant (P = 0.055). ## Kaplan-Meier survival curve analysis The Kaplan-Meier survival curve was drawn according to the category of hct to show the 30-day survival rate of patients with sepsis. The results showed that the difference between HCT level and the 30-day mortality rate of sepsis was statistically significant (P = 0.040), red blood cell ratio Content is related to the prognosis of patients with sepsis. # Discussion Recently, sepsis remains the main cause of death triggered by the infection in ICUs despite of considerable breakthroughs in the exploration of sepsis, including broad-spectrum antibiotics, supportive treatment, and even precision medicine and monitoring. Infection and immune response disorders have been recognized as risk factors for organ dysfunctions. The poor functional status has been reported as a risk factor for sepsis and an expected consequence based on the long-term investigation of sepsis over the past decade. Therefore, it is essential to be able to early predict and evaluate the prognostic factors for sepsis. Infection is one of the crucial reasons for the occurrence and development of sepsis. This is because it can touch the systemic inflammatory response and cause the formation of inflammasomes in the body, thereby inducing the release of numerous pro-inflammatory factors. Therefore, it is imperative to use effective biomarkers to monitor the prognosis of sepsis. Whole blood parameters have been partially studied as biomarkers in the diagnosis, treatment, and prognosis of sepsis. Whole blood viscosity, red blood cell aggregation, and red blood cell deformability may be risk factors for sepsis and septic shock mortality. In addition, the study conducted by Gong Yan et al. has elucidated that elevated RDW can remarkably predict disease progression and poor clinical outcomes in patients with sepsis. Currently, several emerging studies have been conducted to ascertain the influence of HCT on the prognosis of sepsis. It has been elaborated that HCT is correlated with all-cause mortality in end-stage renal disease, heart failure, coronary heart disease, cancer, and inflammatory states. However, HCT has not been determined as an independent influencing factor for clinical outcomes of diseases. In this study, 2057 patients with sepsis were evaluated to figure out the correlation between the HCT level and the prognosis of sepsis. The results illustrated that compared with septic patients with the regular HCT level, the 30-day mortality rate was enhanced in septic patients with a low HCT level measured within 24 hours after entering the ICU (male ≤ 42% and female ≤ 37%), accompanied by conspicuously augmented Simplified Acute Physiology Score II scores. Following gradual regression and correction of various potential confounding factors, we obtained reliable results from the study, suggesting that HCT might be an independent risk factor for the prognosis of patients with sepsis, which is the same as the previous group of sepsis anemia patients and Western countries. The results of a small-scale clinical trial are concordant. HCT is a whole blood parameter that reflects the ratio between red blood cells and plasmas. It is closely linked to the prognosis of critically ill patients and indicators for the fluid resuscitation treatment. However, the relationship between HCT and the prognosis of patients with sepsis remains poorly understood. Acute systemic infections lead to inflammatory responses that result in sepsis, which dramatically diminishes the number of red blood cells entering the blood circulation. The production of reactive oxygen species may contribute to the repression of the ability of red blood cells to transport oxygen and the deformities of red blood cell membranes. During the entire process of inflammatory response and oxidative stress, the number of red blood cells is diminished and the blood dilution induced by liquid expansion leads to a reduction in the HCT level. It is necessary to further prove these research hypotheses. It has been documented that inflammatory factors like tumor necrosis factor α, interleukin (IL)-1, IL-2, IL-6, and IL-8 trigger the adhesion of neutrophils and endothelial cells, leading to the formation of microthrombi. This mechanism may be related to the rapid removal of red blood cells from the circulating blood caused by inflammation and oxidative stress in sepsis, which is manifested by anemia and accelerated red blood cell apoptosis. In addition, red blood cell infusion has been highlighted to improve oxygen transport and metabolism, thereby alleviating microcirculation disorders in patients with sepsis and septic shock. Therefore, the hypotheses about the HCT were proposed on the basis of the aforementioned research. As reflected by the results of this study, patients with the low HCT level exhibit low platelets, hemoglobin, and the high 30-day mortality rate. The mechanism mentioned above can provide a basis for this theory. Paolo Boffetta et al. unraveled that HCT is also associated with the mortality of ordinary people. A retrospective study by Zhang Xin et al. unveiled that the low HCT also correlated to the poor prognosis of patients with lung cancer and ovarian cancer. These findings suggest that HCT may reflect the severity of a wide range of diseases. This study was a large-scale retrospective study, in which plenty of factors related to the death of sepsis were harvested and adjusted to evaluate their impacts on the prognosis. The results manifested that HCT could be an independent risk factor for the prognosis of patients with sepsis. Although HCT is not highly specific according to our results, it is a simple laboratory parameter that is easier to obtain. The HCT levels obtained in our research were all information of patients with sepsis on the first day after entering the ICU, minimizing the changes in the HCT levels caused by the progression of the disease and the treatment process. However, there also exist several limitations in this study: (1) The Medical Information Market Intensive Care Database is a sizeable single-center database that lacks diversification. Therefore, our results are influenced by unity and may have inevitable bias. (2) Due to the singularity of the database, we can only conduct observational researches on HCT and mortality, so further studies are needed to understand the underlying pathophysiological mechanism in the future. (3) Taking the problem of the content of the data sample into account, the sample is not classified by gender, but a general study is carried out. In summary, HCT is associated with the prognosis of septic patients during admission to the ICU. Septic patients with the low HCT level presented with enhanced disease severity and high mortality rate. HCT can be applied as an independent risk factor for the prognosis of patients with sepsis. Further study is required to investigate the pathophysiological and immunological mechanism of the relationship between HCT and clinical outcomes in septic patients. A perfect biomarker for sepsis has not yet been identified. More immunological experiments and multi-center studies on this easily available parameter will be of great significance for the early prediction of the outcomes of sepsis, which must be beneficial for the treatment of sepsis. # Conclusion In summary, the hematocrit in the first 24 hours after ICU admission was independently associated with increased 30-day all-cause mortality in adult septic patients but of limited sensibility and specificity. Further extensive multicenter prospective studies are needed to confirm the relationship and validate whose clinical significance. 10.1371/journal.pone.0265758.r001 Decision Letter 0 Cortegiani Andrea Academic Editor 2022 Andrea Cortegiani 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 Jan 2022 PONE-D-21-35992Association between hematocrit and the 30-day mortality of patients with sepsis: a retrospective analysis based on the large-scale clinical database MIMIC-IVPLOS ONE Dear Dr. luo, 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.  Two Reviewers experts in the field and on big data management reviewed the paper. I agree on their comments, especially about the methods used, which I ask you to fully solve Please submit your revised manuscript by Feb 28 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, Andrea Cortegiani, M.D. 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  <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.  Please include your tables as part of your main manuscript and remove the individual files. Please note that supplementary tables (should remain/ be uploaded) as separate "supporting information" files. 3\. Thank you for stating the following financial disclosure:   (No) At this time, please address the following queries: a\) 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.  b\) 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.” c\) If any authors received a salary from any of your funders, please state which authors and which funders. d\) 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. 4\. Thank you for stating the following in your Competing Interests section:    (No) Please complete your Competing Interests on the online submission form to state any Competing Interests. If you have no competing interests, please state "The authors have declared that no competing interests exist.", as detailed online in our guide for authors at [http://journals.plos.org/plosone/s/submit- now ](http://journals.plos.org/plosone/s/submit-now )  This information should be included in your cover letter; we will change the online submission form on your behalf. 5.  We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. 6\. 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> 7\. Please include your tables as part of your main manuscript and remove the individual files. Please note that supplementary tables (should remain/ be uploaded) as separate "supporting information" files. 8\. Thank you for stating the following in the Funding Section of your manuscript:  (This study was supported by (the Chengdu Science and Technology  Bureau (No.2017FZ0058)) and (the Health Department of Sichuan Province  (No.17PJ474))) We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.  Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:  (No) Please include your amended statements within your cover letter; we will change the online submission form on your behalf. \[Note: HTML markup is below. Please do not edit.\] 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: 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: 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: Thank you for the opportunity to review this study. The authors have conducted a retrospective observational study looking for an association between hematocrit and 30 day mortality in a septic population. The method proposed is a multivariate model of a single center retrospective dataset (MIMIC IV). They found a correlation between low HCT value at admission and increased 30d mortality. The study finding are not very original. I have the following concerns and comments: 1\) Introduction paragraph (line 81-93) seems repeated 2\) reporting checklist is missing, I suggest to use the RECORD guidelines (The REporting of studies Conducted using Observational Routinely-collected health Data) Please refer to <http://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1001885> 3\) Please indicate in detail how do you selected the sepsis population in the MIMIC database. Did you used the "concepts" script for mimic-iv in order to improve the reproducibility of the analysis? (Johnson, A. E., Stone, D. J., Celi, L. A., & Pollard, T. J. (2018). The MIMIC Code Repository: enabling reproducibility in critical care research. Journal of the American Medical Informatics Association : JAMIA, 25(1), 32–39. or an extraction based on ICD-9 codes (Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of care. Angus DC, Linde-Zwirble WT, Lidicker J, Clermont G, Carcillo J, Pinsky MRCrit Care Med. 2001 Jul; 29(7):1303-10.) and Martin et al. (The epidemiology of sepsis in the United States from 1979 through 2000. Martin GS, Mannino DM, Eaton S, Moss M N Engl J Med. 2003 Apr 17; 348(16):1546-54) 4\) Line 119-120 report the use of non parametric test, please specify the distribution of data Reviewer \#2: Comments for authors: Dear authors, You have presented a large-scale retrospective analysis on the association between hematocrit and mortality in patients with sepsis. The topic is interesting and it would be important to find a real contribution in this setting. Please give your statement to the following points: 1\. Abstract The abstract is clear enough. 2\. Introduction \- The introduction was repeated twice; delete the second repetition. 3\. Materials and Methods \- It has not been included if a sample size estimation has been carried out; has it been calculated? \- It is not clear enough to me whether the selected patients were already diagnosed with sepsis upon entering the ICU. 4.Results \- In the figure 1, please add legend. 5\. Discussion \- It is written that "The HCT levels obtained in our research were all information of patients with sepsis on the first day after entering the ICU, which suppressed the progression of the disease and changes in laboratory parameter values induced by fluid resuscitation therapy"; but the patients who had already been diagnosed with sepsis upon admission to the ICU could not have already started fluid resuscitation therapy? Or could not they have been more fragile patients already in fluid overload? \- Please specify better the clinical message that the authors want to send. 6\. Conclusion \- In my opinion the words "hematocrit during ICU" is misleading; better to specify “hematocrit in the first 24 hours after ICU admission”. 7\. Tables \- Table 2 appears to be repeated; please check the data. \- Tables 1 and 2 lack legends; please enter. 8\. References Please check the journal’s guidelines It would be necessary to revisit the English language. Best regards \*\*\*\*\*\*\*\*\*\* 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: **Alberto Noto 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.0265758.r002 Author response to Decision Letter 0 10 Feb 2022 Dear editors and reviewers: Thanks for your letter and the reviewers' comments about our manuscript entitled “Association between hematocrit and the 30-day mortality of patients with sepsis: a retrospective analysis based on the large-scale clinical database MIMIC-IV”.All these suggestions are valuable for me to revise my manuscript. We have studied comments carefully and have made corrections which we hope meet with approval. Suggestions from editors: The topic is interesting and it would be important to find a real contribution in this setting. Please give your statement to the following points: 1\. Abstract The abstract is clear enough. 2\. Introduction \- The introduction was repeated twice; delete the second repetition. 3\. Materials and Methods \- It has not been included if a sample size estimation has been carried out; has it been calculated? \- It is not clear enough to me whether the selected patients were already diagnosed with sepsis upon entering the ICU. 4.Results \- In the figure 1, please add legend. 5\. Discussion \- It is written that "The HCT levels obtained in our research were all information of patients with sepsis on the first day after entering the ICU, which suppressed the progression of the disease and changes in laboratory parameter values induced by fluid resuscitation therapy"; but the patients who had already been diagnosed with sepsis upon admission to the ICU could not have already started fluid resuscitation therapy? Or could not they have been more fragile patients already in fluid overload? \- Please specify better the clinical message that the authors want to send. 6\. Conclusion \- In my opinion the words "hematocrit during ICU" is misleading; better to specify “hematocrit in the first 24 hours after ICU admission”. 7\. Tables \- Table 2 appears to be repeated; please check the data. \- Tables 1 and 2 lack legends; please enter. 8\. References Please check the journal’s guidelines It would be necessary to revisit the English language. Dear editors,the revised portion are marked in red in our paper. Responses to Reviwers: Reviewer \#1 1. -Introduction paragraph (line 81-93) seems repeated We are very sorry for our neligence of this paragraph.And then the repetitive sentence in the introductory paragraph has been deleted. 2. -reporting checklist is missing, I suggest to use the RECORD guidelines (The REporting of studies Conducted using Observational Routinely-collected health Data) Please refer to <http://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.100188-> Considering the reviwer’s suggestion,we have made some modifications based on the guidelines. 3.-Please indicate in detail how do you selected the sepsis population in the MIMIC database. Did you used the "concepts" script for mimic-iv in order to improve the reproducibility of the analysis? (Johnson, A. E., Stone, D. J., Celi, L. A., & Pollard, T. J. (2018). The MIMIC Code Repository: enabling reproducibility in critical care research. Journal of the American Medical Informatics Association : JAMIA, 25(1), 32–39. or an extraction based on ICD-9 codes (Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of care. Angus DC, Linde-Zwirble WT, Lidicker J, Clermont G, Carcillo J, Pinsky MRCrit Care Med. 2001 Jul; 29(7):1303-10.) and Martin et al. (The epidemiology of sepsis in the United States from 1979 through 2000. Martin GS, Mannino DM, Eaton S, Moss M N Engl J Med. 2003 Apr 17; 348(16):1546-54) Firstly,we download the zip package of the database from the official website for a series of installation, and then enter the interface of PostgreSQL 6.0. We use SELECT\*FROM statement to find the icd-9 code of sepsis in d_icd_diagnoses, and then use SELECT WHERE statement to find the related icd-9 code of sepsis in diagnoses_icd (99591/99592)We used the "concepts" script for mimic-iv in order to improve the reproducibility of the analysis. 4.-Line 119-120 report the use of non parametric test, please specify the distribution of data Considering the reviewer’s suggestion the data distribution has been shown in Table 1. Reviewer \#2 1.Introduction \- The introduction was repeated twice; delete the second repetition. Thank you for your careful review.The repetitive sentence in the introductory paragraph has been deleted. 2.Materials and Methods \- It has not been included if a sample size estimation has been carried out; has it been calculated? \- It is not clear enough to me whether the selected patients were already diagnosed with sepsis upon entering the ICU. This sample size has been calculated.The data of sepsis patients we present in the MIMIC IV database by SQL statements are diagnosed on admission to the ICU.This can be determined by referring to the r elevant details of this database on the official website.(<https://physionet.org/content/mimiciv/0.4/>). 3.Results \- In the figure 1, please add legend. Thank you for your careful review.We have added the legend.(Details can be found in the revised version) 4.Discussion \- It is written that "The HCT levels obtained in our research were all information of patients with sepsis on the first day after entering the ICU, which suppressed the progression of the disease and changes in laboratory parameter values induced by fluid resuscitation therapy"; but the patients who had already been diagnosed with sepsis upon admission to the ICU could not have already started fluid resuscitation therapy? Or could not they have been more fragile patients already in fluid overload? Thank you for the reviewer's suggestion, We have reworked it and marked in red. 5.Conclusion \- In my opinion the words "hematocrit during ICU" is misleading; better to specify “hematocrit in the first 24 hours after ICU admission”. Thank you for the reviewer's suggestion, I have adopted your revision and made the statement changes. 6.Tables \- Table 2 appears to be repeated; please check the data. \- Tables 1 and 2 lack legends; please enter. The table has been made in detail。（Details can be found in the revised version） 7.References Please check the journal’s guidelines It would be necessary to revisit the English language. Changes have been made to the reference specification. Thank you for your careful review. We really appreciate your efforts in reviewing our manuscript during this unprecedented and challenging time. We wish good health to you, your family, and community. Your careful review has helped to make our study clearer and more comprehensive. 10.1371/journal.pone.0265758.r003 Decision Letter 1 Cortegiani Andrea Academic Editor 2022 Andrea Cortegiani 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. 8 Mar 2022 Association between hematocrit and the 30-day mortality of patients with sepsis: a retrospective analysis based on the large-scale clinical database MIMIC-IV PONE-D-21-35992R1 Dear Dr. luo, 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, Andrea Cortegiani, M.D. 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: (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 \#1: **Yes: **Alberto Noto 10.1371/journal.pone.0265758.r004 Acceptance letter Cortegiani Andrea Academic Editor 2022 Andrea Cortegiani 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 Mar 2022 PONE-D-21-35992R1 Association between hematocrit and the 30-day mortality of patients with sepsis: a retrospective analysis based on the large-scale clinical database MIMIC-IV Dear Dr. luo: 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. Andrea Cortegiani Academic Editor PLOS ONE [^1]: No

# Introduction Around the globe, an increase of forced migration since 2015 has raised the urgent question of how migration-receiving countries respond best to the associated challenges. The successful integration of the newcomers into the receiving countries’ labor market is among the most pressing related issues as many refugees struggle with finding regular employment. Besides other factors, such as the host countries’ labor market policies or asylum seekers’ professional qualifications, the integration of asylum seekers into the labor market also crucially depends on the willingness of potential coworkers to collaborate with asylum seekers at the workplace (see also). However, as one German survey indicates, there are widespread concerns regarding the potential consequences related to the immigration of asylum seekers, such as increased competition in the job market. These findings add fuel to the debate that increasing diversity may deteriorate intergroup relations and social cohesion more generally, and illustrate the need to investigate predictors of employees’ willingness to collaborate with asylum seekers at work. In this study, we propose that employees’ team properties–specifically, the team’s diversity—affects team members’ willingness to collaborate with asylum seekers. This is because the team constitutes employees’ immediate social environment at the work place with profound effects on team members’ perceptions and motivations. By investigating the effects of work team diversity on employees’ willingness to collaborate with asylum seekers, we attempt to contribute to several streams of literature. Firstly, we connect research on work team diversity, which has primarily investigated the diversity-performance link (see, for a review), with research on the societal consequences of diversity such as the perception of social groups. Secondly, we contribute to research on intergroup contact by proposing that experiences in diverse work teams generalize to being more open towards other diversity types as well. Thirdly, most studies conducted in organisations have focused on the effects of objective diversity (e.g., the divergence on demographic characteristics such as age, ethnicity, or gender, cf.). Studies that have been conducted in other contexts shows, however, that the *perception* of diversity may also play a prominent predictor of intergroup attitudes, and thus intentions towards groups. Consequently, we contribute by exploring both the effects of facets of *objective* and *perceived* team diversity on the willingness to collaborate with asylum seekers. Fourthly, researchers have suggested boundary conditions governing the effects of team diversity. Adding to this research, we also propose and test potential moderators of the influence of diversity on the willingness to cooperate with asylum seekers. The proposed moderators are based on Allport’s original formulation of the intergroup contact hypothesis and were adapted to the work context: cooperation within the contact situation (task interdependence of team members), support by local norms (pro-diversity team norms), and status (anticipated position of the asylum seeker within the work team). ## Intergroup contact effects on attitudinal outcomes The intergroup contact hypothesis proposes that positive interactions between members of different social groups ameliorate intergroup relations. Allport formulated four conditions under which intergroup contact unfolds its beneficial effects on attitudes: equal status, common goals, intergroup cooperation, and support of authorities, norms, and customs. A large body of empirical evidence supports these postulations, while the conditions for positive effects of intergroup contact are rather conducive than indispensable. Many studies have subsequently applied the intergroup contact hypothesis to the contextual level. According to these authors, individuals within contextual units with larger shares of outgroup members are provided with more opportunities for intergroup contact experiences. Consequently, these individuals should have more favorable attitudes towards the respective outgroup than individuals in units with lower shares of outgroup members. Indeed, contextual outgroup exposure has been linked to opportunities for intergroup mixing, which can lead to positive intergroup encounters and more favorable attitudes. For instance, the ethnic composition in social units has been linked to higher levels of acceptance of immigrants or ethnic minorities (\[, –\] but see, e.g.,). Thus far, research in this domain has been rarely linked to contact in professional settings (for an exception, see). However, researchers have argued that smaller social units, such as one’s immediate work context, lend more meaningfulness to individuals than larger, commonly-studied units, such as one’s neighborhood or county. This is because most people negotiate their everyday relationships in small-scaled contexts. Therefore, we aim to add to this literature by examining diversity effects in the work context. In doing so, we focus on work teams, i.e., groups of people perceiving themselves as such due to their work on a shared goal, since teamwork is the most common form of organizational collaboration. Another branch of intergroup contact research suggests that the beneficial effects of intergroup contact are not limited to the groups involved in the encounter. That is, the effect of intergroup contact on attitudes is not restricted to the respective outgroup present in the contact situation, but also affects attitude levels towards groups not involved; a process that has been coined the secondary transfer effect. Empirical studies support this postulation. Work in this realm suggests a generalization gradient, in a way that the transfer effect is stronger for outgroups that are similar or overlapping to the ones involved in the intergroup encounter. To our knowledge, none of the existent studies has tested secondary transfer effects on attitudes towards potential colleagues in work settings. Intergroup contact requires some degree of diversity among those involved. Diversity can be defined as differences between individuals on any attribute that may lead to the perception that another person is different from oneself, such as nationality, migration background, age, or gender \[cf. 14–16\]. Following the intergroup contact hypothesis and Blau’s proposition that diversity serves as a facilitator of integrative group processes, we argue that diversity in teams is equivalent to regular collaborative intergroup contact in the work setting. Consequently, higher levels of diversity in work teams should be linked to more intergroup contact, unlike in other settings, where cross- group contact can also be avoided. Hence, higher levels of diversity should lead to more positive attitudes towards the involved groups. Through the secondary transfer effects, this process should expand to a reduction in prejudice towards asylum seekers, expressed in an increased willingness to collaborate with them in one’s work team. Although objective diversity is a necessary precondition for *intergroup* contact to occur, some researcher argue that the *perception* of diversity may be a more important determinant. We contribute by exploring both the effects of commonly-studied facets of *objective* diversity–nationality, migration background, age, and gender–alongside general *perceived* team diversity on the willingness to collaborate with asylum seekers: 1. *H1a*: *Objective team diversity in nationality*, *migration background*, *age*, *and gender is positively associated with the willingness to collaborate with asylum seekers in teams*. 2. *H1b*: *Perceived team diversity is positively associated with the willingness to collaborate with asylum seekers in teams*. ## Moderators of the relationship between team diversity and willingness to collaborate with asylum seekers The direct and indirect effects of intergroup contact are highly sensitive to contextual factors (for a review, see, e.g.,). Hence, this study further investigated the conditions that may influence the effects of work team diversity on collaboration intentions towards asylum seekers via secondary transfer effects of varying types of diversity. Such conducive conditions have already been introduced by Allport, as specified above. To our knowledge, these conditions have not been tested with regard to work team diversity. We are also unaware of empirical studies investigating Allport’s optimal conditions as moderators of secondary transfer effects. Therefore, we followed recent calls for empirical tests of these optimal conditions: equal status (operationalized as status of the asylum seeker in the team hierarchy), cooperation (operationalized as task interdependence), and support by authorities, laws, and custom (operationalized as pro-diversity team norms) by considering them as moderators of the relationship between team diversity and willingness to collaborate with asylum seekers. In doing so, we acknowledge that common goals are already an integral part of the definition of teams (e.g., and thus refrained from putting this condition to test. ### Equal status–status of the asylum seeker in the team hierarchy Allport considered status differences as an important factor inhibiting the emergence of positive intergroup relations. Status can be defined as “any characteristic, such as gender or skin color, on which people are hierarchically ordered as higher or lower” (, p. 59). As such, status serves as a fundamental organizer of social perception in contemporary societies. Social status is associated with competence perceptions related to respect and the outgroups’ perceived capability and agency. Task-related qualities, such as competence, are typically required and desired in the work context. In organizational contexts, perceived status and competence of social groups meet hierarchical structures within teams, which are themselves associated with perceptions of status and competence (e.g., team leader/supervisor roles vs. assistant roles). Role congruity theories suggest that perceived incongruity between the social status of a social group and of a position can result in a perceived inadequacy for members of that social group to take on such positions. Whereas leadership positions are associated with high status, esteem, competence, and power, majority members typically attribute low social status and low competence, and high need for assistance to asylum seekers. These perceptions are likely to affect performance expectations and thus (un)suitability perceptions for high competence positions. Additionally, majority members oppose status challenges from low-status groups. Accepting asylum seekers in a higher status position than one’s own can be seen as such a challenge. Thus, asylum seekers’ anticipated position in the team hierarchy (same status position vs. superior status position) should have a moderating impact on the relationship between team diversity and one’s willingness to work with asylum seekers. 1. *H2a*: *The positive association between objective diversity in nationality*, *migration background*, *age and gender and willingness to collaborate with asylum seekers will be stronger for same-status positions (colleague) compared to superior status positions (supervisor)*. 2. *H2b*: *The positive association between perceived team diversity and willingness to collaborate with asylum seekers will be stronger for same- status positions (colleague) compared to superior status positions (supervisor)*. ### Cooperation–task interdependence Group processes have a significant impact on team-related outcomes. One determining factor of intra-team cooperation is task interdependence, i.e., the extent to which team members have to rely on each other to accomplish their work. Within the intergroup contact literature, it has been suggested that intergroup cooperation in interdependent tasks is beneficial for positive intergroup attitudes to emerge. This contention has received widespread empirical support across a diverse array of team contexts and compositions. Although a minimal level of cooperation among team members, who–by definition–share common goals, can be seen as given in teams, levels vary as a function of task properties. Task interdependence is causally related to cooperation and is contingent for positive diversity effects to emerge. Tasks with higher task interdependence require more frequent communication and sharing of knowledge or resources to achieve group outcomes. Consequently, higher task interdependence enhances contact and cooperation levels between team members, which directly translates into higher levels of intergroup contact in diverse teams compared to homogenously composed teams. Thus, we argue that task interdependence in diverse teams, compared to homogenous teams, is related to higher levels of intergroup contact, which in turn should lead to stronger contact effects. 1. *H3a*: *Higher task interdependence increases the positive association between objective team diversity in nationality*, *migration background*, *age*, *and gender and willingness to collaborate with asylum seekers*. 2. *H3b*: *Higher task interdependence increases the positive association between perceived team diversity and willingness to collaborate with asylum seekers*. ### Pro-diversity team norms (support by authorities, laws, and customs) Allport suggested that contexts welcoming cross-group interaction are beneficial for positive intergroup attitudes to emerge. Recent studies suggest that social norms in a given context welcoming or rejecting intergroup interactions strongly influence prejudice levels (e.g.. Whereas descriptive norms refer to “the perception of what most people do” (, p. 202), injunctive norms refer to “norms that characterize the perception of what most people approve or disapprove” (, p. 202). Both norms play a vital role in coloring intergroup relations. For instance, an upward shift of positive attitudes towards refugees and other migrant groups in Canada just after President Trudeaux (an arguably more liberal and immigration-welcoming authority) took office can be explained by a shift of descriptive and injunctive norms. In line with these observations, individuals without personal intergroup encounters express more positive intergroup attitudes in contexts where intergroup contact is the norm. Typically, highly prejudiced individuals show a greater reduction of prejudice in contexts where intergroup encounters are the norm. Moreover, research has shown a positive trend between the effect sizes of intergroup contact on intergroup attitudes and the emergence of norms valuing equality between groups in the US. Consequently, we expect that pro-diversity work team norms will moderate the relationship between diversity and willingness to collaborate with asylum seekers. 1. *H4a*: *Higher pro-diversity team norms increase the positive association between objective team diversity in nationality*, *migration background*, *age*, *and gender and willingness to collaborate with asylum seekers*. 2. *H4b*: *Higher pro-diversity norms in teams increase the positive association between perceived team diversity and willingness to collaborate with asylum seekers*. # Method ## Sample and Procedure Low risk studies, like ours, do not require a formal clearance from an internal review board in Germany, where this study has been conducted. All procedures were performed in accordance with the ethical guidelines of the Deutsche Gesellschaft für Psychologie (German Society for Psychology). Data was collected online, and consent given in written form (per check box). We recruited 470 participants (*M*_*age* = 34.36, *SD*_*age* = 11.92; 55.74% female, 41.06% male, 3.2% missing; 69.79% with professional qualification, 28.94% without professional qualification, 1.27% missing; 93.83% German nationality; 4.04% non-German nationality, 2.13% missing; 16.80% with migration background, 79.36% without migration background, 3.8% missing) nested in 106 project and work teams (*M*_{*Team size*} = 10.57, *SD*_{*Team size*} = 8.49) in Germany in spring 2016. Thus, whereas the gender ratio in our sample was balanced and there was substantial variability in participants’ age, our data comprised participants of mostly German nationality, and participants without migration background. The participating teams were tasked with a wide range of different assignments, and stemmed from diverse sectors (19.4% public sector; 27.7% voluntary work; 0.03% police; 33.8% private sector; 10.6% other; 5.5% missing) and organizations (ranging from large retailers to small project teams). As an incentive, teams were provided with a summary of study results upon request. This data was collected in a larger collaborative data collection effort. Two further empirical articles were based on the collected data. These articles addressed entirely different research questions (focusing on the relation of team composition and team member’s mental health). Hence, both the research question and constructs used are unique to this article and do not constitute dual publication. We have uploaded the data set that contains all variables relevant to this project as well as a data transparency table on an open science framework page, <https://osf.io/8v6tm/>. ## Measures ### Objective team diversity Objective diversity in teams is often quantified using the standard deviation of continuous attributes (e.g., age) and the Blau Index for categorical variables (e.g., gender). The Blau Index quantifies the probability that two randomly selected team members would have different attributes. Following the literature on objective diversity, we calculated the level of dispersion in the teams using the standard deviation for age, and Blau Index for gender, migration background, and nationality. ### Perceived team diversity To measure perceived diversity, we adapted a scale from Meyer, Shelma, and Schermuly. On a scale from 1 = *completely disagree*, to 5 = *completely agree*, α =.81, participants indicated their agreement to four items (e.g., “Regarding its composition, my team is diverse”). ### Willingness to collaborate with asylum seekers We measured the individuals’ willingness to cooperate with asylum seekers on a self-developed scale ranging from 1 = *completely disagree*, to 5 = *completely agree*, with three items for each of the two status dimensions (colleague vs. superior). The same-status subscale included the following items: “I can hardly imagine an asylum seeker as a colleague in my team” (reverse coded), “I approve of an asylum seeker becoming part of my team”, “I can imagine well working together with an asylum seeker in my team”, α =.87. The higher status subscale included the following items: “I can hardly imagine an asylum seeker as a leader of our team”, “I approve of an asylum seeker as a leader of our team”, “I can imagine well working together with an asylum seeker as the leader of our team”, α =.90. ### Task interdependence To measure task interdependence, we adapted a scale from Van der Vegt and Janssen. On a scale from 1 = *completely disagree*, to 7 = *completely agree*), participants indicated their agreement to four items, such as: “I have a one- person job; it is not necessary for me to coordinate or cooperate with others” (reverse coded), α =.75. ### Pro-diversity team norms We used a four-item scale by Meyer & Schermuly to measure individual pro- diversity beliefs. The scale ranged from 1 = *does not apply at all*, to 5 = *applies fully*. Reliability analyses revealed inadequate internal scale consistency, α =.29. In a subsequent principal component analysis, two items (“Solving complex problems requires teams with different backgrounds and experiences”and “I prefer to work with people I consider similar to myself” (reverse coded)) loaded on a common factor, which we used in subsequent analyses. Since Cronbach’s Alpha is dependent on scale length, internal consistency of the 2-item version of the scale was acceptable, α =.51. Next, we group-mean centered individuals’ responses to aggregate them the team-level, representing pro-diversity norms within the team context. ### Control variables Besides the demographic variables age, gender, and the type of organization (1 = public sector; 2 = voluntary work; 3 = police; 4 = private sector; 5 = other), we also controlled for personal intergroup contact experiences with asylum seekers and individual-level pro-diversity beliefs. For the former, we used an adapted version of the one-item measure by Barlow et al.. On a scale from 1 = *never*, to 5 = *often*, we asked: “How often do you generally have contact with asylum seekers?”. For the latter, we used the participants’ deviation from the team’s group mean-centered pro-diversity norm score. # Results ## Preliminary analyses All analyses were performed using the R version 1.2.1 statistical environment using the lme4 package. The syntax for our main analyses can be found on this project’s open science framework page, accessible under <https://osf.io/8v6tm/>. Means, standard deviations, and bivariate correlations between measures on the individual level are summarized in. ICCs were calculated for the unconditional model including only the two random effects. The cross-level interaction accounted for about 40% of variance in willingness to work with asylum seekers (ICC =.399). Team ID accounted for about 29% of variance in willingness to work with asylum seekers (ICC =.285) suggesting that there is ample variance at both levels. Following recommendations, we started with the maximal random effects structure for each model given our sample size and reduced it from that point as necessary. As random-intercept random-slope models did not converge, all reported models are random-intercept models. ## Main analyses The results of all main analyses are summarized in. According to the first set of hypotheses, objective team diversity in nationality, migration background, age, gender (H1a), and perceived diversity (H1b) are positively associated with the willingness to collaborate with asylum seekers. To test these hypotheses, we computed linear mixed-effects models with random intercepts for subject and team. Data were nested in two levels (individuals within teams) which included the control variables team size, team type (dummy coded, with public sector as reference category), and pro-diversity norms on the team level, as well as age, gender, intergroup contact, mean-centered task interdependence, and mean- centered diversity beliefs on the individual level (Model 1). Next, we ran a model in which we additionally included all measures of team diversity (nationality diversity, migration background diversity, age diversity, gender diversity on the team level, and mean-centered perceived diversity on the individual level) simultaneously. Only the 95% confidence intervals of age diversity excluded zero, indicating a significant effect. Contrary to our expectations, higher age diversity was associated with *less* willingness to collaborate with asylum seekers, *b* = -0.04, 95% CI \[-0.07, -0.01\]. Therefore, our results did not support Hypotheses 1a and 1b. According to the second set of hypotheses, the positive association between objective diversity in nationality, migration background, age, and gender (H2a) and willingness to collaborate with asylum seekers, as well as perceived diversity and willingness to collaborate with asylum seekers (H2b), should be stronger for same-status positions (colleague) compared to superior status positions (supervisor). Hence, we included interaction effects between all diversity measures and the status of asylum seekers in Model 3. Only the 95% confidence interval of the interaction effect between age diversity and asylum seekers’ status on the willingness to collaborate with asylum seekers excluded zero, *b* = -0.03, 95% CI \[-0.04, -0.01\]. To facilitate the interpretation of the significant interaction, we plotted willingness to collaborate with an asylum seeker as a function of age diversity and status position following the guidelines provided by Preacher, Curran, and Bauer (, see). Contrary to our hypothesis, age diversity was *more negatively* related to the willingness to collaborate with asylum seekers as a superior, *b* = -0.05, *p* \<.001, compared to as a colleague, *b* = -0.03, *p* =.068. Overall, the results neither supported H2a, nor H2b. According to the third set of hypotheses, higher task interdependence increases the positive association between objective team diversity in nationality, migration background, age, and gender (H3a) and willingness to collaborate with asylum seekers, as well as perceived diversity (H3b). Thus, in Model 4, we dropped the interaction effects of Model 3, and included interaction effects between all diversity measures and task-interdependence on the willingness to work with asylum seekers. Only the 95% confidence intervals of the interaction effects of nationality diversity and task interdependence, *b* = -0.69, 95% CI \[-1.34, -0.03\], as well as migration background diversity and task interdependence, *b* = 0.44, 95% CI \[0.06, 0.83\], excluded zero. Again, significant interactions are illustrated in Figs and following Preacher et al.. Unexpectedly, none of the simple slopes involving nationality diversity were significant. In teams with low task interdependence, migration background diversity and willingness to work with asylum seekers were not significantly related, *b* = -0.12, *p* =.758. In teams with high task interdependence, migration background diversity and willingness to collaborate with asylum seekers were positively associated, *b* = 0.95, *p* =.014. Thus, the pattern of results showed partial support for H3a when considering migration background diversity, whereas H3b was not supported. According to the fourth set of hypotheses, higher pro-diversity norms in teams increase the positive association between objective team diversity in nationality, migration background, age, and gender (H4a) and willingness to collaborate with asylum seekers, as well as perceived diversity and willingness to collaborate with asylum seekers (H4b). Thus, in Model 5, we dropped the interaction effects of Model 4 and included interaction effects between all diversity measures and team-level pro-diversity norms on the willingness to work with asylum seekers instead. None of the 95% confidence intervals of the interaction effects excluded zero, indicating that both H4a and H4b were not supported by our data. As the relationship between objective and perceived diversity and willingness to work with asylum seekers was not moderated by team-level pro-diversity norms, we conducted further exploratory analyses with individual-level (group-mean centered) diversity beliefs. Results of these analyses are reported in the online supplementary materials. # Discussion The successful integration of newcomers in the labor market of receiving countries is among the most pressing issues associated with increased forced migration worldwide. Acknowledging that this integration can only be successful if employees are willing to collaborate with asylum seekers at work, we investigated the effects of objective and perceived diversity in work teams on employees’ collaboration intentions with asylum seekers. We tested our predictions in Germany, a country that was a major refugee-receiving country within Europe since 2015. In line with the intergroup contact hypothesis, we expected higher levels of team diversity to lead to more positive attitudes towards the respective outgroup, which in turn should make members of diverse teams more open to other newcomers as well (H1a, H1b). However, contrary to our expectations, team composition was not significantly associated with willingness to collaborate with asylum seekers overall. Only age diversity was significantly related to the willingness to collaborate, yet, contrary to our expectations, *negatively* so. As such, our results are at odds with research suggesting that diverse contexts are on average more welcoming of newcomers. Although the body of literature that finds positive diversity effects is large, prior work has found that diversity may also have negative effects on the openness to newcomers and suggested boundary conditions under which diversity may unfold more likely conducive or detrimental effects. Thus, one explanation might be that the moderators might have been aligned in such a way in our sample that the net effect of diversity was neutral or, in the case of age diversity, negative, which might have not been the case in previous studies we used to deduce our first set of hypotheses from. We propose an additional post-hoc explanation for the negative effect of age diversity. Highly diverse teams typically consist of members that are both young and old. Age is generally associated with more conservative views and prejudice towards outgroups. Consequently, older team members may, on average, express stronger reservations to collaborate with asylum seekers. This is supported by the negative correlation between age and willingness to work with asylum seekers (see). Younger team member may be aware of the likelihood that their older team members might be less accepting of asylum seekers than others. As a consequence, those younger participants, who may otherwise be more accepting of asylum seekers in their teams, might have reservations to welcome asylum seekers into their age diverse team. Irrespective of this speculation, our finding indicates that age diversity as a facet of team diversity can have harmful consequences for the willingness to work with asylum seekers. Adding to the debate of moderators, we also proposed and tested moderator candidates of the influence of team diversity on the willingness to collaborate with asylum seekers. Based on Allport’s original formulation of the intergroup contact hypothesis, we tested to what extent the status of an asylum seeker in the team hierarchy (equal status vs. higher status), task interdependence (cooperation), and pro-diversity team norms (support by authorities, laws, and customs) qualified the relationship between team diversity and willingness to collaborate with asylum seekers. As for task interdependence (Allport’s cooperation), we had expected that higher task interdependence increases the positive association between team diversity and willingness to collaborate with asylum seekers (H3a, H3b). As expected, in teams that reported higher task interdependence, migration background diversity and willingness to collaborate were more closely and positively associated than in teams that reported lower task interdependence. This finding supports the idea that positive interdependent cooperation in diverse settings fosters more welcoming attitudes towards newcomers. Although a significant interaction effect with nationality diversity did emerge, none of the simple slopes reached statistical significance. Additionally, a closer inspection of this diversity variable revealed that our sample was highly skewed towards national homogeneity, reducing the trustworthiness of these results. Future research with a more nationally diverse sample may produce more robust findings. We could not replicate the moderating effect of task interdepence on the association between migration background diversity and willingness to work with asylum seekers across other diversity dimensions. One explanation might be that–in line with work on the secondary transfer generalization gradient–the other diversity categories could be too dissimilar from those of asylum seekers. Consequently, daily interactions in work settings with people with migration background might make people especially more welcoming towards outgroups that have a migration history when team member need to rely on each other to get their work done. Overall, this findings demonstrates that the migration background diversity facet of team diversity can have beneficial effects on the willingness to work with asylum seekers, yet only in teams with high task interdependence. As for the status of the position of the asylum seeker in the team hierarchy (Allport’s equal status), we had expected accepting asylum seekers in a higher status and power position than one’s own can be challenging. Thus, we expected that the *positive* association between diversity and willingness to collaborate with asylum seekers should be stronger for same-status positions (colleague) compared to superior status positions (supervisor; H2a, H2b). Contrary to our expectations, the status of the position of the asylum seeker did not play an important role in qualifying the effects of diversity on willingness to work of asylum seekers most of the time. When it did, results were not in line with our expectations either. This is, because the *negative* association between age diversity and willingness to collaborate with asylum seekers was stronger for an asylum seeker as a superior compared to an asylum seeker as a colleague. Our speculations about the particularities of age diversity elaborated on above might help explain these, on the first glance surprising, findings: Whereas older team members of highly age diverse teams may express their reservations to collaborate with asylum seekers, younger team member that may have had more positive views towards asylum seekers may have adjusted their answers to account for this circumstance, rejecting asylum seekers in a superior position more strongly than in a same-status position. As for the moderator pro-diversity norms (Allport’s support by authorities, laws, and customs), we expected that higher pro-diversity norms in teams increase the positive association between team diversity and willingness to collaborate with asylum seekers (H4a, H4b). Among other things, these postulations were based on the idea that welcoming norms are one of the important ingredients for positive intergroup interactions in diverse teams to emerge, which in turn should make them more welcoming contexts for newcomers. This, however, was not the case. Overall, our study provided unique insights into the relationship between various facets of team-level diversity and individual-level willingness to work with asylum seekers in teams in a number of ways. As such, we added to the literature that proposes and debates factors that determine the successful integration of asylum seekers at the workplace. Our findings demonstrate that, most of the time, team diversity plays no central role in shaping the willingness to work with asylum seekers. Also, we did not find evidence for the idea that the perception of diversity might systematically play a more prominent role in shaping intergroup attitudes than objective diversity, as prior literature suggested. Nonetheless, our results indicate that team diversity can have both harmful and beneficial consequences for the willingness to work with asylum seekers, depending on the kind of diversity considered, and its boundary conditions. We recruited a large number of teams across different industries to address research questions of high academic, social, and political relevance during a time in which the need for successful integration of asylum seekers in the labor market was particularly high (i.e., in 2016, during the heydays of the so-called European “refugee crisis”). Given this circumstance, we advise future research to replicate our study to test the robustness of our findings in different time and country settings. Moreover, we used a cross-sectional design, which limits our ability to draw causal conclusions. Although alternative causal relationships are implausible for most of the variables we included in our analyses (e.g., it is rather unlikely that willingness to work with asylum seekers predicts diversity in teams), we encourage future research to account for this limitation. Lastly, our research revealed a number of unexpected relationships between the constructs we investigated that invited us to speculate about them. We want to stress the need for future research to follow up on these findings to advance our understanding of what kind of team diversity under which conditions has positive, negative, or no substantial effects on willingness to work with asylum seekers. [^1]: The authors have declared that no competing interests exist.

# Introduction Nitrogen contamination is one of the main concerns in wastewater treatment. The complete elimination of nitrogen from water involves many metabolic steps, which are catalysed by microorganisms with complex metabolic requirements. Biological treatments have provided practical solutions to these difficulties and have become reliable and highly cost-effective methods for nitrogen removal in wastewater treatment plants (WWTPs). Ammonium is the major nitrogen compound in wastewater, and sequential aerobic and anaerobic processes are needed to ensure its complete removal using a coupled nitrification-denitrification process. However, in wastewater exhibiting a low carbon to nitrogen ratio, the conventional heterotrophic process of denitrification becomes challenging. This carbon dependence reinforces the need for the development of autotrophic processes to avoid the additional costs involved in adding organic matter to water treatments. Denitrifying microbial fuel cells (MFCs) are among the most promising technologies that are currently used to partially suppress the carbon dependence of denitrification. In a MFC, organic substrates are oxidised by exoelectrogenic bacteria, which produce electrons that are transferred to an anode electrode and then flow to a cathode. The anode and cathode are linked by a conductive material containing a resistor. Protons produced at the anode migrate through the solution across a cation exchange membrane to the cathode chamber where electrotrophic bacteria combine protons with a reducible compound and electrons. Complete denitrification, using either nitrate or nitrite as electron acceptors, has been achieved using biocathodic bioelectrochemical systems. Chemolithoautotrophic denitrifiers regulate the denitrification process in the cathode of MFCs when no organic matter is present, similarly to what have been proposed for nitrogen-contaminated groundwater. Nitrate reduction on the cathode in a microbial fuel cell requires significantly lower influent COD/N ratios than for traditional nitrogen removal in wastewater treatment plants. In the latter, COD/N ratios between 7 and 10 are typically required. Virdis *et al*. found that the ratio between the COD removed at the anode and the nitrate removed at the cathode, was equal to 4.48 g COD g⁻¹ NO^3–N. Previous studies have shown that the cathodes of MFCs harbour complex bacterial communities, including members of *Proteobacteria, Firmicutes* and *Chloroflexi* phyla as the most abundant species. Wrighton *et al.* showed that the observed changes in the dominant members of the bacterial community did not correspond with the changes in reactor functioning. This observation potentially reflects the fact that most of the observed phylotypes did not exhibit relevant denitrification activity, suggesting that information about functional groups is required for a better understanding of the process. Denitrification is an anaerobic respiration pathway catalysed by taxonomically diverse bacteria and archaea. Thus, molecular studies are often directed to the use of functional genes. The commonly used markers for nitrate reduction include either membrane-bound (Nar type) or soluble periplasmic (Nap) nitrate reductases, encoded by *narG* and *napA* genes, respectively. Dissimilatory nitrite reductases (Nir) exist as two functionally equivalent enzymes: a cytochrome *cd₁*-type and a copper containing encoded by the *nirS* and *nirK* genes, respectively, which have been thoroughly used as molecular markers for denitrification. The reduction of nitrous oxide that occurs during the last step in the denitrification pathway has received most of the attention in molecular studies. Nitrous oxide reductase mediates the production of nitrogen gas and is coded by the *nosZ* gene. The sequential nature of the denitrification process, in which the products of a reaction serve as substrates for the next reduction step, makes this process an excellent example to study cooperation through electron transfer between different bacterial cells. The direct electron transfer between cells of different species has been demonstrated in the anodes and cathodes of MFCs. In fact, the functioning of MFCs partly relies on the capacity of microorganisms to accept electrons from the cathode directly via membrane bound cytochromes or indirectly via added (exogenous) or secreted (endogenous) mediators. Our goal was to identify the relevant players of nitrate, nitrite and nitrous oxide reductions as key metabolic steps in the denitrification process in the biofilms generated in the cathode of a MFC at different operating conditions. The different conditions were determined in order to assess the effect of different electron donors (electron transfer between the electrode or organic matter) and different electron acceptors (nitrate or nitrite) in the microbial community. We examined the community composition and abundance of nitrate, nitrite and nitrous oxide reducing bacteria using five functional genes of the denitrification pathway. Abundances of functional genes were analysed through quantitative PCR, and the structure of the nitrate reducer, nitrite reducer and nitrous oxide reducer bacterial communities were assessed using a cloning- sequencing approach. MFC performances, in terms of nitrogen removal, power density generation and nitrous oxide accumulation, were compared for the different electron acceptors and donors used and related to prevailing denitrifiers. # Materials and Methods ## Experimental Set-up The MFC consisted of an anode and a cathode placed on opposite sides of a single methacrylate rectangular chamber (dimensions, 29×26×400 mm; empty bed volume of 265 mL and 385 mL for anode and cathode chambers, respectively). The anode and cathode chambers were filled with rod graphite (Alfa Aesar, Germany), which reduced the compartment volumes to 120 and 145 mL (net anodic and cathodic compartments, NAC and NCC), respectively. Two thinner graphite electrodes (Sofacel, Spain) were connected to an external resistor (100 Ω) to close the electric circuit. A cation exchange membrane (CEM, Nafion 117, Dupont) was placed between the anode and cathode frames. The medium was continuously fed into a recirculation loop to maintain well-mixed conditions and avoid concentration gradients. The system was thermostatically controlled at 23±2 °C. Prior to treatment, the MFC was inoculated with 50 mL of effluent from the anode of a parent MFC that was previously used to treat synthetic wastewater primarily composed of sodium acetate and a buffer solution. ## Microbial Fuel Cell Operation The denitrifying MFC was operated for one year to treat acetate-enriched wastewater in the anode. The biocathode was fed with different electron acceptors (nitrate and nitrite) in two different media (mineral media and effluent from an air-cathode MFC treating urban wastewater). shows the main influent characteristics of the anode and cathode compartments during the experimental period. To determine the influence of the electron acceptor on the MFC performance and the cathode community dynamics, the experimental study was divided into three different periods. Period 1 was operated at constant conditions for 76 days, period 2 for 105 days, and period 3 for 167 days. The cathode was fed differently depending on the operational period. In periods 1 and 3, the enriched mineral medium contained 0.488 g L⁻¹ NaHCO₃, 0.2 g L⁻¹ NaNO₃, 0.92 g L⁻¹ NaH₂PO₄ 2H₂O, 0.0056 g L⁻¹ CaCl₂ 2H₂O, 0.036 g L⁻¹ MgSO₄·7H₂O, 0.0052 g L⁻¹ KCl and 0.1 mL L⁻¹ of a microelements solution (SL10). The nitrate substrate was added in period 1 as an electron acceptor that was subsequently replaced by nitrite in period 3. These periods are hereafter respectively referred to as autotrophic with nitrate and autotrophic with nitrite. During period 2, the effluent from an air-cathode MFC typically used for treating urban wastewater was used. This feeding condition is referred as heterotrophic with nitrate in the results section. During period 1, the cathode of the MFC was fed with 2.0±0.3 L·d⁻¹ of nitrate-enriched mineral medium (26.6±1.3 mg N-NO₃⁻·L⁻¹). After 77 days of operation, 1.4±0.5 L·d⁻¹ of the effluent from an air-cathode MFC used to treat urban wastewater was fed to the cathode of the MFC. The effluent of the air- cathode MFC also contained residual organic matter (64±21 mg COD·L⁻¹). This concentration was similar to that of the non- biodegradable organic matter in urban wastewater. The third period began on day 183 in which 1.2±0.2 L·d⁻¹ of nitrite-enriched mineral medium (20.3±1.7 mg N-NO₂⁻·L⁻¹) was fed without any organic matter content. The anodic feed consisted of a nitrogen-purged medium enriched with acetate. The medium contained 1.44 g L⁻¹ NaCH₃COOH; 0.488 g L⁻¹ NaHCO₃; 0.03 g L⁻¹ NH₄Cl; 0.92 g L⁻¹ NaH₂PO₄·2H₂O; 0.0056 g L⁻¹ CaCl₂·2H₂O; 0.035 g L⁻¹ MgSO₄·7H₂O; 0.0052 g L⁻¹ KCl; 0.044 g L⁻¹ NaNO₃; and 0.1 mL L⁻¹ of a microelements solution (1 g L⁻¹ FeSO₄·7H₂O; 70 mg L⁻¹ ZnCl₂; 100 mg L⁻¹ MnCl₂·4H₂O; 6 mg L⁻¹ H₃BO₃; 130 mg L⁻¹ CaCl₂·6H₂O; 2 mg L⁻¹ CuCl₂·2H₂O; 24 mg L⁻¹ NiCl₂·6H₂O; 36 mg L⁻¹ Na₂Mo₄·2H₂O; and 238 mg L⁻¹ CoCl₂·6H₂O). ## Analyses and Calculations Liquid-phase samples for organic matter (chemical oxygen demand, COD) and nitrogen (ammonium: N-NH₄⁺; nitrite: N-NO₂⁻ and nitrate: N-NO₃⁻) were sampled regularly and analysed according to the Standard Methods for the Examination of Water and Wastewater. The levels of nitrous oxide (N-N₂O) production were estimated according to the electron balance at the cathode following the methodology of Virdis *et al.*. Experiments carried out using liquid- and gas-phase N₂O analysers demonstrated excellent fits between measured data and estimated data using the electron balance. The nitric oxide (NO) production was considered to be negligible. To close the mass balance, the level of dinitrogen gas in the effluent was calculated from the current produced according to the following equation. *ΔNO₃^–* is the nitrate consumption rate, whereas *ΔNO₂^–*, *ΔNO*, and *ΔN₂O* are nitrite, nitric oxide, and nitrous oxide production rates, respectively. *I* is intensity, *V* voltage applied, and *F* the Faraday’s constant. The cell potential (V) in the MFC circuit was monitored at one-minute intervals using an on-line multimetre (Alpha-P, Ditel) equipped with a data acquisition system (Memograph M RSG40, Endress+Hauser). The current (I) was calculated according to Ohm’s law. The current density was calculated by dividing the current by the net cathodic volume (A·m⁻³ NCC). The Coulombic efficiencies for nitrate and nitrite reduction were calculated according to Virdis *et al*.. ## Community Sampling and DNA Extraction Two graphite rods (6×38 mm) were collected aseptically from different positions in the cathode chamber at steady state conditions at days 51 (autotrophic growth with nitrate), 106 (heterotrophic growth with nitrate) and 350 (autotrophic growth with nitrite) of operation. The rods were replaced with an equal number to maintain the working volume of the cathode chamber. The graphite rods were chilled on ice after collection and processed as independent samples for DNA extraction within two hours after sampling. Bacterial biofilms were detached from the graphite rods using the following procedure. The rods were washed three times in Ringer solution (Sharlau®, Barcelona, Spain). Subsequently, the rods were immersed in 4 mL of 0.1 M of sodium pyrophosphate (Na₄P₂O₇·10H₂O), and the biofilm dislodged using three consecutive sonication rounds for 20 seconds followed by 30 seconds on ice. The suspended bacterial cells were pooled and centrifuged at 10,000 rcf for 2 minutes. The DNA was extracted using the FastDNA® SPIN Kit for soil (MP, Biomedicals) following the manufacturer’s instructions. The obtained DNA was quantified using a NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Inc., Wilmington, DE) and stored at −20°C. ## Amplification and Quantification of Genes Using PCR Five functional genes of the denitrification pathway (*narG*, *napA*, *nirS*, *nirK* and *nosZ*) were PCR amplified using primers and conditions previously published with minor modifications. In all cases, the PCR reactions were performed in a total volume of 50 µL containing: 1X PCR buffer, 0.2 mM of each deoxynucleotide triphosphate and 1 U of Taq polymerase. All chemicals and reagents were obtained from Qiagen (Qiagen, Germany). The PCR amplifications were performed in a Gene Amp® 2700 thermal cycler (Applied Biosystems). The correct size of amplification products was verified using electrophoresis on a 1.5% agarose gel and visualised through ethidium bromide staining. The gene abundances of denitrifying bacteria were determined using quantitative PCR (*q*PCR). The *q*PCR amplification was performed for the functional genes *narG*, *napA*, *nirS*, *nirK* and *nosZ*. Additionally, the bacterial 16S rRNA gene was also quantified and used as a proxy for bacterial abundance. All reactions were performed in a 7500 Real Time PCR system (Applied Biosystems) using the SYBR® Green PCR Mastermix (Applied Biosystems). The reactions were performed with a 20 µl final volume containing 1X SYBR® Green Master Mix, 1 µg/µl BSA, 10 ng of DNA, and 1 µM of each primer. All primers were obtained from Biomers. The standard curves were generated using serial dilutions (from 10² to 10⁹ copies/reaction) of plasmids containing known sequences of the targeted genes. The PCR efficiency ranged between 87.0 and 104.9%. The negative controls resulted in undetectable values in all *q*PCR reactions. An inhibition test was performed before the *q*PCR assays were done. Every sample was evaluated for inhibition independently. To detect possible inhibitory effects, 10⁵ copies of the plasmid DNA (pGEM®-T Easy, Promega, Madison, WI) were mixed with 10 ng of sample DNA and quantified with plasmid specific primers T7 and SP6. In all samples, the number of copies of the plasmid compared to the 16S rRNA concentration was sufficiently low (2 orders of magnitude) to detect significant variations of cycle thresholds (Ct) in case of inhibition. In all cases Ct values when the plasmid was measured alone differed for less than 0.08±0.23 when compared with values obtained in inhibition tests. No significant differences in the inhibition tests were observed between samples of the three periods. The relative contributions of the functional genes (*narG*, *napA*, *nirS*, *nirK* and *nosZ*) compared with the16S rRNA gene were calculated as a proxy for denitrifying bacteria abundance. The gene abundances and gene ratios were log transformed to ensure a normal distribution of the data. The normality was assessed for all variables, except for the abundances of the *narG* and *nirK* genes, using the Shapiro-Wilk tests. One-way ANOVA and post hoc tests (Tukey) were used with log-transformed data to characterise the effects of feeding regimes applied to the cathode on the abundance of different genes and ratios when equal variance of data was observed. Alternatively, non-parametric analyses (Kruskal-Wallis test) were used. All statistical analyses were performed using SPSS for Windows 15.0 (SPSS, Inc). ## Cloning of Functional Genes and Phylogenetic Analysis The PCR products for the *narG*, *napA, nirS*, *nirK* and *nosZ* genes were purified using the QIAquick PCR purification kit (Qiagen, Germany) and cloned using the TOPO TA Cloning® Kit for Sequencing (Invitrogen, Eugene, OR) according to manufacturer’s instructions. At least 300 clones per gene were individually selected and screened by PCR using the primers M13. Amplicons of the expected size were sequenced at Macrogen (Macrogen, the Netherlands). The sequences were examined for chimeras using the UCHIME algorithm and manually refined using the BioEdit v7.0. The reference sequences for the *narG*, *napA*, *nirS*, *nirK* and *nosZ* genes were obtained from completely sequenced genomes in the GenBank database, aligned using CLUSTALW and used as a template file to define Operational Taxonomic Units (OTUs) using Mothur v.1.22.1. OTUs were defined at threshold values of 33% (*narG*), 18% (*nirS*), 17% (*nirK*) and 20% (*nosZ*). The threshold cut-off values for *napA* were set at 21%. This value was estimated as the species level cut-off value according to pair-wise comparisons of *napA* and 16SrRNA gene sequences of 23 completed genomes deposited in the GenBank database (results not shown) using a previously described computation procedure. Defined OTUs were used to calculate rarefaction curves and estimate the richness (Chao1) and diversity indices (Shannon). Deduced amino acid sequences were obtained for the representative sequences of each OTU and functional gene and aligned using the ClustalW algorithm in MEGA v.5.0. Phylogenetic trees were reconstructed by neighbour-joining using the pair-wise deletion and p-distance methods. The tree topology was evaluated using bootstrap analysis with 10,000 replicates. Differences in the community composition based on the phylogeny of the *narG*, *napA*, *nirS*, *nirK* and *nosZ* genes were analysed from the tree topologies using a weighted Unifrac test. ## Nucleotide Sequence Accession Numbers The sequence data derived in this study have been submitted to the GenBank database under the accession numbers JX236709 - JX236736 (*napA* gene), JX236737 - JX236898 (*narG* gene), JX237055 - JX237212 (*nirS* gene), JX236899 - X237054 (*nirK* gene) and JX237213 - JX237355 (*nosZ* gene). # Results ## Denitrifying Cathode Performances Under Different Feeding Conditions The MFC operated for one year under different cathodic feeding periods: autotrophic with nitrate (Period 1), heterotrophic with nitrate (Period 2) and autotrophic with nitrite (Period 3). summarises the different performances of the MFC at those conditions. During the experimental study either nitrate or nitrite were removed in the cathode and a current was subsequently produced. The highest current production (15 A·m⁻³ NCC) and cathode Coulombic efficiencies (CE, 85%) were achieved when nitrate was used as the electron acceptor under strictly autotrophic conditions (Period 1). When organic matter was also introduced into the cathode (Period 2), electrotrophy was affected, decreasing the current production (11 A·m⁻³ NCC) and cathode Coulombic efficiency (58%). Nitrate was partially removed via conventional heterotrophic denitrification (using organic matter instead of electrons from the anode as the electron donor). The use of nitrite as an electron acceptor without the presence of organic matter resulted in an increase in current production (14.1 A·m⁻³ NCC) but cathode Coulombic efficiency was decreased (41%). The decrease of the CE suggested the presence of denitrification intermediates (nitrite or nitrous oxide) in the effluent. shows the concentrations of the intermediates in the effluent of the cathode. As expected, nitrous oxide was produced at higher rates during Period 3. In this period, up to 70% of the nitrogen removed was converted into nitrous oxide. In contrast, when nitrate was feed during period 1, nitrous oxide level accounted for only 14% of the nitrogen removed. ## Quantification of the *narG, napA, nirK, nirS, nosZ* and 16S rRNA Genes The abundance of the 16S rRNA gene varied between 1.89×10⁶ and 5.07×10⁶ copies/ng DNA and was significantly higher (*p*\<0.001) than most of the functional genes analysed. During the heterotrophic period, the abundance of 16S rRNA gene was significantly lower (*p*\<0.026) compared to the autotrophic periods. The abundance of all functional genes varied between 0.3±0.0×10⁴ and 2.2±0.1×10⁴ gene copies/ng DNA, except for *nirS* and *nosZ*, which were detected at higher concentrations. The amount of *nirS* varied from 171.0±48.4×10⁴ to 264.8±32.0×10⁴ gene copies/ng DNA, but no significant differences were observed between periods. On the contrary, the abundance of *narG*, *napA*, *nirK* and *nosZ* showed significant differences (*p*\<0.005) according to the operating conditions. The relative contribution of the functional genes (*narG*, *napA*, *nirS*, *nirK* and *nosZ*) compared with the16S rRNA gene varied according to the feeding conditions. The q*narG*/q16S rRNA ratio was significantly higher (*p*\<0.01) when nitrate was used as the sole electron acceptor. The q*nirK*/q16S rRNA ratio was significantly different at all conditions tested and higher during the heterotrophic period. The q*nirS*/q16S rRNA ratio varied from 0.4 to 0.9, which were higher during the heterotrophic period (*p*\<0.05). The q*nosZ*/q16S rRNA varied from 0.028 to 0.054 and was significantly lower (*p*\<0.026) during heterotrophic growth. The gene ratios were used to evaluate the capacity of the system to reduce nitrate to nitrogen gas during the sequential process of denitrification. (q*narG*+q*napA*)/q*nosZ* and (q*narG*+q*napA*)/(q*nirS*+q*nirK*) were consistently below a value of one, indicating a lower amount of nitrate reductases compared with the other steps in the denitrification pathway. Significant differences of both ratios were observed in relation to feeding conditions. The (q*nirS*+q*nirK*)/q*nosZ* ratio varied from 7.99 to 32.45, indicating a great potential for the accumulation of intermediate gases. ## Community Structure of Denitrifying Bacteria A total of 619 sequences, 162 for *narG*, 158 sequences for *nirS*, 156 for *nirK* and 143 for *nosZ*, were obtained from the cloning assay and used in the present study. Unfortunately, only 28 high-quality sequences could be obtained for the *napA* gene, although the cloning effort included the screening of 490 clones, which were too few to obtain an adequate analysis of the *napA*-containing community, thus the results have only been included for comparison with other genes. The diversity and phylogenetic analyses were conducted on the basis of operational taxonomic units (OTUs). Rarefaction curves were constructed to visualise the saturation of the bacterial diversity. Except for the *napA* gene, the coverage values for all samples were higher than 90%, indicating that a representative portion of the bacterial diversity was covered. The maximum richness (number of defined OTUs) was estimated according to Chao1 and varied from 5 to 32. Maximum values were identified for *nirS*-containing denitrifiers. The higher complexity of the *nirS*-containing community was confirmed from estimates of the Shannon diversity index. In contrast, the lowest diversity was observed for nitrous oxide reductase (*nosZ*) under all conditions. The *narG* gene sequences were grouped into 11 different OTUs. OTU 1 and 2 were the most abundant, which comprised 76 and 53 sequences, respectively. OTU 1 was almost exclusively found during the autotrophic periods, whereas OTU 2 was predominantly found during the heterotrophic period. The OTU 1 and 2 representative sequences were approximately 81% similar to the betaproteobacterium *Thiobacillus denitrificans* and the alphaproteobacterium *Methylobacterium nodulans*, respectively. OTU 3 (19 sequences) was exclusively observed during the heterotrophic period and showed a low sequence similarity with most cultivated bacteria. Maximum similarities (73%) were observed with *Polaromonas naphthalenivorans*. The sequences of the periplasmic nitrate reductase (*napA*) were distributed into 10 different OTUs. The most abundant OTU (10 sequences) was shared between autotrophic and heterotrophic periods supplemented with nitrate and showed the highest sequence similarity (85%) to *Dechlorosoma suillum*. OTU 2 was similar to *Sinorhizobium fredii* (78%) and was exclusively observed during the autotrophic period with nitrite. The *nirS* sequences were assigned to 31 different OTUs without a clear dominance. The most abundant OTUs, 1, 2 and 3, were affiliated with betaproteobacteria and showed relatively high similarities (\>84%) with *Dechlorosomonas* sp., *Thauera* sp. and *Cupriavidus pauculus*, respectively. OTU 1 was exclusively observed under nitrate feeding conditions whereas OTUs 2 and 3 were found at all operating conditions. OTUs 4, 8 and 11were observed almost exclusively during nitrite feeding conditions. According to BLAST searches with the reference genomic sequences database, the highest similarities of these sequences were found with the alphaproteobacterium *Paracoccus denitrificans* (OTU 8) and the gammaproteobacterium *Rhodanobacter* sp. (OTUs 4 and 11). The gene encoding the copper-containing nitrite reductase, *nirK,* showed a different distribution between samples. Two out of a total of 22 OTUs were clearly dominant (up to 70 sequences) during autotrophic periods supplemented with nitrate and nitrite. The representative sequences of OTUs 1 and 2 were similar to *Sinorhizobium fredii* (84%) and *Rhodopseudomonas palustris* (85%), respectively. In contrast, during the heterotrophic period, *nirK* sequences distributed into 18 different OTUs. The most abundant OTU (82% similar to *Mesorhizobium* sp. 4FB11) comprised only 10 sequences. The OTU distribution of *nosZ* genes revealed a relatively homogenous community for all periods. Almost 90% of sequences were grouped into a single OTU with a relatively high similarity (85%) to the predicted nitrous oxide reductase gene of *Oligotropha carboxidovorans*. The significance of the observed differences between the microbial communities under different operating conditions in the MFC, was analysed using pair-wise weighted UniFrac analysis for the 5 molecular markers. The UniFrac values confirmed the observed differences in the community composition, although statistically significant differences were only observed for *napA* and *nirS* containing communities. In both cases, significant differences were observed for the community of the autotrophic with nitrite period compared to the other periods. Low UniFrac values were obtained in all pair-wise comparisons for the *nosZ* community indicating a highly similar and stable community in all operating periods. # Discussion ## Influence of Cathode Feeding Characteristics on the Denitrifying MFC Performance The power production and efficiency of nitrogen removal were influenced by the three cathodic influents (nitrate, nitrate plus organic matter and nitrite). The highest current production (15 A·m⁻³ NCC) was achieved when nitrate was used as the electron acceptor. In contrast, when organic matter was added, heterotrophic denitrification was kinetically favoured over autotrophic denitrification, and current production was reduced to 11 A·m⁻³ NCC. The use of nitrite as the initial electron acceptor without the presence of organic matter increased the current production to 14.1 A·m⁻³ NCC. Bacteria gain energy by transferring electrons from a reduced substrate at a low potential to an electron acceptor with a higher potential. The nitrate reduction potential (E°′ = +0.433 V *vs*. standard hydrogen electrode, SHE) is higher than that for nitrite (E°′ = +0.350 V *vs*. SHE), therefore cathodic denitrification from nitrate is thermodynamically favourable under autotrophic conditions. The amount of energy available for the bacteria to grow in the cathode compartment using nitrate is estimated to be an 11% higher than nitrite. The use of nitrate was not only beneficial for energy production, but also for the minimisation of nitrous oxide production, a gas with a strong greenhouse effect, and the increase of the maximum Coulombic efficiency. Apart from N₂ production, N₂O represents a major sink during nitrate removal in the MFC. The reasons for the changing potential of N₂O emissions are multifactorial and include differences in the prevailing physicochemical conditions (*i.e.* temperature, pH and C/N ratio), the nature of the main electron acceptor and the composition of the denitrifying community. Under conditions of low electron availability (*i.e.* in the presence of organic matter), a lower affinity of the N₂O reductase towards the electron donor facilitates the accumulation of this intermediate. In addition, other studies suggest that the carbon source significantly impacts on the net N₂O emission but exhibits a relatively minor effect on N₂O production. When nitrite was used as an electron acceptor in the denitrifying MFC, the Coulombic efficiency decreased to a 41%, compared to nitrate, and the N₂O concentration in the effluent increased to 13 mg N-N₂O·L⁻¹ (a 70% of produced gases). It was proven that the use of nitrite as an electron acceptor and its accumulation in biological wastewater treatments cause an increase of NO and N₂O emissions during denitrification, agreeing with the results obtained in the MFC. ## Influence of the Cathode Feeding Characteristics on Denitrifier Communities The MFC set-up provided excellent conditions to assess how changes in the main electron acceptors (nitrate *vs*. nitrite) and donors (cathode *vs*. organic matter) affected the composition of the denitrifier community and how it was related to the MFC performance. In the present study the PCR primers used are biased towards detecting mainly *Proteobacteria* (*nirK* and *nirS*) and the newly called Clade I *nosZ* gene, thus underestimating the actual diversity and abundance of nitrate and nitrous oxide reducers. However, previous works analysing the bacterial diversity on MFC cathodes have shown these groups as particularly dominant in the biofilm community thus minimizing the impact of primer biases. Analysis of functional denitrification genes may reflect discrepancies, especially at the species-level, with the phylogeny of bacteria due to horizontal gene transfer and gene duplication events that may have occurred during evolution, and taxonomic inferences have to be taken cautiously. The abundance of *narG* and *napA* containing nitrate reducers increased during nitrate addition thus showing the importance of these communities in the first reduction step. Changes in the abundance were accompanied by changes in the OTU composition. A large number of the retrieved *narG* sequences during autotrophic treatments supplemented with either nitrate or nitrite showed a high similarity to *narG* of the obligate chemolithoautotrophic bacterium *Thiobacillus denitrificans*. *T. denitrificans* has an optimal pH for growth around 7.5–8.0 and high denitrification rates (0.78 g NO₃⁻ g cell⁻¹·h⁻¹), which fall in the same range of those estimated in the MFC according to bacterial abundances. During the heterotrophic treatment, in contrast, analyses of the *narG* containing community revealed a higher relative abundance of sequences related to *Methylobacterium nodulans*, a bacterium able to grow using one carbon compounds and reducing nitrate to nitrite, and *Polaromonas naphthalenivorans,* a facultative chemolithotroph found in polluted habitats. These two bacteria partially substituted obligate autotrophs during heterotrophic conditions. The *nirS*-containing community showed highest similarities when nitrate was used as an electron acceptor despite the addition of organic matter. *nirS* sequences similar to those found in members of the family *Rodocyclaceae* were the most abundant. *Rodocyclaceae* have been found as the dominant bacterial population in industrial WWTPs. OTU 6, with a high similarity to *Rubrivivax gelatinosus nirS* gene, accumulated during heterotrophic conditions. *Rubrivivax gelatinosus* has been described as an obligate nitrite reducer able to use different carbon sources. In contrast, when nitrite was used as the first electron acceptor, *nirS* sequences similar to those found in Gammaproteobacteria, in particular *Rhodanobacter* sp., were the most abundant. A recent analysis of complete genome sequences of six *Rhodanobacter* strains isolated from soils have revealed that at least three of them lack the ability to reduce nitrate. Similarly to what has been observed with the *nirS* gene, bacteria enriched when nitrite was used suggest the exclusive use of nitrite as electron acceptor. This is the case for OTU 2 (85% similar to *nirK* sequence of *Rhodopseudomonas palustris*). *Rhodopseudomonas palustris* lacks an ortholog of a dissimilatory nitrate reductase in its genome, suggesting that nitrate reduction cannot be done in this bacterium. The addition of nitrite as initial electron acceptor impacted the composition of *nirS*- and *nirK-*type denitrifiers in the MFC, and possibly caused an enrichment of selected obligate nitrite reducers in view of sequence similarities with the detected functional genes. Contrasting to the previous genes, the *nosZ*-containing community remained almost invariable during all conditions. Sequences with a high similarity to the *nosZ* gene of *Oligotropha carboxidovorans*, a carboxidotrophic bacterium, clearly dominated the *nosZ* community. The presence of *Oligotropha* like *nosZ* sequences has also been detected as major components of the nitrous oxide reducing communities in samples of acidic peat soils and in the cathode of a denitrifying MFC. Moreover, gene abundances during autotrophic conditions supported the idea of *nosZ* community minimally affected by the initial electron acceptor. ## Denitrifiers Affect N₂O Accumulation in the MFC Denitrification is considered the main source of N₂O accumulation at a global scale. However, no consensus exists whether to consider either the *nirS* or the *nirK* community as the main responsible for N₂O accumulation, since the abundance of these two type denitrifiers can vary significantly from environment to environment. In the cathode of the MFC, NirS- type denitrifiers outnumbered NirK-type denitrifiers by two orders of magnitude at all working conditions. High q*nirS*/16S rRNA values were observed, suggesting a clear implication of the *nirS*-type denitrifiers in the denitrification potential and the accumulation of N₂O. Different alternatives were considered to explain the high q*nirS/*q16S rRNA ratio found. First, the presence of multiple copies of *nirS* gene in a single genome, *i.e*. *Thauera* sp., *Thiobacillus denitrificans*, *Dechloromonas aromatica* or *Magnetospirillum magneticum*, was evaluated. Second, an overestimation of the *nirS* abundance due to *q*PCR bias was considered. However, this possibility was excluded after the examination of dissociation curves and the cloning of random *q*PCR products, which led us to confirm the specificity of the reaction. The observed prevalence of *nirS* over *nirK* denitrifiers may be the result of a selective enrichment of the former due to a putative enhanced capacity of electron harvesting by cytochrome *c* family mediators. This feature has been confirmed in *Geobacter sulfurreducens* ATCC 51573 by the analysis of a GSU3274 deletion mutant. GSU3274 is a gene coding for a putative cytochrome *c* family protein. Despite the occurrence of some limitations, such as the presence of multiple gene copies per genome and differences in specific activity, *q*PCR analyses of functional genes provide significant data to infer community dynamics. The ratio between the abundance of nitrite reductases and nitrous oxide reductase allowed us to estimate the potential to reduce completely nitrite to N₂. In the MFC and at the working conditions used in this study, the estimated N₂O accumulation significantly correlated (r² = 0.992) with the (q*nirK*+q*nirS*)/q*nosZ* ratio. Higher accumulations of N₂O were observed when nitrite was used as the electron acceptor, similarly to what has been described in other environments. In conclusion, the cathodic biofilm of the MFC was dominated by *nirS*-type denitrifiers at all conditions tested and its abundance relative to nitrous oxide reducers highly correlated with N₂O emissions. The denitrifying bacterial communities identified affected the electrochemical performance increasing the current density for about 25% in autotrophic conditions. Also the suspected relevant players in nitrate and nitrite reduction have been identified on the basis of functional gene similarities. Their relative dominance at each period was highly affected by the changes of the electron acceptor or electron donors. Contrarily, the *nosZ* community remained almost invariable during all periods tested. Obtained *nosZ* sequences showed a high similarity to nosZ gene of *Oligotropha carboxidovorans*, suggesting that may have an active and preponderant role in electron harvesting in the cathode surface. This may raise new questions, such as which mechanisms are involved in electron transfer and what the location of *O. carboxidovorans*-like bacteria in the biofilm is, that will be investigated in the near future. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JC LB SP MDB. Performed the experiments: AVS RT SP. Analyzed the data: SP AVS RT AGL. Contributed reagents/materials/analysis tools: JC LB. Wrote the paper: AVS SP LB JC.

# Introduction Video endoscopes, such as laparoscopes, gastroenteroscopes, and bronchoscopes, are used in many fields of medicine to help clinicians diagnose and treat patients less invasively with shorter hospital stays. However, video endoscopes are difficult to operate, and it takes time for clinicians to learn how to interpret endoscopic images. One of the difficulties is the discrepancy between the spatial orientation of the endoscopic images and the endoscopist’s working environment. For example, it has been reported that a surgeon’s performance decreases when the optical axis of the endoscope equipment does not match the direction of gravity. Attempts to solve this problem have included pattern recognition, electromagnetic field sensors, and accelerometers. With the pattern recognition method, various computer algorithms were used to extract the features from fiducial markers or anatomical structures on the endoscopic image; however, this method cannot be used when there is no target in the line of sight. The electromagnetic method uses a current induced by a coil moving through a magnetic field. This technique is widely used as a tracking system for image-guided interventions, but it can also be used for image orientation correction. However, conductive surgical instruments or electromagnetic interference, from x-ray fluoroscopy for example, can distort the magnetic field and influence the measurements. The accelerometer method uses a low-cost, small inertial sensor to correct the direction of endoscopic images; however, it cannot distinguish between translational acceleration and gravity. Therefore, an accelerometer should be used in combination with a gyrosensor to correct the direction of a rapidly rotating endoscope. The gyrosensor has a fast response time that is not influenced by translational acceleration. Because the accelerometer fails to differentiate acceleration from gravity, the gyrosensor measures the rotational angle more accurately than the accelerometer when the endoscope rotates. Moreover, the integration error of the rotational angle from the gyrosensor that accumulates over time can be easily corrected using an accelerometer signal. Therefore, these two sensors are often used to complement each other. In this study, we developed and evaluated a real-time system for correcting endoscopic image orientation using an accelerometer and gyrosensor. # Materials and methods ## Development of the system To obtain the direction of gravity and angular velocity, an accelerometer and gyrosensor were attached to the handle of the endoscope. We assumed that the endoscope handle was a rigid body and that there would be no twisting from the handle to the tip of the endoscope. Quaternion, direction cosine matrix, Euler angles, and axis-angle representation can be used to signify the posture of a rigid body. Due to its simplicity and clear physical meaning, we used the axis- angle to represent the posture of the endoscope handle. The axis was defined as a unit vector indicating the direction of the advancing endoscope. The rotational angle θ was defined as 0 when the sensor was in the opposite direction of gravity and increased when the endoscope rotated in the clockwise direction. Because a singularity occurs when the operator holds the endoscope handle perpendicular to the ground, this action was prohibited during the study. ## Calculation of the rotational angle The sensor was attached so that the *x* axis pointed to the right as seen by the operator, the *y* axis pointed toward the head of the endoscope, and the *z* axis point in the direction of the advancing endoscope. The rotational angle with respect to the direction of gravity was calculated as: $$\theta_{accel} = atan2\left( {a_{x}, - a_{y}} \right)$$ $$\theta_{zyro} = {\int r_{z}}$$ where *θ*_accel and *θ*_zyro are the rotational angles from the accelerometer and gyrosensor, respectively; *a*_*x*,*y* is the acceleration from the accelerometer, *r*_*z* is the angular velocity from the gyrosensor, and atan2 (*y*,*x*) is a function that returns the arc tangent of *y*/*x* within the range of −*π* and *π* depending upon the sign of *x* and *y*. The rotational angles from the accelerometer and gyrosensor were calibrated using a Kalman filter. The filter used the output of the gyrosensor and. $$Model\mspace{180mu} equation:\mspace{180mu}\theta_{Kalman,k} = \theta_{Kalman,k - 1} + \left( {T \times r_{k}} \right) + w_{k}$$ where *θ* _Kalman is a state variable–the rotational angle after the Kalman filter (*θ* _Kalman,0 = *θ* _accel,0), *r*_k is the angular velocity from the gyrosensor, *T* is the time from the last measurement, *w*_k is the processing error, and *θ* _accel is the rotational angle from the accelerometer. Eq represents the measurement equation of the Kalman filter using the accelerometer output. $$Measurement\mspace{180mu} equation:\theta_{accel,k} = \theta_{Kalman,k} + z_{k}$$ where *θ* _accel is a measurement variable–the rotational angle from the accelerometer, *θ* _Kalman,k is a state variable, and *z*_k is the measurement error. In the Kalman filter algorithm, w_k and z_k were assumed to be zero mean Gaussian white noise with a covariance of Q_k and R_k. The smaller the Q_k value, the greater the importance of the gyrosensor output; the smaller the R_k value, the greater the importance of the accelerometer output. We set the R_k value to give less weight to the accelerometer output if an additional acceleration was observed. $$Q_{k} = 0.1$$ $$\left. R_{k} = 500,\mspace{180mu} when \middle| a_{k} \middle| < 1.0001g \right.$$ $$\left. R_{k} = 50,000,\mspace{180mu} when\mspace{180mu} 1.0001g < \middle| a_{k} \middle| < 1.02g \right.$$ $$\left. R_{k} = 1,000,000,\mspace{180mu} when\mspace{180mu} 1.02g < \middle| a_{k} \right|$$ where *a*_k is the acceleration from the accelerometer and *g* is the acceleration of gravity. The predicted rotational angle was calculated using $${\hat{\theta}}_{\mspace{180mu} Kalman,k} = {\hat{\theta}}_{\mspace{180mu} Kalman,k - 1} + \left( {T \times r_{k}} \right)$$ $$P_{k} = P_{k - 1} + Q_{k}$$ where $\hat{\theta}$ _Kalman,k is the estimated rotational angle from the Kalman filter, r_k is the angular velocity from the gyrosensor, and P_k is the estimated covariance (P₀ = 1). Finally, the states and the estimated covariance were updated for the next iteration as shown in. $$\left. K_{k} = {P_{k}/\left( {P_{k} + R_{k}} \right.} \right)$$ $${\hat{\theta}\mspace{180mu}}_{Kalman,k} = \left( {1 - K_{k}} \right) \times {\hat{\theta}\mspace{180mu}}_{Kalman,k} + K_{k} \times \theta_{accel,k}$$ $$P_{k} = \left( {1 - K_{k}} \right) \times P_{k}$$ where K_k is the Kalman gain. Subsequently, the rotational angle from the Kalman filter was postprocessed to correct for the delay between the image acquisition and sensor measurement. Because the imaging device and the sensor have different processing and data transfer times, when the image arrives at the processor, the last measurement data from the sensor should not be used without correction. If we assume that the time difference between the last available sensor output and image frame does not change during the exam, the correction should be as shown in : $$\theta_{postproc,k} = \theta_{Kalman,k} - \left( T_{d} \times r_{k} \right)$$ where *θ* _postproc,k is the final rotational angle after postprocessing, r_k is the angular velocity from the gyrosensor, and *T*_d is the time difference between the last available sensor output and image frame. *T*_d was determined by a least square error method using the pre-examination data from the system with. $$T_{d} = \left( \mathbf{r}^{T}\mathbf{r} \right)^{- 1}\mathbf{r}^{T}\left( \theta{}_{\mathbf{K}\mathbf{a}\mathbf{l}\mathbf{m}\mathbf{a}\mathbf{n}} - \theta_{\mathbf{o}\mathbf{p}\mathbf{t}\mathbf{i}\mathbf{c}\mathbf{a}\mathbf{l}} \right)$$ where **r** is the vector of the measured angular velocity, ***θ*** _**Kalman** is the rotational angle vector from the Kalman filter, and ***θ*** _**optical** is the optical rotational angle vector (method will be described in the Technical evaluation section). ## Devices used for the system A 3-Space Sensor™ Bluetooth (Yost Engineering Inc, Portsmouth, OH, USA) was used for the system. According to the manufacturer’s specifications, the accelerometer and gyrosensor precisions are 0.0024 m/s² and 0.070°/s, respectively. The sensor device also has a magnetometer, although it was not used in this study. The video was captured by a charge-coupled device (CCD) camera attached to the eyepiece of the endoscope and transferred to a PC at 30 frames per second using a USB-ECPT video capture device (Sabrent, Los Angeles, CA, USA). The image correction software was developed in C++. The sensor data were acquired by a polling method. ## Technical evaluation To evaluate the performance of the image orientation correction system, a board with a black upper half and a white lower half was placed 30 cm from the PC camera (S101 Web Cam; Kodak, Rochester, NY, USA), and the sensors were attached to the camera. The rotational angle (*θ*_optical) was automatically calculated from the image by recognizing the black and white boundaries. While the camera was rotated randomly for 30 seconds, the rotational angles from the accelerometer (*θ*_accel), gyrosensor (*θ*_gyro), and Kalman filter (*θ*_Kalman), the postprocessed rotational angle (*θ*_postproc), and the optical rotational angle (*θ*_optical) were recorded. Every measurement was repeated using the sensor’s wired and wireless connection. ## Clinical utility test A prospective, randomized observational study using an airway model was used to test the clinical utility of the image orientation correction system. Because the experiment was performed on the airway model and the participants were enrolled as testers, written informed consent was waived. The model simulated human tracheal rings and the angles of the carina and right upper bronchus. Fifteen residents from the Department of Anesthesiology and Pain Medicine at Seoul National University Hospital, each of whom had performed more than 30 bronchoscopic exams, participated in this study. Before the experiment, 5 minutes were allotted for each participant to learn how to use the equipment and practicing with the device. For the experiment, each participant performed two simulated endoscopic exams using two airway tree models in a random order, one with and one without the image orientation correction system. The order of exams was determined by a computer-generated random number. A wired connection was used between the sensor and the computer during the exam with the image orientation correction system. During each exam, participants verbally reported the printed numbers in the order of left main, right main, and right upper bronchi of the airway model. The correctness and the total time taken to report the numbers were recorded. The numbers inside the airway model were reset for each exam. For a statistical power of 0.8 and a significance level of 0.05, the estimated sample size was 12, with the assumption that the mean difference and standard deviation between two measurements was 20 seconds and 25 seconds, respectively. Fifteen participants were needed to compensate for possible dropouts. The times required to finish the exam with and without using the image orientation correction system were compared using the Mann-Whitney *U* test. All statistical analyses were performed using SPSS software (version 21; SPSS, Inc., Chicago, IL, USA). *P* \<.05 was considered statistically significant. Data are expressed as median (interquartile). # Results ## Technical evaluation The optical rotational angle measured from the captured image during the random rotation of the camera ranged from −83.1° to 88.0° with the wired sensor connection and from −108.0° to 90.9° with the wireless connection. The optical angular velocity of the camera ranged from −184°/s to 182°/s with the wired sensor connection and from −251°/s to 201°/s with the wireless connection. The mean sampling rate of the sensor was 333.4 Hz and 27.7 Hz for the wired and wireless connection, respectively. Using only the accelerometer output, the maximum difference between *θ*_optical and *θ*_accel was 12.79° with the wired connection and 15.08° with the wireless connection. Using only the gyrosensor output, the maximum difference between *θ*_optical and *θ*_gyrol was 8.87° with the wired connection and 14.18° with the wireless connection. After the Kalman filter and postprocessing, the maximum difference between *θ*_optical and *θ*_postproc was 5.00° with the wired connection and 7.48° with the wireless connection. As shown in, the minimum root mean square (RMS) error between the calculated and measured optical rotational angles was obtained using the wired connection for both the accelerometer and the gyrosensor with Kalman filtering and postprocessing for the time delay. The wired connection provided better results than the wireless connection for both sensors in most cases. The time differences between the last available sensor output and image frame measured by were 0.025 and 0.002 seconds for the wired and wireless connection, respectively. ## Clinical utility test All participants successfully completed the experiments. The time required to finish the bronchoscopic exam decreased significantly using the image orientation correction system (median, 52 seconds; interquartile range, 32–74 seconds) compared with not using the system (median, 76 seconds; interquartile range, 59–128 seconds; *P* =.012). There was one case when the image orientation correction system was not used in which the number printed at the opposite bronchus was reported. # Discussion In the present study, we developed and evaluated an endoscopic image orientation correction system using an accelerometer and gyrosensor. As a result, there was a clear benefit from using the orientation correction system. The direction of a flexible endoscope needs to be changed by rotation, because the tip can only be bent up or down. During ultrasonography, the probe should be positioned obliquely in order to transmit the beam to the target. These changes cause unwanted rotation of the image orientation. By correcting the image orientation relative to the direction of gravity, the image can be interpreted more intuitively. Video clips of the airway model bronchoscopic exam with and without image orientation correction are provided as supplementary files ( and Video Files). The orientation corrected image was much easier to interpret because the direction of the image did not change during the exam. To correct the endoscopic image orientation, some researchers have reported using an accelerometer alone. Although the accuracy of the device in that report was within 1°, this result was achieved only in a stationary state. When using the accelerometer alone in our study, the error was much greater than 1° while the endoscope was rotated. This error decreased significantly when the gyrosensor and accelerometer were used simultaneously. Use of accelerometer and gyrosensor in endoscopic guidance have been reported already. Behrens and colleagues used accelerometer and gyrosensor for endoscope navigating system. In the report, the mean error of the position measurement was between 1 and 4 degrees. Ren, et al. reported that accelerometer and gyrosensor can improve the accuracy of existing electromagnetic tracking system. However, relatively little effort has been made to evaluate the clinical usefulness of this technology. A novel postprocessing method was also suggested in this study, wherein the time difference between the last available sensor output and image frame multiplied by the angular velocity was subtracted from the Kalman filter output. In the devices used in this study, the time difference between the last available sensor output and image frame was 0.025 and 0.002 seconds with the wired and wireless connection, respectively. Not compensating for this time difference can result in large error, especially with the wired connection. The possibility of postprocessing using the angular velocity is another benefit of using the gyrosensor. The main advantage of the system developed in this study is that the sensor module can be attached to the handle of the endoscope. While this can cause an error when the tip and handle of a flexible endoscope twist, this is not common in most clinical situations. Also, sensors attached to the tip may increase the tip size and the risk of aspiration when the equipment is damaged. Additionally, attaching the sensor to a conventional endoscope handle is more cost-effective than purchasing new equipment. This endoscopic image orientation correction system can also help clinicians to objectively analyze the endoscopic images after the procedure. This is particularly useful when the endoscopist and the operator are not the same person (e.g., laparoscopic surgery or transesophageal echocardiography-guided valvuloplasty), when there is no information about the image orientation (e.g., intravascular sonography or epiduroscopy), or when communicating with another specialist for consultation or education. This study does have some limitations. A camera was used instead of an endoscope during the technical evaluation phase. However, measuring the optical rotational angle from with an actual endoscope may have induced an error due to the lower resolution. Second, an error may have occurred during the technical evaluation if the center of the camera did not align exactly with the center of the target board. Third, an airway model was used instead of an animal or human airway for the clinical utility test. During a bronchoscopic exam of a human patient, the shape of tracheal rings and muscle stripes can help orient the operator; however, the airway model used in this study did not have such detailed structures. In conclusion, the endoscopic image orientation correction system using both an accelerometer and a gyroscope resulted in greater accuracy than using an accelerometer alone. This system also significantly decreased the time needed to perform a bronchoscopic exam, which would be quite valuable in the clinical setting. Although these results were obtained in a limited situation, this orientation correction system will likely help clinicians interpret and analyze endoscopic images in clinical practice. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Alzheimer’s disease (AD) often results in communicative breakdowns that negatively impact the person with AD (pwAD) and caregivers, including by discouraging interactions that could help to preserve dignity and maintain relationships. These breakdowns appear at least partially attributable to declining amounts of informative content in speech produced by pwAD. In connected speech, pwAD have been found to produce fewer words than controls, with total output decreasing as the disease progresses. In the speech they do produce, pwAD describe events less accurately, producing fewer information units than controls and thus omitting relevant information. These declines in informative content have resulted in characterizations of spoken discourse by pwAD as vague or empty. Knowledge of how specific lexical-semantic changes affect the informative content of speech produced by pwAD may help improve diagnosis and monitoring of the disease. An improved understanding of the nature of communicative breakdowns can also facilitate interventions that improve communication between pwAD and their caregivers. Consensus findings from studies of single-word tasks suggest impaired production of both verbs and nouns by pwAD compared to controls. Within groups, pwAD tend to be less accurate with verbs than nouns. Impairments in discrete word production reflect effects of connectedness in the semantic network and contributing psycholinguistic properties. Word frequency and age of acquisition (AoA) are believed to influence semantic network development, and pwAD perform less accurately on naming tasks requiring retrieval of words that are less frequent or are acquired later in life. Better performance with nouns has been attributed to stronger relationships within this word class; however, psycholinguistic effects are unclear, as verbs tend to be more frequent than nouns, but are acquired later in life. Lexical-semantic changes are also present in discourse produced by pwAD. Consensus findings indicate that pwAD produce more verbs and fewer nouns than controls in both picture descriptions and spontaneous speech \[, –\]. Declines in noun production are accompanied by increased reliance on pronouns, which function as less specific noun substitutes. Findings from picture descriptions suggest that pwAD also use a less diverse range of words than controls in discourse. These changes are accompanied by psycholinguistic differences in the groups’ speech samples. In picture descriptions and story retellings, pwAD rely on simple, generic words including copulas and verbs and nouns of higher frequencies than those used by controls. Interestingly, findings from picture description tasks have not indicated that pwAD rely on words of lower AoA. Fraser et al. included AoA in their investigation of language features that might help distinguish pwAD from healthy older people but did not find it to contribute to identification of pwAD. Yeung et al., meanwhile, actually found pwAD to use words of higher AoA than controls, a finding they attributed to a tendency by pwAD to produce utterances that were inappropriate to the task. However, consensus findings on word use have led to suggestions that breakdowns in semantic representations of “advanced,” or less well-connected, verbs and nouns in AD lead to their replacement with easily accessible, more well- connected, but less specific alternatives. This process would result in frequent reuse of less advanced verbs and nouns and increased use of pronouns, suggesting that changes in POS reliance, lexical diversity, and psycholinguistic properties of words produced by pwAD are related. In everyday communication, this combination of changes would contribute to perceptions of speech by pwAD as uninformative, inaccurate, or vague. However, little information is available on lexical diversity or psycholinguistic properties of words produced by pwAD in spontaneous speech. Where reported, these findings have rarely been broken down by POS. Findings on lexical diversity include just one report of reduced overall diversity (*n* = 8 pwAD) and one of reduced diversity among both nouns and verbs (*n* = 10 pwAD). Despite the findings of word frequency and AoA effects on performance in single-word tasks, these measures have not been considered in analyses of spontaneous speech, either overall or by POS. Breakdowns of diversity, frequency, and AoA of nouns and verbs produced in spontaneous speech would reveal whether the differential effects seen in discrete word production manifest in communicative situations. This knowledge would be useful in the design of targeted interventions. POS breakdowns would also help determine the extent to which overall changes in diversity, frequency, and AoA are attributable to changes in POS quantities. This is of interest especially considering the likely overuse by pwAD of a limited number of high-frequency, low-AoA pronouns. The present study therefore aims to investigate whether pwAD differ from controls in the use of words of different POS in conversational speech. Five hypotheses are investigated. - H1 predicts that pwAD will use significantly fewer nouns and significantly more verbs and pronouns than controls, with these changes resulting in significantly lower N/V ratios. - H2 predicts that pwAD will exhibit significantly decreased lexical diversity compared to controls on three measures—overall, for nouns, and for verbs. - H3 predicts that pwAD will produce significantly more copulas than controls. - H4 predicts that pwAD will use nouns and verbs of significantly higher frequencies than controls. - H5 predicts that pwAD will use nouns and verbs of significantly lower AoAs than controls. # Methods ## Dataset Language data used in this study come from the Carolinas Conversations Collection (CCC), a digital archive of recorded interviews with American men and women aged 65 and older. This can be found at <https://carolinaconversations.musc.edu/ccc/about/>. Data collection was approved by institutional review boards at the University of North Carolina at Charlotte and the Medical University of South Carolina. Participants who were able to write provided written informed consent to recording, as did guardians of all participants. Each participant also provided oral assent to recording. The CCC includes interviews with participants diagnosed with dementia by a physician. Prior to interviewing participants, researchers agreed not to conduct further cognitive testing; however, pwAD are uniformly described as being in moderate to late disease stages. The CCC also includes interviews with participants who were screened to rule out dementia. Conversations revolve around daily life and health. Interviewers disclose their own typical daily activities before asking interviewees to do the same. Where an interviewee refers to a health condition, the interviewer responds with a set of semi- structured questions. Recorded conversations were transcribed by trained medical transcribers according to a protocol established in conjunction with researchers. ## Participants and language samples Interviewees were considered for inclusion in our AD group only if they met the following criteria: native English speaker; dementia specified as AD; dates of birth and conversation available; sex- and age-matched control available. These criteria resulted in 16 potential participants with AD. For each, the earliest dated intelligible audio-recorded conversation was extracted, edited, and coded according to CHAT guidelines. This process resulted in speech samples that varied in length from 224 to 1,899 words. Sajjadi et al., noting uncertainty around the word count necessary for a connected speech sample to realistically reflect language production, found 150- and 600-word transcripts comparable for analyses of connected speech by pwAD. However, longer texts are preferred when using lexical diversity to draw conclusions on a speaker’s vocabulary. As such, four transcripts with fewer than 500 words were excluded from analyses. The twelve remaining transcripts were cut to the end of the utterance containing the participant’s 560^th word, matching the length of the shortest transcript. Interviewees were considered for inclusion in our control group only if they met the following criteria: native English speaker; not reported to have AD or other neurological or psychiatric condition; dates of birth and conversation available. A control group was selected to match the group of pwAD in number, sex, and age. Groups were not matched for education because the CCC provides this only in broad ranges that do not facilitate meaningful matching—e.g., 1–8 years, 9–12 years, etc. Conversations for the control group were extracted, edited, coded, and shortened using the above process. ## Language analysis Transcripts were analysed using CLAN, software designed to analyse transcripts conforming to CHAT guidelines. Initial POS tags generated in CLAN were reviewed and inaccuracies that would affect analyses were corrected—for example, *hog* tagged as a verb when in context it had been used as a noun. H1-H3 are on noun, verb, and pronoun quantities and N/V ratios, lexical diversity overall and for nouns and verbs, and copula use. Noun, verb, and pronoun quantities are measured via token counts. CLAN’s default noun token counts were used in analyses. They include gerunds, \~*ing* verb forms functioning contextually as nouns and tagged as such in CLAN. Noun token counts do not include pronouns. For consistency with CLAN’s default N/V ratio calculations, verb token counts include lexical verbs, copulas, and participles. Inclusion of copulas and participles in these counts requires a modification to CLAN code. CLAN’s default pronoun token counts and N/V ratios were used in analyses. Lexical diversity of overall speech samples, of nouns, and of verbs are measured via type-token ratios (TTRs). TTR comparisons may be affected by differences in sample size—samples with more words may have lower TTRs due to increased likelihood of word re-use. Speech samples in this study are matched for overall token count. However, they vary naturally in noun and verb counts. CLAN offers two alternative lexical diversity measures: moving average type- token ratio and VocD. Due to limitations in CLAN functionality for deriving these measures by POS, in this study, noun and verb diversity are measured via TTRs. For consistency and because overall token counts are matched, overall lexical diversity is also measured via TTRs. Overall TTRs consider all words in a speech sample. Noun TTRs consider all nouns and verb TTRs consider all verbs included in token counts. All TTRs used here are lemma-based, so that, e.g., *brother* and *brothers* are counted as two occurrences of the same word rather than as different words. Generation of lemma-based TTRs requires a modification to CLAN code. Copula production is measured via copula counts and ratios. CLAN’s default copula, lexical verb, and participle counts were used to calculate copula-to-verb ratios as copula / (copula + lexical verb + participle). H4 and H5 are on noun and verb frequency and AoA. Wordlists were generated in CLAN and used to obtain frequency and AoA measures from external databases. Wordlists included all tokens appearing in token counts except for gerunds, which were excluded due to discrepancies between the word form’s POS in speech samples and the lemma’s POS in databases. Word frequencies were obtained from WebCelex, the web-based interface to the CELEX lexical database. Frequencies are reported as lemma appearances per million words. Regarding AoA data, few datasets account for distinctions between multiple POSs of a given word form, for example *walk* used as a verb or a noun. Viably scaled datasets that do this tend to be small. This study reports group-level statistics based on mean AoAs extracted from the 30,121-word dataset of Kuperman et al., which does not distinguish word forms by POS. Kuperman et al. asked participants to estimate AoA as the age at which the participant believed they had learned the word. Ratings range from 1.5 to 25 years old and are rounded to two decimals. This study also reports advanced statistics based on ratings from the 2,694-word dataset of Bird et al., the largest viably scaled AoA dataset to account for POS \[cf. ; see also Section 4.1.2\]. Bird et al. asked participants to estimate AoA on a 7-point Likert scale where each point corresponded to a period of two years including the age at which the participant believed they had learned the word (e.g., a rating of one indicating an age between 0 and 2 years old), with a rating of seven for any age over 13. Those authors then multiplied mean ratings by 100, so that final ratings are between 100 for a low AoA and 700 for a high AoA. ## Statistical analysis The statistical software environment R was used for all statistical analyses. H1-H3 were addressed using two-tailed independent samples *t*-tests that compared group mean noun, verb, and pronoun token counts, N/V ratios, TTRs overall and for nouns and verbs, and copula counts and ratios. Data used in these analyses met test assumptions. Alphas for *t*-tests were not adjusted for multiple comparisons because data analysed resulted from observations of natural phenomena. Effect sizes are reported in Cohen’s *d*. Due to effects of sample length on TTR (see Section 2.3), reporting on TTR comparisons considers results of token count comparisons. H4 and H5 were addressed using linear mixed effects models to test predictors of frequency and AoA of nouns and verbs used. Model structures were determined *a priori* based on predictors of interest and included fixed effects of Group, POS, and Age. The continuous variable Age was centred but not scaled. Sex was not included as a fixed effect because there were only two males in each group. Participant and Word were included in models as nested random effects, with frequency values or AoA ratings for a given word appearing as many times as the word appeared in a participant’s speech sample. Because plots of preliminary models did not meet statistical assumptions, dependent variables were converted to natural logarithms for analysis. Group- level comparisons of overall, noun, and verb frequencies and AoAs are provided for descriptive purposes. # Results ## POS production Measures of POS production are presented in. pwAD produced significantly fewer noun tokens than controls (*p* \< 0.01). This result was associated with a very large effect size (*d* = 1.31). Group differences in pronoun use did not reach significance (*p* = 0.08). However, a strong inverse relationship was present between noun and pronoun production by pwAD (*r* = -0.66, *p* = 0.02). The groups produced similar verb token counts and N/V ratios (both *p* \> 0.05). All nonsignificant comparisons of POS production were associated with medium effect sizes. ## Lexical diversity TTRs of pwAD and controls are presented in. Overall TTRs were significantly lower for pwAD (*p* = 0.02), indicating that they used a narrower range of words across all parts of speech than controls. This result was associated with a large effect size (*d* = 1.02). Given the expected inverse relationship between token count and TTR, production of significantly fewer nouns by pwAD (see Sections 2.3 and 3.1) should be accompanied by higher noun TTRs than controls. However, group differences in noun TTRs were not significant (*p* = 0.47), and pwAD produced lower mean noun TTRs than controls. Group differences in verb TTRs were not significant. Both nonsignificant TTR comparisons were associated with small effect sizes. ## Copula production Two copula production measures were analysed—copula counts and copula-to-verb ratios. pwAD produced an average of 23.9 copulas (*SD* = 7.8, range = 12–38) compared with controls’ average of 22.6 (*SD* = 5.8, range = 10–32). Group differences were not significant for this measure (*t* = 0.48, *p* = 0.64, *d* = 0.19). pwAD produced an average copula-to-verb ratio of 0.25 (*SD* = 0.08, range = 0.15–0.4) compared with controls’ average of 0.23 (*SD* = 0.07, range = 0.1–0.31). Group differences were not significant for this measure (*t* = 0.75, *p* = 0.46, *d* = 0.31). These nonsignificant comparisons were associated with small effect sizes. ## Noun and verb frequencies Group-level statistics on the frequencies of words produced are presented in. While pwAD exhibited a tendency to produce words of higher frequencies than controls, two-tailed independent samples *t*-tests indicated that group means did not differ significantly for overall words, nouns, or verbs (*p* \> 0.05). Frequency data were available for 3,664 of 3,716 nouns and verbs spoken by participants. These included 562 of 587 nouns and 1,139 of 1,141 verbs spoken by pwAD (mean frequencies per million words: nouns 354 ± 484, verbs 11365 ± 15700) and 761 of 784 nouns and 1,202 of 1,204 verbs spoken by controls (nouns 310 ± 456, verbs 10635 ± 15175). Within both groups, frequencies were higher and variances greater for verbs than nouns. Qualitative inspection of the data indicated that several verbs, most notably *be* (38301 appearances per million words), were several times more frequent than both other verbs (next most frequent: *have*, 13494) and the most frequent nouns used within either group (pwAD: *one*, 2073; controls: *time*, 1971). Frequency values were converted to natural logarithms for analysis due to heteroscedasticity in the residuals of a preliminary model that used actual values. As seen in, a linear mixed effects model using log frequencies revealed a significant main effect of POS (*p* \< 0.01). Main effects of Group and Age were not significant. Significant two-way interactions were present between Group and POS (*p* \< 0.01) and Group and Age (*p* \< 0.03). The two-way interaction between POS and Age did not reach significance (*p* = 0.07). The three-way interaction between Group, POS, and Age was significant (*p* \< 0.01). Model predictions are presented visually in. Changes with age in frequencies of nouns used are apparent in both groups. Noun frequencies rise with age for controls, indicating use of less advanced nouns. By contrast, noun frequencies decrease with age for pwAD, suggesting use of more advanced nouns. Verb frequencies remain relatively stable with age in both groups, increasing slightly for pwAD, who use verbs of higher frequencies than controls. Participants in both groups use verbs of higher frequencies than their own nouns. Follow-up testing was conducted to explore the unexpected finding that pwAD used nouns of lower log frequencies with age, while controls used nouns of higher log frequencies. Of specific interest was how these trends might relate to noun token counts, as well as how noun token production might relate to age. As seen in, changes with age in noun token production were not significant. Production decreased moderately for controls while increasing moderately for pwAD. ## Noun and verb AoAs Prior to AoA analyses, a Pearson’s correlation was conducted to assess the relationship between frequency and AoA of all words used in this study. This relationship, while significant, was weak (*r* = -0.11, *p* \< 0.01). Group-level statistics on the AoAs of words produced are presented in. On average, pwAD produced nouns and verbs of higher mean AoAs—i.e., more advanced nouns and verbs—than controls. However, two-tailed independent samples *t*-tests indicated that group means did not differ significantly for overall words, nouns, or verbs (*p* \> 0.05). AoA data were available for 2,722 of 3,716 nouns and verbs spoken by participants. These included 276 of 587 nouns and 1,037 of 1,141 verbs spoken by pwAD (mean AoA: nouns 273 ± 72, verbs 280 ± 43) and 329 of 784 nouns and 1,080 of 1,204 verbs spoken by controls (nouns 268 ± 73, verbs 276 ± 43). AoAs were similar between groups for the respective word classes, with variances greater for nouns than verbs. AoA ratings were converted to natural logarithms for analysis due to a positive skew in the residuals of a preliminary model that used actual ratings. As seen in, a linear mixed effects model using log AoAs revealed significant main effects of Group (*p* \< 0.01) and POS (*p* = 0.01). The interaction between these terms was also significant (*p* \< 0.01). The main effect of Age was not significant, and no interactions including the Age term were significant. Model predictions are presented visually in. Between groups, pwAD use nouns of higher AoA than controls, while the groups differ little in AoA of verbs used. Within groups, pwAD use nouns of higher AoA than their own verbs, while controls use verbs of higher AoA than their own nouns. Follow-up testing was conducted to explore the unexpected finding that pwAD used nouns of higher log AoAs than controls. Of specific interest was how this finding might relate to group differences in noun token counts. As such, relationships were examined between participants’ noun token counts and mean log AoAs of nouns used. As seen in, increased use of nouns was not significantly associated with decreases in log AoA of nouns (*p* = 0.08). # Discussion This study aimed to investigate whether pwAD differed from healthy age-matched controls in their use of words of different POS in conversational interviews. Speech samples of 12 pwAD and 12 controls were matched for overall word count and analysed for noun, verb, and pronoun quantities, N/V ratios, overall and noun and verb TTRs, copula counts and ratios, and noun and verb frequencies and AoAs. pwAD produced significantly fewer nouns than controls. They also produced lower overall TTRs, while noun TTR comparisons were not significant despite the group differences in noun production. Differences in pronoun production did not reach significance (*p* = 0.08). The groups did not significantly differ in verb token counts, N/V ratios, verb TTRs, or copula use. Frequency was predicted by a main effect of POS, two-way interactions between group and POS and group and age, and a three-way interaction between group, POS, and age. AoA was predicted by main effects of group and POS and a two-way interaction between group and POS. Findings on nouns and verbs and their implications for research on language in ageing and AD are discussed in Section 4.1. Implications of these findings for communicative interventions are discussed in Section 4.2. Study limitations are discussed in Section 4.3. ## Noun and verb use by pwAD and controls 1. **H1-H3: Word production and lexical diversity.** The finding here of decreased overall lexical diversity in speech by pwAD, indicating use of a narrower range of words, is consistent with consensus findings. Evidence here indicates that this overall reduction is linked to changes in noun use. While direct comparison of noun TTRs suggested group differences were not significant, this result is likely attributable to differences in quantities of nouns produced, as TTRs are inversely related to token counts. pwAD used significantly fewer nouns than controls, which should have resulted in higher mean noun TTRs if the diversity among nouns used by the groups was similar. Instead, noun TTRs of pwAD were lower than those of controls despite their lower token counts, suggesting decreased diversity among nouns produced by pwAD. These findings of decreases in noun production and diversity are consistent with prior findings on noun use in discourse by pwAD. Use of fewer nouns and of a decreased range of words overall, specifically of nouns, by pwAD are likely related through a process by which retrieval issues lead to the replacement of nouns, a wide-ranging, open word class, with a necessarily limited number of words from the closed class pronouns. The inverse relationship here between noun and pronoun use by pwAD is evidence of this replacement process. Its effects on overall lexical diversity are suggested in the combination of decreases in overall TTR and noun token counts alongside nonsignificant increases in pronoun use compared to controls. Relationships between changes in noun and pronoun use and decreased overall lexical diversity in AD are also suggested in prior reporting of similar combinations of findings. Further studies should explore the role of retrieval issues in these changes. In contrast to the differences for nouns, pwAD here did not differ from controls in verb token production, verb TTRs, or copula use. These findings contradict prior indications that in discourse, pwAD produce higher proportions of verbs, particularly generic verbs such as copulas, and a less diverse range of verbs than cognitively healthy controls. Despite group differences in noun production, N/V ratios did not differ by group, likely relating to a nonsignificant decline in verb production by pwAD. N/V ratios can be used to identify a retrieval issue for one of these two major word classes relative to the other \[e.g., \]. However, the implication of this N/V ratio comparison is that no such issue is present for these pwAD. This contrasts with the findings on nouns and pronouns, which appear to suggest noun retrieval difficulties. Direct comparisons of word use by POS—either quantities in standardised samples, as used here, or proportions or percentages of all words—likely provide the better indication of POS-specific retrieval in discourse. 1. **H4 and H5: Noun and verb frequencies and AoAs.** It was expected that pwAD would use words of higher frequencies and lower AoAs than controls. Lower word frequency and higher word AoA lead to difficulties for both the cognitively healthy and pwAD on discrete language tasks. This may relate to the role of these learning history variables in declarative memory and semantic connectedness. Past findings have suggested that this difficulty with more advanced words may translate to reliance by pwAD on simple, generic alternatives in discourse. These findings are frequently derived from simple comparisons of group means \[e.g., \]. Here, preliminary *t*-testing indicated no group differences in frequencies or AoAs of overall words, nouns, or verbs produced. Results of a linear mixed effects model also indicated little overall difference in frequencies of nouns or verbs used by group; however, age affected noun frequencies differently within each group. Frequencies of nouns used by pwAD decreased with age, suggesting use of more advanced nouns, while frequencies of nouns used by controls rose with age. A separate linear mixed effects model found that pwAD produced nouns of higher AoA than controls regardless of age, seemingly indicating consistent use of more advanced nouns. These differences in findings by statistical method highlight the benefits of using more robust tests. Use of mixed modelling here provided information on effects of not only age but also properties of individual words and their use within and across participants. Discussion below of psycholinguistic properties of words produced will focus on results of these models. With age, pwAD produced nouns of lower frequencies and exhibited nonsignificant increases in noun production. This combination of findings appears to suggest improving noun production with age in pwAD. Unfortunately, the CCC lacks demographic information relevant to interpretation of these findings. Participant education levels were only available in broad ranges, so that we were unable to meaningfully explore contributions of education to language production by pwAD. It is possible that older pwAD here were more educated than younger ones, contributing to use of nouns of lower frequencies. Alternatively, the group’s age range (68–94) and the ages of the youngest pwAD (68, 70) at time of conversation suggest the presence within the group of both late-onset AD and the rarer early-onset form (EOAD). EOAD is associated with more rapid cognitive decline and different language symptoms. It is possible that EOAD in younger pwAD contributed to these findings of age effects. However, date or age at diagnosis—information that would have been useful in determining the presence of EOAD—were unavailable. These findings highlight the need for future studies to examine effects of education on noun production in conversation by pwAD and to compare noun production by people with early- versus late-onset AD. The vagueness around interpretation of these findings reinforces the importance of accounting for relevant demographic variables during study design, recruitment, and data analysis. In the present study, cognitively healthy adults over age 70 used nouns of higher frequencies with age in spontaneous speech. A nonsignificant decrease in noun use with age was also present in this group. This combination of findings suggests possible changes in noun use in discourse production in advanced ageing. Healthy older adults are known to experience word retrieval difficulties on single-word tasks. However, it has been suggested that vocabulary growth with age compensates for retrieval difficulties on single-word and discourse tasks. Kavé et al. found that despite increasing noun retrieval difficulty on naming and fluency tasks, healthy speakers aged 20 to 85 produced less frequent nouns with age in picture descriptions. The authors attributed this to older speakers’ larger vocabularies. The present findings are based on production of less constrained discourse by participants older on average than the oldest 30 participants in that study. These findings necessitate larger studies of changes to noun use not across the lifespan, but specifically in advanced ageing. They also underscore the usefulness of detailed reporting on performance by controls, which is at times lacking in studies attributing language changes to AD. Our follow-up frequency analysis considered trends in noun use in both groups rather than exclusively probing the unexpected production of nouns of lower frequencies with age by pwAD. This revealed potential changes in healthy ageing that provided perspective on the apparent changes in pwAD. Future studies attributing language changes to AD should ensure they provide similar context by either including middle-aged comparison groups or analysing changes with age in healthy controls. A follow-up analysis suggests the unexpected finding that pwAD used nouns of higher AoA than controls should not be taken at face value, as this finding may relate to significant group differences in noun token production. An inverse relationship (*p* = 0.08) was present in these data between quantities of noun tokens produced by participants and mean log AoAs of those nouns. Positive skews were also present in AoAs both overall and for nouns prior to conversion for statistical analyses, suggesting heavy reliance on nouns of earlier AoAs. Such distributions are in line with Zipf’s observation that speakers generally prefer less advanced words. While Zipf’s observation was based on word frequencies, it also applies to AoA and, by way of these variables, to semantic connectivity. Thus, production of more noun tokens would likely lead to higher proportions of nouns acquired at earlier ages, potentially explaining the finding here that controls used nouns of lower AoA than pwAD. Low numbers of noun tokens in AoA analyses may have also factored into this finding. Fewer than half the nouns appearing for each group in frequency analyses appeared in AoA analyses (frequency vs. AoA nouns: pwAD 562 vs. 276; controls 761 vs. 329). This likely related to our use of the Bird et al. AoA dataset. While this may be the largest set of viably scaled AoA ratings to account for a word form’s POS (see below), the inclusion of only 2,694 words in that dataset is a limitation that must be considered in interpreting the present findings. Another relevant consideration is the lack of information on participant education—pwAD may have been more educated than controls, and this might have contributed to production of more advanced nouns. Results of linear mixed effects models and *t*-tests suggest that verb frequencies and AoAs differed little by group, with effects of age on frequency less pronounced for verbs than nouns. Together, then, findings from H1-H5 suggest unimpaired verb production in spontaneous speech by pwAD while providing mixed evidence of noun production deficits. These differences in findings by POS are underscored by significant main effects of POS in both frequency and AoA models. Past findings on discrete word production also suggest POS effects in performance by pwAD. However, in those tasks, pwAD have greater difficulty with verbs than nouns. This combination of findings suggests that an advantage for nouns on single-word tasks does not translate to discourse. Findings that pwAD are no more impaired in describing actions than people or objects in pictures may be seen as further evidence of this. Poorer retrieval of verbs than nouns in single-word tasks has been attributed to weaker semantic organizations among verbs. Discourse contexts may mitigate verb retrieval deficits through the necessary interactions between verbs and words around them. These interactions recruit syntactic and morphological processes, which are relatively preserved in pwAD. Task effects arising from differences in processing demands would likely relate further to the engagement of distinct neural areas depending on whether language is produced or comprehended in context or in isolation. Potential effects of task type on retrieval of nouns and verbs by pwAD would be best investigated via within-participant comparisons of performance on single-word and discourse tasks. Analyses of functional imaging would provide further information on neural correlates of language processing and effects on performance resulting from changes to specific neural areas in AD. Findings here on verbs are based on higher and more similar token counts than those on nouns, demonstrating that in addition to processing demands, verbs and nouns also differ in the extent to which they appear in discourse. Speakers’ heavier reliance on verbs, together with the larger number of nouns than verbs in the English lexicon, contribute to much higher mean frequencies for verbs than nouns. In the present data, mean frequencies of verbs were 32 times higher than those of nouns. Despite this, verbs are acquired later in life. POS differences in frequency and AoA have ramifications for interpretations of word use and lexical-semantic decline in AD. While we did not find group differences in overall word frequency, multiple discourse studies have reported higher overall frequencies for pwAD than controls. An example is Kavé and Dassa, who elicited picture descriptions by pwAD nearly twice as long as those of controls. Their groups did not differ in percentages of words that were verbs, suggesting more verbs appeared for pwAD in analyses. pwAD also relied more on pronouns and less on nouns. Thus the finding of higher frequencies among words used by pwAD may simply reflect group differences in numbers of nouns, verbs, and pronouns produced. POS differences also apply to overall AoA comparisons. These were not significant here despite higher noun AoA for pwAD than controls. Yeung et al., though, found pwAD to use words of higher AoA than controls in Cookie Theft picture descriptions. Because this task commonly elicits words of lower AoA, the authors interpreted the finding as a sign that pwAD were more likely to make off-topic remarks. However, the authors also reference rate of noun and verb phrases as an explanatory factor in word-finding difficulties for pwAD. While this is not elaborated on, group differences in noun and/or verb production may have contributed to the overall AoA finding. Accounting for this potential confound in studies of psycholinguistic effects in language production by pwAD would allow for more conclusive testing of whether pwAD replace advanced words of a given POS with less specific ones. Frequency and AoA differences between nouns and verbs are also relevant to the design of single-word studies. Stimuli in those studies are often matched for frequency and AoA, including at times across POS. Druks et al., for example, found both pwAD and controls to name objects faster and more accurately than actions. Object and action stimuli were matched for AoA, with frequencies only marginally higher for verbs. Since verbs tend to have much higher frequencies than nouns, a reduction in these differences may have resulted in use of more advanced verbs, factoring into relative success with object stimuli. A better practice may be to match stimuli according to word class norms, as in White- Devine et al., who found pwAD but not controls more accurate in naming objects. More broadly, differences highlighted here between nouns and verbs—in frequency, AoA, proportion of words spoken, and context-dependent processing demands—are suggestive of deeper differences between the word classes. Those differences should be accounted for as potential confounds in design stages of investigations of word-class specific impairments in pwAD. Conclusions should not be drawn based strictly on within-group comparisons of performance across word classes. POS-specific impairments in pwAD should be framed within the context of any differences, or lack thereof, by word class for controls. Differences were also present here across psycholinguistic variables. pwAD used nouns of similar frequencies but higher AoAs than controls. Age factored into predictions of noun frequencies but not AoAs. Nouns were acquired earlier than verbs despite being less frequent in general usage. Frequency and AoA have been claimed to reflect the same information, a claim supported by tendencies toward earlier acquisition of words of higher frequencies. However, the relationship between frequencies and AoAs of words used in this study was weak, and conclusions that could be drawn on noun use differ by measure. Sailor et al. also found differences in these measures, reporting that words produced on a semantic fluency task were of lower frequencies but lower AoAs than those produced on a letter fluency task. While these findings seem to indicate that the measures are not proxies for one another, they may result from differences in data quality. Frequency values, including those used here, are generally based on counts of a lemma’s appearances in a large corpus. Lemmas correspond to meanings rather than word forms, thereby providing more specific information to corpus users. AoA measures frequently involve ratings based on adults’ perceptions of their own or their child(ren)’s learning. While these methods allow for collection of large datasets, flaws related to subjectivity are inherent \[see \]. More objective data collection methods such as recording or testing children engender separate issues, including difficulty obtaining large amounts of data. Regardless of collection method, AoA data often do not distinguish between meanings of a word form, including multiple POS. This is problematic considering differences in AoA by POS. Brysbaert and Biemiller constructed a comprehensive set of acquisition norms for 44,000 word meanings, including multiple meanings of a word form, from which POS can be inferred. Construction of this dataset involved retrospective changes to data from the testing of US children beginning in grade 4, including the creation of a grade 2 norm. The authors provide a formula by which to derive an AoA from a word’s grade level of acquisition. While these norms attempt to address flaws in prior AoA datasets, the authors acknowledge crudeness in their scale, based as it is on grade, rather than age, of acquisition. The scale is divided into increments of two grades, or years, beginning at US grade level 2, or age seven. It therefore essentially provides a default AoA of seven for any word acquired by this age, at which children generally already possess a large vocabulary. The higher end of their scale is also problematic in that testing of adults was limited to students enrolled in thirteenth to sixteenth years of schooling, with words acquired after year 13 automatically assigned a grade level of 14. Among other potential uses, a large, objective, precisely scaled set of AoA norms that accounts for multiple word meanings would facilitate improved understanding of relationships between frequency and word acquisition. ## Implications for communicative interventions The use by pwAD of nearly twice as many verbs as nouns, together with potentially facilitative effects of context on verb production, necessitate research on communicative interventions targeting verbs. Such interventions have shown success in post-stroke aphasia rehabilitation. Improvements experienced by people with aphasia following verb treatments include greater generalization across word, sentence, and discourse levels than is seen with noun treatments. Improvements appear to be independent of both the underlying verb deficit and the nature of the therapy, though they may be restricted by co-occurring syntactic deficits. Aphasia researchers have recommended not treating verbs in isolation but instead focusing on their argument structures and associated nouns. Such an approach may play to relative strengths of pwAD. As a progressive condition, AD should be considered less conducive to rehabilitative treatments than post-stroke aphasia. However, prior studies on verb acquisition in AD suggest some potential for targeting verbs in language maintenance or restoration. These studies found pwAD able to acquire grammatical features of new verbs. They were less successful with argument structures. This finding, however, may have related to aspects of research methodology—the researchers viewed argument structure as a semantic property of a verb, contradicting the consensus view that it is a syntactic property. Given that syntax is relatively spared in pwAD, a syntax-based approach to acquisition of argument structure may facilitate production of both verbs and nouns in discourse. Study is needed of whether verb treatments for aphasia may, with or without modification, improve communication by pwAD. Decreased noun production by pwAD highlights the need to address noun use in communicative interventions, since these decreases have broad implications for pwAD and their caregivers. Nouns are content words that convey meaningful detail, specifying for example entities that perform or receive an action. Their replacement with less specific pronouns likely factors into perceptions of spoken discourse by pwAD as vague or empty. The replacement of open-class with closed-class words also results in decreased lexical diversity. Adults have been shown to judge children whose speech exhibits low lexical diversity as less appealing, mature, or talkative than those with higher lexical diversity. A question for future research is whether pwAD are judged negatively based on decreased lexical diversity. Such judgments might lead caregivers to engage less in conversation with pwAD, contributing to social isolation and declines in mental health and quality of life for both parties. Mental state and cognitive activity, including social interaction, can affect cognitive abilities, so that this process may speed declines experienced by pwAD. Word retrieval in context and caregiver responses to simplified language production are therefore potential intervention targets. ## Limitations This study has limitations of which the reader should be aware. Severity of cognitive decline in pwAD could only be considered at a general level. While a uniform statement is available on AD stages of CCC participants, information is not provided for individuals. Education level, a predictor of both word use and decline in AD, is provided in broad ranges only and so was not considered here. Due to small group sizes and specifically a lack of male participants, analyses were not controlled for gender, another predictor of decline in AD. Larger studies of the lexical-semantic measures investigated here should account for progression of cognitive decline, education, and gender. Speech samples were included based on a word count threshold and most were artificially shortened. While this practice facilitated statistical analyses, it ignored potential contributors to variation among speech samples, which may include cognitive ability. These data are not controlled for interviewer or conversation topic. While this has the benefit of more closely simulating real-life communication, these factors can influence the communicative participation and word choice of interviewees. Limitations related to existing AoA datasets are discussed in Section 4.1.2. Despite its limitations, this study has provided information on lexical-semantic changes that may accompany ageing and AD while highlighting methodological considerations for improved investigation of these changes. # Conclusion This study compared conversational use of words of different POS by pwAD and healthy age-matched controls. Previous findings have suggested that breakdowns in semantic representations in AD lead to frequent reuse of generic words, including pronouns, copulas, and other high-frequency nouns and verbs. In this study, pwAD produced fewer nouns and more pronouns than controls, leading to decreased lexical diversity. Frequencies of nouns produced did not differ by group, though age affected these differently. pwAD produced nouns of lower frequencies with age, possibly due to differences in education or in age of AD onset. Controls produced nouns of higher frequencies with age, which may suggest difficulty retrieving nouns in conversation in advanced ageing. pwAD used nouns of higher AoA than controls, a finding that may relate to group differences in noun quantities. Use of verbs differed little by group. Overall, aside from increased pronoun use, pwAD did not tend to produce less sophisticated words than controls. Across groups, POS significantly affected word frequency and AoA, with verbs having much higher frequency values than nouns. These findings highlight the importance of breaking down findings by POS when assessing language use. Future reporting on lexical diversity, frequency, and AoA in speech by pwAD should account for POS effects. These findings also provide rationale for targeting both nouns and verbs in communicative interventions. The Medical University of South Carolina (MUSC) and the University of North Carolina at Charlotte (UNCC) are thanked for providing access to Carolina Conversations Collection data. Professor Boyd Davis of UNCC and senior systems engineer Paul Arrington of MUSC are thanked for facilitating its use. University of Canterbury associate professor Daniel Gerhard and senior lecturer Donald Derrick are thanked for their advice on statistical procedures. Professor Sarah Wallace and Dr. Tamiko Azuma are thanked for helpful comments on a previous version of this manuscript. 10.1371/journal.pone.0288556.r001 Decision Letter 0 Hand Christopher James Academic Editor 2023 Christopher James Hand 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. 25 Apr 2023 PONE-D-23-03295Lexical-semantic properties of verbs and nouns used in conversation by people with Alzheimer's diseasePLOS ONE Dear Dr. Williams, 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. I hope that you find the reviewer's comments helpful. Additionally, I would recommend that you consider augmenting certain sections of your literature review to include the work of, for example, Boyd Davis, Paul Darden, et al. Please submit your revised manuscript by Jun 09 2023 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, Christopher James Hand, Ph.D., M.Sc., M.A., PgCAP 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\. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see <http://journals.plos.org/plosone/s/data- availability>. Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. Any potentially identifying patient information must be fully anonymized. Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: <http://journals.plos.org/plosone/s/data-availability#loc- unacceptable-data-access-restrictions>. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access. We will update your Data Availability statement to reflect the information you provide in your cover letter. 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 \*\*\*\*\*\*\*\*\*\* 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: 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: Two small changes: the new standard term is /people living with dementia/ -- I don't think, from personal observation when collecting a lot of the data in the collection, that they ALL had Alzheimer's. Your abbreviation throughout should be PLWD Set off the hypotheses as a bulleted list or it gets confusing Reviewer \#2: The authors aimed to investigate whether people with Alzheimer’s disease differed in their use of words from different grammatical categories compared to age matched controls. Data was from a US-based digital archive of interviews. Speech samples from 12 people with AD and 12 controls were matched for total word count. The AD group was found to produce less nouns than the control group, and type-token ratios were lower, indicating use of a narrower range of words across all word types, compared to controls. Analysis of the frequencies and age of acquisition of words used revealed that, with increasing age, the AD group used less frequent nouns, while the control group used nouns of higher frequency. The authors discuss education level and type of AD as potential reasons for the findings re the effect of age. They make the excellent point that there is a lack of studies of change in noun use in advanced ageing. The authors discuss the fact that their results indicate unimpaired verb use by the AD group, while research on single word naming reveals a verb disadvantage in AD. There is an excellent discussion of potential problems in matching nouns and verbs on age of acquisition and frequency in studies of single word naming. This was a clearly written paper (in the main, please see below), with well- founded measures employed in thorough analyses of the data. Implications of the findings for intervention are clearly drawn at the end of the paper. I believe the study is definitely worthy of publication. I just had some minor points for consideration, Line 74-75 ‘Interestingly, AoA comparisons have not indicated reliance on simpler words by pwAD’ – could the authors please explain what is meant here. Line 140-141 ‘Groups were not matched on education because this was provided in 4-year ranges’ – could the authors please explain what is meant here. \*\*\*\*\*\*\*\*\*\* 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.0288556.r002 Author response to Decision Letter 0 27 May 2023 This information is also copied and pasted in from the response to reviewers (E Williams): A. Editor’s comments: I hope that you find the reviewer's comments helpful. Additionally, I would recommend that you consider augmenting certain sections of your literature review to include the work of, for example, Boyd Davis, Paul Darden, et al. We found the reviewers’ comments extremely helpful. We also appreciated your suggestions of references to help establish the scientific context for our work. We identified two studies by Boyd Davis and colleagues that presented ramifications of changes to language and communication in Alzheimer’s disease. These now appear in the text and reference list as references 2 and 3. Use of these resources has resulted in changes to lines 39-40 (unmarked version) of the Introduction. Reference number 2 also supports and allows for expansion on a point made in the Discussion section; as a result, changes were also made to lines 583-584. B. Reviewer 1’s comments the new standard term is /people living with dementia/ -- I don't think, from personal observation when collecting a lot of the data in the collection, that they ALL had Alzheimer's. Your abbreviation throughout should be PLWD We acknowledge Reviewer 1’s familiarity with the CCC. We are happy to make this change if necessary, pending confirmation of whether the reviewer’s comment considers our description of inclusion criteria in the “Participants and language samples” section. The CCC includes Alzheimer’s disease and dementia as two distinct selection criteria under the Condition heading. We specifically included only data from participants meeting the Alzheimer’s criterion—no one who was only described using the broader term dementia. Our description thus reads, “Interviewees were considered for inclusion in our AD group only if they met the following criteria: … dementia specified as AD…” (bold letters added for emphasis). Can the reviewer please clarify whether this comment applies to participants designated specifically in the CCC search filters as having AD? If it does, we are happy to apply this change. If the comment no longer applies, but the reviewer feels our description could be clarified, we are happy to take suggestions on how to do so. Set off the hypotheses as a bulleted list or it gets confusing We have made this change and agree that it improves readability. Please see lines 95 to 104 of the unmarked version. C. Reviewer 2’s Comments Line 74-75 'Interestingly, AoA comparisons have not indicated reliance on simpler words by pwAD' - could the authors please explain what is meant here. We have revised and added detail to this statement. It now reads, “Interestingly, findings from picture description tasks have not indicated that pwAD rely on words of lower AoA \[8, 19\]. Fraser et al. \[8\] included AoA in their investigation of language features that might help distinguish pwAD from healthy older people but did not find it to contribute to identification of pwAD. Yeung et al. \[19\], meanwhile, actually found pwAD to use words of higher AoA than controls, a finding they attributed to a tendency by pwAD to produce utterances that were inappropriate to the task.” In the current unmarked version, that information can be found in lines 68 to 73. Line 140-141 'Groups were not matched on education because this was provided in 4-year ranges' - could the authors please explain what is meant here. We thank the reviewer for highlighting this. In following up on the comment, we found our initial description not to be fully accurate and we have rephrased as: “Groups were not matched for education because the CCC provides this only in broad ranges that do not facilitate meaningful matching—e.g., 1-8 years, 9-12 years, etc.” As a result of the revision process, this information now appears in lines 141-143. We have also revised wording now in line 411, as that sentence reiterates this point on education. 10.1371/journal.pone.0288556.r003 Decision Letter 1 Hand Christopher James Academic Editor 2023 Christopher James Hand 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. 29 Jun 2023 Lexical-semantic properties of verbs and nouns used in conversation by people with Alzheimer's disease PONE-D-23-03295R1 Dear Dr. Williams, 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, Christopher James Hand, Ph.D., M.Sc., M.A., PgCAP Academic Editor PLOS ONE Additional Editor Comments (optional): I am pleased to report that two reviewers have now endorsed publication of your manuscript. They have provided a little additional feedback; please consider this when preparing the final version of your manuscript for resubmission. Please also carefully check the manuscript and full references for any typographic errors, formatting inconsistencies, etc. 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 \*\*\*\*\*\*\*\*\*\* 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: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes 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 \#1: Yes Reviewer \#2: No \*\*\*\*\*\*\*\*\*\* 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 \#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 \#1: I still think that AOA is not needed although I think you gave a very good discussion of its limitations. That's why I said 'partly' above. Reviewer \#2: Thank you for clearly responding to the two queries. All required questions have been answered and all responses meet formatting specifications. \*\*\*\*\*\*\*\*\*\* 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: **Yes: **Jackie Masterson \*\*\*\*\*\*\*\*\*\* 10.1371/journal.pone.0288556.r004 Acceptance letter Hand Christopher James Academic Editor 2023 Christopher James Hand 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 Jul 2023 PONE-D-23-03295R1 Lexical-semantic properties of verbs and nouns used in conversation by people with Alzheimer's disease Dear Dr. Williams: 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. Christopher James Hand Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Natural disasters can occur rapidly and unpredictably leading to great loss of life and property, increased resource consumption within the family, and family dysfunction. Family dysfunction can in turn result in a range of psychological and behavioral problems, such as alcohol and drug abuse, depression and post traumatic stress disorder (PTSD). For example, Rowe et al. (2010) found that lower family cohesion was a significant risk factor for substance involvement and post traumatic stress in adolescent survivors. Similarly, Wickrama and Wickrama (2008) found that intact family status and family support reduced depression. The results of a qualitative study indicated that, the essential task of damaged families after disasters is to rebuild family life, reduce uncertainty related to disaster, and improve family functioning. Family functioning is defined as the extent to which a family operates as a unit to cope with stressors. The Circumplex model of marital and family systems holds that family functioning is composed of three dimensions: cohesion, adaptability, and communication. Family cohesion refers to the emotional bonding among family members. Family adaptability represents the ability of a family to change its rules, role relationships, and power structure in response to developmental changes or situational stressors. Communication is a facilitating factor to cohesion and adaptability. A review of literature indicates that a great majority of previous studies on family functioning have focused on patients, caregivers, children, and adolescents. Known predictors for family functioning include demographic characteristics, such as gender, age, marital status, educational level, occupation, and economic status. Meanwhile, family variables such as family structure, family support and coping style are also identified as significant predictors for family functioning. In addition, many studies have explored the effects of family functioning on behavioral and psychological problems, such as eating disorder, suicide, depression, and loneliness. Natural disasters can result in deaths of family members, damage to family properties and loss of family security. Disaster exposure can generate more proximal secondary disaster risks, such as post-disaster family problems. For instance, it was reported that 28.6% of disaster-affected families experienced poor family resilience after Cyclone Larry in Australia. However, to our knowledge, few studies have examined perceived family functioning in disaster survivors. Bereavement caused by disasters often results in long-term psychological disorders. For example, a study of bereaved survivors two years after the 2004 Tsunami found that the rates of major depression and PTSD were 25.0% and 34.4%, respectively. Another study of bereaved survivors in the 2008 Sichuan earthquake found that the prevalence of depression was 65.6% and the loss of a child was a significant predictor for psychopathological symptoms. ABC-X model of family stress maintains that whether families survive or fall into X (the crisis) when dealing with situational stressors depends on three factors: A (the event), B (the family’s crisis-coping resources), and C (the family’s evaluation of the event). Because bereavement caused by disasters is a mass stressor, bereaved families without sufficient crisis-coping resources, such as adaptive family functioning, are likely to experience detrimental long-term outcomes after disasters. Furthermore, positive family functioning may be a protecting factor for mental health. However, no study has examined perceived family functioning and its related factors in disaster bereaved individuals. In addition, loneliness defined as a chronic distress without redeeming features includes emotional loneliness and social loneliness. It has been shown that absence of frequent interactions with family members, neighbors or friends, and small social networks can cause loneliness. Studies on loneliness have found that demographic characteristics and personality variables are significant predictors for loneliness. However, these studies have focused on older people and patients with AIDS. Moreover, bereavement as a stressful event can disrupt family life and family functioning, which may in turn result in loneliness. However, no study has investigated the relationship between family functioning and loneliness in disaster bereaved individuals. On May 12, 2008, an 8.0-magnitude earthquake occurred in the northwest of Sichuan province, China. As one of the most severe natural disasters in the history of China, it is reported that 69,227 individuals died, 374,643 individuals were injured and 17,824 individuals were missing in the disaster- affected regions. A great number of survivors lost their beloved family members in the catastrophic event, and bereavement may have led to severe damage to family structure and family dysfunction. However, no study has examined perceived family functioning in bereaved individuals after disasters. Therefore, the aims of this study were to 1) examine perceived family functioning in bereaved individuals eighteen months after the 2008 Wenchuan earthquake; 2) explore the effects of demographic characteristics and disaster-related variables on family functioning; 3) analyze the relationship between perceived family functioning and loneliness. # Methods ## Samples A cross-sectional research design was used in the study. A convenience sample of 274 bereaved individuals was recruited via door-to-door interviews (each individual was drawn from a different bereaved family). Participants lived in the three hardest-hit regions in the Wenchuan earthquake (Dujiangyan, Mianzu and Li county) and all of them resided in the rural areas. The inclusion criteria were as follows: 1) the loss of biological family members in this earthquake (e.g., spouse, children, grandchildren, parents, grandparents and siblings); 2) aged 18 and older; 3) agreed to participate in the study. Participants who had cognitive impairment were excluded. The Chinese Mini–Mental State Examination (MMSE) was used to screen for cognitive impairment. According to the results, three different cut-off points were used depending on the respondent's educational level with a score \>17 (illiteracy), \>20 (primary school), and \>24 (junior high school or above) indicating no cognitive impairment. As a result, 10 potential participants were excluded based on their MMSE scores. ## Measures ### Family APGAR Index The scale assesses the extent to which individuals are satisfied with their family functioning. It is composed of five items: adaptation, partnership, growth, affection, and resolution with a three-point scale ranging from 0 *(hardly ever)* to 2 *(almost always)*. The total scores range from 0 to 10 with higher scores indicating higher levels of satisfaction with family functioning. A score of 0–3 indicates severe family dysfunction, 4–7 moderate family dysfunction, and 8–10 positive family function. Cronbach’s alpha coefficients reported across studies ranged from 0.80 to 0.85. In the study, Cronbach’s alpha coefficient for this scale was 0.88. ### Family Adaptability and Cohesion Evaluation Scale α (FACESα) The scale includes 30 items and two dimensions: family cohesion (16 items) and family adaptability (14 items). It is evaluated by a five-point scale ranging from 1 *(almost never)* to 5 *(almost always)*. The total scores on family cohesion and adaptability range from 15 to 80 and 15 to 70 respectively, with higher scores suggesting better family cohesion and adaptability. Four family types are determined by participants’ combined scores on cohesion and adaptability subscales, namely balanced family type (cohesion: 71–80; adaptability: 55–70), moderately balanced family type (cohesion: 60–70; adaptability: 46–54), mid-range family type (cohesion: 51–59; adaptability: 40–45), and extreme family type (cohesion: 15–50; adaptability: 15–39). Cronbach’s alpha coefficient for the overall scale is 0.90, and test-retest reliability is 0.84. In the present study, Cronbach’s alpha coefficient for this scale was 0.82. ### Emotional and Social Loneliness Scale (ESLS) The 10-item scale consists of two subscales: emotional loneliness and social loneliness. Each subscale includes five items. A five-point scale from 1 *(hardly ever)* to 5 *(often)* is used. The total score for each subscale ranges from 5 to 25 with higher scores indicating greater loneliness. Cronbach’s alpha coefficients for emotional loneliness and social loneliness are 0.78 and 0.76, respectively. Examination of the relationship between the ESLS and UCLA (University of California, Los Angeles Loneliness Scale) demonstrated its good concurrent validity. In the present study, Cronbach’s alpha coefficients for emotional and social loneliness were 0.76 and 0.90, respectively. Referring to recent studies on natural disasters, a number of items with fixed responses were developed in this study to evaluate disaster-related experiences during the Wenchuan earthquake, including types of dead family members in the earthquake, financial loss during the earthquake, housing conditions, types of family structure, self-rated health status and fecundity status after the earthquake. For instance, the items related to types of dead family members were: did your biological family members die in the earthquake? If yes, which family members died in the earthquake? The choices were spouse, children, grandchildren, parents, grandparents, and siblings. Responses to these choices were dichotomous (yes/no). With regard to the post-earthquake fecundity status of bereaved parents, the following four items were used: did your children die in the earthquake? If yes, did you and your spouse want another baby after the earthquake? If yes, did you and your spouse have another baby after the earthquake? If no, were you or your spouse pregnant after the earthquake? Responses to the above four items were also dichotomous (yes/no). In addition, we also collected demographic data such as gender, age, religious beliefs, educational level, and marital status. ## Procedure Prior to this study, two research assistants each with a master’s degree in nursing science participated in two training sessions to ensure that the measures would be administered reliably. As disaster-related experiences would be explored in bereaved individuals during the investigation, two experienced psychologists were invited to train the two research assistants effective communication skills to minimize respondents’ emotional distress. Three weeks after training, a pilot study was conducted on 23 bereaved individuals in Mianzu under the supervision of the psychologists. It was found that the data collection method was feasible, the two research assistants were able to make effective communication with the bereaved, and all the respondents understood the questionnaires. After contacting with the local governments and community service organizations, this survey was carried out from December 1, 2009 to January 31, 2010 in Mianzu, Dujiangyan and Li county, which were the hardest-hit areas in the Wenchuan earthquake. Respondents were required to complete the questionnaires independently according to their actual feelings. With regard to respondents who could not complete the questionnaires by themselves due to physical illnesses, the investigators recorded their responses. As for those respondents who were illiterate, the investigators read the questions and responses word-for-word and recorded their answers. The questionnaires were collected immediately after completion and checked for incomplete items. ## Data analyses The statistical analysis package used in the study was SPSS 16.0. Means and standard deviations (SD) on family functioning were examined. Independent samples t-tests or one-way analyses of variance were used to test differences in family functioning scores between two or more subgroups. Three multivariate regression analyses were run to identify significant predictors for family functioning, with family function, cohesion, and adaptability as the dependent variable, respectively, and demographic characteristics and disaster-related variables as the independent variables for each analysis. Among these independent variables, types of dead family members in the earthquake (reference group: other family members: parents, grandparents, siblings and grandchildren), types of family structure after the earthquake (reference group: single-parent family), and post-earthquake fecundity status of bereaved parents (reference group: not being pregnant) were included in regression equation as dummy variables. The relationships between family function, cohesion, adaptability and loneliness were evaluated by Pearson correlation analyses. *P* value \< 0.05 was considered statistically significant. ## Ethical statement Prior to the investigation, ethical approval was obtained from the Human Subjects Ethics Sub-committee of Sichuan University. Informed consent was obtained from each participant. Participants were assured of anonymity, confidentiality and their rights to withdraw from the study at any time. # Results ## Characteristics of the sample Of the 264 respondents in the study, 45.1% were male and 54.9% were female. The age ranged from 16 to 98 with an average age of 45.87 years (SD  =  13.08). The majority of the respondents were married (90.5%). 131 respondents (49.6%) experienced severe financial loss in the earthquake and 115 respondents (43.6%) lived in temporary post-disaster houses. The rates of post-earthquake nuclear family and extended family were 32.6% and 61.0%, respectively. In the earthquake, 142 respondents (53.8%) only lost their children, 15 respondents (5.7%) only lost their spouses, 57 respondents (21.6%) only lost other family members, such as parents, siblings, grandparents, or grandchildren, and 50 respondents (18.9%) lost two or more than two types of biological family members. Overall, 190 respondents lost their children in the earthquake and 91.2% of the respondents (177/190) wanted another baby. ## Family functioning in bereaved individuals The mean score on family function (Family APGAR Index) was 6.5 (SD  =  2.7). 50% of the respondents (n  =  132) reported positive family function, 37.1% (n  =  98) moderate family dysfunction, and 12.9% (n  =  34) severe family dysfunction. The mean scores on family cohesion and adaptability (FACES??) were 64.2 (SD  =  9.2) and 41.8 (SD  =  6.6), respectively. Compared to the Chinese norms on the FACES?? (cohesion: 63.9±8.0, adaptability: 50.9±6.2), the mean score on family adaptability in the study was significantly lower (t  =  –22.22, *p*  =  0.00), and the mean score on family cohesion was slightly higher but not statistically significant (t  =  0.56, *p*  =  0.58). Regarding the score on each item of the FACES??, the five items with the highest scores and the lowest scores are shown in. All the five items with the highest scores belonged to the family cohesion subscale. Based on linear scoring and interpretation of the FACES??, 15.9% (n  =  42) of the respondents belonged to extreme family type, 33.0% (n  =  87) mid-range family type, 46.2% (n  =  122) moderately balanced family type, and 4.9% (n  =  13) balanced family type. ## The relationships between demographic characteristics and disaster-related variables, and family functioning The results of bivariate analyses found no statistically significant differences in family function (Family APGAR Index), cohesion, or adaptability (FACES??) between male and female respondents. No statistically significant differences were found in family function or family adaptability scores among different types of dead family members in the earthquake. Whether bereaved parents had another baby or were pregnant or not, and age groups were not significantly related to family adaptability. However, statistically significant differences were found in family function, cohesion, and adaptability depending on different educational levels, religious beliefs, marital status, financial loss during the earthquake, post-earthquake housing conditions, self-rated health status, and types of family structure. The results of multivariate regression analyses found that less financial loss in the earthquake was significantly related to positive family function (β =  –0.20, *p*  =  0.00). Relatively better health status after the earthquake was significantly related to positive family function, cohesion, and adaptability (β =  0.20, *p*  =  0.00; β =  0.22, *p*  =  0.00; β =  0.17, *p*  =  0.00). Scores on cohesion and adaptability among the respondents from nuclear families or extended families were significantly higher than those from single-parent families (β =  0.60, *p*  =  0.00; β =  0.56, *p*  =  0.00; β =  0.55, *p*  =  0.00; β =  0.48, *p*  =  0.00). Scores on family function and cohesion in bereaved parents who had another baby were significantly higher than those not pregnant (β =  0.13, *p*  =  0.04; β =  0.15, *p*  =  0.01). ## The relationships between family functioning and loneliness As shown in, positive family function, cohesion, and adaptability were significantly related to less emotional and social loneliness (r ranged from –0.31 to –0.53, *p* \< 0.001), indicating that the better their perceived family functioning, the less emotional and social loneliness they experienced. # Discussion ## Family functioning in bereaved individuals Our study found that half of the bereaved individuals reported family dysfunction and 48.9% of the respondents reported a mid-range or extreme family type. An epidemiological study conducted by McDermott et al. (2010) found that three months after Cyclone Larry in Australia, the prevalence of family dysfunction in disaster-affected families was 28.6% , which is lower than our findings. The disparity may be attributed to the detrimental effects of disasters-induced bereavement on family functioning. Disaster deaths can disrupt family life. However, because no previous study has examined perceived family functioning in disaster bereaved individuals, our results cannot be compared meaningfully with findings from previous studies. Further studies are needed to compare perceived family functioning between bereaved and non-bereaved groups after disasters, and analyze whether bereavement caused by disasters is a risk factor for family health in disaster survivors. We also found that all the five items with the highest scores in the FACES?? belonged to family cohesion dimension and the cohesion score was higher than the Chinese norms on the FACES??. This is consistent with Drabek’s view (1984) that emotional bonding, mutual support, as well as family ties tend to be reinforced after disasters. Our findings are also in line with crisis theory developed by Andrew (1976), which suggests that although specific life events such as death or separation of family members can impede individuals’ satisfaction with basic needs, it can also improve family cohesiveness. Another finding of this study was that the mean score on family adaptability of the sample was significantly lower than the Chinese norms on the FACES??. This concurs with the result of Dyregrov’s study (2001) that bereavement can influence family adaptation to traumatic events and grief over time. Moreover, bereavement can interfere with individuals’ normal life. The relationship between the dead and the bereaved does not disappear in a short time, and strong feelings for the lost person may last for many years, which may undermine ability of family members to transform existing family roles, role relationships, and power structure. How to accommodate to a new circumstance without the loved one or ones is a great challenge for bereaved individuals after disasters. Besides, all the respondents in our study experienced financial loss during the earthquake and over 40% of them lived in temporary post-disaster houses. This may contribute to the loss of safety and stability in the lives of disaster-affected families and result in difficulties for them in adapting to the tremendous family misfortune caused by a devastating natural disaster. In response, after the 2008 Wenchuan earthquake, central and local governments and nongovernmental organizations in China took various measures to help residents cope with disaster-induced health problems, such as sending rescuers and medical care teams to the earthquake-affected areas, transferring the wounded to other provinces for medical treatment, establishing mental health services stations and training residents for psychological relief. However, the majority of the mental health services were on the individual level but not on family health, suggesting the need to direct post-disaster psychotherapy, such as family therapy or family counseling, towards families and to nurture adaptation and development in families. ## Predictors for family functioning in bereaved individuals Our study found that less financial loss during the earthquake was a significant predictor for positive family function. The result is similar to Norris and Uhl’s finding that greater financial loss predicted increased marital stress and reduced family support after hurricane Hugo. Financial loss leads to economic pressure in the family, which could affect family functioning and individual adjustment. Although emergency assistance in food, drinking water, clothing, and medical care was provided by central and local governments immediately after the earthquake, long-term efforts to help survivors improve their economic situation are needed, especially for those with severe financial loss. Perceived health refers to an individual's general perception of health, including biological, psychological, and social dimensions. The present study found that better perceived health after the earthquake was a significant contributing factor to positive family function, cohesion, and adaptability. The result is consistent with the findings of Cano et al. (2003), who reported that poor self-rated health was a significant risk factor for family dysfunction in patients with chronic illnesses, and with the findings of McDermott and Cobham (2012) who reported that increased emotional distress in survivors was a significant risk factor for family dysfunction three months after Cyclone Larry in Australia. Poor perceived health may increase family vulnerability with less employment opportunity, higher levels of family conflict, and increased resources consumption within the family. Improving the health status of family members might be an effective way to facilitate family functioning after disasters. Our study also found that the respondents from nuclear families or extended families reported higher family cohesion and adaptability scores than those from single-parent families. This is consistent with Youngblut and Brooten’ study (2006). Nuclear family is characterized by independence and flexibility with more equal relationships shared by family members. These characteristics may contribute to changes in relationship roles and family rules, and reduction in family conflict when traumatic events occur. In addition, in Chinese society, especially in the rural areas, extended family is the dominant family structure. In extended families, older parents usually cohabit with their children with strong emotional bonding. They are financially supported by their children and shoulder the responsibility for taking care of their grandchildren. This kind of family structure is conducive to developing close family ties and improving mutual negotiations and adaptability when an unpredictable crisis occurs. These findings suggest that more post-disaster efforts such as family therapy focusing on improving relationship patterns among family members should be provided for bereaved families, especially for single-parent families due to their lack of normal family structure and limited emotional support. Moreover, our study found that those bereaved parents already having another baby after the earthquake reported significantly higher scores on family function and cohesion than those not pregnant. It was reported that about seventy thousand people died in the 2008 Wenchuan earthquake, and the vast majority of them were children and adolescents. For the bereaved parents, the birth of another baby can bring hope and enhance emotional bonding among family members. However, as for the bereaved families who were still not pregnant eighteen months after the earthquake, their expectations of pregnancy were not satisfied. As a result, they might be filled with enormous stress, anxiety, and depression, which may impede emotional communication among family members and reduce family cohesion. ## The relationships between family functioning and loneliness Although no previous study has explored the relationship between perceived family functioning and loneliness among disaster-affected individuals, the results of our study are in line with studies focusing on older people and patients with AIDS. Wu et al. (2010) found that positive family functioning and social support in the empty nest elderly were significant predictors for less loneliness. Similar results were also found in the study of Sun et al. (2009) among AIDS patients. These findings suggest that individuals with positive family functioning may obtain more emotional and spiritual support from family members and thereby reduce their loneliness. Bereavement can lead to family dysfunction. Impaired family functioning can in turn result in loneliness. Further studies are needed to explore the moderating effect of family functioning on the relationship between disaster-related bereavement and loneliness. Several limitations are identified in the study. First, participants of this study were drawn from the hardest-hit areas and a convenience sampling method was used, which might have undermined the representativeness of the selected sample. Second, the study adopted a cross-sectional design, which excluded the possibility of identifying possible developmental variations in perceived family functioning. A longitudinal study is warranted. Third, perceived family functioning of the participants before the earthquake was not assessed. This made it impossible to compare the reported prevalence of family dysfunction with that before the earthquake. Therefore, the reported prevalence of family dysfunction should be interpreted with caution. Finally, only a small number of predictors for family functioning in bereaved individuals were examined in our study. Future research needs to investigate other possible predictors, such as coping styles and social support. Despite these limitations, to our knowledge, this is the first study exploring family functioning and its predictors as well as the relationship between family functioning and loneliness in disaster bereaved individuals. Results of this study provide new insight into the psychosocial aftermath of catastrophic natural disasters. Specifically, the study revealed a high prevalence of family dysfunction in disaster bereaved individuals eighteen months after the Wenchuan earthquake and identified four significant predictors for perceived family functioning. These findings can provide useful implications for post-disaster rebuilding and relief work, particularly for healthcare providers to identify high-risk bereaved populations and to develop strategies to help them. Such strategies may include helping disaster survivors with financial difficulties to improve their economic situation, providing long-term health care to those with poor health status, providing professional reproductive assistance to bereaved parents, and initiating family psychotherapy such as family counseling for single-parent families to improve family functioning and reduce loneliness for bereaved individuals. We would like to thank our research assistants, participants, officials in the local governments, and community workers for their cooperation and assistance. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: XYC XLJ XLL. Performed the experiments: XYC XLJ XLL RL. Analyzed the data: XYC MJHL RL. Contributed reagents/materials/analysis tools: XLJ MJHL. Wrote the paper: XYC XLJ XLL MJHL RL.

# Introduction Oxidative stress in the central nervous system (CNS) plays a critical role in the pathogenesis of many acute as well as chronic neurological disorders, such as stroke, Alzheimer’ s disease, Parkinson’s disease, Huntington’s disease, and amyotrophic lateral sclerosis (ALS). We have previously identified the novel small molecule compound, 2-\[mesityl(methyl)amino\]-N-\[4-(pyridin-2-yl)-1H-imidazol-2-yl\] acetamide trihydrochloride (named WN1316), as a unique oxidative stress suppressor and a therapeutic candidate for the treatment of ALS, by employing a new combined approach for the drug screening with *in silico* drug designing and *in silico* drug screening by an algorithm of quantitative relationship between the chemical structure of the compound and its anti-oxidative neuronal cell death activity via up-regulation of neuronal apoptosis inhibitory protein (NAIP). However, the precise molecular as well as pharmacological mechanisms by which WN1316 protects neurons from oxidative stress through suppression of IL1-β and gliosis *in vivo* have been remained to be elucidated. During the course of study, we found that the concentration of WN1316 in the brain after oral administration in wild-type mice was approximately three times higher than that in serum, despite the fact that almost all WN1316 molecules bound to serum proteins. By contrast, the concentrations of the WN1316-sister compounds in the brain and serum after oral administration in mice were most similar. It is generally noticed that the dose range of a drug required to produce *in vivo* efficacy is substantially higher than *in vitro* potency. However, the efficacy of WN1316 in ALS mouse models emerged at a sub-nanomolar range, which was rather hundred times lower than its potency in cultured human neuronal cells. Differences in the efficacy dose of WN1316 between *in vitro* and *in vivo* implies the presence of a mechanism by which WN1316 is preferentially enriched in the brain. Remarkably, it has been demonstrated by our preliminary study that fetal bovine serum (FBS) is a requisite for the cytoprotective activity of WN1316 *in vitro*. These led us to hypothesize that some of serum proteins closely linked to the WN1316-dependent cytoprotective activity. In this study, we attempted to identify the biological WN1316-activating factors in FBS. We successfully identified nine candidate serum proteins by using a proteomic approach in conjunction with quantitative WN1316 cytoprotective activity analysis. Among them, two *N*-linked glycoproteins, alpha-2-HS- glycoprotein (AHSG) and hemopexin (HPX), markedly exhibited the WN1316-mediated cytoprotection against oxidative injury in differentiated neuroblastoma SH-SY5Y cells. Thus, these two glycoproteins play a key role in exerting the WN1316-mediated cytoprotective activity *in vitro*, suggesting that they are ever-unidentified possible endogenous factors implicating in the effective drug delivery to the CNS. # Materials and methods ## Cells and reagents Human neuroblastoma SH-SY5Y cells (CRL-2266) were obtained from the American Type Culture Collection (Manassas, VA). WN1316 was synthesized by Wakunaga Pharmaceutical Co., Ltd. (Hiroshima, Japan). The purity of the compounds is higher than 99.2%. We purchased AlamarBlue from Invitrogen (Carlsbad, CA), PNGase F, and Protein Deglycosylation Mix from New England Biolabs (Tokyo, Japan), XerumFree FBS replacement from TNCBIO (Eindhoven, Netherlands), WEPRO1240G Expression Kit from Cell-Free Sciences (Yokohama, Japan), and all-*trans* retinoic acid and Dulbecco’s modified Eagle’s medium (DMEM) from Wako Pure Chemical Industries, Ltd. (Osaka, Japan). General laboratory reagents were obtained from Nacalai Tesque (Kyoto, Japan) and Sigma (St. Louis, MO). All other laboratory reagents were from commercial sources and of analytical grade. ## Anti-oxidative stress-induced neuronal cell death (AOND) assay SH-SY5Y cells were grown in DMEM supplemented with 10% non-heat inactivated FBS (HyClone, Thermo Fisher Scientific, Waltham, MA), 100 U/ml penicillin, and 100 μg/ml streptomycin at 37°C in 5% CO₂. SH-SY5Y cells were seeded at a density of 0.75 × 10⁴ cells per well in 96-well plates (Primaria; BD Falcon, Franklin Lakes, NJ) for 1 day and were differentiated in the medium containing 10 μM all-*trans* retinoic acid (RA) for additional 5 days. Under such culture conditions, majority of SH-SY5Y cells showed the neuron-like morphology with growing neurites. We also confirmed a significant increase in the expressions of neuron specific enolase (NSE) and neuronal nuclei (NueN) in RA-differentiated SH-SY5Y cells by Western blotting (data not shown). WN1316 was dissolved in dimethyl sulfoxide (DMSO) to prepare 50 mM stock solution. Differentiated SH-SY5Y cells were then treated with 8 μM WN1316 for 3 h followed by 12 h of chase incubation without the compound. Cells treated with DMSO only were handled as vehicle control in this assay. Then, cells were exposed to 40 μM menadione for 4 h. Cell viability (the percentage of viable cell numbers under menadione-treated / -untreated conditions) was measured by AlamarBlue assay as described previously. In this study, the WN1316-mediated cytoprotective activity was expressed as a relative value of the cell viability in WN1316-treated cells for that in DMSO-treated control. For WN1316-activating factor additive experiments, we used serum-free medium containing 10% XerumFree FBS replacement instead of 10% FBS to exclude the influence of FBS in the assay. ## Separation of active protein fractions from FBS ### Ammonium sulfate precipitation FBS (HyClone, 7 ml) was fractionated by addition of pulverized ammonium sulfate to make the final concentration of 50% on ice with gentle stirring for 30 min. The 0–50% ammonium sulfate precipitate (ASP_0-50%) was collected by centrifugation at 13,000 x *g* for 30 min. The following steps were performed at 4°C unless otherwise stated. ### HiTrap Blue HP column chromatography ASP_0-50% was dissolved in approximately 1 ml of 20 mM HEPES buffer (pH 7.4), and dialyzed against 20 mM HEPES buffer (pH 7.4). After dialysis, sample (total protein amount of 42 mg) was applied to a HiTrap Blue HP column (5 ml, GE Healthcare, Tokyo, Japan) equilibrated with 20 mM HEPES buffer (pH7.4) using a manual syringe. After a 50-ml wash with 20 mM HEPES buffer (pH 7.4), the elution of proteins bound to the column was performed with a step-wise elution with 20 mM HEPES buffer (pH 7.4) containing 2 M NaCl, and 1 ml of fractions were collected. Aliquot of fractions from the column unbound (Blue pass) and bound elutes (Blue elute) were subjected to the AOND assay and protein content analysis. ### Bio-Scale Mini CHT Type I column chromatography The active pass-through fraction (total protein amount of 26 mg) from HiTrap Blue HP column was dialyzed against 5 mM phosphate buffer (starting buffer, pH 6.5) and loaded onto a Bio-Scale Mini CHT Type I column (5 ml, Bio-Rad, Hercules, CA) equilibrated with starting buffer using an AKTA Prime Plus system (GE Healthcare). After a 50-ml wash with starting buffer, protein was eluted with a linear gradient of 0–3 M NaCl in starting buffer (100 ml) at a flow rate of 0.5 ml/min, followed by the elution with 50 ml of 300 mM phosphate buffer (pH 6.5), and 1ml of fractions were collected. Aliquots of fractions were subjected to the AOND assay and protein content analysis. The fraction numbers 6 to 14 (named as W1) from washing fraction, 61 to 62 (named as E1) from elution fraction of 0–3 M NaCl in starting buffer, and 119 to 122 (named as C1) from elution fraction of 300 mM phosphate buffer (pH 6.5) were pooled, respectively. Each pooled fraction was concentrated by an Amicon ultra-15 (Millipore, Bedford, MA), and stored at 4°C until use. Protein concentration was quantified by Pierce 660 nm Protein Assay kit (Thermo Fisher Scientific) using bovine serum albumin (BSA) as a standard. In this study, one unit of the WN1316-activating factor activity was defined as the amount of protein required to increase 1% cell viability when using FBS. ## Proteomics analysis In this analysis, fraction E1, C1, and W1 from Bio-Scale Mini CHT Type I column chromatography were rendered to two-dimensional polyacrylamide gel electrophoresis (2DE). These proteomics analyses were conducted by Towa Environment Science Co., Ltd. (Hiroshima, Japan). Briefly, the sample (300 μg) was loaded onto reswollen gel strip with immobilized pH gradient (IPG) (pH 4 to 7, 24 cm long, GE Healthcare) at 20°C for overnight. The proteins were then separated by isoelectric focusing (IEF) using a CoolPhoreStar IPG-IEF Type-P (Anatech, Tokyo, Japan) at 20°C and focused with the following program: 500 V for 1 min, 500 to 3,500 V for 90 min, and 3,500 V for 14.5 h. After completion of IEF, the strips were sodium dodecyl sulfate (SDS) equilibrated, reduced and alkylated with dithiothreitol and iodoacetamide, followed by separation on a 9–18% gradient SDS-polyacrylamide gel electrophoresis (SDS-PAGE) gels (26 x 22 cm, Laboratory-made) using the Anderson ISO-DALT electrophoresis system (Hoefer) at 80 V constant for 16 h. All the experiments were run in quadruplicate for each sample to ensure reproducibility. The gels were stained with SYPRO Ruby protein gel stain (Molecular Probes) and scanned with Molecular Imager FX (Bio- Rad). Image analysis and spot matching were performed using the ImageMaster 2D Platinum (GE Healthcare). The quantitative level of each protein spot was determined by the relative volume of each spot in the gel and expressed as %Volume (spot volume/Σ volumes of all spots). Protein spots were excised from the 2DE and digested with trypsin (37°C, 16 h). The resulting peptides were concentrated and desalted using Zip Tip C18 (Millipore), and mixed with α-cyano-4-hydroxycinnamic acid (Wako) and spotted into wells of a MTP Anchorchip 600/384 (Bruker daltonics). Mass spectra were acquired in reflector positive ion mode using an Ultraflex MALDI-TOF/TOF mass spectrometer (MS) (Bruker daltonics). Selected peptides were fragmented in the second dimension, and the proteins were identified by mass searches in the NCBInr database using the MS/MS ion search software Mascot (Matrix Science, <http://www.matrixscience.com>). The identified proteins were analyzed its properties and functions by using the the UniProt Knowledgebase (UniPortKB) database (<http://www.uniprot.org>), which is a central hub of protein knowledge by providing a unified view of protein sequence and functional information. ## Separation of *N*-linked glycoproteins using a Con A lectin column FBS was buffer-exchanged against 20 mM HEPES, 0.5 M NaCl, 1 mM MnCl₂, 1 mM CaCl₂ (binding buffer, pH 7.4) by Zeba Spin Desalting Columns (Thermo Fisher Scientific). The buffer-exchanged FBS (1 ml) was loaded onto a HiTrap Con A 4B column (1 ml, GE Healthcare) equilibrated with binding buffer using a manual syringe. After a 25-ml wash with binding buffer, the elution of proteins bound to the column was performed with 25 ml of 20 mM HEPES, 0.5 M NaCl, 0.5 M methyl-α-D-glucoside (pH 7.4), and 0.5 ml of fractions were collected. Aliquot of fractions from column unbound (Con A pass) and bound elutes (Con A elute) were subjected to the AOND assay and protein concentration analysis. ## Deglycosylation of proteins Removal of glycans from target proteins in Blue pass fraction was carried out by PNGase F, which cleaved *N*-linked oligosaccharides from glycoproteins, and Protein Deglycosylation Mix (EnzMix) containing PNGase F, *O*-glycosidase, neuraminidase (sialidase), β1–4 galactosidase, and β-*N*-acetylglucosaminidase, which removed both *N*-glycans and some *O*-glycans. Under non-denaturing conditions, Blue pass fraction (200 μg) was incubated with or without PNGase F (1 unit) or EnzMix (1 unit) in 50 μl of 50 mM sodium phosphate (pH 7.5) at 37°C for 16 h. The molecular masses and homogeneities of the proteins, with or without the enzyme treatment, were analyzed by Western blotting with anti-AHSG antibody as a marker of glycoproteins. ## Western blotting Proteins (0.2 μg/lane) were resolved by SDS-PAGE on a 5–20% gradient gel (ATTO, Tokyo, Japan) and transferred onto polyvinylidene difluoride membrane (Bio-Rad). Membranes were blocked with Blocking-One reagent (Nacalai Tesque) for 1 h at room temperature and then incubated with the anti-AHSG antibody (1:1,000 dilution; ab112528; Abcam, Cambridge, UK) in Can Get Signal Solution 1 (TOYOBO, Osaka, Japan) overnight at 4°C. The membrane was washed with TBS-T (50 mM Tris- HCl, pH 7.4, 150 mM NaCl, 0.1% Tween 20) three times and incubated for 1 h at room temperature with horseradish peroxidase-conjugated secondary anti-rabbit IgG (#NA934, GE Healthcare) in Can Get Signal Solution 2 (TOYOBO). The membrane was washed as described above, and then signals were detected using Immobilon Western HRP Substrate (Millipore) and X-ray film. ## Plasmid construction For the synthesis of glutathione S-transferase (GST) fusion proteins by the wheat germ cell-free system, three human cDNAs, *AHSG* (FXC03806), *SERPINA1* (FXC03481) and *RBP4* (FXC03871), were obtained from Flexi ORF clone (Promega, Madison, WI) and were amplified by PCR in a GeneAmp PCR system 9700 (Applied Biosystems, Tokyo) with PrimeSTAR Max DNA Polymerase (TaKaRa) according to the manufacturer's instructions. Specific primer sets were as follows: *AHSG*, `5′-AAAACTAGTATGAAGTCCCTCGTCCTGCTCCTTTG-3'` and `5′-ATGCTAGTTATTGCTCAGCGG-3'`; *SERPINA1*, `5′-AAAACTAGTATGCCGTCTTCTGTCTCGTGGGGCA-3'` and `5′-ATGCTAGTTATTGCTCAGCGG-3'`; *RBP4*, `5′-AAAACTAGTATGAAGTGGGTGTGGGCGCTCTTGC-3'` and `5′-AAACTCGAGTTAAACCAAAAGGTTTCTTTCTGAT-3'`. Four human cDNAs, *A1BG*, *CFB*, *ITIH4* and *HPX* were amplified from Human Total RNA Master Panel II (Clontech, Palo Alto, CA) by RT-PCR with iScript cDNA synthesis kit (Bio-Rad) and PrimeSTAR Max DNA Polymerase (TaKaRa), following the manufacturer's instructions. PCR primers specific to each cDNA were as follows: *A1BG*, `5′-ATGTCCATGCTCGTGGTCTTTCTC-3'` and `5′-AAAAGTCGACTCAGCTTTCTGCCACCAGGAGCTC3'`; *CFB*, `5′-ATGGGGAGCAATCTCAGCCCCCAA-3'` and `5′-AAAAGTCGACTCATAGAAAACCCAAATCCTCATC-3`'; and *HPX*, `5′-ATGGCTAGGGTACTGGGAGCACCC-3'` and `5′-AAAAGTCGACCTAGATGTGGGCTACAGCGCATAT-3'`. The *ITIH4* primers were designed to amplify the 70-kDa mature form of human ITIH4 (residues 85–1983) and they were as follows: `5′-AAAACTAGTATGGAAAAGAATGGCATCGACATCTA-3'` and `5′-AAAAGTCGACCTACAGCTCCACAGACCAGCAGGA-3'`. Amplified fragments were digested with *Sal*I (*AHSG*, *SERPINA1*, *A1BG*, *CFB*, *ITIH4* and *HPX*) or *Xho*I (RBP4), and then cloned into the *Sma*I-*Sal*I or *Sma*I-*Xho*I sites of pEGST expression vector, respectively. The resulting plasmids were designated pEGST- AHSG, pEGST-SERPINA1, pEGST-A1BG, pEGST-CFB, pEGST-ITIH4, pEGST-HPX, and pEGST- RBP4, respectively. ## Cell-free protein synthesis Wheat germ cell-free translational system is most suitable for the expression of unmodified full length eukaryotic proteins from plant and animal sources, and is capable of producing proteins with proper folding. Cell-free protein synthesis was performed with WEPRO1240G Expression Kit according to the manufacturer's instructions. GST fusion proteins synthesized by the cell-free system were purified through a glutathione-Sepharose 4B column (GE Healthcare) according to the instructions of the manufacturer. ## Statistical analysis Data in this study were presented as mean ± standard deviation (SD). Statistical analyses were conducted using PRISM 5 (GraphPad). Statistical significance was evaluated by ANOVA (analysis of variance) followed by Dunnett’s method for multiple comparisons between groups. We considered *p*-values \< 0.01 to be statistically significance. # Results ## Protein factors in FBS exert WN1316-mediated cytoprotective activity In our preliminary study, we found that FBS was essential to exhibit the cytoprotective activity of WN1316 *in vitro* (data not shown). To verify this, we investigated whether FBS affected the cytoprotective activity of WN1316 by AOND assay. The WN1316-meidated cytoprotective activity was observed in a FBS dose-dependent manner, but not in serum-free conditions. The results suggest that actual component in FBS is vital for the cytoprotective activity of WN1316. To investigate the nature of the component, we next analyzed whether the WN1316 cytoprotective activity was affected by the heat-denature treatment of FBS. In this study, we used mild conditions (65°C), which was higher than those in heat- inactivation of FBS (56°C), to avoid the heat-induced protein precipitation. The WN1316 cytoprotective activity was decreased under the cell culture conditions with heat-denatured FBS, while it was partially recovered by the renatured FBS. These results indicate that thermolabile protein(s) plays a pivotal role in the WN1316-mediated cytoprotection. Here, we designated these components as WN1316-activating factors. Since it was revealed that the effective concentrations of FBS ranged from 1 to 5%, we used 2% of FBS, as a positive control, in all the following assays for the WN1316-mediated cytoprotective activity. ## Separation of WN1316-activating factors from FBS To identify the WN1316-activating factors, we sieved proteins, which were not associated with the WN1316 cytoprotective activity, out of FBS. Our preliminary study showed that bovine serum albumin (BSA), a major serum component, had no relation to the WN1316 cytoprotective activity (data not shown). Therefore, we firstly removed BSA from FBS by 50% saturated ammonium sulfate precipitation and following HiTrap Blue HP chromatography. The WN1316 cytoprotective activity was concentrated in the ammonium sulfate fraction and recovered in the pass-through fraction (Blue pass) of HiTrap Blue HP column. Subsequently, the Blue pass fraction was subjected to a Bio-Scale CHT Type I hydroxyapatite affinity chromatography. An approximately 80% of the loaded protein passed through the column and was found no WN1316 cytoprotective activity. Thus, fraction numbers from 6 to 14 were pooled as a non-activity fraction W1. Although the WN1316 cytoprotective activity was broadly eluted out, the highest activity was eluted in fraction numbers 61 and 62, and they are pooled as a high-specific activity fraction E1. Residual WN1316 cytoprotective activity in the wash out fractions (fraction numbers 119 to 122) were also pooled as a low-specific activity fraction C1. The specific activity of each purification step, 50% saturated ammonium sulfate precipitation fraction, Blue pass fraction, and high-specific activity fraction E1 on Bio-Scale CHT Type I hydroxyapatite affinity chromatography, was 20.6, 26.6, and 979 units/mg, respectively. Through these purification steps, WN1316-activating factor was concentrated 117-fold in E1 with a final yielded of 17%. ## Identification of WN1316-activating factor candidates by proteomics analysis To identify WN1316-activating factor, the high-specific activity fraction E1 was further sieved and analyzed by a proteomic approach based on 2DE/MS, while non- activity fraction W1 and low-specific activity fraction C1 were used as references for this proteomics analysis. The protein spots in the 2DE gel were classified into 5 groups. We sorted the spots for further analysis using the following two step ways: Step 1, choosing the E1 specific spots by comparing %volumes of the spots between E1 and the reference W1; and Step 2, selecting out the E1-increased protein spots by comparing %volumes of the spots between E1 spots chosen in Step 1 and the reference C1. As a result, 104 spots were identified in Step 1 as E1 specific signals, and in the following Step 2, 93 spots were selected as E1 increasing signals. Among these spots, 42 spots satisfied with the criteria of sufficient amounts for MS analysis. MS/MS ion search using MALDI-TOF/TOF mass spectrometry classified all the identified proteins, which were multiple spots with different isoelectric points and molecular weights, into 9 clusters; alpha-2-HS-glycoprotein (AHSG), retinol- binding protein 4 (RBP4), hemopexin (HPX), and transferrin (TF), which are related to the transport system; alpha-1-antiproteinase (SERPINA1), a protease inhibitor; gelsolin, which is involved in actin filament severing; complement factor B (CFB), a control factor involved in complement activation; inter-alpha- trypsin inhibitor heavy chain H4 (ITIH4), and anti-inflammatory protein; alpha-1B-glycoprotein (A1BG), glycoprotein with unknown function. ## *N*-linked glycoproteins exert WN1316-mediated cytoprotective activity We searched the characteristics of WN1316-activating factor candidates with the UniProtKB database, and found that seven out of them, except for gelsolin and RBP4, were glycoproteins with potential *N*-glycosylation sites. To investigate whether WN1316-activating factors are glycoproteins, we separated *N*-linked glycoproteins in FBS by using Con A lectin column and measured its WN1316-mediated cytoprotective activity. The Con A-elute as a *N*-linked glycoprotein fraction, but not the Con A pass-through as a non-glycoprotein fraction, showed the WN1316-mediated cytoprotective activity. These results suggest that WN1316-activating factors are *N*-linked glycoproteins. Thus, non- glycoproteins RBP4 and gelsolin were excluded from candidates. ## Deglycosylation does not affect WN1316-mediated cytoprotective activity To confirm whether the glycan chains on target proteins were associated with the WN1316-mediated cytoprotective activity, we carried out AOND assay using deglycosylation-treated FBS. The glycan chains on the glycoproteins were enzymatically digested by glycosidases. The deglycosylation was validated by SDS-PAGE and Western blotting with the commercially available antibody against AHSG as a marker of glycoproteins in FBS. We observed the differences in the protein migration patterns from FBS by the treatment with or without glycosidases. We also confirmed that the AHSG signals after the treatment with glycosidases were shifted to lower molecular weight compared to non-treated control. Intriguingly, WN1316-mediated cytoprotective activity was not affected by the removal of glycans on serum proteins when compared to non-treated control. This suggests that the sugar chains conjugated on WN1316-activating factors are not necessary for the WN1316-mediated anti-oxidative stress function *in vitro*. ## Preparation of human homologues for WN1316-activating factor candidates The final goal of our research is to obtain the human-derived endogenous WN1316-activating factor for elucidating drug efficacy mechanism of WN1316 in human. As shown in, since glycan chains play no major role in the WN1316 cytoprotective activity *in vitro*, we sought to evaluate the WN1316-mediated cytoprotective activity by using human homologs for the WN1316-activating factor candidates. We amplified the cDNAs of human homologue corresponding to 7 bovine WN1316-activating factor candidates (AHSG, SERPINA1, RBP4, ITIH4, A1BG, CFB, and HPX) and generated the cell-free expression constructs. Non-glycoprotein RBP4 was used as a negative control for AOND assay. Each GST-fusion protein synthesized by the wheat germ cell-free translation system was purified by a glutathione-Sepharose 4B column and analyzed the molecular masses by SDS-PAGE. We confirmed that sizes of GST-AHSG, GST-SERPINA1, GST-RBP4, GST-ITIH4, GST-A1BG, GST-CFB, and GST-HPX were 67, 74, 50, 98, 81, 56, and 113 kDa, respectively, as expected. We also utilized the commercially available human TF. ## Two *N*-linked glycoproteins, alpha-2-HS-glycoprotein and hemopexin, are identified as WN1316-activating factors To identify the genuine WN1316-activating factor(s) from the synthesized candidate proteins, we performed AOND assay using differentiated SH-SY5Y cells. We confirmed no WN1316 cytoprotective effect both in RBP4 and GST. Among cell- free synthesized candidates, two proteins, AHSG and HPX, significantly exerted the WN1316 cytoprotective activity when compared to RBP4. In particular, AHSG showed among the highest cytoprotective activity (1.7-fold) compared to the vehicle control. Finally, to verify the WN1316 cytoprotective activity of TF, we carried out AOND assay in the presence of various concentrations of TF. The WN1316-mediated cytoprotective activity was significantly increased at extremely high concentrations of TF in a dose-dependent manner. However, no significant cytoprotective activity was observed at lower TF levels, even though those concentrations of TF were still approximately 10 times higher than the effective concentration of either AHSG or HPX; 10–25 μg/ml for TF v.s. 1.7–1.9 μg/ml for AHSG or HPX. Therefore, we excluded TF from the candidate of WN1316-activating factor. # Discussion In this study, we attempted to unveil the biological WN1316-activating factors in FBS, and successfully identified nine candidate serum proteins by using a proteomic approach in conjunction with quantitative WN1316 cytoprotective activity analysis. Among them, two *N*-linked glycoproteins, alpha-2-HS- glycoprotein (AHSG) and hemopexin (HPX), markedly exhibited the WN1316-mediated neuroprotection against oxidative injury in a glycan-independent manner, indicating that they are endogenous factors implicating in small molecule compound-mediated neuroprotection. To our knowledge, this is a first report that shows concrete evidences for the novel glycan-independent function of serum glycoproteins in drug efficacy. Novel amino imidazole derivative, WN1316, is a high blood-brain barrier (BBB) permeable small molecule compound that can alleviate disease progression via suppression of glial inflammation in ALS mouse models. Two unexpected findings were obtained from our previous studies. First, the efficacy of WN1316 in ALS mouse models emerged is hundred times lower concentration than its potency in cultured human neuronal cells *in vitro*. Second, FBS in culture media is a requisite for the cytoprotective activity of WN1316 *in vitro*. These findings combined led us to hypothesize certain serum factors being involved in exerting the WN1316-mediated anti-oxidative cell death activity. In this study, we have successfully identified two requisite serum factors, AHSG and HPX, in FBS for the WN1316-mediated cytoprotective activity, demonstrating direct proof for the existing of such factors. Additionally, the human homologues of the glycan free AHSG and HPX proteins synthesized *in vitro* worked as well. Thus far, there has been no report showing the evidences that AHSG and HPX are directly and/or indirectly involved in neuroprotection in the CNS. AHSG, commonly referred to as fetuin-A, is a carrier plasma glycoprotein produced in the liver. AHSG is known to be a multi-functional protein and interacts with various proteins, such as transforming growth factor-β (TGF-β), bone morphogenetic protein (BMP), cytokines, and insulin receptor. Among these functions, it is noteworthy that AHSG plays a role in membrane trafficking. Since AHSG binds to lipids and is internalized into cells via endocytosis, it is possible that endocytosis is the mechanism of action involving in WN1316 uptake into cells. Other candidate protein, HPX, is a plasma glycoprotein with a high heme binding affinity, which is mainly produced in the liver but also in the CNS and the peripheral nervous system. Interestingly, the heme-HPX complex is internalized via low-density lipoprotein receptor-related protein 1 (LRP1)-mediated endocytosis into cells. Several studies have demonstrated that LRP1 is highly expressed in brain microvessels, and can transport several ligands across the BBB. Remarkably, it has been reported that the transmembrane LRP1 is expressed in SH-SY5Y cells, which were used in the present study, and plays a crucial role in the cytotoxicity of amyloid-β oligomers. These findings suggest that the HPX molecules are, at least in part, implicated in delivering WN1316 into SH-SY5Y cells via the LRP1-mediated transport system. One important question arising from this study is as to how AHSG and HPX deliver WN1316 into cells. Thus far, it has not yet been demonstrated whether WN1316 can form the complex with AHSG or HPX. However, since the WN1316-mediated cytoprotective activity is totally dependent on the presence of either purified AHSG or HPX, it is conceivable that these specific serum factors directly bind to WN1316 and act as carriers for transporting WN1316 to the target site within cells. Since the degree of WN1316-mediated activation by purified AHSG or HPX was lower than those by FBS, it is still possible that additional factors are needed to fully exert the WN1316-meidated activity. One alternative explanation is that some enzymatic activities present in FBS, which are mediated by factors other than AHSG and/or HPX, increase the WN1316-mediated cytoprotective activity by modifying the chemical structure of WN1316. However, since the HPLC-detected peak retention times of WN1316 in the brain tissues and serum were same as that in the purified compound, such metabolic effects might be marginal if existed. Nonetheless, when taken into account the fact that the effective concentration of WN1316 *in vitro* is extremely high when compared *in vivo*, it is reasonable that either other non-serum factors or serum factors are not present in FBS, or both of them play important roles in WN1316-mediated neuroprotection. Further investigations on the mode of interaction between WN1316 and AHSG or HPX, the route of the internalization of WN1316 mediated by AHSG and HPX, and the cooperative functional interactions of AHSG or HPX with other endogenous factors are required. Another question arising from this study includes whether AHSG and HPX act specifically on WN1316. WN1316 is the only compound implicating in the serum factor-dependent cytoprotective function, which we have identified thus far. However, it does not necessarily mean that WN1316 is a sole amino imidazole derivative possessing such function, since we have not conducted a comprehensive screening of the compounds based on the AHSG/HPX-dependent cytoprotective activity yet. Future studies will uncover the relationship between the chemical structural feature of the small molecule compound and the AHSG/HPX-mediated activity. At present, one of the most promising approaches for drug delivery to the brain is to utilize the receptor-mediated transcytosis (RMT) system. Although it has been tried to utilize various receptors that are expressed in the BBB for the RMT system, the efficiency of CNS drug delivery via RMT depends on the expression level and distribution of target receptor in the BBB. Notably, utilization of ubiquitously expressed receptors for BBB transport causes incorrect delivery of CNS drug to other tissues and organs and could induce undesirable side effects. Therefore, discovery of the new BBB delivery route will give additional choices of delivery methods in terms of selectivity, capacity, and immune tolerance of CNS drugs. Although it is currently still unknown whether AHSG, HPX, and/or other factors are required to exert the WN1316-dependent neuroprotective activity *in vivo*, the administration of WN1316 can, indeed, alleviate the disease progression in ALS mouse models without any adverse side effects or complications. It is plausible that the AHSG/HPX-mediated drug delivery could become an alternative means to treat the CNS diseases. Future pharmacological studies on WN1316 in conjunction with the use of AHSG/HPX conditional knocked-out animals will provide a novel alternative drug delivery pathway and the improvement of the cell membrane permeability of the CNS therapeutic agents. In conclusions, we identified two glycoproteins, AHSG and HPX, that played a key role in the WN1316-mediated cytoprotective activity *in vitro*, suggesting that they could be ever-unidentified endogenous factors implicating in neuroprotective drug efficacy. Since WN1316 is a high BBB permeable small molecule compound and a promising therapeutic candidate for ALS, we have completed the first-in-Man clinical test of WN1316 and shown to be safe in healthy adult persons (unpublished data). To further proceed the clinical trials, it is extremely important to elucidate pharmacodynamics of WN1316 *in vivo* for its safer therapeutic application to the CNS diseases. # Supporting information We thank Kyo Wakasa at Neugen Pharma and Akihisa Mitake at CMIC Pharma Science for their invaluable comments throughout this study. [^1]: TK, KY and KT were employees of Neugen Pharma at the time this research was performed. JEI owns stocks or shares in Neugen Pharma. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials. SH declares that no competing interests exist. [^2]: Current address: CMIC Pharma Science Co., Ltd., Hokuto, Yamanashi, Japan

# Introduction Approximately one in six U.S. children age three to 17-years have a reported developmental disability. While clinically common, there is a remarkable number of genetic abnormalities known to be causal for developmental disorders (DD) that include autism spectrum disorder (ASD) and intellectual disability (ID). Some conditions, such as Down syndrome due to trisomy 21, are relatively straightforward when it comes to a genetic diagnosis, but others are not. Results from a number of studies, for example, suggest ASD is between 56–95% genetic in origin, while another recent paper reported that 552 genes are implicated in ASD–a number that will increase as more genes are identified. In recent years, these findings have led to the recommendation that genetics evaluations be incorporated into clinical practice guidelines for individuals with developmental disabilities. In 2010, the American College of Medical Genetics (ACMG) recommended chromosomal microarray (CMA) as the first-tier diagnostic genetic evaluation when investigating patients with DD, ASD or ID. The American Academy of Pediatrics recommendation that CMA be considered as part of the initial diagnostic evaluation of patients with disorders of postnatal development including DD and ASD, was superseded by a formal guideline issued in 2014 clearly designating CMA as the first-line test, replacing the standard karyotype and fluorescent in situ hybridization tests for the child with intellectual disability of unknown etiology. These recommendations are mirrored by similar guidelines from the American Academy of Neurology and Society of Children Neurology, and the Academy of Child and Adolescent Psychiatry. A confirmed genetic diagnosis means that associated life-threatening medical risks (seizures, tumors, renal disease, heart disease) can be better anticipated and managed, specific medications and therapies administered, and contraindicated treatments can be avoided with a predictability that was not possible before CMA testing. Despite clear recognition of these new technologies and the clinical insights they provide, genetic testing likely remains underused. This means fewer specific diagnoses are made and corresponding appropriate treatments offered. Despite uniform opinion amongst the relevant societies as to the importance of CMA, a variety of CMA tests are currently on the market, which differ distinctly in their genomic coverage, both qualitatively and quantitatively, and the clinical interpretations which can be made from the various tests. This impacts the breadth of genetic detection and, ultimately, the clinical utility that testing offers to the patient and the provider. We undertook a large-scale, experimental study to assess the clinical utility of CMA testing in general and a new 2.8MM probe-CMA test (FirstStep^Dx PLUS; Lineagen, Inc.), which was developed consistent with ACMG Guidelines on CMA design, through the addition of 88,435 probes of content validated to have sound clinical correlates with ASD and other ID/DD associated genes. The 2.8MM probe-CMA test has a higher overall detection rate than currently available CMA tests and includes communication of test results to ordering physicians, and to families, via a comprehensive report that includes clear and actionable recommendations. The goal of this study was to determine the overall evidence- based practice changes that occur when a CMA test was ordered, as well as the specific impacts on diagnostic and treatment quality. This analysis is unique in that it overcomes the enormous practical problem of identifying and aggregating uncommon genetic conditions by using simulated patients to determine clinical utility. This approach allows us to look at the variability of clinical practice–what clinicians do—while holding constant the variability of (hard to find) patients with individually rare disorders. # Methods ## Overview We conducted a longitudinal randomized controlled study of practicing general and sub-specialty-trained pediatricians from across the U.S. without prior experience using the 2.8MM probe-CMA product. Using a “before and after” design, pediatricians were asked to care for online simulated patients, using Clinical Performance and Value (CPV®) vignettes via web-based interactive sessions. Physicians were randomized into one of two study arms: control and intervention. Physicians completed randomly assigned vignette cases at baseline (Round 1) and again 4 weeks later (Round 2). This study, which ran from August 2014 to January 2015, was conducted in accordance with ethical standards and approved on June 16, 2014, by the Chesapeake Institutional Review Board (IRB), Columbia, MD. All participants provided written consent to participate. The trial was voluntarily listed in clinicaltrials.gov (Identifier: NCT02414438). Due to the brevity of the trial, the voluntary nature of the registration (studies not involving drugs, biologics, or devices are not required to register) and a clerical error, enrollment for this trial was underway before being logged in the registry. Notwithstanding, IRB approval was secured prior to conducting the study. The authors confirm that all ongoing and related trials for this intervention are registered. ## Eligibility and Selection of Physicians Recruitment of pediatrician participants was carried out by a mail campaign directed at a random subset of nationally representative lists of 25,000 general and 5,000 sub-specialty practice physicians. Recruit ran from August to September 2014, and letters were sent to 1,000 physicians randomly selected from the lists, explaining the purpose of the study and inviting physicians to participate. Eligibility for participation in the study was assessed using a physician questionnaire. Physicians had to be (1) either board-certified pediatric neurologists, developmental pediatricians or general pediatricians, (2) English-speaking, (3) practicing in a community/non-academic based setting, (4) accessible by Internet, and (5) have no prior experience with the 2.8MM probe-CMA test. Additionally, physicians who met the initial eligibility requirements were required to complete both a physician survey and two rounds of cases to remain in the study. In total, 232 respondents fulfilled the eligibility criteria and consented to participate. These physicians were randomized into control and intervention groups, using a 1:2 control to intervention ratio. Assignment into the arms was based on simple random allocation generated from Microsoft Excel’s random function. Between Rounds 1 and 2 of data collection, physicians in the intervention arm received a 25-minute webinar on the 2.8MM probe-CMA test and were provided with the 2.8MM probe-CMA test results specific to their patient if they chose to order and/or view the results in Round 2. ## Clinical Performance and Value® Vignettes Clinical utility was measured at baseline (Round 1) and again after four weeks (Round 2) using CPV® vignettes for both groups. These vignettes are a widely- used, validated means for assessing differences in clinical practice and inherent variation in care, independent of case. In each vignette, the general and sub-specialist pediatricians were asked to provide open-ended responses regarding the clinical care they would provide for the simulated patient. Responses were scored in 5 domains: taking a medical history, performing a physical examination, ordering appropriate laboratory and imaging tests, assessing disease activity and physical function, and prescribing treatment. Explicit quality of care scoring criteria were established prior to study administration and were derived from the peer-reviewed medical literature and expert opinion by a physician board certified in Medical Genetics and Pediatrics in conjunction with a panel of clinical physicians. All participant responses were scored against these explicit scoring criteria by clinical physicians blinded to the participant’s identity. Overall CPV scores and scores with each domain represent the percentage of physician answers that addressed each of these evidence-based criteria, and range from 0 to 100%. CPVs have been linked to outcomes data in other studies, wherein a 3–5% increase in scores is associated with an improvement in outcomes and is the threshold used for clinical significance. ## Cases The vignettes used in this study simulated clinical encounters involving simulated pediatric patients presenting with an atypical phenotype and/or developmental or intellectual deficits. Patient vignettes in this analysis were developed around five clinical syndromes describing a range of abnormal phenotypes with underlying genetic diagnoses causal of developmental or intellectual delay. Attention was given to conditions in which an accurate diagnosis would lead to change in treatment. These cases were used to evaluate the impact of CMA testing and the 2.8MM probe- CMA testing on patient diagnosis and management. The order of the cases was randomly assigned. To assess physician, patient and practice characteristics, as well as physicians’ familiarity and experience with genetic testing, we also administered a questionnaire to all physicians prior to baseline. All participants in both rounds could order and receive results from existing CMA tests; provision of the 2.8MM probe-CMA test results was restricted to Round 2 intervention group physicians. CMA testing options available to the physicians included 180K array, high resolution CMA, and the 2.8MM probe-CMA test. The five cases comprised three groups for CMA testing results: the first category of cases yield similar results regardless of the CMA selected; the second yields normal (i.e., no abnormality) or incomplete results for the 180K test, but abnormal results (i.e., genetic findings) using high resolution CMA and the 2.8MM probe-CMA test; and the third group yields normal 180K and high resolution CMA results but the 2.8MM probe-CMA test detects the abnormality. ## The 2.8MM Probe-CMA Testing and Reporting Process The 2.8MM probe-CMA test is a whole-genome CMA that is optimized for clinical use in the setting of DD/ID and ASD in part through the addition of an additional 88,435 probes in genes/regions documented in the clinical literature to be associated or causative of these conditions. Compared to other CMA testing platforms, it has a two-fold increase in detection rate for variants that underlie ASD (5–7% on typical CMA platforms to 12–14% on the 2.8MM probe-CMA test), and an overall increase in detection rate from 15–20% to over 30% for all suspected DD/ID cases including ASD. The 2.8MM probe-CMA test results are provided in a detailed report that includes both physician-specific and parent- specific sections with a clear description of the underlying abnormality, suggestions for medical management and/or further testing, and recurrence risk. The increased detection rate and resulting clinical management recommendations from Lineagen’s 2.8MM probe-CMA test should increase the clinical utility of CMA testing when this format is used. ## Intervention Before Round 2 data collection, subjects in the intervention arm were given an educational component describing CMA testing in general and the 2.8MM probe-CMA test specifically. This was a 25-minute web-based pre-recorded video, which participants watched at their convenience and provided an overview of genetic testing for ID/DD and described the 2.8MM probe-CMA test. Nearly all (138 of 149 intervention arm participants) confirmed viewing the intervention video. After viewing the short video, the intervention group completed three additional (randomly) administered cases and now had the opportunity to order and review the 2.8MM probe-CMA test results if they chose to do so. Physicians in the control arm also completed a second round of cases, but did not receive any information about the 2.8MM probe-CMA test and could not order or receive the 2.8MM probe-CMA test results. ## Analysis First, we compared the intervention and control groups, and summary statistics were assessed for baseline provider and practice characteristics. For continuous and interval data, we used an independent samples t-test to determine the p-value. For categorical data, Fisher’s exact test was used to compare the distribution of the two groups. We then tested the main hypothesis: Does CMA testing in general and/or the 2.8MM probe-CMA testing specifically improve clinical utility? We measured changes in clinical utility using quality of care scores from the CPV® vignettes (e.g., ordering the proper diagnostic work, making appropriate referrals), which quantify the percentage of pre-determined evidence-based care criteria which were appropriately addressed by the physician while caring for the virtual patient. We further hypothesized that the pediatricians who ordered CMA, including the 2.8MM probe-CMA test, were more likely to improve their diagnostic accuracy and deliver specific treatments that were appropriate for the patient’s underlying condition. To test this, we estimated the intervention’s effect by performing a logistic regression and modeling the marginal effects for the specific diagnosis or treatment option of interest. We compared the change (differences) in the outcome scores between Rounds 1 and 2 for both the intervention and the control groups. Univariate regression models using physician and practice characteristics were performed. Our *a priori* assumption was that all physician and practice characteristics would be of interest in the final model, and thus, our decision was to include all these variables in the final model, regardless of significance in the univariate model. After including these variables, we added in the relevant CMA, round, and intervention variables. 2.8MM is the explanatory variable describing those who ordered the 2.8MM probe in Round 2, and CMA is an explanatory variable for those who ordered CMA testing (other than the 2.8MM probe-CMA test) in either round. The 2.8MM variable measures the utility of the 2.8MM probe-CMA testing after its introduction in Round 2. We also analyzed the interaction of ROUND x CMA, to measure the differential change in the utility from CMA testing in Round 2 versus Round 1 and a possible learning effect spillover from the 2.8MM probe-CMA test to all CMA testing. The interaction terms between the 2.8MM probe-CMA test x Group 2 and the 2.8MM probe-CMA test x Group 3 tells us the differential effects on cases types of the 2.8MM probe-CMA test compared to any other CMA tests. <img src="info:doi/10.1371/journal.pone.0169064.e001" id="pone.0169064.e001g" /> Y j t = Vignette score at time t , physician j Y j t ∼ N o r m a l ( μ j t , σ e 2 ) μ j t = β 0 \+ β 1 CMA \+ β 2 ( 2.8 M M ) \+ β 3 ROUND × CMA \+ control variables # Results ## Provider Characteristics After randomization, there were 147 participants in the intervention arm and 85 in the control arm. Ultimately, 202 physicians completed both rounds of data collection from September 2014 through January 2015, the predefined study period. 30 participants were dropped from the study (22 from the controls and 8 from the intervention arm) because they were determined to have not met initial inclusion criteria (n = 5), asked to be withdrawn from the study (n = 3), or did not complete the additional study requirements (n = 22). There was no significant difference between the control and intervention groups by provider characteristic. Overall, the majority of the pediatricians were part of a hospital/health care system (61%). Over 90% were employed by their practice. 59% were pediatric specialists and 41% were general pediatricians. Practice characteristics, particularly as they relate to genetic testing, were varied. On familiarity with genetic testing, the average score was 2.9 on a scale of 1 to 5 and 44% rated themselves a 3. However, 85% felt that a geneticist was readily accessible to them if they needed consultation. Their patients were almost evenly split between commercial and public insurance patients (mostly Medicaid) and their practices were active with more than 58% of respondents reporting seeing over 50 general pediatric patients per week, while 30% reported seeing more than 20 patients with developmental disabilities per week. ## Ordering CMA Diagnostic Testing Overall 36% of all physicians ordered CMA testing of any kind in Round 1 of the study and 44% ordered CMA testing in Round 2. Not surprisingly, almost no physicians in any of the study arms ordered the 2.8MM probe-CMA test in Round 1 (n = 5 or 1.1% of the sample) a pattern which persisted in the control arm in Round 2, where only 1 physician ordered the 2.8MM probe-CMA test. If the 2.8MM probe-CMA test was ordered in Round 1 or by a control arm physician in Round 2, the 2.8MM probe-CMA test results were delivered. Almost all of the increase in CMA testing in Round 2 came from physicians in the intervention group ordering specifically the 2.8MM probe-CMA test, where 20% (n = 60) of physicians ordered the test. ## Link between Diagnostic Testing and CPV® Vignette Scores CMA testing was associated with improved medical management, taken here as a surrogate for clinical utility. Those that ordered CMA testing had overall CPV quality scores 5.4% higher (p\<0.001) than those that did not, indicative of clinically significant improvements in medical management. This difference did not change, in either magnitude or significance, in the second round of testing. However, the 2.8MM probe-CMA testing, which was only available in Round 2, resulted in CPV quality scores 7.2% higher than cases in which CMA was not ordered. This represents an additional 1.8% improvement in quality of care over cases in which other CMA tests were ordered, which is statistically significant (p\<0.001). We looked at the three clinical groups detailed in to determine if this improvement was from the additional clinical information provided by the 2.8MM probe-CMA test. By design, the 2.8MM probe-CMA test would provide similar diagnostic information as other CMA tests in Group 1, a more specific diagnosis compared to the 180k test in Group 2, and a more specific diagnosis compared to both the 180k and high resolution CMA in Group 3. A regression model including the interaction of the 2.8MM probe-CMA test with Groups 2 and 3 confirmed that overall CPV scores increased by 10.9% in Group 2 and 8.7% in Group 3 CPVs, compared to Group 1 (p = \<0.001 for each Group). Two additional covariates were also associated with higher scores: generalist physicians (p = 0.008) and female physicians (p = 0.003). For the five cases where the 2.8MM probe-CMA test offered a specific diagnosis, this was also associated with a statistically significant, 10.9% greater accuracy (p\<0.001) for combined diagnosis and treatment scores. This difference accounted for the majority of overall improvement in clinical utility noted in. Conversely, other CMA testing (defined as 180K, regular or high resolution CMA, but excluding the 2.8MM probe-CMA test) was not linked to significantly higher scores in the combined diagnosis and treatment domain, showing an increase of only 2.7% in the diagnosis and treatment scores (p = 0.122). In this analysis, neither covariate of generalist physician or gender of physician was associated with a significant increase in the combined score. However, those physicians who did not own their practice were more likely to accurately treat and diagnose their patients (p = 0.001). The effect of using the 2.8MM probe-CMA test was strong and positive in both Group 2 and Group 3. Finally, we completed a secondary analysis, conducting a case-by-case evaluation of diagnostic and treatment practice changes associated with more accurate the 2.8MM probe-CMA test diagnosis. Despite the lack of statistical significance, the disaggregated data were directionally suggestive of several benefits. For two out of the five conditions (Hunter syndrome and mosaic Turner syndrome), the 2.8MM probe-CMA test led to a statistically higher rate of accurate diagnosis, as seen in the improved diagnosis scores in the intervention group. Correct secondary diagnoses, such as gross motor delay for Hunter syndrome and uncontrolled myoclonic seizures for Dravet syndrome, were also more likely to be made for four of the five conditions. For almost every condition studied there were some statistically significant differences in treatment. Examples of these treatment differences include: evaluation and therapy for gross motor delays (Hunter syndrome); anticipatory guidance and screening for heart disease (Turner syndrome); recommending early cognitive intervention (FOXG1-related disorder); referral to child neurologist and commencement of neuroleptics (Dravet syndrome). Many other treatments were directionally positive but not statistically significant and thus not causally established. However, in the Turner syndrome case, one important treatment that showed there was no difference between the intervention and control groups in either ordering or discussing the performing a prophylactic gonadectomy, despite the risk of a gonadoblastoma in these patients. # Discussion Appropriate assessment of new health care interventions, particularly diagnostic technologies, requires high quality information on the clinical utility, economic value, affordability, and population health benefits of the technology in question. As a reflection of this, assessing clinical utility—demonstrating the usefulness of a test in clinical practice—is now a significant hurdle required by payors before offering insurance coverage for most new technologies. As government and commercial payors struggle to make decisions regarding reimbursement, demonstrating that additional clinical information is useful and leads to changes in what a provider does is critically important for any vulnerable population. Developmental delays, including autism spectrum disorder, have both a common phenotypic expression and are a reflection of myriad different genetic conditions, each with its own clinical characteristics that require unique therapeutic interventions. Assessing these patients clinically, however, is time consuming and difficult, leading to missed diagnoses for thousands of children every year. To overcome this, professional societies recommend that genetic testing, specifically CMA, be done as first tier diagnostic evaluation of these conditions. We sought to clarify the evidence around the utility of CMA testing for patients with developmental delay associated with five specific genetic copy number variants (CNVs) detected by CMA testing. We did this by using an innovative measurement tool, CPVs, which controls for patient case mix, overcomes the problem of collecting an adequate number of patients with a specific (rare) genetic disorder and allows us to conduct a randomized controlled experiment of related physician behavior and patient management. CPVs are a validated measure of clinical practice that has been widely used to serially measure clinical performance. The tool has proven particularly valuable in clinical utility studies determining whether new technologies change clinical decision making and associated spending. Using this tool, we examined whether CMA testing in general, and the 2.8MM probe-CMA test testing specifically, changed provider practice for five different patient genotypes. We found that CMA testing did indeed improve the quality of care and clinical decision making among more than 200 general and sub-specialty pediatricians. When physicians used the 2.8MM probe-CMA test, an optimized whole-genome high resolution CMA, we found an increased causal relationship between testing and improved clinical utility. Physician users of the 2.8MM probe-CMA test had a statistically and clinically significant increase of 7.2% in their overall utility scores over those physicians who use no CMA testing at all and a 1.8% marginally higher score over those who use standard CMA testing. The effect of the 2.8MM probe-CMA test is even more notable on diagnosis and treatment scores, arguably the most critical part of the care process. Scores of the 2.8MM probe- CMA test users for the combined domain score for diagnosis and treatment were 10.9 percentage points higher than those who used no CMA testing at all. The 2.8MM probe-CMA test was particularly useful in individuals who harbor a pathogenic CNV in regions covered by the custom probes contained on that CMA to optimize clinical interpretation in cases of DD, ASD and ID, including but not limited to the genes for GAMT deficiency (GAMT) and congenital Rett syndrome (FOXG1-related disorder). Most of the commercially available CMA platforms cannot detect these, and other genetic abnormalities targeted by the 2.8MM probe-CMA test. The increased clinical utility of the 2.8MM probe-CMA test also extends to individuals affected by syndromes such as Mosaic Turner syndrome with an XY cell line and Dravet syndrome, which can be diagnosed by some other CMA tests available on the market. We hypothesize that this increase in clinical utility is driven by the comprehensiveness and clarity of the 2.8MM probe-CMA test report, which includes specific treatment recommendations. We observed opportunities to improve the quality of care delivered to all patients with DD/ID and ASD. At baseline, only 37% of physicians ordered any CMA for their patients. Following the addition of the 2.8MM probe-CMA test, after providing a simple intervention, CMA testing increased significantly to 44%. CMA testing at baseline was surprisingly low suggesting that professional medical guidelines are either not widely known or are not used. By analyzing the specific aspects of care provided to these patients, we found that the 2.8MM probe-CMA test provided benefits in many areas such as identifying the primary diagnosis of the DD/ID syndrome, managing secondary diagnoses, and defining treatment. The treatment improvements led to instances of proper treatment and presumably earlier therapy. Indeed, successful management of ID/DD/ASD is predicated on recognition of any underlying syndrome. Other studies will be needed to determine if earlier, more precise, diagnosis from CMA testing leads to less unnecessary testing or less testing overall. This study has several limitations. The control group received neither orientation on genetic testing nor the 2.8MM probe-CMA test results, thus it is difficult to tease apart the effect from the 2.8MM probe-CMA test results from that of the educational intervention on genetic disorders and testing. Second, the five diseases that were the focus of the CPVs are each relatively rare like most DD, ASD and ID syndromes, thus results here may not necessarily be extrapolated to all underlying syndromes. We do suspect however that regardless of the syndrome, the challenges of arriving at a genetic diagnosis and developing an appropriate therapeutic plan for the individual are common across a number of disease types. Thirdly, this study did not address issues related to variants of unknown significance, additional secondary possibly actionable variants, or point mutations necessitating genomic sequencing approaches to detect. As genetic tests become increasingly detailed and complex, these issues will benefit from future follow-up work. Finally, this analysis focused on CMA testing, which in general will detect cases of these conditions due to larger deletions and duplications, not point mutations. Future studies of this type will be needed to address the contribution of point mutations by emerging next- generation sequencing technology. Studies to determine the clinical utility and impact of diagnostics and interventions for rare diseases are particularly challenged by the difficulty in locating and identifying an adequate number of patients for study. Those who attempt to undertake assessments of clinical utility using patient level data are additionally challenged by difficulty controlling for case mix which is a routine issue in any patient level study. This challenge is further compounded when dealing with rare diseases and small numbers of study participants. The CPV vignettes provide a standardized method to analyze how physicians with different backgrounds and from various settings assess and treat the same patient. While the CPV approach does not use actual patient level data, they uniquely provide a standardized method to analyze how physicians with different backgrounds and from various settings would assess and treat the same patient. Our results show that CMA in general and specifically the 2.8MM probe-CMA test provides improved medical management in children with developmental delay, and therefore meet criteria for clinical utility. # Supporting Information Jeri Harashima, MS, LCGC, provided help in enrolling physicians into the study. [^1]: The authors have the following interests: MM, MP, RV, and EBW are employees of Lineagen, Inc; LD, JF, DP, and TB are all employees of QURE Healthcare. JP is the President of CPV Technologies, which owns the CPV IP used in the study. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: **Conceptualization:** JP MM MP. **Data curation:** JF DP. **Formal analysis:** JF DP. **Funding acquisition:** MP RV ERW. **Investigation:** JP LD TB. **Methodology:** JP LD JF. **Project administration:** MM LD. **Resources:** JP MP RV. **Software:** JP LD. **Supervision:** JP MP RV ERW. **Validation:** JP DP. **Writing – original draft:** JP MM LD MP RV. **Writing – review & editing:** JP MM LD JF DP MP RV ERW TB. [^3]: ‡ These authors also contributed equally to this work.

# 1 Introduction According to data from World Health Organization, it is estimated that at least 2.2 billion people have visual impairment or blindness, of which at least 1 billion have some visual impairment that could have been avoided or has not yet been addressed. The five leading causes of blindness are: uncorrected refractive errors, cataracts, age-related macular degeneration, glaucoma, and diabetic retinopathy. Age-related macular degeneration (AMD) is an eye disease that affects people over 50, being one of the main causes of visual impairment in elderly patients. It is a progressive and degenerative condition, resulting from aging, related to changes in the structure of the retina, causing symptoms of loss of central vision in one or both eyes due to photoreceptor dysfunction or loss. It can be presented in two forms, the most common being dry or non-exudative, identified in 85 to 90% of cases. This form is characterized by the presence of drusen and/or pigmentation alteration and causing slower visual loss. The neovascular or exudative form occurs in 10 to 15% of cases, however, it is responsible for 90% of blindness due to AMD. Druses are extracellular materials accumulated between the retinal pigment epithelium (RPE) layer and the Bruch’s membrane layer (BM) layer. The presence of a certain number of drusen is normal with advanced age. However, the abundant presence of drusen is a common initial indicator of AMD. The U.S National Library of Medicine (NLM) estimates that AMD affects about 170 million people worldwide. The prevalence rate is expected to increase in the coming decades, as the proportion of elderly people in the population increases. The estimate for 2040 is that 288 million people will be affected by the disease. The clinical diagnosis of AMD is commonly made through retinography (fundus photograph), however, optical coherence tomography (OCT) has been widely used in conjunction with retinography. OCT has emerged as a viable option for diagnosing retinal changes. It is a non- invasive imaging technology that allows the visualization of biological tissues in vivo in sectional sections with high resolution. Thus, it soon gained great importance in the diagnosis of macular diseases such as AMD. Druses, neovascularizations, fibrovascular scars, and intraretinal or subretinal macular fluid are typical changes in the retina that are detectable in optical coherence tomography. The optical coherence tomography image allows the evaluation of changes in the retinal pigment epithelium (RPE) and, hence, of drusen. Compared to the other recommended tests, it was evidenced that the OCT is more sensitive to the detection of pathological changes related to AMD. It was found that 57% of the evaluated eyes showed changes perceived only through OCT and not seen through other exams, such as angiofluoresceinography. Another advantage of retinography is the rapid acquisition, eliminating the use of contrasts and thus the risks of allergic reactions. Early diagnosis of AMD is helpful for monitoring the disease progression and for complications including the onset of wet AMD. When this diagnosis is made from the OCT, it takes place with the visual analysis of the various images generated, and if we take into account that a specialist attends and accompanies several patients per day, technological support for clinical practice is necessary to minimize evaluation errors. Therefore, in this work, we focus on the segmentation of the limits of the ILM, RPE, and Bruch’s membrane of OCT B-scan images of AMD and normal patients. There are two reasons for the focus on these layers. The first reason is that, as the ILM layer is the upper limit of the retina and the BM layer is the innermost limit of the retina, the physiological aspects and morphological structure of the retina can be altered by AMD being detected and visualized in these layers. The second reason is that drusen are formed between RPE and BM, so the boundaries of the RPE and BM layers can indicate data such as size, shape, location, and quantity of drusen. Therefore, we verify the importance of segmentation of these layers to support the detection and diagnosis of AMD. Given this scenario, this article proposes an automatic method of segmentation of the ILM, RPE, and BM layers of the retina in OCT B-scans images. The proposed method is carried out in four stages: (a) pre-processing (using the bilateral filter); (b) definition of the area of interest (using various image processing techniques); (c) initial segmentation (based on U-net, deep neural network (CNN) model); (d) final segmentation (based on DexiNed, deep neural network architecture for edge detection). Among the contributions of this work, we highlight the proposal of a robust methodology capable of segmenting the layers of the retina in a fully automated manner. This article is organized into 5 sections. Section 2 describes the proposed method. Section 3 presents the results, Section 4 discusses the results. Finally, Section 5 presents a conclusion about the method developed. ## 1.1 Related works Segmentation of the retina’s boundary layers in OCT images is a challenging task since it is a complex image that presents a large amount of noise and deformations caused by pathologies (drusen, drusenoid PED, geographic atrophy) that change the characteristics of the edges of the layers. Several studies have been published demonstrating methodologies for segmenting the edges of retinal layers. In a method is presented for segmentation of the total retina, which lies between the internal limiting membrane (ILM) and the retinal pigment epithelium (RPE), and the RPEDC, which lies between the RPE and the Bruch’s membrane. To validate the method, 220 B-scans extracted from 20 volumes are used. All patients had age-related macular degeneration (AMD) in the intermediate stage and with druses, and some with geographic atrophy. The method is based on graph cutting. As a pre-processing step, noise reduction is performed through a rectangular mean filter. From the resulting image, a graph is constructed with the pixel values as weights. The search for each edge is performed sequentially. Considering all layers, the algorithm showed an average difference of 0.95 pixels, a value even smaller than the difference in segmentation between two specialists. Together with the one developed by, these works refer to the development of the DOCTRAP software (Duke OCT Retinal Analysis Program) designed for use in OCT segmentation research. In, a framework is presented that combines convolutional neural networks (CNN) with graph search algorithms to segment nine retinal edges in OCT images. The method was validated with 60 volumes (2915 B-scans) of twenty eyes from people with non-exudative AMD. The CNN is trained with the characteristics of the layers’ edges to estimate the positioning of the edges of the eight layers. Then, these values are put through a graph search algorithm for the final definition of the boundaries. The results found were compared to the results obtained by the DOCTRAP and OCTExplorer segmentation software. These results were superior to those obtained by the OCTExplorer software but still inferior to those obtained by the DOCTRAP software. presents a method for segmenting the edges of retinal layers in OCT images. Three public databases () were used, two with a total of 210 B-scans of patients without a pathology and the other with 220 B-scans of patients with intermediate AMD. Two models based on the combination of Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) are created, 1 to segment eight retina layers and another to segment three layers. CNN is used to extract a region of interest and detect the layers’ edges, while LSTM is used to track the layer boundary. This model is trained with a mix of normal and AMD cases. The results show an average absolute error (EMA) in pixels of 1.30±0.48 *μm*, less than the error of marking the bases of 1.79 ± 0.76 *μm*. In, a new linearly parameterized conditional random field model (LP-CRF) is proposed to segment the retinal layers’ edges OCT images. Two public databases () were used, one with 107 B-scans from patients without pathology and the other with 220 B-scans from patients with intermediate AMD. The proposed LP-CRF comprises two convolution layers to capture each region’s appearance and layer boundary, the relative weights of the previous form, and an additional term based on the similarity of appearance of the adjacent boundary points. All types of energy are learned in conjunction with a Support Vector Machine. The proposed method segments all the retina’s limits in a single step, and for the images without pathology, eight edges are extracted and with AMD three edges. The mean absolute error reached was 1.52±0.29 *μm* pixels for the segmentation of 8 boundaries in the Normal data set and 1.9±0.65 *μm* pixels for three edges in the combined AMD and Normal data set. proposed a method for segmenting the edges of retinal layers in OCT images. They used two image databases, one for pediatric patients without a history of pathologies and intermediate AMD patients and pathologies. The method used a trained recurrent neural network (RNN) as a fragment-based classifier to segment seven layers boundaries in the first base and three edges in the second base. The results indicated that the RNN architecture is a viable alternative to CNN for image classification tasks that exhibit a clear sequential structure. Compared to CNN, RNN showed a slightly smaller absolute mean error. A method for segmenting the edges of the ILM, RPE, and BM layers is presented in. The method was tested on two public databases with normal patients and with intermediate AMD, one provided by with 384 volumes, the other account with 19 volumes made available by. The deep neural network architecture, capsule network, was used to do the initial segmentation of the edges of three layers of the retina, where it classifies fragments extracted from OCT images in the four classes (ILM, RPE, BM, and fundus). To refine the segmentation, the author made use of the graph cut technique, which is applied to the neural network probability maps. As a result, the method reached, respectively, for the two bases, 0.75 and 2.04 average of the mean absolute error for the ILM, 0.93 and 1.97 for the RPE, 1.09 and 2.05 for BM. In, a new method is presented to segment ILM, RPE, and BM in OCT images. Three image bases were used to train and validate the method, with healthy patients’ images and intermediate AMD. The authors use an ensemble method called deepForest, which uses several Random Forests to generate the classification of patches extracted from OCT images. As pre-processing, normalization of pixel intensities is applied. The patches are divided into four classes, internal limiting membrane, retinal pigment epithelium and druse complex (RPEDC), Bruch’s membrane, and fundus. After classification, a graph-based technique developed by the author is applied to refine segmentation. As a result, the method achieved average errors of 0.81, 1.35, and 1.23 pixels for the three bases, respectively. developed a method for segmentation of retinal layers (ILM, RPE, and BM) in OCT images based on the wave algorithm, a mathematical model of the equation of the potential fluid energy in fluid mechanics. The method uses the base provided by and uses the absolute mean error to evaluate the results. The method obtained an EMA of less than 1.5 pixels in all evaluations. In the above-mentioned works, we can highlight that everyone uses some graph search techniques to generate the final segmentation of the edges of the retinal layer. The method proposed in this work uses a deep neural network specialized in edge detection to make the final segmentation step. # 2 Materials and methods The proposed method for segmentation of retinal layers (ILM, RPE, and BM) in OCT images have four steps: pre-processing, the definition of the area of interest, initial segmentation, and final segmentation. shows the method flowchart. ## 2.1 Image dataset The OCT image base used in is composed of 269 volumes of dry AMD eyes and 115 control volumes (eyes considered normal). The images were acquired in four clinics by SD-OCT scanners manufactured by the company Bioptigen, Inc (Research Triangle Park, NC). The base is composed of individuals who met the following inclusion criteria: between 50 and 85 years of age, exhibiting intermediate AMD with large drusen (≥125 *μm*) in both eyes or large drusen in one eligible eye and advanced AMD in the other eye, without a history of vitreoretinal surgery or ophthalmic disease that may affect the accuracy of any eye. Control subjects were enrolled with the same inclusion criteria for the AREDS2 protocol, except that they should have no evidence of macular druse or AMD in either eye at the initial visit or in the follow-up years. The volumes were acquired in a rectangular region of 6.7 *mm* × 6.7 *mm* centered on the fovea with a rapid scan protocol, resulting in volumes of 1,000 × 512 × 100. The SD-OCT exams provided also provide the ground truth image of the borders that limit the total retina (between the internal limiting membrane and Bruch’s membrane), as well as delimiting the retinal pigmented epithelium associated with the drusen complex (when existing). The segmentation was initially performed in an automated way by the DOCTRAP software, developed by Duke University, and submitted to review and adjustments by specialists certified by the Duke Advanced Research in Spectral Domain OCT Imaging Laboratory. After segmentation, the specialists manually marked the central point of the fovea and defined as a region of interest in each volume, a cylinder centered on this point, with a diameter of 5 *mm*. ## 2.2 Pre-processing To attenuate the noise present in the images of the OCT exam and to highlight the edges, we used the bilateral filter, proposed in. It is a non-linear smoothing filter, which preserves the edges and blurs other regions. To reach this result, each pixel of the image is replaced by a weighted average of the neighboring pixels, considering the geometric proximity and its photometric similarity. The bilateral filter parameters used are: *d*= 15; *σ*_*c* = 100, *σ*_*s* = 100. We choose these parameters based on tests performed during the experiments. We chose the parameter from a predefined interval of values (d in, *σ*_*c* and *σ*_*s* in \[60, 100\]) and applied it to the next step for validation, choosing the value that leads to better results. presents an example of the result of applying this filter. ## 2.3 Definition of the region of interest This step has the objective of delimiting the image of each slice of the volume (B-scan) in which region the processing will be applied to identify the edges. This avoids unnecessary processing of regions that are very distant from the retina that represents the “background” of the image. The process for defining the region of interest is done in three stages. The first one consists of applying the segmentation method of Otsu to obtain an initial region. The second stage consists of the application of the morphological operations of opening (with 5 × 5 element and two iterations) and of expansion (with 5 × 5 element and seven iterations). Finally, to reach the final region, we use an elimination filter by area, where we exclude very small objects (1000 pixels). shows the final result of the process. ## 2.4 Initial segmentation This step aims to further reduce the search region for the next step and separate areas of the ILM layer from the region of the RPE and BM layers. For this, we use U-Net, a fully connected deep neural network architecture, developed for the segmentation of medical images. The network architecture has four downsampling layers with an activation function of the average type, where the average intensity of the window pixels is maintained. The transposed convolution is used for the enlargement stage. The convolution layers are followed by a batch normalization layer and a dropout layer, with a 20% chance of neuron deactivation. Finally, the segmentation is generated through a convolution (1 × 1), with a softmax activation function. The network is trained with the Dice Loss (S) loss function, as shown in. $$\begin{array}{r} {S = 1 - \left( \frac{2VP}{2VP + FP + FN} \right)} \\ \end{array}$$ The U-Net training is carried out with B-scan images resulting from the stage of defining the region of interest. The images have the ground truth for the ILM, RPE, and BM borders. Morphological dilation operations are applied to the ground truth images, so there are two larger areas for detection. Such operations were applied since the markings are very thin, and, in the process of downsampling, the network tends to lose such information. Therefore, the dilation process tries to increase the size of the ground truth’s markings, causing it to better maintain itself during the downsampling phases. The labels inserted to the network are the borders (ILM, RPE, and BM) and bottom. For the network’s input, we resized the images to a resolution of 128 × 128. To reach the image size used, we previously tested with various sizes (64 × 64, 128 × 128, 256 × 256, 512 × 512). We verified that the size used does not significantly impact the results. This happens because this step aims to reduce the search area for the next one, performing a rough segmentation. What is sought is that the resulting area contains the layers intended to be segmented in the final step. The resize is done using the smallest image size capable of generating a rough segmentation useful for the next step, which does not create any performance loss in the segmentation process. The computational effort aspect also influenced the smallest image size. With the reduced size, we were able to train the network with more images and thus have a greater diversity of cases in the set. The training process is shown in, in the form of flow, indicating the order of the steps. After the images are processed, they are submitted to U-Net for training, that is, the adjustment of the weights, so the network can correctly segment the pixels corresponding to the edges. ## 2.5 Final segmentation The DexiNed neural network was used in this work to generate the final segmentation of the edges of the layers which are the objective of the method. The architecture used was the same proposed in. DexiNed is a deep neural network architecture proposed for edge detection. It consists of a stack of filters that receive an image and provide a border map. DexiNed can be seen as two subnets: a Dense extreme inception network (Dexi) and the upsampling block (UB). While Dexi is fed with the image, UB is fed with resource maps for each Dexi block. The resulting network (DexiNed) generates thin edge maps, avoiding lost edges in deeper layers. shows the structure of the network, which is composed of an encoder with 6 main blocks. The network outputs present maps in each of the main blocks to produce intermediate edge maps using an upsampling block. All edge maps resulting from the sampling blocks are concatenated to feed the stack of filters learned at the end of the network and produce a fused edge map. All six upsampling blocks do not share weights. The blue blocks consist of a stack of two convolutional layers with a core size of 3 × 3, followed by batch normalization and ReLU as an activation function (only the last convolutions in the last sub-blocks do not possess this activation). Max pooling is used with a kernel size of 3 × 3 and stride of 2. As the architecture follows the learning process at various scales, an upsampling process is added (horizontal blocks in gray). As many convolutions are carried out, all deep blocks lose important edge resources, and only one main connection is not enough, as, from the fourth convolutional layer, the loss of edge resources is more chaotic. Therefore, since block 3, the output of each sub-block is the result of the average with the border connection (orange squares). After max pooling and before adding to the main connection, the edge connector is defined as the average of each output sub-block (see rectangles in green, bottom side). The output of each of the main blocks feeds an upsampling block that produces an intermediate border map, and the merging of each intermediate map will generate the final result. As an input to the network, we use the images delimited by the result of U-Net, thus reducing the space for searching and processing. In both training and testing, the input images were scaled from the original resolution (1000 × 512) to 992 × 128. Several tests were carried out until reaching such an image size. We realized that if the image’s width was reduced, the segmentation did not faithfully represent some curves present at the layers’ edges, especially in the RPE that presents more deformations caused by Druses and geographical atrophies. Hence, we leave the maximum width size since Dexined only works with multiples of 16. Three models were trained, one for each edge that we intend to segment (ILM, RPE, and BM). Each model was trained with 1000 B-scans. The network parameters were the same as those used in the proposed standard architecture. A special loss function is used in this network, it was proposed by and aims to adjust the network to a particular characteristic of the segmentation of edges, where 85 to 90% of the image are not edges. ## 2.6 Validation After the segmentation of the edges of the retinal layers, it is necessary to validate the results. This work used metrics commonly used in methods that aim to achieve this objective, the mean of the Mean Absolute Error and the standard deviation (std). MAE is a performance metric that measures the accuracy of the segmentation of the retinal layer, which means absolute distances between the edge of the segmented layer and the ground truth in each column of the image. The average value and standard deviation of the T OCT scan images (size of *m* × *n*) is calculated by Eqs and [3](http://3) $$\begin{array}{r} {mean(X,Y) = \frac{1}{T}\sum\limits_{t = 1}^{1}\left( \frac{1}{n}\sum\limits_{i = 1}^{n} \middle| X_{t}^{i} - Y_{t}^{i} \middle| \right),} \\ \end{array}$$ $$\begin{array}{r} {std(X,Y) = \sqrt{\frac{1}{T}\sum\limits_{t = 1}^{1}\left( \frac{1}{n}\sum\limits_{i = 1}^{n} \middle| X_{t}^{i} - Y_{t}^{i} \middle| {- mean(X,Y)} \right)^{2}},} \\ \end{array}$$ where *X* and *Y* are different segmentation results, $X_{t}^{i}$ is the position on the *y* axis of A-Scan in the *t* image of OCT by the *X*, $methodY_{t}^{i}$ is the position on the *y* axis of A-Scan in the *t* OCT image by the *Y*, $X_{t}^{i} - Y_{t}^{i}$ is the absolute positioning error on the *y* axis in A-Scan *i* and B-Scan *t*. ## 2.7 Training and testing environment All tests were implemented in the Python programming language with the help of the Keras deep learning library with TensorFlow-GPU as a backend. For image processing the *Opencv* library was used. For training and testing, a computer with an Intel I7 processor, 128 GB of RAM, 8 GB GeForce 1080, and Windows 10 operating system was used. The Unet model’s training lasted approximately 360 s per epoch, and the Dexined approximately 120 s. With trained models, the system lasts about two min from the B-scan’s entry to the exit of the segmented layers with quantification metrics. After several training sections and hyperparameters tuning, the chosen hyperparameters in the training phase were: number of epochs equal to 300, size of batch equal to 4, Adadelta optimizer with initial learning rate equal to 0.0001, decay equal to 0.095. In the experiments, we tried to use data augmentation to verify improvements in the method’s training phase. However, increasing the sample set by applying transformations like zoom, flip, and others, did not improve layer segmentation procedure. OCT images may show a small rotation in some B-scans. There are few cases, so the magnification operations (zoom, flip, rotation, and translation) probably cannot simulate the exam. The imaging database for patients contains only cases of intermediate AMD, thus limiting the method’s applicability to other stages of AMD. # 3 Results This section shows the results of experiments performed with the segmentation method of the edges of the retinal layers presented in Section 2. The image dataset was divided into two groups. A group composed of 308 volumes of OCT was allocated to training, and another group with 76 volumes was reserved for testing. The data for each group were selected randomly, however respecting the proportion of the distribution of exams with AMD and the control group existing in the base (269: 115). Therefore, the set of images for training consisted of 220 volumes with AMD and 88 from the control group. It was used to train the networks that make up the initial and final segmentation stages. The test set was made up of 50 volumes with AMD and 26 from the control group. For each volume of the training set, 5 B-scans were selected to train the networks, comprised of 1540 images, where 540 of which were used for validation. Since each base volume contains 100 B-scans, but not all possess a ground truth, the most central ones were chosen since it contains complete markings of the edges of the retina layers. The segmentation tests evaluated 3,600 B-scans. The initial segmentation step has the objective of separating two areas in the image where the edges of the layers are positioned. We can verify by the structure of the retina that the ILM is separated from the other two (RPE, BM), as explained in Section 2.4. The U-net was used for this task and generated desired results, reaching a Dice of 97% in the test. In, we can see examples of the results generated by the step. shows the results obtained for the segmentation of the ILM edge, the inner edge of the RPE, and the outer edge of the BM, with the results of calculating the error between our automatic segmentation method and the ground truth provided by the dataset. The values of the average of the MAE are considered low, presenting a higher value for the edges of the RPE and BM due to the fact that in retinas with AMD, drusen that cause deformations between these two layers are formed. The results showed low standard deviation values, showing that the method is very consistent, even between classes. In Figs and, we can see cases where the method managed to generate a result very close to the ground truth, a feature that stands out is the ability to detect the curvatures present at the edges of the layers, especially in the RPE and BM, as they suffer from deformations caused by drusen. In, we have cases of error, where the method generated some flaws in the segmentation of the edges. In these errors, the method did not classify these pixels as belonging to the class of that edge, this happens when the noise level in that area makes analysis difficult, and the noise removal filter is not always able to remove and at the same time maintain the characteristics of the edges. # 4 Discussion The proposed method showed great robustness in the task of segmenting the edges of the retinal layers in OCT images. It is a fully automated method. The difficulty of this task is widely recognized. The majority of studies obtained inferior results using the same dataset. The use of a public dataset is another important feature of this work, as it allows other authors to repeat the experiments and establish comparative analysis. In this work, we use the data set from, which is a dataset of OCT exams with patients with intermediate AMD and healthy manifestations. It is worth mentioning that the results obtained in this experiment are very promising, and they stood out among other works proposed for the same task. In the proposed method, segmentation of the edges of the layers is a fully automated process performed entirely by two CNN models, which are very robust tools that perform implicit extraction and selection of features. This is a positive aspect because it eliminates the need to empirically determine the set of characteristics to be used in the learning process and techniques to be used in the selection of resources. The final segmentation step is performed by DexiNed, which is a convolutional neural network created specifically for edge detection. Finally, the combination of all the techniques in this work provided a better segmentation as the results presented in showed. Despite presenting several positive factors, the proposed method also has some limitations. The dataset is a very complex set, and the images possess a lot of noise, the pre-processing step that uses the bilateral filter seeks to mitigate this, but it still causes the segmentation process to be very difficult. The image base for patients with AMD only contains intermediate cases, thus limiting the method’s performance. Deep learning models, such as the CNNs used in this work, usually requires a large number of parameters. This set of parameters is often empirically determined by the user, which may not provide optimal results. This specific situation is another limitation of the proposed method. All the comments mentioned have added value to this work. The many positive aspects listed in this section have allowed the proposed method to achieve good results. We emphasize that, despite some limitations, the proposed method was able to achieve significant results, and, for these reasons, we believe that the proposal provides an important contribution, as it is an automatic, innovative, robust, and promising segmentation method. ## 4.1 Comparison with related works presents a comparison of the proposed method with other works that used the same image dataset. The results are promising and demonstrate that the proposed method allows us to perform the segmentation of the edges with good precision. Our methodology was able to achieve values of MAE lower or very close to all the results of the related works, obtaining a standard deviation much smaller than most, showing the robustness and capability of generalization of the method. For BM and RPE, our method has the lowest values of mean absolute error and standard deviation. For ILM, the works of have slightly better MAE values. However, the variation is greater, reaching maximum values of 0.59 and 1.3 respectively, they also used a very small amount of images to test the methods proposed in their work, using 200 and 300 slices respectively. # 5 Conclusion Early diagnosis of age-related macular degeneration is very important as it is an eye disease that can cause visual impairment and blindness. One of the main characteristics of the disease is the formation of druses, which cause deformations in the retina between the layers of the retinal pigment epithelium and the Bruch’s membrane. In this article, we proposed a methodology for segmenting the edges of three layers of the human retina; lower limiting membrane, retinal pigment epithelium, and Bruch’s membrane, which are important for the diagnosis of AMD. The proposed method has four steps, pre-processing with a bilateral filter that eliminates noise and highlights the edges, two steps to define the area of interest, and a final segmentation step with DexiNed to find each layer’s limits. The experiment results show that our method is robust for healthy retinas and age-related macular degeneration. The statistics for mean absolute errors and standard deviation are better than those found in state of the art for the layers of the retinal pigment epithelium and Bruch membrane. Our method differs from the literature in that it presents a model specialized in segmenting each of the edges, ILM, RPE, and BM. Despite having obtained promising results, the proposed method can be improved in some aspects. To overcome some of the current limitations of our method, we present the following research directions. Search for other databases with severe dry AMD and wet AMD cases to test the method in such conditions. One of the aspects of research that can contribute to the improvement of the proposed method is the investigation of other pre-processing techniques for removing adaptive noise such as the Wiener filter since the excess noise and variation between images is a large issue. The visualization of the edges of the ILM, RPE, and BM layers is also very important for the diagnosis of another disease, diabetic retinopathy, so this method can be incorporated into a method for the diagnosis of this pathology from images of exams of patients affected by this illness. Finally, as the models of deep neural networks generally have a large number of empirically determined parameters, we suggest the use of hyperparameter definition and optimization strategies to obtain more tuned architectures for both models. [^1]: NO authors have competing interests.

# Introduction Nef is a 27kDa, N-terminal myristoylated accessory protein of HIV-1, involved in disease progression and pathogenesis. It is expressed in early stage of viral lifecycle and is one of the first proteins to be detected after host cell invasion. It manipulates the cellular environment by MHC class I downregulation, CD4 downregulation, and benefits the virus by increasing its propagation and activating anti-apoptotic machinery. Nef interacts with signal transduction proteins of host cell and provides long term survival of infected T cells and induces destruction of non-infected T cells through apoptosis. Nef also promotes the endocytosis and downregulation of cell surface proteins, including CD4 and MHC proteins. This action possibly impairs cytotoxic T cell function, and advances host immune evasion by virus thereby establishing state of infection. The multifunctional protein thus helps the virus overcome host immune defenses, maintain high viral loads, contributing to disease progression. Nef promotes survival of infected cells and favours viral infection and replication through interaction with cellular proteins involved in both trafficking of cell-surface receptors and signal transduction molecules. Accessibility of the conserved domain of Nef controls successful interaction of proteins to the host protein. There are several mutations in the conserved region and regulates functionality of Nef as the disease progresses. The mutation in the proteolytic cleavage region does not allow Nef to come in cytoplasmic region. As a result, the protein interaction in cytoplasm is not possible. Myristoylation of Nef favours its association with the cytoplasmic leaflet of cellular membranes, and is essential for nearly all of its major functions. Thus, sequence variation will contribute to structure based functional analysis of Nef and in understanding how the Nef variants vary in their regulatory role with host proteins. Nef contributes to disease progression and different stages of infection reveal specific Nef mutations. Several studies have analysed Nef gene in HIV-1 patient’s population. Krichhoff *et al*. confirmed that deletion or gross abnormalities in Nef were rarely present but comparison of Nef alleles derived from long term non-progressor and individuals with progressive HIV-1 infection revealed that specific variations were associated with different stages of infection. Sequence heterogeneity of Nef transcripts in HIV-1 infected subjects at different stages of disease was observed. These results demonstrate that various Nef activities were modulated during the course of HIV-1 infection to maintain high viral loads at different stages of disease progression. Thus, it is likely that Nef variants influence the expression of host proteins in a different way and a comparative proteomic analysis will capture such alterations in cellular proteome, regulated differently by Nef variants. Not surprisingly, persons infected with HIV-1 strains that have deletions of the nef gene, show delayed AIDS symptoms and progression, than those infected with standard HIV strains. Nef protein enhances pathogenesis by increasing viral replication. The deleted Nef viral strain, upon infection in monkey, showed low viral load, high CD4+ Tcells and inhibited progression towards AIDS. Moreover, Nef-deleted provirus could be detected in individuals with slow AIDS progression. Nef may therefore be a valuable target for pharmaceutical intercession in AIDS progression. Nevertheless, by all accounts, Nef is a pleiotropic protein, capable of executing diverse intracellular functions that augments the overall pathogenicity of the virus including the enhancement of virion infectivity, disruption of T cell signaling, regulation of apoptotic pathways, and perhaps most importantly, alteration of cell-surface receptor expression. Although it is widely known that HIV-1 initiates a profound global host cell response to infection, it is difficult to trail specific expression and pathway alterations occurring at protein level. HIV-1 infection of cells lead to production of foreign viral RNAs and proteins, alterations in cell cycle, morphology, physiological stress, impaired protein synthesis and degradation, hence it is reasonable to anticipate large differences in protein expression between a normal and infected cell. Proteomic techniques are able to characterize these overall changes, which lead to the understanding of the role of viral proteins in manipulating the cell. Cofactors involved in HIV-1 infection can be identified via the molecular interactions between viral and cellular proteins. Knowledge of the interplay between cellular and viral factors involved in pathogenesis plays a key role in the understanding of disease progression and such proteins could potentially arise as an early diagnostic for a disease state, or a potential therapeutic target. Many proteomic studies have been reported revealing the altered protein expression upon HIV-1 infection and its associated disorders. Proteomic studies in T cells revealed changes in protein expression and compared the differences between uninfected and HIV-1-infected T cells. Classes of proteins and general pathways that were altered upon HIV infection were identified. Identification of these unique proteins may uncover the changes in viral state and indicate phenotypic variability in response to infection in a given population. Proteins showing up regulation and/or down regulation upon HIV-1 infection have been found to be associated with alteration in cellular pathways causing metabolic rerouting. Pathway analysis revealed that the differential expression of proteins targeted selected biological pathways. Such studies help to identify surface markers on HIV-1 infected cells by performing comparative proteomics on HIV-1 infected eukaryotic cell and control cells. As proteins do not function in isolation, alterations in protein networks revealed by understanding the pathway helps unveil the host mechanisms contributory to pathogenesis. Comprehensive information of HIV-1-mediated effects on host cellular protein networks and unique protein targets for the design of therapeutics is required. Nef being a key contributor to pathogenesis, is anticipated to show modulation of host proteins. Recently a comparative proteomic analysis showed Nef dependent increase of virus infectivity. Another study revealed interacting partners in Nef mediated nanotube formation. Scoring Nef mutations and observing difference of Nef variants over host protein modulation would be instrumental in the research process. The present study is aimed to unveil how the HIV-1 pathogenicity factor Nef and mutations occurring in it are involved in such alteration in host protein network. # Materials and Methods ## 1. Materials and Reagents RPMI1640 (Sigma), Antibiotic antimycotic solution (Sigma),Fetal bovine serum (FBS) Gibco BRL, Pure link RNA extraction kit (Invitrogen Life Technologies), Maxima SYBR Green/ROX qPCR Master Mix(Fermentas), RevertAid Premium first strand cDNA synthesis kit(Fermentas), IPG strips (GE Healthcare), Urea, Thiourea, CHAPS, EDTA, PMSF,dithiothreitol (DTT), Protease inhibitor cocktail, Iodoacetamide, Bradfoerd, Silver nitrate, Ponceau, BSA, poly-L lysine and paraformaldehyde were purchased from Sigma Aldrich. Passive Lysis Buffer (Promega), PVDF membrane (Millipore), Luminata forte Western HRP substrate (Millipore), anti alpha-tubulin monoclonal antibody (Invitrogen Life Technologies), DAPI (Invitrogen Life Technologies), Primary antibodies- anti- Cyclophilin A antibody (ab58144), anti-eIF5A antibody \[EP527Y\] (ab32407), anti-ENO1 antibody (ab85086), anti-OTUB1 antibody (ab98280), anti-RhoGDI antibody \[1F2\] (ab118159), and anti-VDAC1/Porin antibody—N-terminal (ab135585) were purchased from abcam, UK. ## 2. Selection of Nef variants after sequencing of *nef* from HIV-1 infected patients Sixteen HIV-1 infected individuals from King George Medical College, Lucknow, India were included in the study for sequencing of nef gene. Two Nef variants: Nef RP14 and Nef RP01 were selected for study on basis of sequence variations (NCBI Accession No.:GQ184340 and GQ184335 respectively). ## 3. Cell culturing and transfection SupT1 (T lymphoblastoid cell line) cells were obtained from our Institutional cell line repository. Cell culture of SupT1 cells was grown in RPMI media supplemented with 10% fetal bovine serum and antibiotics at 37°C and 5% CO₂. For transfection, 4 million cells were seeded per sample. After 24 hours, cells were centrifuged, washed twice with incomplete media and resuspended in 400 μl incomplete media. 20 μg of respective plasmid was added in the cells and electroporation was carried out at 290 V pulse (low-voltage mode). Cells were immediately transferred to flasks for further culture. After 2 hrs complete media with 20% FBS and antibiotics was added. Cells were harvested after 48 hrs and subjected for further experiment ## 4. RNA extraction and quantitative real-time PCR Real time PCR analysis was done in SupT1 cells (vector control vs Nef transfected) with the 6 genes keeping GAPDH as housekeeping gene. Total RNA was isolated from control and Nef treated cells using Purelink RNA mini kit (Invitrogen) following manufacturer’s protocol. DNase enzyme digestion of the isolated RNA was done to remove any DNA present in it. RNA was quantified and 10 μg of RNA was taken for DNase digestion. 1 unit of DNase (Fermentas) was added to the RNA in presence of 1xDNase buffer (Fermentas) and incubated at 37°C for 10 min and then transferred to 70°C for 10 min in order to denature the DNase. DNase digested RNA (5 μg) was reverse transcribed using two step cDNA synthesis Kit (Fermentas). Briefly, real-time PCR was performed on Roche\`s 480 Real Time PCR Instrument using SYBR green qPCR mix supplied with two step qRT-PCR Kit (Invitrogen) with 25 ng cDNA and 1 ul of 10 pM of specific primer sets for gene to be amplified in a total volume of 20 μl reaction mix. The thermal cycling conditions comprised 2 min at 50°C, 10 min at 95°C, followed by 40 cycles at 95°C for 30 s, 57°C for 30 s, and 72°C for 30 s. All the reactions were performed in triplicate. Relative quantification of the target mRNA was done using delta delta CT method followed by normalization with the level of the internal control GAPDH mRNA level. ## 5. Two-dimensional gel electrophoresis 2-DGE ### Sample preparation Transfection of Nef (RP14) was carried out in SupT1 cells. Rate of transfection was seen through expression of transfected cells after 24 hrs till 48 hrs. Sup-T1 cells with vector (pYFP-N1) transfection were kept as control. After 48 hours incubation, cells were harvested by centrifugation at 1500 rpm for 5 min at RT. Cells were washed twice with 1X PBS with centrifugation at 1500 rpm for 5 min at RT. Traces of PBS were completely removed and cells were immediately subjected to lysis for protein extraction. 10 million cells were lysed by 3 ml of Urea Lysis Buffer(7 M urea, 2 M thiourea, 4% (w/v) CHAPS, 1mM EDTA, 1mM PMSF,100 nM dithiothreitol (DTT) and protease inhibitor cocktail 10μl/ml) with updown mixing by pipetting. Cells were kept for lysis on rotor for 1 hr at RT, followed by centrifugation at 14000g for 20 min at 4°C-8°C. This was followed by ultracentrifugation at 1 lac g for 1 hr. Pellet was discarded and supernatant was subjected to acetone precipitation to further remove the impurities. 4V of pre-chilled acetone was added to the supernatant, taken out in separate tubes and incubated overnight at -20°C. This acetone mix was centrifuged at 13000g for 10 min at 4°C, supernatant was decanted and pellet further washed with 90% pre chilled acetone. After two washings pellet was air dried properly and resuspended in Rehydration Buffer (7 M urea, 2 M thiourea, 4% (w/v) CHAPS). Protein was quantified by Bradford method. ### 2D Gel electrophoresis Two-dimensional gel electrophoresis 2-DGE was carried out using IPG strip gels (GE Healthcare Bio-Sciences) on an IPGphor unit followed by the second dimension using hand-cast polyacrylamide gels on a sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) vertical electrophoresis unit. Rehydration mix containing 75 μg of total soluble protein was prepared with rehydration buffer containing 2% DTT, 0.5% (v/v) pH 3–10 IPG buffer in a final volume of 300 μl. A trace of bromophenol blue (BPB) was added and the whole mixture kept at room temperature (RT) for 5 min followed by pipetting into a 13 cm strip holder tray. IPG strips (pH 3–10; 13 cm) were carefully placed onto the protein samples, covered with a lid, and placed into the IPGphor unit. The IPG strips were placed gel-face down onto the protein samples avoiding air bubbles, followed by overlaying with1 ml cover fluid and allowed to passively rehydrate with the protein samples overnight. Strips were then transferred on an IPGphor unit for IEF. The whole procedure was carried out at 20°C, and a total of 69562 Vh (volt hour) were used for the 13 cm strips. Following IEF, the IPG strips were removed from the strip holder and the cover fluid adsorbed on filter papers. The strips were then immediately used for the second dimension. The strip gels were incubated in equilibration buffer (6 M urea, 75 mM Tris-HCl pH 8.8, 29.3% glycerol, 2% SDS, 0.002% bromophenol blue, 200 ml) containing 2% (w/v) DTT for 15 min with gentle agitation, followed by incubation in the same equilibration buffer supplemented with 2.5% (w/v) iodoacetamide for the same time periods as above at RT. For the second dimension, the IPG strips were rinsed with running buffer (0.025 M Tris, 0.192 M glycine and 0.2% (w/v) SDS), placed onto pre- casted 12% polyacrylamide gels and overlaid with agarose solution (60 mM Tris- HCl, pH 6.8, 60 mM SDS, 0.5% (w/v) agarose, 0.01% (w/v) BPB). The gel cassettes were placed onto the PAGE running assembly. Electrophoresis was carried out at 80V for first 2 hrs and then increased to 120 V till end. To visualize the protein spots, the polyacrylamide gels were stained with silver. Staining with silver nitrate was performed as described in the instructions provided with the Plus One Silver Staining Kit (GE Healthcare). ### Image acquisition and analysis Protein patterns in the gels were recorded as digitalized images using a digital scanner. The stained gels were scanned wet and all images saved were cropped to include only the resolving area of the gels. The gels were quantitated in profile mode as instructed in the operating manual of the ImageMaster 2D Platinum software (GE Healthcare). All the gels were aligned against a prominent common spot and images were analysed to identify valid spots. Spots were numbered and spot density graphs were acquired through the software techniques. Three gels of treated and control group were analyzed together to find significant up regulation or down regulation of spots. ## 6. Identification of Proteins in 2D gel spots The spots showing significant down regulation upon Nef infection were identified by LC-MS/MS and further studied through bioinformatics tools. Silver stained protein spots showing differential expression pattern were picked and sent to (Proteomics International Pvt. Ltd, Nedlands, Western Australia) for doing mass spectrometry. In brief, protein samples were trypsin digested and peptides extracted according to standard techniques. Peptides were analysed by electrospray ionisation mass spectrometry using the Ultimate 3000 nano HPLC system \[Dionex\] coupled to a 4000 Q TRAP mass spectrometer \[Applied Biosystems\]. Tryptic peptides were loaded onto a C18 PepMap100, 3 mm \[LC Packings\] and separated with a linear gradient of water/acetonitrile/0.1% formic acid (v/v). The amino acid sequence tag obtained from each peptide fragmentation in MS/MS analyses was analysed to identify proteins of interest using Mascot online search engine ([www.matrixscience.com](http://www.matrixscience.com/)) against the taxonomy set to Homo sapiens (human). ## 7. Western blotting Western blotting was done in Sup-T1 cells. Transfected cells (Nef RP14 and Nef RP01) and control cells (pYFP-N1 vector) were harvested by centrifuging cells at 1500rpm for 5minute at 4°C. Cells pellet was washed twice with 1X PBS by centrifugation at 1500 rpm for 5min at 4°C. PBS was removed and pellet was resuspended in 100μl of 1X Passive Lysis Buffer (Promega) and sonicated at 30 amplitude giving 4 pulses of 5 second on and 5 second off. Cell lysate was centrifuged at 12000rpm for 20min at 4°C to remove cell debris. Supernatant was collected and protein estimation was done using Bradford method. 100μg of protein sample was mixed with 6X Laemmli sample buffer. The mixture was heated for 2 min in boiling water bath and total protein was run on 12% SDS- polyacrylamide gels in a vertical slab gel unit was assembled using the manufacturer’s instructions (Amersham Biosciences). The gel was electrophoresed initially at 10 mA till the samples reached resolving gel and thereafter at 20 mA till the end. The gel was removed from the apparatus and western blots were prepared by transferring the separated proteins from the polyacrylamide gel onto a PVDF membrane (Millipore) in a semi dry transfer carried out at 15 V for 3 hrs. Post transfer the membrane was stained with Ponceau stain to check protein transfer and rinsed twice with triple distilled water to remove the stain. Membrane was incubated in blocking buffer (5% BSA in 0.01% PBST-pH 7.4) for 2 hrs at room temperature with gentle shaking followed by incubation with appropriate dilution of the primary antibody in 1% BSA containing PBST buffer dilutions according to manufacturer's instructions) overnight at 4°C. The blots were washed 6 times with 0.05% PBST buffer, for 5 min each and were then incubated with appropriate dilution of the respective secondary antibody in PBST buffer (dilutions according to manufacturer's instructions) for 2 hours at room temperature with gentle shaking. The blot was washed 7–8 times with 0.05% PBST buffer (pH 7.4), for 10 min each. Protein was detected using chemi-luminescent substrate provided in Luminata forte substrate kit (Millipore-1 ml/membrane). Substrate was added to the membrane kept for 1 min, extra solution was drained out, exposed to UV-light and image was captured. Blots were then stripped and probed with an anti alpha—tubulin monoclonal antibody (Invitrogen) as control for equal loading of proteins. ## 8. Immunofluorescence Study Immunoflurescence studies were conducted in SupT1 cells. Control and Nef transfected cells were harvested, resuspended in media and overlaid on 0.01% poly-L lysine coated coverslips placed in 12 well cell culture plates. Cells were allowed to adhere on these coverslips for 6 hrs at 37°C. Media was removed and cells were washed thrice with 1X DPBS. Cells were fixed by adding 4% paraformaldehyde for 20 min at 37°C, washed and were then blocked by 5% BSA in 1X PBS for 1 hr to remove non-specific binding sites. Primary antibodies at recommended concentrations in PBS were added over the cells and incubated overnight at 4°C with gentle shaking. Cells were washed and respective Secondary fluorescent antibodies were added at 1:500 dilution. Plate was covered with foil to maintain dark and kept for shaking for 2 hrs at RT. Washing was done and coverslips were mounted with DAPI on a slide and sealed with nail paint. Slides were incubated in dark for 1 hr at RT and Immunofluorescence was detected by Zeiss Confocal microscope using 65X objective lens with 4X and 2X zoom and analysed by Zeiss LSM Image viewer software. ## 9. Nef RP01 mutant generation with mutation at 55–61 aa Nef RP01 mutant was constructed by replacement of 55CAWLEAQ61 amino acids to 55 AAAAAAA61. For mutant generation nef gene was amplified using two primer sets as shown in. First set of primers amplified the N- terminal fragment of Nef, while second primer set amplified the C- terminal fragment of Nef. These two PCR amplified products with mutated overlapping regions, were mixed and used as template for amplification of final mutated nef gene with primers: Nef RP01 FP and Nef RP01 RP. The amplified mutated nef gene product was cloned in pYFP-N1 for further experiments. Desired mutation was confirmed by DNA sequencing. ## 10. Bioinformatic analysis The sequence alignment of protein translates of Nef variants was done online by Clustal Omega software (<http://www.ebi.ac.uk/Tools/msa/clustalo/>). Sequences were uploaded in FASTA format as Protein Sequences and analysed further for variability. The function of the identified underexpressed proteins was elucidated by SWISS-PROT database and the interaction between these proteins and Nef was checked by HIV interaction database <http://www.ncbi.nlm.nih.gov/RefSeq/HIVInteractions>. A protein-protein interaction network of the 6 proteins with closest interacting partners was generated through STRING 9.1.([string-db.org](http://string-db.org/)) Partners with high confidence level and score were interconnected. ## 11. Statistical Analysis Numerical data was tested for statistical significance using paired student’s t test (two-tailed). Differences were considered significant when p ≤ 0.05(\*), and very significant when p ≤ 0.01(\*\*). All data was analyzed by graphpad- prism 5. ## 12. Ethics Statement The Ethical clearance for the study was granted by Human Ethical Committee of Research Cell of King George Medical University, Lucknow, India (Ref Code: *XVIII ECM/P9*). The written informed consent was obtained from all study participants and this consent procedure was approved by the Ethics Committee. # Results ## 1. Analysis of sequence variability in HIV-1 Nef variants The preliminary objective of the analysis was to identify sequence variability, and major variations in Nef sequences to understand the differential functions of Nef in HIV-1 patients. The two Nef variants taken in this study were RP14 and RP01. Besides other single point variability, RP14 exhibit a nine amino acid insertion at 21–29 aa position and a significant deletion of 7 aa at 55–61 position. Interestingly, this deletion involves the CD4 downregulation and proteolytic cleavage site ⁵⁵CAWLEAQ⁶¹ that is known to be well conserved in Nef. Hence, RP14 was considered mutant form, while RP01 with sequence close to wild type Nef was taken as wild type Nef. The sequence alignment showing the variability between these variants is shown in. This unique mutation has not been correlated to disease progression in the literature so far and could not determine the stage of HIV-1 pathogenesis so its implications have to be investigated. The two Nef variants cloned in pYFP-N1 displayed full biological activity as shown in our previous reports, and were thus functional. These were studied further through proteomics approach. ## 2. Difference in host protein expression mediated by Nef as revealed by 2DGE 2D gel electrophoresis offers a suitable method to understand the modulation of host proteins as mediated by Nef, through quantification of the differences in protein expression of control and Nef transfected cells. To evaluate the host cellular factors being affected by Nef, proteome analysis of Nef transfected SupT1 cells was carried out through 2D gel electrophoresis. Nef RP14 transfected cells and pYFP-N1 transfected vector control cells were harvested after 48 hours of transfection, whole cell protein lysates of each were subjected to 2D gel electrophoresis. The gels with spots are shown in. Gels were scanned and analysed through Image Master 2D Platinum software (GE Healthcare). 2D gel platinum software Spot a corresponding to β-actin (PI 5.29 and MW 41.7 KDA) was considered as one of the marker protein, selected for aligning the gels and normalization of loading. Experiment was performed in triplicate. On image acquisition, total 516 spots were identified in 2D gel of cells, of which, 375 spots were found to be common in both the gels. 14 unique spots were found to show differential protein expression upon Nef transfection. Density of the spots was calculated through software and comparison of spot densities revealed that Nef significantly altered the expression of host proteins. These proteins were found to be downregulated upon Nef transfection. 6 spots (spot no. 25, 31, 98, 125,157 & 244) showing major downregulation were picked and sent for LC/MS-MS. Other spots could not reach significant statistical value due to variation in spot densities among three different experiments and were thus not considered for further analysis. Enlarged sections of the region of gel showing altered expression and the degree of difference in spot density and peak area of 3D images are displayed in. ## 3. Identification of differentially expressed proteins Differentially expressed proteins of interest were excised and successfully identified through LC-MS/MS followed by mascot software analysis. The proteins were considered identified on basis of high sequence coverage and score based on peptide matching. The spots with more than one protein hit identified through software, protein hit with highest score and sequence coverage were considered as protein identified. The proteins were identified as Cyclophilin A, (Peptidyl- prolyl cis trans isomerase), Eukaryotic translation initiation factor 5A-1 isoform B, rho GDP-dissociation inhibitor 1 isoform a, voltage-dependent anion- selective channel protein 1, Ubiquitin thioesterase protein OTUB1,and α-enolase isoform 1(ENO1). The molecular weight and PI of each protein was determined. As shown in and, all the 6 proteins identified had 3 or more peptides matched upon spectrum analysis and showed high sequence coverage. The matched peptides are shown bold red. According to annotations from Uniprot (Swiss-prot/EMBL) and descriptions provided, identified proteins were found involved in various cellular functions like host-virus interaction, Protein and mRNA transport, apoptosis, signal transduction, ion transport, protease activity, GTPase activation, immune response, plasminogen activation and energy associated pathways like glycolysis. The detailed description of the identified proteins is shown in. ## 4. Real Time PCR for analysis of transcriptional profile of the differentially expressed proteins Quantitative Real Time PCR was done in order to check the transcriptional alterations of the 6 genes, keeping GAPDH as control housekeeping gene. Significant down regulation of 5 genes was obtained in Nef treated SupT1 cells as compared to control and were consistent with the result of 2DGE. In contrast, VDAC1, showed different result and was found to be upregulated at transcript level. As shown in, both Nef variants caused significant downregulation of rest of 5 genes with Nef RP14 causing greater than Nef RP01 in case of CYPA, EIF5A, RhoGDI and OTUB1, however the downregulation of ENO1 was greater by Nef RP01. Fold change in mRNA expression of the genes was normalized with GAPDH and calculated from three different experiments. ## 5. Validation of the identified proteins by western blotting analysis To further validate the proteomics result and ensure that identified proteins were underexpressed by Nef, western blotting was performed with specific antibodies against the proteins of interest in control and Nef transfected SupT1 cells. All the 6 proteins showed downregulation by Nef variants. Interestingly, there were striking differences in the effect of the Nef variants upon the expression of these proteins. As demonstrated from, α-enolase (ENO1), VDAC1, and ubiquitin thioesterase (OTUB1) were more underexpressed by Nef RP01 with fold decrease 3.55, 1.95, and 1.63 respectively (p\<0.01); whereas the downregulation of Cyclophilin A and RhoGDI was greater by Nef RP14 with fold decrease 1.43 and 2.12 respectively (p\<0.01), as calculated from three different blots normalized with alpha tubulin as loading control. Underexpression of EIF5A-1 was found to be similar by both Nef forms. Thus, western blotting results confirmed the 2DGE results and showed difference of Nef variants upon expression of these proteins. ## 6. Confirmation of proteomics result and comparison of the surface expression of proteins through immunofluorescence study Control and Nef transfected cells were fixed, permeabilised and labelled with specific antibodies against the six proteins followed by incubation with fluorescent secondary antibodies and fluorescence was captured through confocal microscopy to compare the expression of the 6 proteins in SupT1 cells being transfected with Nef variants. As seen in the results of immunofluorescence were found to be consistent with proteomics. On comparison of fluorescence intensity of the labelled protein over cells there was significant downregulation seen in the protein expression by Nef. α-enolase (ENO1), VDAC1 and OTUB1 were underexpressed substantially by Nef RP01, whereas EIF5A1, Cyclophilin Aand RhoGDI were underexpressed by both Nef variants. Cells with vector/Nef expression (Green channel) specific protein expression (Red Channel) and DAPI stained nucleus are shown in with statistical significance of fold decrease being calculated from fluorescence of 5–8 transfected cells taken in 6 different fields of which single cell image has been shown in figures. ## 7. Mutation of 55CAWLEAQ61 in RP01 reversed its downmodulating effect on α-enolase (ENO1) and VDAC1 To investigate whether the differences observed in effect of two Nef variants over expression of the proteins, could be attributed to the unique deletion of proteolytic cleavage site in RP14, a mutant of Nef RP01 was constructed in which the amino acids of 55CAWLEAQ61 domain of Nef RP01 were replaced with Alanine. Western Blotting and immunofluorescence studies were performed to show the effect of this RP01 mutant upon expression of α-enolase (ENO1) and VDAC1 as both these proteins were significantly downregulated by RP01 but not by RP14. As shown in, RP01 downregulates α-enolase (ENO1) and VDAC1 but RP01 mutant does not and its effect is comparable to that of RP14. This indicate the significance of 55CAWLEAQ61 site of RP01 upon these protein expression, as mutation of this site in RP01 nullified its downmodulating effect on α-enolase (ENO1) and VDAC1. Similarly, Nef RP01 mutant was employed for IF studies in α–enolase (ENO1) and VDAC1, and showed reverse effect on their expression as compared to RP01. ## 8. Bioinformatic analysis and protein interaction pathway for Nef variants with host proteins The six proteins were found to play critical roles in regulation of different cellular functions. The Nef variants showed difference in effect over expression of six proteins. The downregulated proteins and the protein interacting partners connected in a pathway were determined by STRING 9.1 database shown in. Direct interaction of the 6 proteins and their close interacting partners in the network generated by STRING was checked by HIV human protein interaction database and were found to be associated with Nef. Among the six proteins being downregulated, Cyclophilin A was found linked with Nef, whereas from the close interacting partners BCL2L1, CDC42 and RAC1 were associated with Nef. As demonstrated from this network Nef might be affecting the associated pathways and Nef variants must be regulating different signal transduction pathways and hence affecting viral pathogenicity differently. Further mechanism behind the regulation of these proteins by Nef and their downstream pathways, which may favour viral replication, needs to be explored extensively. Possibility of a common signalling event mediating the downregulation at transcriptional level was pursued through in silico studies of promoter sequences of the 6 genes. Sequence alignment showed 37–58% similarity among the promoters (data not shown) but no common transcription factors or significant sequence similarity was seen through this analysis. # Discussion HIV-1 Nef, a key determinant contributing to viral pathogenesis, binds to a diverse group of host cell signalling molecules and regulates the surface expression of numerous T-cell molecules involved in HIV-1 infection and T-cell functions. Several conserved motifs of Nef mediate association with cellular factors and modulates protein expression through interaction with cellular kinases and signalling molecules. Nef is known to induce a state of activation in the host T-cell, the principal target of HIV-1 infection, which leads to enhanced viral replication. The present study uses a proteomics approach to investigate pathways associated with Nef regulated host proteins and understand how the sequence variability in Nef, altered its ability to modulate the protein expression in T-cells. For the study, Nef was sequenced from randomly selected HIV-1 infected patients but do not elicit any specific mutations, that could correlate to specific stage of infection. However, a conserved domain was found to be deleted in many HIV-1 infected patients. The Nef protein having deletion of conserved domain was pathogenic and higher order tetramer association was found. Interestingly insertion of 9 amino acids in this mutant form of Nef, RP14, was found in a variable region of patient’s nef gene. The deletions and insertions being present in N-terminal unstructured region conserves the overall length of Nef and explains that the length of N-terminal region could be critical for its functionality and restoration of Nef activities. This unique deletion of 7 amino acid ⁵⁵CAWLEAQ⁶¹ in unstructured region of patients nef gene compromises two functions of Nef, that is, CD4 down regulation and liberation of membrane Nef to cytosolic Nef, which is demonstrated by mutational studies in this region. Till date, in-vitro studies showed that Nef was cleaved by viral proteases at cleavage site located between 57W/L58, and determine the modular organization of Nef, separating it into anchor and core domain. This indicates that in HIV-1 infected condition, membrane bound Nef is liberated to cytosolic Nef by cleaving at this site. The studies have shown that the anchor domain N-myristoylation region of Nef, places it to cell and associates with rafts where several signalling molecules present interact for activation of T-cells. The two variants have been investigated for functional aspects of Nef in our previous reports. One study identified the interacting domain of Nef (RP01) and ASK1 where their physical interaction led to inhibition of ASK1 enzymatic activity. Nef RP14 has been reported to show tetramer association and showed pathogenesis in *C*. *elegans* reported by our group. Two Nef variants, Nef RP14 and Nef RP01, were selected for further study with an aim to explore how mutations occurring in Nef contribute to its functionality. Comparative proteomic study was carried out focused on the target proteins of these Nef variants and analysed the modulation of host protein expression in Sup-T1 cells to observe the difference in functional aspect of these Nef variants. Proteomic profiling allows formation of new hypothesis regarding the functionality of Nef. It is conceivable that, besides known interactions of Nef, there may be some proteins that are differentially regulated by Nef and have undescribed effects over viral life cycle. To study the T-cell–Nef interaction, SupT1 cells cultured in-vitro were used, which likely differ from primary T-cells found in-vivo, but were effective at representing the principal effects of Nef. Proteomic profiling of SupT1 cells upon Nef transfection was done by 2D gel electrophoresis with Nef RP14, the mutant form. Differentially expressed proteins were observed and host proteins downregulated by Nef were picked and identified for further study. Six proteins with major downregulation were identified as α-enolase isoform 1 (ENO1), ubiquitin thioesterase, cyclophilin A, voltage-dependent anion-selective channel protein 1, Eukaryotic translation initiation factor 5A-1 isoform B and rho GDP- dissociation inhibitor 1 isoform a. The underexpression of these proteins by Nef needed to be further validated, and also the transcriptional alteration of these genes was to be checked to observe the level of modulation. Further confirmation involved both Nef variants, to elucidate their difference over host protein expression. Q-PCR studies were carried out to check whether this alteration was occurring at transcript level or a post-transcriptional change. All genes were found to be downregulated by both Nef forms, except VDAC1 which was found to be enhanced at transcript level.In some cases change in mRNA levels and protein levels did not correlated that well mainly due to the regulation control at different stages. Increased mRNA level parallel with downregulation of protein expression is generally seen among proteins being downregulated by ubiquitination. Western blotting analysis of the six proteins by both variants showed interesting result and gave striking difference in effect of both variants over the expression of these proteins. This contrasting effect of Nef variants was further confirmed by immunofluorescence studies. This also compared the cell surface expression of the proteins. Downregulation of α-enolase (ENO1), VDAC1 and OTUB1 was more significant by Nef RP01 whereas, Cyclophilin A EIF5A-1 and RhoGDI were found to be downregulated by both Nef variants more by Nef RP14. Consistent results were obtained in Western and IF analyses for all the proteins. The sequence variability in Nef variants RP14 and RP01 could be targeted to predict the possible mechanism behind the differential regulation of host proteins. To answer the reason behind observed difference in effect of Nef forms upon expression of host protein, especially in α-enolase (ENO1) and VDAC1, the 55–61 amino acid site representing proteolytic cleavage domain of Nef was exploited. RP14 shows natural deletion of this site apart from over 20 other point mutations as compared to RP01. In our later part of study we constructed 55CAWLEAQ61 to 55AAAAAAAA61 mutant of RP01 and checked its effect upon protein expression. Further studies showed that effect of RP01 got reversed significantly upon protein expression after generation of this mutation. It is possible that small differences between protein expression profiles reflect a high degree of specificity in mode of activation and influence cellular outcome. The proteins were found to be involved in several cellular pathways and associated with relevant host functions. The proteome of Nef infected T-cells was changed particularly in networks associated with energy release, transport, apoptosis, cell migration and signal transduction. Most significantly underexpressed protein α-enolase (ENO1) is known to induce cell migration and tissue invasion. In a previous report, α-enolase (ENO1) was found to be associated with HIV-1 revealed by proteomic analysis of HIV-1 virions produced by monocyte derived macrophages. In present study, it was found to be downregulated by Nef, which may alter the migratory property of immune cells and affect their invasiveness. Another protein VDAC-1 (voltage dependent anion selective channel protein 1) plays role in ion transport and regulates apoptosis. This protein is believed to form the major pathway for movement of adenine nucleotides through the outer membrane and to be the mitochondrial binding site for hexokinase and glycerol kinase. It is proapoptotic energy associated regulatory protein. Nef downregulates it and alters the energy requirement of cells that favour establishment of infection stage. Both VDAC1 and α-enolase (ENO1) are energy associated proteins. Several reports prove that cell metabolism is being affected by HIV. In a study, effect of HIV on glucose metabolism was measured in human intestinal epithelial cells. Glycolytic dysregulation was detected in neuronal cells upon treatment with HIV gp120 protein. These studies depicted disturbed glycolytic and oxidative activities. Decreased glycolysis should induce the pentose phosphate pathway which can also be used for the synthesis of new viral particles and stress response. Decreased α-enolase (ENO1) would lead to decreased T-cell migration. Inhibition of T-cell migration would protect the infected cells from host immunological attack and help in immune evasion. The decreased energy state of infected T-cell leading to inhibition of their migration goes in agreement with the observed downregulation of α-enolase (ENO1) and VDAC1. VDAC1 also regulates cell volume and apoptosis. In fact, apoptotic proteins have been observed differentially expressed upon infection. BCL-2 appears to regulate cell death by blocking the voltage- dependent anion channel (VDAC) by binding to it and preventing the release of the caspase activator, cytochrome c, from the mitochondrial membrane. The Bcl-X(S) isoform promotes apoptosis. Many proteins employ ubiquitination and subsequent proteolysis for cell cycle regulation and apoptosis. OTUB1, another Nef downregulated protein, is a hydrolase playing regulatory role in preventing protein degradation by removal of conjugated ubiquitin, prevent proteolytic degradation. This isoform has been found to regulate T-cell anergy, in which T-cells were rendered unresponsive to antigen rechallenge. OTUB1 was found to be overexpressed in viral infected cells treated with cyclosporine A, showing its possible role in causing the inhibitory effect. Decreased OTUB1 would lead to increase of protein degradation in host and regulates various cellular events and host-virus interactions. For the first time, α-enolase (ENO1), VDAC1 and OTUB1 have been reported to be modulated by HIV-1 Nef through our study. Known HIV-1 host-virus interactions were also observed. Cyclophilin A has known interaction with Nef and is important for viral packaging and entry. Cyclophiln A is known to bind human HIV-1 gag protein and is required for the HIV-1 infectivity probably by causing uncoating of the virion coat. Viruses lacking cyclophilin A completely fail to attach to target cells and are mandatory for viral adsorption on host. It is essential for viral attachment to host cells and thus for the infectivity of virus. Preventing CypA packaging, either by the addition of CsA to particle producer cells or by the introduction of mutations in the binding region of CA, inhibits virus infectivity, demonstrating a strict requirement for CypA in HIV-1 replication. It was found to be downregulated by Nef in T cells, inferring the prevention of superinfection by virus. Another protein eIF-5A, the eukaryotic translation initiation factor is an mRNA-binding protein involved in translation elongation and is also described as a cellular cofactor of HIV-1 Rev protein, essential for mRNA export of retroviral transcripts and HIV-1 replication. Rev is essential for virus replication and is required for the cytoplasmic expression of unspliced and singly spliced viral mRNAs encoding the viral structural proteins. It was found to be downregulated upon viral infection. It plays vital role in pathways involved in stress response and cell cycle progression maintaining cell wall integrity. Study also found Rho GDI 1 (Rho GDP dissociation inhibitor) member of the family of small GTP binding proteins down-regulated. It regulates GDP/GTP exchange reactions of Rho proteins by inhibiting the dissociation of GDP and causing subsequent binding of GTP to them. Thus, its down-regulation may promote viral replication through constant activation of GTPases, as they are known to be necessary for HIV-1 replication. Rho-GDI complex expression was also found to be down- regulated in Jurkat-Tat cells and had role in signal transduction and communication. Rho GDI, which has been involved in actin cytoskeleton organization, was found to be downregulated upon HIV-1 infection. The predicted protein-protein network generated by STRING, shows connection of these underexpressed host proteins with proteins related to signal transduction receptors, glycolysis, apoptosis, GTP binding proteins, cell adhesion molecules, tumour progression, cell cycle, cell differentiation and cell migration. Although some of these proteins had known interactions with HIV-1 but their Nef mediated alteration and probable association with Nef is attributed through our study. The network showed association of Nef with host proteins and their close interacting partners. Cyclophilin A, BCL2L1, CDC42 and RAC1 were found to be directly linked with Nef. Nef shows interaction with cyclophilin A and RAC1. In a previous study it has been reported that Cyclophilin A binds to HIV-1 Nef using a biochemical assay, however the biological significance of this interaction is currently unknown. Nef binds with DOCK2-ELMO1 complex to interact with RAC1 and inhibits lymphocyte chemotaxis. RAC1 inhibits Nef mediated internalization of CD80 and CD86. Active RAC1 levels are increased by Nef mediated activation of Vav. HIV-1 Nef upregulates active CDC42 levels in dendritic cells. CDC42 and RAC1 are required for activation of Nef associated kinase and HIV-1 replication. Nef enhances apoptosis in lymphocytes by downregulation of anti-apoptotic proteins Bcl-2 and Bcl-X. How all these changes fit in with energy state of cells, migration, apoptosis, cell cycle arrest, viral production, and/or transport remains to be established and how the observed changes correlate to pathogenicity for the infected individual is also unclear. This first 2DGE analysis of changes in the T-cell proteome upon Nef infection and comparative proteomic study of Nef variants, not only uncovered the difference in effect of the Nef variants over host protein expression but also allowed to identify many proteins potentially involved in the virus-host interactions. The presence of such host proteins being modulated by Nef variants is often the most accurate reflection of cellular response to an infection and can be correlated to disease progression. The mutational study proved the importance of proteolytic cleavage domain among other mutations between the two Nef variants and strengthened its role in affecting the functionality of the Nef variants through differential expression of host proteins. Moreover, cell culture studies are more potent at signifying the primary effects of an infection. HIV-1 infection leads to profound changes in the cellular transcriptome and proteome. Here, the study focused on quantifying perturbations caused by Nef and its mutations and gaining insight into the molecular mechanisms behind this modulation of proteins, in further studies. # Conclusion In the present study, Nef-induced modulation of T-cells was characterized in terms of the resulting alterations in cell proteome. The study adopted a gel based proteomic approach to probe changed proteins that allowed for identification of physiologically relevant targets of signal transduction pathway. In this way, many of the proteins modulated by Nef, provide indications for uncovering potential virulence mechanisms. The subversion of the T-cell by Nef and alteration of cellular factors is likely to represent an important mechanism facilitating increased virion production. We have identified cellular proteins downregulated by Nef and subsequently, on basis of contrasting effects of Nef variants upon protein expression we found a specific domain of Nef (55CAWLEAQE61) regulating this differential expression. The difference in effect of Nef variants over protein expression reflect the probable differences in these Nef forms in terms of their implications over cellular functions. Nef, by downmodulation of these proteins must be compromising the associated host functions and create an environment favourable for viral propagation. The sequence variations leading to decreased Nef pathogenicity could be further explored for identifying factors regulating the protein modulation, which would help in better understanding of Nef functionality and design inhibitors. The proteins could be explored for their effect over viral infectivity and in deciphering how Nef is affecting HIV-1 infection through underexpression of these host factors. These altered proteins can serve as future diagnostic and/or therapeutic markers and be studied as potential drug targets with relevance to different stages of disease progression. As a reference to the functionality of Nef, its large scale protein expression profile may aid a full understanding of the contribution of this accessory gene in pathogenesis and could be used in the development of inhibitors of Nef action. Thus, this study can be useful in revealing the regulatory aspect between HIV-1 Nef and its variants with host cell. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: RS RKT. Performed the experiments: RS SG. Analyzed the data: RS RKT. Contributed reagents/materials/analysis tools: KS AKT KM. Wrote the paper: RS RKT.

# Introduction First-generation drug-eluting stents (DESs) reduced the incidence of in-stent restenosis compared to bare metal stents (BMS) in symptomatic coronary arterial disease by inhibiting smooth muscle cell proliferation and preventing intimal hyperplasia. Stents that eluted antiproliferative, immunosuppressive, and anti- inflammatory drugs reduced target vessel failure and consequently improved clinical outcomes. However, DESs were also associated with an increased incidence of stent thrombosis, which is mainly attributed to the prolonged endothelialization process and the resulting impairment of vascular healing. Patients with acute ST-segment elevation myocardial infarction (STEMI) were excluded from the first DES trials due to the uncertainty of the thrombotic effect of DESs on vessel patency. Therefore, DESs were used “off-label” in STEMI until late-2000s. In the early and mid-2000s, it was proposed that BMS and Genous stents be implanted during primary percutaneous coronary intervention (pPCI). Notably, the Genous stent was developed in 2001, as the world’s first pro-healing intracoronary stent. The first-generation Genous bioengineered stent (OrbusNeich Medical Technologies, Fort Lauderdale, FL, USA) was a passive-coated 316L stainless steel stent covered with a biomaterial that contains monoclonal anti-humanCD34+ antibodies, which bind circulating endothelial progenitor cells (EPCs) at the stent surface. When the EPCs were concentrated near the vessel injury site, they helped accelerate endothelialization by differentiating into active endothelial cells. The rapid re-endothelialization of the stented vessel surface limited smooth muscle cell proliferation and theoretically decreased intimal hyperplasia as well as stent thrombosis. Several small and larger cohort and randomized studies that have included patients with stable coronary lesions have demonstrated the efficacy of the Genous stent in reducing target vessel revascularization, restenosis, and stent thrombosis. The use of Genous EPC-capturing stents in STEMI was proposed because acute myocardial infarction causes the mobilization of EPCs, which, in turn, can accelerate the healing of the stented vessel injury. Orally administered statins would act to increase the number of circulating EPCs. The idea was that using the Genous stent, with its ability to bind EPCs and its non-eluting passive coating, in combination with the maximum dose of oral statins, might be useful for pPCI in STEMI to prevent target vessel failure. The first cohort and randomized studies that included patients with STEMI who received Genous stents reported favorable short- and mid-term outcomes for clinical endpoints. The aim of this study was to compare the long-term (9-year) clinical outcomes of patients with STEMI who received either Genous or TAXUS Liberté stents during pPCI. # Methods ## <span class="smallcaps">S</span>tudy design This study was a retrospective, single-center, 9-year follow-up trial conducted at the Department of Cardiology at the Medical University of Vienna. The study protocol was approved by the Ethics Committee of the Medical University of Vienna and complied with the principles of the Declaration of Helsinki. ## <span class="smallcaps">S</span>tudy population Between February 2006 and December 2008, consecutive patients with STEMI who received either the Genous or TAXUS Liberté stents during pPCI were included in the analysis. The pPCI procedures were performed in accordance with the contemporary guidelines for the management of patients with STEMI and the choice of stent was left to the physician’s discretion. Use of both BMS and DES was recommended for pPCI without Level A evidence of superiority of DES to BMS, with off-label indication of DES and lack of experience of Genous, as passive coated BMS for pPCI during the study initiation phase. The inclusion criteria were STEMI according to the STEMI definition guidelines, age \>18 years, pPCI of the infarct-related artery within 12 hours of symptom onset, and use of either Genous or TAXUS Liberté stenting in the culprit lesion of the infarct-related artery. The exclusion criteria were cardiogenic shock at the time of clinical presentation of the patient, concomitant disease influencing the duration of the dual antiplatelet therapy, or requiring triple antithrombotic treatment (such as chronic atrial fibrillation) and any other disease that limited life expectancy, such as progressive cancer, chronic inflammatory disease, significant valve disease, planned coronary or other cardiovascular surgery, allergy or intolerance for the dual antiplatelet therapy, or significant hematologic disorder. The following information was recorded: clinical information (age; gender; prevalence of atherosclerotic risk factors, such as diabetes mellitus, hypertension, hypercholesterolemia or smoking; anterior location of the infarction; previous infarction or coronary artery intervention; statin treatment; and door-to-needle time); procedural information (stent size, length, number of stents implanted in the infarct-related artery, pre- and post- dilation, and stent implantation pressure); and follow-up clinical information (stent thrombosis, death, re-AMI, stroke, target lesion and vessel restenosis, target vessel failure, and re-intervention) at the pre-defined time-points. ## <span class="smallcaps">S</span>tudy endpoints The primary endpoint of the study was the patient-oriented composite clinical endpoint with the incidence of major adverse cardiac and cerebrovascular events (MACCE), including all-cause death, recurrent acute myocardial infarction (re- AMI), target vessel revascularization (TVR), and stroke during the 9-year follow-up, occurring in hierarchical order. The secondary endpoints included the a) separate primary endpoint events at the 9-year follow-up; b) procedural, in-hospital, 6-month, and yearly MACCE rates; c) acute, subacute, and late stent thrombosis (definitive, probable, or possible); d) target lesion restenosis; target lesion revascularization (TLR); target vessel, non-target lesion, and non-target vessel revascularization (non- TVR); and e) major bleeding. All endpoint definitions were in accordance with those of the Academic Research Consortium (ARC). ## <span class="smallcaps">S</span>tenting procedure and follow-up The Genous stent device has been described elsewhere. In cases with multiple lesions, the infarct-related artery was determined based on the coronary morphology, the presence of possible vessel thrombus, and the location of the ECG changes. The infarct-related lesion was treated either with a Genous or a TAXUS Liberté stent. The second lesion stenting or staged procedure was left to the physician’s discretion; all non-target lesions were preferably treated with the same stent type, but this was not mandatory. Pre- and post-stent dilations were performed at the physician’s discretion. Loading doses of 300 to 600 mg of clopidogrel and 250 mg of aspirin were administered before the procedure or immediately after diagnostic coronary angiography and before the start of PCI. Use of GPIIb/IIIa was left to the physician’s discretion. During the first year of the study, all patients were treated with dual antiplatelet therapy (DAPT) (a 75-mg dose of clopidogrel plus a 100-mg dose of aspirin daily) in accordance with the guidelines in place at the time. Patients who received a Genous stent were treated with the maximal statin dose immediately after pPCI, irrespective of their previous statin treatment dose or blood lipid levels or existing hyperlipidemia. Follow-up was performed to document all cardiac and non-cardiac events, including a clinical visit, an inspection of the medical record of each patient, or telephone contact after hospital discharge at 6 months, at 1 year, and then yearly for up to 9 years. ## <span class="smallcaps">S</span>tatistics Descriptive statistical analyses were performed using continuous variables expressed as mean values ± standard deviations (for parameter with normal distribution) or median with interquartile ranges (for parameter with skewed distribution) and using categorical variables presented as the percent frequencies. For comparisons between groups in the pre-specified sub-analysis, two-tailed t-tests or Kruskal Wallis non-parametric test were used for the continuous variables with normal distribution or skewed distribution, respectively. Chi-squared tests were used for categorical variables. A propensity score was calculated for each patient after identifying the significant predictors of MACCE by fitting a stepwise logistic-regression model with candidate variables. These variables included patient- and procedure- related predictors, such as age, gender, hypercholesterolemia, hypertension, diabetes, smoking, previous myocardial infarction, number of diseased vessels, and door-to-needle time. For the primary endpoint, the historical first event was calculated. For the secondary endpoints, all events were calculated, even if a single patient experienced more events. The time-dependent 9-year cumulative incidence rates of composite MACCE, all-cause death, and TVR were compared using Kaplan-Meier estimate supplemented by log-rank statistics. Due to the influence of age at the time of study inclusion (mean age, 62 years) and the 9-year duration of the follow-up, with age and follow-up time expected to have significant influence, the Kaplan-Meier analyses of MACCE and all-cause death were adjusted for age (≥ 62years). The time dependency of the selected covariates of the Cox regression was tested by creating interactions of the predictors and survival time function, and as expected, the age proved to be a time-dependent covariate. Since this parameter was time-dependent for both groups, we have excluded this parameter from the Cox regression model and performed the proportional hazard assumption to evaluate the effects of potential risk factors on event-free survival. For event-free and survival analyses, the following variables were considered in the multiple regression models: gender, diabetes, hypertension, hypercholesterolemia, smoking, type of implanted stent (Genous or TAXUS Liberté), left anterior descending coronary artery (LAD) as the infarct-related artery, and the number of implanted stents. For TVR, the following parameters were considered as potential predictors: gender, diabetes, hypertension, hypercholesterolemia, smoking, type of stent implanted (Genous or TAXUS Liberté), LAD as the infarct- related artery, proximal lesion, cumulative stent length, and the number of implanted stents. The estimated relative risk (hazard ratio) with 95% confidence intervals were calculated. The patient-years were calculated as the final follow-up time (days) for the individual patients (considering the time to event) and divided by 365. The number of patients at risk per year was calculated as number of events (MACCE) of the group divided by the patient-year (denominator) and given as percentage. Post-hoc analysis was performed to analyze the effects of prior statin treatment on the incidences of the following: acute and subacute stent thrombosis (both definitive); in-hospital, 6-month, 1-year, and 9-year death; re-AMI; TVR; and MACCE in patients who received a Genous stent. We considered p\<0.05 to be statistically significant. Statistical calculations were performed with SPSS for Mac version 24. # Results ## <span class="smallcaps">P</span>atients Between February 2006 and December 2008, 203 patients were included in the study who received either a Genous (n = 102) or a TAXUS Liberté (n = 101) stent. The baseline parameters did not differ between the groups. Propensity score analysis revealed a normal distribution of the individual scores in both groups and an expected non-significant difference between the groups. ## <span class="smallcaps">A</span>ngiographic and procedural data The angiographic and procedural data are presented in. The time from symptom onset to pPCI and the door-to-needle time were similar in the groups. In the Genous and TAXUS groups, 7 and 5 patients, respectively, experienced outdoor cardiopulmonary resuscitation (CPR) before transport to pPCI, with complete hemodynamic stabilization at the time of the clinical presentation. lists the angiographic and procedural data of the patients included into the study. The number of the implanted stents were 1.4±0.4 vs 1.5±0.8 in patients receiving Genous vs Taxus stents. Two patients of each group received stent type other than the original stent type on the infarct-related artery: one Cypher (Cordis, Fremont, California, U.S) and one Endeavor (Medtronic, Minneapolis, MN, U.S.) in group Genous and one rapamycin-coated Yukon (Translumina Therapeutics LLP, New Delhi, India) and 2 Endeavor (Medtronic, Minneapolis, MN, U.S.) stents in the Taxus group, due to missing size or length of the required stent; otherwise, the same stent type was used. Staged procedures were performed with one Endeavor and otherwise Genous stents in the Genous group, and only Taxus stents in the Taxus group. Stent delivery failure occurred in 1 patient in each group. No significant differences were found regarding the number of implanted stents, stent diameter and length. The lesion pre-dilation balloon diameter (1.6±1.8 vs 1.7±1.0 mm), pre-dilation pressure (6.9±5.5 vs 8.4±5.0 atm), stent inflation pressure (13.9±2.1 vs 14.0±2.5 atm), post-dilation pressure (5.1±7.6 vs 3.9±6.9 atm) were not different between Genous and Taxus groups. Glycoprotein IIb/IIIa (ReoPro) was administered due to suspected vessel thrombosis of the infarct-related artery in 18.6% and 5.0% of patients who received Genous and TAXUS Libertéstents, respectively (p\<0.001). In the Genous stent group, a total of 11 patients (10.8%) experienced acute hemodynamic instability that required peri-procedural CPR due to malignant arrhythmia or acute vessel closure post-stenting. ## <span class="smallcaps">C</span>linical outcomes The primary endpoint and the main secondary endpoints (separate primary endpoints events and the procedural, in-hospital, 6-month, and yearly MACCE rate) of the groups are presented in and. The patient-years were 595 vs 695 for Genous and TAXUS groups, and the patient- at-risk for MACCE rate per year were 7.6% vs 4.5%, respectively. lists the non-serious adverse events, stent thrombosis and bleeding events (secondary endpoints). In the Genous and TAXUS groups, 7 and 1 patients, respectively, had acute stent thrombosis (p\<0.001). There was a trend towards higher in-hospital MACCE in the Genous group versus the TAXUS group (10.8% vs. 4.0%, respectively). At the 6-month follow-up, the MACCE rate was significantly higher in the Genous group. The in-hospital, 6-month, and 1-year mortality rates were 7.8%, 8.8%, and 9.8% in the Genous group, and 2.0%, 3.0%, and 4.0% in the TAXUS group. Adverse events were rare after hospital discharge, and the incidence was similar in the two groups. The 9-year mortality was 23.5% vs. 16.8%, and the 9-year cumulative MACCE was 44.1% vs. 30.7% (p\<0.05) in the Genous vs. TAXUS stent groups, respectively. The Kaplan-Meier 9-year MACCE-free analysis revealed significantly worse outcomes for patients in the Genous group as compared to the TAXUS group. Cox regression analysis revealed both the use of the Genous stent and greater patient age influenced the time-dependent MACCE. Adjustment of the Kaplan-Meier MACCE-free analysis with age, however, did not influence the finding that use of the Genous stent was a predictor for MACCE. A trend towards a higher incidence of time-dependent mortality was observed in patients in the Genous group. According to Cox regression analysis, age significantly influenced mortality. Adjustment with age revealed a trend towards higher mortality in older patients who received a Genous stent as compared with patients with a TAXUS Liberté stent (p = 0.054). Kaplan-Meier analysis did not reveal differences between the Genous and TAXUS stent groups in terms of time-dependent TVR, which was not influenced either by age, by implantation of the Genous stent, or by other factors. Notably, target lesion restenosis with consequent target lesion revascularization was significantly higher in the Genous group as compared to the TAXUS group at the 1-year follow-up. ## <span class="smallcaps">S</span>ubanalyses Since all patients in the Genous group received maximal dose of statin post- procedure and were treated more often with GPIIbIIIa, which factors might have an influence on the post-procedural clinical events, both of these factors were also entered into the Cox regression model (post-hoc subanalysis). However, none of these factors had effect on the clinical outcome or proved to be an independent predictor for long-term event. Post-hoc analysis comparing patients in the Genous group with or without prior statin treatment showed similar rates of definitive stent thrombosis, in- hospital deaths, re-AMI, TVR, and MACCE. After pPCI, all patients with a Genous stent took the maximum recommended dose of oral statins. The clinical outcomes of patients who received the TAXUS Liberté stent was not influenced by prior statin treatment (data not shown). # Discussion To the best of our knowledge, this is the first study to investigate the long- term (9-year) safety and efficacy of Genous and TAXUS Liberté stents in patients with STEMI who underwent primary PCI. There were four main findings: 1) Patients with STEMI in the Genous group had a higher incidence of acute and subacute definitive stent thrombosis and target lesion revascularization. 2) Patients in the Genous group had a higher incidence of cumulative MACCE at all follow-up times, which was due mainly to a higher death rate, especially in-hospital and at the 6-month and 1-year follow-up. 3) In both stent groups, the event rate was very low after the first year of follow-up in patients who survived the peri- infarct and in-hospital periods. 4) Implantation of the TAXUS Liberté stent in patients with STEMI showed a long-term safety profile with low mortality rates and low adverse cardiac and cerebrovascular event rates. ## Analysis of the Genous stent results in the context of the literature Tables and summarizes the main data from the literature on Genous stents and lists the main outcomes. Most studies were conducted in patients with noncomplex coronary lesions who underwent elective PCI. The safety and efficacy of Genous stent implantation was first confirmed in the HEALING-FIM study, which included 16 patients with stable or unstable angina or silent ischemia and reported just 1 adverse event, a TVR, at the 9-month follow-up. The subsequent HEALING II study (n = 63) and e-HEALING registry (n = 4939) showed similar beneficial results, with low incidence of TVF and a stent thrombosis rate of 1.1%. However, the first randomized studies in patients with de novo lesions with predicted high-risk restenosis (chronic total occlusion, lesion length \>23 mm, vessel diameter \<2.8 mm, or any lesion in a diabetic patient) requiring non- urgent PCI failed to reach the primary endpoint and reported higher late lumen loss at 6months and a higher incidence of TVR and TVF at the 1- and 2-year follow-ups. In addition to reporting that the Genous stent proved to be non- inferior in comparison to the TAXUS Liberté stent, the authors found no stent thrombosis in the Genous group. Co et al. reported the first prospective observational registry data in STEMI patients who received Genous stents and concluded that this procedure is safe, with very low mortality rates of 0.8%, 2.4%, 3.3%, and 3.3% for in-hospital, 1-month, 6-month, and 1-year mortality after STEMI. The only exclusion criterion of this study was cardiogenic shock. Patients received aspirin indefinitely but clopidogrel for only 1 month. Platelet glycoprotein IIb/IIIa inhibitor was used in 14.2% of the patients, and 58% received adjunctive thrombectomy device therapy before stent implantation. In contrast, an Indian pilot study with angiographic follow-up of 11 patients with STEMI reported a high in-stent restenosis rate and high late lumen loss of the infarct-related artery that was treated with the Genous stent. These results were confirmed by the larger angiographic follow-up study by Low et al. that reported a binary restenosis rate of 28% of Genous stent 1-year post-infarction, even though no stent thrombosis occurred. In contrast, implantation of Genous stents in our patients with STEMI resulted in an unusually high rate of periprocedural stent thrombosis, with consequently higher mortality up to the 1-year follow-up. In our study, DAPT was prescribed for 1 year after STEMI; thus, discontinuation of DAPT could not have influenced the 1-year results. Since our STEMI patients who received the TAXUS Liberté stent in the same timeframe showed favorable short- and long-term clinical outcomes, the clearly worse outcomes of the Genous stent patients cannot be explained by differences in study design and population, the use of thrombectomy device, or platelet glycoprotein IIb/IIIa inhibitors. Notably, our patient population was older than the populations in other studies. Similar to the results of the E-Healing study, the 1-year results of the Genous stent in an elderly population showed a significantly higher event rate compared with younger patients. This was mainly driven by higher mortality. After the first year, adverse events were rare in our patients, and it seems likely that such events were due mainly to the multiple comorbidities of the aging patients. ## Comparison of our data with literature data regarding long-term clinical outcomes It is difficult to compare our data with data from other studies because there are few reports of long-term clinical outcome data of patients with STEMI and primary PCI with either balloon dilation or implantation of any kind of stents. The 9-year mortality of patients with STEMI and primary PCI was 27% in an unselected cohort in the Warsaw STEMI registry, which was an all-comers study that had no exclusion criteria. The registry also included patients who were in cardiogenic shock. TIMI flow 0 or 1 was observed in a high proportion of patients (82%), and 77% of patients received an intracoronary stent in the infarct-related artery. A subgroup of patients in this registry with total ischemic time\< 3 h (i.e. the time interval between the onset of symptoms and first balloon inflation) had mortality rates of 7% and 27% at 1 and 9 years, respectively. In contrast, all of our patients received intracoronary stents, and complete or functional occlusion of the infarct-related artery (TIMI flow 0 or 1) was less frequent, with higher post-PCI TIMI 3 flow grade. For comparison, the patients in our study had 1- and 9-year mortality rates of 9.8% and 23.5% in the Genous group and 3.9% and 16.8% in the TAXUS group. The SIRTAX VERY LATE trial included patients with selected coronary lesions who underwent elective and urgent PCI with first-generation sirolimus- or paclitaxel-eluting stents. That trial reported higher rates of ischemia-driven target lesion revascularization and late and very late definitive stent thrombosis (1.9% and 5.3% at the 1- and 10-year follow-up) with mortality rates of 2.6% and 24.2% at the 1- and 10-year follow-up. Our STEMI patients who received the Genous stent showed a high incidence (12.4%) of definitive stent thrombosis, which included acute, subacute, late, and very late stent thrombosis; of the 12.4%, 9.8% occurred within the first month after primary PCI. This can only be explained in part by STEMI that was accompanied by elevated acute thrombotic risk in the coronary arteries. In contrast, implantation of the TAXUS Liberté, a second-generation DES, resulted in a 5% definitive stent thrombosis rate during the entire 9-year follow-up. ## The Genous endothelial progenitor cell capturing stent: pro-healing technology In the E-healing registry, it was recommended that patients with planned Genous stent placement receive at least 2 weeks of statin therapy prior to PCI and DAPT for at least 1-month post-procedure with aspirin taken indefinitely. This recommendation was based on the observation that statins increase the peripheral concentration of circulating EPCs. Patients with low levels of circulating EPCs responded poorly to the Genous stent in the HEALING II study. However, the HEALING-IIB study revealed that treatment with the maximum daily dose of atorvastatin 2 weeks prior to and at least 4 weeks after placement of the Genous stent did not reduce the restenosis rate. Subgroup analysis of our cohort who received Genous stents found that prior statin treatment had no influence on the clinical outcome. In fact, complete coverage of the stented artery seems to be independent of the number of the circulating EPCs. The role of EPCs in endothelialization and antithrombotic processes is complex. EPCs can affect the expression or downregulation of thrombotic and coagulation markers, attract CD3 cells, and influence the activity of the vascular smooth muscle cells. However, accelerated endothelialization is still not completely understood, and there are conflicting experimental reports. In addition, intravenous infusion of bone marrow-derived mononuclear cells aggravates the formation of atherosclerotic lesions and contributes to a more vulnerable atherosclerotic plaque that has increased microvascularization and enhanced lipid cores covered by thinner fibrous caps. It seems that formation of a functional endothelial layer from EPCs requires a sequence of signaling events that ranges from cell mobilization, migration and adhesion to cellular differentiation to vascular endothelial cells that form a functional endothelial layer. However, neither our study nor other studies have confirmed that rapid endothelialization has a beneficial effect on stent thrombosis or that it decreases restenosis. It seems likely that factors other than rapid endothelialization play important roles in suppressing thrombus formation. ## Clinical importance of the findings One strength of the current study was the fairly long observation period of STEMI patients who received either a first-generation Genous stent or a second- generation paclitaxel-eluting stent. Since millions of people have received the Genous and TAXUS Liberté stents and continue to live with them, it is helpful to know that the TAXUS Liberté stent is safe. The occurrence of adverse events after the first year post-STEMI was associated with aging, even in patients who received Genous stents. Due to the reported higher TVF in TRIAS-HR study, the Genous stent platform has been changed, from stainless steel to cobalt-chromium, and also to drug-eluting stent system (COMBO Dual Therapy stent, which combines sirolimus with the endothelial progenitor cell-capturing layer), resulting in fewer thrombotic and other cardiac adverse events in clinical studies that are currently on-going or recently published. ## Limitations We did not calculate TVF, which is a composite of cardiac deaths, target vessel- related AMI, and TVR. This is because we did not always have the exact cause of death after hospital discharge. However, MACCE is a similar clinical outcome as TVF in terms of cardiac and cerebrovascular events. This was a retrospective cohort study that was not randomized. However, at the time of patient inclusion, the use of DES was not recommended (although it was not prohibited) for pPCI in STEMI; accordingly, there were no data from randomized clinical studies that used Genous stents in STEMI. Importantly, propensity score analysis confirmed that the Genous and TAXUS Liberté stent groups were comparable. In fact, a strength of our study was that this group of patients collectively represent a real-world clinical scenario, since the study included consecutive patients and only excluded cardiogenic shock or limited life expectancy status in the chosen time frame. **In conclusion**, first-generation Genous stents used for primary PCI in STEMI had higher procedural and peri-procedural mortality plus higher in-hospital and short-term (1-year) mortality than TAXUS Liberté stents. The TAXUS Liberté stent had more favorable 9-year clinical outcomes. # Clinical Perspectives ## What\`s known? Several small and larger cohort and randomized studies that have included patients with stable coronary lesions have demonstrated the safety and efficacy of the Genous stent in reducing target vessel revascularization, restenosis, and stent thrombosis. ## What\`s new? This is the first study to investigate the long-term (9-year) safety and efficacy of Genous and TAXUS Liberté stents in patients with STEMI who underwent primary PCI. Patients with STEMI in the Genous group had a higher incidence of acute and subacute definitive stent thrombosis, target lesion revascularization and incidence of cumulative major adverse cardiac and cerebrovascular events. Implantation of the TAXUS Liberté stent in patients with STEMI showed a long- term safety profile with low mortality rates and low adverse cardiac and cerebrovascular event rates. ## What\`s next? The platform of the Genous stent has been changed from stainless steel to cobalt-chromium, and also to drug-eluting stent system, resulting in fewer thrombotic and other cardiac adverse events in clinical studies that are currently on-going. # Supporting information AMI acute myocardial infarction BMS bare metal stents CPR cardiopulmonary resuscitation DAPT dual antiplatelet therapy DES drug-eluting stents EPC endothelial progenitor cells LAD left anterior descending artery MACCE major adverse cardiac and cerebrovascular events pPCI primary percutaneous coronary intervention STEMI ST-segment elevation myocardial infarction TLR target lesion revascularization TVR target vessel revascularization TVF target vessel failure [^1]: The authors have declared that no competing interests exist.

# Introduction As researchers miniaturize biosensors and microfluidic devices down to submicron scales\[–\], high-resolution biomolecule conjugation compatible with these processes has become desirable. Several protein patterning techniques have been demonstrated, including gasket-based patterning, microcontact printing (μCP)\[–\], and dip pen lithography (DPL)\[–\]. Gasket-based patterning is the easiest and arguably most popular protein patterning approach in academic work. In this approach, removable polymer gaskets, usually made of polydimethylsiloxane (PDMS), are used as a mask on the conjugation surface to localize surface functionalization reagents. Since alignment is usually performed manually, it is difficult to precisely align the polymer gaskets with existing surface structures. This method is suitable for low resolution and low throughput research projects, but is difficult to scale up. DPL and μCP are both based on the contact between small probe tips with surface chemistry reagents and a substrate surface. μCP uses PDMS stamps with micron or, more recently, submicron features to pattern functionalization reagents onto a substrate. This method exhibits higher resolution and throughput than that of gasket-based methods, but the alignment between desired protein patterns and existing features is limited by the alignment tool. DPL involves using an atomic force microscopy (AFM) tip to trace out the functionalization patterns and can achieve nanoscale resolution with submicron alignment, but usually requires custom AFM systems and is thus difficult to scale up for bulk manufacturing. To achieve high resolution, precise alignment and bulk manufacturability, engineers have long turned to photolithography techniques to pattern surface moieties. For example, lithographically patterned gold films can be functionalized with thiol groups to pattern proteins onto gold surfaces selectively\[–\]. However, gold thin films may have strong interference with fluorescent signals, so this process may not be suitable for some cases requiring a fluorescent readout. Another common technique involves depositing a sacrificial layer, such as common photoresists, to occlude a region of the surface where conjugation is unwanted, and subsequently dissolve the sacrificial layer with a biocompatible solvent. However, photoresists can be problematic for surface chemistries that require organic solvents, since photoresists dissolve quickly in most of the commonly used organic solvents (e.g. ethanol, acetone, isopropyl alcohol (IPA) and dimethyl sulfoxide (DMSO)). For example, silane is commonly used for silica surface functionalization, but due to the hydrolysis of silane groups in aqueous solution, an organic solvent like ethanol or DMSO is preferred for silane dissolution in order to maintain the activity of silane groups. A sacrificial layer which will dissolve in a biocompatible solution but not in commonly used organic solvents would enable precise surface conjugation across a wider range of surface chemistries. While there are alternatives to photoresist such as using a dissolvable metal as a sacrificial layer in protein lift off, the dissolution rates of these metals are low at neutral pH, they are difficult to deposit, and some are flammable. Here, we demonstrate a simple, inexpensive and manufacturable protein patterning technique with submicron resolution and only nanoscale misalignment using germanium (Ge), a semiconductor material, as a sacrificial layer for a lift-off process. This method is suitable for creating patterns of a target material on a surface using a sacrificial material. Ge has been studied for decades in the field of microelectromechanical systems (MEMS) and has been used as a sacrificial layer due to its solubility in aqueous solutions. Ge dissolves slowly in aqueous solution, but is not affected by most organic solutions. Germanium is first oxidized into germanium dioxide by oxidizing agents (e.g. dissolved oxygen molecules or aqueous hydrogen peroxide (H₂O₂)). H₂O₂ accelerates the oxidation of Ge since it is a much stronger oxidizing agent than oxygen molecules. The spontaneous dissolution of germanium dioxide in water (containing some oxidizer) exposes more Ge for further oxidation until complete dissolution of the Ge sacrificial layer occurs. In brief, our process is as follows: A thin layer of germanium is first patterned. The entire surface (containing germanium and non-germanium coated surfaces) is then functionalized and the germanium is then dissolved away, leaving biomolecules only on the surfaces that originally were uncoated by germanium. The whole process is scalable and easy to apply in any microfabrication lab, as it relies on standard thin film techniques. In addition, germanium and its dissolution products are biocompatible, which enables the use of this technique for both *in vitro* and *in vivo* devices. In this paper, we report protein patterning results from two different Ge deposition processes, E-beam evaporation and low temperature chemical vapor deposition (LPCVD). Since E-beam evaporation is more directional (more material deposited on horizontal surface than on vertical sidewalls) and LPCVD is more conformal (same film thickness on both vertical and horizontal surfaces), LPCVD protects vertical walls from biomolecule attachment while E-beam evaporation enables sidewall functionalization, which is a topic of recent interest. Both biotin and streptavidin surface activity were characterized with various incubation periods in 0.35% H₂O₂. To our knowledge, this study is the first demonstration of protein nanoarray fabrication with Ge as a sacrificial layer. # Experimental methods Substrate features were first patterned onto p-type silicon wafers by conventional photolithography techniques. Deep ultraviolet (DUV) photoresist (Dow UV210-0.6) was spun at 7000 rpm for 420 nm final thickness, exposed in ASML 5500/300, developed in MF26A for 1 min and hard baked for 2 h at 120°C (all DUV photoresists were patterned with these conditions unless otherwise indicated). Wafers were then etched in a Lam TCP 9400SE Etcher to provide 120 nm-deep alignment markers, which enabled alignment of protein patterns with substrate features in subsequent germanium patterning. In this paper, we etched 500 nm deep square wells into the substrate in a Lam Etcher system to demonstrate alignment accuracy and sidewall functionalization. After well etch, we deposited a layer of high-temperature oxide (HTO) in a TYTAN Diffusion Furnace System at 900°C. HTO was not necessary for our protein patterning technique with a Ge mask, but provided a blank surface with binding sites for silane chemistry. If other surface chemistries are used, the HTO step can be skipped and the Ge process applied after substrate fabrication. There are two Ge fabrication methods presented in this paper. In method 1, we used an E-beam evaporation and lift-off process to pattern Ge. DUV photoresist (Dow UV210-0.6) was spun at 1480rpm to form a 900nm sacrificial layer and patterned to cover the areas to be functionalized. In this work, we patterned micron scale squares inside the previously etched square wells as a demonstration. 50 nm thick germanium was evaporated on top of the patterned photoresist in a CHA E-beam Evaporator at 10⁻⁶ Torr base pressure. The surface conjugation areas were then exposed by stripping photoresist in 1-methyl-2-pyrrolidone (NMP). In method 2, we deposited Ge first with LPCVD and then etched away Ge in surface functionalization areas. 50 nm thick germanium was deposited in a TYTAN Diffusion Furnace System at 340°C after depositing a layer of 2 nm amorphous silicon as an adhesive layer between HTO and Ge. Then, 420 nm DUV photoresist was patterned over the germanium. Germanium exposed through the photoresist was completely etched away in a Lam TCP 9400SE Etcher to open windows for protein patterning in the next step. The plasma etching of germanium had high selectivity against HTO so that HTO also worked as a stop layer during germanium etching. After Ge patterning with either method, we brought the substrates in contact with 25 mg/ml silane-PEG-biotin (Nanocs, New York, NY) in ethanol or dimethyl sulfoxide (DMSO) and incubated for 30 min at room temperature. After washing away excess silane solution, wafer substrates were immersed in deionized (DI) water with 0.35% hydrogen peroxide to strip Ge. After the dissolution of germanium, wafers with biotin patterns were washed in DI water and blown dry with nitrogen, followed by storage in a 4°C fridge for future use. Germanium dissolution rate as a function of H₂O₂ concentration, temperature and pH had been well characterized in literature. Our previous work also demonstrated that nanometer-thick Ge dissolved in low concentration H₂O₂ in minutes while no significant dissolution discovered in all tested organic solvent. Before surface conjugation with silane chemistry, dissolution of Ge was tested in 0.35% H₂O₂ and surfaces were imaged with a Gemini Scanning Electron Microscope (SEM) before and after dissolution. Substrates were first patterned with 25 μm by 25 μm square wells that were etched 500 nm deep, followed by 100 nm HTO deposition. A layer of 50nm Ge was deposited in an LPCVD furnace and 3 μm by 3 μm squares were etched inside the square wells. The etching recipe had a 20 s overetch after the 50nm Ge etch to ensure all Ge and amorphous Si on top of the HTO was etched away. Substrates were then imaged in SEM to identify the Ge surface versus oxide surface. After immersion in 0.35% H₂O₂ for either 10 min or overnight at room temperature, substrate surfaces were imaged by SEM again to characterize the surface after Ge dissolution. It was very easy to distinguish oxide from Ge in SEM, since the oxide surface was smoother and brighter than the Ge surface which had micron- size grains. Prior to functionalization, 25 μm by 25 μm and 42 μm by 42 μm square wells were etched 500 nm deep into the silicon wafer, followed by 100 nm HTO deposition. All samples were separated into two groups. Group 1 and 2 were fabricated and functionalized using Method 1 and 2, respectively. 3 μm x 3 μm and 5 μm x 5 μm squares patterned inside 25 μm x 25 μm and 42 μm x 42 μm square wells (16 squares per well) were biotinylated using silane-PEG cross linker (see *Fabrication*). Wafers were then incubated with 1% bovine serum albumin (BSA) blocking solution for 1 hour at room temperature to prevent nonspecific binding. Streptavidin conjugated with Alexa-Fluor 647 (Strep-647) was used to fluorescently label the biotinylated surface by incubating 10 μg/ml Strep-647 in 1X PBS with substrate surfaces for 30min at room temperature, after which Ge was removed by 10 min-immersion in 0.35% H₂O₂ with mild stirring. Since Ge is dissolved after Strep-647 labelling, it was not strictly necessary to incubate in blocking solution; the step was carried out to be certain that Strep-647 bound to biotin specifically (as streptavidin is known to bind nonspecifically to surfaces). We did not find significant signal decrease due to oxidation in H₂O₂, so all Strep-647 labeling was done before Ge dissolution (with the exception of the experiment to test biotin activity in H₂O₂). Fluorescent images were taken using a Nikon Eclipse Ti fluorescent microscope after washing and blowing dry. The resolution of proteins patterned with Ge hard masks was limited by the photolithography process and exposure tools. In order to show that the limit was attributable to lithography limits, we pushed the resolution of our protein patterning down to 250 nm, the limit of our ASML model. For this experiment, 500 nm deep square wells were not etched before HTO deposition, since a 500nm height difference disrupted the uniformity (even distribution of photoresists) of DUV photoresist coating (420 nm thick) and affected submicron lithography. Thus, protein nanoarrays were fabricated on flat silicon wafers with 100 nm HTO and Method 2 was used to pattern 50 nm Ge film with 1 μm x 1 μm, 500nm x 500nm, 300nm x 300nm and 250nm x 250nm openings for conjugation. Biotinylation of nanoarrays was performed as per *Fabrication.* Strep-647 was then used as described above to label biotin molecules on the surface. Fluorescent images were taken with a Carl Zeiss Elyra SR.1 Super Resolution Microscope in order to resolve the 250 nm squares. As the concentration of H₂O₂ used for Ge dissolution was very low, both biotin and streptavidin surface exhibited sufficient activity for the following experiments. Biotin-conjugated chips were fabricated as above. Biotin activity was demonstrated by incubating biotin conjugated chips in 0.35% H₂O₂ for various periods from 10 to 360 min at room temperature. For streptavidin-activity tests, biotin conjugated chips were firstly incubated with 100ug/ml non-fluorescent streptavidin solution for overnight to get streptavidin surfaces. Streptavidin conjugated chips were immersed in 0.35% H₂O₂ for different incubation periods as per biotin groups. (Note, sufficient H₂O₂ solution and a sealed the container should be used for long incubations to prevent drying. If the H₂O₂ solution dries or too much water evaporates during incubation, neither biotin nor streptavidin survive). After H₂O₂ incubation, biotin and streptavidin chips were blocked in 1% BSA for 1h at room temperature before brought in contact with Strep-647 and Atto-488 Biotin (Biotin-488), respectively. Fluorescent images were then taken with a Nikon Eclipse Ti fluorescent microscope after washing and blowing dry. # Results & discussion Germanium dissolved rapidly in 0.35% H₂O₂ at room temperature. As shown in, each 25 μm x 25 μm well had sixteen 3 μm by 3 μm squares for functionalization. The exposed oxide area was smoother than the Ge layer, for which micro-scale grains were evident. After 10 min incubation in 0.35% H₂O₂, it was very clear that all Ge grains were completely removed and overnight incubation didn’t remove more material. Any residual Ge on the oxide surface would be very obvious in SEM. Thus, a 10min incubation in 0.35% H₂O₂ was sufficient to remove a 50nm Ge sacrificial layer. After Ge dissolution, 3 μm by 3 μm squares were still visible under SEM. This arose because the additional over-etch step in the Ge etching process etched away the 2 nm Si layer in those 3 μm by 3 μm conjugation regions, while the 2 nm Si layer elsewhere did not dissolve in H₂O₂ during Ge dissolution. As Strep-647 was dissolved in 1X PBS, the germanium layer oxidized and dissolved slowly during the 30min labeling process. The Ge dissolution rate in 0.35% H₂O₂ was accelerated if done after a 30min incubation in PBS. Importantly, a 30min incubation in PBS was insufficient to dissolve a 50 nm Ge layer. The germanium did not exhibit appreciable etching until 0.35% H₂O₂ was added. This is evidenced by the low background intensity in as the intensity of the unconjugated area was not higher than that of a blank silicon wafer. On the other hand, the Ge surface integrity was less important since a biotin pattern was formed on chip surface during the silane chemistry step. Even if the Ge layer completely dissolves before streptavidin conjugation, streptavidin should only be immobilized on biotin surfaces. Thus, after biotin conjugation, Incubating chips in aqueous solution is not a problem. All biotinylated regions were successfully conjugated by fluorescent streptavidin. The results for both Ge patterning techniques, Method 1 & 2, are shown in, respectively. The fluorescent sidewalls in pointed out by red arrows indicate sidewall conjugation (in contrast to, which did not exhibit this). The intensity profiles along yellow dash lines in are shown in respectively. Four major peaks represent 4 squares labeled with fluorescent streptavidin. There are two additional little peaks caused by sidewall conjugations on each side of the 4 major peaks, highlighted with red arrows in. These additional intensity peaks have much lower intensity than the major peaks, but still high enough to be distinguished from background, which are however not exhibited in at all. This sidewall conjugation arose because of the non-conformal deposition of Ge evaporation: evaporated ultrathin Ge films (e.g. 50 nm) covered only horizontal surfaces and left vertical sidewalls exposed for biotin conjugation. In contrast, LPCVD Ge films of the same thickness protect vertical sidewalls from surface chemistry. The use of modern deep ultraviolet (DUV) lithography enabled nanoscale protein patterning. As shown in, all features ranging from 250–1000 nm were successfully resolved and conjugated with biotin. No nonspecific binding was evident. The feature rounding on smaller features was an artifact of submicron lithography. Both biotin and streptavidin could maintain activity after H₂O₂ incubation, as shown in. The activity was measured in terms of the number of fluorescent biotin or streptavidin molecules bound on streptavidin and biotin surfaces, respectively. Thus, the fluorescent intensity was proportional to the remaining activity of the surface molecule. With 10min 0.35% H₂O₂ incubation, both biotin and streptavidin showed \~80% activity. Even after immersing in H₂O₂ for 6h, biotin and streptavidin surface still had more than 50% activity remained. Longer incubation was not required for 50nm Ge dissolution; here the longer incubation periods were used to show biotin and streptavidin stability in H₂O₂ in case a longer incubation was needed for other applications. # Conclusion In this work, we demonstrate the use of germanium sacrificial layers to achieve biomolecule patterning which is compatible with nano/microfabrication, scalable to high resolutions with precise alignment, and ultimately suitable for low cost and high volume manufacturing of biosensors. In addition, the method is applicable to a wide range of biomolecules and surface chemistries. It is of particular utility for surface conjugation chemistries which require organic solvents. Examples include the commonly used primary amine reactive chemical groups for protein labeling, such as N-hydroxysuccinimide ester (NHS), which hydrolyzes quickly in water and the conjugation of proteins onto silicon dioxide surfaces using silane-PEG-NHS. Because both functional groups hydrolyze in water, organic solvents are usually required and a germanium sacrificial layer would be of significant use. Two processes are reported in this paper, each of which has specific advantages. Method 1 uses evaporation and a lift-off process to pattern the Ge hard mask. This process avoids the high temperature deposition process of Ge and the aggressive etching process in Method 2. For LPCVD processing, all wafers must be cleaned in piranha and hydrogen fluoride (HF). This mandates that there be no organic materials on substrate, which is often not the case for biosensor applications. Since Method 1 uses a lift off technique instead of etching, there is no need to have a stop layer (such as the HTO layer in our case) which is compatible with both the surface chemistry and the etching process. Thus, Method 1 is compatible with low-melting-point biomaterials underneath the Ge layer. We demonstrated nanoscale protein patterning only with Method 2. This is due to lift-off process in Method 1 requiring submicron pillars, which are much harder to resolve than submicron openings with positive photoresists. If an appropriate negative photoresist is available, Method 1 can also be used to resolve nano- features. Furthermore, sidewall functionalization can be achieved and controlled by using a non-conformal Ge deposition process (as in Method 1). On the other hand, the Ge film in Method 2 is much more uniform over the whole surface due to the conformal LPCVD process. It can successfully avoid protein immobilization on sidewalls, which may be preferred for some biosensor applications. This work employed biotin/streptavidin conjugation to demonstrate the germanium sacrificial proof of concept; both biotin and streptavidin surfaces were stable enough to maintain activity after 0.35% H₂O₂ incubation. We also used streptavidin as an example of the method’s use for protein patterning. Of note, streptavidin can be directly used as a cross linker to immobilize biomolecules. For example, after Ge dissolution, a protein can be immobilized on a biotin surface by conjugating a target protein with streptavidin first. Alternatively, streptavidin can be immobilized onto the surface first to capture biotin conjugated proteins onto the surface after Ge dissolution. We believe this method could have wide applicability across a variety of protein patterning applications. This work was funded by Berkeley Sensor and Actuator Center (BSAC) funds and DARPA Advanced Study HR0011- 16-C-0023. Michel M. Maharbiz is a Chan Zuckerberg Biohub investigator. [^1]: The authors have declared that no competing interests exist.

# Introduction Leucine-rich repeats (LRRs) are 20–29-residue sequence motifs present in a number of proteins with diverse functions. In the 3D structures, each LRR corresponds to one coil of the solenoidal fold. The coils consist of a β-strand and mostly α-helical elements (can also be 3₁₀ helix or polyproline helix) connected by loops. The coils are arranged so that all the strands and helices are parallel to a common axis, resulting in a non-globular, horseshoe- shaped molecule with a curved parallel β-sheet lining the inner circumference of the horseshoe and the helices flanking the outer circumference. In LRR proteins, a six-residue motif LxxLxL is conserved (x can be any amino acid and L-positions can be occupied by Leu, Val, Ile, and Phe), and in the known structures corresponds to a turn and a consecutive β-strand; whereas the remaining parts of repeats may be very different. While the invariant motif of the β-region is a characteristic feature of the entire LRR superfamily, the consensus sequences of the variable part suggest several specific subfamilies. LRR proteins can be subdivided into at least seven subfamilies. The repeats from different subfamilies retain a similar solenoidal fold and non-globular horseshoe shape but differ by 3D structures of individual repeats. Based on sequence analysis, it was concluded that LRRs from different subfamilies never occur concomitantly within one LRR protein. This observation is explained by mutually exclusive inter-coil packing arrangement of LRRs from different subfamilies. Such a relationship for LRRs suggests that LRR proteins of different subfamilies most probably have emerged independently during evolution rather than descended from a common ancestor. In line with this conclusion, the described LRR subfamilies could be assigned to a specific subgroup of eukaryotes or prokaryotes, and share similar functions and cellular locations. For example, the bacterial LRR subfamily with the shortest known LRRs contains only extracellular proteins of Gram-negative bacteria. The Plant-Specific LRR (PS-LRR) subfamily has exclusively extracellular proteins from plants. Proteins of ribonuclease inhibitor-like LRR (RI-LRR) subfamily are intracellular and all belong to the Metazoa kingdom. Since 1998, when these conclusions were formulated, a large number of new LRR proteins have been identified and several new 3D structures of LRR proteins have been determined. After a lapse of nine years, the classification of the LRRs and most of the previously made conclusions, including the mutual exclusive rule, withstand the test of time. At the same time, the analysis of some newly identified LRRs shows that their assignment within the existing classification of the LRR subfamilies may lead to confusion. Recently, it was shown that the phytopathogenic bacterium *Ralstonia solanacearum* encodes several type III effectors, called GALA proteins, that contain F-box and LRR domains. The F-box domain enables the interaction with SKP1 in the SCF-type E3 ubiquitin ligase protein complex. Their LRRs (hereafter GALA-LRR) have a specific consensus pattern with characteristic differences from the previously described consensus sequences of LRR subfamilies, especially from the known bacterial LRR subfamilies. On the other hand, among the LRR subfamilies that are closest to *R. solanacearum* GALA-LRRs there is the Cysteine-Containing LRR (CC-LRR) subfamily of plant, animal and fungi proteins which can also contain the F-box domains and, therefore, may have a similar function. Thus, it was not clear, whether the GALA-LRR proteins are members of the CC-LRR subfamily or they should be assigned to a new LRR subfamily. Here we clarify this ambiguous case by using sequence analysis and molecular modeling. We also focus our analysis on the origin and evolution of GALA proteins from *R. solanacearum*. # Results and Discussion ## Sequence analysis of GALA LRRs Analysis of F-box containing GALA proteins from *Ralstonia solanacearum* shows that their 24-residue long LRRs have a specific consensus pattern that has characteristic differences from the previously described LRRs. Comparison of GALA-LRRs with the other known 24-residue LRRs such as typical LRRs, PS-LRRs shows that GALA-LRRs frequently have Ile instead of Leu in position 5, Gly or Ala instead of Leu in position 9, Ala instead of Pro in position 10, and do not have a conserved Leu in position 16. The GALA-LRR consensus motif also has some differences with the 26-residue CC-LRR motif. For example, positions 3 and 16 of the GALA-LRR motif do not have a conserved Cys and position 6 is frequently occupied by Gly instead of Thr. Using the generalized profile technique, a sensitive method for sequence database searches, we found the GALA-LRR type of repeats in about 40 proteins including proteins of the cucurbit crops pathogenic β-proteobacterium *Acidovorax avenae subsp. Citrulli*, the human pathogen, and γ-proteobacterium *Legionella pneumophila*, as well as in the aquatic planctomycete bacterium *Gemmata sp.* Wal 1. These proteins, unlike *R. solanacearum's* GALA proteins, don't contain an F-box domain. Sometimes their entire sequence corresponds to the LRR domain. Some proteins have LRRs that are similar to GALA-LRR (GALA-like or GL-LRR hereafter), however, their consensus sequence has several characteristic differences from GALA-LRRs such as Val instead of Ile in position 5, Leu instead of Ala in position 10, presence of conserved Leu in position 16. Remarkably, isolated examples of GALA-LRR are found in GL-LRR domains of two F-box containing proteins from plants. Sequence database searches with generalized profiles revealed GL-LRRs in more than a hundred LRR proteins. Among them are plant proteins, and also proteins from bacteria *(Gemmata sp. Wa1-1*, *Parachlamydia sp.*, *Legionella pneumophila*, *Rhodopirellula baltica)*, protists (*Entamoeba histolytica*, *Leishmania*, *Trypanosoma, Dictyostelium*) and animals (*Danio, Tetraodon, Drosophila, Anopheles gambiae, Xenopus, Strongylo, and Homo sapiens*). Interestingly, some of the GL-LRR proteins from plant, animal and protista also contain F-box domains. ## Place of GALA-LRRs and GL-LRRs in the classification of LRR proteins Although, the newly identified GALA-LRRs and GL-LRRs do not perfectly fit any of the previously described consensus sequences of seven LRR subfamilies, they have some similarities in the consensus sequences with CC-LRRs. In particular, a characteristic ITD-motif of CC-LRRs (positions 5 to 7) is aligned with similar Igd- and Vtd-motifs of GALA- and GL-LRRs respectively. Furthermore, conserved apolar residues in positions 5, 10, 13, 19, 22 and 24 of the 24-residue-long GALA- and GL-LRRs can be aligned to the 26-residue-long CC-LRRs by deleting a residue in each of the two connecting loop regions of the CC-LRRs. These loop regions are known to be the most accommodative for such length differences. Interestingly, many of the CC-LRR proteins, similarly to GALA-LRR and GL-LRR proteins, contain F-box domains. Hence they can share functional similarity in that they recruit proteins, *via* their LRRs, to the SCF-type E3-ubiquitin ligase complex. On the assumption of the membership of GALA and GL-LRRs in the CC-LRR subfamily, the previously proposed CC-LRR consensus requires modifications. The updated CC- LRR consensus sequence is shown on. In this motif, Cys is not the only residue that occurs in positions 3 and 16: the other frequently occurring residues are Thr and Asn in position 3 and Leu, Asn and Ser in position 16. The updated CC- LRR often has Gly in addition to Thr in position 6. Finally, position 9 is frequently occupied by Gly. Database searches with the updated CC-LRR generalized profiles were able to detect such domains in heterogeneous group of about thousand proteins. Among these newly defined CC-LRR proteins there are not only proteins of animal or plant origin, but also proteins from pathogenic or non-pathogenic microorganisms such as bacteria (*Parachlamydia sp.),* protista (*Dictyostelium)* and fungi (*Cryptococcus, Candida albicans, Candida glabrata, Neurospora*). The sequence profiles and search results can be viewed on a dedicated web-page <http://bioinfo.montp.cnrs.fr> (Tools\>Profiles\>Show Profiles). Interestingly, about 600 of them have an F-box domain. This strongly supports a similar role for these specific LRRs in binding the protein substrate that is then recruited by the SCF-type E3-ubiquitin ligase. ## Structural implications The knowledge of even one 3D protein structure in a given sequence alignment provides a powerful means to test the correctness of the whole alignment. In our case, the 3D structure of one CC-LRR protein, the human F-box protein Skp2 is known, and was used to verify the alignment of GALA- and GL-LRRs with CC-LRRs. The analysis shows that all conserved and apolar residues of GALA- and GL-LRRs in the suggested alignment (positions 5, 10, 13, 19, 22 and 24) correspond to the residues of Skp2 CC-LRRs that form the hydrophobic core inside of the structure. In the conserved position 3, GALA-LRRs have an Asn residue and GL- LRRs have a Thr residue instead of a Cys in CC-LRRs. This is an additional support for the alignment, because, in general, position 3 of LRRs tolerates a few amino acid residues including mentioned Asn, Thr and Cys. These residues being in position 3 can form specific hydrogen bonds with the peptide groups of the backbone. The two extra residues in the typical 26-residue-long CC-LRRs compared to the 24-residue-long GALA- and GL-LRRs, in the alignment are located in the loop regions of CC-LRRs connecting α-helices and β-strands. The LRRs of Skp2 are variable in length and some of them are 1–2 residues shorter than typical 26-residue CC-LRR. The superposition of the 3D structures of these LRRs revealed that the loops are the most variable regions. In particular, missing residues of the short 24- and 25-residue LRRs of Skp2 are located in the loops. These structures represent good examples of how each of two loops of the CC-LRR can accommodate the loss of one residue. One of these short LRRs from Skp2 crystal structure (residue 2211 to 2235) was used as a template for construction of the GALA-LRR model (see for details). shows structural models of a single GALA-LRR and a complete set of LRRs (12 repeats) from GALA4 of *Ralstonia solanacearum* (strain MolK2, personal communication C. Boucher and S. Genin). In, the superposition of the GALA-LRR model and the crystal structure of Skp2 demonstrates that the difference between them is in the loops connecting α-helices and β-strands. The conserved asparagine of GALA-LRR (position 3) similarly to the majority of the other LRR structures, located right after the β-strand so that it is able to form a network of specific hydrogen bonds with these NH and CO groups, thus satisfying their hydrogen-bonding potential in the hydrophobic core of the structure. The conserved bulky aliphatic residues form the hydrophobic core of the structure. The conserved small Ala and Gly residues allow a tighter side-by-side packing of α-helices. The modeling also shows that GALA-, GL-, and CC-LRRs can be packed well together and therefore, in contrast to the other LRR subfamilies, do not have mutually exclusive relationships with CC-LRRs in terms of inter-LRR packing. This conclusion is based on the following analysis. The conserved β-structural parts of the known crystal structures of LRR domains from different subfamilies and the model of GALA-LRRs were superimposed with the CC-LRR domain and the side-by- side packing of variable LRR fragments was analyzed (see for details). The analysis shows that only the α-helices of GALA-LRR and CC-LRR have an energetically favorable “knobs-into-holes” interface while the superposition of LRRs from the other analyzed subfamilies with the CC-LRR results in “knobs-into- knobs” packing with steric tensions and voids. For example, the RI-LRR and CC- LRR, PS-LRR and CC-LRR, and SDS22-LRR and CC-LRR interfaces have distances between Cα and (or) Cβ atoms of 2.1-2.7 Å that are 0.5–1.1 Å closer than normally allowed limits for such distances. This steric tension could be alleviated by a deformation of the LRR β-structure, but the distortion of the β-structural H-bonds would eventually also lead to the loss in the structure stability. The superposition of the typical LRR and CC-LRR domains does not lead to such close contacts, however, it results in an energetically unfavorable “knobs-into-knobs” packing with voids. Thus, our analysis suggests that some LRRs with different sequence motifs have an energetically favorable (“permissive”) packing, while simultaneous occurrence of the other ones in the same structure results in unfavorable (“mutually exclusive”) packing. The permissive packing of repeats with different consensus sequences may serve as a criterion for their membership in the same subfamily, at the same time as the mutually exclusive packing defines the boundaries between the LRR subfamilies. It is worth mentioning that GALA-LRRs are erroneously assigned to the RI-LRR subfamily in the annotation of protein databases on the NCBI Web site (<http://www.ncbi.nlm.nih.gov/entrez>). In order to dissipate any doubt, displays the apparent difference of GALA-LRR and RI-LRR through the backbone superposition. ## Inferring origin of *R. solanacearum* GALA proteins GALA F-box domains are functionally related to plant F-box domains. Although some bacteria have a proteasome-like compartmentalized protease system, they do not have an ubiquitin-dependent protein degradation system like in eukaryotes. Still several bacteria have in their genome typical eukaryotic E3 ubiquitin ligase-like proteins among which F-box proteins, like the GALA proteins from *R. solanacearum*. These bacterial F-box proteins also often contain eukaryote-like protein-protein interaction domains like LRR, ankyrin and WD40. We systematically searched all the sequenced eubacterial genomes available (353 genomes available through TIGR Comprehensive Microbial Resource, release 24.0) for the presence of the F-box domain (automatic search with Pfam Hiden Markov Model for F-box (PF00646) service available at TIGR CMR). We only found F-box domains present in one chlamydiae species out of 11 complete sequence available (*Candidatus Protochlamydia amoebophila* strain UWE25) and in 9 proteobacteria out of 184 complete sequences available (alphaproteobacteria: *Mesorhizobium loti, Agrobacterium tumefaciens*; betaproteobacterium: *Ralstonia solanacearum*, gammaproteobacteria: *Pseudomonas syringae, Sodalis glossinidius, Coxiella burnetii, legionella pneumophila, Xanthomonas campestris* and *X. axonopodis*). All these positive hits correspond indeed to the presence of a canonical F-box domain. The evidence for functional F-box domains is available for both *A. tumefaciens* and *R. solanacearum* F-box containing proteins. A few low scoring hits (in proteins from *Borrelia burgdorferi, B. garinii, Chlamydophila caviae, Rhizobium etli, Salmonella tiphimurium* and *Streptococcus pneumoniae*) were inspected and clearly ruled out as being F-box domains (by constraints in primary sequence, see). Within the proteobacteria phylum, 9 out of 184 completely sequenced bacteria clearly contain at least one F-box-containing predicted protein. Among the 175 negatively scoring bacteria, we believe we can rule out the presence of “remnants” of F-box domain, which could have been indicative of gene loss. Considering such sporadic presence of this F-box domain, the scenario of systematic gene loss appears very unlikely The F-box domain has its only described function in eukaryotic cells and is overrepresented in this kingdom (interpro F-box domain (IPR001810) hits: 735 in *A. thaliana,* 428 in *Caenorhabditis elegans,* 120 in humans, and only 46 hits among all bacteria sequence available, mostly in proteobacteria, see above). It is interesting to mention that all the bacteria containing F-box domains in their genome intimately interact with eukaryotes. For example, *P. amoebophila, S. glossinidius* and *M. loti* are symbionts of amoeba, insects and plants; *A. tumefaciens, R. Solanacearum, P. syringae, X. campestris* and *X. axonopodis* are plant pathogens and *C. burnetii* and *L. pneumophila* are human pathogens. Finally, for several of these F-box-containing bacterial proteins injection into their host cells (via specialised bacterial secretion systems) has been proven (*M. loti, A. tumefaciens, R. solanacearum*) – or predicted (*L. pneumophilae*). Among the seven *GALA* genes from the *R. solanacearum* genome (strain GMI1000, (<http://bioinfo.genopole-toulouse.prd.fr/annotation/iANT/bacteria/ralsto/>), *GALA1(*RSp0914) is located in an alternative codon usage region, *GALA2*(RSp0672) is flanked by a region duplicated elsewhere in the genome and *GALA3*(RSp0028) is flanked at either side by an alternative codon usage region. These genomic characteristic have been previously identified as potential signatures of LGT. Furthermore, considering the capacity of *R. solanacearum* to uptake DNA, it is natural to suggest a lateral gene transfer (LGT) from host plant DNA that gave rise to the F-box domain (and possibly the LRRs) of the GALA proteins. One way of testing such a hypothesis is through phylogenetic analysis of the protein origins to identify putative donor for a potential LGT. Apart from the F-box domain, the F-box-containing proteins are constituted of repeat domains (mostly LRRs). Thus phylogenetic inference on the whole protein for large numbers of divergent taxa appears highly problematic. Instead we focused the analysis on the F-box domain, using 50 aa on their own, as well as together with 150 aa from the downsteam F-box-adjacent region or a shorter region containing 2 or 3 LRRs. Fifteen large datasets were assembled to include homologues from the broadest possible species range. The datasets varied by the similarity thresholds used in the profile searches, criteria used to align the sequences (profile, penalties), amount of gaps, and included from 217 to 2853 taxa (and from 64 to 217 aa in the alignment). Trees were inferred with neighbour-joining and maximum likelihood methods. The resulted (approx.) 30 inferred phylogenies had low average branch supports, which demonstrated that phylogenetic inference for our data (with many short sequences of deep divergences and large percentage of gaps) is indeed problematic, even when considering the conserved F-box domain. The choice of the analysis model did not influence the inference much but considerably more variation in branching order was observed when analyzing different sets of sequences and alignments. In particular, varying the amount of gaps in the alignment had an impact. However, most phylogenies favoured the scenario where all *R. solanacearum* GALA genes clustered together and with *Arabidopsis thaliana* or with *Oryza sativa* as their closest basal lineages (for a representative example of an inferred Maximum Likelihood (ML) tree see that is a supplemental file phymlGALA.tre, which can be viewed with ATV-forester from [www.phylosoft.org/atv/](http://www.phylosoft.org/atv/)). Only in some rare cases we also observed that one or two GALA genes were grouped with a non-plant lineage (but with low clade supports). Overall, the underlying phylogenetic signal appears to be in favour of the postulated LGT from a plant lineage. At the same time, the limited accuracy of inference does not enable us to confidently suggest a putative donor. Our conclusion that the GALA-LRRs belong to the CC-LRR subfamily is consistent with the hypothesis of the lateral transfer. The updated CC-LRR subfamily has proteins from a very heterogeneous group of organisms including animals, plants, fungi, protista, and bacteria. By its wide taxonomic distribution it resembles the TpLRR subfamily that includes proteins found in all three domains of life. The broad distribution of TpLRR proteins also has been explained by LGT. Furthermore, the analysis of TpLRRs suggested that genes linked to pathogenicity can be shared between parasitic bacteria and parasitic eukaryotes. The results of our present analysis of CC-LRRs agree with this hypothesis. The updated CC- LRR subfamily includes many proteins from bacteria (*Gemmata sp., Parachlamydia sp., Legionella pneumophila, Rhodopirellula baltica, Ralstonia solanacearum),* protista *(Entamoeba histolytica, Leishmania, Trypanosoma, Dictyostelium)* and fungi (*Cryptococcus, Candida albicans, Candida glabrata, Neurospora*), among which many are parasitic organisms colonizing plants and animals. Despite the similarity of GALA-LRRs and CC-LRRs, they have some systematic differences and it needs to be explained. Usually, only about half of residue positions of LRRs remain conserved over a long evolutionary period. The conservation usually reflects the importance of these residues for the preservation of the structure. However, GALA-LRRs are nearly perfectly repeated and this suggests that they emerged relatively recently. On the other hand, our molecular modeling indicates that GALA- and CC-LRRs fold in very similar structures that can be compatible and well-packed together, if a CC-LRR is inserted between GALA-LRR or *visa versa*. This conclusion is supported by the fact that two plant F-box-LRR proteins have a couple of GALA-LRRs inserted in GL-LRR tandem arrays. Considering that CC-LRRs are much more abundant in plants than GALA-LRRs, and based on the above-mentioned facts, we propose the following sequence of evolutionary events that could “transform” the CC-LRR into GALA-LRR tandem arrays. First, the accumulation of point mutations may lead to the spontaneous occurrence of the first GALA-LRR and due to the structural complementarities between this new LRR and the CC-LRRs (see previous section) the occurrence of GALA-LRR does not significantly affect the overall structure and stability of the CC-LRR domain. Second, it is known that repetitive sequences can evolve more rapidly than non-repetitive ones,. This applies both to the repeat multiplication and to the repeat deletion. Therefore, once appeared, GALA-LRRs can multiply and CC-LRRs disappear. As a result the plant CC-LRR genes, being acquired by a bacterium, may shed their CC-LRRs replacing them by GALA-LRRs seeded in their original sequences. Currently it is not clear why *R. solanacearum* may prefer to generate arrays of GALA-LRRs instead of CC-LRRs. It has been shown that GALA are type III effectors required for virulence of *R. solanacearum* on three different plants, namely Arabidopsis, Tomato and *Medicago truncatula.*. Furthermore it is very likely that the GALA type III effectors participate in virulence through their action in plant cells as the adaptors in SCF-type E3-ubiquitine ligases. In the SCF- type E3 ubiquitin ligase complex F-box containing proteins interact through their LRR (or other protein-protein interaction domains) with the protein targets to be ubiquitinated. We propose that a possible conversion from an original F-box and CC-LRR protein to an F-box and GALA-LRR protein was dictated by functional constrains. It is possible that the new GALA-LRRs have better plant-protein target recognition and are more versatile adaptors suitable to detect protein targets from diverse host plants. ## Testing GALA-LRRs for positive selection in an attempt to establish their functional binding sites To gain insight into the function of the GALA proteins we examined whether adaptation could have acted on a proportion of protein residues during the evolution of GALA LRRs and identified positions of such sites, using likelihood ratio tests (LRTs) and the Bayesian prediction,. In the agreement with the evolutionary scenario suggested by us for GALA-LRRs, data sets containing aligned LRRs were used to analyze the strength of selective pressure across its residues since their common ancestor sequence was acquired by the bacterium. To some extent, the evolution of LRRs in one particular GALA protein may be likened to the evolution of the gene family members after a duplication event whereby paralogous genes originate from the common single ancestral sequence. Using this analogy we studied the process of the accumulation of substitutions at each site in a single LRR by comparing the codon substitutions at the homologous sites in other repeat sequences from the group of orthologous GALA proteins in four different *R. solanacearum* strains: GMI1000, RS1000, UW551 and MolK2 (C. Boucher and S. Genin, personal communication). The first two strains belong to the phylotype I and the others to the phylotype II, among the four phylotypes defined previously for this species complex. The full coding DNA alignment of all LRRs from all the available GALA proteins (\>400 sequences) was analyzed and none of the tests gave a significant evidence for positive selection, although parameter estimates hinted that this possibility could not be ruled out. This could mean that only for some GALAs the LRRs accumulated changes due to positive selection, as by averaging over all domains we loose power to detect positive selection affecting only certain GALA proteins. To test this we subdivided the aligned LRR sequences, so that only sequences from the orthologous GALAs were analyzed together. Parameter estimates from these alignments showed that for all GALA proteins 50–70% of the LRR positions are rather conserved while substitutions at remaining sites generally have neutral effect on the fitness of the protein. However for GALA2 both LRTs for positive selection were highly significant (with *P*-values \<0.01). Estimates suggested that 8% of sites evolved under positive selection. For GALA2 LRRs, the Bayesian approach detected positions 8 and 15 (numbering as in) with high probability (e.g., using model M8, the corresponding probabilities were 0.97 and 0.99, see and of Supporting Information for details). In accordance with the modeled GALA-LRR structure these residue positions are located in the α-helical region and exposed to the solution. In the LRR domains these positions are located on the convex surface of the horseshoe shaped structure. For GALA7 LRRs, only the LRT comparing M7 vs. M8 supported positive selection (*P*-value \<0.05), but the estimate of the ω ratio (describing selective pressure) was only slightly higher than 1 (ω≈1.15), indicating the lack of clear support for positive selection signal. Model M1a that does not allow positive selection described data equally as well as model M2a that allows positive selection. The Bayesian inference suggested that positions 4, 8, 11, 15 and 17 had a slightly elevated ratio of nonsynonymous to synonymous changes. Changes at these sites at the very least should be neutral to the fitness of the protein but may have a mild advantageous effect, possibly indicating a recent increase of adaptive pressure. The same can be concluded about position 4 in GALA1 and GALA3 LRRs and position 11 in GALA5 LRRs. If mapped on the structural model of the GALA-LRR, most of them (positions 8, 11, 15 and 17) are located on the external side of the α-helix and on the convex surface of the LRR solenoid. The side-chain in position 4 belongs to the loop connecting β-strand with the α-helix and also is exposed to the solvent. To see if signature of positive selection on GALA 2 and 7 is detectable on the level of the entire LRR domain of these proteins, we analyzed separately the groups of four GALA2 and GALA7 orthologous sequences from the different strains of *R. solanacearum*. Analysis of both GALA2 and GALA7 LRRs returned highly significant results for both tests (with *P*-values from 0.0014 to 0.025), providing the evidence of positive selection on both genes. The lack of the strong evidence for positive selection in GALA7 LRRs in the previous analysis suggests that positive selection may affect only certain (orthologous) repeats of GALA7 while the homologous sites in other repeats of this protein evolve neutrally. In this last analysis the number of sequences is too low for the Bayesian prediction to be accurate, and so the results of such inference are used only in an explorative manner, to see if the predicted positive selection sites correspond to any particular repeats and where such sites could be located. Mapping of the predicted sites of GALA2 onto the repeats of the LRR domain (with probability \>0.85 from BEB) shows that they are located in position 15 in four LRRs and in position 21 of one LRR and dispersed over the LRR domain mostly on the convex surface (see and of Supporting Information). Interestingly, the analysis of individual LRRs of GALA2 also pointed at position 15, that is the most represented in the analysis of the entire LRR domain. In the GALA7 LRR domain, these sites are also dispersed over the convex surface of the protein and are found in exactly the same positions as predicted in individual LRRs of GALA7 (positions 4, 11, 15 and 17). Thus, it is encouraging to observe that most inferred positions throughout the LRR domain of orthologous GALAs coincide with those inferred in individual LRR repeats analysis on groups of GALA orthologues. It is important to mention that all positions, that are inferred to be under positive selection, are found on the surface of the structural model of GALA LRR domain. This grants additional support for the GALA-LRR structure prediction described above. Furthermore, our analysis suggests that the convex surface of the horse-shoe shaped GALA-LRR domain is more prone to positive selection than its concave one. It is tempting to propose that the selective pressure leading to an increase of variability on such residues could be the site of protein binding, relevant to the adaptor function of the F-box GALA proteins in the SCF-type E3 ubiquitin ligase. This study provides a strong background for further functional studies. ## Conclusions The GALA-LRR example shows that the differences in LRR consensus motifs can be not only mutually exclusive in terms of inter-LRR packing as it was observed in LRRs from different subfamilies, but also permissive as we found in the case of GALA-LRRs and CC-LRRs. The permissive packing may serve as a criterion for the affiliation of LRRs having different consensus sequences with the same LRR subfamily. Therefore, one may expect to find a subfamily of evolutionary related proteins that share similar functions, cellular location, globular domains and, at the same time, having quite different repetitive consensus patterns. The relationships between GALA-, GL- and the other CC-LRRs suggest that structural constraints, namely, permissive packing of repeats may control the limits of the LRR variability within a subfamily. This result provides new insight into the fascinating interplay between the structural constraints and unusual evolutionary dynamics of LRRs and can be used to classify other newly identified LRR domains. The *R. solanacearum* GALA proteins are bacterial F-box proteins containing a new kind of LRR, which can be found in other bacteria, plants and unicellular eukaryotes. These GALA-LRRs and related GL-LRRs are part of the CC-LRR subfamily, which is generally associated with an F-box domain. The presence of this F-box domain in GALA proteins is indicative of the probable ancient lateral transfer from eukaryotic (possibly plant) genes into a *R. solanacearum* ancestral recipient strain. We further looked into the GALA-LRR in all the GALA sequences available and found that for some GALA proteins there is a strong signature of positively selected residues and only on the convex side of the GALA-LRR structure. This suggests that the GALA proteins, probable E3-ubiquitin ligase adaptors, necessary for the virulence of *R. solanacearum* on its host plants, could bind their potential protein ligand on the convex side of the LRR domains, similarly to the *A. thaliana* TIR1 F-box type E3 ubiquitin ligase. As we have an experimental system that enables us to test for the functionality of GALA proteins in virulence, this current study provides a strong theoretical background for testing the relevance of specific GALA-LRR residues to pathogenesis. # Methods ## Sequence profile search The generalized sequence profile method and the *pftools* package were used. Since a single LRR would be unlikely to form a stable structure on its own, we limited the search to proteins containing at least three tandem repeats, thus increasing the selectivity of the search. Several profiles corresponding to different types of LRRs were constructed (GALA-LRR, GL-LRR and updated CC-LRR). The probability that a match is a product of chance alone was calculated by analyzing the score distribution obtained from a profile search against a regionally randomized version of the protein database, assuming an extreme value distribution. All database searches were performed with a nonredundant data set constructed from 2006 releases of non-redundant protein sequence database including GenPept, Swissprot, PIR, PDF, PDB and NCBI RefSeq and available on the NCBI FTP site ([ftp://ftp.ncbi.nih.gov/blast/db](http://ftp://ftp.ncbi.nih.gov/blast/db)). The sequence profiles and results of the search can be viewed on a dedicated web- page <http://bioinfo.montp.cnrs.fr/profiles> (Tools/Profiles/Show Profiles). ## Molecular Modelling The initial template for GALA-LRRs was taken from a 24-residue LRR (residues 2211 to 2235) of the known crystal structure of human Skp2 protein, (see Results and Discussion) using the Insight II program. The amino acid sequence of the template was edited in accordance with the GALA sequences using the homology modeling option of Insight II program. The structure was further refined by the energy minimization procedure based on the steepest descent algorithm implemented in the Discovery subroutine of Insight II, and tethering heavy backbone atoms to their starting conformations with force constant K = 100. The 300 steps of minimization led to a maximum RMS derivative of 0.4 kcal/(mol\*Å). The next stage of minimization was 500 steps of the conjugate gradients algorithm, tethering the backbone atoms with lower force (K = 50), and then 300 steps with K = 25. The tethering was accompanied by setting the distance constraints at K = 50, in order to improve the geometry of H-bonds. To allay the concern that these constraints generated significant tensions in the minimized structure, the last calculation was performed without any restrictions, to an RMS derivative of 0.3 kcal/(mol\*Å). The CVFF force field and the distance dependent dielectric constant were used for the energy calculations. The program PROCHECK was used to check the quality of the modeled structure. In accordance with the PROCHECK results all residues of the LRR domain of the GALA4 model have backbone conformations from allowed regions of the Ramachandran plot; and G-factors of the polypeptide stereochemistry (a log-odds score based on the observed distribution of the covalent geometry) equal to −0.15. The overall average G-factors for the model is −0.49, values that would be expected for good-quality model. To examine the side-by-side packing of LRRs from different subfamilies the following procedure was used. First, fragments of the LRR domains corresponding to different LRR subfamilies were extracted from the known crystal structures (CC-LRR, pdb code 2AST; Typical LRR, pdb code 2O6Q; Bacterial LRR, pdb code 1G9U; SDS22-like LRR, pdb code 1D0B; RI-LRR, pdb code 2BNH; PS-LRR, pdb code 1OGQ). Second, each of these structures and the structural model of GALA-LRRs were superimposed with CC-LRR domain. For the superposition, the conserved β-structural parts of the LRRs were used. Two adjacent β-strands of CC-LRR and the analyzed structures were superimposed and the side-by-side packing of the variable LRR fragments was analyzed. ## Tests for positive selection The selective pressure at the protein level was measured by the ratio of nonsynonymous to synonymous rates *ω = d*_N/*d*_S, with *ω*\<1, = 1, or \>1 indicating conserved, neutral or adaptive evolution respectively. Selective pressure was evaluated using the probabilistic Markov models of codon substitution. Such models describe the substitution process based on a multiple alignment tree. The transitions from one codon state to another are described by the transition probability matrix over time *t* as *P*(*t*) = exp(*Qt*). The generator matrix *Q* = {*q_ij*} defines the instantaneous substitution rates at site *s* from codon *i* to codon *j* : Here *π_j* is the frequency of codon *j*, parameter *κ* is the transition/transversion ratio, and *ω*^(*s*) is the *ω* ratio for site *s*. The codon substitution process is assumed independent among sites, and model parameters are estimated by maximizing the log-likelihood function of sequence data *X* = {*x*_s} given a phylogeny with branch lengths τ and a model Μ: The models used in the analysis differed by statistical distributions of the *ω* ratio used to describe the variation of selective pressure along a sequence. Likelihood ratio test (LRT) for positive selection compares maximum log- likelihoods of two nested models, one of which allows sites under positive selection while another does not. To test that a model allowing positive selection describes data significantly better, twice the log-likelihood difference is compared to the χ²-distribution with degrees of freedom equal to the difference in the number of free parameters between the two models. We performed two LRTs for positive selection, comparing models M2a and M8 that allow sites with *ω*\>1 (alternative hypotheses) with simpler models M1a and M7 respectively that do not allow sites with *ω*\>1 (null hypotheses). Model M1a (nearly-neutral) assumes two site classes in proportions *p*₀ and *p*₁ = 1–*p*₀: one with *ω*₀ ratio estimated between 0 and 1, and the other with *ω*₁ fixed at 1. The alternative model M2a (positive selection) extends the null model M1a by adding a proportion *p*₂ of positively selected sites with *ω*₂\>1, estimated from data. The second LRT uses the null model M7 (beta) that assumes the *ω* ratio is drawn from a beta distribution defined between 0 and 1. The alternative model M8 has an extra class of sites under positive selection with *ω*\>1. We also considered two other codon models: the most simple one-ratio model M0, where *ω* is assumed to be constant over all sites in the sequence, and the discrete model M3 that allows three discrete classes of sites with ratios *ω*₀, *ω*₁, and *ω*₂ occurring in proportions *p*₀, *p*₁ and *p*₂ = 1−*p*₀−*p*₁. Models M0 and M3 are also nested, and can be used to perform the LRT for heterogeneity of selective pressure along the sequence. This test is often significant, as most coding data has significantly heterogeneous selective pressures acting on different sites of the sequence, according to their functional importance and the role in the protein folding and stability. In comparison with models M8 and M2a, model M3 better combines the algorithmic simplicity with sufficient complexity necessary to reflect heterogeneity of selection pressure in nature. This model is often used to evaluate the underlying distribution of the selective pressure across sites in a sequence. Inconsistencies in estimates under different models may be a sign that the algorithm has not converged to a global optimum. To insure proper convergence, we performed repeated runs for each model (with different starting values) and confirmed that the distribution of selective pressure described by estimates under models M2a and M8 were compatible with the distribution estimated under M3 for all datasets analyzed. Where a LRT for positive selective pressure was significant, we used the Bayesian inference to calculate posterior probabilities that a site belongs to a particular site class. The posterior distribution of the parameter of interest (in our case ω) is proportional to the product of its assumed prior distribution and the likelihood of the observed data given this prior. In this study we used the Bayesian Empirical Bayesian approach, where the posteriors are obtained by integrating over the prior distribution of selection-related parameters, while setting other model parameters to their maximum likelihood estimates. Sites with high posteriors probabilities (\>0.95) of coming from a class with ω\>1 are likely to have evolved under positive selection. Anisimova *et al*. showed that the major factor affecting the accuracy Bayesian site prediction is the diversity of the data set and the number of sequences used. The sequence length was shown to have little effect on the accuracy of this method. We therefore believe that our LRRs analysis should have good accuracy as short sequence length may be compensated by the diversity and large numbers of sequences used in this study. Several site-by-site studies support this notion. While extreme levels of sequence divergence do not seem to compromise the accuracy of the LRTs, the Bayesian prediction becomes unreliable. In this study the divergence levels ranged from 0.21 to 0.34 nucleotide changes per codon per branch (calculated as the tree length divided by the number of branches in the unrooted tree; see). This corresponds to the optimal divergence levels and so insures good accuracy of Bayesian prediction results reported here. All ML phylogenies were inferred using PHYML program. The phylogenies used to perform the selection tests was inferred using coding sequences under HKY+Γ, and all phylogenies used for testing LGT hypothesis were inferred under WAG model. Phylogenies for testing LGT hypothesis were also inferred using fast neighbor- joining method implemented in BIONJ,. Branch supports were inferred using approximate LRT and 100 bootstrap replicates where computation permitted. # Supporting Information [^1]: Analyzed the data: NP MA. Wrote the paper: NP AK MA. Other: Bioinformatics analysis of amino acid sequences, molecular modeling: AK. [^2]: The authors have declared that no competing interests exist.

# Introduction Over 7.6 million individuals are receiving antiretroviral therapy (ART) in sub- Saharan Africa (SSA) as a result of unprecedented global health initiatives to scale-up ART delivery since 2003. Uganda was one of the first countries to roll-out public-sector HIV and ART programs on a large scale, commencing in 2004. These efforts were mainly supported and funded by the World Health Organization (WHO), Global Fund to Fight AIDS, Tuberculosis and Malaria (GFATM), and the United States President’s Emergency Plan for AIDS Relief (PEPFAR). After an initial roll out in national and regional referral hospitals, access to antiretroviral treatment was gradually expanded to peripheral facilities with more than 400,000 HIV positive individuals receiving ART by the end of 2012. Short and medium term studies from sub-Saharan Africa show that despite high early mortality rates due to late presentation, substantial loss to program, and high rates of drug substitution due to toxicity, antiretroviral treatment programs have achieved outcomes comparable to those in developed settings especially in terms of CD4 count gains and viral suppression. Furthermore adherence levels in sub-Saharan Africa, once a big concern before the scale up of ART, have been reported to be similar or higher than those achieved by patients in resource-rich settings. However, these studies were done for a short duration and were unable to account for long term outcomes of patients as they stayed longer on treatment and as their well -being improved. The Infectious Diseases Institute at the Makerere University College of Health Sciences in Kampala was one of the first HIV care centers to receive free antiretroviral drugs from GFATM and PEPFAR, therefore some of the active patients attending the adult clinic have been followed up for more than ten years. We have previously described the short term (one-year) and medium term (four-years) outcomes from a well characterized ART cohort within a large urban HIV treatment center at the National Referral Hospital in Uganda. The objective of this analysis was further to evaluate the long term outcomes in this cohort of patients initiated on ART at the beginning of free treatment roll out in Uganda, and followed up for a decade. Specifically we assessed four outcome indicators: retention in care, virologic suppression and immune-recovery, ART regimen changes. # Methods ## Ethical statement The study was reviewed and approved by the Makerere University Faculty of Medicine Research and Ethics Committee (Approval number: 016–2004) and the Uganda National Council for Science and Technology (Approval number: MV 853). All patients provided written informed consent. ## Study site and population The Infectious Diseases Institute (IDI) is a large urban HIV center of excellence located at Mulago National Referral Hospital in Kampala, Uganda. IDI began providing free antiretroviral treatment through the GFATM and PEPFAR since April 2004. At present more than 30,000 patients have been registered and over 15,000 have ever been started on ART. ART efficacy is monitored through bi- annual CD4 counts and viral load measurements were performed ad hoc in patients with immunologic failure according to the WHO guidelines; ART toxicity was not routinely monitored but safety laboratory test were performed upon clinician’s discretion. In this context a cohort of consecutive adult patients starting ART was initiated and 559 patients were enrolled in a research cohort between April 2004 and April 2005. The main objective of this study, hence forth referred to as the IDI Research Cohort, was to assemble and to characterize a cohort of patients on ART in order to address pertinent epidemiologic, clinical and psychosocial research questions relevant to the national and sub-Saharan African ART roll- out. ## Study procedures At enrollment and every three months patients were evaluated by a dedicated study clinician and counselor. ART was started according to WHO and Uganda Ministry of Health guidelines in patients with an AIDS defining illness or a CD4 count \<200 cell/μ with a combination of stavudine (weight-adjusted), lamivudine and nevirapine or zidovudine, lamivudine and efavirenz. Detailed study procedures for this is cohort are described elsewhere, but in summary: at enrollment and during the follow up study visits information was collected on patients’ demographic, previous and current opportunistic infections as well as non-HIV related clinical events, WHO staging, vital signs and ART regimen, and physical examination was performed. During the follow up visits adherence was assessed using the visual analog scale and ART toxicity and reasons for ART substitution were recorded. Treatment regimen was changed with a single drug in patients who experienced a grade 3 or 4 toxicity according to the AIDS Clinical Trials Group (ACTG) classification. Efavirenz was substituted with nevirapine (up to 2012) in women found to or those planning to become pregnant, while efavirenz was substituted with nevirapine among patients diagnosed with tuberculosis. Patients with 2 consecutive viral loads greater than 1000 copies/ml were eligible for switch to protease inhibitors (PIs) based regimens if second line drugs were available. Patients with a first viral load greater than 400 copies/ml were offered extra adherence counselling. However, in our context, where second line drugs are not always readily available, it is not uncommon that patients with detectable viral load but low levels of viremia are maintained on failing first line regimens, with additional adherence counseling. Laboratory investigations were performed every 6 months and these included complete blood cell count, liver and renal function tests, CD4 count by FACSCount (Becton Dickinson, San Jose, CA and, more recently, by FACSCalibur, Becton Dickinson), HIV viral load by Amplicor HIV-1 Monitor PCR Test version 1.5 and more recently, COBAS Ampliprep/COBAS Taqman HIV-1 Test Ver.2.0 (Roche Diagnostics, Indianapolis, IN) and storage of 5 ml of plasma at -80 C for future testing. Laboratory testing was performed at the Makerere University–Johns Hopkins University Core Laboratory, which is certified by the College of American Pathologists. Tracking and complete ascertainment of patients who did not attend their study appointment was carried out in real time by phone call and/or home visit. In case of death the cause was determined from a combination of source documents and a structured interview with the patient’s next of kin (“verbal autopsy”). Generally ascertainment of outcomes of nearly all study subjects was achieved and lost to follow up was negligible. ## Data collection and data quality Data was collected on a study specific designed questionnaire by the study counselor and the study clinician. The data was entered into a database designed using Oracle up to year 2011 and in ICEA (Integrated Clinic Enterprise Application), an application developed by a team of software developers based at the IDI, which ensures good quality of collected data. The entered data was subsequently validated by a senior data entrant before being approved. Laboratory results were obtained directly from the laboratory database. The data was also regularly checked for quality, completeness and discrepancies by a data quality assurance team and a data manager. ## Definitions A patient was considered *retained* in the cohort if he/she was alive and still enrolled in the cohort. A patient was defined as *lost to follow up* if two consecutive study visits were missed, i.e. the patient did not attend the clinic for more than 6 months. *Viral suppression* was defined as the attainment of HIV viral load \<400 copies/ml after starting ART in patients with at least 6 months of follow up, and in those with documented viral suppression, *viral failure* was defined as 2 consecutive viral load measurement \>400 copies/ml or 1 viral load\> 5,000 copies/ml if a subsequent viral load was not available. *Treatment failure* was defined as the lack of attainment of viral suppression or the occurrence of viral failure after initial suppression. *Treatment change* was defined as the occurrence of drug substitution or regimen switch. *Drug substitution* was defined as the modification of at least one drug within first line regimen, while *regimen switch* was defined as a change from a first to a second line regimen. ## Statistical analysis Baseline characteristics were described using medians and proportions and stratified by gender. Gender categories were compared using chi-square test for categorical variables and Wilcoxon rank-sum test for not normally distributed variables. To describe retention in the cohort we reported the number and proportion of patients who were no longer enrolled in the cohort for any reason including loss to follow up, death, transfer out, and withdrawal of consent at month 6, 12, and year 2, 5 and 10. We reported mortality over a 10-year time period by cause: HIV related, other medical, and accidental and we calculated the probability of death using Kaplan Meier estimates with right censoring at time of death, stratified by gender and CD4 count μL at ART start (\<50 cells/μL, 50–99 cells/μL, and ≥100 cells). We described the increase in CD4 count from baseline to all subsequent 6-monthly CD4 count readings and the overall increase from baseline to the last visit in the cohort. We also reported the proportion of patients at each time point with a CD4 count \>400 cells/μL, which is the lower normal limit of CD4 count in Ugandan adults. We reported the proportion of patients who attained viral suppression and of those who experienced viral failure. We also calculated the cumulative probability of treatment failure stratified by gender and CD4 count at ART start at different thresholds. We described the ART regimens prescribed during the 10 years categorized as stavudine, zidovudine tenofovir based for first line regimens, and PI based. We described all causes of drug substitution categorized as toxicity, treatment failure, new episode of tuberculosis, pregnancy/intended pregnancy and others, and we calculated the yearly and cumulative probability of first drug substitution for any reason while of first line. All differences in survival probabilities were compared using the log rank test. The analysis was performed using STATA® version 12.2, Texas USA. # Results ## Baseline characteristics Between April 2004 and April 2005, 559 consecutive patients starting first line antiretroviral treatment were enrolled in the IDI Research Cohort. The baseline characteristics and the characteristics by gender are summarized and compared in. The majority (69%) of the patients was female, in WHO clinical stages 3 and 4 (89%), with a median CD4 count of 98 cells/μL (IQR: 21–163). The characteristics of the patients in the cohort were similar to the characteristics of other patients started on ART at IDI but not enrolled in the study. Overall male patients were older, and presented with more advanced WHO stage, had lower body mass index (BMI) and median CD4 count; a higher proportion of male was started on efavirenz-based regimen due to concurrent tuberculosis (TB) treatment or being under investigation to rule out TB. ## Survival and retention into care The median follow up time was 110 months (IQR: 25–111). shows the follow up status of the patients during the follow up. Overall 127/559 patients (22.7%) died. of which 63% died during the first year on ART, and 31 (5.5%) were lost to follow up. The probability of being retained in the study was 0.81 (95%CI: 0.77–0.84) by year 1, 0.70 (95%CI: 0.66–0.73) by year 5, and 0.64 (95%CI: 0.60–0.68) by 10 years on ART. The probability of death was 0.15 (95%CI: 0.12–0.18) by year 1, 0.21 (95%CI: 0.18–0.25) by year 5, and 0.24 (95%CI: 0.21–0.28) by 10 years on ART. The majority (86, 67.7%) of the deaths were AIDS events, followed by other medical condition (24, 18,9%) and 1 (0.8%) accidental death; the cause of death could not be determined for 16 (12.6%) patients due to lack of relative contact or non-conclusive “verbal autopsy”. Patients with a CD4 count \<50 cells/μL at ART start had a significant higher overall probability of dying, 0.31 (95%CI: 0.25–0.39) as compared to patients with a CD4 count ≥100 cells/μL, 0.18: (95%CI: 0.14–0.23) (P value 0.001). We did not find an independent association between survival and gender (P value 0.7). ## Immune-recovery The median CD4 count increased from 98 to 589 cell/μL (IQR: 450–739 cell/μL) over the 10 year period with a median increase of 357 cells/μL (IQR: 128–600 cells/μL). The proportion of patients with a CD4 count \>400 cells/μL rose from 0.9% at baseline to 83.8% after ten years on ART, with majority of patients (50.9%) being above this threshold by year 5 on ART. The cumulative probability of attaining a CD4 count \>400 cells/μL was 0.17 (95%CI: 0.14–0.21) at year 1, 0.72 (95%CI: 0.68–0.77) at year 5, and 0.92 (95%CI: 0.89–0.94) at year 10 on ART. ## Treatment failure Four hundred seventy four patients were followed up for at least 6 months, and of these 35 (7.4%) never attained viral suppression; the cumulative probability of attaining viral suppression was 0.94 (95% CI: 0.92–0.96). Among the 439 (92.6%) patients who attained viral suppression 139 (31.7%) experienced subsequent viral failure with a cumulative probability of 0.34 (95%CI: 0.30–0.39). Five of those who never attained viral suppression, and 14 and those who experienced viral failure subsequently died. Overall the median time to treatment failure was 30 months (IQR 11–66) and the cumulative probability of treatment failure over 10 years was 0.38 (0.34–0.43).The probability of treatment failure was similar in males as compared to females (0.37, CI 0.32–0.43 versus 0.45, CI 0.36–0.53), P = 0.08) and we did not find any difference in the probability of treatment failure across the CD4 count strata at ART start (P = 0.52) ## Treatment regimens and treatment change shows the proportion of treatment regimens received by the patients in the cohort during the 10 years of follow up. While the majority (74.4%) of patients were started on stavudine based regimens, after 7 years of follow up none of the patients was receiving stavudine; conversely the proportion of patients receiving zidovudine increased from 15.6% to 69.4% and of those receiving tenofovir from 0% to 13.5%; in the 10^th year of follow up 17.1% of the retained patients were switched to second line PI-based regimens. Three hundred and two patients (54%) had at least one drug substitution while on first line for a total of 378 drug substitutions. Two hindered forty four (64.6%) had at least one drug substitution, 46 (24.3%) two substitution and the remaining ones up to 5 substitutions. The most common reason (141, 46.7%) for the first drug substitution was the 2008 Ministry of Health recommendation to replace stavudine with a less toxic nucleoside reverse transcriptase inhibitor in patients still receiving it, followed by ART toxicity (115, 32.9%), occurrence of a new episode of tuberculosis (24, 8%), pregnancy (14, 4.6%), and other reasons (8. 2.6%). The median time to the first drug substitution was 40 months (20–44); the cumulative probability of drug substitution was 0.77 (CI: 0.72–0.81) with a higher probability of substitution in females (0.81, CI: 0.76–0.85) as compared to males (0.67, CI 0.58–0.76, P = \<0.001). Overall 66 (11.9%) of the patients were switched to a second line PI-based regimen due to confirmed treatment failure. # Discussion To our knowledge this is the first comprehensive analysis of 10-year outcomes in a well characterized African ART cohort. We observed that despite a high early mortality, the overall retention and viral suppression rates were high with remarkable immunologic recovery over the 10 year period. These findings are encouraging given the context of an initially highly immune-deficient population. Unfortunately this success was overshadowed by high rates of early mortality due to late presentation and high rate of toxicity resulting in drug substitution. Overall we observed that our cohort consisted of young, predominantly female and severely immunosuppressed individuals, representative of African ART patients at the start of the ART roll-out. The population in this cohort started ART at very advanced stage of disease, with 89% of the patients being in 3 or 4 WHO stages, and with a quarter of them having a CD4 count \<21 cells/μL. Male patients were generally older and had a lower BMI probably due to a later presentation into HIV care. The overall rate of loss to follow up was very low, but this may not reflect the attrition rates observed in other settings, or even for patients routinely followed up in our adult clinic, but not enrolled in a research study. The patients enrolled in this cohort had initial and ongoing intensive counseling, they were reminded the day before the study visit about their appointment by phone call, and were systematically traced if the appointment was missed, or visited at home if the attempt of contacting them was unsuccessful. Moreover patients in this cohort had 6 monthly viral load measurements with review of these results with the patient at subsequent visits. The reported level of retention may not be achieved in routine busy settings where resources are minimal and health care workers are overwhelmed, but we do demonstrate that with a dedicated clinical team and routine viral load monitoring high retention into care can be attained. The findings are at present even more encouraging with the current efforts to roll out routine viral load monitoring in resource limited settings, which in the Uganda case have commenced at the larger HIV treatment centers. Despite a low baseline CD4 count, we report remarkable immune recovery (median CD4 count after 10 years: 589 cells/mL) with CD4 count levels continuously increasing throughout the 10 years follow up period up and with 83% of the patients attaining a CD4 count above the lower normal limit. The reduction in incidence of OIs including tuberculosis (data not shown) and mortality is to a large extent attributable to this immune recovery. In this analysis we included all patients regardless of viral suppression and ART regimen in order to demonstrate the general effect of ART on population immune recovery over time. We also observed a high initial viral response comparable or higher than observed in on treatment analysis from similar African and non-African settings with only 7% of the patients not achieving viral suppression; it is likely that in this cohort acquired HIV resistance was extremely low due to limited number of patients on ART in Uganda before 2004, and therefore the lack of achievement of suppression should be mostly attributed to early suboptimal adherence. Thirty two percent of the patients with viral suppression experienced subsequent viral failure; our results may not be comparable with others from sub-Saharan Africa due to the longer follow up and variation of definition of viral failure. Similarly to what was observed in a cohort from similar settings the probability of treatment failure rose more sharply during the first 2 years on ART as compared to subsequent follow up period. Viral suppression is associated with reduction in OI incidence, as well as non-communicable diseases-related to the markedly reduced immune-activation. This observation of high viral suppression is additionally important from the perspective of HIV prevention and underscores the gains of treatment as prevention (TasP). Similarly to other reports of patients started on ART at the beginning of the scale up, a high proportion of patients died. Seventy five percent of the deaths occurred within two years of follow up and sixty three percent during the first year; the majority of the deaths were HIV related and associated with low CD4 count at ART start suggesting that those patients died as result of late presentation with underlying opportunistic infections, rather than inadequate quality of care. More than half of the patients required at least one drug substitution while on first line; this was mostly due to the inclusion of stavudine in the majority of the ART regimens prescribed at treatment start. Most of the substitutions due to toxicity were attributable to stavudine toxicity while the most common reason for drug substitution was the Ministry of Health recommendation to phase out stavudine from the ART national program; the majority of the patients subsequently received zidovudine since tenofovir is more expensive and was reserved for patients intolerant to zidovudine. Similarly to other studies women had a higher probability of having drug substitution, mostly due to stavudine toxicity. In summary, although the success of ART was jeopardized in the early years by early mortality and ART toxicity, the outcomes from this cohort are encouraging, particularly the remarkable and incremental immune-recovery and a satisfactory rate of virologic suppression. Patients presenting with less advanced disease and those started on ART later during the scale up when stavudine was no longer part of the preferred regimens are likely to experience even more favorable outcomes. The authors would like to acknowledge the study staff and participants. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: BC A. Kiragga. Performed the experiments: BW JW. Analyzed the data: A. Kiragga FM JM. Contributed reagents/materials/analysis tools: A. Kambugu MRK. Wrote the paper: BC A. Kiragga A. Kambugu. Reviewed the manuscript prior to submission: BW JW MRK JS.

# Introduction Establishing the position of Cetacea (whales, dolphins and porpoises) within Mammalia has long been a focus of mammalian systematists. The transition from a primitively quadrupedal terrestrial ancestor to a convergently ‘fish-like’ modern mammal species involved changes in numerous character systems. Almost all anatomical systems of living cetaceans are highly modified for an aquatic lifestyle, with dramatic changes seen in areas such as the ear region, skin, limbs, and cranium relative to terrestrial mammals. The study of phylogenetic data that fossilizes (primarily skeletal and dental morphology) has been particularly important because it is by studying extinct species that we can reconstruct the order of character acquisition that led to the origin of Cetacea (see review of studies in). Continued discovery of fossils that capture transitional stages in cetacean evolution provides critical new data on how the stem lineage to Cetacea transformed. By incorporating new fossils into increasingly large total evidence (character congruence) analyses, we are beginning to develop a firm understanding of the evolutionary history of this clade and can start testing explicit hypotheses concerning character transformation. For example, ‘Did whales develop ear bones for underwater hearing while still able to easily move on land?,’ or ‘What came first in the whale lineage - dietary change to aquatic carnivory or committed life in water?’ None of these hypotheses can be assessed without a robust test of the sister taxa to the clade Cetacea. Subsequent to the last large scale total evidence analyses of the position of cetaceans among mammals new specimens of the extinct †raoellid artiodactylan, †*Indohyus*, were described that potentially offer critical information about the phylogenetic position of Cetacea. Among the new specimens is a skull that preserves quadritubercular dentition (4 major cusps, found in herbivores and omnivores) and a pachyostotic auditory bulla (found in mammals derived for underwater hearing). These features had not previously been recorded in the same individual. Thewissen et al. argued on the basis of these new data that †*Indohyus* was the sister taxon of living and extinct whales. Many aspects of their published phylogenetic tree were, however, highly incongruent with other recent studies, and their phylogenetic results were subsequently challenged. We reevaluate the position of this significant fossil in the context of the largest total evidence analysis of Artiodactyla and relatives to date. We generate 49 new DNA sequences from five nuclear loci and expand our taxonomic sample to include living and extinct Carnivoramorpha (cats, dogs and fossil relatives) and †Creodonta (archaic extinct carnivorous mammals). Carnivoramorpha and †Creodonta may be critical for determining the position of the wholly extinct clade †Mesonychia, which has played a pivotal role in our understanding of the pattern of character evolution in Cetacea (see discussion). In particular, we are interested to know how the carnivorous (or hypothesized to be carnivorous) taxa (Carnivoramorpha, †Creodonta, †Mesonychia) are related to Cetacea, a highly-specialized carnivorous/piscivorous lineage that is nested within a clade composed primarily of herbivores (Artiodactyla). Inclusion of a variety of taxa such as these, that have dental similarities to early whales, could directly influence tree topology and interpretations of dental evolution on the stem lineage of Cetacea. Because the association of diagnostic †raoellid cranial fossils with postcranial remains has not been convincingly established (noted), we also examined how exclusion of postcranial information affected the phylogenetic position of †*Indohyus*. To facilitate discussion of key transitional fossils on the stem lineages of living clades, we revise the higher-level taxonomy of Artiodactyla. Although the focus of this paper is to examine the phylogenetic relationships of †Mesonychia and the †raoellid †*Indohyus*, this study has implications for Ferae (which we recognize as including only Carnivora plus †Creodonta, following, ; we do not follow the more inclusive Ferae of). The monophyly of both Ferae and †Creodonta has been questioned. Despite the long-standing grouping of Ferae , this taxon has never been the subject of a rigorous phylogenetic test in a cladistic framework. The modern concept of the †Creodonta, a wholly extinct carnivorous group, includes two sub-clades: †Hyaenodontidae and †Oxyaenidae, but the relationships of these taxa needs further testing. The most recent phylogenetic studies of †Creodonta, have concentrated on subclades within the group, and did not address the relationships of †Creodonta in a broader framework. Other analyses that included †creodonts have utilized the taxa as outgroups, and did not specifically test the relationship between the two †creodont families. The expansion in taxon sampling of this study not only benefits our understanding of relationships among Cetacea, other artiodactylans, and †Mesonychia, but also those of Ferae and its component clades. ## Revised Taxonomy for Artiodactyla It has become increasingly important to have phylogenetic names for the ingroup in question, Artiodactyla, to discuss character evolution unambiguously. Other groups have greatly benefited from a revision of taxonomic nomenclature to reflect phylogeny. Previous taxonomies for Artiodactyla have not been based on robust phylogenetic results or have ignored extinct diversity. This has led to confusion in discussions of evolutionary relationships in the clade. This disorder can be rectified, in part, by a new taxonomy that utilizes “crown clades” and “total clades”. A crown grouping is based upon a cluster of extant species, and a total clade is the crown group plus the paraphyletic series of fossils at its base. The core of our new system is a set of traditional, commonly-used taxonomic names that in the past have been applied variously as stem, crown, node, or apomorphy based groups. Furthermore, our taxonomy incorporates information from recent combined phylogenetic studies that have shown a highly consistent set of hierarchical relationships among major clades of Artiodactyla (reviewed in). All higher-level artiodactylan names proposed are being submitted to the Companion Volume to the Phylocode. By formally naming these clades, we hope to assist in providing a unified system of nomenclature for Artiodactyla, consistent with those applied to other mammalian orders. In the new taxonomy, we utilize the name Artiodactyla as a crown clade, the monophyletic group that includes the last common ancestor of cattle, antelope, deer, giraffes, musk deer, chevrotains, hippos, pigs, peccaries, and camels, and all of its descendants. Many analyses have supported the nesting of Cetacea several nodes within Artiodactyla. This prompted Montgelard et al. to rename the combined group ‘Cetartiodactyla.’ Despite our prior use of the term ‘Cetartiodactyla’, the topological change of placing Cetacea within Artiodactyla was never grounds to retire the name, Artiodactyla, according to rules of phylogenetic nomenclature. ‘Cetartiodactyla’ has gained some traction in the literature, especially among molecular workers, but here we formally retain the name Artiodactyla following the logic entailed in the Phylocode. All groups that we name as crown clades have been robustly supported by combined phylogenetic analyses of molecules and morphology from living and extinct taxa, and this study). These include Cetacea, Hippopotamidae, Cetancodonta, Ruminantia, Cetruminantia, Suina, Camelidae, and Artiodactyla, which are found in all minimum length trees (even if the strict consensus is sometime unresolved due to unstable fossils). As suggested by de Queiroz, we have applied widely-used artiodactylan names to crown clades. We then added a standard suffix (“-morpha”) to the crown names to identify corresponding total clades. This allows those who are more familiar with extant diversity, presumably the majority of scientists and laypersons, to link extinct diversity broadly to better known extant species. For example, as defined here, Cetacea includes all descendants of the last common ancestor of *Tursiops truncatus* and *Balaena mysticetus*. Most biologists are familiar with whales and dolphins, and these species have been referred to as ‘cetaceans’ for a very long time. Assigning the traditional name ‘Cetacea’ to this crown clade informs the user that extinct crown cetaceans are close relatives of whales and dolphins and likely have many of the synapomorphies shared by all extant cetaceans (e.g., obligately-aquatic lifestyle, reduced hindlimbs, tail flukes, flipper-shaped forelimbs, pachyostotic ear bones, etc.). In our new taxonomy, Cetaceamorpha is the total clade defined as Cetacea plus all extinct taxa more closely related to extant cetaceans than to any other living species. This replaces the use of ‘Cetacea’ as a stem clade (sensu). Thus, in addition to the crown clade, Cetaceamorpha includes fossil stem taxa that are successive outgroups to crown Cetacea. Crown and total clade-based taxonomy provides a consistent reference system for both specialists and those less familiar with the systematics of a given clade. As emphasized by, a further significant reason to use crown clades and total clades is their unambiguous representation of data directly available for study. Many types of data (e.g., molecular, soft tissue, behavior) are rarely preserved for direct study outside the crown clade, and thus cannot be optimized below the common ancestor of a crown clade (see also Level 1 inference of). Here we also apply no formal taxonomic rank (i.e., Family, Subfamily, etc.) to the names proposed in this paper. ‘Whippomorpha’ was proposed as the name for “Cetacea +Hippopotamidae”. Subsequently, Cetancodonta was offered as a replacement for ‘Whippomorpha’. We support the formalized use of the term “Cetancodonta” for the crown grouping, based upon arguments made by Arnason when the name was first proposed and the problematic nature of using a term ending in -morpha for a crown clade. We formally define Cetancodonta as a node-based taxon, including all species that are descendants of the most recent common ancestor of *Hippopotamus amphibius* and *Tursiops truncatus*. Cetancodontamorpha is applied to the total clade that includes Cetancodonta and all extinct species more closely related to extant cetancodontans than to any other living species. In addition to artiodactylan taxa, we define the completely extinct clade †Mesonychia as a node based taxon that is the common ancestor of †*Hapalodectes leptognathus* and †*Mesonyx obtusidens* and all of its descendants. We have not altered the nomenclature of taxa utilized as outgroups to Artiodactyla in the present study. The sampling for these clades is not comprehensive enough to warrant the re-examination of taxonomic terms in the present study. We should note, however, that Carnivora, as a node-based crown clade, has been defined elsewhere, as has the total clade, Carnivoramorpha. Our treatments of Artiodactyla and Cetacea mirror this utilization of a traditional ordinal level name (Carnivora) applied formally to a crown clade. Additional outgroup names used in the discussion below are: crown group Perissodactyla, †Creodonta , Ferae, and Lipotyphla. For the above terms that have not yet been formally defined cladistically, their current compositions should be unambiguous given the provided references and the taxa in our analysis. # Results and Discussion ## Minimum Length Trees and Comparisons to Previous Hypotheses The total evidence matrix includes 12,222 parsimony-informative characters (603 phenotypic \[osteology, dentition, soft tissue and behavior\], 11,619 molecular - see). Parsimony analyses in both PAUP\* and TNT recover 20 most parsimonious trees of 57,269 steps. The strict consensus of minimum length trees is fairly well resolved. We rooted the trees with the tubulidentate, *Orycteropus*, as it has been found to be outside of a Carnivora + ungulate clade in a number of studies (e.g.,). The lipotyphlan *Erinaceus* (hedgehog) is positioned at the base of the tree, followed by a split between a monophyletic Ferae and a clade of ungulates. Within Ferae, †Creodonta, Carnivoramorpha, and Carnivora are recovered. In †Creodonta, however, monophyly of the subclade †Hyaenodontidae is not supported, as the one included †oxyaenid , †*Patriofelis*, nests within †Hyaenodontidae. Ferae is the sister group to a diverse clade that includes †Mesonychia, archaic ungulates of uncertain affinities, Perissodactyla and Artiodactyla. †Mesonychia is monophyletic with a basal split between the one included †hapalodectid (†*Hapalodectes*) and †Mesonychidae (†*Pachyaena*, †*Dissacus*, †*Harpagolestes*, †*Mesonyx*, †*Sinonyx*). Resolution within †Mesonychidae is limited. †Mesonychia is the sister clade to the remaining taxa in our analysis (however, see discussion below regarding the instability of this node and its effect on the overall tree). Among the remaining taxa, four archaic ungulates (†*Protungulatum*, †*Hyopsodus*, †*Phenacodus*, †*Eoconodon*) are basal to a clade composed of Perissodactyla plus Artiodactylamorpha. Within Perissodactyla, *Equus* is sister to a Rhinocerotidae plus *Tapirus* clade. †*Hyracotherium* falls outside the crown clade Perissodactyla. Basal relationships of Artiodactylamorpha are poorly resolved in the strict consensus. Four major artiodactylan clades and three extinct species form a polytomy in the strict consensus. The ungrouped species are two †anthracotheriids (†*Anthracokeryx ulnifer*, †*Microbunodon minimum*) and †*Gobiohyus orientalis* (a †helohyid according to, and Artiodactyla *incertae cedis* according to). Cetancodontamorpha, Ruminantiamorpha, Suinamorpha, and Camelidamorpha contribute to this polytomy as well. In all of the most parsimonious trees, Camelidamorpha includes †oreodontoids (†*Merycoidodon*, †*Agriochoerus*), a †cainotheriid (†*Cainotherium*), and the extinct stem camelidamorphan †*Poebrotherium*, which is sister to the two included living camels, *Lama* and *Camelus*. Within the second major clade, Suinamorpha, †*Perchoerus* is the sister to the remaining taxa. The living tayassuid (*Tayassu*) is positioned in a polytomy with †*Xenohyus* and extant Suidae (*Sus*, *Babyrousa*, *Hylochoerus*, *Potamochoerus*). A third large clade is Ruminantiamorpha, which here includes taxa traditionally classified as ruminants, as well as three †anthracotheriids (†*Bothriogenys*, †*Libycosaurus*, †*Elomeryx*) and a †protoceratid, †*Protoceras*. In crown clade Ruminantia, *Tragulus* is sister to extant pecorans (*Antilocapra*, Giraffidae, *Cervus*, *Odocoileus*, *Ovis*, *Bos*, *Moschus*) and the extinct †leptomerycid, †*Leptomeryx*. A close relationship between the moschid (*Moschus*),†*Leptomeryx*, and Bovidae (*Bos*, *Ovis*) is supported. Within Artiodactylamorpha, the fourth large clade recovered is Cetancodontamorpha. The †anthracotheriid †*Siamotherium* is sister to the remaining cetancodontamorphans. Among these, two †entelodontids (†*Brachyhyops*, †*Archaeotherium*) cluster with the †helohyid †*Achaenodon* and †*Andrewsarchus*, the latter being a relatively incomplete fossil from Mongolia that has been historically difficult to classify. Within the crown group Cetancodonta, Hippopotamidae (*Hippopotamus*, *Choeropsis*) clusters with the †anthracotheriid, †*Merycopotamus*, and this clade is in turn sister to Cetaceamorpha. The basal division in Cetaceamorpha is between a clade composed of the †helohyid †*Helohyus* plus the †dichobunid genus †*Diacodexis*, and a second clade that includes †*Indohyus*, *Pakicetus*, †*Ambulocetus*, †*Rodhocetus*, †*Artiocetus*, †*Dorudon*, †*Basilosaurus*, and crown Cetacea (Mysticeti, Physeteridae, Ziphiidae, *Pontoporia*, *Inia*, Monodontidae, Delphinidae). Within this second clade, †*Indohyus* is the sister taxon of others, and there is a pectinate arrangement of extinct taxa at the base of crown Cetacea. Within crown clade Cetacea, Mysticeti is the sister group to Odontoceti. Among odontocetes, Physeteridae is basal, followed by Ziphiidae. *Pontoporia* plus *Inia* cluster, and this group is sister to a clade composed of Monodontidae and Delphinidae. Unresolved sections of the strict consensus result from various equally- parsimonious placements of nine fossil taxa. If the unstable positions of these taxa are ignored, relationships for the remaining 72 taxa in the analysis are consistent across all 20 minimum length trees. This “maximum agreement subtree” reveals additional resolution at the base of Artiodactyla. Among the primary divisions of extant artiodactylans, Cetancodonta (Cetacea + Hippopotamidae) groups closest to Ruminantia (together, Cetruminantia) with Suina, and Camelidae branching as successively more distant relatives of Cetancodonta. This basic pattern is consistent with several previous phylogenetic analyses of molecular and combined data (see discussion). The wholly extinct †Mesonychia, a group of apparently carnivorous mammals from the Paleocene-Eocene periods (∼60-40 mya), has been implicated in the early evolutionary history of Cetacea. †Mesonychians traditionally have been assigned to the stem lineage of Cetacea, , or alternatively have been positioned completely outside of Artiodactyla. The earliest combined phylogenetic analyses of molecules and fossils, included information from two different morphological matrices with extensive DNA sequence data, but this work could not place †Mesonychia consistently. Equally parsimonious trees put this critical taxon deep within Artiodactyla and close to Cetacea, or completely outside of Artiodactyla and distant from Cetacea. This internal conflict resulted in a lack of resolution in strict consensus trees. O'Leary and Gatesy presented the most recent and extensive compilation of evidence bearing on whale origins, over 600 phenotypic characters and \>40,000 molecular characters. In that study, the balance of evidence tipped toward a close relationship between Cetacea and †Mesonychia, with this grouping nested at least five nodes within Artiodactyla. The present study returns a phylogenetic hypothesis that contradicts O'Leary and Gatesy. Instead we find minimum length trees in which †Mesonychia is the sister group to a large clade composed of Artiodactyla (including Cetacea), Perissodactyla, †*Hyopsodus*, †*Protungulatum*, †*Phenacodus*, and †*Eoconodon*. The change in topology can be attributed to a more comprehensive sampling of characters and taxa in the present analysis. The critical importance of sampling is also emphasized when our results are compared with those of Thewissen et al.. Despite the incorporation of critical new character data for the †raoellid †*Indohyus*, our combined analysis of molecules and morphology supports trees that are highly incongruent with the morphological analysis of Thewissen et al. and more similar to those found by other authors. The close relationship between †*Indohyus* and Cetacea is the primary agreement among all of these analyses. The inclusion of this newly discovered †*Indohyus* material is particularly important for the impact it has on the position of †Mesonychia (see below). It is important to note that before hypotheses supporting a close relationship between †Mesonychia and Cetacea, †mesonychians were included in †Creodonta. The modern concept of †Creodonta is more restricted, excludes †Mesonychia, and is composed of two sub-clades: †Hyaenodontidae and †Oxyaenidae. †Creodonts, in turn, have been grouped with Carnivoramorpha (cats, dogs, and close fossil relatives) in a more inclusive clade, Ferae. In our total evidence analysis †Creodonta, Carnivoramorpha and Ferae are all supported, and there is no support for including †Mesonychia within †Creodonta or Ferae. Supplementary lists synapomorphies for several key clades examined in this study. Cetacea is supported by 24 unambiguous synapomorphies, primarily from the cranium. Cetaceamorpha is also supported primarily by cranial synapomorphies. Examination of the node allying †*Indohyus* with other cetaceamorphans indicates that presence of the pachyostotic bulla is one key feature supporting this clade. The base of Cetaceamorpha is united by the presence of a third trochanter and the absence of a meatal tube on the auditory bulla. The condition of the auditory tube in basal cetaceamorphans is only recorded for †*Diacodexis pakistanensis*, (other taxa are represented by “?” for this feature) as inferred from a line drawing in the cranial description of this specimen (the original specimen is lost to science, personal communication, J. G. M. Thewissen). This drawing suggests that the meatal tube is essentially absent, a very rare feature for noncetacean artiodactylans. It would be extremely important to corroborate this observation by discovering additional specimens of †*Diacodexis*. Hippopotamidamorpha is supported by 12 unambiguous synapomorphies from the cranium, the dentition, and the postcranial skeleton; Cetancodonta is diagnosed by 8 synapomorphies that are primarily cranial. Regarding the outgroup taxa sampled, we recover a suite of synapomorphies for Ferae from different anatomical systems (Supplementary). Wyss and Flynn previously suggested that some of these characters (carnassial shear and features of the ankle) are Ferae synapomorphies, but the majority represent new features, not previously discussed, that link †Creodonta and Carnivoramorpha. †Creodonta also is supported by a diverse set of synapomorphies, which have not previously been identified. However, five synapomorphies of †Creodonta are not found in †*Patriofelis*. If the tree is constrained for †hyaenodontid monophyly, these five characters serve as synapomorphies for the clade Hyaenodontidae, with the remaining synapomorphic characters optimizing as †creodontan synapomorphies. If the position of †*Patriofelis* is ignored, the topology for the included †hyaenodontids agrees with but not Gunnell. ## Nodal Support and the Instability of †Mesonychia We used three different approaches to describe the stability of our phylogenetic results: branch support, linked branch support, and selective removal of taxa and characters. The first two methods summarize the net amount of character evidence for a particular clade or set of clades. The third assesses the phylogenetic impact of new taxa sampled here and provides insight into contrasting signals from different types of character data partitions. Branch support scores for nodes found in the total evidence parsimony analysis range from +1 to +11. Clusters of relatively high branch support generally are confined to subclades of the strict consensus that contain only extant lineages. The crown groups Cetacea, Ruminantia, Perissodactyla, and Carnivora each are characterized by at least two clades with branch support greater than +5, and Cetacea includes six groups with branch support of at least +9. A grouping of living and extinct whales (†*Pakicetus*, †*Ambulocetus*, †*Rodhocetus*, †*Artiocetus*, †*Dorudon*, †*Basilosaurus*, Mysticeti, Physeteridae, Ziphiidae, *Pontoporia*, *Inia*, Monodontidae, Delphinidae) has branch support of +3, and the †raoellid †*Indohyus* is sister group to this clade with a branch support of +2. Crown group Ruminantia is well-supported (+6), as is Camelidae (+7), Hippopotamidae (+4), Carnivora (+7), the wholly-extinct †Creodonta (+7), and separation of Lipotyphyla plus *Orycteropus* from the remaining taxa (+8). Ferae (+2), †Mesonychia (+2), Artiodactyla (+1), Cetaceamorpha (+1), Hippopotamidamorpha (+2), and Cetancodontamorpha (+1) are resolved, but are not particularly robust nodes as assessed by branch support. As noted above, the phylogenetic relationships of †Mesonychia have been particularly unstable in recent phylogenetic analyses. Here, †Mesonychia occupies a basal position in our most parsimonious total evidence trees, falling completely outside of a large clade that includes Artiodactylamorpha, Perissodactyla, and a variety of archaic ungulate genera. Multiple nodes separate †Mesonychia from Cetaceamorpha, but examination of slightly suboptimal topologies reveals a set of trees in which †Mesonychia assumes an apical position in the tree, nested within Cetaceamorpha, Cetancodontamorpha, Cetruminantiamorpha, and Artiodactylamorpha. Displacement of †Mesonychia from the base of the tree disrupts eight basal nodes supported by the total evidence, including the close relationship between the †raoellid †*Indohyus* and Cetacea. The sum of branch support scores for the eight nodes is +10, but the simultaneous collapse of all eight nodes in a single tree requires only two extra steps. Linked branch support for the entire set of eight clades is therefore +2. This pattern of interdependent support for adjacent nodes suggests that homoplasy is clumped and not dispersed evenly across the tree. In other words, there is conflicting character support for two very different sets of alternative topologies. †Mesonychia either falls within Cetaceamorpha or is completely excluded from Artiodactyla, but all other possible placements of †Mesonychia are less parsimonious than these two, highly discrepant alternatives. The pattern implies profound character conflict relating to the position of this one group, and the volatility of this critical fossil taxon limits branch support scores at multiple nodes within Artiodactylamorpha. Note, however, that the movement of †Mesonychia in these topologies does not affect the robustly supported relationships among extant artiodactylans in our total evidence matrix (thick gray branches). The instability of †Mesonychia also is apparent from parsimony analyses in which taxonomic sampling is perturbed. Carnivora, †Creodonta, and Lipotyphla (*Erinaceus*) were removed successively from the total evidence matrix, in a variety of combinations, and parsimony searches re-run. Whenever Carnivora is deleted, †Mesonychia groups close to Cetacea and †*Indohyus* clusters with Hippopotamidae in contrast to the total evidence result. For the analysis in which both Carnivora and Lipotyphla are extracted, †Creodonta groups with †Mesonychia as the sister group to Cetacea. Removal of either Lipotyphyla or †Creodonta alone does not result in a repositioning of †Mesonychia within Artiodactyla, but deletion of both Lipotyphla and †Creodonta gives an ambiguous answer. There are two equally parsimonious sets of very different trees; one set positions the †raoellid †*Indohyus* close to Cetacea with †Mesonychia completely excluded from Artiodactyla, while the other set joins †Mesonychia with Cetacea in a clade that is deeply nested within Artiodactyla. This pattern mirrors that seen in examination of nearly optimal trees. Overall, the phylogenetic placements of †Mesonychia and †*Indohyus* are highly sensitive to the particular outgroup taxa included in analysis. Representatives from Carnivora, †Creodonta, and Lipotyphla are all required to give clarity to character polarities. If only one of these groups is present, the positions of †Mesonychia and †*Indohyus* changes dramatically. This underscores not only the importance of broad taxon sampling, but also the instability of †Mesonychia and †*Indohyus*, particularly when compared to the relatively stable phylogenetic ‘backbone’ of Artiodactyla that is supported by data from extant lineages. Excluding the controversial postcranial evidence for †*Indohyus* has little effect on phylogenetic results; nine optimal trees were recovered that were a subset of the 20 minimum length trees for the total evidence matrix. The strict consensus of these nine trees is slightly more resolved than the strict consensus derived from the complete combined data set. Removal of †*Indohyus* entirely has a much more profound effect. †Mesonychia again moves from a basal position to a highly nested placement within Artiodactyla, close to Cetacea. This suggests that the unique combination of characters in the skull of †*Indohyus* has a very large influence in determining results in combined analysis of molecules and morphology. Additional perturbations of the total evidence matrix included analysis of only skeletal and dental characters (characters that fossilize), and a search that considers only molecular, soft tissue, and behavioral characters (those for which we generally lack data for the fossils sampled). Analysis of molecular characters, soft anatomy, and behavior from extant taxa yields a tree that is generally consistent with the total evidence analysis. Separate analysis of characters that commonly fossilize supports a very different result. The traditional artiodactylan clades Selenodontia (Ruminantia + Camelidae) and Suiformes (Hippopotamidae + Suina) are supported, and †*Indohyus* does not group close to Cetacea. Instead, Cetacea clusters within a paraphyletic †Mesonychia, and this grouping is excluded from crown clade Artiodactyla. This means that the recent discovery of fossils such as †*Indohyus* and †*Rodhocetus*, which are critical taxa in early cetaceamorphan evolution did not result in congruence between phylogenetic data from the fossil record and data from living taxa (see also discussion). The skeletal and dental characters alone do support Ferae, Carnivora, Caniformia, and †Creodonta), but overall, there is extensive conflict with the total evidence analysis. If the combined data matrix is fit to the topologies generated by the skeletal+dental data, these trees are at the least 1,500 steps beyond the minimum length. ## Selected Character Optimizations for Cetancodonta ### Character reconstructions based on parsimony alone Since the recognition that whales are highly-derived artiodactylans, it has been of interest to understand how an aquatic, carnivorous clade, Cetacea, evolved within a predominantly terrestrial, herbivorous clade, Artiodactyla. Our primary means of reconstructing characters in hypothetical common ancestors, and of reconstructing soft tissue/behavior characters (that are not directly observable) in fossil taxa, is to use parsimony. We have coded the following behavioral characters in our matrix: aquatic habitat (character 618 \[state 1\]), herbivorous diet (character 658 \[state 2\]), and ability to interpret the direction of sounds under water (character 659 \[state 1\]). These three character states optimize unambiguously to the common ancestor of Cetancodonta. This corroborates predictions about the origin of an aquatic lifestyle as having occurred once in the common ancestor of Hippopotamidae and Cetacea. This common ancestor of Cetancodonta had the derived behavior of spending at least 10% of its time in water (character 618 \[state 1\]). Parsimony indicates that this state is shared by all extant members of Cetancodonta, and is reconstructed for all taxa nested within Cetancodonta (including basal cetaceamorphans, such as †*Indohyus*, †*Diacodexis* and †*Helohyus*). Gatesy et al. had previously suggested overturning a parsimony-based optimization for some extinct taxa nested in this clade based on absence of osteological correlates for aquatic behavior, but we do not advocate that position here (also see). Using parsimony we reconstruct all fossil hippopotamidamorphans as herbivorous (character 658 \[state 2\]). The dietary behaviors of taxa along the stem to Cetacea (i.e., noncetacean cetaceamorphs) are, however, equivocal based on optimization of states seen in extant taxa. Somewhere on the stem to Cetacea, diet changed from herbivory to aquatic carnivory (character 658 \[state 3\]), but using parsimony alone we cannot reconstruct unambiguously where the behavioral change occured Parsimony-based optimization also implies that all living and extinct cetancodontans shared a derived ability to hear underwater sounds at least as well as extant *Hippopotamus*. This is noteworthy because several cetancodontans (including living members of Hippopotamidae) lack a pachyostotic auditory bulla (involucrum) in the ear. To summarize, in the minimum length trees, the †raoellid †*Indohyus* is reconstructed to have spent at least 10% of its time in water and to have had the derived behavior of directional underwater hearing, but reconstruction of its diet is equivocal. In alternate trees that are two steps longer, parsimony recovers the same character state reconstructions for †Mesonychia, because this taxon is a close relative of Cetacea in slightly longer trees. ## Extended character reconstructions: inferring behavior from osteology and dentition As discussed by, inferences about behavior in fossil taxa, which go beyond parsimony, can be made if there is “compelling morphological evidence” that a certain fossilized trait is strictly correlated with a certain behavior (e.g., distinctive coiling of the cochlea in bats indicating echolocation). These deductions should, however, be clearly delineated from reconstructions based on parsimony. Here we discuss such inferences related to diet and hearing in Artiodactyla. Molars that have a tall, angular protoconid and a compressed talonid are typically associated with carnivorous diets in mammals, and molars with low- crowned, quadritubercular cusps are associated with herbivory/omnivory. Reconstructing behavior from fossilized tooth shape, we would infer that several cetaceamorphans (†*Diacodexis*, †*Helohyus*, and †*Indohyus*) are herbivorous/omnivorous because they have quadritubercular teeth (hypocone on M2, character 419). It is noteworthy that prior to the description of a relatively complete †*Indohyus* skull there were no cetaceamorphans that had both the pachyostotic ear region and quadritubercular dentition. Molar shape would suggest that cetaceamorphans from †*Ambulocetus* through more highly nested taxa were carnivorous due to the presence of narrow talonids on m2 (character 364, state 1) and of tall protoconids on m1 (358, state 1; here technically †*Pakicetus* and more highly nested taxa due to missing data). According to minimum length trees for the combined data, these character states were independently derived within Cetaceamorpha and in †Mesonychia. In the slightly longer topology, however, both of these dental characters are synapomorphies uniting †Mesonychia and Cetacea. Based on the results of the total evidence analysis, we can return to the question posed in the introduction, ‘Did aquatic carnivory precede committed life in the water?’ Inferring carnivory from tooth shape and inferring committed life in the water from detachment of the sacral vertebrae from the pelvis (character 488, state 0), carnivory did precede committed life in the water. Not until the last common ancestor of †*Dorudon*, †*Basilosaurus*, and crown Cetacea did cetaceamorphans lose the articulation between the pelvis and the vertebral column, but dentition suggesting carnivory appears at a more basal node (see also). The presence of the pachyostotic bulla and what it implies about hearing has also been of interest because this is a relatively rare anatomical feature among mammals. This structure has been argued to indicate an ear region derived to process underwater sounds, a behavior that has evolved in Cetaceamorpha. Luo and Gingerich (:89) stated that acoustic isolation of the left and right ears creates density differences analogous to those created by a pachyostotic bulla, and may confer directional hearing underwater. Interestingly, *Hippopotamus* lacks a pachyostotic bulla despite the fact that this species has an ability to hear underwater sounds exceeding that of typical terrestrial mammals. Pachyostosis of the ear region, therefore, does not appear to be essential for certain derived types of underwater hearing. Pachyostosis may indicate instead an even more derived level of underwater hearing than previously recognized but confirmation requires further functional studies. Reconstructing auditory function from osteology, we would infer that the presence of a pachyostotic bulla (character 59, state 1) in †*Indohyus*, and all more highly nested cetaceamorphans, potentially indicates an even more derived state of underwater hearing than that which developed in the common ancestor of Cetancodonta. If the alternate topology only 2 steps longer obtains in future studies, (with †Mesonychia closer to Cetacea and †*Indohyus* a more distantly related cetaceamorphan), then the pachyostotic bulla developed two times independently in Cetaceamorpha: once in the common ancestor of †*Ambulocetus* and Cetacea, and once in †*Indohyus*. Furthermore, optimization by parsimony indicates that underwater hearing is a feature shared by all cetancodontans, thus this character state appeared at a more basal node than detachment of the sacral vertebrae from the pelvis and committed life in the water (character 488, state 0). In summary, these inferences imply that the history of Cetaceamorpha included both carnivorous and herbivorous species. All cetancodontans were at least as aquatic as living hippos and exhibited some ability to hear underwater sounds, even though several cetancodontan species lack a pachyostotic bulla. Appearance of the pachyostotic bulla may indicate a shift to a yet more derived degree of directional underwater hearing, a hypothesis that requires further investigation. In slightly longer trees, in which †Mesonychia groups close to Cetacea, the derived bulla of †*Indohyus* would be interpreted as convergent with that of cetaceans. Finally the shift to carnivory within Cetaceamorpha preceded the loss of limbs that functioned in terrestrial locomotion in this clade. ## Conclusions This study highlights the importance of expanded taxon sampling when examining complex questions of relationships and underlying patterns of character evolution. The addition of carnivorans and †creodonts, taxa not traditionally included in discussions of artiodactylan/cetacean phylogeny, has a significant impact on the resultant tree topology. The complex combined data set compiled for the present study underscores some of the key issues remaining in studies of cetacean origins. Future analyses should continue to expand taxon sampling. The current matrix, although relatively large, remains highly unstable to slight perturbations in taxon sampling. Initiatives underway, such as the mammalian component of the Assembling the Tree of Life project, seek to increase both the number of taxa and characters utilized in combined data phylogenetic analyses. Such comprehensive studies will more fully explore the influence of outgroups on deeply nested ingroup relationships. †Mesonychia is only distantly related to Artiodactyla in our shortest trees, with †*Indohyus* grouping as a close relative to living cetaceans. However, in trees just two steps longer than minimum length, we find the more ‘traditional’ arrangement of †Mesonychia positioned close to Cetacea. In these trees, †*Indohyus* is a cetaceamorphan but is not as closely related to Cetacea as is †Mesonychia. The lack of abundant support for either topology and the outstanding incongruence between data that fossilize and those that do not, suggests that many key fossils remain to be discovered. Ferae and †Creodonta are both solidly supported clades in our total evidence analysis, each with multiple synapomorphies. This study is the largest test of the relationships of these taxa to date and utilized many characters that had never before been applied to members of the Ferae in a cladistic context. Additional work is needed to test the monophyly of these groups, such as the addition of several taxa suggested in the past to be closely related to †Creodonta (e.g., †Leptictidae) as well as much more comprehensive sampling of the †creodonts themselves. However, despite the need for further research, this analysis shows that discounting Ferae and †Creodonta as monophyletic groups, is premature. # Materials and Methods ## Taxon and Character Sampling The large morphological character matrix previously compiled by O'Leary and Gatesy included 71 taxa (28 extant, 43 extinct) and 635 characters (310 cranial osteology, 147 dental, 123 postcranial osteology, and 55 soft-tissue/behavior). This data set for Artiodactyla and close relatives was used as a starting point for the present analysis. We generally chose representatives of extinct groups based on the relative completeness of fossil material. Four members of the extinct order †Creodonta were included: three †hyaenodontids (†*Thinocyon*, †*Sinopa*, †*Hyaenodon*) and one †oxyaenid (†*Patriofelis*). Five carnivoramorphans were added: one basal extinct taxon (†*Vulpavus*), and within Carnivora two extant feliforms (*Nandinia* and *Felis*), and two extant caniforms (*Canis* and *Ailurus*). We also sampled an extant lipotyphlan insectivore (*Erinaceus*). We based morphological character codings for all extant genera on examinations of single species, but when compiling DNA sequences (see below), monophyly of extant genera was assumed to reduce the overall percentage of empty cells for molecular characters in the matrix. To limit missing data for the five fossil taxa, observations from multiple species were merged for each extinct genus. All characters were scored based upon direct examination of specimens. Morphological character codings were collected and archived using the web application Morphobank. A primary objective of this study was to assess the phylogenetic impact of newly described †raoellid artiodactylan fossils in the context of a very large combined matrix of molecular and morphological characters. In addition to the ten taxa added to the matrix, we augmented our previous set of character state observations for †*Indohyus*. M. O'Leary examined †*Indohyus* specimens in the collection of H. Thewissen at Northeastern Ohio Universities College of Medicine (NEOUCOM). We were permitted only to corroborate matrix scores for the 196 characters published in because the †raoellid fossils are still being examined by Thewissen and colleagues. Thus there are missing data for this taxon that could have been collected if we were allowed to make new observations for our full matrix of 661 morphological characters. We increased morphological character sampling slightly relative to the analysis of O'Leary and Gatesy. Approximately five morphological characters were added based upon previous systematic work on Ferae. This count is not exact because many characters were at first appended to the previously published matrix, but then later subsumed into existing characters once overlaps in character states were identified. Delimitations of some characters in the matrix from were revised based upon new information from Ferae/Lipotyphla. There is an overall increase of 26 morphological characters relative to our previous matrix due to the addition of characters and re-defining of previous characters. We also augmented the extensive molecular data (40,928 characters) compiled by. First, data for the five new extant genera sampled here (*Canis*, *Felis*, *Nandinia*, *Ailurus*, *Erinaceus*) were added to our previously published alignments; four mitochondrial genomes and information from 31 nuclear loci at the Genbank database were added to the overall matrix. We also included additional new data from Genbank that have been published since ; for example, sequences from the nuclear genes *TBX4* and *SRY* were concatenated to the existing molecular data set. Finally, 49 new sequences from five nuclear genes (*ZP3*, *BDNF*, *ATP7A, AMEL, RNASE1*) were generated in our lab for this study (Genbanks \#s GQ487580-GQ487628) using PCR, cloning, and sequencing methods described in. PCR/sequencing primers for the *ZP3* gene were (5′ to 3′): ZP3L1 - GACCAACTAAACAAAGCCTG, ZP3L2 - CAGCAAGTCCTCCAACAGGT, ZP3ODOL1–GAGACCAGATTGGACATAAC, ZP3ODOR1–GCACACAGGGTGGGAAGCAG, and ZP3R3–TATTGGGAAGCAGACAC. Published primers were used for the *ATP7A*, *BDNF*, *AMEL*, and *RNASE1* genes. Recently deposited data in Genbank and sequences from our lab generally were aligned to our previously published matrix with the introduction of very few new gaps \[e.g., see 1\]. However, several gene segments were re-aligned using CLUSTALW with gap opening cost of five and gap extension cost of 1; some adjacent gaps in the resulting multiple-sequence alignments were consolidated using SeqApp 1.9a as in. All newly-incorporated loci (*TBX4*, *SRY*, *ZP3*) also were aligned in this way. The final molecular data set exceeded that of O'Leary and Gatesy by more than 5,500 aligned nucleotides. The 661 morphological characters were downloaded from Morphobank and merged with the revised molecular matrix of 46,587 characters in PAUP\* 4.0b10. The total combined data set for this study has been stored at Morphobank (project \#48). The main matrix in this project file is the morphology component of this study, and the total evidence nexus file is in the documents folder for this project. The nexus file records all Genbank numbers for molecular sequences in the matrix. This nexus file is also available as supporting information for this article:. ## Phylogenetic Analyses Parsimony analyses of the total evidence data set were undertaken using both PAUP\* 4.0b10 and TNT. In PAUP\*, searches were heuristic with 1000 random stepwise additional replicates and tree-bisection-reconnection (TBR) branch swapping. All character state changes were given equal weight, all characters were unordered, gaps were treated as missing data, and the amb- option was used so that internal branches were collapsed if minimum length was zero. The search strategy employed in TNT was to first analyze the data under the ‘New Technology search’ option, selecting the sectorial search, rachet, and tree fusing search methods, all with default parameters. Under this setting, iterations were run until the minimum length tree was found in 500 separate replicates, to try to hit as many islands of trees as possible. The generated trees were then analyzed under traditional search options (using TBR) in order to more fully explore the discovered tree islands. Strict consensus trees and agreement subtrees were used to examine topological conflicts among multiple most parsimonious trees. Branch support was used as a measure of nodal stability for all groups resolved in the strict consensus of minimum length trees, and linked branch support was estimated to summarize the interdependence of character support for multiple nodes supported by the total evidence. Branch support for a particular clade can be defined as the length of the shortest tree that does not include the clade, minus the length of the shortest tree that includes that clade. Estimates of branch support were derived from additional heuristic searches in PAUP\* that incorporated “anti-constraints,” and also by searches in TNT that retained sub- optimal trees and determined at what length each recovered clade was lost in a strict consensus. Linked branch support for a particular set of supported clades can be defined as the length of the shortest tree that does not include *any* of those clades, minus the length of the shortest tree that includes *all* of those clades. Further PAUP\* searches with “anti-constraints” enforced were used to estimate linked branch support for particular groups of nodes. Given extensive blocks of missing data in the combined matrix, character resampling (bootstrap) was not utilized to assess nodal support in this study, and estimates of branch support in our trees may be lower than the estimates calculated here. The total evidence analysis of all data was considered the best test of phylogenetic relationships, but we explored other search strategies to examine different signals in the combined matrix, and to test the overall stability of our results. A variety of searches were conducted, with different subsets of taxa and characters activated in each analysis. These were: 1) All characters and all taxa included. 2) Only skeletal and dental characters for all taxa; this search summarized the strongest hierarchical signal in traits that commonly fossilize. 3) Only molecular, behavioral, and “soft anatomical” characters for extant taxa; this run shows the pattern supported by information that generally can only be coded from extant taxa. 4) All taxa and most characters, with postcranial characters from †*Indohyus* excluded. This analysis assessed whether or not our total evidence results are dependent on possible misassignment of bones to this critical taxon. 5) All taxa included, except for †*Indohyus*; this analysis determined the influence of this critical ‘intermediate’ taxon on phylogenetic results. 6) All characters and most taxa, with some outgroup taxa deleted from analysis. All combinations of three higher-level groups (†Creodonta, Carnivoramorpha, Lipotyphyla) were successively deleted to see the effects of outgroup sampling on phylogenetic results, in particular placement of Cetacea relative to †Mesonychia. Characters were optimized onto all minimum length trees using the map characters option in TNT. For critical nodes supported by the total evidence, character state changes that mapped unequivocally onto all optimal trees were noted; these are listed in Supplementary. We also used parsimony to map characters onto suboptimal hypotheses to identify transformations that support conflicting relationships regarding †Mesonychia and †*Indohyus*. # Supporting Information We thank J. Geisler, A. Berta, O. Rieppel, K. de Queiroz, and J. Gauthier for helpful discussions on taxonomy, S. Parent for assistance with morphological data collection, and J. Robalino and A. Giallombardo for general discussions. We thank J. Finarelli, A. Farke, and one anonymous reviewer for helpful comments that greatly improved this manuscript. The New York Zoological Society, Southwest Fisheries Science Center, (specimen ID \#s Z176–*Lissodelphis borealis*, Z4502–*Balaenoptera musculus*, Z21–Kogia breviceps, Z505–*Inia geoffrensis* from Smithsonian Institution Division of Mammals, Z7349 - *Pontoporia blainvillei*), The Zoological Society of San Diego, O. Ryder, G. Amato, G. Schaller, P. Morin, A. Dizon, K. Robertson, M. Milinkovitch, H. Rosenbaum, K. Hecker, M. Cronin, M. Springer, and P. Arctander provided DNA samples. B. Kong and H. Lee helped with sequencing the *ZP3* gene data. Carl Buell executed artwork (full body reconstructions), and L. Betti-Nash drew the skeletons. [^1]: Conceived and designed the experiments: MS MAO JG. Performed the experiments: MS MAO JG. Analyzed the data: MS MAO JG. Contributed reagents/materials/analysis tools: MS MAO JG. Wrote the paper: MS MAO JG. [^2]: The authors have declared that no competing interests exist.

# Introduction Since its domestication in China more than 5,000 years ago, the genetic variability of cultivated peach, *Prunus persica* (L.) Batsch, has been shaped by three major forces. The first is inbreeding, favored by the self-pollinating system of peach, an unusual situation compared to most other *Prunus* species that have an efficient gametophytic self-incompatibility system. Selfing, plus the directional selection of the domestication and the selective processes of plant breeding, have resulted in a dramatic loss of variability when compared to other cultivated species of the genus. The second force is random drift, as a consequence of the small number of parents used in the breeding programs that started in the US in the middle of the 20th century. A limited number of founders may also have been used in Asian breeding programs, as inferred from the work of Li et al. The third force is heterosis that, because peach is easily propagated by grafting, can be exploited in commercial breeding. The usual breeding scheme of peach and most fruit trees is the selection of individuals arising from the first generation of crosses between two accessions that are themselves, and therefore their offspring, partly heterozygous. The consequence of all this is that cultivated peach has a low level of genetic variability, high linkage disequilibrium conservation and, in most modern cultivars, a relatively high level of heterozygosity. This has been demonstrated using sets of SSR markers with good genome coverage, and, more recently, using resequencing data. Due to its economic value (second, after apple, among temperate fruits) and its self-pollinating behavior, peach has been one of the tree species most amenable to genetic studies, becoming a model system for genomics research in the Rosaceae. In particular, the genomes of different *Prunus* species are highly conserved, allowing many major genes and QTLs from peach and other *Prunus* to be positioned on a single genetic map, using a set of high quality and comparable linkage maps, most developed at the end of the last century. Recently, the whole genome sequence of peach has been released, including the resequencing data of several peach accessions, opening a new era for the development of genetic studies and their application to modern breeding in this species. It has been possible to estimate the SNP variability of this species and a 9k SNP array v1 has been developed by the International Peach SNP Consortium (IPSC). This SNP array is currently being used for genetic analysis. The high abundance and relatively low cost of SNPs compared to SSRs allows for a complementary and deeper study of peach germplasm variability. In this paper, we used the IPSC 9K SNP array to genotype a large collection of peach accessions including an ample set of Occidental and Oriental materials. Our results confirm previous findings on variability parameters and subpopulation structure, provide a finer analysis of linkage disequilibrium (LD), and allow the whole-genome association analysis of a set of morphological characters with Mendelian inheritance. # Results ## IPSC 9K SNP array performance We genotyped 1,580 *Prunus* accessions (1,576 *P*. *persica* plus four *Prunus*-related species) using the IPSC peach 9K SNP array v1. The average call rate (number of calls divided by the number of assayed SNPs) was 0.839 and, after exclusion of 40 samples with the poorest quality (call rate lower than the average call rate minus 0.1), the mean call rate increased to 0.854. The list of varieties with genotypic data is presented in. The 8,144 SNPs assayed were divided into five classes from A to E depending on their performance. The SNPs classified as A (no-Call \< 5% and all three possible genotypes, AA, AB, and BB), represented 53.2% of the total (4,330). After removing those with minor allele frequency (MAF) lower than 0.05 and those with poor clustering in GenomeStudio Data Analysis software (Illumina Inc.), the final set of class A SNPs was 4,271 (52.4%). The percentage of polymorphic SNPs increased to 65.5% when the SNPs in class B (null allele or preferential annealing) and C (duplicated sequences/genes) were included. Only the 4,271 class A SNPs were used for subsequent analysis, corresponding to an average of 18.6 SNPs/Mb (considering the total peach genome size of 230 Mbp) and approximately equivalent to 8.2 SNPs/cM of the reference *Prunus* map (519 cM as in Dirlewanger et al.). ## Germplasm characterization We compared all pairs of accessions to exclude known and unknown sport mutations or synonyms from the analysis, and detected 173 groups encompassing 473 cultivars with two or more cultivars with identical genotype (more than 98% of their SNP genotypes identical). Consequently, 300 cultivars were removed, retaining only one accession for each unique genotype. The largest group of clones contained 11 genotypes, two had nine, eight, seven and six genotypes, respectively, and the rest were distributed in four groups of five, 17 of four, 28 of three and 115 of two genotypes. Most (76 out of 83) of the samples with the same name had the same genotype, including all those introduced as controls (‘Rich Lady’ and ‘Diamond Ray’ with five replicates each, and ‘Maycrest’, ‘Big Top’, ‘Elegant Lady’ and ‘Jing Yu’ with four replicates, each from a different germplasm bank). Five pairs of accessions with the same name from two different germplasm collections (‘Akatsuki’ ‘Siberian C’, ‘Suncrest’, ‘Fei Cheng Bali’ and, ‘Wu Yue Xian’), expected to be identical, were different. Additionally one replicate out of three of ‘Armking’ and one out of four of ‘Rubirich’ didn’t present the expected genotype. Some known sports were included in the panel. In all cases they grouped together. The largest group of identical genotypes included all known sports of ‘Springcrest’: ‘Maycrest’ (four replicates, each from a different repository), ‘Early Maycrest’, ‘Queencrest’, ‘Springbelle’, ‘Springold’ and ‘Springlady’, plus two cultivars of unknown origin, ‘Harken’ and ‘San Giovanni’. Their genotypes were identical for all markers analyzed. Another group included ‘O’Henry’ and its sports ‘John Henry’ and ‘Summer Lady’. In this case, differences were identified in 11 SNPs, with ‘O’Henry’ differing from ‘John Henry’ by nine SNPs and 11 from ‘Summer Lady’, with ‘John Henry’ and ‘Summer Lady’ only differing by one SNP. All these SNPs were located at a region of 5.4 Mb at the top of chromosome 6 (between markers SNP_IGA_604703 and SNP_IGA_621593), where there was a large amount of missing data for many other SNPs in the derived sports ‘John Henry’ (41 missing data out of 92 SNPs) and ‘Summer Lady’ (38 out of 92). A similar pattern was observed for the same genomic region in ‘Redskin’ and its sport ‘Redkist’ (13 differences; 44 missing data out of 92 SNPs in ‘Red Kist’) and in the two groups of cultivars classified as having the same genotype and of unknown relationship: ‘Vittorio Emanuele’ and ‘Paola Matteucci’ (differing by 8 SNPs in the same region and with ‘Paola Matteucci’ having 60 missing data out of 92 SNPs) and four Italian accessions with different names ‘Pieri 81’, ‘Bella Lucia’, ‘Vaccaro Roccalmunto’ and ‘Zingara Nera’, where the former three were identical and the latter differed by four SNPs in this region and had 46 out of 92 SNPs with missing data. In all cases, the differences in SNPs occurred at loci that were heterozygous in the complete accessions and homozygous in those having missing data. A similar case but involving a different region was observed between cultivars ‘Springtime’ and ‘Starcrest’. These were identical in all markers except for eight SNPs located in the upper part of chromosome 4, within a region of 17.9 Mb where there was also a large amount of missing data (179 in ‘Springtime’, 231 in ‘Starcrest’, 178 of them in common with the total 581 markers in the region). As in the previous cases, all distinctive SNPs were homozygous in one variety (‘Starcrest’) and heterozygous in the other. ## Cultivar variability and population structure A total of 1,240 unique accessions were genotyped with the 4,271 class A SNPs. The observed mean heterozygosity (H_o) per individual was 0.286, ranging from 0.003 in ‘Burrona di Rosano’ to 0.680 in ‘Zheng huang 4’. H_o per SNP ranged from 0.046 in SNP_IGA_762094 to 0.781 in SNP_IGA_550718. Chromosome 6 (Ho = 0.28) was the least heterozygous while chromosome 1 had more loci in heterozygosis (Ho = 0.33). The mean expected heterozygosity (H_e) was 0.39, ranging from 0.055 to 0.500. The mean average inbreeding coefficient (F = (H_o—H_e) / H_o) was 0.27, ranging from -0.705 to 0.9922. The vast majority of the markers (98.5%) showed significant (p \< 0.001) deviation from the Hardy- Weinberg Equilibrium (HWE). A phylogenetic dendrogram with the 1,240 unique accessions clearly revealed two main groups, one with Oriental accessions and the other including those from Occidental origin. Two clusters could be distinguished within the Occidental group, where the traditional/non-breeding accessions were differentiated from those derived from breeding programs. The position of each accession in the dendrogram is provided in the. A first approximation to the study of population structure was obtained using principal component analysis (PCA) for the complete set of SNPs, which separated the Oriental and Occidental accessions. The Occidental accessions were clearly separated into two major groups, one including the traditional/non-breeding accessions and the other with varieties used and obtained in modern breeding programs. The separation in three clusters was not absolute and a discrete number of accessions occupied a centric position. The majority of these accessions are known to come from crosses between accessions in different clusters. The software STRUCTURE was used to obtain a more detailed picture of the stratification in the panel. In agreement with the results of the PCA, the most probable number of subpopulations inferred by this method (K) was three. Considering only the accessions with a subpopulation membership coefficient higher than 0.8, the three subpopulations are: population 1, the Occidental materials used in modern breeding programs (P1OC_B, with 352 accessions), population 2, the traditional/non-breeding occidental accessions (P2OC_T, 165 accessions) and population 3, the oriental accessions (P3OR, 58 accessions). At K = 2, the division between the Oriental and Occidental cultivars was already evident. Increasing the number of subpopulations to four the Occidental breeding accessions were divided into two subgroups: the first comprising the vast majority of nectarines and the second the vast majority of peaches. Further increments of K maintained these four subpopulations, with additional ones empty or including only a few individuals with a membership coefficient ≥ 0.8. At all K values, more than half of the individuals (665) were not assigned to a single population and were considered as admixed (ADM). The position in the UPGMA tree of the accessions assigned into each subpopulation at K = 3 is presented in, where accessions belonging to P1OC_B, P2OC_T and P3OR are colored in blue, green and red, respectively. Gray color indicates accessions in the ADM subpopulation. Distance between populations was calculated through the Fst statistic. Genetic differentiation between P2OC_T and P3OR and between P1OC_B and P3OR were moderate (Fst = 0.14 and Fst = 0.13, respectively), while P1OC_B and P2OC_T showed a greater differentiation (Fst = 0.18) ## Linkage disequilibrium (LD) The LD, measured as r², was calculated in the three subpopulations and in the admixed accessions separately, as well as in all 1,240 accessions together using a modified r² that corrects for population structure and relatedness. Analyzing the SNP data in each subpopulation, 188, 949 and 353 out of the 4,271 SNPs in P1OC_B, P2OC_T and P3OR, respectively, failed the frequency test (MAF \< 0.01) and/or the non-calling cut-off of 5% and were discarded. The average value of intra-chromosomal r² was 0.096 in P1OC_B and P2OC_T, 0.133 in P3OR, and 0.082 in the admixed individuals. As expected, the average value of inter-chromosomal r² was smaller than the intra-chromosomal and was 0.016 in P1OC_B and admixed individuals, 0.018 in P2OC_T, and 0.042 in P3OR. The LD decayed with distance between markers in all subpopulations. The average value of r² dropped below 0.2 at 1.4, 1.0, 1.8, and 0.8 Mb in P1OC_B, P2OC_T, P3OR and ADM, respectively. Chromosomal heat-maps showed the presence of several blocks with a high level of LD spread in almost all the chromosomes, generally spanning hundreds of kilobases (– Figs). Three of these blocks were common to all subpopulations, two at the top end of chromosome 4, of approximately 1.2 and 2 Mb, and one in chromosome 5 spanning about 2Mb. The organization of the P1OC_B genome in LD blocks was highest, especially chromosomes 1, 2, 3, 4 and 5. The largest block of SNPs in high LD (r² \>0.5) was observed in LG1, starting at 8 Mb from the top of the chromosome and spanning 12 Mb. This block was followed in length by a 6 Mb block in the same chromosome at 32 Mb from the top. In P2OC_T and P3OR, the SNPs at the bottom of chromosome 6, spanning 2 Mb, were in strong LD. The chromosomal heat-maps of the ADM population revealed only a few blocks of SNPs in high LD, most disappearing when the whole collection of 1,240 cultivars were analyzed with an LD statistic that corrects for population structure and relatedness, only the three common to all four subpopulations remaining. ## Genome-wide association and haplotypes for major genes Between 832 and 1,071 of the 1,240 genotyped plants were phenotyped for the following traits, known to be determined by single Mendelian genes: fruit pubescence (*G/g*), fruit shape (*S/s*), fruit flesh color (*Y/y*), non- melting/melting flesh (*M/m*), titrable acidity (*D/d*), leaf gland type (*E/e*), and showy/non-showy flower type (*Sh/sh*) (see for details). For all except fruit shape, the two phenotypic classes were homogeneously distributed. For the trait of fruit shape, the flat individuals (57) were underrepresented compared with round ones (1,003). Association between single SNPs and traits was identified, taking into account kinship and structure. The average value of kinship was 0.58, on a matrix scaled at 2, corresponding to about one quarter of the genome equivalent for example to a grandparent-grandchild, or a half sib relationship. We found significant association for all the traits (-log₁₀(p) \> 5.63). ### Titratable acidity (TA) The acid/sub-acid fruit trait has been mapped at the beginning of LG5, where the sub-acid allele behaves as dominant. A total of 405 and 427 accessions were classified as acid (TA\>8.2 mEq/100 ml.) and sub-acid, respectively. Twenty- three SNPs at the beginning of chromosome 5 were associated with fruit acidity. The associated SNPs cover a region of about 1.8 Mb (scaffold_5: 467,067..2,270,122) in agreement with the reported position for the D locus. SNP_IGA_544640 (scaffold_5: 629,641) showed the strongest association with the trait, with a p-value of 1.5E-25 and MAF of 0.21. The most frequent allele was observed in 71.4% of sub-acid varieties, in either homozygosity or heterozygosity, while this allele was only present in 3.2% of the acid varieties. The genetic variance (heritability) explained by the model adopted for the association was 0.66. In total, the 23 SNPs produced 99 haplotypes, imputed using fastPHASE, with an average heterozygosity of 17.13%. Four haplotypes showed a frequency higher than 5%, two being strongly associated with the sub-acid phenotype (chi-squared p-value \< 2.2E-16). ### Fruit flesh color At least three different recessive alleles of the CCD4 gene (ppa006109; scaffold_1:25,639.445..25,641.500) are responsible for the yellow flesh in peach, one produced by a mutation in a microsatellite at the first exon of the gene, one by an SNP in the second exon and one by the insertion of a transposable element. In our collection, a total of 525 accessions were yellow- fleshed (*y*) while 499 were white (*Y*). The chromosomal heat-maps showed that the SNPs flanking this gene were in strong LD in a region 2 Mbp long (– Figs). This block was shorter in P2OC_T and disappeared in P3OR (where most of the cultivars had white flesh) and ADM. Association analysis identified 21 SNPs located in a 5.2 Mb region in linkage group G1 (scaffold_1:23,352,245..28,551,537) associated to this trait. The SNP with the strongest association was SNP_IGA_90213 (scaffold_1: 26,479,525), with a p-value of 8.03E-22 and MAF of 0.21. The most frequent variant at this position was strongly linked to white flesh, being present in 61.9% of the white and 6.8% of the yellow varieties. The heritability explained by the allele was 0.65. The 21 associated SNPs were distributed in 152 imputed haplotypes. Six of them showed a frequency higher than 5%, all associated with the trait (Chi-square p-value \<0.01), and an average heterozygosity of 92%. ### Fruit hairiness The recessive glabrous fruit trait (nectarine) has recently been associated to a retrotransposon insertion in an MYB gene (ppa026143; scaffold_5:15,897,836..15,899,002). Of the 1,071 accessions studied, 792 were hairy (peaches, carrying the *G* allele) and 290 glabrous (nectarines, homozygous for the *y* allele). We found 19 SNPs associated with the trait, mapping in G5 (scaffold_5: 13,952,707..16,774,236). The average MAF of the associated SNPs was 0.31 and the heritability 0.73. SNP_IGA_603047 (scaffold_5:16,774,236) showed the highest association (p-value = 2.5E-19, MAF = 0.37 and R² = 0.66). The most frequent allele was equally present in peaches and nectarines while the less frequent was almost exclusively in peach (98.2% of the varieties with this allele were peaches). PHASE imputed 123 haplotypes using the 19 associated SNPs, five with a frequency \> 5% and an average heterozygosity of 8.15%. The most frequent was in high association with the recessive trait: 98% of the accessions with this haplotype in homozygosis had the recessive phenotype. ### Flat fruit shape Flat fruit shape is caused by a dominant allele in obligated heterozygosity, mapped at the end of LG6. In our panel, only 57 cultivars were flat whereas 1,003 were round. The low number of flat accessions hampered the analysis and it was not possible to identify significant association using the full dataset. This situation could be partially explained considering the relatedness of some flat individuals, for example those from the series UFO (12.3% of flat varieties), and the strong structuring of the trait. In order to solve these drawbacks, we studied 15 independent subsets containing the 57 flat and 100 randomly chosen round cultivars from P1OC_B and admixed accessions. In all the subsets, a single peak of association at the end of chromosome six was evident (scaffold_6:23,101,004..26,601,733). The number of associated SNPs varied between subsets, ranging from five to 22. Five were associated with the trait in all runs (SNP_IGA_683904, SNP_IGA_684085, SNP_IGA_685825, SNP_IGA_696241, SNP_IGA_696280), with an average MAF of 0.29 and heritability of 1.0. The first two were in complete LD and showed, on average, higher association with the flat trait. Most of the flat varieties (87.7%) were heterozygous at these two loci while this percentage decreased to 8% in the round varieties, consistent with the genetics of the trait. ### Texture An endoPG gene, mapping at the end of LG4, has been reported to control melting flesh (M) in peach. In our germplasm collection, 213 accessions were classified as non-melting and 746 as melting. We found two SNPs significantly associated in LG3 12 kb apart (scaffold_3:12,836,182..12,878,608) and four in LG4 covering a region of 11.6 kb (scaffold_4:23,497,381..23,508,950). The two SNPs associated in LG3 (SNP_IGA_341962; p-value 2.1E-9, and snp_3\_12878608; p-value 2.4E-8) were in high LD (r² = 0.909) and the association was due to an excess of the most frequent allele in homozygosis in the melting varieties. All four SNPs associated in LG4 (SNP_IGA_477941, SNP_IGA_477945, SNP_IGA_477951, SNP_IGA_478039) were in high LD, the first two being in complete LD (r² = 1) (SNP_IGA_477941 and SNP_IGA_477945; p-value = 5.1E-11). The association was due to an excess of the less frequent allele in homozygosis in the non-melting fruits. Four and six haplotypes were imputed using the two and four associated SNPs of chromosomes 3 and 4, respectively. Two haplotypes had a frequency higher than 5% in both chromosomes. The average heterozygosity was 10.6% in chromosome 3 and 18.9% in chromosome 4. ### Flower type (showy/non-showy) The showy/non-showy flower locus (*Sh*) maps in LG8, with the non-showy allele being dominant. Two-hundred and sixty of the accessions analyzed had non-showy flowers while 681 had showy flowers. We found 18 SNPs associated with the trait in a region of 4.3 Mb of chromosome eight (scaffold_8: 13,204,775..17,476,655). The average MAF of the associated SNPs was 0.31 and the heritability was 0.62. The most highly associated SNP was SNP_IGA_864149 (scaffold_8: 13,756,987) with a p-value of 2.73E-23, MAF of 0.36 and an r² of the model with the SNP of 0.42. The homozygous genotype of the most frequent allele of this SNP was more frequent in the showy (66.2%) than in the non-showy (10%) varieties. A total of 162 haplotypes were imputed using the 18 SNPs associated, six with a frequency higher than 5%. ### Leaf gland type The *E* locus, determining the globular or reniform shape of the leaf glands, maps in LG7. The globular leaf gland phenotype was present in 190 accessions and the reniform in 750. A single SNP (SNP_IGA_776161, scaffold_7:14,870,521) was associated to the leaf gland type. The associated SNP had an MAF = 0.39, R² = 0.10 and heritability of 0.31. # Discussion Analysis of 8,144 SNPs, represented in the IPSC 9k chip, in a diverse sample of 1,580 peach accessions revealed polymorphisms in 5,378 SNPs (66%), while 727 (9%) were false (monomorphic) and 2,037 (25%) failed. These data confirm the accuracy of the validation by Verde et al. using a subset of 96 SNPs of this array, where they found similar proportions (57%, 19% and 24%, respectively). Only 4,271 of the polymorphic SNPs were retained after exclusion of 1,107 that were suspected to contain null/preferential amplification alleles, correspond to duplicate sequences or had MAF \<0.05. Such an array of polymorphic markers allows comparison of genotypes at the whole-genome level, identifying features not possible to discern with the low density marker analysis assays used to date. One of them concerns the differences observed between some of the known and inferred groups of sports analyzed. In four cases, we observed that SNP differences among sports were concentrated at one 5.3 Mb region in the distal end of chromosome 6. These differences involved changes of marker genotype (in 1–11 SNPs) and missing data in many of the 92 markers included in our SNP array in this region (between 38 and 60 SNPs with no data). The pair of cultivars, ‘Springtime’ and ‘Starcrest’, presented a similar pattern, although in this case affecting the upper part of chromosome 4. The same region of chromosome 6 is involved in a reciprocal translocation with chromosome 8 in the red-leaved rootstocks ‘Nemared’, ‘Akame’ and ‘Rubira’. Our results suggest that this region of the genome may be particularly unstable, with major rearrangements occurring frequently. They also suggest that the phenotypic differences between an original cultivar and its sports may be caused by genes located in these unstable DNA fragments. The UPGMA dendrogram separated Oriental accessions from those cultivated in Occidental countries. Breeding accessions were also clearly separated from those cultivated locally. The kinship analysis revealed that accession pairs were related on average at the level of grandparent-grandchild or half-sibs. This was in agreement with the results from PCA and the STRUCTURE software, identifying three subpopulations within the peach germplasm analyzed: P1OC_B, containing most accessions from modern breeding programs, essentially the gene pool derived from that used by early US breeders; P2OC_T, including the old European accessions, most of them non-melting peaches that were seed- propagated for a long time and as a consequence highly homozygous; and P3OR that encompasses the majority of Oriental accessions. Previous studies with SSRs also showed that *P*. *persica* has a strong population stratification and identified the same groups. A recent study identified a different stratification in a panel of 84 wild and cultivated peach accessions: i) wild relatives (species), ii) edible and iii) ornamental accessions. However, only a few Western and breeding accessions were included in the panel. The subpopulations detected are in general consistent with the known history of peach domestication and dispersal, where P3OR represents the original Chinese gene pool from the domestication process, and the old European materials (P2OC_T) are those derived from the migration of the Chinese accessions throughout central Asia towards Western Europe. Some of these materials were taken to America by European settlers and were the founders, with a few accessions coming directly from China, of the US breeding programs that started in the early 20th century using the concepts of modern genetics. These programs produced a new wave of successful cultivars that are currently the basis of European and North-American peach production, constituting subpopulation P1OC_B. The population structure identified in this study is also consistent with the genetic bottlenecks described by Verde et al. using a set of whole genome SNPs in a panel of 12 Occidental and Oriental accessions. The first bottleneck clearly denotes the domestication event and is related to the P3OR subpopulation. The second bottleneck is related to peach dissemination in the West and to the modern breeding activities that began in the US and Europe in the last century. This bottleneck is represented by the P2OC_T and P1OC_B subpopulations. In their study Cao et al. suggested the existence of a single bottleneck related to the domestication event failing in identifying dissemination and breeding bottlenecks. However, they included only a few Occidental and breeding accessions in their panel; this lack likely hampered the detection of Western dissemination and modern breeding bottlenecks. A subdivision within this population that is also evident in the SSR studies separates peaches from nectarines, which emphasizes the fact that breeding of these two fruit types has often been separate. A large fraction of the accessions analyzed (54%) were admixed between the subpopulations identified, suggesting that breeding programs also look for useful variability crossing the boundaries of each subpopulation. The mean observed heterozygosity was H_o = 0.28, a lower value than that usually reported with SSRs: H_o = 0.34 in a collection of Occidental cultivars and H_o = 0.47 for a larger collection of Oriental and Occidental materials. This was expected considering the higher numbers of alleles per locus of SSRs. The proportion of markers with MAF\< 0.2 was low when considering all the genotypes studied in the admixed group (17.3–20.0%), and much higher for the three subpopulations (46.8% for P1OC_B, 54.2 for P2OC_T and 61.3% for P3OR). This situation can be explained considering that the SNPs were identified in resequencing data of about 50 peach accessions, mostly from the P1OC_B and the admixed groups, and by the larger size of these groups (352 and 665 accessions, respectively) compared to the other subpopulations. Population structure affects the LD throughout the genome. LD decayed with distance in all subpopulations, being faster in the ADM subpopulation. Accessions in this population contain variability from genetically distinct founding populations with different allele frequencies: thereafter they should retain the LD of the parental populations. Additionally the amount of admixture- generated LD between any two loci is highly dependent on the proportion of each population in the admixture process. Faster decay of LD in ADM could be due to the remaining subpopulation stratification in P1OC_B, P2OC_T and P3OR or to a relatively low participation of these three populations in the admixture process. Previous studies comparing LD dynamics have detected faster decay in Oriental compared to Occidental germplasm. One reason this was not observed in our sample could be the origin of the accessions of the population; here P3OR only included mostly Chinese breed varieties, and no wild germplasm or landraces. This indicates that breeding efforts in China and Occidental countries have resulted in similar LD patterns and, according to Xie et al., in a similar reduction of variability. As expected, the LD pattern was not homogeneous across the chromosome, containing areas with extensive linkage disequilibrium. These areas may correspond to recombination cold-spots and so should be observed in all populations as well as in linkage maps. Two LD blocks in LG4, one at the top and one at the bottom, and one in LG5, were observed in all populations. These regions were also observed by Verde et al., who also plotted the relation between genetic and physical distance in peach chromosomes. A low ratio between genetic and physical distance, indicating a low recombination rate, was only observed for the block at the end of LG4. Selection may explain the LD blocks of the other two regions. Artificial selection has a strong effect on LD. Mosaics of large LD blocks have been found in regions carrying agronomic-related genes. Another example of LD generated by selection detected in our study is the long LD block around the *Y* gene in P1OC_B, with most of the varieties having the recessive yellow flesh phenotype. This disappears in P3OR, where most varieties had white flesh. Association mapping exploits natural diversity to search for functional variation. This method presents several advantages over linkage mapping but usually they are used together to validate or discard spurious associations. Here we applied association mapping to validate its use in peach. For some of the traits evaluated the genes have been cloned (as is the case of the *Y*, *G* and *F* loci) or mapped in a short interval. In all cases, the SNPs found to be associated with the traits were located in the known region. For the known genes ppa006109 for the *Y* locus, ppa023143 for *G* and ppa006839 for *F*, the most associated SNPs were 838, 875 and 849 Kb upstream of the 5’-UTR of the gene, which corresponds approximately to 1.9 cM (considering the *Prunus* genome of 519 cM and 230 Mb). We found SNPs tightly linked to the traits considered, but their use in marker assisted selection (MAS) is not straightforward and still needs to be validated. For most of the traits we found haplotypes that explain a large fraction of the phenotypes, however the association is not complete. For example three causal alleles have been reported for the yellow flesh trait. All three are loss-of- function alleles from different mutation events and, consequently, have different SNPs (or haplotypes of SNPs) associated to each allele. Here we showed the considerable extent of LD in peach and that, in populations, SNPs are structured into haplotypes which may be better at predicting the phenotype than single SNPs. Our results do strongly suggest that the use of haplotypes is more powerful than single tightly-linked SNPs in MAS. However, in most cases, these haplotypes will be population-dependent, i.e. they have to be chosen based on the frequency of the haplotypes in the parental lines used in the breeding program. # Conclusions A collection of 1,580 Occidental and Oriental peach accessions was analyzed with an SNP array covering the whole peach genome. Results showed that the levels and organization of variability are consistent with the known major events in the history of this species after domestication, and generally agree with published information obtained with SSR markers and SNPs. Given the high levels of LD conservation of the peach genome, whole-genome association analysis with a relatively small number of loci (4,271) showed that SNPs significantly associated with seven major genes were always in the regions predicted by classical linkage analysis. For the genes still not cloned, our data provide sets of tightly linked markers that can be the starting point to identify candidate genes and eventually used for their cloning and molecular characterization. The results also suggest that similar analysis with characters of complex inheritance and high value, such as those related with fruit organoleptic quality and extended post harvest life, will be able to identify genomic regions containing the most important genes responsible for variation of the traits. The information has other applications for peach breeding and germplasm management, as many commercial cultivars from private and public breeding programs have been characterized genetically to an unprecedented level of detail, enabling the study of the variability within each program and comparison of the variability between breeding programs on a whole-genome basis. This information is useful for parent selection and to search for variability at specific genomic regions not present in the set of parents used in a specific breeding program, but present in the global peach gene pool. Moreover, our results reveal that, for certain characters, the MAS strategy not only requires one or several markers sufficiently close to the gene of interest, but also identification of the different haplotypes for various markers in the region, selecting those that are diagnostic for the character based on the haplotypes of the set of parents used by each breeding program. # Materials and Methods ## Plant materials and phenotyping A set of 1,580 accessions of *Prunus* were selected to be representative of five germplasm collections in four different countries. Of these accessions 1576 were *P*. *persica* and four were of other related *Prunus*. In detail, 420 accessions were from the peach germplasm collection of Centro di Ricerca per la Frutticoltura (CRA-FRU, Roma), 112 and 239 were provided by the National Institute of Agronomic Research (INRA) of Avignon (France) and Bordeaux (maintained in the Prunus Genetic Resources Center, France), respectively, 367 were from the peach germplasm collection of IRTA at the experimental stations of La Tallada d’Empordà and Gimenells (Spain), 160 were obtained from the collection of University of Milan (Italy), and 282 from the peach collection of Zhejiang University (Hangzhou, China). The complete list of the accessions is reported in. Between 50 and 100 mg of young leaves from each accession were freeze-dried and used for genomic DNA extraction with the DNeasy 96 Plant Mini Kit (Qiagen, Valencia, CA, USA) following the manufacturer’s protocol. The phenotype of a variable number of plants was available for seven monogenic traits: 1,071 for fruit pubescence (peach/nectarine), 1,060 for fruit shape (round/flat), 1,024 for fruit flesh color (yellow/white), 959 for fruit texture (melting/non-melting), 832 fruit acidity (acid/sub-acid), 940 for gland type (round/reniform), and 741 for flower type (showy/non-showy). All the traits were coded as qualitative. ## SNP genotyping 1,580 *Prunus* accessions were genotyped using the International Peach SNP Consortium (IPSC) peach 9K SNP array v1. SNP genotypes were scored with the Genotyping Module of the GenomeStudio Data Analysis software (Illumina, Inc.) using the default parameters. The SNPs were divided into five categories: A, B, C, D, and E. The first three included the polymorphic SNPs and the last two the non-polymorphic. The SNPs with GenTrain higher than 0.4 and GeneCall 10% higher than 0.2 and at least two genotypic classes were classified as polymorphic. The three classes of polymorphic SNPs were defined as follows: 1. A. SNPs with less than 5% of No Call (failed genotyping) and all three possible genotypes (AA, AB, BB) defined. 2. B. SNPs with a potential null allele or preferential annealing. Between 5 and 50% of individuals showed a normalized signal intensity value R \< 0.2 and the remaining individuals represented at least two of the three genotypic classes. 3. C. Probable presence of duplicated sequences/genes. These SNPs were characterized by the absence of one homozygous cluster with an over- representation of heterozygous individuals, and a percentage of No Call \<5%. The non-polymorphic SNPs were divided as: 1. D. False SNPs. Characterized by a single genotypic class and a percentage of No Call \< 5%. 2. E. Failed SNPs. Having more than the 50% of No Call and/or with a GenTrain \< 0.4 and/or a GeneCall 10% \< 0.2. All further analysis were performed using all Class A SNPs with a minor allele frequency (MAF) higher than 0.05. Genotypic data have been uploaded to GDR database (<http://www.rosaceae.org/>) under the accession number tfGDR1013 and in the FruitBreedomics database (<http://bioinformatics.tecnoparco.org/fruitbreedomics/>). ## Phylogenetic tree The distance matrix was calculated using TASSEL, as 1—IBS (identity by state) similarity, being IBS the probability that alleles drawn at random from two individuals at the same locus are the same. For clustering, the distance of an individual from itself was set to 0. The clustering was performed using the R package hclust with the UPGMA method. ## Population Structure and Fst PCA analysis was performed using the “prcomp” R package, freely available at <http://stat.ethz.ch/R-manual/R-patched/library/stats/html/prcomp.html>, after genotype numericalization as follows: The following approach was followed for numericalization: the higher value (2) was assigned to the heterozygous state, the lower value (0) was assigned to the less frequent homozygous state, and the intermediate state (1) was assigned to the most frequent homozygous state. Population structure was studied with the “prcomp” function of R and with the Structure v.2 software. This program uses a clustering method that identifies K subgroups of individuals with distinctive allele frequencies. Individuals can be members of multiple subpopulations with a different coefficient, with the sum of all being equal to 1. To check for population stratification, the program was run under the admixture model assumption with correlated allele frequencies. The run used 100,000 interactions after a burn-in of 10,000 for a value of K ranging from 2 to 20. To avoid bias due to tightly linked markers and to speed up the computation time, the SNPs were pruned based on LD using PLINK to give a final 1,506 SNPs. Pruning was using the “indep” function with a window size of 50 SNPs, five SNPs shift of the window at each step, and a variance inflation factor (VIF) threshold of 10 (equivalent to prune to an LD of 0.75). The VIF is 1/(1-R²) where R² is the multiple correlation coefficient for an SNP being regressed on all other SNPs simultaneously. The most probable number of populations was estimated using the method proposed by. The fixation index (Fst) was calculated as in using a custom python script. ## HWE and MAF The Hardy Weinberg equilibrium and the minor allele frequency were calculated for each SNP using PLINK. The SNPs showing severe distortion of the HWE (p\<10e-4) and/or MAF lower than 0.05 were discarded from further analyses. ## Linkage disequilibrium Given that the phases between alleles at two heterozygous loci are unknown, the linkage disequilibrium (LD) was calculated using the squared correlation based on genotypic allele counts as implemented in PLINK. Additionally, the LD was calculated using the R-package LDcorSV, able to correct the classical r² measure for population structure and relatedness. To measure LD, SNPs with a minor allele frequency (MAF) lower than 5% were discarded. The curve representing the observed r² values was computed using a polynomial fit as implemented in the “numpy” module (<http://docs.scipy.org/doc/>) of python v.2.7.3. ## Kinship The kinship matrix was computed using the VanRaden algorithm as implemented in the GAPIT R package. ## Genome wide associations Association analysis was using the compressed mixed linear model implemented in the GAPIT R package. Association tests were run using two approaches. One was a global approach where an MLM model for all individuals was run using population structure (k coefficients of ancestry, Q), and a genetic covariance matrix (kinship matrix, K) used as cofactor for SNP effects. The other was a population level approach where each sampling population was run in a separate analysis using the MLM as outlined above. Significant associations were determined using both a B–Y FDR P-value adjusted for multiple testing and a Bonferroni adjustment at the α = 0.01 level. Prior to the analyses, SNP markers were filtered for MAF of ≥ 0.05, and a minimum genotyping success of 90%. The SNPs associated to each trait were phased independently using fastPHASE. For each run of fastPhase 20 random starts and 100 iteration of the EM algorithm were used. To test the association of the inferred haplotypes with the analyzed phenotypes a chi-squared test was performed for each haplotype. The null hypothesis was that the number of accessions showing the two alternative phenotypes was the equal. # Supporting Information We acknowledge Andrea Caprera (PTP, Lodi, Italy) for his contribution to the organization and standardization of phenotypic datasets; Zhi-jun Shen (Jiangsu Academy of Agricultural Sciences), Ke Cao (Zhengzhou Fruit Research Institute), Hui-juan Jia, Xian-qiao Meng (Zhejiang University), Ming Xie (Zhejiang Academy of Agricultural Sciences) and Jian-bao Tian (Shanxi Academy of Agricultural Sciences) for collecting materials and collecting phenotypic data; Martina Lama (CRPV, Cesena, Italy) for her participation in lab work and the technical staff of the “Biologie du Fruit et Pathologie” unit of the INRA, in particular Hélène Christmann, for her technical help in DNA extraction, and Jean Claude Barbot, Hélène Christmann, Anthony Bernard and Aude Mear for their active collaboration in the phenotyping of the INRA-Bordeaux peach collection. We thank the INRA’s “Prunus Genetic Resources Center” for preserving and managing the peach collections and the Experimental Unit of INRA-Toulenne (UEA) for growing the trees. This work has been funded under the UE seventh Framework Programme, the views expressed in this work are the sole responsibility of the authors and do not necessarily reflect the views of the European Commission. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: IV LR DB BQ-T TB PL TP Z-SG FL PA MJA. Performed the experiments: DM IV MTD SM VA SF CSL LR IP EB TB PL X-WL II JC MJA. Analyzed the data: DM NN MT MTD IV LR PA MJA. Contributed reagents/materials/analysis tools: IV LR DB PL TP MT TB Z-SG L-RW R-JM II JC PA MJA. Wrote the paper: DM PA MJA. Built the bioinformatic infrastructure for data access: NN. Critically revised the manuscript: IV MTD SM LR DB IP MT BQ-T TB PL TP Z-SG FL. [^3]: Current address: Research and Innovation Centre, Fondazione Edmund Mach (FEM), San Michele all'Adige (TN), Italy [^4]: Current address: INTA, Universidad de Chile, Santiago de Chile, Chile

# Introduction Promotional activities within pharmaceutical industries were relatively high in the past few years. In 2012, the pharmaceutical industry in the USA spent more than US\$27 billion on drug promotion. In Canada, promotional activities were estimated to cost US\$30000 per physician per year. Pharmaceutical companies appear to spend much more on promotion than they do on research and development (R&D). For example, a study based on annual reports of pharmaceutical companies found that ten of the largest global pharmaceutical companies spent a total of US\$739 billion on ‘marketing and administration’ between 1996 and 2005 compared to US\$288 billion on R&D for the same period. The industry claims that the promotional activities aim to provide health care professionals with scientific and educational information. Also, surveys suggest that many physicians believe that marketing does not influence their prescribing habits or acknowledge that it may have an influence on some physicians but not on themselves. Despite these claims, there is evidence suggesting that the interaction of pharmaceutical companies with physicians may have a negative effect on their clinical practice. The last identified systematic review assessing the interaction of pharmaceutical companies with physicians was published by Spurling et al. in 2010. The population under study included both physicians in practice and residents in training. The review found high degrees of heterogeneity that may have been due to the diverse populations under study (i.e., both practicing and in-training physicians). The review by Spurling et al. could not reach definitive conclusions about the degree to which information from pharmaceutical companies decreases, increases or has no effect on the quality, cost or frequency of prescribing. Since the publication of Spurling’s review, at least eight original studies have been published. One of these was a large study of the association between physicians’ receipt of meals from industry and the rates of prescribing the promoted drug to Medicare patients. That study appears to be at lower risk of bias than the previously published studies, and thus would contribute the improving the quality of the evidence. Therefore, the objective of our study was to systematically review the association between physicians’ interactions with pharmaceutical companies and their clinical practice. # Methods ## Protocol We followed a detailed methodology that we describe in the protocol included in. The review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. ## Eligibility criteria The inclusion criteria were: - Type of study design: observational design (e.g., cohort, time series analysis, before-after design, case control, cross sectional), and experimental design (non-randomized controlled trials, and randomized controlled trials); - Type of participants: practicing physicians as defined in the primary studies - Type of exposure: targeted interaction between physicians and pharmaceutical companies, where there is direct interaction with the physician. Direct interactions could include individual invitation to a continuing medical education (CME) event; active presentation of industry-related information to the physician; or provision of gifts to the individual; - Type of control: either no interaction or a lower level of interaction. The intention was to capture studies that stratified the levels of interaction of physicians with medical representatives, e.g., according to the number of visits within a specified period of time. - Types of outcomes: ○Knowledge of physicians (e.g., accuracy of knowledge related to a specific medication); ○Attitude of physicians (e.g., perceived influence of information from pharmaceutical company on their behavior); ○Behavior of physicians (e.g., prescribing quality; prescribing quantity/frequency; reliance on pharmaceutical companies for drug information; giving drug sample to patients; submitting a formulary request for a drug made by a specific company); ○Financial outcomes (e.g., patient out of pocket expenses; prescription costs); ○Patients’ clinical outcomes. The exclusion criteria were: - Qualitative studies, ecological studies, econometric studies, editorials, letters to the editor, and non-English studies. - Studies focusing on medical students and physicians in training. - Studies assessing the relationship between attitudes and behavior - Studies assessing non-targeted interactions (e.g., journal advertisement) and research funding - Studies assessing interventions to share industry-independent drug information or interventions to reduce interactions between physicians and pharmaceutical companies. The latter has been addressed in a recent systematic review. We did not exclude studies based on date of publication. We did not exclude any study based on risk of bias. Instead, we conducted sensitivity analyses excluding studies at high risk of bias, and also took risk of bias into account when grading the quality of evidence using GRADE approach. ## Search strategy We used OVID interface to electronically search MEDLINE and EMBASE in July 2016. The search included both free text words and medical subject headings. It combined terms for physicians and pharmaceutical industry and did not use any search filter. A medical librarian assisted with designing the search strategy (see supporting file for full search strategy). In addition, we reviewed the references lists of included and relevant primary studies and literature reviews. ## Selection of studies Two reviewers screened the title and abstracts of identified citations for potential eligibility in duplicate and independently. We retrieved the full text for citations considered potentially eligible by at least one of the two reviewers. The two reviewers then screened the full texts in duplicate and independently for eligibility. The reviewers resolved disagreement by discussion or with the help of a third reviewer. They conducted calibration exercises and used a standardized and pilot tested screening form. ## Data collection Two reviewers abstracted data from eligible studies using a standardized and pilot tested screening form with detailed written instructions. When needed, disagreement was resolved with the help of a third reviewer. The data abstracted included: type of study, funding source, characteristics of the population, exposure, outcomes assessed, and statistical data. For studies including both attending physicians and residents, we attempted to contact authors for data relating to the former group. ## Assessment of risk of bias in included studies Two reviewers assessed the risk of bias in each eligible study in duplicate and independently. They resolved disagreements by discussion or with the help of a third reviewer. We used the tool suggested by the GRADE working group for assessing the risk of bias for observational studies. We calculated the risk of bias using the following criteria: - Failure to develop and apply appropriate eligibility criteria (e.g., no clear eligibility criteria, convenient sampling, under- or over-matching in case-control studies, selection of exposed and unexposed in cohort studies from different populations, and low response rate (\<60%) with no attempts to compare non-respondents to respondents). - Flawed measurement of exposure (e.g., differences in measurement of exposure such as recall bias in case- control studies, and subjective or self-reported assessment of exposure) - Flawed measurement of outcome (e.g., differential surveillance for outcome in exposed and unexposed in cohort studies, and subjective or self-reported assessment of outcome) - Failure to adequately control confounding (e.g., failure of accurate measurement of all known prognostic factors, absence of control group in a before-after study, failure to match for prognostic factors and/or adjustment in statistical analysis - Incomplete follow-up or failure to control for loss-to-follow up We graded each potential source of bias as high, low or unclear risk of bias. We used unclear when the authors did not report enough information for us to make the judgment. ## Data synthesis We calculated the kappa statistic to assess the agreement between reviewers for full text screening. We conducted a meta-analysis to pool the results across studies for the association between ‘targeted interactions with physicians’ as the exposure of interest, and ‘changes in physician prescribing behavior’ as the outcome of interest. We contacted the authors of studies that appeared to have measured the outcome and exposure of interest, but did not report data (such as odds ratio or standard error) that we could include directly. We received responses from authors of 7 out of 9 relevant studies. The authors provided us with the needed information for only two out of the seven studies. Please refer to for a summary of author contacts. We used the following a priori plan for choosing which data to include in the meta-analysis: - For studies reporting on more than one type of exposure (e.g., gifts, detailing), we treated each exposure as a separate unit of analysis. - For studies measuring the same outcome at several points in time, we chose the first time point to avoid any potential confounding effects from subsequent measures. - For studies assessing the association of interest for more than one drug (i.e., reporting more than one association), we included the value that is the closest to the mean of all reported values amongst those associations. We used the generic inverse variance technique with a random-effects model to pool the association measures across included studies that reported the needed statistical data. We carried out statistical analysis using RevMan (version 5.2). To take into account the heterogeneity introduced by the different types of exposures (i.e., gifts, detailing, and CME), we stratified the meta-analyses by type of exposure. We tested the results for homogeneity using the I2 test and considered heterogeneity present if I2 exceeded 50%. In addition, we conducted three post-hoc sensitivity analyses by respectively excluding: - Studies at high risk of bias; - Studies funded by pharmaceutical industry; - Studies measuring the outcome of interest as ‘changes in generic prescription’ or ‘formulary request’ (as these were considered indirect measures compared with the ‘changes in the prescribing of promoted drug’). Although we had planned to construct funnel plots to assess publication bias, the number of included studies in the meta-analyses was too low to allow for that. Indeed, funnel plots are encouraged for interventions that include at least 10 studies, with a substantially higher number required if significant heterogeneity is present.. We used the GRADE approach to assess the quality of the body of evidence. The GRADE methodology involves rating the initial quality of evidence for an association as high (with observational data), followed by downgrading based on five criteria (risk of bias, inconsistency, imprecision, indirectness and publication bias), and upgrading based on three criteria (large effect size, dose-response gradient, and plausible confounding). We narratively reported any additional results that we were not able to include in the meta-analysis from eligible studies (this includes studies that could have contributed data to the meta-analysis). Whenever provided, we included the p-value to denote significance of results. # Results ## Selection of studies shows the study flow. Of the 12, 400 article titles identified by the electronic literature search, 20 articles reporting on 19 studies met our inclusion criteria (two articles reported on different outcomes for the same study). A list of the excluded studies along with reasons for exclusion is provided in. The kappa statistic value for full text screening was 0.89, suggesting high levels of agreement. ## Characteristics of included studies shows the characteristics of the 19 included studies. The design for the majority of studies was cross-sectional (n = 13). Four studies reported using pre-post study designs, one was a retrospective cohort study and one was a nested case-control study. The sample sizes in these studies varied between 10 and 279, 669 with a median of 206. One study did not provide adequate information on the exact sample size. The included studies were conducted in the USA (n = 13), Australia (n = 1), Spain (1), Denmark (n = 1), Germany (n = 1) and the Netherlands (n = 2). The publication year ranged from 1972 to 2016. The specialties of the physicians in the majority of studies were primary care (including general practitioners, family medicine, internal medicine) (n = 8); obstetrics and gynecology (OB/GYN) (n = 1); dermatologists (n = 1); mix (n = 3); and unclear (n = 6). The types of exposure assessed were sales representatives’ visits or detailing (n = 10), industry-funded continuing medical education including travel funds (n = 4), and receiving free gifts (e.g. drug samples, meals, gifts in the form of office stationery, and grants and payments) (n = 11). One study assessed a mix of exposures without reporting data specific to each exposure. The types of outcomes assessed were physicians’ prescribing behaviors (n = 19), physicians’ beliefs (n = 1), physicians’ knowledge (n = 1) and financial outcomes (n = 4). Physicians’ prescribing behavior was defined as the changes in quantity or quality of prescriptions. None of the studies assessed clinical outcomes. ## Risk of bias The detailed judgments about each risk of bias item for included studies are displayed in. shows the corresponding risk of bias summary for these studies. For the majority of the studies, the risk of bias was judged to be low for ‘appropriate eligibility criteria’, ‘measurement of intervention’, and ‘measurement of outcome’, except for the ‘completeness of data’ that was judged as unclear. For ‘controlling for confounding’ the risk of bias was judged as low for nine studies, unclear for four studies, and high for six studies. ## Findings of studies provides a summary of the outcomes and the statistical results reported for each included study. Out of the 19 included studies, six reported data in a format that could be included in the meta-analysis of the association between the exposure and the behavior (i.e., reported odds ratio or risk ratio or provided raw data allowing the calculation of an odds ratio). Below, we present the results of those six studies and their meta-analysis. We then narratively report any additional results that we were unable to include in the meta-analysis from eligible studies. ## Results of the meta-analysis The study design of the six included studies were retrospective (n = 2), nested- case control (n = 1), pre-post (n = 1) and cross-sectional studies (n = 2). These studies assessed the following types of interactions (with some studies reporting on more than one type): detailing (n = 2); industry-funded continuing medical education including travel funding (n = 2); and receiving free gifts (drug samples and meals) (n = 3). Sondergaard et al. conducted a retrospective cohort study and reported a statistically significant effect of the first visit of drug representatives on the general practitioner’s drug preference favoring the marketed drug (odds ratio (OR) 2.39; 95% confidence interval (CI) 1.72–3.32). The effect on drug preference increased further after the second visit (OR = 1.51; 95% CI: 1.19–1.93), but no significant change was noted after the third visit (OR = 1.06; 95% CI: 0.94–1.20). We considered the data for the first visits only as the analysis for subsequent visits may be confounded by the effect of the previous visits as highlighted by the authors: “the effect of promotional visits could in part be caused by representatives selecting practices with a higher probability of adopting the promoted drug. Although we have controlled for the time until first visit, a selection effect cannot be excluded.” Also, none of the remaining studies included in the meta-analysis reported data for subsequent visits. Bowman et al. conducted a pre-post survey showing the effects of attendance of three continuing medical education courses, each subsidized by a single but different drug company, and the changes in rate of prescribing by physicians of course-related drugs. We excluded the results for courses I and II, given data were not matched. In course III, Diltiazem was the sponsoring company's drug. Prescribing Diltiazem most frequently to new patients statistically increased from 22.3% to 33.9% pre-post course (p\<0.05). In addition, the number of new prescriptions for Diltiazem increased statistically from 31.4% to 50.1% pre-post course (p\<0.05%). We considered the first outcome and conducted sensitivity analysis that demonstrated no important changes in results. Pinckney et al. examined in a cross-sectional study the interaction between the availability of medication samples in the clinics and the prescription preference of primary care prescribers (stated as the name of the medication) in response to two clinical vignettes. Clinicians who did not have samples in their offices were more likely to prescribe hypertension medication according to clinical practice guidelines (p\<0.01), and more likely to prescribe a depression medication that was generic (p = 0.02). Multivariable regression models were conducted only for the hypertension vignette. The findings showed that clinicians with samples were still less likely to select thiazide diuretic that is the preferred treatment for hypertension (OR = 0.15; 95% CI: 0.04–0.56). Miller et al. conducted a retrospective study to look at the association of a pre-post removal of drug sample closet and the percentage of medications prescribed to uninsured or Medicaid patients as generics. Following a logistic regression model, the authors found that the absence of the sample closet was associated with uninsured patients receiving a generic prescription (OR = 4.54; 95% CI: 1.37–15.0). In a nested case-control study, Chren el at found that physicians who had met with pharmaceutical representatives were significantly more likely to have requested that drugs manufactured by specific companies be added to the formulary, than other physicians (OR = 3.4; 95% CI 1.8–6.6). Similar increased odds of formulary requests were obtained for physicians who had accepted money from those companies to attend educational symposia (OR = 7.9; 95% CI: 1.1–55.6), or to speak at educational symposia (OR = 3.9; 95% CI: 1.2–12.7). We treated each exposure as a separate unit of analysis in the meta-analysis. Dejong et al. examined in a cross-sectional study the association between the receipt of an industry-sponsored meal promoting the drug of interest and prescribing rates of promoted drugs compared with alternatives in the same class. Physicians who received a single meal promoting the drug of interest had higher rates of prescribing Rosuvastatin over other statins (OR = 1.18; 95% CI: 1.17–1.18), Olmesartan over other ACE inhibitors and ARBs (OR = 1.52; 95% CI: 1.51–1.53), Nebivolol over other β-blockers (OR = 1.70; 95% CI: 1.69–1.72), and Desvenlafaxine over other SSRIs and SNRIs (OR = 2.18; 95% CI: 2.13–2.23). We included the value that is the closest to the mean of all reported values amongst those associations. We pooled the results for the six studies in a meta-analysis, stratified by type of exposure. Please refer to for a summary of all decisions and their rationale with respect to the statistical data included in the meta-analysis. The pooled estimate showed a statistically significant association between interaction with pharmaceutical industry and physicians’ prescribing behaviors (OR = 2.52; 95% CI: 1.82–3.50). The heterogeneity was considered high with I² of 64%. The test for subgroup effect did not identify any subgroup difference by type of exposure (Test for subgroup differences: P = 0.88). shows the risk of bias summary for these six studies. We judged the overall risk of bias in four of these studies as low. Following the GRADE methodology we downgraded the quality of evidence for the outcome ‘behavior of physician’ from high to moderate for risk of bias and inconsistency. There was no major concern with imprecision, indirectness of the evidence, or publication bias warranting further downgrading. The first sensitivity analysis excluded two studies judged as having an overall high risk of bias. The analysis found no important change in the pooled effect estimate (OR = 2.55, 95% CI: 1.75–3.71). In the second sensitivity analysis, we excluded the one study funded by pharmaceutical industry. The analysis resulted in a non-substantial increase in size of the pooled effect estimate (OR = 2.74, 95% CI: 1.78–4.23). The third sensitivity analysis excluded two studies measuring the outcome of interest as ‘changes in generic prescription’ and ‘formulary requests’, respectively. The analysis resulted in the pooled effect estimate decreasing slightly in size but remaining statistically significant (OR = 2.11; 95% CI: 1.62–2.74). The details of the analyses are provided in, and Figs, respectively.” ## Narrative summary of studies not included in the meta-analysis As noted above, fourteen eligible articles reporting thirteen studies were not included in the meta-analysis. In addition, we narratively summarized the findings of the study by Pinckney et al (included in the meta-analysis), albeit only for the outcome ‘beliefs of physicians’. The study designs were cross-sectional (n = 12), pre-post survey (n = 1) and retrospective cohort study (n = 1). These studies assessed the following types of interactions between physicians and drug representatives (with some studies reporting on more than one type): detailing (n = 8); continuing medical education including travel funding (n = 2); receiving free gifts (e.g. drug samples, meals, grants and payments) (n = 9); and a mix of interactions (n = 1; this study did not report data specific to each exposure). We summarized the results narratively, stratified by type of exposure and within each exposure by risk of bias. ### Detailing (or pharmaceutical representative visits) Nine articles reporting on eight studies evaluated the interactions between physicians and pharmaceutical sales representatives \[, –\]. All of these studies assessed prescribing behaviors. In addition to assessing prescribing behavior, one assessed financial outcome and one assessed physicians’ knowledge. Five of the eight studies found an association between detailing and increased prescribing frequency, lower prescribing quality, higher expenditure, or earlier awareness of promoted drugs. Three studies found both associations with higher prescribing frequency or expenditures and lack of significant associations, for the different types of drugs or exposures examined. The overall risk of bias was judged as high for two of the eight studies. Figueiras et al. and Caamano et al. conducted a cross-sectional study to examine the association between detailing and the amount, expenditure and quality of drug prescribed by primary care physicians. The quality of prescription was reflected via three indicators which were combined to produce a global indicator variable. Figueiras et al. showed that using information obtained from pharmaceutical representatives was associated with a higher percentage of prescription drug not included in the primary care formulary and with a higher “global indicator variable”, thus reflecting lower prescription quality. Camaano, et al., found that the utilization of the visiting marketers’ information was significantly associated with higher amounts of prescription (p = 0.048) and higher expenditure per physician (p = 0.035). On the contrary, the number of sales representative visits was not statistically associated with prescription amount or expenditure. Pedan et al. conducted a retrospective cohort study to assess the impact of detailing on new prescriptions for three different statin brands, namely, Lipitor, Crestor and Vytorin. The authors provided data for promotional activities of ‘own brand’ and the promotion of ‘competitive brand’. The findings for ‘own promotion’ indicated that detailing produced a highly significant positive impact on new prescriptions for Lipitor and Crestor (p\<0.05), while results were not significant for Vytorin. ‘Competitive’ detailing had a significant negative impact on new prescription for Vytorin (p\<0.05) whereas results were not significant for the other two brands. Muijrers et al. conducted a cross-sectional study and found that more frequent visits from pharmaceutical industry representatives had a significant negative correlation with the quality of prescribing of general practitioners measured as “adherence to guidelines” (p\<0.05). The latter was calculated as a weighted average score on 20 prescribing indicators based on general practice guidelines of the Dutch College of General Practitioners. Mizik et al. conducted a pooled time series cross-sectional study to evaluate the association between pharmaceutical sales representative visits per month and the number of new prescriptions issued by a physician per month for three studied drugs. The average number of new prescription per month by drug, were: 1.56 (95% CI: 0.80–2.23) for drug A; 0.32 (95% CI: 0.22–0.43) for drug B; and 0.15 (95% CI: 0.11–0.20) for drug C. Becker et al. conducted a cross-sectional study and found that the use of detail men as sources of prescribing information concerning new drugs was significantly associated with higher prescription of the drug chloramphenicol by primary care physicians (p\<0.01), and a poorer rating of prescription quality (p\<0.01) by two panels of experts relative to five common complaints and five common illnesses. The findings of the remaining two studies with overall high risk of bias are discussed below. Peay et al. conducted a cross-sectional study and found that doctors (specialists and general practitioners) who had contact with detail men regarding the promoted drug Temazepam reported earlier awareness of it (p \< 0.041), were more likely to rate it as a moderate advance (as opposed to a minor advance or no advance at all) (p\< 0.028), were more likely to have prescribed it (p\< 0.0031), reported prescribing it earlier (p\< 0.005) and were more likely to prescribe it routinely in preference to alternatives (p\< 0.014). In the multivariate analysis, contact with detail man regarding Temazepam and first news of drugs from detail man remained significantly related to four of the five dependent variables correlating with prescription of Temazepam. In another cross-sectional study, Lieb et al. found that frequently visited practices (i.e. daily or 2–3 times per week) had a significantly higher number of prescriptions (p = 0.005) and total daily doses per patient (p = 0.003) compared to practices visited less frequently by representatives. However, they did not have a higher total expenditure per patient (p = 0.115). The effects of the “frequency” of pharmaceutical sales representatives visits on prescribing behavior were no longer significant after taking into consideration the number of patients per office. ### Continuing medical education Two studies assessed the effects of industry-funded continuing medical education on physician prescribing behaviors and expenditure on off-patent branded drugs. Both studies found significant associations between attending industry-funded continuing medical education and higher prescribing frequency, lower prescribing quality, or increased prescription cost. The overall risk of bias was judged as high for one of them. A pre-post study tracked the prescribing pattern for two drugs before and after physicians attended symposia sponsored by a pharmaceutical company. The ‘expense-paid seminar at a resort’ was associated with a significant increase in the prescribing of the two promoted drugs within a few months of each symposium compared to their use before the symposium (p\<0.001). The cross-sectional study with high risk of bias found that compared to doctors who frequently, occasionally or rarely took part in sponsored CME events, doctors who mentioned that they never took part in such events had a lower number of on patent-branded drug prescriptions per patient (mean ± SD; 1.05±0.35 vs. 1.27±0.55; p = 0.005, a higher proportion of generics (83.28±7.77% vs. 76.34±13.58%; p\<0.0005) and lower expenditure on off-patent branded drugs per patient (£27.36±23.23 vs. £43.75±43.22; p = 0.002). ### Receiving free gifts Nine studies evaluated receiving free gifts as the exposure of interest. All of these studies assessed prescribing behaviors. In addition to assessing prescribing behaviors, two assessed financial outcomes and one assessed physicians’ beliefs. For the latter study, we did not include the findings pertaining to prescribing behavior as these were already included in the meta- analysis. Five of the nine studies found an association between receiving free gifts and increased prescribing frequency, lower prescribing quality, or increased prescription cost. Each of three studies found both associations with higher prescribing frequency and lack of significant associations, for the different types of gifts examined in the studies. The remaining study found that prescribers with access to samples were significantly more likely to believe that samples benefited patients. The overall risk of bias was judged as high for four of the nine studies. Pedan et al. conducted a retrospective cohort study to evaluate the effects of sample dispensing and meals on total monthly number of new prescriptions written by a physician for the three leading statin brands (Crestor, Lipitor and Vytorin). The authors provided data for promotional activities of ‘own brand’ as well as the promotion of ‘competitive brand’. The findings for ‘own promotion’ showed that sample dispensing had a significant positive effect on new prescription for Crestor (p\<0.01) and Vytorin (p\<0.05), but results were not significant for Lipitor. They also found that free meals had a significant positive impact on new prescription for all three statin brands: Lipitor (p\<0.05), Crestor (p\<0.05) and Vytorin (p\<0.01). ‘Competitive’ sampling had a significant negative impact on new prescription for Lipitor (p\<0.05) and Vytorin (p\<0.05) but results were not significant for Crestor. ‘Competitive’ free meal-related promotions had a significant negative impact on Lipitor only (p\<0.05) The authors concluded that, while on average the marketing efforts affect the brand share positively, the magnitude of the effects is very brand specific. Mizik et al. conducted a pooled time series cross-sectional study to examine the association between receiving free sample medications and the number of new prescriptions issued by a physician per month for three drugs. They observed statistically significant but small effects of sample dispensing on prescription behavior for all three types of drugs. In a cross-sectional study, Symm et al. examined the prescription claims data for 25 medications in one “clinic X” where sample medications were dispensed compared to two clinics, Y and Z, which did not dispense free sample medications. They showed that, first, family physicians in clinic X significantly wrote the largest proportion of prescriptions for study medications (p \<0.0001) versus non-study medications. Second, family physicians in clinic X significantly prescribed the lowest proportion of preferred name brands among study medications (p \<0.0001). Third, the average cost of a 30-day prescription differed significantly by clinic (p \<0.0001), being the highest in clinic X. Hurley et al. conducted a cross-sectional study and found a strong correlation (r = 0.92) between the increase in provision of samples with a prescription by dermatologists and increased use of branded generic drugs promoted by these samples. For physicians at local academic centers where free samples are prohibited, only 17% (230 of 1364) of the commonly prescribed medications were for branded or branded generic drugs compared to 79% for office-based dermatologists on a national level where free samples are available. Additionally, the national mean total retail cost of prescriptions was “conservatively” estimated to be twice as higher (roughly \$465 nationally versus \$200 at an academic medical center where samples were prohibited). Yeh et al. found that among physicians with industry payments in the Massachusetts database, every \$1000 in total payments received was associated with a 0.1% increase in the rate of brand-name statin drug prescribing (95% CI: 0.06%-0.13%; P \<.001). Receiving payment for educational training was associated with an average 4.8% increase in prescribing of brand-name drugs (95% CI: 1.55–7.95; p = 0.004), but the other types of payment (i.e. food, grants/education gifts, and bona fide services) were not. The findings of the four studies with overall high risk of biases are summarized below. Peay et al. examined in a cross-sectional study the association between receiving a sample of Temazepam and physicians’ prescription rate and preference for it over the alternatives. They demonstrated that physicians (specialists and general practitioners) who had received a sample of Temazepam, compared to those who had not, were more likely to have prescribed it (p \< 0.001) and more likely to say that they now usually prescribe it rather than the alternatives (p \< 0.006). Pinckney et al. examined the association between the presence of samples in offices and prescribers’ beliefs about the use of samples. They found that prescribers with samples were significantly more likely to believe that samples: are liked by patients; expedite treatment; help patients who cannot afford their medication; reduce patient costs; and help physicians assess the efficacy of medications. Nonetheless, most prescribers with samples still agreed that samples alter treatment plans and increase the costs of care. In another cross-sectional study, Lieb et al. found that physicians who always or frequently accepted gifts in the form of office stationery prescribed higher daily dose totals per patient (mean ± SD; 491.97±158.95 vs. 420.53±140.57; p = 0.003) and more generics (mean ±SD; 385.52±147.52 vs. 319.43±133.69; p = 0.004) in comparison to physicians who only occasionally, rarely or never accepted stationery. We did not include the other types of gifts as the categorization of answer options was not conducive to interpretation (e.g., the categories ‘rarely’ ‘frequently’ and ‘occasionally’ were grouped together as exposure group and the category ‘never’ as control group for one type of gift whereas for another type of gift, the categories ‘frequently’ ‘occasionally’ ‘rarely’ and ‘never’ were categorized together as control group). Anderson et al. conducted a cross-sectional study and found that the frequency of eating industry-funded food was associated with greater reliance of obstetrics and gynecology physicians on pharmaceutical representatives for drug information when prescribing new medications. This finding was statistically significant in a first regression analysis (β = 0.16, 95% CI: 0.02–0.31). It became non-significant in a second regression model including as an independent variable “the perceived value of pharmaceutical representatives in helping physicians to learn about new drugs” (β = 0.07, 95% CI: -0.06–0.19). ### Mixed exposures One cross-sectional study with overall high risk of bias examined the association between rational prescribing of general practitioners and a mix of exposures including “manufacturer’s representatives, drug companies’ mailings, use of samples, drug companies’ journals, drug firm meetings and usefulness of drug company information”. The investigators found that reliance on information from pharmaceutical industry was negatively associated with prescribing rationality (p\<0.001). # Discussion ## Summary and interpretation of findings We identified 19 studies looking the association between physicians’ interactions with pharmaceutical companies and their prescribing behaviors. In addition to assessing prescribing behavior, four of the 19 studies assessed financial outcomes, one assessed physicians’ knowledge, and one assessed their beliefs. None of the included studies assessed clinical outcomes. Out of the 19 studies, 15 found a consistent association between interactions promoting a medication, and inappropriately increased prescribing rates, lower prescribing quality, and/or increased prescription costs. Each of three studies found both associations with higher prescribing frequency and lack of significant associations, for the different types of exposures and drugs examined in the studies. Only one study, albeit at high risk of bias, found an association between receiving gifts from the industry and increased generic prescribing whereas the results were mixed for the remaining types of exposures and outcomes assessed in that study. A meta-analysis of six of the included studies provided moderate quality evidence showing more than doubling of inappropriate prescribing rate among practicing physicians. The heterogeneity was considered high. Sensitivity analyses excluding the two studies at high risk of bias and the industry-funded study, respectively, did not substantially change these results. A subgroup analysis did not find a difference by type of exposure. ## Strengths and limitations of the review A major strength of the present study is the use of Cochrane methodology for conducting the systematic review. In addition, this is the first systematic review to focus on practicing physicians, as other reviews included residents and physicians in training as their population of interest. A potential limitation of our review is that we searched only two electronic databases. However, we believe our search was sensitive. Also, Spurling et al. did not include a study that was eligible for our review, and that was missed by our search strategy. Another potential limitation is the exclusion of studies published in a language other than English. Also, all included studies were conducted in developed, high-income countries; therefore, there is a chance that we have missed studies conducted in low or middle-income countries that have been published in a non-English language. However, given the consistency of findings, we expect them to be generalizable to those countries. Another limitation relates to the potential misattribution of the exposure category; however, such cases were few. One example is the study conducted by Yeh et al. which considered “payments for CME” as different from “payments for educational training”. Other limitations relate to the observational nature of all included studies and to our inability to include some studies in the meta-analysis because they did not report data for the association between the exposure and the outcome of interest. ## Comparison to findings of similar reviews The latest systematic review addressing the same topic was published in 2010 and included physicians as well as residents. The chosen populations may have contributed to the high level of statistical heterogeneity reported (I² = 91%). Also, the previous systematic review focused on exposure to information directly provided by pharmaceutical companies, and thus excluded other types of exposures such as gifts, samples, and continuing medical education courses that were funded by unrestricted grants from pharmaceutical companies. On the other hand, our review has included these types of interaction and has covered 8 years of literature since 2008, the year of literature search of the 2010 review. Still, our review included a relatively lower number of studies compared to the previous review due to the stricter eligibility criteria (e.g., we excluded residents and medical trainees; non-targeted interactions such as advertisements in journals or prescribing software, or mailed information; and participation in sponsored clinical trials). Nevertheless, our findings are consistent with those of the previous review in terms of decreased prescribing quality and increased costs associated with exposure to information provided directly by pharmaceutical companies. In addition, we found evidence of similar effects associated with drug samples and industry gifts (both of which were not assessed in the previous review). ## Implications for policy and practice Our findings suggest that the interaction between physicians and pharmaceutical companies should be better managed to reduce any negative effects and promote appropriate drug prescribing. While many policy options have been proposed to manage the interactions between physicians and pharmaceutical companies, the level of evidence supporting them varies, as described below. There is an increasing trend of mandating pharmaceutical companies to disclose payments to physicians. For instance, the Physician Payments Sunshine Act enacted in the USA in 2010 requires pharmaceutical and medical device manufacturers to publicly disclose payments (or transfers of values) exceeding US\$10 per instance or US\$100 per year made to physicians and teaching hospitals. Similarly, Medicines Australia recently revised its code of conduct, requiring pharmaceutical companies to report on payments to individual health professionals for their services, sponsorships to attend educational events, as well as educational grants. However, so far there is no evidence supporting the effectiveness of mandatory disclosure. As an example of regulatory approach, France introduced in 2004 the French Sales Visit Charter which requires sales representatives to provide physicians with “approved product information”. A report by French National Authority for Health pointed to the ineffectiveness of the Charter and the difficulty in supervising the content of verbal information conveyed during representative visits. A potentially effective option would be to restrict physician-industry interactions, particularly given the evidence that restriction policies may have a positive effect on improving prescribing behavior. This could be achieved, at the institutional level by restricting free samples, promotional material, and meetings with pharmaceutical company representatives. For example, Stanford University banned pharmaceutical sales representatives from its hospitals. Similarly, Memorial Sloan Kettering Cancer Center and Brody School of Medicine at East Carolina University banned all industry involvement (including funding) in CME, reportedly with success. At a higher level, policy makers may consider legislation specifying the types of interactions that are permissible, and those that are not. For example, Minnesota enacted a law that bans pharmaceutical industries from providing gifts to physicians with a total annual combined retail value above US\$50. While some may claim that restricting interactions between physicians and pharmaceutical companies could create an ‘information gap’, several studies conducted in different contexts found that sales representatives often did not state the risks and harmful effects of drugs to physicians. Academic detailing has emerged as an effective alternative to industry-dependent drug information. For instance, Canada and some USA states have established nationally-funded academic detailing programs that rely on similar sales tactics utilized by the pharmaceutical industry to influence physician prescribing according to evidence-based guidelines. Considerations should also be given to educating health care providers about the influence of interactions with the pharmaceutical industry, as well as inclusion of courses on industry marketing techniques and conflict of interest in medical curricula. Existing evidence suggests positive effects of educational programs about industry marketing strategies on medical trainees’ attitudes and behaviors . ## Implications for research Given all of the included studies were conducted in high-income countries, future studies should explore the effect of interaction of physicians in low and middle-income countries with pharmaceutical companies on their clinical practices. Also, it would be important to understand the impact of these interactions on clinical outcomes. # Supporting information We would like to thank Ms. Aida Farha for her valuable help in designing the search strategy as well as the Alliance for Health Policy and Systems Research for supporting the work of our group on systematic reviews related to health policy and health systems. We would also like to thank Andrea Darzi (MD, MPH) for editing the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** EAA. **Data curation:** HB RF LA LAK HN. **Formal analysis:** EAA HB RF. **Methodology:** EAA HB RF. **Project administration:** HB RF FEJ LA LAK. **Resources:** EAA. **Supervision:** EAA. **Validation:** EAA HB RF. **Writing – original draft:** EAA HB RF. **Writing – review & editing:** EAA HB RF FEJ LA LAK HN. [^3]: ‡ These authors are first authors on this work.

# Introduction Deforestation within the tropics is one of the primary threats to global biodiversity. In addition to forest reduction, fragmentation results in extended edges that are often considered entirely distinct ecosystems from forest interiors. Though fragments may persist after deforestation, most are unsuitable habitat for forest species. Within Madagascar, more than 80% of forest areas exist less than 1 km from an edge, thus fragmentation is of great concern for the survival of forest fauna and flora species. Decreasing deforestation rates and reforesting fragmented landscapes would help prevent extinctions. Generally, corridors are defined as thin strips of habitat that connect two or more isolated forest fragments, and many studies validate their utilization by organisms (reviewed in). Although fragmentation of populations may result in genetic erosion and increase extinction risk, it has been shown that a mosaic of small, suitable habitat fragments may mitigate these negative effects by acting as a single large habitat if the fragments are linked via corridors. A network of forest fragments linked by corridors allowing species to disperse may act as a means to maintain biodiversity and ecological processes in anthropogenic landscapes. The primates of Madagascar are the most threatened mammalian taxon in the world. Their survival is continually jeopardized by hunting for bushmeat, as well as habitat loss due to shifting cultivation and timber harvest. As habitat destruction continues to isolate the remaining lemurs in forest fragments, the need for regenerating forests and connecting those remaining fragments is crucial. As such, it is imperative to understand the responses of native plants and animals to disturbance if we are to create effective buffer zones and corridors that combine secondary and natural habitats. It is often thought that invasions by exotic species present a critical hindrance to the preservation of endemic biodiversity, as well as ecosystem restoration efforts. In southeast Madagascar, the Mandena littoral forest matrix exists within a seasonally-inundated flood plain that consists of natural littoral swamp and *mahampy* (*Lepironia mucronata*) wetlands, a portion of which remains inundated throughout the year. It is here, and in similarly inundated areas, that the broad-leaved paperbark tree *Melaleuca quinquenervia* (Family Myrtaceae), native to Australia, has been an aggressive disperser. These littoral forests and their animal communities are among the most threatened ecosystems in Madagascar. While the viability of non-native tree plantations has been examined to potentially assist in dispersal and fulfilment of partial habitat requirements for the conservation of lemurs, the role of introduced and possibly invasive tree species has only been minimally examined. This is of considerable interest as littoral forest fragments represent critical refuges for the survival and maintenance of biodiversity in the extremes of climatic variability. Riparian habitats often serve as corridors for multiple taxa; therefore, the conservation of these areas and nearby isolated forest blocks is critical to maintaining resilience. In view of this conflicting situation, we sought to understand the role of an invasive species habitat (*Melaleuca* swamp) on the behavioral ecology of a small-bodied folivore, the southern bamboo lemur (*Hapalemur meridionalis*) in littoral forest fragments of extreme southeastern Madagascar. Growing knowledge of the ecological flexibility of bamboo lemurs makes this species an excellent model with which to examine its ability to utilize distinct habitats, and potentially corridors, within the anthropogenic landscape. Here, we investigated whether *Melaleuca* swamp facilitates movement of *H*. *meridionalis* between littoral forest fragments and/or natural littoral swamp, and whether this invasive habitat provides additional services, e.g., suitable feeding and resting locations. # Materials and Methods ## Ethics Statement This study was conducted under the Accord de Collaboration between the University of Antananarivo and the University of Hamburg. Research protocols were approved and permits authorized by Commission Tripartite of the Direction des Eaux et Forêts de Madagascar (Autorisation de recherché n.240/12/MEF/SG/DGF/DCB.SAP/SCB du 17/09/2012), adhering to the legal requirements of Madagascar. We captured adults via Telinject® blow darts (administered by an experienced Malagasy technician) containing a hypnotic anesthesia (4 mg/kg of ketamine hydrochloride or tiletamine hydrochloride), so that the animals neither suffered nor recalled the capturing process. All animals recovered from anesthesia within 1.5 hours at the capture site, and there were no injuries as a consequence of capture and animals were followed until regaining full mobility. This process was repeated at the end of the study in December 2013 to remove the radio-collars from the bamboo lemurs. ## Study Site and Subjects Our study was conducted in the Mandena Conservation Zone (24°95’S, 46°99’E) in southeast Madagascar, a protected area approximately 10 km north of Fort-Dauphin (Tolagnaro). This area consists of 148 ha of fragmented and degraded littoral forest, which is characterized as occurring within 3 km of the coast and growing on sandy substrates with a typically low canopy, and approximately 82 ha of interspersed natural littoral swamp and invasive *Melaleuca* swamp that separates the two littoral forest fragments. During our study period, temperature (°C) was recorded in 30-mins intervals using Lascar EL-USB-1 data loggers, operated by custom software (EasyLog USB Version 5.45, Lascar Electronics). Precipitation (mm) was measured daily at 6:00h using a rain gauge placed within the study site. Day length (a proxy for season) was calculated as the time between sunrise and sunset, as obtained from the US Naval Observatory Astronomical Calendar (<http://aa.usno.navy.mil/data>), using geographic coordinates for Mandena. Southern bamboo lemurs (*Hapalemur meridionalis*) are relatively small-bodied folivorous primates with a mean body mass of 1.072 ± 0.107 kg (*X* ± SD; *N* = 15) that maintain a cathemeral activity pattern. This species lives in small social groups with one or two breeding females and typically one breeding male. Within Mandena, *H*. *meridionalis* groups average 5.6 ± 1.5 individuals (*X* ± SD; *N* = 5). In addition to southern bamboo lemur, the cathemeral collared brown lemur (*Eulemur collaris*) and nocturnal gray mouse lemur (*Microcebus murinus*), eastern fat-tailed dwarf lemur (*Cheirogaleus medius*), greater dwarf lemur (*C*. *major*), and southern woolly lemur (*Avahi meridionalis*) are present within Mandena. Ten adult *H*. *meridionalis* across four neighboring social groups were captured and habituated between October and December 2012. Data were recorded from January to December 2013. As bamboo lemurs are highly cryptic, individuals were fitted with external radio-transmitters with an archival tag (ARC400, Advanced Telemetry Systems, Isanti, USA) that allowed us to more easily follow groups. ## Habitat Characterization To characterize each distinct habitat, we sampled 25 x 100 m² botanical plots, i.e., 10 in both the littoral forest and littoral swamp, and five in the *Melaleuca* swamp, the latter requiring fewer plots due to its floristic homogeneity. Within each plot we included all trees with a diameter at breast height (DBH) ≥ 5 cm, recording the scientific binomial name and family name of each so as to measure tree diversity, in addition to their height (m) and crown volume (m³). The latter was estimated as an ellipsoid via the crown height and two crown diameters, i.e., maximum and perpendicular widths. We further conducted vertical-line transects within each plot, so as to detail the structure and canopy cover for each these three habitats. Lastly, we calculated the Shannon index (*H′*) to determine the species diversity of each habitat. The Mandena littoral forest and littoral swamp that our focal *H*. *meridionalis* groups inhabit are legally protected forests; however, much of the *Melaleuca* swamp falls outside of this demarcation. As such, local people access these unprotected areas daily to harvest wood. To measure the degree to which this occurs, we included felled trees (via tree stumps) in our botanical plots. Lastly, it should be noted that in order for lemurs to access the *Melaleuca* swamp around Mandena, they must descend and traverse a barren, sandy area that would make them visually conspicuous to any potential predators as they leave the canopy cover of the littoral forest. To examine these crossing sites, we measured the distance (m) traversed where lemurs accessed the *Melaleuca* swamp. ## Behavioral sampling From January to December 2013, we conducted full-day focal follows (sunrise to sunset) with the aim of acquiring 50hrs/month per group for three social groups. We identified individuals using radio-tracking tags with unique-colored pendants. We collected behavioral data via instantaneous focal sampling at 5-min intervals on broad-level activities (resting, feeding, moving, social, and other) and noted the habitat (littoral forest, littoral swamp, and *Melaleuca* swamp). In addition, we collected continuous feeding data each time a focal individual fed, recording the specific food item of the species, and duration of consumption measured to the second. All adult individuals in each group were sampled at least once each month. We further noted each occurrence in which the focal animal utilized the *Melaleuca* swamp corridors connecting the littoral forest fragments. ## GIS analysis We recorded a focal animal’s GPS location in 15-min intervals using a Garmin GPSMAP 62S unit, and noted the specific habitat type. All ranging data were entered into ArcGIS 10.2 (ESRI) using the Geospatial Modelling Environment (GME) spatial ecology interface with R statistical software version 3.1.2. We determined each group’s territory using a 95% kernel density estimate and further estimated the area (ha) of each habitat type. ## Statistical analyses To determine whether the characterization metrics of habitats differed, we used Kruskal-Wallis analyses for tree DBH, height, and crown volumes. We performed non-parametric tests as the data were not normally distributed, even after transformations. To determine the influence of habitat on bamboo lemur activities, a two-way repeated measures ANOVA was performed for each habitat, assessing the monthly proportion of broad-level activities (limited to rest, feed, and travel). Each habitat (littoral forest, littoral swamp, and *Melaleuca* swamp) was treated as the within-subjects factor, with groups acting as the between-subjects factor. Additionally, abiotic factors of total precipitation (mm), mean temperature (°C), and mean day length (h) per month were included in the model as covariates. The model errors for the repeated- measures ANOVA (via unstandardized residuals) were found to be normally distributed using the Kolmogorov-Smirnov test, allowing for the continuation of parametric analyses. Adjusted *p*-values are reported according to the Huynh–Feldt correction when assumptions of sphericity were violated; uncorrected biases from lack of sphericity can otherwise inflate *F*-statistics. All analyses were performed using PASW v. 21.0 and significance was set at *p* \< 0.05. # Results ## Habitats Compared to the botanically diverse littoral forest and littoral swamp, the *Melaleuca* swamp comprised six tree species, each from a distinct family. While 90.02% was *M*. *quinquenervia*, the remainder comprised native *Typhondorum lindleyanum* (7.52%), *Pandanus platyphylus* (2.09%), *Barringtonia racemosa* (0.27%), *Ravenala madagascariensis* (0.05%), and exotic *Acacia mangium* (0.05%). Tree analyses found that the three variables were significantly different between habitats (DBH (cm): Kruskal-Wallis *H* = 363.70, *df* = 2, *p* \< 0.001; height (m): Kruskal-Wallis *H* = 195.43, *df* = 2, *p* \< 0.001; crown volume (m³): Kruskal-Wallis *H* = 350.33, *df* = 2, *p* \< 0.001). As a demonstration of human impact on the *Melaleuca* habitat, we recorded 65 *M*. *quinquenervia* with a mean DBH (*X* ± SD) of 12.85 ± 8.89 cm felled within our five *Melaleuca* swamp botanical plots between January and December 2013. In addition to timber harvesting and significantly different tree metrics, habitats were further distinguished by their vertical structure. The mean distance of the eight confirmed crossing sites that *Hapalemur* groups utilized in order to access the *Melaleuca* habitat from adjacent littoral forest is 9.75 ± 2.71 m (*X* ± SD). ## Spatial analysis The total area (ha) of both home ranges utilized by groups 1 and 2 were even in size, while the home range of group 4 was substantially smaller. The *Melaleuca* swamp habitat constituted large portions of the home ranges of groups 1 and 4, while it appeared to be minimal for group 2. ## Activity and Habitat We observed *H*. *meridionalis* for 1,762 hours between January and December 2013 across 194 focal days. Groups differed in the proportion of time (i.e., percentage) spent resting in each habitat (Forest = 27.99 ± 2.21; Swamp = 7.23 ± 1.35; *Melaleuca* = 4.76 ± 1.02 (*X* ± SE; *N* = 36 months)). There were no differences in overall rates of resting between habitat types. Significant interactions were found between temperature and habitat, day length and habitat, as well as group and habitat. Considering the covariates, resting is significantly affected by temperature, but not affected by seasons (i.e., day length). Post-hoc analyses of groups revealed a significant difference between groups 1 and 4 (*p* = 0.004), while groups 1 & 2, and 2 & 4 were similar in the proportion of time and location they chose to rest. When considering feeding activity, there was no appreciable difference in the mean proportion of time each group fed; however, there were significant differences in the average proportion of feeding between the habitats (Forest = 25.52 ± 2.75; Swamp = 5.72 ± 1.09; *Melaleuca* = 11.35 ± 2.31 (*X* ± SE; *N* = 36)). Significant interactions were revealed between temperature and habitat, day length and habitat, and group and habitat. Furthermore, feeding activity is affected by both temperature and day length, varying seasonally. Post-hoc analyses showed no discernible effect of feeding between groups. Traveling showed no differences in means between the groups, while the main effect of habitat was revealed to have no influence (Forest = 6.69 ± 0.51; Swamp = 2.02 ± 0.37; *Melaleuca* = 1.41 ± 0.21 (*X* ± SE; *N* = 36)). There were significant interaction effects between precipitation and habitat, and group and habitat. Considering the covariates, traveling is affected by both temperature and day length, varying seasonally. Post-hoc analyses showed no discernible effect of feeding between groups. ## *Melaleuca* habitat use Considering individual focal days, *H*. *meridionalis* were observed to access *Melaleuca* habitat on 54.12% of days, although this only constituted 18.55% of our total observation record. Despite this, both groups 1 and 4 accessed this invasive habitat often, while the minimal proportion of *Melaleuca* within the territory of group 2 was still utilized on greater than 20% of observation days. In terms of monthly percentage of time, however, group 2 utilized *Melaleuca* less compared to the other lemur groups. Two of the three bamboo lemur focal groups fed regularly on the flowers of this invasive species when available. Group 1 was observed to feed on the flowers of *M*. *quinquenervia* for 110.65 mins, constituting 0.79% of the annual diet. While group 2 never fed on *M*. *quinquenervia* flowers, group 4 spent 2.43% of their annual total feeding record (316.32 mins) selecting for them. # Discussion Our results show that *H*. *meridionalis* use the introduced stands of *M*. *quinquenervia* substantially. *Melaleuca* and swamp habitats were often inundated by water during the warm/wet austral summer, which may restrict lemur use of these habitats to the cooler/drier months. However, examination of the monthly use of *Melaleuca* habitat by each group shows that while they spend less time here in the warmer months, they are capable of accessing this habitat when inundated. In fact, it is during this inundated period (Oct-Apr) when *M*. *quinquenervia* flowers in short, frequent bursts, but availability of this food item does not appear to influence the proportion of time bamboo lemurs spend in this habitat. This is especially true of groups 1 and 4, which spent considerable time feeding on these flowers when available, something that collared brown lemurs (*E*. *collaris*), eastern fat-tailed dwarf lemurs (*C*. *medius*), and gray mouse lemurs (*M*. *murinus*) have also been observed to exploit. There were larger proportional areas of *Melaleuca* habitat in the territories of groups 1 and 4, thus they spent more time resting, feeding, and travelling in this habitat compared to group 2. Furthermore, these social groups were occasionally found before sunrise sleeping in a *Melaleuca* tree, typically huddled together at an approximate height of 7m. The overall difference in time- budget between the groups (when controlling for the effect of habitat and the covariates) was similar for feeding or travelling activity categories, but displayed appreciable differences for resting. Precipitation was not influential, except in the case of travelling. Additionally, we observed *E*. *collaris*, *M*. *murinus*, and southern woolly lemurs (*A*. *meridionalis*) travelling and sleeping in the *Melaleuca* habitat (see also). Indigenous and/or exotic tree species can provide benefits to both local people and primates; the presence of *Melaleuca* in Mandena has value as habitat and as timber. Local people have begun to harvest these trees daily with the recent legal protection status of the Mandena littoral forest. Gaps in the *Melaleuca* canopy would allow for continued growth of terrestrial swamp vegetation, specifically graminoid species, which constitute a large portion of the *H*. *meridionalis* diet. *Melaleuca* may have value as a temporary, fast-acting solution to connecting fragments while more long-term conservation solutions are being put in place, e.g., the Mandena nursery/reforestation efforts. In the case of Mandena, the exotic *Melaleuca* acts similarly to a plantation forest for native fauna; while not as ideal as natural littoral forest, it provides valuable habitat and may possibly contribute to the conservation of endemic fauna. Many studies from various countries, including Madagascar, have documented that exotic plantation forests can provide habitat for numerous native forest fauna. As an example, threatened bird species such as *Apteryx mantelli*, *Casuarius casuarius*, and *Upupa epops* have been known to occur in substantial populations in some exotic plantation habitats (but see). Furthermore, primates such as black howler monkeys (*Alouatta pigra*) have been reported to thrive in *Eucalyptus* spp. plantations, mantled howler monkeys (*A*. *palliata*) are able to use shade-grown coffee (*Coffea arabica*) as the core of their habitat range, while siamang (*Hylobates syndactylus*) are known to occur in rubber (*Hevea brasiliensis*) and dammar gum (*Shorea javanica*) tree plantations. As for Malagasy primates, many lemur genera (including *Eulemur*, *Hapalemur*, *Indri*, *Cheirogaleus*, *Microcebus*, and *Lepilemur*) are known to use old growth eucalypt (*Eucalyptus* sp.) plantations, while some occasionally utilize and feed from mono-stands of invasive guava (*Psidium* spp.). Bamboo lemurs (both *Hapalemur* spp. and *Prolemur simus*) similarly use an exotic species habitat and appear to be relatively adaptable within anthropogenic landscapes: they have been seen crop-raiding agricultural fields at some sites and even living in a coffee plantation (but see). That the lemurs utilized invasive *Melaleuca* for behavioral activities demonstrates its potential role as a riparian corridor to facilitate dispersal. From October 2012 to December 2013 we confirmed three separate *Hapalemur* dispersals that utilized *Melaleuca* corridors to emigrate from their natal group, while a fourth dispersal remains unconfirmed. While our data indicate that *H*. *meridionalis* are tolerant of habitat degradation and fragmentation, habitat matrix composition and connectivity have been shown to influence dispersal in various birds and mammals, e.g., hazel grouse *Bonasa bonasia*, barred antshrikes *Thamnophilus doliatus*, Angola black-and-white colobus *Colobus angolensis palliatus*, and various marsupials. Furthermore, exotic tree plantations/forests have been demonstrated to facilitate dispersal for a wide range of taxa, for example, dispersal of the chucao tapaculo (*Scelorchilus rubecula*) is facilitated by the vertical structure rather than plant species composition of the corridor, in this case shrub fields dominated by 1–2 m tall invasive *Baccharis magellanica*. While instances of successful dispersal provide a glimmer of hope, the further fragmentation of remaining forests is of great concern if forest species of Madagascar are to persist. Lemurs fulfill important ecological roles, e.g., they are the primary seed dispersers and pollinators, and are essential for maintaining the island’s unique forests; their loss would likely trigger extinction cascades. Although the fate of all lemur species should be considered precarious due to increasing habitat destruction, the knowledge that some lemurs are able to cope with this degradation (to a certain degree) should be seen as positive. Recent studies have begun to alter our view of *Hapalemur* spp. as dietary specialists: they demonstrate dietary flexibility and some populations are able to subsist on items other than bamboo. Some primate species adapted to narrow ecological specializations may be sensitive to natural or anthropogenic habitat perturbations, whereas others have been shown to adjust to these changing environments. The ecological flexibility of the southern bamboo lemurs might provide a model for conservation action to help some of their congeners to survive. Among the most threatened within the genus is the Lac Alaotran gentle lemur (*H*. *alaotrensis*), assessed by the IUCN as Critically Endangered (CR B1ab(iii,v)), due to its greatly restricted range that is becoming increasingly populated while the remaining viable habitat continues to shrink. They subsist on a diet limited to sedges and non-bamboo grasses, similar to *H*. *meridionalis* when it is in the *Melaleuca* habitat. In captivity, however, *H*. *alaotrensis* regularly display a preference for bamboo, suggesting little divergence from congeners, with a flexibility that may allow them to persist in habitats outside of Lac Alaotra. As it would appear that all former subspecies of *H*. *griseus* maintain some dietary plasticity, in that they are not restricted to bamboo forests, perhaps conservationists need to rethink their strategy when considering how to save species of the *Hapalemur* genus. # Supporting Information This work was carried out under the Accord de Collaboration between the University of Antananarivo and the University of Hamburg, as well as a collaboration agreement between TME and QIT Madagascar Minerals (QMM). We thank the Madagascar Ministry of the Environment, Water and Forests for approving protocols and granting permission to conduct research. Special thanks to Jacques Rakotondranary and Tolona Andrianasolo for their logistical support and to Katie Hall and Natalie Breden for their assistance in the field. We acknowledge Johny Rabenantoandro and the QMM biodiversity staff for their assistance and provision of on-site logistical support. We thank Clara Scarry and Shauna Burgess Mora for assistance with the spatial analyses. Finally, we thank Steven Goodman for helpful comments and suggestions. [^1]: The authors have the following interests: This study received funding from the Knowsley Safari Park, through a grant awarded by the Primate Society of Great Britain. Co-authors Jean-Baptiste Ramanamanjato, Faly Randriatafika, Laza N. Andriamandimbiarisoa, David Rabehevitra and Robertin Ravelomanantsoa are employed by QIT Madagascar Minerals. 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: TME JUG GD. Performed the experiments: TME RR JBR FR LNA DR. Analyzed the data: TME JUG GD. Contributed reagents/materials/analysis tools: TME JUG GD JBR FR LNA DR RR. Wrote the paper: TME JUG GD.

# Introduction Dysarthria is a neurological speech disturbance characterized by abnormalities in muscle strength, steadiness, tone, speed, range of motion, and/or accuracy of control of speech organs (e.g., tongue, lips, and larynx) for speech production. Among these multidimensional motor disorders, muscle weakness is associated with decreased exercise speed, which is speculated to be associated with slow speech. This hypothesis has received mixed support depending on the correspondence between performance of the syllable repetition task, called oral diadochokinesis (oral-DDK) or alternating motion rate, and muscle weakness. However, current reviews have reported no significant relationship between tongue strength and speech-related indicators, such as speech intelligibility, articulation rate, and oral-DDK rate. Low levels of orofacial muscle strength are required to generate utterances. The orofacial muscle strength required for normal speech is at most 10%–20% of the maximum muscle strength. In speakers with atrophic lateral sclerosis with orofacial muscle weakness (i.e., bulbar paralysis), the ratio of tongue-to- palatal contact pressure during speech to maximum isometric tongue muscle strength is 2%–8%. This ratio is not significantly different from that of healthy participants, suggesting that the maximum muscle strength and tongue- palate contact pressure during speech decrease proportionally. Additionally, speakers with dysarthria in whom tongue muscle strength is lower than the lower limit of normal speakers have moderate-to-severely reduced articulatory precision and overall severity (including speech intelligibility and naturalness). In a previous cross-sectional study, dysarthria speakers (n = 8) with severe anterior tongue elevation muscle strength had an oral-DDK rate of \<5.8 syllable/s for the syllable /tʌ/. In contrast, 44.6% of the remaining speakers with dysarthria had an oral-DDK rate of \>5.8 syllables/s. These findings suggest that a severe decrease in orofacial muscle strength adversely affects speech intelligibility. Abnormal articulation is the main cause of poor speech. In particular, the tongue (among the articulatory organs) has a strong effect on articulation. Elevation strength of the anterior tongue is correlated well with audibly acquired articulatory precision compared with speech intelligibility. However, since audibly articulatory precision and speech intelligibility are qualitative assessments, they affect the distribution of the data (ceiling or floor effect). A second formant (F2) slope is a quantitative evaluation of articulation. It is a quantitative acoustic measurement based on connected speech (word or sentence level). The F2 slope changes almost in response to the back and forth movements of the tongue, and frequency trajectories, such as diphthongs, move up and down relatively rapidly. Therefore, the F2 slope is speculated to reflect the movement speed of the tongue during articulation. A correlation exists between perceptually measured vowel accuracy and the F2 slope. In addition, the F2 slope of the F2 transition is also correlated with speech intelligibility. The clear explanation for the decrease in the F2 slope in speakers with dysarthria is the relatively slow changes in tongue shape. Specifically, the back-and-forth motion of the tongue during articulation is slower and/or the range of movement is narrower, resulting in a longer and thinner change in the F2 movement. The hypothesis that the tongue strength affects the speech speed does not appear to be supported by the weak correlation between the tongue strength and the articulation and oral-DDK rates. However, these rates quantify how fast the syllables are generated, and the accuracy of articulation is not considered much. In addition, there is a trade-off between articulation accuracy and speed. The index for measuring the rate of syllable generation varies from person to person in terms of the actual movement speed of the tongue during articulation (the relationship between movement range and required time). A study investigated the correlation between the range of motion of the oral articulator during articulation and tongue strength. Speakers with oculopharyngeal muscular dystrophy (n = 12) showed no correlation between vowel space area or vowel F2 range and tongue muscle strength. This result may be due to the fact that the range of motion of the oral articulator can be compensated for by slowing down the speech speed. Given the above limitations, additional research is needed on the relationship between tongue strength and slowed tongue movement during speech. This study aimed to further elucidate the relationship between tongue muscle weakness and dysarthria in an adult multidisciplinary group. Speech-related indicators have been extended from the well-studied oral-DDK rate and speech intelligibility to include the F2 slope. As mentioned above, the F2 slope directly measures tongue movement during articulation from the viewpoint of range and speed. Furthermore, the anterior tongue elevation and anterior tongue consonant tasks are related. Our interest in the relationship between tongue strength and the F2 slope was also motivated by the specificity of this site. This study may provide suggestions for adaptation and efficacy verification of strength training for dysarthria. # Materials and methods ## Participants This cross-sectional study population included speakers with dysarthria who were admitted to acute and convalescent hospitals between September 2017 and June 2020 and were continuously evaluated by speech-language-hearing therapists (SLHTs). The eligibility criteria were as follows: 1) request made for speech rehabilitation from a doctor; 2) Japanese as the first language; 3) absence of severe cognitive impairment or psychiatric disorders that may hinder speech assessment, 4) no complications of respiratory function that may affect speech, such as pneumonia and asthma; and 5) no dentition defects that affect the production of lingual–alveolar consonants or tongue pressure measurement. Background information, such as age, sex, height, weight, albumin, and diagnosis were obtained from the medical records of the participants. Body mass index was calculated as weight (kg) divided by height in meters squared (m²). These factors were considered to account for the possible effects on tongue muscle strength and speech caused by factors other than the primary disease causing dysarthria. Classification of the dysarthria type was diagnosed by SLHTs using the Mayo Clinic classification system. Thirty neurologically normal speakers (14 men and 16 women) aged 19–85 years (median, 22 years; interquartile range, 21.0–23.8 years) were also included. This group comprised 25 participants reported in a previous study and five selected in their preliminary experiments. It also included one participant with esophageal cancer (76 years old, male) and one with a trochanteric fracture of the femur (85 years old, female). These participants comprised the control group for the tongue pressure and speech-related indicators in this study. Their heights, weights, and albumin levels were not included. ## Tongue strength A balloon-type tongue pressure measurement device (TPM-01; JMS Co. Ltd, Hiroshima, Japan) was used to measure tongue strength. Maximum tongue pressure (MTP) measurements were performed by six SLHTs, including the trained author (WY). The reproducibility and reliability of this device have been validated in a previous study. The measured values were calculated according to a previously established methodology. The TPM-01 comprises a disposable probe, an injection tube as a connector, and a hard ring (bite block; length, 8.5 mm; thickness, 0.5 mm; diameter, 6.0 mm) device. The participants were instructed to put the balloon in their oral cavity in a sitting position. They held the probe at the midpoint of the central tooth. The participants were asked to maintain this position while the measurer adjusted the probe and confirmed the correct position. Measurements were performed thrice with 1-min rest and one preliminary exercise. The maximum value of the three measurements was defined as the MTP in kilopascals (kPa). In this measuring device, a balloon, which is fixed in front of the tongue, is compressed with the tongue toward the palate. Thus, it is speculated to reflect the anterior tongue strength. In addition, since the bite block is fixed by the incisors, the compressive force of the temporomandibular joint does not affect the MTP measurement. ## Speech analyses The speech test was evaluated using two tasks: reading aloud a long sentence “The north wind and the sun” and oral-DDK of the lingual–alveolar consonant /ta/ (/a/ corresponds to /ʌ/ in Japanese). In the reading aloud task, patients were instructed to “read aloud using the volume, pitch, and speed as when normally speaking (without intentionally speeding up or slowing down).” A practice reading was performed before recording. In oral-DDK, the participants were instructed to “repeat /ta/ at maximum speed without taking a breath,” and two measurements were performed. The speech of the participants was saved as an uncompressed file, with a sampling frequency of 44.1 kHz with 16-bit quantization using a digital voice recorder (R-05; Roland, Shizuoka, Japan). The recording was conducted in a quiet room with a noise level of 30 dBA or less. The microphone to mouth distance was 15 cm, and the input level was kept constant. ### Auditory perceptual assessment We evaluated speech intelligibility, which is one of the most important indicators of speech disorder severity. Three certified SLHTs (TT, YT, and YW) blindly and audibly evaluated the recording of long sentences that were read aloud. The representative value of speech intelligibility was the average value from the three evaluators on a nine-point speech intelligibility scale. This evaluation system is widely used in Japan, and its reliability was confirmed previously. Speech intelligibility was scored from 1 to 5 in 0.5 increments. A score of 1 indicates normal; 5, severe; and 2–4, cases with speech disorders between the two points. ### Oral-DDK rate To reduce the effects of speech irregularities such as freezing, slurring, or syllable prolongation during speech onset, or respiratory dysfunction during the second half of the task, we extracted \~3 s of recorded data from the middle parts of the audio for the analysis. The acoustic analysis was performed by the author TT using the acoustic analysis software Multi-Speech 3700 with Motor Speech Profile 5141 (KayPENTAX, Lincoln Park, NJ, USA). The maximum repetition rate of oral-DDK (unit: syllables/s) was calculated. The representative value of the oral-DDK is the average value of the two measurements. ### Second formant transition The acoustic analysis was performed by the author TT using Praat acoustic analysis software (ver. 6.0.50; Boersma & Weenink, University of Amsterdam). From the recorded voice of the long sentence “The north wind and the sun,” the following three parts were analyzed: /ai/ and /jo/ of “太陽” /taijo:/ means “sun,” and /ai/ of “外套”/ɡaito:/ means “coat.” /taijo:/ and /ɡaito:/ appeared three times each in the long sentence, which were all analyzed (3 parts × 3 times = 9 times in total). For the analysis object, the transition section of the F2 of the two-vowel and semivowel sequences was extracted, and the F2 movement part was used as the measurement target. Details of the measurement targets: 1) /taijo/’s /ai/ transition, 2) /taijo/’s /jo/ glide, and 3) /gaito:/’s /ai/ transition. shows an example of the /taijo:/ measurement. F2 movement was measured based on a previous report measuring diphthong /aɪ/ and semivowel /jæ/. The movement duration (ms) was set from the start (F2 onset) to the end (F2 offset) of the F2 movement. The F2 linear predictive coding tracks on a wide- band spectrogram (analysis bandwidth, 300 Hz) were manually edited and identified. The 20/20 rule (specifying a frequency change of ≥20 Hz during 20 ms in the transition onset and offset) was applied. If the F2 track was unclear due to hoarseness or other noises, we identified it by referring to the changes in F1 (opening and closing of the mandible) that were almost synchronized with F2 in the /ai/ and /jo/ sequences. Thus, relatively stationary portions of the vowel before or after the target transition were not included in the analysis. Movement extent (Hz) is the difference in the F2 between the beginning and end. The F2 slope (unit: Hz/ms) was obtained by dividing the movement extent by the movement duration. Therefore, the F2 slopes are expressed in absolute values (Hz/ms) throughout this manuscript to eliminate the positive/negative sign caused by the target-inherent F2-movement direction. Based on a previous study of dysarthria, we averaged all the F2 slopes with different acoustic properties. This aimed to reduce the error from the actual syllable-specific tongue motion velocity, considering the possibility that mandibular opening and closing could be related to the context. ## Statistical analyses Data are expressed as the median (minimum, 25th percentile, 75th percentile, maximum). SPSS v. 27.0 (IBM Corp., Armonk, NY, USA) was used for the statistical analyses. Nonparametric tests were selected for smaller sample sizes in the control group and for effect size comparisons between indicators. The Mann–Whitney U test was used to assess the difference between a speaker with dysarthria and a neurologically normal speaker. The effect size of the difference was evaluated by calculating r from the obtained z-value and the number of participants. The χ-square test was used to evaluate the differences in the sex ratio. Pearson correlation was used to examine the relationships between continuous variables and MTP results for speakers with dysarthria. Spearman rank correlation analysis indicated the strength of the associations between the MTP and speech intelligibility because of non-normal distributions for speakers with dysarthria. Normal distribution was confirmed using the Shapiro–Wilk test. Since a difference exists between male and female individuals in the MTP and F2, subgroup analysis was also performed. In addition, a subgroup correlation analysis by the dysarthria subtype and MTP severity (divided into two groups by median) was performed. For subgroup analysis, a nonparametric method was selected from a small sample size. Statistical significance was set at p \< 0.05. The sample size required to detect an association between speech-related variables and the MTP was 59 or more participants, with an effect size of 0.35 and a power of 0.8, based on a previously established methodology. In a previous study, the effect size of anterior tongue pressure and auditory perceptual assessment ranged from 0.35 to 0.52. Cronbach’s α was used to evaluate the reliability of the evaluators of speech intelligibility. The in-session reproducibility of the oral-DDK rate within the participants was evaluated using intraclass correlation coefficients (ICCs). The ICC was used to evaluate the reliability within measurement of the F2 slope. It evaluated a group of speakers with dysarthria who may have more variability in measurements than healthy individuals. ## Statement of ethics Consent from participants was obtained in writing and verbally. Informed consent was obtained from all participants. All procedures were approved by the ethics committees of the Uonuma Kikan Hospital (Approval no.: 30–007) and Nagaoka Nishi Hospital (Approval no.: 29–02). Consent was also obtained from all participants regarding the secondary use of data. We guaranteed the participants of their rights to withdraw from the study using an opt-out procedure. # Results The final analysis included 63 participants (median age, 68 years; interquartile range, 58–77 years; 44 men and 19 women). There were 72 entries in this study. However, seven speakers with severe cognitive impairments were excluded. Sixty- five speakers with dysarthria met the eligibility criteria. Of these, two were excluded due to lost data. A significant difference in the age and sex was found between neurologically normal speakers and speakers with dysarthria (age: p \< 0.001, sex: p \< 0.031). Details of all the participants, including healthy speakers, are shown in the. The diagnoses were as follows: neurovascular events (41), progressive neurological diseases (16) (amyotrophic lateral sclerosis, 3; multiple system atrophy, 5; Parkinson’s disease, 2; myasthenia gravis, 2; corticobasal degeneration, 1; progressive supranuclear palsy, 1; spinocerebellar ataxia, 1; and hereditary spastic paraplegia, 1), and other neurological diseases (6). The breakdown of the dysarthria subtypes in the cases was as follows: spastic, 3; flaccid, 7; ataxic, 7; hypokinetic, 6; hyperkinetic, 1; unilateral upper motor neuron, 16; mixed, 14; and undetermined, 9. ## Tongue strength and speech measures shows the results of the MTP and speech-related evaluations measured in this study. The interrater reliability assessed using Cronbach’s α for speech intelligibility among the three raters was 0.931, indicating high reliability. The two measurements of the oral-DDK rate showed high reproducibility of the ICC (0.985). In addition, of the 63 speakers with dysarthria, the recorded speeches of 19 randomly selected speakers (30.2%) were measured again by the same examiner who had performed the first measurement (6 months previously), and the ICC of the two measurements showed high reliability (F2 slope, 0.904). ## Relationship between tongue strength and speech measures The MTP and F2 slope were significantly associated among speakers with dysarthria (r = 0.368, p = 0.003). No significant correlation was detected between the MTP and oral-DDK/speech intelligibility. In the analysis by sex groups, a significant correlation was observed between the MTP and F2 slope (male, rs = 0.397, p = 0.008; female, rs = 0. 479, p = 0.038). The correlation between MTP and speech intelligibility was significant only in males (rs = -0. 328, p = 0.030). There was no significant correlation between MTP and the /ta/ DDK rate in either sex. In the analysis by the dysarthria subtype, a significant correlation was observed between the MTP and F2 slope (Flaccid, rs = 0.786, p = 0.036; Mixed, rs = 0.640, p = 0.014;). There was no significant correlation between any of the combinations in the other subtypes. Note that spastic (n = 3) and hyperkinetic (n = 1) types were excluded from the analysis due to their small sample size. In the analysis by the groups categorized according to maximum tongue pressure, a significant correlation was observed between the MTP and all speech-related variables only in the lower MTP group (Speech intelligibility, rs = 0.397, p = 0.008; /ta/ DDK rate, rs = 0. 479, p = 0.038; F2 slope, rs = 0. 479, p = 0.038; ). # Discussion In this study, we measured the tongue pressure in speakers with various types of dysarthria, following which, we conducted a correlation analysis between MTP and speech-related indices. The tongue pressure was not significantly associated with the oral-DDK rate and speech intelligibility. However, the tongue pressure was significantly associated with the F2 slope. In addition, a significant difference in the MTP and all speech-related indicators was noted between speakers with dysarthria and neurologically normal speakers, with moderate to large effects. To date, many studies have investigated the relationship between tongue muscle strength, speech intelligibility, articulatory precision, and the oral-DDK rate. To the best of our knowledge, this study is the first to investigate the relationship between tongue strength and the F2 slope. These findings provide some implications for understanding tongue muscle strength in patients with dysarthria. The effect size of the difference in the oral-DDK rate between speakers with dysarthria and those in the control group was large. However, in speakers with dysarthria, the oral-DDK rate and MTP were not correlated. Previous studies have shown contradictory results regarding the relationship between the oral-DDK rate and tongue strength. For example, in patients with oculopharyngeal muscular dystrophy (n = 12), the MTP and oral-DDK rates were significantly reduced compared with healthy participants, although no correlation was found. On the other hand, conventional speech rehabilitation therapy and tongue strength exercises for speakers with cerebrovascular disorders result in a significantly faster oral-DDK rate at /tʌ/. One of the factors behind these contradictory results is the inadequacy and imbalance of the participants. Specifically, depending on the threshold of the tongue muscle weakness affecting speech, the correlation may not be clear when only a small number of patients has the most severe muscle weakness. In addition, there may be a difference in the degree of contribution to the speech impairment between diseases in which muscle weakness is the main symptom and other diseases. A previous study including speakers with different types of dysarthria (n = 55) described a weak correlation between the oral-DDK rate for /tʌ/ and MTP (r = 0.247). In the current study, not much difference was observed in the correlation coefficient between the MTP and oral- DDK rate (r = 0.142) compared with a previous similar study. Therefore, the results of this study do not support a strong relationship between MTP and the oral-DDK rate. In this study, the MTP was not correlated with speech intelligibility in speakers with dysarthria. A previous study described a moderate correlation between speech intelligibility (%) and the MTP (rs = 0.349) in speakers with different types of dysarthria (n = 55). It also showed that the subgroups (n = 8) with severely reduced MTP included three participants with amyotrophic lateral sclerosis and five with sustained combat injuries (many with polytraumatic injuries). This may have affected the results, as our study did not include participants with sustained combat injuries. Combat injuries with orofacial injuries may have a greater impact on the tongue function necessary for speech. In addition, speech intelligibility is evaluated using auditory (qualitative) techniques and is not highly sensitive to mild cases. Therefore, the difference in the correlation coefficient could not possibly detect a significant correlation because few participants with severe dysarthria were included in this study. Furthermore, comparing our study with previous studies, which showed a strong correlation between word intelligibility and tongue muscle strength, the heterogeneity of the target disease is considered to have an effect. For example, Searl et al. found that 13 participants with amyotrophic lateral sclerosis (n = bulbar type 8, spinal type 5) have a strong correlation between tongue strength and speech intelligibility. Amyotrophic lateral sclerosis is correlated with mixed flaccid–spastic type of dysarthria, and muscle weakness has a significant effect on speech. Therefore, in cases where muscle weakness extends to the whole body, the relationship between speech intelligibility and tongue muscle strength becomes stronger. In our results, speech intelligibility and oral-DDK rate were not significantly correlated with MTP in all subtypes. However, the flaccid type (n = 7) had the highest correlation coefficient among all types (speech intelligibility, rs = -0.436; /ta/ DDK rate, rs = 0.517). Additional subtype-specific studies are warranted. In this study, the F2 slope and MTP were significantly correlated in speakers with dysarthria (r = 0.368). Solomon et al. concluded that auditory articulatory precision is suitable for assessing the association between tongue strength and speech function. The main cause of reduced speech intelligibility is abnormal articulation, although speech intelligibility and articulatory precision are not always the same. Focusing on the adequacy of articulation is reasonable to more directly evaluate the tongue, which plays a major role in articulation. F2 roughly corresponds to the back-and-forth movement of the tongue. The F2 slope, calculated from the movement duration and extent of the F2, is an acoustic index that correlates with the perceived accuracy of vowels. The F2 slope is speculated to reflect the speed of the tongue movement during articulation. In this study, the correlation between the F2 slope and tongue muscle strength was similar to all speakers with dysarthria in the sex-separated subgroups. In previous studies, sex does not significantly affect the relationship between the F2 slope and speech intelligibility in speakers with dysarthria. As for the MTP, sex differences disappear beyond the age of 50 years. In summary, in speakers with dysarthria, tongue strength may be associated with tongue movement velocity during articulation. Our findings partially support the hypothesis that “muscle weakness is associated with slow speech”. On the other hand, the effect size of the difference between the speakers with dysarthria and healthy participants on the MTP and F2 slope in this study was moderate. However, the effect of the difference in oral-DDK rate and speech intelligibility between healthy speakers and speakers with dysarthria was large. In a previous study, Japanese speakers with mild dysarthria (n = 16) showed no significant difference in the MTP compared with speakers without dysarthria (n = 29). In addition, the F2 slope was lower in speakers with moderate to severe dysarthria. Thus, MTP and F2 slope are less capable of differentiating between healthy speakers and speakers with mild dysarthria than oral-DDK rate and speech intelligibility. Nevertheless, the findings of this study showed a moderate correlation between the F2 slope and MTP in speakers with dysarthria. The F2 slope is correlated with tongue muscle strength, which is stronger than the relationship between tongue muscle strength and other speech-related indicators. In addition, the present study showed different results for the F2 slope and oral-DDK rate, which are related to tongue movement speed during articulation. This result is worthy of special mention. The DDK rate of /ta/ also depends on the speed of the voice on/off and mandibular opening/closing. In addition, the oral-DDK rate counts the number of syllables produced per second. Some articulation distortions (i.e., under shooting) are not reflected in the DDK rate measurements. Thus, both clear and unclear articulations are counted as one syllable (i.e., narrow or wide range of movement). The F2 slope, on the other hand, acoustically isolates the back-and-forth movement of the tongue during articulation and measures its speed. In a previous study, an oral-DDK task was not possible at maximum articulatory velocity. Therefore, it is likely that the F2 slope better reflects the speed of the tongue movements than the oral-DDK rate, and this may have affected the results. This study has some limitations. First, it was a cross-sectional study; thus, the causal relationship between tongue muscle strength and the F2 slope is unknown. Longitudinal studies are useful for verifying causal relationships. Second, it did not consider the speech treatments the participants received in this study (e.g., non-speech oral motor exercises, speech rate modification, and Lee Silverman voice treatment). Third, the control group was not strictly set. In this exploratory research, we included young people to clearly show the difference between speakers with dysarthria and healthy speakers. A comparative controlled study involving participants matched for age and sex would be useful for a more detailed understanding. Finally, because only few participants with severe dysarthria were included, the data were not sufficient to adapt to people with severe disease. JMS devices have bite blocks and cannot be used in patients with missing teeth, which indicates that collecting data from elderly and severely ill patients may have been difficult. For studies involving speakers with dysarthria that require those with severe dysarthria, considering devices without bite blocks and sensor-type pressure gauges is necessary. Hence, further studies are warranted to overcome these limitations and to assess the broader applicability of the MTP and F2 slope. # Conclusions The MTP was not significantly correlated with the oral-DDK rate and speech intelligibility. However, using correlation analysis, we confirmed that the F2 slope and MTP were related. This suggests that the maximum isometric tongue strength is associated with the tongue movement speed during articulation. From a clinical point of view, tongue strength training may be considered for dysarthria speakers with a reduced F2 slope (that is, the appropriateness of articulation and speed) and tongue strength. From a research point of view, the F2 slope may be useful for verifying the effect of tongue strength training and the appropriateness of articulation and speed. # Supporting information The authors would like to thank Nagaoka Nishi Hospital (Nanae Koyama, Sanae Saito, and Hiromi Takeuchi) and Uonuma Kikan Hospital (Yuuki Kobayashi and Koki Maruyama) for their assistance with data collection. 10.1371/journal.pone.0264995.r001 Decision Letter 0 Finley Sara Academic Editor 2022 Sara Finley 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 2021 PONE-D-21-32058Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthriaPLOS ONE Dear Dr. Tamura, 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 see the editor and reviewer comments, below. Please submit your revised manuscript by Jan 03 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, Sara Finley, Ph.D. 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\. Thank you for stating the following financial disclosure: “This work was supported by the Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research (KAKENHI) (<https://www.jsps.go.jp/j-grantsinaid/>) under Grant number JP20K19324.” Please state what role the funders took in the study.  If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." If this statement is not correct you must amend it as needed. Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. 3\. We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. 4\. 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>. Additional Editor Comments: I have received 2 reviewer reports for the paper, “Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthria”. Both reviewers note that the research question is of potential interest to a wide range of scholars, and that the methods and interpretation of results are generally sound. However, both reviewers expressed some concerns related to the details of how the F2 slope was calculated. Given that this is the major contribution of this paper, I agree that more details need to be included in your revisions about how F2 slope was calculated, particularly the rationale for the method chosen. For example, Reviewer 2 asks questions about whether the F2 slope was calculated across the entire word, or within each vowel separately, so more details on how the vowels in the target words were analyzed distinctly from the consonants in the words is needed. An additional figure, as suggested by reviewer 1, would be welcome. Reviewer 1 also suggests several places where the writing could be improved for clarity, and I suggest following this advice, particularly where redundancies can be avoided (also noted by Reviewer 2), or jargon can be removed or explained (particularly in the abstract). Reviewer 2 also suggests that some analysis be conducted related to the type and severity level of dysarthria, which would be welcome, given the title of the paper. \[Note: HTML markup is below. Please do not edit.\] 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: 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 \*\*\*\*\*\*\*\*\*\* 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: Relationships between maximum tongue pressure and second formant transition in　speakers with different types of dysarthria General comments: This paper has a new important issue on speech evaluation for dysarthria patients. But the outside reader of your manuscript it could be a little hard to understand correctly. Could you look at the points below: Abstract 1\. P3 L22-23 “The tongue pressure in speakers with dysarthria was significantly associated with the second formant transition slope.” This sentence seems to have the same meaning as the one before this sentence. I recommend that you delete this sentence. 2\. P3 L23-24 “This result suggests that the maximum isometric tongue strength is associated with articulation severity and tongue movement speed during articulation.” The results showed that MTP was not correlated with speech intelligibility. The "the maximum isometric tongue strength is associated with articulation severity" in this sentence is inconsistent with the result. Introduction 1\. P5. L56-57 “these are qualitative assessments” Does this “these” mean "Elevation strength of the anterior tongue"? If so, the elevation strength of the anterior tongue is quantitative assessment, not qualitative assessment. 2\. P3 L57 “they affect the distribution of data (ceiling or floor effect).” Please add references. 3\. P6 L67-69 “The hypothesis that … oral-DDK rates.” Please add references. Material and Methods 1\. P9 L121 “thickness, 0.5mm” I think it would be more appropriate to write the diameter rather than the thickness of the bite block. 2\. P12 L177-179 I suggest you should make a figure for analysis methods about onset time, offset time, movement duration, and movement extent. 3\. P12 L176-177 In the previous study, /aɪ/ was used to analyze the F2 movement. Why did you include /jo/ as well as /ai/ in the analysis in this study? 4\. P12 L184-185 The F2 slope values of /ai/ and /jo/ were calculated as one average value in this study. Why did you calculate the two types of two-vowels (/ai/ and /jo/) with different acoustic characteristics together as one average value? Results 1\. P14 L223-224 This cross-sectional study population included speakers with dysarthria who were admitted to acute and convalescent hospitals between September 2017 and June 2020. Please describe how many participants were entered, and how many participants did not match the inclusion criteria and were excluded, resulting in 65 participants matching the inclusion criteria. In addition, please describe the reasons for the exclusion. 2\. P14-15 L225-227 “A significant difference in age was found between neurologically normal speakers and speakers with dysarthria (p \< 0.001)”. This sentence only showed the difference in age between neurologically normal speakers and speakers with dysarthria. Please describe the statistically analyzed differences in sex between neurologically normal speakers and speakers with dysarthria. 3\. P17 L244 “The average MTP of a speaker with dysarthria was 32.3 ± 9.9 kPa.” The median of MTP was shown in table 2, so this statement did not need to show the mean of MTP. Discussion 1\. P21 L303-304 “One of the factors behind these contradictory results is the inadequacy and imbalance of the subjects \[10,13\].” Please add a specific explanation for “inadequacy and imbalance of the subject”. 2\. P21 L309-310 “Therefore, the results of this study support the weak correlation between the oral-DDK rate and tongue muscle strength.” In this study, there was no significant correlation between the oral-DDK rate and tongue muscle strength. I consider that this statement is over-interpreted. 3\. P21 L316-317 “This may have affected the results, as our study did not include participants with sustained combat injuries.” Please describe how the non-inclusion of the participants with sustained combat injurie affected the results in this study. 4\. P22 L321 were included →　were included in this study. 5\. P23 L354-355 “Therefore, the smaller effect size of the difference between the F2 slope and MTP suggests that speakers with mild dysarthria have lower discriminative function.” This sentence is difficult to understand. Please describe why the smaller effect size of the difference between the F2 slope and MTP suggests that speakers with mild dysarthria have lower discriminative function. 6\. P24 L357 between the F2 slope and MTP. →　between the F2 slope and MTP in speaker with dysarthria. 7\. I consider that the oral-DDK rate and F2 slope are associated with the tongue movement speed during articulation. However, the MTP was significantly related to the F2 slope but not significantly related to the oral-DDK rate in this study. Please describe why the oral-DDK rate and F2 slope, which are associated with the tongue movement speed during articulation, had different results from each other. Reviewer \#2: This paper investigates the relationships between maximum tongue pressure and speech-related features, which include speech intelligibility, /ta/ DDK, and F2-slope. While speech intelligibility and /ta/ DDK were revealed to have no significant correlations with maximum tongue pressure, F2-slope was significantly correlated with the maximum tongue pressure (r=0.37, p\<0.05). This paper proposes that the F2- slope may be useful at verifying the effect of tongue strength training. 1\. Is the manuscript technically sound, and do the data support the conclusions? \- The idea of using F2-slope which can see both articulation accuracy and articulation rate, is convincing. \- F2-slope is usually used for analyzing diphthongs(one vowel), and sometimes vowel sequences. However, I am concerned about this experiment design, which analyze the F2-slope in a word level. Are the features extracted from the start of /a/ and end of /o/? If this is the case, the design should be revised in a major manner. Especially, for word /gaito/, the consonant is in between the two vowels, which must interrupt in extracting the F2-slope. Even for /taijo/, the difference of F2 between /a/ and /i/ are much larger compared to difference of F2 between /a/ and /o/. Analyzing the F2-slope for each vowel sequence/semi- vowel may be more persuasive. 2\. Has the statistical analysis been performed appropriately and rigorously? Mann-Whitney U test is applied to investigate the difference between healthy speakers and dysarthric speakers. Pearson Correlation is used to examine the relationships between MTP results and speech function features. Subgroup analysis by gender are appropriately performed. Further subgroup analysis by dysarthric subtypes and severity levels should also be considered. In particular, the authors argue that the reason the study results do not agree with the previous results is because of the different distribution of speakers. Hence this analysis is necessary. 3\. Have the authors made all data underlying the findings in their manuscript fully available? \- The datasheet used for the statistical analysis is uploaded. All features used in the analysis are included - TP, F2-slope, /ta/DDK, speech intelligibility. \- However, each notation must be explained of its meaning. For example, what does POST and typeNo imply? \- Further, each notations should be explained why each feature is included in the data. It is hard to understand why are height, weight, BMI, Albumin are included for the analysis. \- Brief information of dysarthric speakers should be stated in the manuscript. (subtypes) 4\. Is the manuscript presented in an intelligible fashion and written in standard English? \- The manuscript is presented in an intelligible fashion and well-written in standard English. \- Though there are some redundant sentences that should be deleted for the final submissions. \- Abstract is quite confusing, especially the last sentence : "However, based on the degree of these correlations, the hypothesis that the relationship between the maximum force of the tongue and speech function is weak is also strengthened." This sentence blurs the main idea of this paper, which suggests the significant correlation between maximum force of the tongue and F2-slope. \- line 338 : F2-slope -\> F2 (second formant) \- line 340 : It -\> F2-Slope \*\*\*\*\*\*\*\*\*\* 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.0264995.r002 Author response to Decision Letter 0 12 Dec 2021 Joerg Heber Editor-in-Chief Sara Finley, Ph.D. Academic Editor PLOS ONE Dear Editor: I wish to re-submit the manuscript, titled “Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthria.” The manuscript ID is PONE-D-21-32058. I thank you and the reviewers for your thoughtful suggestions and insights. The manuscript has significantly benefited from these insightful suggestions. I look forward to working with you and the reviewers to move this manuscript closer to publication in PLOS ONE. The manuscript has been rechecked and the necessary changes have been made in accordance with the reviewers’ suggestions and marked in red. The responses to all the comments have been prepared and presented below. Thank you for your consideration. I look forward to hearing from you. Sincerely, Toshiaki Tamura Department of Speech, Language, and Hearing Sciences, Niigata University of Health and Welfare, Niigata city, Niigata 950-3198, Japan Tel: +81 (25) 257-4507 Fax: +81 (25) 257-4507 Email: <toshiaki-tamura@nuhw.ac.jp> Point-by-point responses to comments from the Editor and Reviewers Thank you for reviewing our work. We appreciate all your comments and suggestions. We have revised the manuscript accordingly. Our point-by-point responses are presented below. Response to Journal Requirements Comment 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> Response 1: We are very grateful for your guidance. We have ensured that the manuscript meets the style requirements of PLOS ONE. Comment 2: Thank you for stating the following financial disclosure: “This work was supported by the Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research (KAKENHI) (<https://www.jsps.go.jp/j-grantsinaid/>) under Grant number JP20K19324.” Please state what role the funders took in the study. If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." If this statement is not correct you must amend it as needed. Please include this amended Role of Funder statement in your cover letter; we will change the online submission form on your behalf. Response 2: Thank you for your valuable comment. Kindly revise the data availability statement accordingly for us. We have also added a statement to the cover letter. “This work was supported by the Japan Society for the Promotion of Science Grants-in-Aid for Scientific Research (KAKENHI) (<https://www.jsps.go.jp/j-grantsinaid/>) under Grant number JP20K19324. The funders had no role in the study design, data collection and analysis, decision to publish, and preparation of the manuscript.” Comment 3: We note that you have stated that you will provide repository information for your data at acceptance. Should your manuscript be accepted for publication, we will hold it until you provide the relevant accession numbers or DOIs necessary to access your data. If you wish to make changes to your Data Availability statement, please describe these changes in your cover letter and we will update your Data Availability statement to reflect the information you provide. Response 3: The data required for the analysis of this paper are included in the supplementary data. Therefore, we did not provide data repository information. Please change the data availability statement accordingly for us. We apologize for the inconvenience. We have also added a request for change in the cover letter. “Furthermore, the data required for the analysis of this paper are included in the submitted supplementary data. Therefore, we did not provide the data repository information.” Comment 4: 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>. Response 4: We have added captions at the end of the manuscript according to the supporting guidelines. In addition, we have updated the citations in the text. New additions are shown in red. Page 37, Lines 59–596 Supporting information S1 File. Dataset with all the participants. Re-measurement: result of a re- measurement by the same inspector \> 6 months after the original measurement (30% of speakers with dysphoria). Page 15, Lines 253–254 Details of all the participants, including healthy speakers, are shown in the S1 File. Response to the editor Comment: I have received 2 reviewer reports for the paper, “Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthria”. Both reviewers note that the research question is of potential interest to a wide range of scholars, and that the methods and interpretation of results are generally sound. However, both reviewers expressed some concerns related to the details of how the F2 slope was calculated. Given that this is the major contribution of this paper, I agree that more details need to be included in your revisions about how F2 slope was calculated, particularly the rationale for the method chosen. For example, Reviewer 2 asks questions about whether the F2 slope was calculated across the entire word, or within each vowel separately, so more details on how the vowels in the target words were analyzed distinctly from the consonants in the words is needed. An additional figure, as suggested by reviewer 1, would be welcome. Reviewer 1 also suggests several places where the writing could be improved for clarity, and I suggest following this advice, particularly where redundancies can be avoided (also noted by Reviewer 2), or jargon can be removed or explained (particularly in the abstract). Reviewer 2 also suggests that some analysis be conducted related to the type and severity level of dysarthria, which would be welcome, given the title of the paper. Response: We appreciate the reviewer for these constructive comments. We agree with all your suggestions; we have tried to reduce redundancy, particularly in the Abstract section. The authors’ responses to the comments are as follows: Reviewer \#1: General comments: This paper has a new important issue on speech evaluation for dysarthria patients. But the outside reader of your manuscript it could be a little hard to understand correctly. Could you look at the points below: Response: Thank you very much for your supportive comments and valuable suggestions for improving the quality of our manuscript. Abstract Comment 1: P3 L22-23 “The tongue pressure in speakers with dysarthria was significantly associated with the second formant transition slope.” This sentence seems to have the same meaning as the one before this sentence. I recommend that you delete this sentence. Response 1: Thank you for your helpful suggestion. We have omitted the duplication. Comment 2: P3 L23-24 “This result suggests that the maximum isometric tongue strength is associated with articulation severity and tongue movement speed during articulation.” The results showed that MTP was not correlated with speech intelligibility. The "the maximum isometric tongue strength is associated with articulation severity" in this sentence is inconsistent with the result. Response 2: Our data did not show that the articulation severity was related to MTP. Therefore, we have omitted "articulation severity and" from the sentence as indicated. Page 3, Lines 37–38 This result suggests that the maximum isometric tongue strength is associated with tongue movement speed during articulation. Introduction Comment 3: P5. L56-57 “these are qualitative assessments” Does this “these” mean "Elevation strength of the anterior tongue"? If so, the elevation strength of the anterior tongue is quantitative assessment, not qualitative assessment. Response 3: The word "these" meant "audibly articulatory precision and speech clarity." We have modified the sentence as follows: Page 4, Lines 67–68 However, since audibly articulatory precision and speech clarity are qualitative assessments, they affect the distribution of the data (ceiling or floor effect) \[13\]. Comment 4: P3 L57 “they affect the distribution of data (ceiling or floor effect).” Please add references. Response 4: Thank you for your kind suggestion. We have cited the reference for this sentence as follows: Page 4, Lines 69 “they affect the distribution of data (ceiling or floor effect) \[13\].” Comment 5: P6 L67-69 “The hypothesis that … oral-DDK rates.” Please add references. Response 5: Thank you for your valuable suggestion. We have added the reference to this sentence: Page 5, Lines 79–81 The hypothesis that the tongue strength affects the speech speed does not appear to be supported by the weak correlation between the tongue strength and the articulation and oral-DDK rates \[13\]. Material and Methods Comment 6. P9 L121 “thickness, 0.5mm” I think it would be more appropriate to write the diameter rather than the thickness of the bite block. Response 1 : Thank you very much for your valuable suggestion. We have added the diameter of the bite block as follows: Page 8, Lines 131–133 The TPM-01 comprises a disposable probe, an injection tube as a connector, and a hard ring (bite block; length, 8.5 mm; thickness, 0.5 mm; diameter, 6.0 mm) device (Fig 1). Comment 2. P12 L177-179 I suggest you should make a figure for analysis methods about onset time, offset time, movement duration, and movement extent. Response 2: Thank you very much for your valuable suggestion. We have created and included Figure 2 in our revised manuscript. Page 13, Lines 202–207 Fig 2. A spectrogram of the word /taijo:/ produced by a man with stroke and mild unilateral upper motor neuron dysarthria. This figure illustrates a major two vowel sequence and semivowel F2 movements in /ai/ and /jo/. The black lines running through the estimated centers of the first and second formants (F1 and F2) illustrate the formant tracing. The F2 slope was determined or the time interval, Duration, and the corresponding F2 frequency change, Extent. Comment 7. P12 L176-177 In the previous study, /aɪ/ was used to analyze the F2 movement. Why did you include /jo/ as well as /ai/ in the analysis in this study? Response 7: Thank you very much for your pertinent comment. We apologize for not explaining this point accurately. There is no consensus on the existence of double vowels in Japanese. In a previous study of a limited number of native Japanese speakers of dysarthria, vowel-to-vowel sequences were used (e.g. Tamura et al, 2021); however, there were concerns regarding the measurement errors. On the other hand, semivowels are also present in Japanese, and since we could obtain formant trajectories relatively close to the double vowels, we decided to add not only the /ai/ vowel sequence but also the semivowel /j/, which has characteristics similar to double vowels, to increase the data and reduce the measurement error. In addition, we included /jo/, which is different from /ai/, owing to the possibility that the opening and closing of the mandible could affect the formants (Yunusova et al, 2012). We have added the following explanations to our manuscript, which also address Comment 8. Page 11, Lines 190–192 F2 movement was measured based on a previous report measuring diphthong /aɪ/ \[33\] and semivowel /jæ/ \[22\]. Page 11–12, Lines 197–200 Based on a previous study of dysarthria \[19\], we averaged all the F2 slopes with different acoustic properties. This aimed to reduce the error from the actual syllable-specific tongue motion velocity, considering the possibility that mandibular opening and closing could be related to the context \[22\]. Comment 8. P12 L184-185 The F2 slope values of /ai/ and /jo/ were calculated as one average value in this study. Why did you calculate the two types of two-vowels (/ai/ and /jo/) with different acoustic characteristics together as one average value? Response 8: We wanted to minimize the error in the measurements, as shown in Response 7; therefore, we included /ai/ and /jo/, in our analysis, which have different acoustic properties, and both show a rapid rise and fall in F2. By mixing the measurement targets, we thought we could adjust the error of each measurement value. Kindly refer to our previous response. Results Comment 9. P14 L223-224 This cross-sectional study population included speakers with dysarthria who were admitted to acute and convalescent hospitals between September 2017 and June 2020. Please describe how many participants were entered, and how many participants did not match the inclusion criteria and were excluded, resulting in 65 participants matching the inclusion criteria. In addition, please describe the reasons for the exclusion. Response 9: Thank you for your valuable suggestion. We have added the number of entries, the number of people excluded, and the reason for exclusion as follows: Page 14, Lines 249–250 There were 72 entries in this study. However, seven speakers with severe cognitive impairments were excluded. Comment 10. P14-15 L225-227 “A significant difference in age was found between neurologically normal speakers and speakers with dysarthria (p \< 0.001)”. This sentence only showed the difference in age between neurologically normal speakers and speakers with dysarthria. Please describe the statistically analyzed differences in sex between neurologically normal speakers and speakers with dysarthria. Response 10: Thank you for your helpful comment. The χ-square test showed a significant difference in the sex ratio between the groups. We have added the following sentence to the manuscript: Page 12–13, Lines 216–217 The χ-square test was used to evaluate the differences in the sex ratio. Page 14–15, Lines 252–253 A significant difference in the age and sex was found between neurologically normal speakers and speakers with dysarthria (age: p \< 0.001, sex: p \< 0.031) Comment 11. P17 L244 “The average MTP of a speaker with dysarthria was 32.3 ± 9.9 kPa.” The median of MTP was shown in table 2, so this statement did not need to show the mean of MTP. Response 11: Thank you for pointing this out. We have omitted this sentence. Discussion Comment 12. P21 L303-304 “One of the factors behind these contradictory results is the inadequacy and imbalance of the subjects \[10,13\].” Please add a specific explanation for “inadequacy and imbalance of the subject”. Response 12: Thank you for your helpful suggestion. It has long been pointed out that dysarthria occurs in a variety of disease backgrounds (i.e., Parkinson's disease, ALS, stroke) and that different diseases produce different speech symptoms (Darley et al., 1969; Duffy 2020). In addition, various organs are involved in speech production, and speech characteristics may also differ depending on the damaged site. In order to correctly interpret the results of this study, it is necessary to collect and analyze a large number of cases for each of these diseases and disorders. Speech is reportedly affected the most by the severe loss of tongue muscle strength (Neel et al, 2015, Solomon et al, 2017). We have added these statements to the manuscript: Page 23, Lines 358–362 　　　 　Specifically, depending on the threshold of the tongue muscle weakness affecting speech, the correlation may not be clear when only a small number of patients has the most severe muscle weakness \[11,13\]. In addition, there may be a difference in the degree of contribution to the speech impairment between diseases in which muscle weakness is the main symptom \[4,10\] and other diseases \[3,13\]. Comment 13. P21 L309-310 “Therefore, the results of this study support the weak correlation between the oral-DDK rate and tongue muscle strength.” In this study, there was no significant correlation between the oral-DDK rate and tongue muscle strength. I consider that this statement is over-interpreted. Response 13: Thank you for your suggestion. We agree with your opinion, and have modified the sentence as follows: Page 23, Lines 366–367 Therefore, the results of this study do not support a strong relationship between MTP and the oral-DDK rate. Comment 14. P21 L316-317 “This may have affected the results, as our study did not include participants with sustained combat injuries.” Please describe how the non-inclusion of the participants with sustained combat injurie affected the results in this study. Response 14: Thank you for your helpful suggestion. Solomon et al. (2017) included 16/55 people with combat injuries. A subgroup analysis with severely reduced tongue pressure also included 5/8 patients with brain injury as well as structural destruction of the orofacial region by blast explosion or gunshot wound and subnuclear paralysis. We believe that the orofacial disruption strengthens the relationship between the tongue muscle strength and speech function. Solomon et al. (2017) themselves mention a possible effect on the distribution in the combat injuries. Accordingly, we have added the following explanation: Page 24, Lines 374–376 Combat injuries with orofacial injuries may have a greater impact on the tongue function necessary for speech. Comment 15. P22 L321 were included →　were included in this study. Response 15: Thank you for your valuable suggestion. The sentence has been revised as follows: Page 24, Line 377-379 Therefore, the difference in the correlation coefficient could not possibly detect a significant correlation because few participants with severe dysarthria were included in this study. Comment 16. P23 L354-355 “Therefore, the smaller effect size of the difference between the F2 slope and MTP suggests that speakers with mild dysarthria have lower discriminative function.” This sentence is difficult to understand. Please describe why the smaller effect size of the difference between the F2 slope and MTP suggests that speakers with mild dysarthria have lower discriminative function. Response 16: Thank you for your valuable suggestion. To clarify our intentions, we have revised the text as follows: Page 26–27, Lines 411–422 On the other hand, the effect size of the difference between the speakers with dysarthria and healthy participants on the MTP and F2 slope in this study was moderate. However, the effect of the difference in oral-DDK rate and speech intelligibility between healthy speakers and speakers with dysarthria was large. In a previous study, Japanese speakers with mild dysarthria (n = 16) showed no significant difference in the MTP compared with speakers without dysarthria (n = 29) \[40\]. In addition, the F2 slope was lower in speakers with moderate to severe dysarthria \[21,41\]. Thus, MTP and F2 slope are less capable in differentiating between healthy speakers and speakers with mild dysarthria than oral-DDK rate and speech intelligibility. Nevertheless, the findings of this study showed a moderate correlation between the F2 slope and MTP in speakers with dysarthria. The F2 slope is correlated with tongue muscle strength, which is stronger than the relationship between tongue muscle strength and other speech-related indicators. Comment 17. P24 L357 between the F2 slope and MTP. →　between the F2 slope and MTP in speaker with dysarthria. Response 17: Thank you for your valuable suggestion. We have incorporated the following changes: Page 26, Lines 419－421 Nevertheless, the findings of this study showed a moderate correlation between the F2 slope and MTP in speakers with dysarthria. Comment 18. I consider that the oral-DDK rate and F2 slope are associated with the tongue movement speed during articulation. However, the MTP was significantly related to the F2 slope but not significantly related to the oral- DDK rate in this study. Please describe why the oral-DDK rate and F2 slope, which are associated with the tongue movement speed during articulation, had different results from each other. Response 18: Thank you for your helpful suggestion. F2 slope reflects the rate of back-and-forth movement of the tongue during articulation. The DDK rate, on the other hand, is affected by the on/off of the voice and the opening/closing of the mandible. Moreover, the DDK rate does not take into account the distortion of the target articulation. Therefore, we believe that the F2 slope better reflects the motor function of the tongue during articulation. We have added an explanation in the discussion section as follows: Page 27, Lines 423–434 In addition, the present study showed different results for the F2 slope and oral-DDK rate, which are related to tongue movement speed during articulation. This result is worthy of special mention. The DDK rate of /ta/ also depends on the speed of the voice on/off and mandibular opening/closing. In addition, the oral-DDK rate counts the number of syllables produced per second. Some articulation distortions (i.e., under shooting) are not reflected in the DDK rate measurements. Thus, both clear and unclear articulations are counted as one syllable (i.e., narrow or wide range of movement). The F2 slope, on the other hand, acoustically isolates the back-and-forth movement of the tongue during articulation and measures its speed. In a previous study, an oral-DDK task was not possible at maximum articulatory velocity \[42\]. Therefore, it is likely that the F2 slope better reflects the speed of the tongue movements than the oral-DDK rate, and this may have affected the results. Reviewer \#2 This paper investigates the relationships between maximum tongue pressure and speech-related features, which include speech intelligibility, /ta/ DDK, and F2-slope. While speech intelligibility and /ta/ DDK were revealed to have no significant correlations with maximum tongue pressure, F2-slope was significantly correlated with the maximum tongue pressure (r=0.37, p\<0.05). This paper proposes that the F2- slope may be useful at verifying the effect of tongue strength training. Response: Thank you very much for your supportive comments and suggestions for the improvement of the quality of our manuscript. Comment 1: Is the manuscript technically sound, and do the data support the conclusions? \- The idea of using F2-slope which can see both articulation accuracy and articulation rate, is convincing. \- F2-slope is usually used for analyzing diphthongs(one vowel), and sometimes vowel sequences. However, I am concerned about this experiment design, which analyze the F2-slope in a word level. Are the features extracted from the start of /a/ and end of /o/? If this is the case, the design should be revised in a major manner. Especially, for word /gaito/, the consonant is in between the two vowels, which must interrupt in extracting the F2-slope. Even for /taijo/, the difference of F2 between /a/ and /i/ are much larger compared to difference of F2 between /a/ and /o/. Analyzing the F2-slope for each vowel sequence/semi- vowel may be more persuasive. Response 1: In this study, we did not analyze the whole word, but only the vowel-vowel transitions and semivowel (glide) portions. We analyzed three F2 slopes for each of these three locations: 1) the /ai/ transition in /taijo/, 2) the /jo/ glide in /taijo/, and 3) the /ai/ transition in /gaito:/. We have added the details of how we measured the F2 slope, as shown below, and created a measurement example in Figure 2. Page 11, Lines 188–190 Details of the measurement targets: 1) /taijo/'s /ai/ transition, 2) /taijo/'s /jo/ glide, and 3) /gaito:/'s /ai/ transition. Figure 2 shows an example of the /taijo:/ measurement. Page 11, Lines 195–197 Therefore, the F2 slopes are expressed in absolute values (Hz/ms) throughout this manuscript to eliminate the positive/negative sign caused by the target- inherent F2-movement direction. Comment 2: Has the statistical analysis been performed appropriately and rigorously? Mann-Whitney U test is applied to investigate the difference between healthy speakers and dysarthric speakers. Pearson Correlation is used to examine the relationships between MTP results and speech function features. Subgroup analysis by gender are appropriately performed. Further subgroup analysis by dysarthric subtypes and severity levels should also be considered. In particular, the authors argue that the reason the study results do not agree with the previous results is because of the different distribution of speakers. Hence this analysis is necessary. Response 2: Thank you for your helpful suggestion. We performed additional analyses grouped by dysarthria subtypes and by maximum tongue pressure. For the subtypes, there were significant correlations between F2 slope and MTP for Mixed and Flaccid, with Flaccid having the strongest correlations with oral-DDK rate and speech intelligibility compared with the other subtypes. In the group with the lower maximum tongue pressure, all the speech indices were significantly correlated with maximum tongue pressure. We have added the following text to the Methods, Results, and Discussion section, respectively. Table 4 and Figure 4 have been added to the Results section. Methods Page 13, Lines 224–225 In addition, a subgroup correlation analysis by the dysarthria subtype and MTP severity (divided into two groups by median) was performed. Result Page 20, Lines 311–318 　　　　　In the analysis by the dysarthria subtype, a significant correlation was observed between the MTP and F2 slope (Flaccid, rs = 0.786, p = 0.036; Mixed, rs = 0.640, p = 0.014; Table 4). Note that spastic (n = 3) and hyperkinetic (n = 1) types were excluded from the analysis due to their small sample size. Table 4. Spearman’s rank correlation coefficients (two sided) between the maximum tongue pressure and speech measures for each subtype of speakers with dysarthria (Total n = 59) Page 21–22, Lines 323–334 In the analysis by the groups categorized according to maximum tongue pressure, a significant correlation was observed between the MTP and all speech-related variables only in the lower MTP group (Speech intelligibility, rs = 0.397, p = 0.008; /ta/ DDK rate, rs = 0. 479, p = 0.038; F2 slope, rs = 0. 479, p = 0.038; Fig 4). Fig 4. Bivariate scatter plot and best-fit regression line of the maximum tongue pressure and speech-related variables in the two speakers with dysarthria groups divided according to the maximum tongue pressure by median Discussion Page 25, Lines 388–390 　　　　　　The results of this study also showed that the correlation between the MTP and all the speech evaluations was stronger in the flaccid type, where muscle weakness was the main symptom, than in the other types. Comment 3: Have the authors made all data underlying the findings in their manuscript fully available? \- The datasheet used for the statistical analysis is uploaded. All features used in the analysis are included \- However, each notation must be explained of its meaning. For example, what does POST and typeNo imply? Response 3: Thank you for your valuable suggestion. First, we changed "POST" to "Re-measurement" in the supporting information file. We have also added an explanation for " Re-measurement." In addition, I deleted "typeNo" since it was not relevant information for the readers. Page 37, Lines 593–596 　　　　　S1 File. Dataset with all the participants. Re-measurement: Result of a re- measurement by the same inspector \> 6 months after the original measurement (30% of speakers with dysarthria). Comment 4: Further, each notations should be explained why each feature is included in the data. It is hard to understand why are height, weight, BMI, Albumin are included for the analysis. Response 4: Thank you for your valuable suggestion. The authors have reviewed the data and determined that the height, weight, BMI, and albumin do not need to be analyzed. Therefore, we have removed them from the data in the supporting information file. Comment 5: Brief information of dysarthric speakers should be stated in the manuscript. (subtypes) Response 5: Thank you for your valuable suggestions. We have made the following additions: Page 16, Lines 266–268 The breakdown of the dysarthria subtypes in the cases was as follows: spastic, 3; flaccid, 7; hypokinetic, 6; hyperkinetic, 1; ataxic, 7; unilateral upper motor neuron, 16; mixed, 14; and undetermined, 9. Comment 6: 4. Is the manuscript presented in an intelligible fashion and written in standard English? \- The manuscript is presented in an intelligible fashion and well-written in standard English. \- Though there are some redundant sentences that should be deleted for the final submissions. Abstract is quite confusing, especially the last sentence : "However, based on the degree of these correlations, the hypothesis that the relationship between the maximum force of the tongue and speech function is weak is also strengthened." This sentence blurs the main idea of this paper, which suggests the significant correlation between maximum force of the tongue and F2-slope. Response 6: Thank you for pointing this out. We have removed the aforementioned sentence and modified the entire abstract to correct the redundancy. Comment 7: line 338 : F2-slope -\> F2 (second formant) Response 7: Thank you for your valuable suggestion. We have split the sentence into two and modified it as follows: Page 26, Lines 399–402 F2 roughly corresponds to the back-and-forth movement of the tongue \[15\]. The F2 slope, calculated from the movement duration and extent of the F2, is an acoustic index that correlates with the perceived accuracy of vowels \[16\]. Comment 8: line 340 : It -\> F2-Slope Response 8: Thank you for your valuable suggestion. We have corrected it accordingly. Page 25–26, Lines 402–404 The F2 slope is speculated to reflect the speed of the tongue movement during articulation. In this study, the correlation between the F2 slope and tongue muscle strength was similar to all speakers with dysarthria in the sex-separated subgroups. 10.1371/journal.pone.0264995.r003 Decision Letter 1 Finley Sara Academic Editor 2022 Sara Finley 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. 10 Feb 2022 PONE-D-21-32058R1Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthriaPLOS ONE Dear Dr. Tamura, 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 Mar 27 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, Sara Finley, Ph.D. 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): Both original reviewers have read the revised manuscript and believe that their concerns have been met in the revision. However, both reviewers have minor suggestions for clarification and wording. These are important in order to be sufficiently clear for a broad audience. I also have some minor comments related to clarity and wording: Line 120: "Comprised 20 participants" should be 'comprised of 20 participants" Line 130: "highly guaranteed" seems somewhat superlative. I would suggest toning this down, or noting that the measurement is believed to be reproducible and reliable Lines 206-207- "The F2 was determined or the..." This sentence is somewhat confusing to read. Please make sure that capitalization is consistent. Line 319: "Bold value indicates and" This sentence is confusing. It seems that a word is missing here. Value should be 'values'? Line 333. Make sure to have consistent capitalization for 'Lower'. \[Note: HTML markup is below. Please do not edit.\] 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: (No Response) 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 \#1: Yes Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: Yes 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 \#1: Yes 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 \#1: Yes 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 \#1: Relationships between maximum tongue pressure and second formant transition in　speakers with different types of dysarthria General comments: This paper has been properly revised from the first manuscript. But a part of your manuscript remain a little hard to understand correctly. Could you look at the point below: Material and Methods 1\. P12 L202-207 “Fig 2. A spectrogram of the word…. corresponding F2 frequency change, Extent.” In Figure 2, the analysis methods of F2 onset and F2 offset for /ai/ and /jo/ were unclear. Please explain the detailed definition of F2 onset and F2 offset for /ai/ and /jo/. 2\. P12 L206-207 The F2 slope was determined or the time interval, Duration, and the corresponding F2 frequency change, Extent. → The F2 slope was determined or the time interval, duration, and the corresponding F2 frequency change, extent. Reviewer \#2: The authors have adequately addressed the given comments raised in a previous round of review. I feel confident that the manuscript has become much clearer. The manuscript is technically sound, and the data support the conclusions. The statistical analysis has been performed appropriately and rigorously (Mann-Whitney U test - assessment of the difference between dysarthric speakers vs healthy speakers; Pearson correlation - correlations between MTP and speech-related features (speech intelligibility, /ta/ DDK, F2-slope) within speakers with dysarthria. However, there are some minor comments for the improvements of clarity. \# Abstract line 31: Some speakers with dysarthria -\> Speakers with dysarthria \- The paper reports the significant difference between dysarthric speakers and healthy speakers. To state as 'some speakers' seems unnecessary. line 38: This result suggests that the maximum isometric tongue strength is associated with tongue movement speed during articulation. \- oral DDK is also known to represent the tongue movement speed during articulation. Elaborating on the difference between oral DDK and F2-slope may help the readers better understand the paper's main contribution. \# Introduction line 49-50: does speech clarity align with speech intelligibility? Please elaborate on this terminology. It is difficult to relate 'severity' with speech- related indicators (compared to articulation rate, oral-DDK rate). Please elaborate. line 60: overall severity (of dysarthria? muscle strength?) line 62: what does /ta/\<5.8 syllables/s infer? Stating the oral-DDk rate for healthy speakers or mild dysarthria is necessary. line 63, line 67: please explain what speech audibility is. The authors are using 3 terminologies in the manuscript - speech intelligibility, speech clarity, speech audibility. line 71 : F2 slope -\> F2 (second formant) line 77-78: relatively slow changes in tongue shape. Please elaborate. How is tongue shape related to the decrease in the F2-slope? In addition, tongue shape seems not to be related to movement speed, which is the essence of this paper. \# Materials and Methods - Participants \- Information of dysarthric speakers (how many speakers, Gender, Age, etc) is omitted. line 115: please account for the reason why height, weight, albumin, BMI are collected for this study. \# Materials and Methods - Oral-DDK rate line 173: 3s of the central parts: Please explain why the authors decided to use only the central parts of the audio, rather than the full audio. line 192: Specific explanation of how the authors determined the onset and offset of F2 is needed for replication. \# Materials and Methods - Second formant transition fig 2 - fast formant \> first formant \# Results line 301-302 : Correlation between MTP & speech intelligiblity, MTP & /ta/ DDK by sex groups should also be presented by gender. line 311- 314: Descriptions for the analysis of other subtypes are missing. (hypokinetic, ataxic, etc.) \# Discussion line 337-347: The fact that the authors analyzed the correlations between MTP and speech-related indicators within dysarthric speakers does not stand out in this paragraph. line 388-390: According to the results, speech intelligibility and oral-DDK measurements did not show significant correlations with MTP. line 407: some speakers with dysarthria - please concretely describe the corresponding speakers. \# Conclusion line 458: the appropriateness of articulation and speed \*\*\*\*\*\*\*\*\*\* 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 \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.\] 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.0264995.r004 Author response to Decision Letter 1 18 Feb 2022 Point-by-point responses to comments from the Editor and Reviewers Thank you for reviewing our work. We appreciate all your comments and suggestions. We have revised the manuscript accordingly. Our point-by-point responses are presented below. Response to Journal Requirements Comment 1: 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 1: Thank you for your suggestion. The doi of references 19 and 41 in the reference list were incomplete and have been corrected. The retracted papers are not cited in this paper. Other minor corrections have also been made. Response to the editor General comments: Both original reviewers have read the revised manuscript and believe that their concerns have been met in the revision. However, both reviewers have minor suggestions for clarification and wording. These are important in order to be sufficiently clear for a broad audience. I also have some minor comments related to clarity and wording:. Response: We appreciate the reviewer for their constructive comments. The authors’ responses to the comments are as follows: Comment 1: Line 129: "Comprised 20 participants" should be 'comprised of 20 participants" Response 1: Thank you for the detailed confirmation. Our manuscript was submitted for professional English editing and proofreading, and we were informed that “comprised of” is a grammatically incorrect phrase. We hope that retaining the original phrase here will not be an issue.. Taken from the Merriam-Webster website: “Although comprised of is an established standard for "being composed or constituted of," it is often liable to criticism and scrutiny. The correct version put forward by grammar guides is to use "composed of" or "comprises" such as "the cake is composed of flour and eggs" or "comprises flour and eggs." Comment 2: Line 139: "highly guaranteed" seems somewhat superlative. I would suggest toning this down, or noting that the measurement is believed to be reproducible and reliable Response 2: Thank you for your helpful suggestion. Accordingly, we have made the following revision in the sentence: Page 8, Line 139 The reproducibility and reliability of this device has been validated in a previous study \[26\]. Comment 3: Lines 225-226: "The F2 was determined or the..." This sentence is somewhat confusing to read. Please make sure that capitalization is consistent. Response 3: Thank you for your helpful suggestion. We have made the following revision in the sentence: Page 13, Lines 225-226 Duration is the time interval of the F2 movement. Extent is the range of frequency change in the F2 movement. Comment 4: Line 344: "Bold value indicates and" This sentence is confusing. It seems that a word is missing here. Value should be 'values'? Response 4: Thank you for your helpful suggestion. We have made the following revisions in the text: Page 22, Line 344: Values in bold and with asterisks indicate that rs is significant (p \< 0.05) (two-tailed). Comment 5: Line 357: Make sure to have consistent capitalization for 'Lower'. Response 5: Thank you for your helpful suggestion. Accordingly, we have removed the capitalization from the sentence: Page 23, Lines 357-359 　　　　Speakers with lower maximum tongue pressure and dysarthria (lower group) tended to have lower speech-related variables, such as lower maximum tongue pressure. Reviewer \#1: General comments: This paper has been properly revised from the first manuscript. But a part of your manuscript remain a little hard to understand correctly. Could you look at the point below: Response: Thank you for your supportive comments and valuable suggestions for improving the quality of our manuscript. The authors’ responses to the comments are as follows: Material and Methods Comment 1: P13 L221-225 “Fig 2. A spectrogram of the word…. corresponding F2 frequency change, Extent.” In Figure 2, the analysis methods of F2 onset and F2 offset for /ai/ and /jo/ were unclear. Please explain the detailed definition of F2 onset and F2 offset for /ai/ and /jo/. Response 1: Thank you for comment. We have adjusted the sound spectrogram in Fig. 2 to improve clarity regarding the whole picture. We have also made the following revisions in the text: Page 12, Lines 205–212 The F2 linear predictive coding tracks on a wide-band spectrogram (analysis bandwidth, 300 Hz) were manually edited and identified. The 20/20 rule (specifying a frequency change of ≥20 Hz during 20 ms in the transition onset and offset) was applied \[41\]. If the F2 track was unclear due to hoarseness or other noises, we identified it by referring to the changes in F1 (opening and closing of the mandible) that were almost synchronized with F2 in the /ai/ and /jo/ sequences. Thus, relatively stationary portions of the vowel before or after the target transition were not included in the analysis. Page 13, Lines 226–228 The F2 movement for each syllable is determined by a change in frequency of at least 20 Hz during a 20 ms period, not including the relatively stationary portions before and after the syllable. Comment 2: P12 L225-226 The F2 slope was determined or the time interval, Duration, and the corresponding F2 frequency change, Extent. → The F2 slope was determined or the time interval, duration, and the corresponding F2 frequency change, extent. Response 2: Thank you for your helpful suggestion. We have revised this for more clarity. Page 13, Lines 225-226 Duration is the time interval of the F2 movement. Extent is the range of frequency change in the F2 movement. Reviewer \#2 General comments: The authors have adequately addressed the given comments raised in a previous round of review. I feel confident that the manuscript has become much clearer. The manuscript is technically sound, and the data support the conclusions. The statistical analysis has been performed appropriately and rigorously (Mann- Whitney U test - assessment of the difference between dysarthric speakers vs healthy speakers; Pearson correlation - correlations between MTP and speech- related features (speech intelligibility, /ta/ DDK, F2-slope) within speakers with dysarthria. However, there are some minor comments for the improvements of clarity. Response: Thank you for your supportive comments and suggestions for the improvement of the quality of our manuscript. The authors’ responses to the comments are as follows: \# Abstract Comment 1: line 31: Some speakers with dysarthria -\> Speakers with dysarthria \- The paper reports the significant difference between dysarthric speakers and healthy speakers. To state as 'some speakers' seems unnecessary. Response 1: Thank you for your helpful suggestion. We have incorporated your suggestion in our paper. Comment 2: line 39: This result suggests that the maximum isometric tongue strength is associated with tongue movement speed during articulation. \- oral DDK is also known to represent the tongue movement speed during articulation. Elaborating on the difference between oral DDK and F2-slope may help the readers better understand the paper's main contribution. Response 2: Thank you for your helpful suggestion. We have made the following revision in the sentence: Page 3, Lines 37–39 The oral diadochokinesis rate, which is related to the speed of articulation, is affected by voice on/off, mandibular opening/closing, and range of motion. In contrast, the second formant slope was less affected by these factors. \# Introduction Comment 3: line 51-53: does speech clarity align with speech intelligibility? Please elaborate on this terminology. It is difficult to relate 'severity' with speech-related indicators (compared to articulation rate, oral-DDK rate). Please elaborate. Response 3: Thank you for your question. -Yes, speech clarity is the same as speech intelligibility. For more clarity, the term “speech intelligibility” has been used consistently in the paper. In addition, “severity” has been deleted because it includes impressions other than speech. We have made the following revision in the sentence: Page 3, Lines 51–53 However, current reviews have reported no significant relationship between tongue strength and speech-related indicators, such as speech intelligibility, articulation rate, and oral-DDK rate \[6–8\]. Comment 4: line 63: overall severity (of dysarthria? muscle strength?) Response 4: Thank you for your comment. Overall severity is an auditory measure of dysarthria, and corresponds to a comprehensive assessment of speech intelligibility and naturalness. We have added the following to the sentence to make it clearer: Page 4, Lines 61–64 Additionally, speakers with dysarthria in whom tongue muscle strength is lower than the lower limit of normal speakers have moderate-to-severely reduced articulatory precision and overall severity (including speech intelligibility and naturalness) \[13\]. Comment 5: line 65: what does /ta/\<5.8 syllables/s infer? Stating the oral-DDK rate for healthy speakers or mild dysarthria is necessary. Response 5: Thank you for your valuable suggestion. We have made the following additions: Page 4, Lines 66–67 In a previous cross-sectional study \[13\], dysarthria speakers (n = 8) with severe anterior tongue elevation muscle strength had an oral-DDK rate of \<5.8 syllable/s for the syllable /tʌ/. In contrast, 44.6% of the remaining speakers with dysarthria had an oral-DDK rate of \>5.8 syllables/s. Comment 6: line 68, line 72: please explain what speech audibility is. The authors are using 3 terminologies in the manuscript - speech intelligibility, speech clarity, speech audibility. Response 6: Thank you for pointing this out. We were using these terms in the same context. Therefore, we have unified all three terms into “speech intelligibility.” Page 3, Lines 53; Page 4, Lines 68, 72; Page 5, Lines 73 Comment 7: line 76 : F2 slope -\> F2 (second formant) Response 7: Thank you for the detailed confirmation. We have defined the second formant as F2 in the sentence preceding this one(line 74). We hope that retaining the original phrase here will not be an issue.. Comment 8: line 82-83: relatively slow changes in tongue shape. Please elaborate. How is tongue shape related to the decrease in the F2-slope? In addition, tongue shape seems not to be related to movement speed, which is the essence of this paper. Response 8: Thank you for your comment. The tongue is a muscle without any joints. Therefore, “relatively slow changes in tongue shape” has the same meaning as “slow movement of the tongue.” In other words, the back-and-forth motion of the tongue during articulation is slower, resulting in a smaller F2 slope. To make it clearer, we have added the following sentence. Page 5, Lines 81–85 The clear explanation for the decrease in the F2 slope in speakers with dysarthria is the relatively slow changes in tongue shape \[22\]. Specifically, the back-and-forth motion of the tongue during articulation is slower and/or the range of movement is narrower, resulting in a longer and thinner change in the F2 movement. \# Materials and Methods - Participants Comment 9: Information of dysarthric speakers (how many speakers, Gender, Age, etc) is omitted. line 120: please account for the reason why height, weight, albumin, BMI are collected for this study. Response 9: Thank you for your valuable suggestion. These factors were considered to account for the possible effects on tongue muscle strength and speech caused by factors other than the primary disease causing dysarthria. The following sentences have been added to the manuscript. Page 7, Lines 122–124 　　　　These factors were considered to account for the possible effects on tongue muscle strength and speech caused by factors other than the primary disease causing dysarthria. \# Materials and Methods - Oral-DDK rate Comment 10: line 183: 3s of the central parts: Please explain why the authors decided to use only the central parts of the audio, rather than the full audio. Response 10: Thank you for your valuable suggestion. We used a 3-second duration of the middle part of the audio to reduce the effects of speech irregularities during speech onset and respiratory dysfunction during the second half of the task. In the syllable repetition task, the first syllable is often uttered longer than the following syllable in normal subjects (Ackerman et al, 1995). In speakers with dysarthria, it is possible that the initial effect is due to a freezing at movement onset (e.g., hypokinetic type). In addition, speakers with dysarthria are more likely to be affected by breath-holding in the second half of oral-DDK due to reduced respiratory function. The following sentences have been added to the manuscript. Page 11, Lines 183–186 　　　　To reduce the effects of speech irregularities such as freezing, slurring, or syllable prolongation during speech onset, or respiratory dysfunction during the second half of the task, we extracted \~3 s of recorded data from the middle parts of the audio for the analysis. Comment 11: line 204: Specific explanation of how the authors determined the onset and offset of F2 is needed for replication. Response 11: Thank you for your comment. Based on your suggestion, I added the method of determining the onset and offset. Page 12, Lines 205–212 The F2 linear predictive coding tracks on a wide-band spectrogram (analysis bandwidth, 300 Hz) were manually edited and identified. The 20/20 rule (specifying a frequency change of ≥20 Hz during 20 ms in the transition onset and offset) was applied \[34\]. If the F2 track was unclear due to hoarseness or other noises, we identified it by referring to the changes in F1 (opening and closing of the mandible) that were almost synchronized with F2 in the /ai/ and /jo/ sequences. Thus, relatively stationary portions of the vowel before or after the target transition were not included in the analysis. \# Materials and Methods - Second formant transition Comment 12: fig 2 - fast formant \> first formant Response 12: Thank you for pointing this out. We have corrected this. \# Results Comment 13: line 322-323 : Correlation between MTP & speech intelligibility, MTP & /ta/ DDK by sex groups should also be presented by gender. Response 13: Thank you for your valuable suggestion. We have added the following sentence: Page 20, Lines 324-326 The correlation between MTP and speech intelligibility was significant only in males (rs = -0. 328, p = 0.030). There was no significant correlation between MTP and the /ta/ DDK rate in either sex. Comment 14: line 335- 337: Descriptions for the analysis of other subtypes are missing. (hypokinetic, ataxic, etc.) Response 14: Thank you for pointing this out. We have added the following sentence: Page 21, Lines 337–338 There was no significant correlation between any of the combinations in the other subtypes. \# Discussion Comment 15: line 362-373: The fact that the authors analyzed the correlations between MTP and speech-related indicators within dysarthric speakers does not stand out in this paragraph. Response 15: We have added the following sentence: Page 23, Lines 363–364 In this study, we measured the tongue pressure in speakers with various types of dysarthria, following which, we conducted a correlation analysis between MTP and speech-related indicators.. Comment 16: line 414–417: According to the results, speech intelligibility and oral-DDK measurements did not show significant correlations with MTP. Response 16: Thank you for pointing this out. We have made the following revision: Page 26, Lines 414–417 　　　　In our results, speech intelligibility and oral-DDK rate were not significantly correlated with MTP in all subtypes. However, the flaccid type (n = 7) had the highest correlation coefficient among all types (speech intelligibility, rs = -0.436; /ta/ DDK rate, rs = 0.517). Additional subtype- specific studies are warranted. Comment 17: line 432: some speakers with dysarthria - please concretely describe the corresponding speakers. Response 17: Thank you for your valuable suggestion. We have reconsidered the results of this study and believe that it is not possible to determine the specific characteristics of the speakers. Therefore, we have deleted “some” and modified the sentence as follows: Page 27, Lines 432–435 In summary, in speakers with dysarthria, tongue strength may be associated with tongue movement velocity during articulation. Our findings partially support the hypothesis that “muscle weakness is associated with slow speech” \[2\]. \# Conclusion Comment 18: line 483: the appropriateness of articulation and speed Response 18: Thank you for your suggestion. We have added this to the text. 10.1371/journal.pone.0264995.r005 Decision Letter 2 Finley Sara Academic Editor 2022 Sara Finley 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 Feb 2022 Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthria PONE-D-21-32058R2 Dear Dr. Tamura, 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, Sara Finley, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Thank you for carefully revising and considering each comment from the editors and reviewers. I believe that this revision has successfully addressed all concerns and the paper should be published. Reviewers' comments: N/A 10.1371/journal.pone.0264995.r006 Acceptance letter Finley Sara Academic Editor 2022 Sara Finley 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 Feb 2022 PONE-D-21-32058R2 Relationships between maximum tongue pressure and second formant transition in speakers with different types of dysarthria Dear Dr. Tamura: 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. Sara Finley Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Prior to 2003, H7 subtype avian influenza virus causing human infection cases were very rare and mostly caused by occupational accidents or laboratory exposures. For the last decade, more than 500 cases of human infections with H7 subtypes have been documented. Some of which include the H7N7 (NL/219/03) virus outbreak (89 human infected, one died) in the Netherlands and H7N9 outbreak (more than 440 human infected, out of which with 155 fatal cases) in China. H7N7/NL/219/03 virus is a highly pathogenic avian influenza that can infect both mice and ferrets without prior adaptation. Additionally, the H7N7/NL/219/03 viral attachment pattern and replication efficacy in mammalian respiratory tracts showed great similarity to H5N1 viruses. Moreover, the replication patterns resembled that of H5N1 virus. The broad host spectrum, unusually high zoonotic potential, as well as its ability to suppress host immune responses in a similar way to 1918 H1N1 virus are raising concerns for potential future influenza pandemics. Hence, the vaccine development against H7N7/NL/219/03 is of high priority to our defense against any possible H7 pandemic. In our previous vaccine study, mice that were intranasally immunized with live Bac-HA were protected from lethal H7N7 viral challenge. However, no protection was observed when Bac-HA or inactivated H7N7 virus were administered subcutaneously, possibly due to diminished immunogenic nature of the H7N7 (H7N7/NL/219/03) virus. The effect of glycan shielding on H7N7HA protein could be another plausible explanation too. Previous studies have reported that glycosylation of HAs could result in poor neutralizing antibody titer. Both influenza and human immunodeficiency viruses (HIV) were found to employ glycan masking as a strategy for blocking antibody-epitope interactions. Several studies have also explained the impact of HA glycosylation on the antigenicity, pathogenicity and evolution of influenza virus. Interestingly, in our recent vaccine study, mice that were subcutaneously immunized with live Bac-HA (H7N9) survived in both H7N9 and H7N7 virus challenge. Comparing H7N7HA1 and H7N9HA1 amino acid sequences, there were 15 amino acid positions differ were identified. Among these 15 positions, in this study we selected three positions, namely (i) 143, (ii) 198 and (iii) 211,(numbering of amino acid on HA sequence starts from ‘‘ATG” and includes the signal peptide) of the H7N7HA1 protein. These three positions are located within or near the receptor binding site of H7N7HA protein. Moreover, amino acid threonine at position 143 of NL/H7N7HA generated a potential N-linked glycosylation at position 141 of the H7N7 HA protein. Following the selection, we generated a total of six mutant constructs by amino acid substitution at these three positions either singly (T143A, T198A and I211V) or doubly (T143A-198A, T198A-I211V and I211V-T143A) to the corresponding amino acids found in H7N9HA protein by site directed mutagenesis. The H7N7HA wild-type and all the H7HA mutants were expressed on the surface of baculovirus via Baculovirus Expression Vector System (BEVS). These HAs were further immunized subcutaneously into BALB/c mice, before they were intranasally challenged with mouse adapted HP RG-H7N7 virus. The immune responses and protections in BALB/c mice were then subsequently monitored to evaluate the efficacy of these expressed HAs as vaccines. # Materials and Methods ## Ethics statement All animal experimental protocols were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of the Temasek Life Sciences Laboratory, National University of Singapore, Singapore. (IACUC approval number TLL-14-020). All animal experiments were carried out in animal BSL3 containment facility in compliance with CDC/NIH and WHO recommendations. Mice were housed in individually ventilated cages (Tecniplast Sealsafe) provisioned with water and standard food, and monitored daily for health condition. More than 25% body weight loss was used as criterion for early euthanasia. The animals were euthanized by CO₂ inhalation for five minutes. ## Viruses and cells Reassortant influenza virus RG-H7N7 (A/Netherland/219/03) and RG-H7N9 (A/Shanghai/2/2013) were generated by reverse genetics as described previously. Viruses were propagated in 10 day old specific pathogen free embroyonated chicken eggs at 37°C. The tissue culture infectious dose 50 (TCID₅₀) of reassortant viruses were calculated by the Muench-Reed method (1938). All experiments were performed in a biosafety level 3 (BSL-3) containment laboratory in compliance with CDC/NIH and WHO recommendations. Spodoptera frugiperda (Sf9II) cells (ATCC) were maintained in Sf900II serum free medium (Gibco BRL, USA) at 28°C for wild-type baculovirus (wt-Bac) and recombinant baculovirus production. ## Preparation of recombinant baculovirus subunit vaccine The full length of H7N7HA (A/NL/219/03) cDNA gene was commercially synthesized based on the sequences from the NCBI influenza database (GenScript, USA). cDNA was used for Bac-HA and Bac-HA_m constructs preparation. For preparing the mutant constructs (single mutants T143A, T198A, I211V, double mutants T143A-T198A, T198A-I211V and I211V-T143A), we substituted the amino acid alanine at 143, alanine at 198 and valine at 211 coding gene inH7N7HA wild-type cDNA by using quick change site directed mutagenesis kit (Agilent Technologies). Transfection and viral amplification were carried out according to the Baculovirus Expression System Manual (Invitrogen Life technologies).The live baculovirus content of each vaccine constructs were determined by TCID50 method as described Reed and Munch (1938). HA content of the each vaccine construct was determined densitometrically by comparing against known concentrations of the insect cells expressed H7N7HA protein(Sino Biologic, Beijing, China) using ImageJ (National Institutes of Health) as described Miriam et al.(2014). ## Immunofluorescence assays to detect HA expression in insect cells The Sf9 cells were infected with Bac-HA or Bac-HA_m constructs or wt- Bac at 0.5 MOI. Two days post infection, the cells were fixed with 4% formaldehyde and stained with mouse anti H7N7-HA monoclonal antibody (8H9), followed by FITC conjugated goat anti mouse immunoglobulin (Dako Cytomation, Denmark). The fluorescence signal was detected with an inverted fluorescence microscope (Olympus, UK) and the images were captured by a digital imaging system (Nikon, USA). ## Glycosylation pattern confirmation by western blot analysis Amino acid substitution at T143A in H7N7HA wild-type protein inhibits synthesis of N- glycosylation motif in A141. It was confirmed by western blot analysis. Briefly, the Bac-HA or Bac-HA mutant infected cell supernatant was mixed with Laemmli sample buffer and resolved in 12% SDS–PAGE. The gel was transferred to a nitrocellulose membrane and Western blotting was performed as described previously Kumar et al (2013).The anti-mouse rHA polyclonal antibody (TLL, Singapore) at a dilution of 1:250 was used as the primary antibody and rabbit anti-mouse Ig (Dako Cytomation, Denmark) at a dilution of 1:3000 was used as a secondary antibody. The protein bands were visualized by chemiluminescence kit (Amersham, UK). ## Mice immunization Six to eight weeks old specific pathogen free female BALB/c mice were used in all experiments. Mice (10 mice per group/9 groups) were subcutaneously immunized with 100 μL of purified live Bac-HA or Bac-HA_m constructs (T143A or T198A or I211V, T143A-T198A or T198A-I211V or I211V-T143A) containing 2.5μg of HA content (approximately1 X10⁷to 1X10^7.5 pfu/ml recombinant Baculovirus—Live baculovirus content of each recombinant constructs were determined by using TCID50 method as described Reed and Munch (1938)) or wt-Bac (10⁸ PFU) and PBS control. Four weeks after the first immunization, a second dose of vaccine was given to all immunized mice groups. Two weeks after the second immunization (42^ndday), serum samples were collected from each experimental mice group (Five mice/Group). All samples were stored at -20°C for measuring the hemagglutination inhibition (HAI) and microneutralization (MN) assay against RG-H7N7, RG-H7N9 viruses. ## Microneutralization (MN) assay The MN assay was performed as described previously. Briefly, serial two fold dilutions of heat inactivated (56°C for 45 min) 42nd day immune serum samples were mixed separately with 100 TCID₅₀ of RG-H7N7 or RG-H7N9 viruses and incubated at room temperature for 1 h. The mixtures were added to the MDCK monolayers in triplicate wells. One hour after infection, the virus-serum mixture was removed, and serum free DMEM was added to each well. Cytopathic effect was observed 3 to 4 days after incubation at 37°C. The highest serum dilutions that completely protected cells from cytopathology were considered to be the MN titer. ## Hemagglutination inhibition (HAI) assay Functional serum antibody titers were determined by hemagglutination inhibition assay.Serum samples were treated with 4 volumes of a receptor-destroying enzyme of *Vibrio cholera* filtrate (RDE Denka Siken Co., Japan) for 18 h at 37°C. After addition of 3 volumes of 2.5% (v/v) sodium citrate, the serum samples were incubated at 56°C for 30 min and diluted with PBS to yield a 1:10 dilution of the original serum sample. Serum samples were 2-fold serially diluted in PBS (25 μL sample volume) in Nunc 96-well polystyrene V-bottom microwell plates (Thermo Fisher Scientific, Denmark). Approximately four HA units of each H7 subtype (RG-H7N7 or RG-H7N9) viral antigen was incubated with the serum for 30 min at room temperature followed by the addition of 1% chicken RBCs and incubation at room temperature for 40 min. The inhibition of hemagglutination at the highest serum dilution was considered the HAI titer of the serum. ## Mice viral challenge To assess the protective efficacy of the vaccines, immunized mice groups were anesthetized intraperitoneally with ketamine (100 mg/kg)/Xylazine (20 mg/kg) and intranasally challenged with 50 μL (25 μL per naris) of 5 MLD₅₀ of mouse adapted RG-H7N7 virus. Mice were observed daily to monitor body weight, clinical signs of disease (ruffled fur, lack of mobility, labored breathing) and mortality. Mice were humanely euthanized with CO₂ inhalation if their body weight dropped to 75% of baseline weights. For determination of lung viral titers after challenge, three mice from each vaccinated group was euthanized by CO₂ inhalation on day 5 post-challenge as described Prabakaran et al.. ## B cell and T cell epitope prediction The immune epitope Database (IEDB) analysis resources were used for predicting the B—cell and T-cell epitopes in hemagglutinin molecule of A/Netherland/219/2003 and H7N9 A/Shanghai/2/2013 viruses for searching the cross reactive epitopes between the viruses. ## Statistical analysis All data are expressed as arithmetic mean ± standard deviation. The significance between the groups was calculated using unpaired two-tailed Student’s t test and significance was expressed as \*\* *P*\<0.001. # Results ## Baculovirus-HA vaccine construction The Bac-HA and other Bac-HA_m protein expressions were analyzed by immunofluorescence staining and their native structure were analyzed by hemagglutination activity with Chicken RBCs. Fluorescence signal and agglutination activity (32 HA units) were detected only in baculovirus expressed H7N7HA (Bac-HA or Bac-HA_m constructs) infected cells. In contrast, no fluorescence signal and agglutination activity was observed in cells infected with wt-Bac. ## Western blot confirmation of HA protein with lacking of N-linked glycosylation Amino acid substitution at position 143T inhibits synthesis of N linked glycosylation motif in insect cell expressed H7NHA protein which was confirmed by western blot through molecular weight differentiation. The result showed that molecular weight of single mutated (T143A) or double mutated (T143A-T198A, T143A-I211V) HA protein was reduced (approximately 64–66 kDa) compared to wild- type HA protein (approximately 67–70kDa). However, other mutated HA protein (T198A or I211V or T198A-I211V) did not show any differences in molecular weight. This shows that amino acid substitution at T143A result in lack of N-glycosylation motif in A141. ## B cell and T cell epitopes prediction We predicted B-cell and T-cell epitope in HA molecule of H7N7 and H7N9 viruses by using immune epitope analysis resources as described previously. Here, we showed B-cell and T-cell epitope located in the mutant region (143,198 and 211) of HA molecule. Based on the epitope prediction result, both viruses have a B-cell epitope at region 138–150 (T143A), 194-205(T198A) and T-cell epitope at position 203–211. ## Systemic immune response The functionality of serum antibodies was tested in hemagglutination inhibition (HAI) assay. Sera of mice immunized s.c. with Bac-HA_m construct I211V-T143A exhibited slightly enhanced HAI titer but no significant difference compared to mice immunized with T143A or T198A-I211V Bac-HA_m construct. However, significant difference was observed between mice immunized with Bac-HA_m construct T143A or I211V-T143A or T143A or T198A-I211V groups and Bac-HA immunized mice groups. Further, the ability of the serum antibodies to neutralize live influenza virus was tested in a microneutralization assay. The neutralizing antibody titers of sera from s.c. immunized mice groups at day 42 after the first immunization were measured against 100 TCID₅₀ of homologous H7N7 reassortant virus. The sera obtained from the s.c. vaccinated Bac-HA_m construct T143A (1:208), T198A-I211V and I211V-T143A showed a higher neutralizing titer (1:256) against RG-H7N7virus compared to other vaccinated groups. Also, we tested cross- neutralization activity of vaccinated sera against RG-H7N9 virus. Interestingly Bac-HA_m construct T143A(1:128) or T198A-I211V (1:112) vaccinated group showed a higher neutralization activity against H7N9 compared to other vaccinated groups but no significant difference compared to mice immunized with T143A-I211V. Bac-HA vaccinated mice group showed significantly low level of neutralization activity to H7N7 (1:28) or H7N9 (1:10) RG viruses compared toT143A or T198A-I211V or T143A-I211V Bac-HA_m construct vaccinated group. ## Mice challenge study The mice viral challenge study result showed that s.c. immunized with Bac- HA_m construct T143A or T198A-I211V or I211V-T143A mice groups achieved 100% protection throughout the 14 day observation period with a negligible weight loss. Other Bac-HA_m constructs T198A (33.33%), I211V (83.33%), and T143A-T198A (50%) also give partial protection against lethal H7N7 challenging, but Bac-HA administered by s.c. route, failed to protect the mice. These animals showed obvious clinical symptoms such as ruffled fur and hunching shoulders which started at day 4 post challenges. Also, a significant weight loss of up to 25% to 35% was observed before they were sacrificed at day 10. Also, the mice lung viral titer result showed that the H7N7 influenza virus titer was significantly reduced in the s.c. immunization with Bac-HA_m construct T143A or I211V or T198A-I211V or I211V-T143A groups compared to other Bac-HA_m vaccinated or Bac-HA vaccinated groups. # Discussion Avian influenza viruses, particularly H7 subtype, have been responsible for large outbreaks of diseases among poultry, as well as in human beings, in recent years. In our previous vaccine study, mice subcutaneously immunized with live baculovirus or inactivated RG-H7N7 virus did not provide protection against the lethal H7N7 influenza virus. Such lack of protection could be due to the less immunogenic nature of H7N7 virus. Additionally, several previous studies have suggested that amino acid substitutions in critical region of the antigen improve immunogenicity. In our studies, the amino acid position 143, 198 and 211 of H7N7HA protein were selected and substituted, either singly or doubly, to the corresponding amino acids found on H7N9HA wild-type sequence using site directed mutagenesis. Six H7N7 HA mutant constructs, namely the single mutants: (i) T143A, (ii) T198A and (iii) I211V, and the double mutants (iv) T143A-T198A, (v) T198A-I211V, (vi) I211V-T143A) were generated. All these constructs were expressed on the baculovirus surface using insect cell line and the protective efficacy of these vaccine constructs were evaluated using a mouse model, where the mice were challenged with RG-H7N7 viruses. The HAI activity of the sera derived from mice vaccinated against H7N7 virus were subsequently tested. The sera of the mice vaccinated with Bac- HA_m constructs T143A, I211V, T198A-I211V and T143-I211V showed higher HAI titers when compared to that of mice vaccinated with other Bac- HA_m constructs. These higher HAI titer observed could be the result of the respective mutations being located within or near the receptor binding site, or even in antigenic dominant region of the HA. In support of our explanation, Medina et al. has previously reported that adding one or two glycosylation on residues around dominant antigenic region was sufficient to significantly reduce the HAI activity of the polyclonal response elicited by the wild-type unglycosylated virus. This higher HAI titer was also true when we tested the neutralizing activity of sera obtained from mice immunized with Bac-HA_m constructs. The result showed Bac-HA_m constructs (T143A, T198-211, and I211V-T143A) had significantly higher neutralizing activity against homologous RG-H7N7 virus when compared to the sera from mice vaccinated with Bac-HA and other Bac- HA_m. This higher neutralizing response against H7N7 virus could be attributed to the amino acid substitution at position 143, which inhibits the N-linked glycosylation with Asn141 in mutant HA (143A) protein. The lack of glycosylated motif may, in turn, expose the hidden neutralizing epitope at region 138–150, thus, inducing a higher neutralizing antibody titer. On the contrary, wild-type HA protein has the potential to glycosylate at position 141, creating a motif region 138–150, thereby hiding the neutralization epitope. This, therefore, leads to a low neutralizing antibody titer. Supporting our hypothesis, Ma et al. has reported that partially deglycoslated envelope would bind better than fully glycosylated gp41 envelope to specific naıve B cells, hence improving immunogenicity. Interestingly, the sera of the Bac- HA_m construct I211V vaccinated mice also showed higher HAI and neutralization titer against H7N7 virus than that of Bac-HA. The exact mechanism as to how H7N7HA mutant (211V) induces higher HAI and neutralization titer is yet to be known and we hope that the mechanism could be elucidated in our future studies. A significant association between complete protection against RG-H7N7 influenza virus and titers of HI and neutralization antibody was observed. Previous study has also shown that higher neutralizing antibody and HI titer does protect mice from H7N9, hence, supporting the association seen in our experiments. Cross neutralizing activity was analyzed by MN assay against virus of H7 subtype H7N9. Sera of mice immunized with Bac-HA_m constructs T143A, T198A-I211V and T143A-T211A showed notably higher neutralizing activity against RG-H7N9 virus than those sera drawn from mice vaccinated using Bac-HA and other Bac- HA_m. Based on the epitope prediction result (, courtesy: Larisa Rudenko et al 2013), both H7N7and H7N9 viruses have a B-cell epitope at region 138–150 and T-cell epitope at position 203–211. The difference between the amino acid sequence of the B-cell epitope of H7N9 and H7N7 is at position 143. Amino acid Thr at position 143 of NL/H7N7HA generated an N-linked glycosylation with Asn141, creating a glycan mask on the predicted B-cell epitope located at position between 138 and 150. H7N9 virus, on the other hand, have alanine at position 143, and this inhibits the N-linked glycosylation. The unglycosylated motif of H7N7HA_m (T143A) protein, similar to those found in H7N9 virus, results in the exposure of the B-cell epitope, between region 138 and 150. This explains why antiserum of mice vaccinated with H7N7_m constructs has higher cross neutralizing activity against H7N9 virus. Based on T-cell epitope prediction, H7N7 and H7N9 viruses, have a T-cell epitope at position 203–211 of their HA proteins. One difference between the amino acid sequence of H7N9 and H7N7 T-cell epitope is at position 211. Several previous studies have suggested that T-cell epitope containing amino acid valine at position 9 acts as a dominant C-terminal anchor residue that provides higher binding affinity to MHC, hence inducing greater antigen-specific T cell immune responses. Similarly, H7N7HA mutant I211V carrying the amino acid valine (at position 9 in its T cell epitope) may provide such greater MHC binding affinity, thereby, inducing higher peptide-specific T-cell immune responses. This might be the reason for the cross-protection against H7N9 influenza virus when immunized with I211V of H7N7HA. The protective efficacy of vaccine candidates against RG-H7N7 influenza virus was determined in a mouse challenge study. In this study, subcutaneous vaccinations using Bac-HA_m constructs T143A, I211V-T143V and T198A-I211V completely protects the mice that were challenged with 5MLD50 of RG-H7N7 influenza virus. The complete protection observed could be the effect of higher HAI titer and neutralization antibody titer against H7N7 influenza virus. In addition to the 3 constructs just mentioned, vaccinations using 2 other Bac- HA_m constructs T198A and T143A-T198A have provided 33.3% and 50% protection respectively. Moreover, immunization of mice using mutant (I211V) HA construct provides sufficient protection (83.33%) when challenged with RG-H7N7. Such reasonable protection could be the effects of higher T-cell responses against H7N7HA. In contrast, no protection was observed in mice vaccinated with Bac-HA, likely a result of the significantly lower level of HAI titer induced by subcutaneous vaccination with Bac-HA construct. Furthermore, the neutralizing titer of sera, from mice vaccinated with Bac-HA_m constructs T143A, T198A-I211V and T143A-I211V were significantly higher than that of Bac-HA vaccinated mice. This, therefore, shows a significant correlation between neutralization titer and animal protection. Besides the higher neutralizing titer, mice vaccinated subcutaneously with Bac-HA_m constructs T143A, T143A-I211V, and T198A-I211V also showed a reduced lung viral titer, compared to all other immunized groups. This reduced titer could be the effects of successful induction of neutralizing antibodies. In summary, mice that were subcutaneously vaccinated with Bac-HA_m constructs T143A, T143A-I211V and T198A-I211V induced higher HAI and neutralizing antibody titers than those that were Bac-HA vaccinated. Such higher HAI and cross-neutralizing antibody titers, in turn, provide complete protection against RG-H7N7influenza virus. In conclusion our results indicated that amino acid substitution at positions T143A or I211V significantly enhance the immunogenicity of H7N7HA vaccine against H7N7 and protect mice from lethal H7N7/NL/219/03 virus challenge. Also, we needed further studies to know the role of other amino acid positions and substitution of other amino acids at the same position on immunogenicity of H7HA antigen. Our findings suggest that we can enhance the immunogenicity of poorly immunogenic conventional influenza vaccines (live attenuated or heat killed) or subunit vaccines or viral like particle vaccines by identifying and inhibiting the potential glycan motif-epitope masks formation in HA protein by amino acids substitution. Even though we could not provide more and precise mechanism regarding how I211V mutation enhancing the immunogenicity, the animal protection against lethal challenge with H7N7 virus data suggest that we can improve antigen specific T cell efficacy by amino acids substitution at T cell epitopes, which provide higher binding affinity to MHC and induce higher antigen specific T cell response. We are grateful for the financial support received from Temasek Life Sciences Laboratory, Singapore. We thank Mr. Subramanian Kabilan and Mr. Govindarajan for animal work. We thank Ruben Donis, Influenza Division, Centers for Disease Control and Prevention, Atlanta, GA, for providing the plasmids for reverse genetics. We are thankful to Mr. Ng Qing Yong and Mr.Vishweshwaran Sridhar for manuscript editing. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JK SR MP. Performed the experiments: SR MP. Analyzed the data: SR MP. Contributed reagents/materials/analysis tools: JK FH KVA. Wrote the paper: SR MP.

# Introduction Upper limb prostheses have advanced in the past decade, with the commercial release of devices that have multiple powered degrees of freedom (DOF). While these devices enable more movements that more closely resemble the movements of the anatomical limb, they are challenging to control. Historically, powered upper limb prostheses have been controlled using dual myoelectric (EMG) inputs. Each input controls a single direction of movement, and switching between joints is accomplished by using an alternate signal. Another option for prosthetic control is foot control using inertial measurement units (IMU). The use of IMU controls has been spearheaded in the DEKA Arm, an FDA approved upper limb prosthesis that has recently become commercially available. To date, all studies of the DEKA Arm reported on subjects who utilized IMUs worn on the feet as the primary control method. IMUs were used in combination with pressure transducers and traditional EMG controls and switches, as needed or desired. The IMUs are mounted on a clip that is attached to the top of the shoe and utilize gyroscopes and micro electromechanical systems (MEMS) accelerometers to sense small movements of the foot/ankle in space. The user commands motion of the prosthesis by tilting their foot (and the IMU) in various directions. However, foot controls cannot be used for persons who are missing both lower limbs or who have lower limb paralysis or serious impairment. Furthermore, for safety reasons the IMUs are designed to automatically stop functioning during ambulation due to a “walk detect” feature which minimizes the possibility of inadvertently moving the device while walking. The DEKA Arm also has a “standby” feature, which enables the user to disable the controls without powering off. This is typically operated using a pressure transducer. There are three configuration levels of the DEKA Arm; radial configuration (RC) for persons with transradial amputation (TR); the humeral configuration (HC) for persons with transhumeral amputation (TH); and the shoulder configuration (SC) for persons with a short residual limb, a TH amputation, and shoulder disarticulation or forequarter amputation. All configurations use control inputs for the hand and wrist including: grip selection (to choose between six different grip patterns) wrist flexion and extension, wrist pronation and supination, and opening and closing of the hand. All configurations of the device have a wrist display which shows the grip pattern that is currently selected as well as the mode of operation. The HC DEKA Arm uses control inputs for the movements of elbow flexion and extension and humeral internal and external rotation; the SC DEKA Arm uses a unique control strategy called Endpoint control. Both the RC and HC devices utilize an external battery pack, typically worn on a belt secure around the waist. However, the HC device also has an internal battery and because of this the HC device may be disconnected from the external battery and operated without it for brief periods. Internal battery life varies depending upon the amount of motor usage but typically ranges from 1-2 hours. Regardless of the control strategy, there is substantial cognitive load required for the complex task of pre-planning and sequentially controlling multiple joints/actuators. We believe that this cognitive burden may have been increased when substituting foot movements for upper body movements. This may explain, in part, why a kinematic study reported that using the DEKA Arm was slower than using conventional myoelectric devices for TR amputees and that these subjects, despite training, did not take full advantage of the wrist motion available. Recent commercial release of an EMG pattern recognition (EMG-PR) system holds promise for decreasing cognitive burden of using the DEKA Arm. In the past decade or so, there have been major advances in EMG-PR control, culminating in the launch of the first commercial EMG pattern recognition product, CoAPT COMPLETE CONTROL™ system in 2014 ([http://www.coaptengineering.com](http://www.coaptengineering.com/)). In EMG- PR, the controller deciphers the user intent based on activation patterns of residual muscle when the user imagines and “executes” the motion of the missing limb. The controller then selects the corresponding motor and controls the speed in proportion to the sum of the EMG intensity of all monitoring signals. Compared to the conventional direct EMG control scheme, EMG-PR does not rely on independent muscle contraction. Specifically, EMG-PR utilizes the activation patterns of muscles, chosen by the individual patient, that are mapped to prosthesis movements through the process of calibration. We anticipated, based on results reported in several recent studies, that an EMG-PR control scheme may enable users to control multiple DOFs of the DEKA Arm more efficiently, be easier to learn to use, and be less cognitively demanding, as compared to IMU foot controls in combination with direct EMG control. However, this control method was new and unstudied. Thus, the purposes of our study were to: (1) utilize EMG-PR to control the DEKA Arm; and (2) describe participants’ experience of controlling the DEKA Arm using this control strategy. # Materials and methods ## Study design The study of EMG-PR was conducted as part of the VA Home Study of an Advanced Upper Limb Prosthesis (Home Study), a quasi-experimental study that used a time series design. The research was approved by the Institutional Review Boards of the Providence VA Medical Center, the James Haley VA, the VA NY Health Harbor System, and the Center for the Intrepid at Brooke Army Medical Center. The study consisted of in-laboratory training (Part A); and up to 12 weeks of home use with in-person re-evaluations at study sites, at monthly intervals (Part B). ## Coapt – DEKA arm integration The study team worked with Coapt LLC (Coapt LLC, Chicago, IL) and DEKA Integrated Solutions (DEKA Integrated Solutions Corp, Manchester, NH) to create an integration of Coapt EMG-PR system (referred to as EMG-PR or Coapt in this paper) and the DEKA Arm for this study. Two prototypes of the EMG-PR controls were utilized in this study: Prototype 1 and Prototype 2. Major differences between prototypes are highlighted in. Prototype 1 could be configured to utilize up to 8 distinct movement patterns and allowed mode switching to expand the number of DOF and functions that could be controlled. Thus, Prototype 1 with mode switching could control up to 4 DOFs in the DEKA Arm (3 for an RC device and 4 for an HC device). Prototype 1’s input for mode switching could be used to toggle between hand grip patterns to select the desired grip (aka grip selection) in both the RC and HC configurations. Mode switching was a selection function only, it did not involve movement of the prosthesis joints. In Prototype 1 the wires for the 8 dome electrode pairs, and 1 reference electrode were connected to the Coapt control system processor and then to the DEKA Arm through an multi-connection interface cable (i.e. mating cable), which connected to DEKA Arm Control Interface (ACI) units mounted on the external socket. The DEKA Arm tactor is an output device which utilized vibrations that provided tactile and auditory (i.e. the user can hear the vibrations) feedback to the user about grip force and mode switching. The tactor as well as any pressure transducers used for control of grip selection or mode switching, were also connected to the DEKA Arm through an ACI. RC DEKA Arm and socket with EMG-PR prototype 1 controls. Key device components shown in this picture include: the pressure transducer, wrist display, externally mounted ACI, ACI wiring, tactor, and power cable. Prototype 2 could be configured to operate up to 6 functions by using up to12 unique patterns of muscular recruitment (not shown). Thus, Prototype 2 did not require mode switching for grip selection or humeral rotation control. Prototype 2 interfaced with the DEKA Arm directly through use of a single CAN Bus connection. As in Prototype 1, tactors and pressure transducers were connected to an externally mounted ACI. In this study, all participants, regardless of Prototype, opted to use a pressure transducer for control of standby mode of the device. Calibration, which is the process of training the control system to recognize the patterns of muscle signals for specific prosthesis action. During calibration, the control system records the pattern of muscle signals that are used for the movement, and utilizes this information to control the prosthesis. Calibration could be done using the software interface, which displayed a virtual prosthesis, or through prosthesis guided calibration with the DEKA Arm enabled. In prosthesis guided calibration, the prosthesis moves through each movement in sequence, and the user produces the corresponding muscular patterns for each movement. Both Prototypes used the calibration process available in the commercially available EMG-PR system; albeit with a greater number of DOFs and functions that could be calibrated. Study participants were introduced to the calibration process using the virtual prosthesis in the software environment and then, once they were familiar with the process, advanced to calibrating using the prosthesis guided calibration. Study prosthetists and therapists were oriented to the system and EMG-PR and instructed in EMG-PR training principles, provided with an operations manual and reading list. The training was repeated and the operations were manually updated when Prototype 2 became available. Study staff also had telephone and on-site consultation from Coapt specialists and from DEKA as needed to address technical issues. ## Subjects A convenience sample of persons with unilateral or bilateral TR or TH amputation who met eligibility criteria was included. Eligible participants for Part A were at least 18 years old, with no health conditions limiting study participation. At the conclusion of Part A, the Principal Investigator and the local study staff determined whether the subject met eligibility criteria for Part B. Included Part B participants had, in the judgment of the Principal Investigator and study therapist, at least fair functional use of the DEKA Arm, demonstrated safety awareness and sound judgement, and possessed the ability to troubleshoot minor technical issues. ## Data collection At baseline, participants reported the type of prosthesis they used as well as the number of years that they had been using a prosthesis. Skillfulness of personal prosthesis use was measured using the University of New Brunswick Test of Prosthetic Function for Unilateral Amputees. Qualitative data were collected through open-ended and structured survey questions and semi-structured interviews (SGIs) at the End of Parts A and B (Appendix A). The SGIs were audio- taped. The audio files were transcribed verbatim. Audio and video recordings were made of study training and testing sessions, when feasible. Semi-structured interviews were administered by study staff and allowed for additional probing to elicit further information, and/or clarify responses. Surveys were administered by pencil and paper at the end of Parts A and B. Participants were asked about their impressions of the DEKA Arm, if they wanted to receive a DEKA Arm in the future, to rate their skill level with the experimental system, and to rate comfort and weight of the prosthesis. The subset of participants who were prosthesis users at baseline were asked about activities they could do and preferred to do with the DEKA Arm and not their personal device(s) and the reverse. Furthermore, the Part A survey solicited participant’s impression of whether they had enough training. SGI questions (Appendix A) solicited participants’ opinions on their training experience, the DEKA Arm and its components, the EMG-PR control and their function using the experimental system. Participants were asked to compare the experimental system to their personal prosthetic devices and control schemes, Participants were also asked to comment on the need for repair and service of the system. At the End of Part A (EOA), participants were asked if they thought that they were ready to utilize the system at home. At the End of Part B (EOB), participants were asked to comment on their home use experience. ## Data analyses We used a qualitative case series design with a constant comparative approach to identify common categories of experience using the EMG-PR controls. Data sources described above were used in conjunction with individual case summaries developed by study analysts in the qualitative analysis. Case summaries provided important contextual information, gleaned from communications with the study staff and/or videotaped training sessions, not necessarily available through other source materials. Development of the coding structure was an iterative process which occurred during and after data collection. Data analysis was facilitated by the use of NViVO software which was used to manually code and organize text data. summarizes the coding process. Coding began when two qualitative analysts independently coded the data from four participants to identify major categories in the data. The analysts then met with the first author (LR) to discuss the commonalities and differences in the major categories and then write preliminary definitions of each category. This resulted in coding iteration#1. These definitions were then deployed by the analysts as they coded data from four additional participants. The process was repeated, and the coding categories and content compared, resulting in further refinement of definitions, reorganization of several categories and the addition of new categories (coding iteration \#2). Coding iteration \#2 included many sub-categories which were generated to appreciate the scope and content of each category. The revised coding schema was then utilized to code 2 additional cases and the constant comparative process repeated again and coding iteration \#3 was generated. The analysts re-coded the 4 original cases using coding iteration \#3, constantly comparing exemplars within each category to insure appropriate classification. These results were discussed with LR, and together the team determined that no new categories were identified in the data and that saturation of coding categories had been reached. The two analysts then met to apply the coding scheme (iteration \#3) to all 10 cases, resolving any differences of interpretation together during the coding session. Thereafter one additional case, the last subject who had completed Part B, was jointly coded by both analysts utilizing coding iteration \#3. Summary reports of the content of each coding category were written by the analysts and reviewed by the first author who further organized the categories by identifying a central, or primary category, the major categories of user experience, factors impacting experience, and broader contextual factors (coding iteration \#4). Coding iteration \#4 was shared with the qualitative analysts who agreed that organization made sense and that all data was represented in coding iteration \#4. Finally, the data for the last subject in the study, who did not complete Part A, was coded by both analysts using coding iteration \#4. shows the major categories and subcategories of each of the coding iterations. # Results Characteristics of the study sample by amputation level are shown in. Characteristics of individual participants (identified by pseudonyms) are shown in. Briefly, the sample consisted of 12 persons with upper limb amputation (TR = 10; TH = 2; males = 10). Both participants with TH amputation had undergone targeted muscle reinnervation (TMR), a surgical procedure in which severed nerves that previously controlled arm/hand function are transferred onto targeted de-innervated muscles on the residual limb. This procedure allows the user to use physiologically appropriate motor control signals when operating prosthetic myoelectric controls. Eleven participants (92%) were users of a personal prosthesis at baseline; 4 (33%) had prior experience controlling the DEKA Arm using foot controls; and 2 (17%) had prior experience using EMG-PR with another prosthesis. Three participants (all TR) used only Prototype 1 of EMG-PR and DEKA integration, 2 participants utilized both prototypes (1TR, 1 TH). Seven participants (6 TR, 1TH) used only Prototype 2. Eleven participants completed all Part A activities (9 TR, 2 TH); 8 participants began Part B and 7 completed (5 TR, 2 TH). One participant was lost to follow-up and did not complete a SGI or structured surveys in Part A. Because of this, we used mid-study data and selected transcripts from training videos in his analysis. Two Part B completers (1 TH, 1 TR) discontinued using the DEKA Arm at home and another (TR) chose to use the device minimally during much of Part B. The participant with TH amputation stopped using the device for the last 9 weeks; one TR participant stopped using it for the last 6 weeks; and another TR participant wore/used the device for only 1 hour during the last 3 weeks of Part B. In addition, the other subject with TH amputation decreased his prosthesis use during the last 8 weeks of Part B due to hot and humid weather. Seven subjects completed Part B activities. Two completers of Part A (both TR) were deemed ineligible for Part B because they had not achieved sufficient prosthesis control within the training time completed. One had completed 20 training visits (about 40 hours) and the other chose to end study participation after 8 visits (about 16 hours), declining additional training that was offered. One Part A completer (TR) chose not to continue into Part B, stating that he was too busy to participate, however the study staff noted that he had a low frustration tolerance and did not feel comfortable wearing the device in public. ## Changes to EMG-PR settings during study participation Six participants had modifications to their controls during Part A of the study. The most common type of modification was changing from PR control of grip selection to the use of a pressure transducer. Although not shown in, the majority of participants altered or refined the muscular recruitment strategies that they used for specific movement controls. Control of grip selection was used successfully for 3 of 4 participants who used Prototype 1. One participant had difficulty controlling 3 DOF using PR and was provided with an IMU to control the pronation/supination movement. Although it was possible to control grip selection using the EMG-PR controls of Prototype 2 system, by toggling through the 6 grip patterns to select the desired grip, all four subjects who tried doing so encountered difficulties with this function. Three transitioned to using a pressure transducer for grip selection. When these types of changes occurred, the subject had additional practice (training hours) using the new control schemes before final testing. Given the difficulty in controlling grip selection through EMG-PR, and the frustration for study staff and subjects, subsequent subjects using Prototype 2 had grip selection set-up initially with pressure transducer control. One participant (TH) transitioned from Prototype 1 to Prototype 2 as soon as it became available, which occurred after the completion of Part A training and before the initiation of Part B. Given the timing, the participant returned to the study site and had the control system updated. At that time, he had 2.5 hours of training delivered over a two day period. At the end of the two days, the OT reported that he was able to operate all controls well and the subject indicated that he felt ready to take the system home. A second participant (TR) also transitioned from Prototype 1 to Prototype 2 prior to the start of Part B and after the EOA. The OT and prosthetist determined that this subject required no additional training hours given that the control scheme for Prototype 1 and 2 were set up identically. ## Overview of qualitative findings The overall organization of the coding categories are displayed in and the content of each major category is described briefly in. The section below contains descriptions of the findings, organized using coding iteration \#4, with illustrative examples from the rich text data. shows the theoretical model resulting from our analysis. The desirability of EMG-PR and the DEKA Arm system is shaped by the user experience of the system’s complexity, calibration and function, as shown in the central circle of the figure. These aspects are largely explained by the categories shown in the inner ring of the figure: technical issues, fatigue, prosthesis design, EMG-PR prototype, training and acclimation. Finally, personality and environmental factors, as shown in the figure’s outer ring are contextual factors that influence the entire model. ## Experience of EMG-PR control of the DEKA Arm ### Desirability of the system Part B participants had mixed views on the overall desirability and usability of the system. Two subjects (both TR) stated that they wanted an EMG-PR controlled DEKA Arm, while two others (1 TH, 1 TR) stated that they might want one in the future. Three subjects (1 TH, 2 TR) said they did not want one. At EOB, the 6 participants who were prosthesis users at baseline were asked to compare the controls of the DEKA Arm to the controls in their personal prosthesis. Four (66%) indicated that they preferred their own controls, and 2 (33%) did not express a preference for either control type. Several subjects, including both users with TH amputation and a history of TMR, made very positive comments about EMG-PR independent of the interface with the DEKA Arm. > “*I think that is potentially one of the most advancing systems I’ve > seen in…forever*. *I mean*, *that’s definitely I think the next step > from like*, *where the TMR is…Going from the*, *you know*, *basic two > bicep/tricep muscles to the four with the TMR*, *and then to the Coapt > system with the TMR*.” (David, Part B) Some subjects with TR level amputation recognized that the system that they used was an early prototype and spoke about the potential of using EMG-PR with the DEKA Arm “if the fundamental problems can be addressed.” One believed that EMG- PR could eventually be the optimal means of controlling the multiple DOFs of the DEKA prosthesis: > *“…I think that once all of the problems are worked out and the > integration is completely done*, *figuring out how the 2 technologies > work together*, *I think the Coapt system or pattern recognition > system in general will be the best way to move the hand*, *because > there are so many degrees of freedom and it allows for such intuitive > control with that many degrees of freedom*.” (Mary, Part A). Subjects mentioned several features of the DEKA Arm which dampened or added to their enthusiasm for the device. Ten of eleven subjects who completed Part A and 5 of the 7 who completed Part B commented negatively on the weight or bulk of the DEKA Arm system, remarking that it was “too heavy,” that the “weight is very frustrating”, or the device is “bulky” or “cumbersome.” The size of the prosthetic socket with the externally mounted componentry, and the shape of the hand itself made it difficult for some subjects to wear their usual clothing. Subjects made negative comments about the external battery pack and wiring. One subject said there was “too much hanging off it.” Another remarked, > “*It didn’t work near well enough to warrant dealing with the crap > hanging off of it*, *especially the battery*, *belt and cord*. *Try > taking off a coat and using a public restroom wearing that get up*.” > (George, Part B) Participants commented positively on the powered wrist movements of the DEKA Arm. One subject with TR amputation explained that while the need for repeated repairs had a “large effect” on her interest in obtaining a DEKA Arm in the future, the benefit of having a wrist that was able to able to “rotate, flex and extend” was “definitely worth it.” One subject (TR),who called the DEKA Arm “not ready for prime time yet,” reported that he used the Arm only for 1 hour at home during the last 3 weeks of Part B, explaining that the DEKA Arm, despite its multiple DOFs, made him less “productive:” A third subject (TR) summed up the tradeoffs between functionality of the system and need for frequent repair: > *“……I could do chores around the house*, *mowing the lawn and stuff*, > *that was definitely a bonus*, *but every time I mowed the lawn the > hand would break then I’d get frustrated*, *because then I wouldn’t > wear it as long because it was broken so I couldn’t wear it*.*”* > (Martin, Part B) A fourth subject with bilateral lower limb amputation explained his preference for his own simpler terminal device: > *“I just left it (DEKA Arm) there and I was like*, *’I’m putting my > own arm on*.*’ It’s faster*… *easier…After a while you’re like*, > *‘this just takes too long to do stuff with this hand*.*’ And I could > just like throw mine on*, *throw a Greifer on it and call it a > day*.*”* (Charles, Part B) This subject, who sometimes utilized a wheelchair and sometimes walked with a cane, also reported that the external battery pack rubbed against the side of the chair, the device was complicated to wear when seated in the wheelchair, and the wrist bent under the weight of his arm when using a cane. ### Complexity and intuitiveness of controlling DEKA Arm with EMG-PR All subjects who completed Part A commented on the complexity of using the EMG- PR controls to operate the DEKA Arm saying that it was “more complicated”, “harder”, “more difficult”, and “a more tedious process” compared to other prosthetic control systems that they had used. Because of the powered DOFs of the DEKA Arm, subjects required more control inputs; sometimes two or three times more than their personal devices. Subjects needed to remember, and then replicate, distinct patterns of muscular contractions for each control. Those who were already using myoelectric controls needed to use new muscular patterns that may have differed from their current control systems. During Part A one subject who was accustomed to using two-site myoelectric control remarked that learning to use the new controls was, “like a re-wiring of the brain and the muscles.” This process, he said, at EOA, was “exhausting” and “too much,” and he thought that it would require “a lot of training” to become proficient. At EOB, the subject stated that he preferred the control system of his current prosthesis due to “the simplicity,” explaining that using his controls was “easy,” in part because he was familiar with them and that he did not need to calibrate them. Another subject with TR amputation compared the EMG-PR controls to his current dual site myoelectric controls, noting, “it’s a lot harder to use” because “it’s got more complexity.” The complexity, he reported at EOA, “sometimes gets in the way”, causing inadvertent movement. He explained, > “*So it can do things*, *but if I’m focusing on something and*, *this > comes to not having worn it enough*, *and not having used it enough*, > *but*, *you know*, *I will inadvertently open it*, *or I’ll > inadvertently flex or extend the wrist when I’m trying to do > something*.*”* (George, Part A) At EOB, this subject reported that he continued to have difficulty discriminating between all of the control inputs to utilize the system with the DEKA Arm. However, he was very positive about the potential for EMG-PR as a system of control in general, and believed that it would have been easier to use with a prosthesis with fewer DOFs. He remarked, > *“*…*slap the Coapt and Michelangelo on me tomorrow*, *and it’ll go > off like gangbusters*” (George, Part B). This subject was one of two who thought that it would have been easier to master the PR control system if they had been exposed to it at the time of amputation, prior to any other prosthetic training. This subject explained that initially his sense of his phantom limb, which he engaged in using the controls, was stronger. > *“…my hand was amputated closing in on three years ago… I actually was > able to work with it*. *I could feel each individual left finger*. *I > could do things*, *I could move each finger independently*, *‘cause I > still had sensations in them*, *but after two years plus of using a > myo*, *a standard myo with just two sites*, *you lose all those > sensations*. *And so the ability to discriminate between rotating the > wrist*, *flexing*, *or extending and flexing the wrist*, *and opening > and closing the hand - that gets much more difficult than it would > have been for me two years ago*.*”* (George, Part A) Similarly, another TR amputee remarked that he thought it would have become “second nature” to use the controls if he had trained with the system when he was a “brand new amputee....“fresh out the hospital”. The complexity of the control system was even greater for those using an HC DEKA Arm. These subjects needed to operate 4 more controls than those operating an RC Arm. One subject commented: > *“…it takes a lot more thought and a lot more training I feel*, *to*, > *and not just like strength training and stuff*, *but just thinking of > what muscles or what movements you want to make”* (David, Part A) This subject chose to discontinue using the DEKA Arm during Part B, soon after 2 additional distinct controls were added during the system upgrade to Prototype 2. He explained that the additional complexity discouraged him from using the device, leading to abandonment of home use, commenting: > *“…doing the amount of movements that we do already with this > (Prototype 1) is more than enough*. *I mean*, *we’ve got like six > different functions that I’m having to do…and adding two more into > it*, *it changes everything*.*”* (David, Part B) The second HC user, who had prior experience with an EMG-PR system, said that the study experience was, “kind of humbling because I thought I was better at pattern recognition than what I am.” He remarked that he “enjoyed” the process of using EMG PR control because he could “drive the Arm” with his phantom limb, and that it felt “more in-tuned” to him. However, he noted that his current prosthesis only had three functions, and that the DEKA Arm had five functions. At EOB, although he indicated that he had fully acclimated to the EMG-PR controls, he stated: > *“…this arm is very mentally taxing*, *with all the patterns and > everything like that*, *to use it to get it to work and everything > like that*. *It is super*, *it takes a lot of concentration……so it > took a while to figure out how to isolate all those patterns*, *‘cause > you’ve got to figure you’re doing two for humeral rotation*, *you have > two for elbow*, *so every movement you have two patterns for*. *You > know you have the movement and then you have the opposite movement*, > *so there are 11 different patterns or something…”* (John, Part B) Another subject, a TR amputee who had prior experience with EMG-PR control, made similar comments about the complexity of controls due to the greater number of DOFs of the DEKA Arm. He reported that he felt that the EMG-PR control system used in this study was less intuitive than his other prosthesis control systems, and that it was “more difficult” than he anticipated it would be, and “less reliable,” leading to unintentional control activation. He commented: > *“…*… *there was a lot of pattern signals getting confused with each > other – that’s the only way that I can put it…You try to do one thing > and something else happens*, *or if you move your arm in a certain > position*, *and you try to open your hand or you want to close and it > opens up and you drop your utensil or whatever*.*”* (Matt, Part B) Although most subjects spoke about the complexity of the system, 50% of the subjects who used a personal prosthesis at baseline and also completed Part B (N = 3), stated that the EMG-PR system was either more intuitive than their existing mode of control or “basically the same.” One subject who reported that it was more intuitive explained how the system adapted to him: > “*I do stuff and Coapt…Coapt records what I do*. *Coapt learns me*, > *as opposed to me learning the standard myo*. *So it’s far superior > and I think in the long run*, *it’s going to be definitely*, *you > know*, *the go-to control mechanism…”* (George, Part B) Three of the 4 subjects who had previously controlled the DEKA Arm with IMUs made comments comparing the two control approaches. One subject with a TR level of amputation felt the EMG-PR control was more intuitive than IMUs, but stated that she also liked the “simultaneous control” that the IMUs afforded her for wrist and hand movements, which she did not feel was possible with the EMG-PR controls. > *“… I really got used to them (IMUs) and liked the fact that they are > simultaneous control*, *and I could bend my wrist while opening and > closing the hand like that was a really cool part of it*. *But at the > same time*, *it’s frustrating having to move my feet to move my hand*. > *It’s not intuitive obviously*, *so I do like Coapt control overall > because of that*, *but if it had simultaneous control it would make it > better*.*”* (Mary, Part A) Another subject with TR amputation also commented on the advantage of “not having to worry about something else” (i.e. moving one’s feet) for control, but instead depending on “the motion computing from your amputated arm.” This was the only subject in this study who, due to his inability to master the EMG-PR controls and his prior familiarity with IMU controls, was fit with an IMU for some wrist movements during training. He expressed satisfaction with the IMU for that motion and said that he preferred it to using EMG-PR control, stating. “it made it a lot easier” because “I only had to worry about two motions (with EMG- PR) instead of three.” A subject with TH level of amputation, who was highly dissatisfied with foot controls when he participated in prior DEKA studies, continued to express this opinion at the end of this study. He felt EMG-PR was “so much more intuitive” than foot controls. > *”… it beat out the IMUs a billion to one*, *cause I hated the IMUs*. > *For transhumeral*, *I wouldn’t use anything but the Coapt*.*”* > (David, Part B) He appreciated the fact that the EMG-PR controls did not involve a device mounted on his shoes and could be used when barefoot. However he observed that EMG-PR required more training and stated that for “somebody who’s never worn prosthetics before” the IMU system may be easier to learn: > *“The Coapt it takes a lot more thought and a lot more training I feel > and not just strength training but just thinking of what muscles or > movements you want to make*. *It’s a bit more complex but you don’t > have to worry about moving your feet*. *It’s so much simpler and more > complex at the same time using the Coapt as opposed to the IMUs*.*”* > (David, Part A) This subject noted that the weight of the DEKA Arm affected controlling the prosthesis with EMG-PR “a lot, more so than like the IMUs”. But the tradeoff he thought was “well worth it.” ### Calibration All subjects commented on the experience of controls calibration, discussing the reliability and consistency of the calibration process, the degradation of controls after extended use, and the need for and frequency of recalibration. One subject remarked, > *“we calibrated*, *it worked for a good while*. *Then something > happened…the calibration wasn’t*, *adjusting to me or me adjusting to > the Arm*. *It was quite difficult*. *We had to recalibrate*.*”* > (Charles, Part A) Although re-calibration typically improved prosthetic control and subjects said that they would “rather recalibrate” to obtain better controls when there was degradation, the need to calibrate at all was seen as a disadvantage as compared to current myoelectric control. In comparison, direct myoelectric control was perceived as “simple,” did not require recalibration and “will work no matter what.” There were also occasions when the calibration process itself was not successful. One subject commented that she was not “guaranteed that it’s (the calibration process) going to work,” and thus she would need to repeat the process. All subjects spoke about strategies that they used to improve the likelihood of getting a successful calibration. Several said calibration was more successful when they wore the DEKA Arm regularly, paid greater attention during the calibration process, layered calibrations, and/or calibrated with the prosthesis weight supported as well as not supported. One subject at EOA and 4 subjects at EOB reported that they needed to recalibrate often and would layer multiple calibrations to regain control. However, the need for and frequency of calibration varied within subjects. As one subject remarked at EOB, “sometimes I would do it two times a day and sometimes I wouldn’t have to calibrate at all.” Additionally, the number of calibration attempts needed to obtain good control was inconsistent: > “*there were times when I could calibrate once and that’s great*, *but > then there were times when it was really acting up and I’d like do it > six or seven times in a row before I’d get it calibrated*.*”* (Matt, > Part B) Several subjects explained that they found the calibration routine tiring, particularly because, with the prosthesis guided movement, “the Arm is moving around so intensely.” The movement of the prosthesis required the subjects to stabilize their proximal arm and shoulder using their own musculature, or sometimes using their contralateral arm or a table. One subject explained that the movement of the prosthesis on the residuum, > *“alters the way that it’s picking up the recording because it (the > prosthesis on the socket) is like shaking back and forth*.*”* (Suzie, > Part B) Although users could calibrate using the computer software which displayed a virtual arm, during the study calibration using this method could only be done in laboratory with assistance from a staff member. Two subjects stated that they were able to time their muscle contractions using the virtual Arm. Another stated that he didn’t “like to watch on the laptop” and preferred to watch “my real hand” do the work, during calibration. The two subjects who had prior experience with the EMG PR, reported that the calibration process for the EMG-PR DEKA controls was the same as they were accustomed to, but it took longer because “there’s a lot more movements.” These subjects also commented that the calibration process seemed less reliable compared to calibration with their personal prostheses. ### Function Several subjects made general comments describing their enhanced functionality due to ease of use of the EMG-PR system. > *“…because once I have it on I can just do things just more fluidly as > opposed to pressure to open*, *pressure to closing it*.*”* (Suzie, > Part B). Most users commented on the advantages or disadvantages of the combined functionality of the controls and prosthesis. For instance, a TH user remarked that the entire DEKA Arm system with wrist flexion-extension and humeral rotation together with EMG-PR controls provided more “freedom” and “felt natural,” saying “you don’t have to angle yourself” for tasks like he needed to do with his other prosthesis. Another subject commented negatively about the wrist function, after using the prosthesis to remove a T-shirt, > “*When you get \[DEKA\] on all those funny angles it just gets > taxing*.*” (Jason; Part A)* One TR user described how various aspects of the entire system made him slower and less productive than when he used his existing prosthesis. > “*…it became apparent to me through the process of this that putting > the DEKA Arm on made me less productive, it slowed me down…The time it > takes to put it on, make sure it’s working right…I was working around > it rather than using it to just do stuff*.” *(George, Part B)* Subjects who were using a prosthesis at baseline were asked at the end of Parts A and B to comment on activities that they could do or preferred to do with the DEKA Arm compared to their own prosthesis. All prosthesis users who completed Part A (9 TR, 1 TH) used a myoelectric device at baseline. When asked to compare their own device to the DEKA Arm at EOA, 50% said that there were activities they could do with the EMG-PR controlled DEKA Arm that they could not do with their own prosthesis and 30% said that there were activities they could do with their own prosthesis that they could not do with the EMG-PR controlled DEKA Arm. When asked whether there were activities they preferred to do with each prosthesis, 30% said that there were activities that they preferred with the DEKA Arm and 60% said that there were activities that they preferred to do with their own prosthesis. At EOB, 33% of completers said that there were activities that they could do with the DEKA Arm but not their own prosthesis, and 83% stated that there were activities that they preferred doing with their own prosthesis as compared to the DEKA Arm; 50% stated that they preferred to do everything or “just about everything” with their own prosthesis. (.) ## Factors affecting user experience ### Training and acclimation At EOA, all subjects stated that the amount of training they received during the study “was just right” and most indicated that they expected to continue acclimating to the system with “time and repetition.” Both TH amputees commented on the importance of adequate training, remarking that acclimation “takes time.” The TH subject who discontinued usage during Part B estimated that sufficient training “would have taken three or four months” compared to the 20 hours (10 training visits) over approximately 6 weeks that he had initially received, and 2.5 additional hours after the prototype was updated. At EOB he commented, > *“People need more time to train*. *Especially if we’re going to > operate it the way it’s supposed to operate*.*”* (David, Part B) The other subject with TH amputation, who was experienced using EMG-PR controls with his personal prosthesis, said that he had fully acclimated to the device and controls. He reported that the amount of training he had received (about 30 hours over 15 visits) was adequate. He advised others learning to use the controls system to be “patient with it,” noting, “it’s a steep learning curve.” At EOB, when asked to rate their level of acclimation to the controls system, only 2 of the 7 subjects who completed rated their acclimation as complete. “Two others stated that they were “mostly” acclimated, and 3 stated “somewhat.” Subjects who stated that they were other than completely acclimated, made comments such as: it was still “a learning curve,” and “I think I control it as good as I can for what it’s got.” Only 1 of the completing subjects rated his skill level as “excellent”, while 2 rated themselves as “good”, 3 as “fair” and 1 as “poor”. One subject who had persistent difficulty discriminating muscle muscular patterns said: > *“*,*if I had no other responsibilities*, *I could spend my days > just practicing with DEKA and Coapt*, *I might’ve been able to get > that discrimination back*. *I don’t know*, *maybe I could*, *maybe I > couldn’t…”* (George, Part B) Factors that may have interrupted the acclimation process during the home portion of the study included discomfort, pain, or skin rash which prevented wearing the prosthesis, as well as disruptions in prosthesis use due to technical problems which required repair. Three subjects, one who did not continue to Part B and 2 who did, reported that pain, numbness, or discomfort in their residual limb limited or prevented prosthesis wear. A fourth subject reported that her socket was too loose at the beginning of Part B, which limited her use; however it was adjusted during her Week 4 visit to the site. At week 8 this subject developed a persistent skin rash that prevented prosthesis wear for approximately a week. This subject who had multiple interruptions in use at home summarized the advantages of being able to “stay in a routine for wearing” the device. > *“…so it was a lot easier to use as opposed to when I didn’t wear it > for a little bit and then got busy and putting it back on was like > starting fresh again*.*”* (Suzie, Part B) ### Fatigue All subjects who participated in Part A and the majority of subjects who participated in Part B spoke about fatigue resulting from using the EMG-PR controls. Factors contributing to fatigue included; the need to perform distinct new patterns of muscular contractions, repetitions of those contractions, extended periods of prosthesis use, the weight of the DEKA Arm, and the need to recalibrate. Some subjects described a cycle wherein they became fatigued from using the controls, and then muscular fatigue led to increased difficulty reproducing the distinct muscular patterns needed for control. Their muscles were, in some cases, “working twice as much” in order to utilize the controls and this was perceived as “a lot.” One subject stated, > *“the more tired I got*, *the harder it was to give reliably > consistent*, *you know*, *discriminated inputs that the Coapt was > looking for*” (George, Part B). Two transradial amputees explained how fatigue from overactive muscles led them to unintentionally activate the EMG-PR controls. One said, > “*I think what creates the inadvertent activity is muscle fatigue or > muscle*, *overactive muscles*, *and that’s brought on by fatigue”* > (Tom, Part A) Another subject reported, > *“after having it \[the DEKA Arm\] on a couple of hours*, *especially > extended use*, *and I think I found that if I was overdoing it*, *it > would not operate the way I wanted it to*, *so I would do it even > harder…which adds more fatigue*.*”*(Adam, Part A). The two subjects who were experienced EMG-PR users explained that fatigue was caused by contracting unfamiliar muscular patterns to control the prostheses’ DOF. > “*because these are patterns that I’ve never made before*, *there is > definitely significant muscle fatigue so I have a hard time setting > that same threshold for the pattern*.*”* (John Part A). One subject talked about the fatiguing effects of repeating contractions during the calibration layering process, and explained how this was compounded by the weight of the prosthesis. He said, > “*as I added calibrations to it*, *I was getting more tired*, *my > overall calibration to a certain point seemed to like started to > getting worse*…*I don’t know how much of that was me… fatigue…that I > have a harder time discriminating…I was confusing Coapt*. *The weight > of the DEKA influenced that to some degree*. *I just couldn’t tell you > by how much*” (George, Part B). ### Prosthesis design The most common comments about prosthesis design affecting use of EMG-PR related to the weight of the device. Three quarters of subjects in Part A and 71% of subjects who participated in Part B commented about the impact of the prosthesis weight on their controls experience. Three subjects thought that difficulty performing muscle contractions during calibration and control was due to the prosthesis shifting weight during movement, and the resulting changes in socket pressure and surface electrode contact, which sometimes led to involuntarily muscle contraction. As one subject explained, > “*you calibrate in one position and then you start doing activities in > another position and the weight absolutely changes the contact of the > Coapt*…*I think it’s one of the issues why I couldn’t get control*. > *Calibrating in one position start doing activities in another*” > (Adam, Part A). Another subject explained, > *“the weight could have affected with the thing*....*you are holding > your muscles and they tense and you get cramps or something or you get > muscle fatigue and that could send different signals to the different > wrist turning and other motions and sometimes it might confuse it to > where you are ”* (Martin, Part A) The HC subject who was an experienced EMG-PR user talked about the weight shifting of the prosthesis on the residual limb, saying, > *“the weight of the Arm is actually shifted on the residual limb*, *so > it makes more contact if you are reaching out to your front or to your > side on different parts of that limb*, *so it is harder to set the > pattern because the contact isn’t the same from the socket*. *So the > thresholds for the patterns stay the same*, *but the conditions for it > are not the same*. *It becomes harder to replicate because of the > weight of the Arm and whatever you have in your hand*.”(John, Part B). The prosthesis weight also caused proximal muscle fatigue: > “*I have gotten fatigued several times*, *but a lot of time I don’t > think it was as much in my residual limb*, *the fatigue had to do with > my shoulder carrying the weight of the DEKA*. *And occasionally there > was a little bit of fatigue in my residual limb*, *but at that point*, > *I just recalibrate it and it adjusted*. *Like when I’m training it*, > *it adjusts for fatigue so it was never really a problem*” (Mary, Part > A). There were other issues related to the prosthesis design that influenced user experience. Although the power draw from the EMG-PR system was not expressly measured it was the user's impression that it reduced internal battery life of the DEKA arm as indicated by the statement that: > *“it sucks power like crazy*, *so I don’t think you can even last like > a couple of hours*. *With the (EMG PR system)*, *that’s the biggest > thing they have to address*.*”* (David, Part B) ### Technical issues Several subjects commented on inconsistency of controls operation. They believed that the problem stemmed from the controls system itself or integration of the controls with the prosthesis. One subject who did not wish to extend his training in Part A to become more proficient in order to participate in Part B stated: > “*I was unable to get a consistent control after many different > calibrations, many hours of training. I just couldn’t get consistent > control. Sometimes it would work well sometimes it would work kind of > good, and sometimes it would be horrible. And so just from a > consistency standpoint and reliability it was low marks*.” *(Adam, > Part A)* Factors influencing consistency of the EMG-PR controls included sweating and changes in volume of the residual limb that affected electrode contact. One subject who continued to Part B was often unable to wear the Arm during the summer months due to sweating interfering with control reliability. He explained: > *“…it was the user interface where I was sweating so much that*, *I’m > just conducting too much…*.*Because the more you sweat*, *the more > sensitive it becomes which makes it*, *you know*, *you’re constantly > overshooting whatever you target…”*(John, Part B) Another subject who had changes in residual limb volume reported: > *“And with Coapt like just throughout the day and your arm moving*, > *or your residual limb getting bigger or smaller from heat or cold or > whatever*, *affects how you control your arm or if you have to > recalibrate throughout the day*. *I just feel like that changes like > the consistency of control*.*”* (Mary, Part A) A subject who had poor skill after 40 hours of training and was deemed inappropriate to continue to Part B thought that some of the problems were related to socket fit and electrode contact. This subject was also observed by staff to have difficulty with retaining information, and was only able to tolerate wearing the prosthesis for about 40 minutes at a time. The subject reported: > “*They’re (the controls) not hard to learn but what is difficult is > once you learned and you calibrate it, is to transfer that information > immediately. It doesn’t seem to be able to do it immediate all the > time, just some of the time and I don’t know the reason for that yet. > I’m suspicious of what it might be. It seems like maybe the arm swells > up or shrinks or something when you put it on*.” *(Tom. Part A)* Several subjects who participated in Part B reported similar issues of inconsistency in control operations. One stated at EOA that “there was a lot of inconsistency” which she said, “had nothing to do with me as a user.” The other stated he felt the pairing of the EMG-PR system and the DEKA Arm technologies “was kind of rushed” and he felt it would have been better to > “*have had technicians on site with it when we were actually starting > to use it*, *Coapt and DEKA…because we’re putting the two systems > together*.”(John, Part A) Ninety two percent of subjects who participated in Part A and 86% who participated in Part B required at least one repair to their controls or prosthesis during the course of the study. On average subjects required 2.1 (sd 1.2) repairs during Part A, and 3.9 (sd 3.8) repairs during Part B. The range of needed repair episodes was wide, particularly during Part B. Technical issues and repairs were of the controls system, the prosthesis or a combination of both. It was often difficult for users and staff to differentiate the source of the problem. When asked to comment on the frequency of faults or errors in controls and the need for troubleshooting, one subject said “it wasn’t acceptable.” Another subject said that he thought that the faults were “in that Coapt/DEKA integration.” One subject indicated that he thought the DEKA arm system itself was reliable, but that the control system “is not as reliable”. Another stated, > *“It is awesome when it is working well*. *However it is frustrating > when it is not working”* (Jason; Part A) One subject stated that he thought that the DEKA Arm was “too unreliable, kept breaking down” and said he did not want the DEKA Arm and the EMG-PR in the future. The same subject also said that he thought EMG-PR “could be” an effective means of controlling the DEKA Arm, despite the fact that “the wires kept coming apart” during home use. Several subjects said that their main concern about taking the DEKA Arm home was that it could “malfunction”, and that “too many things can go wrong.” During Part B, the need to ship the prosthesis back for repairs created interruptions in study participation. One subject, who experienced 8 repair episodes during Part B, said he used the DEKA Arm “less than I expected” because it “kept breaking down”. He explained > *“it just interrupt the whole thing…that distracts you from wanting to > do it because it’s like*, *“Aww*, *man*! *Here we go again*!*””* > (Matt, Part B) Another subject also commented on the impact of interruptions due to repair episodes, stating: > “*anytime you stop using something your body just gets so used to > using it and your muscles*, *the memories are in there*. *It’s kind of > like riding a bike*, *once you get used to it you know how to do it*.” > (Suzie, Part B) This experience was not universal however. One subject said that he was “impressed with how consistently well” the system worked. However, this was the subject who did not use or wear the device due to “sweating” for 34 days of Part B which occurred during the hot summer months. Although subjects experienced many technical problems during the study, some were still very positive about the system. One stated, > “*when it is working*, *it’s working great*, *because it helps me do > things I need assistance with*” (Martin, Part B) ### Issues related to prototype change or grip select control As mentioned above, two subjects had their controls system upgraded to Prototype 2. One TH amputee, who had two additional control inputs added during Part B, reported that the additional complexity of the new control inputs made him “not want to use the thing at all,” and he discontinued all use after Week 4 of Part B. This subject was also observed by staff to have a low frustration level. The TR user who switched to Prototype 2 experienced multiple problems with the operation of grip select that required multiple visits to the site for adjustments. These problems clearly affected his perception of controls reliability and willingness to use the Arm at home. Two subjects who used Prototype 1 with EMG-PR for grip selection throughout their protocol were generally satisfied with that form of grip control. One subject who used Prototype 1 and changed to pressure transducer during training, stated the manual mode of control was “easier” and that “pattern recognition was a huge pain” because: > *“…moving my arm in any way confuses it*, *I think*, *to where it > thinks that I’m asking it to change the grip and it does when I don’t > want it to*.*”* (James, Part A) No subjects used direct grip select with Prototype 2 without experiencing technical problems. These problems may have been attributable to the experimental nature of the grip select option in these new control prototypes. Despite problems, one subject would have preferred to use EMG-PR for this function if it had been more reliable. > “*II really liked being able to use my Coapt to do the grip select. I > felt like it made it a lot easier. I hate having to use the button. > But I do feel like since we took the button out that the hand has been > cooperating much more than it was before*.” *(Mary, Part A)* The other 2 other subjects who initially used EMG-PR for grip select with Prototype 2 (1 TH, 1 TR) also experienced technical problems and changed to pressure transducer during training. They both expressed satisfaction with this manual mode of control. These types of problems led prosthetists to discontinue utilizing EMG PR for control of grip select. Thus, the 5 final subjects in the study used pressure transducers for grip selection without attempting to control it through EMG PR. ## Contextual factors The study analysts identified a variety of personal and environmental factors that may have influenced subjects’ experience and reporting of their experience. Personal factors, including low frustration tolerance, interfered with learning and willingness to engage in training visits for several subjects. One subject clearly had difficulty in comprehending interview questions, as evidenced by several responses that did not address the questions asked. Several subjects, as mentioned above, had prior experience with the foot controlled DEKA Arm or with EMG-PR control of their personal device. These subjects may have had certain advantages over others in that they had less new information to learn. Residual limb pain, skin problems and intermittent swelling interfered with prosthesis wear time in several cases. It is possible that muscle atrophy or limited residual limb musculature due to extremely short residua or amputation approach may have made it more difficult for some subjects to utilize EMG-PR signals as compared to others. Environmental factors, such as the subjects’ level of rapport with staff; personal issues creating a chaotic home environment, and health concerns for other family members, impacted subjects’ willingness and ability to engage with study activities. A major hurricane with flooding created a long disruption in training for one subject; and very hot and humid summer weather impacted another subject during Part B, who could not tolerate wearing any prosthesis due to excessive sweating. # Discussion Our work adds to the understanding of user experience of EMG-PR control of upper limb prostheses by providing an in-depth analysis of participant perspectives on using a commercially available EMG-PR control system that had been adapted to interface with the multi-DOF DEKA Arm. This system enabled EMG-PR control of up to 4 functions of the RC DEKA Arm, and up to 6 functions for users of the HC DEKA Arm. These functions included powered degrees of freedom of the prosthesis and toggle grip selection to enable switching of the DEKA Arm’s pre-programmed grip pattern. Our study also compared and contrasted participants’ views on EMG- PR control and the DEKA Arm as compared to the controls used with their personal prostheses, the attributes of their personal prostheses, and IMU foot controls of the DEKA Arm. Although EMG-PR control of upper limb prostheses has been studied for decades, few clinical investigations have compared this control method to other control methods. One in-laboratory study, and one take home study, compared clinical outcomes, such as dexterity and activity performance, for transradial amputees who used two or three DOF prostheses. These studies involved very small samples and reported mixed results regarding superiority of one control type over the other. A third study, which used a randomized control design, compared EMG-PR control to direct myoelectric control in TH amputees with TMR who had participated in a 6-8 week home trial. In that study, Hargrove et al reported that functional outcome metrics improved after home use. Furthermore, and of most relevance to our study, they found that 7 of 8 subjects preferred EMG-PR to direct myoelectric control of the same prosthesis. However, they did not provide any detailed qualitative data or explore the reasons why users indicated that they preferred this control method. However, findings from our study are not directly comparable to prior studies of EMG-PR control of upper limb prostheses because of the greater number of DOF of the DEKA Arm. In contrast to Hargrove et al, we found that the majority of prosthesis users who completed an in-home trial expressed a preference for the controls of their personal prosthesis, rather than the EMG-PR control. While our study demonstrated that EMG-PR is a viable option for control of the DEKA Arm or other multi-DOF devices, it also illustrates challenges using this method with multi-DOF devices. We found that only 29% of home users of the EMG-PR controlled DEKA Arm stated that they wished to receive one in the future, and another 29% thought that they might want to receive one. Participants in our study made repeated remarks about the complexity of the control system and the challenges of calibrating due to numerous DOF, the added challenges of using a prosthesis that many perceived was heavy, and the need to repeated calibrations. Although some participants recognized the functional advantages of using the system, others did not. Most commented that they became fatigued when using this controls scheme, and reported that fatigue negatively impacted their ability to use the EMG-PR proficiently. Participants also described a process of learning and acclimation and recognized the need for sufficient prosthetic training. Similar to any surface myoelectric control, it appeared that EMG-PR control was susceptible to degradation caused by excessive sweating, and changes to prosthesis fit. Others have raised concerns about the clinical robustness of surface EMG-PR due to problems with electrode shifting, limb orientation, and degradation of signals with repeated activation. These concerns have led to many attempts at improving robustness of classifiers, as well as recommendations for training EMG-PR by maximizing sources of variation during the calibration (i.e. controls training) process, using visual feedback, and focusing on phantom limb movements. While most of our team was new to training participants using this control methodology, our team, with guidance from Coapt LLC specialists, attempted to utilize this training advice in our protocol to improve robustness of the EMG-PR controls. Nevertheless, our data shows that many of the concerns about reliability and robustness of EMG-PR remain. Others have recognized that training of EMG-PR (calibration) aimed to maximize variation is potentially burdensome and “prohibitively intensive.” Feedback from our subjects demonstrates that many agree that calibrating in many positions is burdensome. The 3 participants in our study who had experienced both control types had mixed viewpoints on EMG-PR compared to IMU controls of the DEKA Arm. Our findings did not indicate, as we had initially expected, that operating the DEKA Arm using EMG-PR would be easier to learn and less cognitively demanding than operating the device using IMU controls. Additional analyses of quantitative metrics directly comparing participants who used EMG-PR to those who used IMU controls are currently underway and will be published elsewhere. Our study has several limitations. The study was not designed to assess the effectiveness of the training approach for use of the DEKA Arm and EMG-PR control. We cannot say with certainty whether or not our study participants would have had more favorable views of EMG-PR control of the DEKA Arm if they had been given additional training and time to acclimate. Despite the fact that all participants rated the amount of training that they had received as “just right”, we found that 4/7 who used the system at home rated their overall skill level as poor or fair. This finding suggests that additional training may have been beneficial. We do not know how our training quantity compared to that provided in the Hargrove et al study, because that study did not describe the amount of training provided. We recommend further study to identify the optimal timing and dosage of prosthetic training required to optimize results for complex multi-DOF devices like the DEKA Arm. Both of our participants with TH amputation emphasized the need for sufficient training to master the extra DOFs using EMG-PR control. The participant who felt he mastered controls completely received 30 hours of training. However, he was an experienced using EMG-PR, albeit with a prostheses that had fewer DOF. This finding suggests that more than 30 hours of training may be needed for similar individuals. Like the participants in Hargrove et al’s recent study, the two TH amputees with a history of TMR in our study were very positive about EMG-PR as a control system. In contrast, person with TR amputation, who needed to control fewer DOFs had more mixed responses to EMG-PR. However, due to our small sample size and the fact that there were no TH participants who had not undergone TMR, we cannot conclude that TH amputation level in and of itself is associated with greater satisfaction with EMG-PR compared to TR amputation level. Because of small sample sizes within sub-groups, we did not try to contrast the experiences of participants by years of prosthesis use, or type of prosthesis used at baseline. Future studies could examine differences in user experience within a variety of sub-groups. Surface EMG-PR is thought to be more robust than direct EMG control because information from patterns of muscular recruitment derived from multiple electrode sites is used. Our findings suggest that prosthetic socket fit and changes in electrode contact within the prosthetic socket may have influenced the consistency of EMG-PR controls. That, said, we did not utilize any standardized metrics to evaluate prosthetic fit in our study, and cannot quanitfy the extent to which problems in control consistency stemmed from socket fit, not the control system in and of itself. Given that fitting of upper limb prosthetic sockets is more of an art than a science and that limb volume and shape may fluctuate which weight gain, fluid retention, and hot and humid conditions, there may be inherent vulnerabilities in utilizing surface EMG-PR systems. We believe that the quality of the prosthetic services delivered in our study were well above what would be available in the wider clinical community. Our study included several very well trained upper limb prosthetists who had expertise in fitting persons with upper limb amputation. Future studies could evaluate the relationship between prosthetic socket fit, consistency of electrode contact, and EMG-PR control consistency. We found it challenging to disentangle participant preferences for control scheme from the characteristics of the DEKA Arm itself. There are unique aspects of the DEKA Arm compared to other prostheses These factors (such as number grips, wrist design, number of DOFs, heavier weight, required external wires and battery pack) may have impacted users (both negatively and positively) and thus colored their perceptions of overall prosthesis desirability and functionality. Given that ours is the first study to carefully describe user experience with EMG-PR controls, we were not able to identify or measure all relevant confounding/contributing issues. Future studies of EMG-PR might include quantitative metrics on the effect of the system on battery life. Our findings are specific to EMG-PR control of the DEKA Arm and not EMG-PR control more broadly. We cannot say whether or not our participants would have preferred EMG-PR for control of their personal devices, or other devices with fewer DOF. We also cannot be sure if the fatigue associated with using EMG-PR might have been less of a factor with a lighter weight device. In our view, some amount of physical fatigue would be expected in the early phases of training as users cope with increased muscular demands. We believe that our participants’ experience was impacted by technical issues originating in the EMG-PR/DEKA interface. Our study was the very first to test this new interface, and thus we identified compatibility issues, which no doubt would be addressed in future prototypes. EMG-PR control of grip select was considered by Coapt LLC to be an experimental feature. To our knowledge, no other research teams have attempted to toggle grip select with EMG-PR. This was inherently difficult, given that there is no natural physiological movement that is linked to this action. This may explain, in part, why there were so many issues using EMG-PR for this function. # Conclusion This study provided an in-depth description of the user experience of operating the multi-DOF DEKA Arm using EMG-PR control. Major aspects influencing the user experience related to the complexity of using EMG-PR system, the process of calibrating, and the functional benefits of using EMG-PR with the DEKA Arm. These aspects were largely explained by the technical issues, muscle fatigue, prosthesis design, technical issues with our EMG-PR prototype, training and acclimation. Factors influencing the user experience included training and acclimation, fatigue, aspects of the prosthesis design such as the weight, technical issues requiring repairs, changes to the controls. Broader contextual factors, both personal and environmental also impacted the user experience. An understanding of the patient experience of using EMG-PR to control a DEKA Arm, and factors that may affect that experience will be helpful for clinicians and patients who are choosing prosthetic control and hardware options. We conclude that operating the multiple DOFs of the DEKA Arm using EMG-PR is more difficult to learn and no less cognitively demanding than operating the device using IMU foot controls. We found that the majority of participants expressed a preference for the controls of their personal prosthesis rather than the iteration of EMG-PR controlled DEKA Arm used in this study. Most home users were also positive about the future potential of EMG-PR control, or about the control scheme used in our study given further development. Both participants with TH amputation and a history of TMR, were particularly enthusiastic about EMG-PR for the number of DOFs required to operate an HC DEKA. Our conclusions are limited to the prosthesis/control combination that we tested. Future studies are needed to evaluate whether the iteration of EMG-PR we tested would be experienced differently by users if it were interfaced with a prosthesis other than the DEKA Arm, particularly one with fewer DOFs. # Supporting information The information in this manuscript does not necessarily reflect the position or policy of the government; no official endorsement should be inferred. The view(s) expressed herein are those of the author(s) and do not reflect the official policy or position of the U.S. Government. [^1]: The authors have declared that no competing interests exist.

# Introduction Alzheimer’s disease (AD) is a neurodegenerative disorder which is characterised by progressive accumulation of amyloid-beta (Aβ) peptides, forming senile plaques in the brain, and neurofibrillary tangles of hyperphosphorylated tau, followed by progressive neuronal and synaptic loss. It has long preclinical and prodromal stages and affects memory, thinking, behaviour, daily functional abilities, and emotion. Alzheimer’s disease may account for approximately 60–80% of dementia cases. The worldwide prevalence of dementia was estimated to be over 46 million people, with studies suggesting that the number of cases is rising exponentially with age and that the above estimate will almost triple by 2050. Apart from the health and social problems, Alzheimer’s, and dementias in general, have huge economic impact, with the estimated worldwide cost of dementia approaching USD 818 billion. Thus, drugs are urgently needed to improve therapies of AD and dementia. There are two competing models proposed to explain the hallmarks of AD: the amyloid hypothesis (the neuron-centric model) and the Inverse Warburg hypothesis (the neuron-astrocytic model). The neuron-centric model suggests that a mutation in the nuclear genome induce overproduction of Aβ and tau that become toxic to neurons. The neuron-astrocytic model contends that the progression of AD is triggered by defects in the normal energy transduction process, a condition induced by mitochondrial dysregulation. Although some clinical trials of metabolic interventions have shown promising results for the improvement of cognitive performance, clinical trials of disease-modifying treatments have consistently failed. A recent study reviewed 413 clinical trials (124 Phase 1 trials, 206 Phase 2 trials, and 83 Phase 3 trials) of AD drug candidates that have been conducted between 2002 and 2012. The study reported that 99.6% of the trials failed or had been discontinued. No cure or other treatments that prevent, halt or reverse the underlying pathology of established AD have been approved as yet. Available treatments, such as the cholinesterase inhibitors tacrine (1993), donepezil (1996), rivastigmine (1998) and galantamine (2001), are approved for mild to moderate AD and can provide only temporary improvement in cognitive and behavioural symptoms, and at best a temporary impact on the progression of the underlying pathology of the disease. The failure rate of AD clinical trials is far higher than that of trials in other therapy areas, and information on the reasons of trial failure or discontinuation is limited. Some of the reasons of high trial failure include the poorly understood nature of AD pathogenesis and progression and poor trial design. For example, the wrong choice of clinical endpoints and markers used as clinical trial entry criteria, as well as the selection of unsuitable individuals for enrolment in the trial. The retention of participants due to the long duration of the AD trials is also an important problem to overcome in the design of successful clinical trials in the AD area. The high failure rate of clinical trials can also be attributed to the inability to detect signals of efficacy, even if the treatment is safe and effective. This could be due to the administration of the drug at a relatively late stage of the disease, at which time the condition of the patient may be irreversible, or due to the high variability in the measurements of different biochemical and cognitive markers that are used as diagnostic markers and endpoints. Mathematical modelling of AD progression and the development of clinical trial simulations are essential tools for exploring the reasons why a clinical trial can fail and for the improvement of the design of clinical trials. The poorly understood nature of AD pathogenesis and progression limits the possibilities of developing robust mechanistic models for the accurate prediction of disease progression. However, a few attempts have been made to develop models that describe the development of AD at the molecular level. Published mathematical models also include semi-mechanistic models developed to describe longitudinal changes in measures of cognition, as assessed by errors in various cognitive tests designed for the assessments of the cognitive capabilities of the patients, such as the Modified Mini-Mental State Examination (MMSE) and the Alzheimer’s Disease Assessment Scale (ADAS-cog). In general, published research on the mathematical modelling of AD has mainly focused on: (1) the study of the epidemiology of AD; (2) the investigation of the progression of the disease based on changes in the cognitive status; (3) on defining the order of the various biomarkers of the disease in predicting progression, and (4) on the description of the pathogenesis and evolution of the disease at the molecular level (e.g.). Most of the models developed so far are stochastic in nature. Such models permit variability in the model parameters and chance elements in AD disease progression. In this study we develop and analyse a simple stochastic model to predict AD progression in a clinical trial. In particular, we define a discrete-time Markov model to describe the movement of individuals through a finite sequence of distinct health and disease states over time (distinct time points at which the state of the individuals is assessed). Markov models are ideal for the modelling of AD, as this disorder can be described as a multi-state disease process, and due to the observational nature of studies of AD in which the condition of individuals is assessed at periodic visits. These are also useful tools for the estimation of the impact of risk factors on the transitions between the states and can easily include reversible states that have been observed in many AD cohort studies. This modelling framework can contribute to the improved design of future clinical trials by providing insights into what to measure, the importance of measurement error in determining efficacy, sample size selection, duration of the trial and other important factors such as demographic characteristics of the sample. Markov models have also been purposed in previous studies for the estimation of the probabilities of transitioning between different stages of AD severity, and the investigation of potential covariates that can influence these probabilities. In this study, we use this framework to develop tools for the assessment and evaluation of the impact of receiving a hypothetical preventative treatment of varying efficacy, at different stages of disease progression, and illustrate the potential problems detecting significant differences between treated and untreated individuals. The specific action of the treatments in the model is irrelevant; the treatments are assumed to reduce the probability of transitioning to a more severe stage of the disease, which consequently slows the development and progression of the disease. The transition probability distributions of the Markov model we describe are estimated using a particular longitudinal cohort dataset provided by the Alzheimer's Disease Neuroimaging Initiative (ADNI, [adni.loni.usc.edu](http://adni.loni.usc.edu)). # Material and methods ## The model The development and progression of AD in a clinical trial was assumed to be a stochastic process. In particular, we developed a discrete-time Markov chain in a random environment defined on a finite state space, with the property that for all states the probability of moving from the current state to the next state is independent of the past. Individuals were classified according to their disease status. In particular, based on a set of criteria individuals were categorised as Cognitively Normal (CN), with Mild Cognitive Impairment (MCI) and with AD. The criteria are based on the level or presence of a number of markers, including but not limited to memory complaints, logical memory II subscale, MMSE, CDR, and NINCDS/ADRDA criteria for probable AD. The states are distinct and do not overlap. In each (discrete) time step of the process, individuals can either withdraw from the study, for example due to death or personal reasons, remain at the same state or move to a different state with some probability (transition probability). Based on the information obtained from the ADNI dataset, transitions are allowed between the states CN and MCI in both directions (i.e. recovery to normal cognition is also allowed) and from the MCI to the AD state. Back-transitions may be due to either measurement error or a genuine ability in some patients to show cognitive improvement (e.g. due to cognitive reserve). In the ADNI dataset, a small number of transitions from AD back to MCI is also observed (a total of 17 transitions in yearly assessments across the ten-year cohort). Some reverse transitions from AD to MCI, as well as from MCI back to CN, have also been recorded in other longitudinal studies. However, due to the progressive nature of AD and the fact that AD diagnosis cannot be confirmed until an autopsy is performed, in the theoretical model that we developed we do not allow transitions from AD back to MCI state, i.e. once patients reach the AD state their condition cannot improve. Even in the presence of candidate treatments that could potentially reverse the pathology at a stage of AD , there is no sufficient evidence that the cognitive functions and clinical outcome of AD individuals can be improved. Due to the stochastic nature of the model, allowing both an MCI to CN and an AD to MCI transition would even mean that there will always be a chance of patients becoming fully recovered from AD, i.e. returning to a cognitively normal state, an outcome that cannot be supported at present. However, it should be noted that incorporating an AD to MCI transition in the model would have a negligible effect on the output due to the small chance of this transition happening. A small number of patients in ADNI also transitioned from CN directly to AD (3 patients in yearly assessments across the ten-year cohort). These transitions have also been attributed to potential misclassifications of patients and not been incorporated into the model. A schematic diagram of the model developed is presented in. ## Transition probabilities For the estimation of the transition probabilities between the different states we followed a piecewise approach. We divided the observation period into one- year intervals and calculated the transition probabilities in each of these intervals (in ADNI, data exists for a total of 10 years). The probabilities of transitioning between states were estimated independently for each time interval, and it was assumed that the process is temporarily homogeneous within these intervals, i.e. the transition probabilities were assumed constant within each of the time intervals. Let $p_{S_{i},S_{j}}(t)$ be the transition probability to state S_*j* at time *t* given that at time *t* − 1 the state is S_*i*, where *i*, *j* ∈{1, 2,…, *K*} and S_*i*, S_*j* ∈*S*. *S* is the state space of the Markov Chain and *K* is the number of states. The maximum likelihood estimate of the transition probability $p_{S_{i},S_{j}}(t)$ is given by $$p_{S_{i},S_{j}}(t) = \frac{\tau_{S_{i},S_{j}}(t)}{\sum_{l = 1}^{K}{\tau_{S_{i},S_{l}}(t)}},$$ where $\tau_{S_{i},S_{j}}(t)`$ is the number of transitions from S_*i* (at time *t* – 1) to S_*j* (at time *t*). ${\sum\limits_{j = 1}^{K}p_{S_{i},S_{j}}} = 1$ for all *i*. In order to assess the uncertainty of estimates of each transition probability we used the bootstrap method. With this method, a number of possible cohort datasets are generated to develop a range of possible transition matrices. Specifically, for each time interval new sets of transitions are generated by sampling with replacement from the original set of transitions (the number of transitions in the generated samples is the same as the total number of transitions in the observational study). Combining all the bootstrapped transition matrices we derive an approximation of the sampling distribution. To smooth the distributions of the transition probabilities we fitted a Normal- distribution using the maximum likelihood estimation method. The accurate estimation of transition probabilities is particularly challenging, mainly due to the heterogeneity of the data, the infrequent observations on the same patient, and missing data points. It should be noted that although the total number of individuals in the study that we used is relatively large (1624 individuals, see subsection ‘Dataset used’), this number decreases significantly over follow-up, either because individuals die or are lost to follow-up. It is also important to note that our analysis uses follow-up time as opposed to chronological time, such that those subjects recruited later (into the newer ADNI-GO or ADNI-2 cohorts, see subsection ‘Dataset used’) will have inherently less follow-up time than those enrolled initially into the study (ADNI-1). Thus, we naturally lose depth in the number of subjects at later follow-up times due to recruitment date. ## Stochastic simulation Within the modelling framework described, we numerically simulated the time progression of the disease in a sample of individuals in a clinical trial. As an example, in this preliminary study the development and progression of the disease was simulated in a general population of 1000 individuals for duration of up to 10 years. The trial duration was discretised into uniform steps. The time-step was defined to be one-month. At the beginning of each simulation, a value for each transition probability was assigned to each individual in the trial. The value was drawn at random from the respective distribution of the average one-year transition probability. In each time-step in the simulation (same for all individuals), each individual moved independently from one state to another state with the defined transition probability. The state of each individual was updated synchronously. For each numerical simulation we performed 10⁴ realisations. We focused on the proportion of AD cases over time. We are mainly interested in the effect of potential preventative treatments, before the occurrence of AD clinical symptoms. For this reason, we considered optimistic treatments focusing on the case where at the beginning of the process all individuals are cognitively normal. The other extreme where all individuals are initially at the MCI state was also considered. It should be noted that, given the assumption that AD is a progressive disorder, and thus AD is an irreversible state, including a number of AD individuals in the initial population would have the same effect as removing these individuals completely from the process. The simulator was developed in MATLAB (R2017b). The computer code can be obtained from the authors upon request. ## Impact of interventions Interventions were applied at different time points of the process. In our analysis we focused on treatments that, irrespective of their specific action, reduce the probability of transitioning from state X to a more severe state Y by a proportion *E*_*X*,*Y* ∈ (0,1). Despite the nature of the dataset we used, we focused on preventative treatments that are administered to a sample of cognitively normal individuals in order to reduce the probability of transitioning to MCI and later from MCI to AD. The same treatments were also assessed in a sample consisting of only MCI individuals. We also simulated clinical trials to evaluate the effect of treatments that are administered to a sample of only cognitively normal individuals, but the treatment becomes effective only at the MCI state and reduces the probability of developing AD. The effect of a treatment that reduces the probability of transitioning from CN to MCI has a negligible effect when administered to individuals that have already reached the MCI state. To investigate the impact of treatments that might not be effective beyond a point, as may happen, for example, due to severe neuronal degeneration, we also considered treatments that are administered before the MCI onset to reduce only the probability of developing MCI, i.e. they are no longer effective beyond the MCI state. Indeed, the model also allows the consideration of the effectiveness of treatments that might vary with time, but we did not consider this scenario in the current study. All the treatment effects were assumed to be homogeneous between individuals. We focused on the ideal scenario where treatments have an immediate effect, but we also considered cases where there are time delays, which may refer either to delays in the administration of treatments that are immediately effective or delays before treatments that are administered at the beginning of the trial start to be effective. Indeed, if a clinical trial involving ideal treatments which have an immediate effect and their effectiveness is not reduced over time is unsuccessful, we expect that clinical trials of less effective treatments (example, clinical trials of treatments which have delayed and/or time-decaying effects) will also be unsuccessful. We assessed the impact of interventions on the proportion of AD cases over time. The effect of a treatment at a specific time point is defined as the mean difference between the proportion of AD cases in the untreated and treated groups. Other potential endpoints that could be studied using the model include the incidence proportion of AD cases, the expected time spent in a disease state, and the expected time and probability to reach a specific state from any other state. Information for each of these endpoints is generated in the simulation. For every simulation output we present the predicted value of the proportion of AD cases at the end of a 5- and a 10-year trial under different intervention scenarios. We calculated the 95% credible interval of the distribution of the proportion of AD cases and the standard deviation. Although outside the scope of the current study, it should be noted that the results of such simulation exercises could also be used as prior information required for sample size calculations before the conduct of a clinical trial. For every case, we performed a hypothesis test for the equality of two binomial proportions, the proportions of AD cases in the treated and untreated groups, i.e. we tested whether the difference of the two proportions is equal to 0. A two-tailed test was performed and statistical significance was calculated using a 95% confidence level. ## Dataset used For the development of the Markov model and the estimation of the model parameters we used data obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database ([adni.loni.usc.edu](http://adni.loni.usc.edu), last downloaded 2016/10/31). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). For up-to-date information, see [www.adni-info.org](http://www.adni-info.org/). The dataset used is a longitudinal observational study consisting of 1737 individuals (from the ADNI-1, ADNI-GO, and ADNI-2 protocols), 106 of which were categorised as ‘Significant Memory Concern’ (SMC) at baseline and excluded from our analysis. In addition, there were 7 individuals that had only a screening visit and they were also excluded. Therefore, in the present study, we analysed data from 1624 participants. The characteristics of the individuals in the dataset considered are presented in Supporting Information. The maximum duration for which a participant in the study has been followed up is 10 years. It should be noted that in the original dataset, individuals in the MCI state are further classified as ‘Early MCI’ (EMCI) and ‘Late MCI’ (LMCI). However, in this analysis we combined the two MCI subtypes. The status of individuals was assessed only at the distinct time points considered in the model. The number of transitions between the states of the Markov model at each time point is presented in Supporting Information. # Results illustrates the distributions of the average one-year transition probabilities (for the parameters of the fitted Normal distributions see)., and show the output of stochastic simulations (expected proportion of AD cases over time in a general population of 1000 individuals) in the untreated and treated groups under different intervention scenarios. In all the examples presented in these figures we assumed that individuals are initially at the CN state. shows the output of simulations for the case where treatments reduce the probability of transitioning from CN to MCI and from MCI to AD, and all individuals are initially at the MCI state. In each of the above figures, the uncertainty in the output of the stochastic simulations in both the treated and untreated groups is also illustrated. The expected values of the proportion of AD cases at the end of a 5- and a 10-year simulated trial where all individuals are initially cognitively normal are provided in. In we present the results obtained at the end of a 5-year simulated trial for the cases where all individuals are initially at the MCI state. For each case, the 95% credible interval and the standard deviation are also presented. In and Tables we present the *p*-values of the hypothesis tests for the equality of the proportions of AD cases in the untreated and treated groups under the different intervention scenarios that have been investigated. It is observed that although treatments might be effective in reducing the chance of developing AD, due to the high uncertainty it might be difficult to assess their real effectiveness in a clinical trial. In particular, it is remarkable that in the sample of cognitively normal individuals alone, the simulation suggests that in a clinical trial of duration of five years, it is likely that the true efficacy of a treatment will not be detected even if the treatment has moderate efficacy (and Tables). As expected, if all the participants are initially at the CN state, a difference between the untreated and the treated group becomes even more challenging if the treatment is effective only when individuals are at the MCI state (and Tables). In particular, in a sample of cognitively normal individuals, treatments that reduce both the transition probability of moving from CN to MCI and the probability of moving from MCI to AD are the most effective, followed by treatments that reduce only the probability of transitioning from CN to MCI and then the treatments that reduce the probability of developing AD from the MCI state. Increasing the duration of the trial and the population size improves the chances of clinical success in all cases (see for example and Tables). The efficacy of a treatment that reduces the probability of transitioning to more severe states may be more likely to be detected in a trial where all individuals are initially at the MCI state (and Tables). However, the proportion of AD cases at the end of such a trial is much higher. As mentioned earlier, in the analysis performed we showed that the effect of a treatment that reduces the probability of moving from CN to MCI state is negligible when it is administered to MCI individuals. illustrates the impact of time delays, before the treatment starts to be effective, on the proportion of AD cases at the end of a 5-year simulated trial. The treatment is administered to reduce the probabilities of transitioning to more severe states (*p*_*CN*,*MCI* only, *p*_*MCI*,*AD* only, *p*_*CN*,*MCI* and *p*_*MCI*,*AD*). The impact of such delays on the proportion of AD cases at the end of a 10-year trial in the same scenario is illustrated in. In both and all individuals are initially cognitively normal. shows the impact of time delays on the effect of a treatment that reduces the transition probabilities from CN to MCI and from MCI to AD when the initial population consists of only MCI individuals., as well as and Figs, can also be thought of as the case where there is a period of delay before the administration of treatments that are immediately effective. Such delays influence the population distribution at the time of treatment administration; the initial number of CN individuals decreases and the number of MCI and AD individuals increases. It is shown that treatments administered to CN individuals are likely to be more effective in slowing AD progression if they are administered at the beginning of the study (*t* = 0) to reduce the probabilities of moving to MCI and later to AD state (see, and Tables). However, the impact of time delays is more pronounced in the cases where treatments given to CN individuals reduce their chance of transitioning to the MCI state (but see ). This is partly due to the fact that delays in the ‘activation’ of the treatment result in a reduction in the number of CN individuals who can benefit from the treatment, as a consequence of disease development between the initiation of the study and treatment activation. Consequently, after some delay, treatments administered to CN individuals to reduce the probability of developing AD from the MCI state may be more effective than treatments administered to CN individuals to reduce only the probability of moving to the MCI state. Hence, in a clinical trial the specific time of intervention is crucial. # Discussion Using the ADNI dataset, we developed a basic framework for the simulation of the development and progression of AD in a clinical trial to facilitate the assessment of the effect of potential candidate treatments. Treatments slow the transition to a more severe state, and thus disease progression. This reduction in the progression rate of the disease can be associated with a significant decrease in the incidence of AD, increase in the duration of individuals’ independence and the time until institutionalisation, and concomitantly a significant decrease in the costs of care. It is interesting to note that if such interventions could delay both disease onset and disease progression by even one year, the global prevalence of AD could decrease by almost 9.2 million, by the year 2050. It has been demonstrated that the chance of identifying the true treatment effect may be challenging due to the high uncertainty in the clinical trial outcome, even if the treatment has high efficacy. This can be one of the dominant causes of trial failure of AD treatments. The high uncertainty, which can be illustrated by the output of the simulations, can be attributed partly to the sparsity and heterogeneity of the available data, and the high variance in the measurements of markers that have been used for the classification of individuals in the different disease states. The simple criteria used for the classification of individuals, such as the performance on cognitive tests, and potential associated measurement error, may influence the correct diagnosis and could be an important reason of the presence of the high variability. Other reasons that can introduce variability in the measurements, both within and between individuals, and influence the diagnosis of the actual clinical stage include underlying medical conditions such as the presence of depression or psychosis, inter-clinician differences in application of diagnostic criteria, and resistance to cognitive decline due to cognitive reserve. As some of these factors can significantly contribute to variance in the model parameters representing disease progression, they should potentially be considered as covariates in extended models. It should be noted that, due to the nature of the ADNI data used in this study, our outputs may not reflect the progression and impact of the disease in the general population, and indeed the output in prevention clinical trials considered here. Participants in this dataset have been selected to mimic that of a population that would be recruited for a clinical trial of a potential AD- drug, such as from memory assessment clinics. For example, as observed in Supporting Information, in the dataset used in this study there is a high proportion of ApoE ε4 carriers, which is a significant genetic risk factor of AD. Thus, these individuals are at higher risk of developing the disease, creating significant selection bias issues. In addition, in the model we assumed long preclinical and prodromal AD stages and ADNI dataset is not sufficient to support this silent ‘incubation period’. Due to the nature of the data, we also believe that the real effect of hypothetical preventative treatments, as assessed in the current study, is underestimated, as CN individuals in ADNI move faster to the MCI state than expected in the general population. However, all this should not limit the applicability of the model and the major outputs of this work which illustrate the usefulness of clinical trial simulations in AD therapy area, and how the high variability in measurements could mask the true efficacy of potential AD treatments. Despite the nature of AD and problems with the available data, the Markov model is a powerful tool for predicting the development and progression of the disease and the assessment of the impact of potential treatments in clinical trials. Ideally, the model described in this paper should be extended so that there are more disease states, given there is enough information for the derivation of reliable estimates of the model parameters. The model should also be extended for the estimation of covariate effects of risk factors on transitions among states and concomitantly on disease progression. Such covariates could include age, genetic background (for example presence/absence of the apolipoprotein ε4 (APOE ε4) genotype), gender and education, all of which are important risk factors for AD. Other factors, such as family history, capability, lifestyle (e.g. alcohol consumption), social status, health and environment can also be influential. Adding such heterogeneity will be described in a subsequent publication. It would also be interesting to incorporate into the model changes in cognitive markers and biomarkers, such as beta-amyloid and total-tau, as well as MRI markers, even though the precise estimates of the transition probabilities over time, conditional on the levels of these markers, might be difficult using the current data. Clinical trials are difficult to conduct and very expensive if carried out in large samples of patients observed over many years. Modelling and simulating AD trials are relatively low cost activities and they can be a powerful tool for improving the design of the actual clinical trials and increase the chances of the accurate assessment of treatment efficacy. Improving the quality of the data is clearly essential for the determination of the importance of variance in clinical trials. The reduction of the variance in currently employed measures (whether biomarkers, brain scans or cognitive scores) will improve trial design, facilitate detecting a signal in the trial, shorten the trial times and reduce the number of patients enrolled. Ideal datasets should include frequent sampling and repeated assessments for the same individual. Utilising the growing number of available AD datasets from many different countries, and performing systematic and standardised analyses, it should be possible to provide substantial information for improving clinical trial design. However, extra caution will be needed when working with the different datasets. Among the available studies there is a variation in the diagnostic tools, in the markers measured and measurement methods used. There is also a little agreement on the definition of the disease states and selection criteria as well as the size and demographic characteristics of the samples. Pooling and standardisation of clinical trial data, as well as standardisation of the methods employed in future studies, will facilitate better prediction of disease course and allow for better interpretation of model predictions. It will help in the identification of the sources of variance, and consequently it will facilitate progress towards the precise assessment of new prophylactic and therapeutic treatments of AD. # Supporting information We would like to thank Kevin McRae-McKee and Stephanie Evans for providing advice regarding the statistical analysis of the results. Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database ([adni.loni.usc.edu](http://adni.loni.usc.edu)). The Principal Investigator of ADNI is Michael W. Weiner, MD (email: <Michael.Weiner@ucsf.edu>). A complete listing of ADNI investigators can be found at: <http://adni.loni.usc.edu/wp- content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf> The investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health ([www.fnih.org](http://www.fnih.org/)). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. [^1]: RMA is a non-executive board member of GlaxoSmithKline (GSK). GSK played no part in the conduct, financing or publication of this research. This does not alter our adherence to PLOS ONE policies on sharing data and materials. [^2]: ¶ Membership of the Alzheimer’s Disease Neuroimaging Initiative is provided in the Acknowledgments.

# Introduction Night-shift work has negative effects on human biological adaption to the natural light and darkness cycles and cause health problems such as sleep disorders, gastrointestinal diseases, cardiovascular diseases, and cancers. Night-shift work is defined as fulltime work that occurs between 24:00 and 05:00. In developed countries, about 20% of all employees perform work exceeding the regular working hours, and this proportion is likely to increase. Therefore, the negative effects of night-shift work on health-related issues are being recognized now. Among the possible negative health effects of night-shift work, mental health has been less investigated compared with physical effects, and previous studies presented inconclusive results. Alcohol use disorders (AUDs) include dependence, abuse, and harmful statements of alcohol use and increase in the proportion of health-care burden worldwide. Night-shift workers who suffered poor sleep quality exhibit higher levels of alcohol consumption and are at risk of AUDs. However, previous studies regarding the association between night-shift work and AUDs showed inconsistent results and remain a source of controversy. Some studies revealed no difference in alcohol consumption patterns between day and night workers, while a study conducted on Norwegian nurses denoted a significantly positive association between hours worked per week and alcohol consumption. A study conducted in an East Asia showed that service and sales workers had a greater risk of alcohol consumption than managers and professionals. Korea is one of the countries in Asia with the highest prevalence of alcohol consumption, with its general population having a high-risk alcohol consumption rate of 15.1%. Health-related quality of life (HRQL) refers to the subjective evaluation of the culture, society, and environment in which people are embedded. One of the major factors affecting HRQL is the working status, such as occupation type, employment status, working hours, workplace condition, and shift work, including duration and frequency. Women involved in night works or rotating shifts showed lower HRQL compared with day workers. However, a study conducted among Croatia hospital nurses denoted non-significantly lower HRQL in shift workers compared with non-shift workers. Although the prevalence of night-shift work has increased and various health- related effects have been suggested, a limited relationship was established between night-shift work and its effects on mental health, especially in the Asian population. Thus, this study aimed to examine the association between night-shift work and mental health focused on AUDs and HRQL applying nationally representative samples in an East-Asian country. # Materials and methods ## Data source and study population Data were obtained from the fourth (2007–2009), fifth (2010–2012), and sixth (2013–2015) Korea National Health and Nutrition Examination Survey (KNHANES) conducted by the Korea Centers for Disease Control and Prevention. The KNHANES is a nationally representative survey and used a multistage stratified cluster sampling method to the access data regarding the health and nutritional status of the Korean population since 1998. Health interview, health examination, and nutrition survey were the three main components of the KNHANES. The details of KNHANES were described elsewhere. From 2007 to 2015, the number of study population was 81,411, with a 75% overall response rate. The study population was restricted to individuals aged 20–60 years, who responded to the questionnaire about working status, including 36,675 adults. Additionally, 9,780 individuals who had not had a job (N = 8,877), engaged in split work (N = 194), had a 24-hour rotating work (N = 154), with irregular working schedule (N = 288), or reported other working schedules (N = 267) during the last 1 year were excluded from the study. A total of 26,895 adults were included in the study. The KNHANES was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention (IRB no. 2007-02CON-04-P, 2008-04EXP-01-C, 2009-01CON-03-2C, 2010-02CON-21-C, 2011-02CON-06-C, 2012-01EXP-01-2C, 2013-07CON-03-4C, 2013-12EXP-03-5C). All procedures were performed in accordance with the Declaration of Helsinki 7^th version and informed consent was obtained from all participants. ## Measurements ### Working schedule and other covariates In the KNHANES, after asking the participants if they have had a job during the last 1 year, the working schedule of the participants who ever had worked during the last 1 year was assessed. The major working schedules were classified as follows: mainly daytime work (6 am–6 pm), evening work (2 pm–midnight), night work (9 pm– 8 am the next day), day-night regular shift work, 24-hour rotating work, split shift, irregular shift work, or others. In this study, participants were considered as having day work if they had worked between 6 am and midnight in combination with daytime (6 am–6 pm) and evening work (2 pm–midnight), and night-shift if they had night work (9 pm–8 am next day) or day-night regular shift work. However, in the analysis, we divided night-shift work into two categories: night work and day-night regular shift work. Participants who reported other working schedules were excluded from the analysis due to the small number of participants. The baseline characteristics and lifestyles were recorded through interviewer-guided questionnaires, including age, gender, education, alcohol consumption, smoking status, average number of hours of sleep, stress, and physical activity. ### Alcohol use disorders The AUDs were assessed using the Alcohol Use Disorders Identification Test (AUDIT), with a scale from 0 to 40. Briefly, this test includes 10 questions related to hazardous alcohol use (frequency of drinking, typical volume, and heavy drinking), symptoms of alcohol dependence (ability to stop drinking, ability to control normal activities, and morning drinking), and harmful alcohol use (remorse after drinking, blackouts, alcohol-related injuries, and other concerns about drinking) as a simple method of screening for hazardous and harmful alcohol consumption in primary care setting. We used the AUDIT cutoff value of ≥8 points to define AUDs. This was proven to give an appropriate level of sensitivity and specificity in unselected populations. Additionally, based on the WHO guidelines, AUDIT scores were categorized into four zones with their corresponding interventions: zone I, 0–7 points: alcohol education; zone II, 8–15 points: simple advice; zone III, 16–19 points: counseling and continued monitoring; and zone IV, 20–40 points: referral to specialists for diagnostic evaluation and treatment. ### Health-related quality of life HRQL was measured by EuroQol-5D (EQ-5D), which was developed by the EuroQoL Group. EQ-5D was descriptively quantified by five dimensions, namely mobility, self-care, usual activities, pain/discomfort, and anxiety/depression. Each of the dimension was described as follows: 1  =  “no problem,” 2  =  “some problems,” and 3  =  “severe problems.” All of these response levels were converted into an EQ-5D index using the weight scoring system of the five dimensions ranging from 0 (worst) to 1 (best). The five dimensions were used as dichotomous variables by merging levels 2 and 3 into “some or extreme problem.” ## Statistical analysis KNHANES data were analyzed, by applying sampling weights based on the inverse of selection probabilities, inverse of response rates, and other post- stratification factors including age, sex, and metropolitan area or province category. The baseline characteristics were displayed by number, weighted percent, and p-values for chi-square tests for each working group. The mean scores of AUDIT and EQ-5D were calculated and compared using weighted ANOVA with the Tukey’s post hoc test. The associations between night-shift work and AUDs as well as problem in HRQL 5 dimensions were evaluated using the complex samples logistic regression model adjusted for age, gender, education, smoking status, average number of hours of sleep, stress, and physical activity for the model of AUD. In the analysis between night-shift work and HRQL, we added alcohol drinking and removed stress as adjusted variables. In addition, we used the complex samples multinomial logistic regression model to analyze the association between night-shift work and the AUDIT of four categories. The interaction between night-shift work and gender was assessed by adding the interaction term into each model. The results of the analyses were expressed as odds ratios (ORs) and 95% confidence intervals (95% CIs). Data analysis was performed using SAS version 9.4 (SAS Institute, Cary, NC, USA). # Results The characteristics of the study population consisting of 13,571 (50.5%) male workers and 13,324 (49.5%) female workers are presented in. There were significant differences in gender, age, education, occupation, alcohol consumption, smoking status, and physical activity between working schedules (P-value \<0.01). The mean age in the study population by working schedule were 41.4±10.6, 38.8 ±12.2, and 39.3±10.3 years for day work, night work, and day-night regular shift work, respectively. Approximately 2.26% of the study population (607 people) engaged in night work, and 3.22% (867 people) in day-night regular shift work. The weighted prevalence of night-shift work in the combined KNHANES IV–VI was 5.19%, with 4.53% in the KNHANES IV, 5.60% in the KNHANES V, and 5.37% in the KNHANES VI. In each survey period, the prevalence of night-shift work was higher among men than among women. Detailed proportions are provided in. The mean and standard deviation of AUDIT and EQ-5D scores among different working groups are presented in. The total population and women who engaged in night work showed significantly higher AUDIT scores (9.930 ± 0.444 and 8.299 ± 0.771) in the post hoc test, than those who engaged in day work and day-night regular shift work, but this association was not observed in men. For the EQ-5D index, day-night regular shift workers had a higher score than those in other working groups. Among women, night workers had significantly lower mean EQ-5D index than day-night regular shift workers (p = 0.0138 by Tukey’s post hoc test). The mean values of EQ-5D index in the three working groups were not significantly different among men (p = 0.1684). Furthermore, EQ-5D index were always lower for women than for men for each working group, with a statistical significance: the p-value for t-test of ED-5D index means by sex were \<0.0001 in every working group. Due to the mean differences in AUDIT score and EQ-5D index among different working schedules according to gender, we evaluated if there was an interaction between gender and working status on AUDIT score and EQ-5D index. The p-values of interaction were 0.0065 for AUDIT and 0.3943 for EQ5D; thus, we conducted both pooled and stratified analyses according to gender. When an AUDIT cutoff value of ≥8 was applied to define AUDs, the weighted prevalence of AUD was 48.8% (63.3% in males and 24.8% in females) with 48.4% in day workers (63.5% in males and 24.3% in females), 59.4% in night workers (65.3% in males and 47.0% in females), and 50.2% in day-night regular shift workers (58.5% in males and 24.5% in females). The prevalence of those who experienced “some or extreme problem” in each dimension of the EQ-5D was 4.9% for mobility, 0.9% for self-care, 3.0% for usual activities, 16.4% for pain/discomfort, and 8.5% for anxiety/depression. The association between night-shift work and AUDs as well as HRQL is shown in. In women, the risks of AUDs were 2.23 times higher among night workers than among day workers. However, in men, no significant association was observed between night-shift work and AUDs. When categorizing AUDIT score into four zones, the risk of AUDs among women who engaged in night work substantially increased for zone II (OR: 1.88, 95% CI: 1.17–3.01), zone III (OR: 2.64, 95% CI: 1.22–5.73), and zone IV (OR: 4.58, 95% CI: 2.24–9.37) compared with zone I. However, no association was observed among men. The p-value for trend was \<0.001 among female night workers compared with day workers. With regard to EQ-5D, only anxiety/depression dimension risk was observed to be significantly increased among night workers (OR: 1.37, 95% CI 1.00–1.89) and decreased among day-night regular shift workers (OR: 0.60, 95% CI: 0.40–0.91). A gender-stratified analysis showed that women engaged in night work had a higher anxiety/depression dimension risk than female day workers, but the significance disappeared when the OR was 1.45 (95% CI: 0.93–2.26). Otherwise, female day-night regular shift workers had a significantly lower anxiety/depression dimension risk than female day workers (OR: 0.46, 95% CI: 0.25–0.85). No significance was observed in the five dimensions of EQ-5D according to work type among men. # Discussion The findings of this study showed a higher risk of AUDs and lower HRQL among night workers compared with day workers, especially in women. The risk of AUDs was 2.23 times higher among female night workers than among female day workers. In addition, an increased trend was observed in women engaged in night work with more severe level of AUDs. Female night-shift workers had a lower HRQL, measured by the EQ-5D index, than female day-night regular shift workers. However, these trends were not observed in men. Furthermore, women had significantly lower ED-5D index than men in every working group. These results suggest that women are more often affected by the effects of alcohol consumption and health-related problems even in the same night working status compared with men. To the best of our knowledge, this study is the first study to investigate the different associations between night-shift work and mental health in terms of AUD and HRQL according to gender. The mechanism of the association between night work and AUDs could be explained by the intermediate effect of sleep disorder. Shift work can cause sleep disorders, circadian rhythm disruptions, and disturbed sleep–wake pattern. Insomnia patients are at a higher risk of AUDs and have higher alcohol consumption in order to improve their sleep. Regarding sleep quality, men do not often suffer disrupted sleep likely because of less commitment to family responsibilities such as child care, but women experience more interrupted sleep due to family care both at night and beginning of the day, irrespective of their working schedule. A gender gap in sleep quality caused by gender differences in working conditions and family responsibilities, such as household tasks and caregiving, has also been reported. However, most of the previous studies showed similar or lower levels of alcohol consumption among night workers than among day workers and it might be caused by that gender stratification was not performed. The individuals investigated in these studies were mostly employed as nurses, and involved a higher proportion of women; thus, stratified analysis could not be conducted. One study, in which 70% of the study population were men, did not conduct stratified analysis and also showed no association. In our nationally representative study, we demonstrated the gender difference in the association between night work and AUDs for the first time. The heterogeneities of AUDs according to gender in the general population were denoted by many authors. Genevieve M et al. reported the four main components that influenced women’s drinking behavior: family drinking patterns, ethnic drinking patterns, gendered drinking norms, and occupational subculture. Furthermore, some studies suggested that the effect of shift work differ according to gender due to the need to balance work and family roles, and women are more likely affected. Moreover, studies have reported that women experience more stress and gender discrimination but less reward than males in the same working conditions which may cause health-related consequences such as unhealthy use of alcohol. Other studies revealed that the increased vulnerability of women to the adverse effects of shift work would be caused by women’s physiological pattern, including more complex circadian and hormonal rhythms. With regard to HRQL, night workers are more likely to adopt to unhealthy lifestyles, which can lead to various adverse health-related outcomes including mental health problems, and decreasing the HRQL of night workers. The higher EQ-5D index and lower anxiety/depression dimension index in day-night regular shift workers could be due to their higher education level. In addition, considering that the study population had a job during the last 1 year at the time of survey, other dimensions such as mobility, self-care, usual activities, or pain/discomfort might not decrease, but the anxiety/depression dimension was the most vulnerable among these dimensions. The gender differences in the association between night work and mental health, especially the higher risk trend among women, could be highlighted in this study despite the inconsistent results from several studies. A cross-sectional study among Korean electronic workers denoted a statistically significant association between depression and shift workers in female but not in male. However, in a large prospective cohort study, higher risks for developing depressed mood and depressive disorder in night-shift workers compared with day workers were not observed among either women or men. In general population, similarly to our study’s findings, the effects of gender on the EQ-5D index have been reported that women had lower HRQL than men. Some studies revealed potential biases due to gender differences when using the quality of life instruments. Men and women’s response to the measurement of the HRQL or health utility were different even if they experienced a similar level of discomfort. In addition, some conditions that only women experience such as abnormal uterine bleeding could affect the differences or responses to the five dimensions in the EQ-5D and its related biases in various ways. Women who engaged in night work experienced problems in their reproductive health that men never experienced. Furthermore, the EQ-5D instrument includes questions about the participants’ experiences during the interview; thus, it may not reflect the HRQL that the participants experienced as a consequence of the night works during the last 1 year. In addition, the insignificant results in the stratification analysis despite of significances in the pooled analysis would be caused by the limited number of participants in each stratum. Hence, we included participants aged 20–60 years, who were considered to be the main workforce of the population. The weighted prevalence of nightshift work was 5.19% during the 2007–2015 period, and men have a higher prevalence (7.37%) than women (2.95%). A survey conducted by the Korean Ministry of Employment and Labor (KMEL) reported that 11% of employees perform night work. This discrepancy may be due to the fact that our interview questions were regarding the last 1-year night-shift exposure, whereas the questions in the KMEL survey asked about the participants’ lifetime exposure to night work. However, the exposure to the night work in Korean population was lower than that in Western countries. The prevalence of AUDs was 48.8% with 48.4% in day workers, 59.4% in night workers, and 50.2% in day-night regular shift workers. Although direct comparisons were limited due to different measurement tools utilized, the prevalence of AUDs in Korean workers was higher than that of workers in other countries. However, the prevalence of the present study was comparable to other studies conducted in Korea. Considering the higher prevalence of alcohol consumption in Korea, it is likely that there could be a higher prevalence of AUDs in Korean workers compared with those of other countries. This study had several strengths. We included a relatively large sample size, and the study population is representative of the general Korean population. Thus, the generalizability of the findings is guaranteed. However, this study also had some limitations. First, no causal relationship was established due to the cross-sectional design of the current study. In addition, because we did not have information about alcohol use and quality of life before the engagement in night work, reverse-causation such that females with AUDs or poor HRQL would be more engaged in night-shift jobs could be possible. Second, we could not assess the impact of frequency and intensity of night-shift work because the data are limited. Third, because we considered the work schedule during the last 1 year, the long-term effect of work schedule on AUDs or HRQL could not be assessed. Fourth, despite adjusted analysis was conducted to identify an independent association between night-shift work, AUDs, and HRQL, some possible variables affecting alcohol drinking patterns such as religion could not be considered because of lack of information. Fifth, despite the efforts to minimize the effect of possible covariates on outcomes such as adjustment and weighted analysis, the differences of selective factors between working schedules may have affected the results. Sixth, although AUDIT has been a widely used screening tool with satisfactory validity and reliability, it cannot replace the gold standard of clinical diagnosis. Thus, the findings of this study might not be concordant with those of clinically diagnosed AUDs. In conclusion, the risk of AUDs was higher in women engaged in night work. The HRQL was lower in women engaged in night work than those engaged in day-night regular shift work. However, these associations were not observed in men. In addition, there was a gender difference in the risk of AUDs and HRQL in the same working status. Further studies should be conducted using a large prospective design and considering the frequency and intensity of night-shift work and risk of AUD and lower HRQL, to provide conclusive evidence for illumining these associations, in consideration with the different associations based on gender and the increased vulnerability in the female population. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Connectivity can be defined as the ability of a landscape to support ecological and evolutionary processes such as animal movements, gene flow, and range shifts. The degree to which a landscape maintains connectivity affects population sizes and genetic structure across a broad range of taxa. Studies have found that as individuals lose the ability to move across their former range population size decreases, genetic diversity decreases and genetic differentiation increases. Complex interactions between demographic characteristics, such as density, and landscape characteristics, such as habitat amount and configuration, can alter rates of dispersal and gene flow between populations. The ability of low density populations to maintain connectivity has not been explicitly explored. Hypothetically, connectivity may be reduced if population density is low and individuals are unable to find mates. This could produce small populations that may experience genetic drift and lose genetic diversity, which is a fundamental component of biodiversity and provides raw material for adaptation and evolution. Our focal species, the Mojave desert tortoise (*Gopherus agassizii*) is federally listed as threatened because of declines in population density and habitat. Range-wide population density monitoring has found that tortoise densities vary across the landscape, are related to habitat, and have been negatively impacted by habitat loss and degradation. Since their listing, desert tortoises have continued to decline. Estimates of an approximately 37% loss of adults range-wide from 2004 to 2014 have been reported. Ongoing recovery actions related to increasing density include installation of tortoise barrier fencing, protecting and connecting habitat within and among conservation areas, and targeted population augmentation. However, the effects of population density on Mojave desert tortoise connectivity have not been considered. Anthropogenic linear features, like roadways and railways modify behavior and restrict movement, with emergent genetic consequences in multiple species. Roads have been found to alter and reduce tortoise movements, resulting in a genetic signal of fragmentation. Roads also cause direct mortality when unfenced and reduce relative abundance within several kilometers of roadways. Reduced abundance of tortoise sign is correlated with habitat degradation along roadsides, with impacts extending farther from roads themselves as traffic load increases. When a natural landscape becomes fragmented, road crossing structures and other corridors may maintain or even restore connectivity. Functional corridors can enhance population viability by allowing individuals to move safely across the landscape. Gopher tortoises (*Gopherus polyphemus*) will cross railroad tracks using trenches dug underneath the rails, improving demographic connectivity. In the Mojave Desert, hydrological culverts along linear barriers are placed in washes, which tortoises often use as movement paths. In semi-natural experiments, Mojave desert tortoises walked hardwire mesh fences and entered culverts under a large highway (Interstate 15). However, many culverts are currently inaccessible due to roadside fencing. Therefore, evaluation of the effectiveness of underground crossing structures, such as culverts, in maintaining population connectivity is needed to guide management decisions concerning culvert accessibility. There are multiple ways to evaluate connectivity, including demographic (the impact of dispersal on population growth rates) and genetic (evolutionary consequences of dispersal or gene flow). Dispersal movements are poorly understood in the desert tortoise. Tortoises are believed to reside in habitat areas with average home-range estimates of 1 km² and have limited dispersal ability. In species with limited dispersal, genetic tools can provide a framework to examine hard-to-observe processes like movement across a landscape. Incorporating genetics with demographic metrics, such as population size, can improve our estimates of connectivity across a landscape. Understanding landscape spatial structure is central in evaluating connectivity because landscape amount and configuration influence population size and genetic structure. The use of spatially explicit simulations can enhance our understanding of the role landscape plays in shaping population size and genetic structure. Spatial agent-based models add realism by explicitly incorporating landscape, allowing for the study of land use changes on the structure of populations. The cost of movement, or ability of an individual to move through its environment, can be represented using a landscape resistance surface. Resistance values (i.e. high resistance may be assigned to urban areas or major roads) for each pixel cell in a gridded raster allow movement to be modeled as a function of features on a map. The lag time associated with detection of genetic patterns following changes on the landscape can make empirical evaluations of gene flow in relation to contemporary disturbance challenging. This is exacerbated by long generation times. Limited dispersal ability has also been found to increase the lag time for detecting genetic differentiation (*F*_*ST*) associated with landscape barriers (up to 200 generations). Because desert tortoises have limited dispersal ability and long generation times, it would likely require decades, or longer, to empirically determine the genetic effects of fragmentation from recent linear barriers, the potential connective value of underground culverts, and the consequences of population density. However, forward-in-time simulations can aid in evaluating how changing landscape features and declining population densities are predicted to influence future populations. Our work asks a novel question in the field of landscape genetics: How do culverts and population densities affect connectivity and potentially drive long-term genetic and demographic patterns? This study used agent-based models to simulate gene flow forward-in-time to better understand tortoise connectivity and population size in four landscapes (three hypothetical and one real): without a linear barrier, with an absolute barrier to movement, with the addition of connectivity culverts along the linear barrier, and in a heterogeneous landscape with multiple complex features. Simulations using hypothetical landscapes were run at variable initial population densities. We tested the following hypotheses: 1) the addition of culverts across roads and railways will maintain or improve gene flow for desert tortoises and 2) reduced population density in the surrounding landscape will negatively influence connectivity. We applied the simulation framework to a real world study landscape that may represent an important population linkage for wild desert tortoises. However, rapid and recent anthropogenic disturbance in the area may reduce population density and disrupt natural patterns of gene flow. Modeling the impacts of crossing structures and population density could provide managers with valuable information to weigh the benefits of investing in improving local population densities and culvert access for desert tortoises. # Methods ## Simulation model We performed agent-based spatially explicit genetic simulations of gene flow for 200 generations (approximately 3400 years), using the program <span class="smallcaps">SimAdapt</span> v.1.8.0, with 30 replicated runs, sampling individuals every five generations (roughly every 85 years). <span class="smallcaps">SimAdapt</span> uses the <span class="smallcaps">NetLogo</span> environment to model mating and dispersal in non-overlapping generations within a georeferenced area with closed boundaries. The simulation modeling platform allows density to vary across the landscape and tracks individuals through time. The program simulates landscape genetic processes with user defined parameters, including initial genetic structure, and records the alleles of all individuals from forward-in-time generations. This allowed us to explore variations in population density and genetic signals to test the influence of barriers and systematically evaluate genetic and demographic connectivity on variable landscapes. Below and in we describe the specific model inputs and parameters. ## Barrier resistance landscapes We characterized landscapes using resistance surfaces constructed and mapped in R 3.5.3 using packages *ggmap* v.3.0.0.901, *raster* v.2.9–5, and *rgeos* v.0.4–3. In preliminary work, we found that the simulated landscape size was sufficient to allow isolation-by-distance (IBD) to form. At larger spatial extents the increased number of individuals resulted in intermittent failure due to limitations with computing genotypes. Resistance values represent the cost of movement across a gridded landscape where ease of movement ranges from no resistance (0) to no movement (1). We created landscapes constructed as a uniform surface with no resistance, a surface with an absolute barrier to movement, and a surface with three culverts embedded within the linear barrier. These resistance surfaces allowed for evaluation of differences in genetic diversity, genetic structure, and simulated population size in simplified scenarios allowing for IBD. We tested specific landscape genetic hypotheses of connectivity across linear barriers and the potential influence of population density. Barriers did not allow movement unless a culvert was present. We defined the culvert resistance value using the inverse of a desert tortoise habitat suitability model. The inverse of habitat suitability has been assumed to be a valid proxy for resistance and may be appropriate for species that require occupancy within their movement areas. Simulated tortoises were present across the range of potential habitat values (0.1 to 1), but with a low probability of presence between 0.1 and 0.3. Because tortoise use of culverts has been documented, but culverts do not represent habitat, we assigned culverts a resistance value of 0.7 (habitat suitability value of 0.3), to allow for a low probability of use (three animals could occupy or move through that cell each generation). ## Ivanpah Valley resistance landscape We created a landscape resistance surface of the Ivanpah Valley, situated along the Nevada/California border and considered central to range-wide connectivity for desert tortoises. Cell values were determined using the inverse of desert tortoise habitat suitability. Because habitat becomes less suitable with anthropogenic disturbance, we reduced the habitat suitability value of cells containing disturbance on our landscape resistance surface. Habitat suitability values were adjusted using conversion factors as numerical estimates of possible degradation. We applied conversion factors to cells with urban and solar disturbance, a railway and interstate, minor roads (secondary and dirt), and utility corridor right-of- ways. We applied a relatively low maximum conversion factor to minor roads based on road density in a cell, so that cells with more dirt roads were associated with penalties at or below the maximum value. Following conversion, we calculated resistance by taking the inverse of the resulting values. Conversion factors were adapted for desert tortoises and calculated with the equation used by Inman et al. as: $$habitat\ suitability–(habitat\ suitability \times conversion\ factor)$$ ## Population density assignments Each cell was assigned a maximum density based on the resistance value. Because empirical density estimates for adult desert tortoises are highly variable (0.2 to 28/km²), we used estimates from 1 km² study plots from ongoing research in Ivanpah Valley to define high density. We represented moderate and low densities based on the relationship between habitat suitability and population density. Individuals were assigned to geographic populations based on their location relative to the linear barrier (i.e. on either side of the road). Habitat in the hypothetical landscape scenarios (no barrier, absolute barrier, and barrier with culverts) was assigned no resistance to movement and modeled at three densities: high, moderate, and low. For the Ivanpah Valley simulation, the carrying capacity ranged from 0 (cells with maximum resistance) to 24 tortoises (cells with no resistance). ## Simulation parameters Because putatively neutral genetic markers like microsatellites are not influenced by selective forces, they are ideal for investigations of gene flow. We simulated 20 variable microsatellite loci from an empirical dataset using a subset of data from a previously identified genetic cluster in the Ivanpah Valley. Detailed information on genotypes selected is available in. We created genotypes for simulation scenarios by randomizing the original samples to remove any potential signal of IBD. We then generated a large genotype file (*N* = 14572) using our simulation framework on a uniform landscape surface with a burn-in of 100 generations. We compared the simulated data (*n* = 750) with the original data (*n* = 170). All loci conformed to Hardy-Weinberg equilibrium following Bonferroni correction (*p* \< 0.003). In our simulations, time was measured in generations as discrete steps. Non- overlapping generations is a simplification that allows patterns to be more clearly detected; however, it may underestimate time to detection. In the simulations all adults died at the end of each time step. At the start of the next time step all surviving offspring became breeding adults. We defined a generation as the average length of time to reach reproductive maturity, estimated at 17 years. Individuals dispersed, reproduced, then died in that order. Individuals were selected at random for dispersal movements, reproduction, and death. As there are no empirical data on an individual’s dispersal likelihood, we used telemetry data from Ivanpah Valley to calculate the percent of animals that left 1 km² study plots without returning to approximate the probability of dispersal (Hromada pers comm 2019). Simulated dispersal was not directional. To allow movement in any direction, including diagonally, individuals were allowed to move to one of eight neighboring cells (Queen’s case), then up to 10 cells (≤ 14 km) per generation. Individuals traveled the maximum distance allowed, provided movement was possible on the landscape. The destination cell was chosen randomly among possible cells. There is evidence that tortoise movements are not random. Tortoises also exhibit social behaviors related to burrow use that may influence mating behaviors; however, how tortoises find mates is not known. In our simulations reproduction depended on finding mates within the same cell. Should no mate be available, individuals could not reproduce. The number of surviving offspring was controlled by a logistic growth function using the number of individuals in a given cell, carrying capacity, and an intrinsic growth rate. For the percent population growth per generation we used a growth rate based on an annual growth estimate of 1% multiplied by 48 breeding years, based on average lifespan minus average age of reproductive maturity (17 years). Transmission of alleles followed Mendelian laws of inheritance. No loci were under selection, and we used a mutation rate of 0.0005 per locus per generation. This value falls within the range of mutation rates for microsatellite loci estimated by Edwards et al. for desert tortoises. A detailed description of simulation methods is available in. ## Analysis of simulation output We examined simulated data through time using outcomes averaged across replicate simulations. We randomly sampled large datasets without replacement to create subsamples for analyses. To ensure our sample size was sufficiently large to capture genetic effects, we plotted pairwise *F*_*ST* values against sample sizes of 100–1200 that were taken from the moderate density simulation with no barrier at generation 200. We selected a sample size (*n* = 750) that fell after estimates of pairwise *F*_*ST* converged. Individuals were grouped based on geographic location relative to the linear barrier. We compared simulated population dynamics by evaluating the total number of individuals through time by landscape and population density scenario. Ivanpah Valley simulation results were compared with original genetic data (current Ivanpah Valley population genetic diversity and structure adapted from Dutcher et al.). To monitor genetic consequences, we calculated effective population size (*Ne*) at the landscape level and for each subpopulation on either side of a linear barrier. Effective population size estimates the size of an idealized population experiencing the same loss of diversity through drift as the population under study. In our models with non-overlapping generations, *Ne* estimated from a single generation or cohort approximates the number of breeders that contributed to the cohort in which it was measured. Effective population sizes were estimated using the linkage disequilibrium method under a random mating model implemented in *NeEstimator* v.2.1. We used 0.05 for the lowest allele frequency with 95% parametric confidence intervals. We also measured genetic diversity as allelic richness (*Ar*), and predicted observed and expected heterozygosity (*H*_*o* and *H*_*e*) at the landscape level and for each subpopulation on either side of a linear barrier. Genetic diversity statistics were calculated with the R package *adegenet* v.2.1.1. We investigated pairwise *F*_*ST* between groups in the R package *hierfstat* v.0.04–22 in time-series, using outcomes averaged across replicate runs, reporting the standard deviation. We selected genotype files best representing the mean *F*_*ST* for downstream analyses. We examined population genetic structure at generation 200 with a Bayesian clustering analysis (<span class="smallcaps">Structure</span> v.2.3.4). We used the admixture model to allow for mixed ancestry, correlated allele frequencies, and location as a prior to improve inference when genetic structure is weak. We estimated the probability of *K* population clusters (ranging from 1 to 10) using 10 replicate runs of 1,000,000 Markov Chain Monte-Carlo iterations following a burn-in of 500,000 iterates. Results were visualized using *PopHelper* v.2.2.9 in R. In the presence of spatial autocorrelation, <span class="smallcaps">Structure</span> may misrepresent genetic clusters. Because the mean log probability of the data (*Pr(X\|K)*) may overestimate clusters with IBD, we report the number of clusters based on the second order rate of change (Δ*K*), except where *Pr(X\|K)* = 1. Additionally, we used spatial principal components analysis (sPCA in R package *adegenet* v.2.1.1, to evaluate cryptic genetic patterns in the presence of IBD. We analyzed generations five and 40 (approximately 85 and 680 years; respectively) because they represent standardized time periods for short and long-term persistence. We also analyzed generation 200, the final generation simulated. sPCA is a multivariate method that maximizes genetic diversity (variance) in individual allele frequencies while accounting for spatial structure (spatial autocorrelation measured by Moran’s *I*). We compared the genetic patterns to 999 randomized Monte-Carlo permutations to test for differences between observed structure and the distribution of random expectations. Mantel tests were performed on sPCA data to determine the significance of the correlation between genetic and geographic distances. Because a Mantel test uses whole data and Moran’s I uses a portion of the data, the former may indicate stronger patterns while the latter may find more cryptic structure; therefore, we report *p*-values for both. # Results Each simulation scenario was seeded based on population density. There were no significant differences in *H*_*o* and *H*_*e* between the original dataset used to create individuals for simulation scenarios and the simulated individuals (*p* \> 0.05, *df* = 19;). ## Barrier and density simulations By the final simulated generation, the total number of individuals was highest with no resistance and lowest with an absolute barrier in all simulations. The absolute barrier represented a major road and decreased movement as well as occupancy. Reduced tortoise abundance has been documented well beyond the footprint of a major road and may reflect decreased habitat quality. Roads can also increase habitat loss and degradation by facilitating the spread of invasive plants, increasing soil erosion, and changing hydrologic patterns. The total number of individuals in moderate and high density simulations remained stable through time with the no barrier landscape. However, in the absolute barrier landscape, and with the addition of three culverts along the 25 km barrier, the average number of individuals decreased by 8%. The population census size decreased by 75% and *Ne* decreased by 98% to 99.9% in low density simulations regardless of landscape because of lack of mates within the dispersal cell. To further characterize the effect of a barrier we compared *Ne* in the barrier and culvert simulations with the no barrier simulation. In the low density scenarios, we predicted relative rates of decline in *Ne* of 80% (barrier) and 9% (culvert). Moderate and high density simulations had the most dramatic decreases in *Ne* in the barrier simulations (72% to 82%), with some improvement when culverts were present (61% to 66%), although losses were still high. Relative rates of *Ne* decline in the moderate density scenarios were 55% (barrier) and 37% (culvert). In the high density scenarios, rates were estimated at 93% (barrier) and 86% (culverts). The reported estimates of relative declines in *Ne* may be steep due to wide confidence intervals at moderate and high densities. Initially genetic diversity (*Ar*, *H*_*o* and *H*_*e*) was comparable among low, moderate, and high density simulations. At moderate population densities, heterozygosity did not differ over time with the no barrier resistance surface (*p* \> 0.05, *df* = 19) and *Ar* had a *p*-value = 0.05. Genetic diversity remained stable with moderate population density between the no barrier and culvert scenarios (*Ar* and *H*_*o* *p* \> 0.05, *df* = 19). Between the no barrier and barrier scenarios heterozygosity remained stable (*p* \> 0.05, *df* = 19), but *Ar* decreased (*p* = 5 × 10⁻⁶, *df* = 19). The high population density simulations saw a significant increase in *Ar* (*p* = 7 × 10⁻⁴, *df* = 19). We found a significant decline in *Ar* between the no barrier and barrier simulations (*p* = 6 × 10⁻⁴, *df* = 19), but there were no differences between the no barrier and culvert scenarios (*p* \> 0.05, *df* = 19). Heterozygosity also remained stable at high densities in the no barrier, absolute barrier, and barrier with culvert scenarios (*p* \> 0.05, *df* = 19). However, at low population densities, genetic diversity dropped with the no barrier resistance surface (*Ar p* = 4 × 10⁻¹¹ and *H*_*o* *p* = 3 × 10⁻¹⁵, *df* = 19). There were no significant differences in heterozygosity between the barrier scenarios (*p* \> 0.05, *df* = 19), but *Ar* did decrease significantly (*p* = 4 × 10⁻⁸). When evaluated by subpopulation *Ar* showed a decreasing trend in the culvert and barrier simulations relative to the no barrier simulation at all densities. Initially, *F*_*ST* was low (\< 0.002) in all simulations, indicating potentially high levels of genetic connectivity. No barrier landscapes displayed stability through time regardless of population density scenario. Our simulation results predicted increases in *F*_*ST* over time in absolute barrier simulations at all population densities. However, at low density, genetic differences were predicted to increase by an order of magnitude from no barrier to absolute barrier simulations. Low density simulations also displayed higher *F*_*ST* values in all landscape scenarios compared with moderate and high densities. Genetic differentiation for the barrier with culverts fell between the no barrier and absolute barrier landscapes at all simulated densities. Population genetic structure was initially absent in all simulated landscapes at all population densities. Over time the no resistance landscape revealed a pattern of IBD at all densities as the result of individual movement abilities. This conformed to expectations in natural populations. However, geographic population structure began emerging in simulations of low density (low density *K* = 3; moderate *K* = 1; high *K* = 1). We found that the absolute barrier created isolated populations regardless of simulated density (low, moderate, and high density *K* = 2). While the addition of connectivity culverts along the barrier did not alter the number of genetic clusters, admixture increased (low, moderate, and high density *K* = 2), indicating gene flow. We were able to detect population genetic structure earlier in time in low density simulations likely because genetic drift increased sensitivity. Spatial genetic patterns of isolation were strongest in low population density simulations, and weakest with high density. By generation five a genetic signal of fragmentation was developing in low population density simulations with a linear barrier to dispersal (Mantel test *p* = 0.59 and *r* \< -0.01, Moran’s I *p* \< 0.01). Culverts also produced a signal of fragmentation at low density by the fifth generation; however, it was weaker than the absolute barrier simulation and allowed for some genetic connectivity (Mantel test *p* = 0.66 and *r* = -0.01, Moran’s I *p* = 0.03). By generation 40 isolation was emergent on the no barrier surface in the low density simulation (Mantel test *p* = 0.08 and *r* = 0.03, Moran’s I *p* \< 0.01). The absolute barrier resulted in a signal of genetic fragmentation in moderate (Mantel test *p* 0.22 and *r* = 0.01, Moran’s I *p* \< 0.01) and high density simulations (Mantel test and Moran’s I *p* \< 0.01, Mantel test *r* = 0.04). Culverts also produced a signal of fragmentation by generation 40 at moderate density (Mantel test *p* = 0.33 and *r* = 0.01, Moran’s I *p* \< 0.01). By generation 200 isolation developed at moderate density (Mantel test *p* = 0.55 and *r* \< 0.01, Moran’s I *p* \< 0.01), but was not apparent at high density (Mantel test *p* = 0.50 and *r* \< 0.01, Moran’s I *p* = 0.06) on the no barrier surface. At high densities fragmentation was found with culverts by generation 200 (Mantel test *p* = 0.31 and *r* = 0.01, Moran’s I *p* \< 0.01). ## Ivanpah Valley simulation In the Ivanpah Valley, anthropogenic disturbance decreased the number of tortoises by 12% in the first five generations and by 15% by generation 40. Population census size remained stable thereafter. At the landscape level, *Ne* decreased 69% by generation five relative to the initial simulated subsample, 82% by generation 40, and 96% by generation 200. Subpopulation *Ne* followed similar patterns, decreasing 38–61% by generation 40 and 65–89% by generation 200 relative to generation five. Despite heterogeneous population density in the Ivanpah Valley simulation, *Ar* most closely aligned with our moderate density absolute barrier simulation (19.1 and 19.5; respectively). It is likely that the current disturbance landscape surface fragmented populations similarly to the absolute barrier simulation. Observed heterozygosity decreased through time by 4% at generation five and 5% at generation 40. Following 40 generations *H*_*o* stabilized in the Ivanpah Valley with existing disturbance. Empirical estimates of genetic diversity in the Ivanpah Valley are higher than predicted, likely due to the relatively recent timescale of development and the lag time to detect changes in genetic patterns. However, within five tortoise generations, we predict a loss of genetic diversity relative to current estimates as a result of recent disturbance. Predicted pairwise *F*_*ST* was measured on either side of a linear barrier (Interstate 15) to infer gene flow. Initially, pairwise *F*_*ST* in the Ivanpah Valley simulation was significantly lower than the current expectation (*p* = 0.02). By generation 40, pairwise *F*_*ST* doubled relative to the empirical estimate. By the final generation simulated, predicted *F*_*ST* was almost an order of magnitude higher than the empirical estimate, similar to that predicted at moderate density with an absolute barrier. Bayesian clustering analysis found an increased number of genetic clusters within the simulated area over time (*K* = 5), compared to the current single genetic cluster within Ivanpah Valley, indicating that populations became progressively isolated. We can currently detect weak genetic patterns on either side of the interstate in the Ivanpah Valley (Mantel test *p* = 0.15 and *r* = 0.04, Moran’s I *p* \< 0.01). Our sPCA results from the Ivanpah Valley simulation indicated similar patterns of weak isolation from disturbance by generation five (Mantel test *p* = 0.84 and *r* = -0.02, Moran’s I *p* \< 0.01) that became stronger with time (generation 40 Mantel test and Moran’s I *p* \< 0.01, Mantel test *r* = 0.05). Linear regression of the results supports spatial population structure in all scenarios, with the slope becoming steeper and spatial isolation visibly apparent in the scatterplots as time increased, indicating greater isolation. # Discussion ## Culverts maintain or improve gene flow Understanding how barriers impact connectivity is crucial to conservation efforts. In our Ivanpah Valley simulation, we found evidence of a reduction in gene flow on either side of the interstate by generation five (roughly 85 years). The loss of gene flow became more pronounced through time, resulting in a strong signal of isolation by generation 40 (roughly 680 years). This corresponds with empirical evidence from the Ivanpah Valley, where reduced gene flow has been associated with Interstate 15 and the Southern Pacific Railroad. The results of our hypothetical landscape simulations provide additional support that barriers restrict gene flow. The genetic signature of fragmentation was strongest with an absolute barrier and increased through time. The addition of culverts restored some genetic connectivity, but did not entirely negate the genetic effects of a linear disturbance. These findings agree with simulation studies from a broad range of taxa where the addition of connectivity routes in fragmented landscapes is predicted to improve gene flow. Tortoises have been documented using culverts and underpasses at variable rates, from no crossings at some locations up to 10 successful individual crossing events in a two-year period. The number of possible crossing structures in our simulated culvert scenario underrepresented the number of existing potential crossing structures in the Ivanpah Valley. Every generation (roughly 17 years) our culvert scenarios allowed up to three individuals to use a crossing structure (a rate of roughly one crossing every 5.7 years per connectivity culvert). With three possible crossing locations in the 25 x 25 km landscape, a maximum of nine individuals could cross every generation (a rate of roughly one crossing every 1.9 years). Even with this conservative rate of culvert use, our results indicated that culverts under anthropogenic barriers, like roads and railways, are likely to improve connectivity for the desert tortoise. ## Population density influences connectivity In simulations of habitat loss and fragmentation, Hand et al. found evidence for strong non-linear critical thresholds in life history traits with relatively small losses in habitat and connectivity. In our moderate and high density simulations, population census size only decreased when a linear barrier was present due to the associated loss of habitat. Habitat loss was due to the wide footprint of the road itself (1 km²), which falls within estimated road effect zones that have found decreased tortoise abundance up to 4 km on either side of unfenced highways. Habitat loss and degradation from current levels of disturbance in the Ivanpah Valley simulation produced fewer individuals and fewer breeding opportunities. The resulting genetic drift dramatically decreased effective population size relative to census size and is likely not significantly influenced by overlapping or non-overlapping generations. We predicted losses of roughly 69% in effective population size and 12% in population census size by the fifth generation. Isolation among populations can develop within five to 10 generations in species with limited dispersal (such as desert tortoises) and with low population density. Population declines associated with numerous unquantified threats have been documented in designated tortoise conservation areas with empirical estimates at or below 3/km². Our results indicated that the greatest changes in connectivity are likely to occur within five generations and at low population density. In our low density simulations (3/km²), negative impacts occurred regardless of landscape surface, likely as the result of individuals moving without finding mates. Reduced adult reproductive rates and survivorship are both implicit drivers of tortoise population declines, although the causes of reduced reproductive output are unclear. At maximum carrying capacity in low population density simulations, not all individuals in a cell could mate, which reduced population census size. This resulted in genetic drift, dramatic reductions in effective population sizes, and isolation. It is plausible that wild tortoises may be unable to find mates if densities fall too low; therefore, population density likely plays a principal role in maintaining connectivity. The present study is an initial examination showing a clear non-linear effect and a potential threshold between medium and low population densities under which connectivity is unsustainable. ## Model advantages and limitations We implemented agent-based spatially explicit forward-in-time simulations of gene flow to track the number of individuals and their genotypes through time. Because gene flow and dispersal are population-level processes that occur over long time periods, our simulation models addressed large-scale spatial patterns over long time periods, rather than fine-scale dispersal. The impacts of landscape change on genetic patterns are associated with a considerable lag time in detection. This can hinder detection following recent habitat disturbance. Therefore, better understanding of time to detection is important to interpreting empirical population genetic data following habitat disturbance. Our simulation framework allowed us to explore possible desert tortoise connectivity outcomes and time to detection. Over the 200 generation simulation period, tortoises will certainly experience additional changes in land use, climate, and vegetation to those presented here. However, our findings have practical implications now and can guide management actions towards the long-term persistence of the species. Our simulation framework had several additional advantages for testing landscape genetic hypotheses. First, focusing on simplified landscapes allowed us to explore variations caused by a linear barrier and a barrier with culverts in the absence of any confounding influence from heterogeneous landscape structure. Second, we were able to evaluate differences arising from systematically varying population density, which impacted population census size. Third, we were able to examine current disturbance in the Ivanpah Valley. This added complexity, but also allowed for hypothesis testing related to the influence of multiple landscape features on gene flow. The sampling grain or raster pixel size should ideally be smaller than an average home-range size and dispersal distance. The mean home-range for desert tortoises is highly variable, but estimated to be roughly 1 km². Our 1 km² grid cells represent areas in which individuals would likely make routine movements, capitalize on resources, and find mates. The cell size is also well below the expected dispersal distance for this species, which has been documented to be up to 13.6 km. Using a sampling grain that is too large may smooth patterns, resulting in a decreased ability to detect the influence of landscape. We were able to detect the influence of landscape, indicating that our grid cell size was appropriate. That said, using a smaller sampling grain, especially in areas of interest such as connectivity culverts along major roads, could reveal additional information relevant to these features. A simplification to consider when interpreting our results is non-overlapping generations. The inclusion of overlapping generations in other platforms results in additional tradeoffs, such as constant population density, simulations of a single locus, or all individuals in a population assigned a single genotype (see: <span class="smallcaps">CDMetaPop</span>; <span class="smallcaps">CDPop</span>; <span class="smallcaps">MetaPopGen</span>). These simplifications would not have addressed our research questions. However, the effect of overlapping generations could be explored in future work to detect genetic signals of fragmentation in long lived organisms, like the desert tortoise. The use of overlapping generations may allow for more realistic movement parameters with more than one movement per generation. In our simulations the destination cell was chosen randomly among possible cells, with breeding occurring when individuals occupied the same cell. Using a distribution function in future simulations to determine individual movements may influence dispersal probability and the rate of genetic exchange. In our low density simulations, mating opportunities were much reduced, reflected in very low effective population size estimates. This may be mitigated to some extent in real tortoise populations by social structure, such as shared burrows and mating behavior. We assumed survivorship and breeding success to be equal for individuals. Building on this work by incorporating risk factors associated with age and habitat along with more realistic movement and reproductive behavior may provide additional information on reproduction and survival. ## Management considerations The threat of habitat loss and fragmentation is likely to intensify in the foreseeable future. A precipitous drop in the survival-habitat relationship indicates that as a population approaches this threshold, additional small losses of habitat have a large impact on the probability of survival. Although this threshold is still unknown for the desert tortoise, habitat loss from development is substantial and the number of individuals is declining. Efforts to recover disturbed Mojave Desert lands may enhance success by increasing habitat amount, but recovery can be slow. Ensuring adequate habitat is available by restricting anthropogenic disturbance to already disturbed areas, then restoring habitat where it has been lost and populations are declining may be an effective recovery strategy and has been proposed as part of a sustained management effort. Roadside fencing reduces tortoise mortalities and may allow animals to reclaim depauperate habitat. Retrofitting existing culverts and tortoise fencing to allow animals access to culverts is a feasible option to optimize safe passage and improve connectivity for tortoises. Future management practices that include optimal design and construction of crossing structures could further connectivity efforts in fragmented landscapes. A unified network of conservation areas and connectivity routes at variable spatial scales, ranging from specific culverts across linear barriers to range- wide, may approximate natural patterns and maintain connectivity. Design and enactment of such a network is challenging, but would have long-term benefits for the persistence of the tortoise and the preservation of the Mojave Desert ecosystem. Existing research can serve to guide conservation and connectivity efforts. For example, Gray et al. found strong evidence that tortoise movement is associated with increased vegetation cover. Therefore, areas with predicted high habitat quality and vegetation cover could be prioritized for protection. Management actions to increase vegetation cover in areas targeted for maintaining connectivity may also be effective. Extinction risk may be underestimated if genetic factors are not considered in conservation strategies. Changes in genetic diversity, differentiation, and structure may indicate more severe effects associated with population decline. We found estimates of population census size and effective population size yielded telling trends in five, or fewer, generations. Effective population size could be a useful metric to include in tortoise population monitoring efforts, along with continued population density monitoring. Incorporating genetic monitoring into management practices prior to disturbance, then periodically thereafter (i.e. every generation following disturbance), would allow patterns to be tracked through time. # Supporting information Discussion with Marjorie Matocq and Scott Bassett contributed to the development of this research. Scott Cambrin and Kimberley Jenkins provided supportive feedback throughout project development. We would also like to thank Anna Mitelberg for assistance with genotyping and Lee Bice for supplying disturbance data layers. We are grateful for those who contributed to the acquisition of original data including Kristina Drake, Felicia Chen, Ben Gottsacker, Amanda McDonald, Jordan Swart, and Sara Murray. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. [^1]: The authors have declared that no competing interests exist.

# Introduction Embryonic stem cells (ESC) can be captured in vitro in two distinct states of pluripotency known as “naive” and “primed”. A distinguishing feature of naive cells is their capacity to integrate into a blastocyst and to contribute efficiently to germline chimeras. Importantly, naive stem cells can be established from mouse strains of non-permissive genetic background and can be modified genetically by homologous recombination. ESC from non-rodent mammals share features of primed pluripotency, characterized by their FGF and Activin/Nodal dependence, and reduced capacity to contribute to chimeras. This has led to the suggestion that establishing naive pluripotency may be beneficial for generating highly competent stem cells amenable to genetic modification. The conversion of primed cells to naive state by switching culture conditions was first reported in the mouse following the dissociation of epiblasts into single cells. Established epiblast stem cell (EpiSC) lines, however, are much more resilient to this switch and cannot be converted to naive ESC simply by changing the culture conditions, unless additional factors are overexpressed. Similarly, human ESC (hESC) cultured under stringent conditions to force the conversion to a naive state are highly unstable and cannot be cultured extensively. These findings demonstrate that the naive state can be captured efficiently by including specific cytokines during the establishment of new cell lines from mouse blastocysts, or shortly thereafter by disrupting cell interactions. However, once a cell exits the naive state and transits to a primed state, the reversion can only be restored artificially by specific factor overexpression. In non-rodent species naive ESC have not yet been isolated from pre-implantation embryos. Stem cell lines derived from primate, rabbit, and pig, embryos share features of primed pluripotency, as described for mouse EpiSC. This seems to suggest that the naive state is either absent or very transient during embryo development in these species. A link between naive pluripotency and physiological diapause has been proposed as the underlying developmental mechanism that might explain why naive stem cells can be established from certain inbred mouse strains, but not from non-permissive mouse strains or from other mammals. Attempts to impose naive culture conditions failed in establishing stem cell lines in the pig, however when ICM cells from early blastocysts (day 5.5) were transduced with *KLF4*, cell lines with some features of naive pluripotency were established. These lines failed to form teratomas when injected into SCID mice, but when *OCT-4* was over-expressed, the cells eventually acquired naive properties as demonstrated by their LIF dependency, teratoma formation capacity, and efficient integration to the ICM of blastocysts. Thus, it seems that naive pluripotency can be imposed in cells derived from porcine embryos, however the optimal conditions for converting to this state may require species-specific considerations. For instance, NANOG protein is not detected in early pig blastocysts, suggesting that entering the naive state in these cells might be compromised due to the lack of a functional pluripotency network. Studies in mouse embryos show that modulation of MEK and Wnt signalling result in an enriched NANOG cell population in blastocysts. Interestingly, these effects were not observed in human and cattle embryos, where hypoblast cells expressing GATA4/6 were still detected,. These interventions before the segregation of the inner cell mass (ICM) from trophectoderm (TE) offer an opportunity for capturing naive cells that may naturally only be present very transiently, if at all, when the ICM arises. The aims of the present study were 1- to study whether modulation of multiple signalling pathways can alter the proportion of NANOG positive cells in the ICM of pig blastocysts, and 2- to determine whether stringent culture conditions that support naive pluripotency in the mouse can be imposed in pig pluripotent cells. # Materials and Methods ## Embryo Collection and In Vitro Culture All the procedures involving animals have been approved by the School of Biosciences Ethics Review Committee (University of Nottingham, UK). Landrace × Large white crossbred sows were artificially inseminated twice over 2 days. Pig embryos were collected at day 4 after insemination. The oviduct and uterine horns were flushed with pre-warmed phosphate-buffered saline (PBS) supplemented with 1% fetal calf serum (FCS). The embryos were placed in an ovum concentrator and rinsed with PBS containing 1% FCS and 25 mM Hepes. Recovered embryos were allocated to either PZM3 or N2B27 culture medium supplemented with 0.3% fatty acid free BSA. Embryos at morula stage were included in the study. Embryos at earlier stages were cultured in PZM3BSA until the compact morula stage and subsequently transferred to the experimental groups. Embryos were incubated in a humidified atmosphere at 39°C, under 5%CO₂ and 5%O_2. The embryos were treated with inhibitors and growth factors at the following concentrations: PD0325901 (PD, MEK inhibitor, Calbiochem) 0.4 µM or 1 µM when combined with GSK3β inhibitor CHIR (GSK3β inhibitor, Selleck) 3 µM; PD173074 and PD161570 (FGF receptor inhibitors, Tocris) 100 nM; SB431542 (Type 1 TGFβ receptor ALK5, Tocris) 20 µM; 42009 (JAKi, JAK/STAT3 Inhibitor 420099, Calbiochem) 0.6 µM; LY294002 (InSolution™ LY 294002, Merck) 10 µM, human recombinant FGF4 (Peprotech) 1 µg/mL and heparin 1 µg/mL, as described by. Heparin was included because it has been shown to bind FGF4, increasing the stability of the ligand-receptor interactions. DMSO was used to dissolve the inhibitors, and was maintained at equal concentrations among groups. Control groups were added DMSO accordingly. ## Porcine Fetal Fibroblasts Isolation, Reprogramming and Cell Culture Primary porcine fetal fibroblasts (PFF) cell lines were isolated from approximately 40 day-old fetuses. PFF were cultured in DMEM containing 10% fetal calf serum (FCS) and supplemented with 1% glutamine, 1% penicillin/streptomycin, 1% nonessential amino acids and 0.1 mM β-mercaptoethanol (culture medium: CM). Induced pluripotent stem (iPS) cells were generated from passage 3 cells. PFF were plated onto gelatinized dishes (0.1% porcine skin gelatine) at a density of 0.1 million cells per 9.6 cm². PFF were infected with virus containing-medium twice at 48 h intervals. Lentiviruses were produced by transfecting HEK293 cells with doxycycline inducible FUW-tetO vectors (Addgene), encoding the human cDNA sequence of the four transcription factors (4 factors) OCT4, KLF4, SOX2 and C-MYC plus FUW-M2rtTA (Addgene). The virus containing- medium was collected 48 h after transfection. The media containing the 4 factors and FUW-M2rtTA virus were pooled in equal volume, filtered through a 0.45 µm filter and supplemented with 4 µg/mL polybrene (Sigma-Aldrich). PFF were passaged 48 h after the second transduction and cultured in CM supplemented with 10³ units/mL mouse LIF (ESGRO, Chemicon International). Cells were supplemented with 2 ug/mL doxycycline (Dox, Sigma- Aldrich). Between 1 and 4 weeks after Dox induction the colonies were picked and plated onto mouse embryonic fibroblast feeders. After two manual passages, piPS cells were dissociated with 1 mg/mL collagenase IV and passaged every 3–4 days at a ratio of 1∶5 or 1∶6. Pig iPS cells required continuous Dox supplementation for survival. A normal karyotype was determined for the cell lines used in this study (2*n* = 38). The FCS + LIF derived piPS were transferred to serum free N2B27 medium supplemented with 0.3% BSA, 10³ units/mL mouse LIF (ESGRO, Chemicon International), 1 µM PD0325901, 3 µM CHIR and 100 nM PD173074, as described previously. ## Immunocytochemistry and Alkaline Phosphatase Activity Before fixation, expanded embryos were treated with 0.5% pronase for 1–2 min to remove the zona pellucida (ZP). Hatched embryos were washed in 1% PBS/BSA and fixed in 2.5% paraformaldehyde (PFA) for 15 min at room temperature (RT). After three washes in 0.2% Tween 1% PBS/BSA (PBST), embryos were permeabilized in 0.2% Triton X-100 PBS for 20 min at RT and blocked for 1 h at RT in 7% BSA and 5% donkey serum in PBST. Pig iPS cells were fixed in 4% PFA and subsequently permeabilized with 0.1% Triton X-100 in PBS and blocked in 5% PBS/BSA. Embryos and piPS cells were incubated overnight at 4°C with the following primary antibodies: NANOG (1∶400; rabbit polyclonal, Peprotech 500-P236), and GATA-4 (1∶200; goat polyclonal, Santa Cruz Biotechnology SC1237), OCT-4 (1∶100; goat polyclonal, Santa Cruz Biotechnology SC8628), STELLA (1∶60; rabbit polyclonal, Abcam ab19878), SSEA-1 (1∶50; Hybridoma bank, Iowa City), all in 5% PBST/BSA. The following antibodies were used in differentiated piPS cells: NESTIN (1∶100, rabbit polyclonal, Abcam), SOX17 (1∶100, goat polyclonal, R&D Systems), CYTOKERATIN-14 (1∶100, mouse monoclonal, Hybridoma bank, Iowa City), βIII- TUBULIN (1∶100, mouse monoclonal, R&D Systems), VIMENTIN (1∶100, AMF-17b, mouse monoclonal, Hybridoma bank, Iowa City). Pig iPS cells incubated with SSEA-1 antibody were not permeabilized. After 4 washes in 0.05% PBST-Triton (PBSTT) embryos and piPS cells were transferred to the appropriate secondary antibody and incubated for 1 h at RT, followed by 3× washes in PBSTT. Embryos and cells were mounted in Vectashield with DAPI (4′-6-diamidino-2-phenylindole; Vector Laboratories). Alkaline phosphatase (AP) activity was analyzed with the AP kit (Sigma-Aldrich) following the manufacturer’s instructions. ## Quantification of Total and ICM Cell Numbers in Embryos The total cell number was obtained by counting all DAPI positive nuclei. Cell counts were performed in triplicates to obtain the average ± SD per embryo. The ICM cell number was calculated as the total number of NANOG and GATA-4 positive nuclei. Fluorescent images were acquired using epifluorescence. ## RNA Isolation and Polymerase Chain Reaction RNA isolation was carried out using RNeasy kit (Qiagen) following the manufacturer’s instructions. RNA reverse transcription was performed using Omniscript synthesis kit (Qiagen). End-point PCR was performed with ReadyMix (Sigma-Aldrich) and 0.4 uM of each primer. PCR products were run in 1.5% agarose gel to determine the amplicon size. Quantitative RT-PCR (qRT-PCR) was performed using SYBR green mix (Roche) and 0.25 nM of each primer. For each gene, the analysis was performed in triplicate. RT-PCR protocol included an initial step of 95°C (5 min), followed by 45 cycles of 10 sec at 95°C, 15 sec at 60°C and a primer extension step of 20 sec at 72°C. Fluorescence data were acquired at 72°C. Melting-curve analysis to confirm product specificity was performed immediately after amplification and the amplicon size was checked by electrophoresis. The relative expression of the target gene was normalized with *GAPDH* and a calibrator sample. Sequence accession numbers were obtained from NCBI, Ensembl and TGI databases. Primers used in this study are listed in. ## In Vitro Differentiation Embryoid bodies (EB) were prepared from piPS cells using the hanging drop method. Briefly, piPS cells were trypsinized before preparing hanging drops at a density of 3000 cells/EB. The differentiation medium consisted of DMEM supplemented with 1 ug/mL Dox (for the first 10 days, and was subsequently removed), 1% glutamine, 1% penicillin/streptomycin, 1% nonessential amino acids and 20% FCS. Aggregated EBs were transferred to non-adherent dish after 4 days. After 10 days, EBs were plated onto gelatine coated plates and cultured for up to 30 days. EBs were collected and frozen at −80°C until analysis. ## Statistical Analysis To evaluate the statistical differences in relative numbers, probability (P) values were established using Two-sided Mann-Whitney U test. Two-tailed Student’s *t*-test was used to compare absolute values between groups. Differences between groups were considered significant when *P*\<0.05. Pearson’s correlations coefficient was used to analyse results from qPCR. # Results ## Culture of Porcine Embryos to the Late Blastocyst Stage To study the segregation of the ICM and TE, morulae obtained at day 4 of in vivo development were cultured in PZM3 medium under low oxygen, which has previously been shown to support in vitro porcine embryo development. After 48 to 72 h, 30.2% (SD ±15.9, n = 74) of embryos hatched from the zona pellucida, and the majority of these (79% ±15.5) had a total cell number \<100 cells. Because these embryos showed very small or no ICM , it was not possible to perform meaningful analysis of lineage segregation. Next we tested N2B27, a serum-free basal medium used for growing ESC and human embryos. After 48 to 72 h in culture, 53.4% (SD ±25.8, n = 182) of embryos hatched, and the total cell number exceeded 100 cells in more than 82% of these embryos. Since a clear ICM was visible in most embryos this culture system was selected for the experiments performed in this study. ## Effect of FGF Signalling During Hypoblast Segregation in the Porcine Embryo NANOG and GATA-4 are expressed in mutually exclusive cellular domains marking the epiblast and the hypoblast, respectively. In the pig embryo these two factors are not found in early blastocysts (days 5.5–6.5), but they can be detected in the early epiblast from day 7.5. Since it is not clear whether NANOG and GATA-4 are expressed in mutually exclusive domains, in vivo retrieved morulae cultured for 48 h until they progressed to the late blastocyst stage were analysed by immunofluorescence (IF). Consistent with previous findings in mice and humans, NANOG positive cells were observed in clusters surrounded by GATA-4 positive cells ( control). Importantly, NANOG and GATA-4 cells were only detected in hatched embryos (late blastocysts) containing \>100 cells (n = 5), all others with \<100 cells (blastocysts and expanded blastocysts; n = 15) stained negative. In four expanded blastocysts we found NANOG positive cells (2, 2, 4, and 7 cells), but no GATA-4 cells (not shown), suggesting that GATA-4 cells may originate from a NANOG positive population in the ICM. These observations, combined with the finding that OCT-4 is expressed in the pig ICM, show that the pluripotency network is established during the transition from early to late blastocyst in this species. To study how this process is regulated, we evaluated this transition in embryos treated with small molecule signalling inhibitors. Experiments in mouse embryos show that MEK inhibition prevents hypoblast differentiation and enhances the number of Nanog expressing cells in the ICM, however in human and bovine embryos no specific inhibition of GATA4/6 was observed under these conditions. We found that a high proportion of pig morulae cultured with the MEK inhibitor PD0325901 progressed to the hatched blastocyst stage (75%; n = 20). These embryos had an average cell count of 197.2 (SD ±44.1, n = 14) cells and a similar proportion of TE cells as the control group, indicating that the viability of these embryos was not visibly compromised by MEK inhibition. NANOG and GATA-4 protein were detected in the ICM of treated embryos, however the proportion of GATA-4 cells was reduced compared to DMSO-only treated embryos. This result indicates that the hypoblast can segregate in the absence of MEK signalling, although the number of GATA-4 cells is reduced. In rodents, FGF initiates hypoblast differentiation via MEK signalling. Therefore, to gain further insight into GATA-4 activation, embryos were cultured with the FGF receptor inhibitor (FGFRi) PD161570. Although the proportion of embryos progressing to the hatched blastocyst stage (55% ±21.2, n = 20) and the total cell count (215.8±42.7, n = 15) was not significantly different to those treated with PD0325901, only few embryos (5/15) formed an ICM, all of which had very few cells. To confirm that this effect was not due to the specific inhibitor, we repeated experiments with another FGFRi PD173074, and similar results were obtained (data not shown). After immunostaining, we observed no differences in the proportion of NANOG and GATA-4 cells compared to the control group. This result indicates that FGF inhibition does not interfere with hypoblast segregation, but might interfere with the formation of the original founder population of the ICM. Since FGF can signal through phosphatidylinositol 3-kinase (PI3K), we evaluated whether inhibition of this pathway affected the founder cell population of the ICM. Embryos treated with PD0325901+ LY294002 hatched at a similar rate to control embryos (41.67%) and showed no differences in NANOG and GATA-4 staining (not shown). We next evaluated the effect of FGF4 supplementation in lineage segregation. Morulae incubated with FGF4 and heparin developed normally to the hatched blastocyst stage (62.5±24.7, n = 20) and had an average total cell number of 209.7 (SD ±78.2, n = 15;). The majority of these embryos formed a clear ICM, although it was significantly smaller compared to the control group. Furthermore, most embryos (9/14, 64.2%) had 100% GATA-4 positive cells, and no NANOG positive cells. This result shows that FGF4 promotes the segregation of the hypoblast without affecting the total cell number of the embryo. These experiments demonstrate that MEK inhibition has only a partial effect in preventing hypoblast segregation in the pig embryo. ## Effect of GSK3β, JAK/STAT3 and ALK5 Signalling Inhibition During ICM Segregation in the Porcine Embryo We next evaluated whether a combination of MEKi (PD0325901) and stimulation of Wnt signalling by inhibition of GSK3β could contribute to an enrichment of the NANOG population in the ICM, as described in the mouse. A moderate increase in the proportion of hatched embryos (80% ±28.3, n = 20) and in the total cell number (220.8±73.8, n = 15) was observed in these embryos. The percentage of ICM/total cells was lower than in control embryos. The proportion of NANOG positive cells showed a median value similar to embryos treated with MEKi, however, due to the high variability between embryos, the observed differences were not statistically significant compared with the control. Thus, stimulation of Wnt signal together with MEK inhibition does not have a synergistic effect on the number of NANOG positive cells present in the porcine ICM. LIF signalling is operative in the mouse ICM, as demonstrated by the requirement of this cytokine for the derivation of ESC. In the pig there is limited information on the role of this cytokine during the formation of the ICM, although recent studies show that LIF receptors are expressed in the TE of late blastocysts, but not in the ICM. We found that hatching rate (7.1% ±10.1) was significantly reduced in embryos treated with JAK/STAT3 inhibitor 420099 compared to the control group, mainly because the embryos degenerated between early to late blastocyst transition. Of the two blastocysts that developed, only one had an ICM with two NANOG positive cells. This result indicates that during the transition from morula to blastocyst JAK/STAT3 signalling plays a critical role in the development of the ICM and the TE. Activin/Nodal signalling controls NANOG expression and is essential for maintaining hESC and mEpiSC in the undifferentiated state. Furthermore, Activin A is expressed in the porcine ICM, suggesting that this pathway may participate in the activation of NANOG. To test this possibility we incubated morulae with the ALK5 receptor inhibitor SB431542, which causes downregulation of Nanog and neurectoderm differentiation in EpiSC. We found that progression to the hatched blastocyst stage (65.5% ±17.3, n = 18) and total cells numbers were not affected by the treatment. Similarly, the proportion of ICM cells and distribution of NANOG and GATA-4 cells was unaffected compared to the control group. This result demonstrates that Activin/Nodal signalling is not required for NANOG activation and for normal segregation of the hypoblast. ## MEK and GSK3β Inhibition + LIF Promotes Naive Pluripotency in Porcine iPS Cells The results above suggest that during the transition from morula to blastocyst the ICM responds to MEK inhibition by reducing the proportion of cells differentiating to hypoblast. The inhibition of this signalling pathway, together with inhibition of GSK3β (known as 2i medium) and activation of JAK/STAT3 pathways, are essential for establishing naive pluripotency in the mouse. Here we applied these culture conditions to embryos treated with 2i between the morula to blastocyst transition. Seven whole-embryos plated onto feeders in 2i + LIF medium attached after 24 h, however in the following 72 h most clumps failed to initiate proliferation and degenerated (not shown). This result suggests that the conditions for the isolation of naive cells directly from embryos are still sub-optimal to support stem cell derivations. We turned to a piPS cell system to study whether the naive state could be imposed in pluripotent cells in this species. A recent report shows that piPS cells become LIF dependent when cultured in normal ESC culture conditions. Here we extended these investigations to determine whether 2i (or 3i) + LIF could be used to force naive features in piPS cells. We generated iPS cells from porcine fetal fibroblasts using *hOCT-4, hc-MYC, hKLF4* and *hSOX2* under the control of doxycycline. Colonies started to appear after 6–7 days, and after 15–30 days strong alkaline phosphatase activity was detected in most cell clumps. Colonies were picked and transferred onto feeder cells in standard ESC culture medium (containing 10% FCS and LIF). Eight cell lines were expanded and further characterized. Cells under these conditions grew as compact colonies which consistently expressed OCT-4 and endogenous NANOG, but only few colonies expressed SSEA1 antigen. All the lines expressed the four exogenous factors, therefore we selected four lines to characterize their gene expression profile in more detail. A distinguishing feature of mouse naive stem cells is the expression of *Stella* and *Rex1*, in contrast to the expression of *Fgf5* and *Nodal*, typical of primed stem cells. An initial analysis of these lines showed that all lines expressed endogenous *OCT-4, NANOG, SOX2, KLF4, NODAL and FGF5,* however *STELLA* and *REX1* were expressed in 2/4 lines. These results show that piPS cells cultured with FCS and LIF are heterogeneous, displaying features of primed and naive pluripotency. We next tested whether switching the culture conditions of piPS cells to 2i and 3i (2i + FGFRi) + LIF promoted features of naive pluripotency, as previously demonstrated for mouse ESC. Switching piPS cell culture medium to N2B27 supplemented with 2i (or 3i) and LIF resulted in some cell death during initial 1–2 passages, however in subsequent passages the cells grew as homogeneous round colonies. To test their LIF requirements cells were cultured either with 2i only or with the JAK/STAT3 inhibitor. In both conditions extensive cell death was observed and most cells were lost within one passage. Interestingly, LIF withdrawal or JAK/STAT3 inhibition of cells grown with FCS resulted in a rapid loss of their compact morphology. These cells showed overt signs of differentiation, which was confirmed by the expression of *SOX17* and *NODAL* and the reduction in *NANOG* expression. Similar results were obtained with cells cultured with 3i (not shown). These experiments demonstrate that piPS cells depend on LIF to maintain the undifferentiated state. We next evaluated whether cells cultured in 2i/3i + LIF activated genes indicative of naive pluripotency. Two cell lines (A-09 and A-12) transferred from FCS + LIF to 2i/3i + LIF showed increased *STELLA* and *REX1* expression within 2 passages. Cells grown in 3i + LIF showed a similar phenotype to cells grown in 2i + LIF, but overall the cultures in 3i were more homogeneous and showed low levels of differentiation (not shown). Immunolocalization of STELLA demonstrated homogeneous staining in colonies of cells grown in 3i + LIF, whereas cells grown in FCS + LIF had a large proportion of cells in the periphery of the colonies that were STELLA negative. Analysis of additional 6 lines shows that 3i + LIF induced an increase in *STELLA* expression in all cell lines, however high variability was detected between different lines. Interestingly, the changes in *STELLA* expression correlated with changes in *NANOG, OCT-4* and *REX1* expression, but were not correlated with *FGF5* expression. ## Naive piPS have Increased Germline Differentiation Potential Mouse naive stem cells can differentiate to all somatic lineages and can generate germline chimeras efficiently, in contrast to primed stem cells. We investigated the capacity of three piPS cell lines maintained in FCS + LIF or in 3i + LIF media to differentiate in EBs. Expression of ectoderm (*PAX6, NESTIN*), mesoderm (*CARDIAC ACTIN*), and endoderm markers (*GATA6*) was determined by RT- PCR in the three cell lines analysed. The expression of markers of ectoderm (β-TUBULIN, NESTIN, CYTOKERATIN 14), mesoderm (VIMENTIN), and endoderm (SOX17, GATA4) was further confirmed by immunohistochemistry. We next compared the germ line differentiation capacity of cells grown in 3i + LIF vs. FCS + LIF. Three cell lines expressing *STELLA* at different levels (A-04*^low* (female line), A-09*^medium* (male line), and A-12*^high* (female line)) when cultured in 3i + LIF, were induced to differentiate for 20 days. The differentiation capacity of these cells was compared to the respective cell lines grown in FCS + LIF before inducing differentiation. *VASA* was induced after 15 days of differentiation in the three cell lines cultured in 3i + LIF, however no *VASA* expression was detected when the same lines were grown with FCS + LIF. After 20 days, *VASA* expression increased further in cells grown in 3i + LIF, while some expression was also detected in cells grown in FCS + LIF at this timepoint. To determine whether the differentiation protocol induced the exit from the naive state, *REX1* was also analysed. Cells cultured in 3i + LIF had higher levels of *REX1* before differentiation, however a sharp reduction was detected after 15 days, and it was not detected in two (A-04 and A-09) of the three cell lines after 20 days. These results show that piPS cells expressing *STELLA* can be efficiently induced to initiate the germ cell differentiation program in vitro. # Discussion This study was designed to investigate the consequences of modulating signal transduction pathways in the establishment of pluripotency during pig embryo development and in piPS cells. The first set of experiments were designed to determine which signals participate in the establishment of NANOG expressing cells from in vivo produced embryos. Since NANOG protein is first detected in pig late blastocysts (from day 7.5), the culture conditions were optimized to ensure a high proportion of embryos reaching this stage in vitro. The majority of embryos cultured in the conventional PZM3 culture medium did not progress beyond the blastocyst stage, had low total cell count, and very small or inexistent ICM. The culture of pig morulae in N2B27 resulted in a significant increase in hatching rate and total cell count, and supported cultures up to the late blastocyst stage in vitro. Furthermore, NANOG and GATA-4 cells were detected in the ICM of most hatched blastocysts. In earlier embryos (non- hatched) GATA-4 was never detected, however NANOG was found in a few isolated cells in 4/32 embryos, suggesting that NANOG expression precedes GATA-4 expression in the founder population of the porcine ICM. Since GATA-4 is a hypoblast marker, this observation also suggests that hypoblast cells (GATA-4 positive) segregate from early epiblast cells (NANOG positive). We next investigated whether interfering with MEK signalling during morula- blastocyst transition prevented hypoblast segregation. In agreement with the observations in human and bovine embryos, MEK inhibition did not completely prevent hypoblast formation in the porcine embryo. Although GATA-4 expression was still detected after MEK inhibition, the number of positive cells was significantly reduced. In contrast, a significantly higher number of NANOG positive cells were observed. Since the number of ICM cells was not affected by the treatment, the results demonstrate that the reduction in GATA-4 was due to a partial interference with hypoblast segregation. To establish whether MEK signal is triggered by a response to FGF, as previously shown in mouse embryos, pig morulae were cultured with an FGFRi. Interestingly, inhibition of FGFR resulted in a significant reduction in the number of ICM cells, however the proportion of NANOG and GATA-4 cells was unaffected. These findings are in agreement with the observations in bovine embryos, although the effect of the treatment on total cell numbers was not reported in this study. Our results show that MEK- independent FGF signalling is important for the establishment of the founder population of the ICM. Since inhibition of MEK and PI3K signalling (an alternative downstream pathway of FGF) did not affect lineage segregation, it is possible that other FGF signalling effectors, such as phospholipase C and protein kinase C, may be playing a role in establishing the ICM. Future studies will focus on the role of these pathways on the formation of the ICM. The role of FGF signalling in promoting hypoblast differentiation was further demonstrated in experiments where FGF4+ heparin were added to developing embryos. Treated embryos developed normally and maintained normal cell numbers in the ICM, however a significant increase in GATA-4 cells was observed. These results are in agreement with observations in mouse, rat, human and bovine embryos, and demonstrate that the role of FGF in promoting hypoblast segregation is conserved in mammals. We next asked whether other signalling pathways may participate in the activation or maintenance of NANOG expression in the ICM. The role of Wnt was investigated first, since maintenance of this pathway in mouse ESC promotes cell proliferation and undifferentiated state when combined with a MEKi in serum free cultures. In the pig embryo, inhibition of GSK3β in combination with a MEKi did not affect the segregation of hypoblast cells, indicating that this pathway is not critical during the establishment of the founder cells of the ICM. The role of Activin/Nodal signalling was also evaluated, based on a previous study showing the transient expression of Activin A in porcine blastocysts around the time when NANOG expression is activated. Inhibition of Activin/Nodal signalling did not affect the activation of NANOG, the number of ICM cells or the segregation of the hypoblast. This demonstrates that an Activin/Nodal- independent NANOG cell population exists during the formation of the porcine ICM, perhaps with features similar to the mouse ICM. Since previous studies showed no increase in blastocyst rates of porcine and bovine embryos cultured with LIF, here we tested the possible involvement of this signalling pathway through the inhibition of the JAK/STAT3 phosphorylation. We found that most embryos cultured in the presence of the JAK/STAT3 inhibitor failed to hatch from the zona pellucida and had significantly lower total cell count and ICM cells, indicating that blocking this signal directly interferes with the cavitation process. Indeed, a recent study described the expression of LIF receptors in the TE, but not in the ICM of porcine late blastocysts. This suggests that LIF signalling may play an important role during the segregation of the TE, rather than with the formation of the ICM. It cannot be excluded that this effect is due to the interference with IL-6 (which signals through JAK/STAT3) that is produced in an autocrine manner, and plays a role during pig early embryo development. These experiments demonstrate that the transition from morula to late blastocyst is characterized by the formation of the ICM, in which NANOG is first activated. The ICM then gradually expands before segregating GATA-4 cells. Although the activation of NANOG in the founder population appears to be independent of MEK, Wnt and Activin signalling, it is clear that FGF plays a crucial role in ensuring the expansion of the starting ICM population. These findings suggest that NANOG activation is a cell autonomous process, and that extrinsic signals (including FGF, TGFβ, Activin, and Wnt ligands) may cooperate during the transition from early to late epiblast to maintain NANOG activity and pluripotency. As defined in ESC, modulation of different signalling pathways supports two known states of embryonic pluripotency: naive and primed pluripotency. A naive state, imposed by MEK and GSK3β inhibition + LIF cannot be used to capture NANOG-only ICM cells in porcine embryos, suggesting that additional signals may be needed to prevent hypoblast formation. Importantly, however, we show that naive state can be imposed effectively in piPS cells. These naive cells are LIF dependent, resembling the cells reported by others, however the analysis of *STELLA* and *REX1* expression showed significant heterogeneity between piPS cell lines. *STELLA* and *REX1* expression are good markers of naive stem cells, and are not expressed in primed cells. In agreement with findings in mouse cells, *STELLA* and *REX1* expression increased in piPS cells transferred to 2/3i + LIF conditions. From our results we conclude that *STELLA* expression is indicative of higher homogeneity in iPS cell cultures, which is reflected in the compactness of the colonies, low numbers of STELLA negative cells, and increased expression of *NANOG, OCT-4,* and *REX1*. A test comparing the differentiation capacity of *STELLA^high*(cells in 3i + LIF) versus *STELLA^low* (cells in FCS + LIF) cells demonstrated that both lines could differentiate to the three somatic lineages, but importantly, *STELLA^high* cells had increased capacity to differentiate to *VASA*-expressing germ cell precursors. These findings lead us to suggest that naive pluripotency in piPS cells can be imposed in 2i and 3i + LIF conditions, conferring these cells increased in vitro differentiation capacity to germ line precursors. Future in vivo studies will determine the capacity of these cells to colonize chimeric embryos. The authors wish to thank G. Wood & Sons ltd. for their help in obtaining the uterine tracts, and N. Pritchard for supplying the animals used in this study. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: AR CA RA. Performed the experiments: AR RA. Analyzed the data: AR RA. Contributed reagents/materials/analysis tools: CA RA. Wrote the paper: AR CA RA.

# Introduction Peste des Petits Ruminants (PPR) is an acute, highly contagious, and febrile viral disease that affects both domestic and wild small ruminants. Clinically, PPR infection is characterized by leukopenia, pyrexia, congestion of mucosal surfaces, ocular and nasal discharge, erosive stomatitis, diarrhea, and suppression of the immune system often leading to co-infection\[–\]. The disease is considered to be a big “dam” to the development of sustainable agriculture across the developing countries as a notifiable disease listed and has recently been oriented by the World Organization for Animal Health (OIE) and the Food and Agriculture Organization (FAO) for eradication with the aim of global elimination of the disease by 2030. The causative agent of the disease, Peste des Petits Ruminants Virus (PPRV), is an enveloped virus with a non-segmented, negative-strand RNA genome, and is classified as a member of the genus *Morbillivirus* along with Measles virus (MV) and Rinderpest virus (RPV). The virus genome length is approximately 16,000 nucleotides (nt) and encodes six essential structural protein such as H and fusion (F) proteins and the two non- structural proteins\[–\]. Unusually noticeably, H and F are two kinds of glycoproteins of the virus. To infect cells, enveloped viruses must interact with their receptors on host cells for enveloped proteins and induce fusion of the viral membrane with the host cell membrane in order to enter host cells. Dramatically, H protein is responsible for attaching the virions to the host receptors and generally regulates viral adsorption and entry, determining pathogenicity, and releasing newly-produced viral particles. In addition, neutralizing antibodies against H protein serve as protective antibodies in PPRV infection. Multiple researches have shown the epitopes for the neutralizing antibodies and T cell determinants against H protein. Further analysis of H protein will further our understanding of the molecular evolution of the virus and the presence of host-specific mechanisms, which will be significant in the control and elimination of the PPR disease. Phylogenetically, PPRV can be divided into four genetically distinct lineages, I-IV, based on the partial region of *N* or *F* gene. The lineages are associated with the geographic distribution of the virus, with lineages I and II exclusively isolated in West Africa, lineage III in Arabia, East Africa, and southern India, and lineage IV in southern Asia, the Middle East, and more recently, northern Africa. Herein, we rebuild the phylogenetic tree of the *H* gene to estimate the evolutionary relationships of various strains in different regions or dates. This analysis will enable a more precise evolutionary and phylogenetic assessment of the relationships among the four PPRV genetic lineages. As a result of the selection pressures of the host immune system, antigenic changes may occur through positively-selected amino acid substitutions. The selective pressures could lead to a change in the antigenic epitopes and may cause a change in the receptor between the host and virus. Indeed, attachment glycoprotein of an essential antigen of respiratory syncytial virus and parainfluenza virus could show frequent positive selections in the antigenic epitopes of the protein. Recently, Muniraju M. evaluated the mean ratios of Nonsynonymous (dN) to Synonymous (dS) substitutions per site of concatenated coding regions of the PPRV genome, which indicated there may be positive selections sites in the coding region of *H* gene. However, codon sites under positive selection are not known. Thus, it may be important to analyze the molecular evolution of *H* gene in PPRV. We estimated positive selection sites in the genes by the CODEML program of PAMLX. Furthermore, single likelihood ancestor counting (SLAC), fixed-effect likelihood (FEL), internal fixed-effect likelihood (IFEL), and random effect likelihood (REL) method in the Date monkey (<http://www.datamonkey.org>) were used to evaluate the result. We used the likelihood method with more effective models of evolution to analyze our collected data. A *Morbillivirus* initially infects a broad range of immune system cells and then spreads to epithelial cells. However, SLAM (signaling lymphocyte activation molecules; also known as CDw150), which appear on the surface of activated T and B lymphocytes, macrophages, and dendritic cells, acts as a receptor for all morbilliviruses, including PPRV. Recent reports determined that the interaction of SLAM with MV was not only the first step of the virus invading the host, but also the important cause of the pathological changes in the host organism and presentation of clinical symptoms. Because of this background, some researchers have explored the interaction between SLAM and PPRV, and recent evidence demonstrated the inference that SLAM serves as the cell receptor for H protein in PPRV. However, there is inconclusive evidence that would reveal more detailed data of the interaction, such as domains or motifs of the protein-protein interaction. As we all known, the small segments of antigen may be the antigenic determinants or the epitopes that are limited in eliciting the preferred immune response. We hope to find the antigen epitopes that have the ability to produce neutralizing antibody with the combination of H and SLAM in response through to understand H protein interaction with SLAM in the three-dimensional (3D) structure of protein. It is well known that the 3D structure of a protein determines its function. Above all, the knowledge about the 3D structure of protein is extremely important for epitopes vaccine design, novel drug discovery, and rational drug design. The 3D structural model of the H protein isn't yet known. Nevertheless, the structure of MV H (the highly homologous protein of PPRV H) was reported by Christopher’s and Yusuke’s groups in 2007, which have provided the greatest aid in studying 3D structure of PPRV H. Therefore, in our study, we used homology modeling to build the structural model of H protein. In order to understand H protein-SLAM interaction, the 3D model of H-SLAM complex was predicted by homology modeling using Discovery Studio (DS) v4.5 (Accelry, San Diego). Hence, our study is of significance to the future exploration of molecular evolution and characterization of H protein, which may provide useful insights in studying molecular epidemiology of the virus and in predicting epitopes for vaccine design against PPR. # Materials and Methods ## Sequence Collection and Multiple Sequences Alignment (MSA) Biochemically characterized *H* gene of PPRV (Nigeria/75/1 strain, GenBank accession No.X74443) as query sequence was obtained from the National Center for Biotechnology Information (<http://www.ncbi.nih.gov/>) database. The full-length coding sequence of *H* gene for available PPRV strains from the disease-endemic countries were downloaded from GenBank at NCBI using inference sequence to search by BLASTn against nucleotide collection (nr/nt) database with default parameters. We concentrated an integrated collection of the full-length coding sequences for *H* gene of PPRV. Sequences with 100% identity and suspicious were excluded from the dataset in further analyses. Additionally, the outgroup, RPV (Kabete O strain, GenBank accession No.X98291; RBOK strain, GenBank accession No.Z30697), was added to the dataset. Multiple sequences alignment (MSA) was the critical step in phylogenetic analysis and the alignment played a crucial role in extracting evolutionary information from a large number of sequences. In this study, MSAs for coding sequences of *H* gene were performed by MUSCLE with the algorithm using default parameters in the software Molecular Evolutionary Genetics Analysis (MEGA) 6.0.6 (<http://www.megasoftware.net/>). The result of test of substitution saturation performed on all sites using DAMBE5 showed that the observed Iss is significantly lower than Iss.c (p = 0.0000), and the estimated Transition/Transversion bias (R) is 4.44. After processing, 36 strains of PPRV were identified in this study, which are shown in. ## Phylogenetic Analysis Bayesian tree was reconstructed based on codon positions using all available sequences we obtained to estimate the evolutionary relationship by TOPALi v2.5. Prior to tree construction, the best-fitting evolution model GTR+I+G was selected Akaike Information Chriterion (AIC) in Modeltest v3.7 (June 2005) and MrMTgui (Nuin, 2007) The Bayesian tree was rebuilt using the following settings: 4 runs, 1,000,000 generations, 100 of sample frequency and 25% burn in. To confirm the topology of the Bayesian tree, Maximum Likelihood (ML) tree was rebuilt using the HIVw+G+F model, which was selected using Akaike Information Chriterion (AIC) in ProtTest 2.4. And the ML tree topology was optimized using NNI method and a BioNJ starting tree and 1000 Bootstrap replicates were used to estimate the reliability of the internal nodes. Bayesian and ML, the different methods, were used to reconstruct phylogenetic tree to ensure the reliability of the tree. ## Selective Pressures Analysis To test positive selection in individual codons of the *H* gene, ω ratios were compared using the two ML frameworks, the CODEML program of PAML X software package and the Hyphy package implemented in the Data Monkey Web Server. In CODEML, site model was selected to detected positive selection. Three pairs of models, Model0 (one ratio) and Model3 (discrete), Model1 (nearly neutral) and Model2 (positive selection), Model7 (β) and Model8 (β & ω), were performed to evaluate our collected data set. M0 *versus* M3 was used to test for rate heterogeneity among amino acid sites; M8 *versus* M7 and M1 *versus* M2 were used to determine the possible sites under selection. LRTs were performed to test the positive selection sites by comparing the nested models. The Bayes empirical Bayes (BEB) analysis and the Naive Empirical Bayes (NEB) analysis in the case of comparing models was used to calculate the Bayesian posterior probabilities (BPP) of the codon sites under positive selection. In this test, the *H* gene tree was employed as the guide tree. In order to further determine the positively selected sites acquired in the CODEML, a battery of ML methods were performed in the Data Monkey Web Server. The automatic model selection tool on the server was used to research the best fitting nucleotide substitution models. Four different codon-based maximum likelihood methods, SLAC, FEL, REL and IFEL were used estimate the dN/dS (also known as Ka/Ks or ω) ratio at every codon in the alignment of all available sequences. The significance level of *p*-values (*p* \< 0.05) for SLAC, FEL, and IFEL was used to determine codons under positive selection and accepted sites, and Bayes Factor \>50 for REL as candidates for positive selection sites. ## Homology Modeling The monadelphous structural model of H protein and PPRVHv-shSLAM (the H protein of PPRV vaccine strain and the sheep SLAM) complex structural model were generated by Homology modeling in the DS v4.5. The sequence of Hv (the H protein of PPRV vaccine strain, GenBank accession No.CAJ01700) used as target sequence to build monadelphous model. The coupling of Hv and shSLAM (GenBank accession No.NP_001035378) were used as target sequence to build a compound model. In terms of the refined model, the Accelrys CHARMm force field was used for the simulation, and Ramachandran plot was selected for the validation. The structures were optimized by Generalized Born (GB) implicit-solvent model. Energy minimization of the selected model was performed *via* the Steepest Descent method (5000 steps). The model structure was refined using a Conjugate Gradient method (2000 steps). For a better understanding of the protein-ligand interactions, three complexes were generated and the DOPE scoring function and PDF Total Energy were used for score calculation for selecting best complex model. Otherwise, Calculate Interaction Energy (Binding) was selected to analyze for binding interaction and binding energy using the following settings: Permittivity = 75 F/m and Nonbond List Radius = 14.0, Nonbond Higher Cutoff Distance = 12.0 and Nonbond Lower Cutoff Distance = 10.0. Thereafter, alanine- scanning mutagenesis for binding interface was used to evaluate the affinity of protein-ligand. Figures were produced using DS v4.5 Client. # Results ## Phylogenetic Analysis Phylogenetic analysis demonstrated that the topology patterns of the phylogenetic tree included four clades, and dendograms with highly coincident was constructed in two cases (Bayesian and ML trees). Obviously, PPRV could be divided into four lineages based on *H* gene. The isolates from India, China, Ethiopia, Morocco, Turkey, and Kurdistan belonged to lineage IV and were confined in a big clade in the phylogenetic tree, which could be further branched into four minor clades. UAE86 strain, Oma83 strain, Eth94 strain, Uga12 strain and Ken11 strain belonged to lineage III. The strains from Nigeria and Gha10 strain, Sen13 strain, and CIV09 strain were confined in a clade, which belonged to lineage II. And Sen69 and CIV89 strains belonged to lineage I. ## Selective Pressures Analysis In the PAML test, phylogenetic tree was utilized to detect the positive selection, and the estimated parameters of different models and the LRT results are provided in. M0 indicated that *H* gene underwent the purifying selection (ω \< 1) and very short regions or on only a few sites underwent positive selection. LRT (-2ΔlnL = 67.9324) of M0 *vs* M3 was performed, and the results (*p* \< 0.001) showed the presence of selection pressure at codon positions. One site (aa246) could be underwent positive selection with a posterior probability larger than 0.99 by BEB and NEB. But, there were no sites identified as being under went positive selection with the posterior probability larger than 0.95 by BEB and NEB. However, the results of the LRT test statistic of M7–M8 comparison was 4.1982 (*p* \< 0.05). M7 is rejected in favor of M8revealing that the model which allowed positive selection better than that which did not allowed positive selection. Additionally, the BEB and NEB approaches detected one site (aa246) under positive selection with BPP values larger than 0.95. To confirm the test, we estimated positive selection sites in the strains using SLAC, FEL, IFEL, and REL methods in the Data Monkey. Detailed data are shown in. Results showed that aa246 underwent positive selection by the SLAC method test; three sites (aa246, aa419, and aa574) underwent positive selection by the FEL and IFEL method test; two sites (aa176 and aa246) underwent positive selection by the REL method test. The common site of positive selection estimated by four methods was aa246 which was also determined by PAMLX. ## Homology Modeling Based upon blast results, MV H (PDB ID:2ZB5) was considered as an ideal homologue and used as a template for homology modeling as both viruses belong to the same genus (genus *Morbillivirus*); therefore, the MVH-maSLAM complex (the MV H-marmoset SLAM complex, PDB ID:3ALZ) was used as a template for complex homology modeling. The result of building model for Hv showed that the fold of H monomer was a β-propeller with six blades (B1∼B6) surrounding a large cavity. Every one of blade contained four antiparallel β-strands (S1–S4), and the six blades were connected through the loops between S1 of the next and S4 of one module The Ramachandran plot (φ/ψ) distribution of the backbone conformation angles for each of the residues of the refined structure revealed that 91.9% and 7.8% were in the favored region and allowed region, respectively. In predicted PPRVHv-shSLAM complex by homology modeling, the PPRVH head domain exhibited the six-bladed β-propeller fold (rainbow colors) and formed a monomer. SLAM-V (purple) exhibited a typical β-sandwich structure with two β-sheets: BED and AGFCC′. In addition, we compared the interface residues of PPRVHv-shSLAM and MVH-maSLAM. To verify the contributions of the selected residues on protein- ligand complex, we designed a series of mutants for both proteins and evaluated their affinity. The residues in protein-ligand binding interface were respectively selected for alanine scanning. In the interface of PPRVHv-shSLAM complex, the mutation energy of Phe552, Arg533, Arg191, Tyr553, Tyr543, Arg503, Asp505, and Pro554 on PPRVHv and Lys78, Lys76, His130, His62, Leu92, Leu64, Val128, and Glu123 on shSLAM were more than 1.00 kcal. # Discussion In this study, we utilized *H* gene of available PPRV strains, which were divided into four lineages based on *N* or *F* gene from different endemic countries, to estimate the molecular evolution of the virus and analyzed the interaction between the virus and host. It well known that PPRV can be divided into four genetically distinct lineages (I-IV). Historically, isolates from Africa were numbered lineage I-III according to the spread of the virus from West Africa to East Africa. Keeping to this viewpoint, West African isolates from Senegal and Ivory Coast belonged to lineage I; the strains derived from Ghana and Nigeria were formed lineage II; and the strains detected in Ethiopia, UAE and Oman were formed lineage III. Because there was no effective action to control PPR in the past for a period of time, the prevalence of the disease has become more complicated. There was the cross of the lineage in different areas. PPRV of lineage IV appear constantly in the Arabian Peninsula, the Middle East, Southern Asia, and across several African territories, and still in rapid spread. Since PPR emerged for the first time in the Tibet autonomous region of China in 2007, the disease has been reported in China. Because of some effective measures were taken, the epidemic was relatively stable. A study about lineage 4 has suggested that the *N* gene is more divergent and therefore more suitable for phylogenetic distinction between closely related circulating viruses. However, H protein is a membrane glycoprotein of PPRV and very conservative. It may provide useful insights on the epidemic of PPR to study epidemiology of PPRV based on *H* gene. Following our research, the phylogenetic analysis on the basis of *H* gene showed highly similar evolutionary relationship contained in four clusters of lineages. Thus, *H* gene could also be used to analyze the evolutionary relationships of different isolates. Therefore, the evolutionary relationship of different isolates based on *H* gene has a guiding significance in epidemiology. In the longest co-evolution between viruses and host, molecular adaptations that optimize the organism’s survival in a given living environment can be accumulated and inherited. To probe deeply into how that evolutionary power drives protein evolution and how gene sequences differ in the various strains or species, phylogenetic analysis by maximum likelihood was selected to detect selection or adaptation in this study. The M0 model (ω = 0.206, ω \< 1) demonstrated that the entire *H* gene underwent the purifying selection, indicating that overall *H* gene was relatively conserved. This might be correlated with the functionalism and structuralism of H protein which mediates the virus binding to host cellular receptors, a first step in the progression of PPRV infection. However, a study by Yang, Z *et al*. pointed out that the average ω ratio for overall sequence was very seldom greater than 1 and the positive selection seldom happened in some structural domains. In our study, some positive selection sites were found; and to the best of our knowledge, this is the first study that reports this finding in PPRV H. Recently, Kimura H. *et al*. reported that there were 8 positive selection sites in 297 strains of *H* gene of congeneric virus (MV) in the prevalent Asian genotypes, D3, D5, D9, and H1. Nevertheless, a previous report showed that MV in 50 strains of the *H* gene of the various genotypes had 14 positive selection sites. Our research about the *H* genes of PPRV indicated that four lineages had one positive selection site (*p* \< 0.05) by BEB analysis. However, four positive selection sites were estimated using the HyPhy package. In the previous report, Sinnathamby *et al*. used autologous skin fibroblast cells to identify a motif from 400 to 423 amino acids (24 amino acids long), involving aa419, in the H protein of PPRV, which carries a CTL epitope, and is highly conserved among morbilliviruses, especially PPRV and RPV. In addition, Renukaradhya *et al*. also mapped the H protein carrying B cell epitopes using mAbs for the presence of immuno-dominant regions, which suggested that the regions from 538–609 amino acids, including in aa574, are immunoreactive. Amino acids 575–583 domain is unique to the RPV H protein and is an immuno-dominant epitope. Amino acids 575 on RPV H protein is arginine. Interestingly, amino acid 575 on most PPRV H protein is also arginine. We believe that aa574 may also be an important position in the epitope of H protein. These amino acid changes at positive selection sites may confer a disadvantage to PPRV vaccine protection. So far, there has been no relevant report about aa176. Bewilderingly, the common site (aa246) of positive selection estimated by both methods is yet to be unraveled. It is unclear whether this site has any relationship with the interaction between the virus and host receptor or is a result of host specificity. Nonetheless, our research about the interaction between H protein and SLAM showed that the interface does not contain the common site. Our study indicated that amino acid change at positively selection sites did not lead to changes in the receptor of PPRV. All the positive selection sites were mapped on the H protein structure. The structure of the measles virus hemagglutinin was reported by Christopher’s and Yusuke’s groups in 2007. The fold of MVH is a β-propeller with six blades (B1\~B6) surrounding a large cavity, which is similar to our predicted structural model. Every one of blade contained four anti-parallel β-strands (S1–S4), and the six blades were connected through the loops between S1 of the next and S4 of one module. In 2011, Yusuke’s group presented crystal structure of MVH-maSLAMV complex (MV H and the V-set Ig domains of marmoset SLAM). The crystal structure suggested that there were four small areas (sites 1–4) of the binding interface between MVH and SLAM. Our results show that the groove in the B4 blade and B5 of the head domain binds to the AGFCC′ β-sheets of the membrane- distal ectodomain of shSLAM, which the C β-sheet, C′β-sheet, and the loop of the two β-sheets are core regions of the interface. From the analysis of acting force, hydrogen bond, and hydrophobic force are the maintenance forces. There is an abundant presence of hydrogen bonds, electrostatic interactions and hydrophobic interactions in the interface of PPRVHv-shSLAM. The complex model showed four small regions of the binding interface: 1) is an intermolecular β-sheet assembled by the polypeptide backbones of the Arg191–Arg195 strand of PPRV Hv and the Val126–Ser132 strand of shSLAM, and there is a hydrogen bond in Arg191 of Hv and Ser50 of SLAM; 2) is formed by salt bridges of Asp505 and Asp507 in the proposed acidic patch of PPRV Hv to Lys78 of shSLAM. In addition, there are two hydrophobic interactions of PPRVHv Arg503 with shSLAM Lys77 and Leu92; 3) is formed by Pi interactions of PPRVH Phe552 to Pro554 and shSLAM Val128 and His130 as well as some hydrophobic interactions of PPRVH Ser550, Phe552, Tys553, Pro554, and Arg556 with shSLAMLys76, His130, Val128, Glu124, and Val126; and 4) are aromatic stacking of Tyr541 and Tyr543 of PPRV H with Phe119, and Pi interaction of Arg533 with His62 of shSLAM, and hydrogen bonds of PPRV H Asp530 and Arg533 with Lys78 and Glu123 of shSLAM. Comparing the interaction energy of the protein-ligand, we found out that mechanism of interaction of both the viruses and the receptors were very analogous, and interaction energy was also similar. Our studies indicate PPRVHv-shSLAM binding interface and MVH- maSLAM binding interface were consistent. Phe552, Arg533, Arg191, Tyr543, Arg503, Asp505 and Asp507 on PPRVHv; and Lys78, Lys76, His130, His62, Leu64, Val128 and Glu123 on shSLAM play a key role in the PPRVHv-shSLAM interaction. Recently, Yusuke’s group used crystal structures of MV-H–receptor complexes and functional studies to reveal the mechanism behind the long-term success of the measles vaccine *via*. The crystal structures of MVH–SLAM and MVH-CD46 complexes showed that both receptors target an exposed small region, and mutagenesis studies demonstrated that Nectin-4 binds to this area. The above studies proved that this region is the compact domain of the epitope. Therefore, it is of great importance to study whether those sites play the crucial role in pathogenicity, interspecies transmission, or immunoreaction, which may contribute in designing novel vaccines against PPRV for elimination of the disease. # Supporting Information The authors would like to thank Dr Gao, Dr Ma and Dr Liu for timely advice and discussion. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YD ZL LC. Performed the experiments: ZL RY LC. Analyzed the data: ZL RY LC. Contributed reagents/materials/analysis tools: ZL XZ. Wrote the paper: ZL RY.

# Introduction Prostate cancer is the second-leading cause of cancer-related deaths in both Europe and the United States. Androgen deprivation therapy (ADT) is considered a key treatment as monotherapy or in combination with other regimens. Most patients initially respond to ADT; however, the intrinsic nature of the heterogeneity of tumor cells results in resistance to treatment and progression into highly morbid disease termed castration-resistant prostate cancer (CRPC) within 18–24 months. End-stage CRPC is commonly associated with osseous metastasis, which causes significant mortality and morbidity with the development of severe skeletal complications in affected patients. Recent clinical approaches using agents that target distinct mechanisms of action, including tubulin-binding chemotherapy (cabazitaxel); immunotherapy (sipuleucel-T); CYP-17 inhibition (abiraterone); androgen receptor (AR) blockade (enzalutamide); and radioisotope therapy (radium-223) although have shown promising results in delaying skeletal complications and also improving overall survival, the management of patients with metastatic CRPC remains a challenge, with a mean survival time of less than 19 months. Thus, the development of new agents with more effective antitumor activity is crucial for treating metastatic CRPC. In particular, drugs are needed that target hormone-refractory prostate cancer cells regardless of differentiation state, with various levels of androgen receptor (AR) and prostate-specific antigen (PSA) expression. Past genetic and molecular studies held that tumor cells are heterogeneous and their subsequent metastases are the results of non-random, sequential and multistep selective processes among preexisting cell populations. However, recent studies have evidenced the intricate intercellular communication between stromal and cancer epithelial cells leading to permanent genetic and behavioral changes not only in the epithelial cells but also in cancer-associated stromal cells that drives further genetic and gene expression changes in prostate cancer cells. Through a series of complex, intimate bi-directional communications between prostate cancer and the host stroma, cancer cells gain additional growth and survival advantages and ultimately disseminate to distant organs with lethal effect. Thus, co-targeting of both the tumor and its supporting stromal cells can improve therapeutic responses and overall survival of patients with prostate cancer. Given that gene therapy has been identified as the preferred treatment for metastatic cancers, developing an effective strategy for the delivery and expression of therapeutic genes in the tumor epithelium and adjacent stroma is essential to making such treatment available. Osteonectin (also known as basement membrane-40 \[BM-40\] and secreted protein acidic rich in cysteine \[SPARC\]) is widely distributed in several tissues during development and cellular injury and plays a major role in regulating cell adhesion, proliferation, migration, and tissue remodeling. In the bone microenvironment, osteonectin is the most abundant non-collage matrix protein which is highly expressed early in osteoblastic differentiation and is critical for the maintenance of bone mass. The role of osteonectin in prostate cancer has been identified as a chemoattractant for bone-invasive prostate cancer cells. High levels of osteonectin expression have been observed in prostate cancer cell lines derived from metastases and in prostate cancer metastatic foci. In addition, elevated osteonectin levels in primary prostate cancer was associated with the subsequent development of metastasis, indicating that prostate cancer cell metastasis to the bone is mediated in part by the osteonectin-mediated promotion of cancer cell migration, protease activity, and invasion. Because osteonectin expression occurs in both tumor epithelial cells and bone cells, the osteonectin promoter could be used to drive a therapeutic gene co-targeting the bone metastatic prostate cancer and its supporting microenvironment, regardless of the basal level of AR and PSA expression. In this study, we sought to create a promoter–mediated therapeutic agent that co-targets prostate cancer and its surrounding stromal cells. We found that osteonectin was upregulated in prostate cancer epithelial cells and cancer- associated stromal cells compared with their normal counterparts. We designed a novel hON promoter (hON-522E) containing positive transcriptional regulatory elements identified in both the promoter and exon 1 region of the hON gene. We also constructed a replication-defective adenoviral vector bearing a herpes simplex virus thymidine kinase (hsv-TK) gene driven by a highly active 580 bp hON promoter (hON-522E). Treatment with this construct, Ad-522E-TK, in combination with the prodrug ganciclovir (GCV) was found for the first time to kill both androgen-independent prostate cancer and bone stromal cell lines *in vitro* and to inhibit the prostate tumor growth in an xenograft model. Because of the heterogeneity of human prostate tumors, Ad-522E-TK may be applied as an adjunct therapy with other AR-targeting modalities for treatment of hormone refractory and bone metastatic prostate cancers. # Materials and Methods ## Cell lines and Cell culture The human prostate cancer cell lines LNCaP, C4-2, C4-2B, PC3, DU145 and PC3M, and a human osteosarcoma cell line MG63 that have been used in our previous studies were maintained in T medium and supplemented with 5% fetal calf serum (FBS). hFOB 1.19 human osteoblast and HS27A human bone marrow stroma cell lines were purchased from ATCC (Manassas, VA, USA) and maintained in a 1:1 mixture of Ham’s F12 Medium/Dulbecco’s Modified Eagle’s Medium and RPMI 1640 medium, respectively, with 10% FBS. The adenovirus packaging 293 cell line (Microbix Biosystems Inc., Toronto, Ontario, Canada) was maintained in Minimal Eagle’s Medium and supplemented with 10% FBS and 2 mM glutamine (Invitrogen, Carlsbad, CA, USA). All cell culture media and reagents were purchased from Invitrogen. All cells were cultured in a 37°C incubator with 5% CO₂ and were passaged upon reaching 90% confluence. ## Human subject and Laser capture microdissection (LCM) Experiments with human samples were reviewed and approved by the institutional review board (IRB) at Taipei Medical University (TMU-JIRB 20131253). A prostate tissue microarray (TMA) containing 49 tissue cores representing samples from 40 cases of prostate cancer and 9 matched normal adjacent tissues was obtained from Super Bio Chips (CA4, Seoul, Korea). Frozen human prostate tissue samples were obtained from the TMU Joint Biobank based at the Taipei Medical University and affiliated hospitals from subjects with written informed consent. LCM was used to isolate selectively pure populations of prostate cancer cells and non- neoplastic epithelial cells as well as the stroma adjacent to Gleason grade 3 and grade 4 glands and stroma adjacent to non-malignant glands from frozen sections of prostatectomy specimens derived from four patients. In brief, eight- micron-thick sections of frozen tissue were stained using the Arcturus HistoGene Frozen Section Staining Kit according to the manufacturer’s instructions. Areas of the selected cell populations were microdissected from the sections and collected using the ArcturusXT system and CapSure HS LCM Caps. The settings of the laser were as follows: spot diameter set at 30 μm, power 70 mW, and pulse duration 25 milliseconds. Total RNA from each microdissected sample was extracted using the PicoPure RNA Isolation Kit following the manufacture’s protocol. LCM instrument and all reagents used in the experiment were obtained from Thermo Fisher Scientific Inc. (Madison, WI, USA). ## Real-time RT-PCR Total RNA was extracted from cells using the High Pure RNA Isolation Kit (Roche, Indianapolis, IN, USA) according to the manufacturer’s instructions. The first- strand complimentary DNA was synthesized using random primers and Moloney murine leukemia virus reverse transcriptase (Invitrogen) and subjected to real-time PCR using the LightCycler 480 with the Light Cycler TaqMan master kit combined with the Universal ProbeLibrary probe (Roche) according to manufacturer instructions. Target genes were amplified using specific primers for hON (forward: 5′-GTGCAGAGGAAACCGAAGAG-3′ and reverse: 5′-TGTTTGCAGTGGTGGTTCTG-3′, probe no. 77), and housekeeping gene, HSPCB (forward: 5′- AGCCTACGTTGCTCACTATTACG-3′ and reverse: 5′- GAAAGGCAAAAGTCTCCACCT-3′, probe no. 55). The relative gene expression of osteonectin in the cell lines is represented as 2^-ΔCT, with ΔCT determined by subtracting the average housekeeping gene HSPCB threshold cycle from the average target gene value. ## Plasmid construction All promoter constructs were generated using the TOPO TA cloning system (Invitrogen) and subsequently digested using appropriate restriction sites in the polylinker to allow insertion into the vector pGL3-basic (Promega, Madison, WI, USA) containing the coding region of the firefly luciferase gene. All promoter constructs had the same 5′ end. The spacer between the GGA-boxes 1 and 2 were deleted by recombinant PCR using the following primer sets: 522-N: (5′ACTAGTAGCAGCTTGTCTTGTC3′), spdel-C: (5′CTTCTCCCCTGTCTCTGTCTT3′), and spdel-N: (5′AAGACAGAGACAGGGGAGAAG3′) combined with downstream primers: Intron-C: (5′TACCTCAGTGGCAGGCAGGCAG3′), Exon-C: (5′CAGGCAGGCAGGCGGCAG′), and Hafner-C: (5′GCGCGCTCTCCGGGCAGTCTG3′) to construct hON-522I, hON-522E, and hON-522H, respectively. Genomic DNA was isolated from DU145 cells for the template. All constructs including PCR-generated DNA fragments were confirmed by sequencing. ## DNA transfection and Luciferase assay Cells were cotransfected with various osteonectin promoter luciferase reporter plasmids and pCMV-βgal (galactosidase) in a 5:1 molar ratio using the lipofectamine 2000 transfection reagent (Invitrogen). After 48 h of incubation, cell extracts were prepared for the luciferase and β-gal activity assessments using the Luciferase Assay System and β-Galactosidase Enzyme Assay System (Promega), respectively, according to manufacturer’s instructions. Relative luciferase activity was calculated as the firefly luciferase relative light units (RLU) divided by the corresponding value for the β-gal activity present in each sample. Three independent experiments were performed in triplicate. ## Design of adenoviral vectors The Ad-522E-TK and Ad-522E-Luc adenoviral vectors (type 5) were designed and mass-produced according to the established protocol. Briefly, the plasmids p522E-TK and p522E-Luc containing a hON-522E promoter and herpes simplex virus TK gene and luciferase gene, respectively, were constructed by inserting the expression cassette into the E1A deleted region of the Ad5 adenoviral shuttle vector pΔ E1sp1B. A replication-defective recombinant Ad522E-TK adenovirus was generated in the 293 cells by co-transfecting these cells with both the expression shuttle plasmid and a circular Ad genome plasmid (pJM17) using the standard calcium-phosphate precipitation method; Ad-CMV-TK and Ad-CMV-Luc were constructed similarly. ## Thymidine kinase activity assay TK activity in Ad-522E-TK- and Ad-CMV-TK-infected cell lines was assayed through phosphorylation of \[³H\]GCV. Briefly, the supernatant fraction of crude cell extracts was mixed with an equal volume of TK assay buffer containing 0.2 μCi \[³H\]GCV (Moravek Biochemicals, CA, USA), 3 mM MgCl₂, 3 mM ATP, 10 μg/μL bovine serum albumin, and 50 mM sodium phosphate buffer (pH 6.5). The reaction mixture was incubated at 37°C for 90 min, transferred to DE-81 discs (Whatman, Hillsboro, OR, USA), air-dried, and washed thoroughly with 50% ethanol. Phosphorylated \[³H\]GCV bound to the discs was determined with a scintillation counter (Beckman Coulter Inc., Schaumburg, IL, USA). Three independent experiments were performed in triplicate. ## *In vitro* cytotoxicity assays Cells were seeded on 24-well plates at a density of 2 × 10⁴ cells per well. After 24 h, the cells were infected with Ad-522E-TK in the range of 0–100 Multiplicity of Infection (MOI). After a 2 h adsorption, the virus-containing medium was replaced with fresh medium. After 24 h, the cells were incubated in the presence or absence of 10 μg/mL GCV for 5 days followed by a crystal violet staining; subsequently, the relative cell number was assessed at an optical density (OD) of 590 nm after staining. Each experiment was performed in triplicate. ## Animal study All animal experiments were approved by and complied with the regulations of the Institutional Animal Care and Use Committee (IACUC) of Taipei Medical University (LAC-2013-0047). Six-week-old male athymic nude mice BALB/cAnN.Cg- Foxn1nu/CrlNarl mice were obtained from the National Laboratory Animal Center (Taipei, Taiwan). The animals were kept under standard pathogen-free conditions and cared for according to the criteria outlined in the National Academy of Sciences Guide for the Care and Use of Laboratory Animals. For analysis of the hON-522E promoter activity *in vivo*, 1 × 10⁵ PC3M cells in 10 μL PBS were injected into the ventral prostates of mice. At 5 days after tumor cell injection, tumor-bearing or untreated mice received intravenous administration of 1 × 10⁹ pfu Ad-522E-Luc or Ad-CMV-Luc through the tail vein (n = 5). Mouse organs and prostate tumor xenografts were harvested for the luciferase activity assay 2 days after viral injection. For assessment of Ad-522E-TK combined with GCV-induced inhibition of tumor growth *in vivo*, 5 × 10⁵ PC3M cells in 50 μL PBS were injected subcutaneously into the flanks of the mice. When the tumor became palpable (3–4 mm in diameter), the animals were randomly assigned to 4 experimental groups (n = 8 for each group): group 1, PBS treatment; group 2, GCV only; group 3, Ad-522E-TK combined with PBS; and group 4, Ad-522E-TK combined with GCV. Ad-522E-TK (50 μL; 2 × 10⁹ pfu) in PBS was administrated through intratumoral injection every other day for 3 times. GCV (100 μL) was administrated daily via intraperitoneal injection at a dose of 40 mg/kg body weight for 2 weeks. Bidimensional tumor measurements were performed twice a week with calipers, and the tumor volume was calculated using the simplified formula for a rotational ellipsoid (L × w² × 0.5236). Animals were sacrificed 5 weeks after therapy using CO₂ for euthanasia, and tumors were excited for histopathologic examination. ## Immunohistochemistry (IHC) IHC staining was performed using the Novolink Polymer Detection System (Leica Microsystems, Newcastle Upon Tyne, UK) as previously described. Ki-67 protein was detected in tumor xenografts with mouse antihuman Ki-67 monoclonal antibody (1:100; NCL-Ki67-MM1, Leica Biosystems). Apoptosis was evaluated using the Apo- BrdU-IHC In Situ DNA Fragmentation Assay Kit (BioVision, Inc., Milpitas, CA, USA) as described. IHC staining for osteonectin was performed on prostate TMA using an anti–human-osteonectin monoclonal antibody (1:50; NCL-O-NECTIN, Leica Biosystems). Each TMA spot was examined by a pathologist (J.L.C.) using Allred scoring system. The numerical value for overall intensity (intensity score) is based on a 4-point system: 0, 1, 2, and 3 (for none, light, medium, or dark staining). The numerical value for percent stained (proportion score) is determined by a geometric division; no stain = 0; ≤1/100 cells stained = 1; ≤1/10 cells stained = 2; ≤1/3 cells stained = 3, ≤2/3 cells stained = 4; all cells stained = 5. Addition of the two values gives the total Allred score. ## Statistical analysis All data are presented as the mean (standard deviation \[SD\]) unless otherwise specified. Analysis was performed using the two-tailed Student’s t-test. P \< 0.05 was considered significant. # Results ## Expression of osteonectin in prostate cancer and stromal cells The association of osteonectin expression with human cancer progression was initially evaluated through real-time RT-PCR on the androgen-responsive LNCaP and androgen-insensitive PC3 prostate cancer cell lines. In the series of LNCaP lineage cell lines, the expression of osteonectin correlated with increased bone metastatic potential, in which the hormone refractory and bone metastatic C4-2B expressed a higher level of osteonectin than did its parental androgen-dependent LNCaP and androgen-independent C4-2 cells. Similarly, PC3M, the highly metastatic derivative expressed 18 times higher levels of osteonectin compared with their parental PC3 cells that was originally derived from bone metastases of prostate cancer. These results revealed a correlation between elevated osteonectin expression and metastatic CRPC progression. As prostate cancer bone metastasis is considered as a microenvironment-driven disease, the higher levels of osteonectin expressed by human bone stromal cells than that of prostate cancer cells was observed using hFOB osteoblast and HS27A bone marrow derived fibroblast cell lines. To assess the differential expression of osteonectin between the noncancerous and cancerous prostate epithelial cells as well as the normal and cancer-associated stromal cells in same individuals, LCM dissected samples from primary prostate tumors were used. Among the pairs of the normal and malignant prostate tissues, 3 of 4 patients’ samples displayed a significantly increased expression of osteonectin in cancer cells and cancer- adjacent stromal cells in comparison with their normal counterpart; and the other one showed a decreased expression pattern in tumor tissues. To further validate the association between the overexpression of osteonectin and prostate carcinogenesis in clinical specimens, immunohistochemical (IHC) analysis was performed on a prostate tissue microarray containing 40 primary prostate tumors and 9 matched normal prostate samples. In each core, immunoreactivity for osteonectin were measured in sections containing normal stroma, normal epithelium, tumor stroma, and tumor epithelium. Although these differences were not statistically significant (P = 0.2654 and 0.4042), the ranks of total Allred score in the overall samples of tumor epithelium and tumor stroma was higher relative to normal epithelium and normal stroma, respectively . The difference was more significant when only matched pairs of prostate tumor and normal prostate samples was analyzed (P = 0.085 and 0.2165 for epithelial staining and stromal staining, respectively). The lack of statistical significance may be due to the small sample size. In addition, strong staining of osteonectin can be detected in the sample of prostate cancer bone metastases. The direction of the effect suggests the autocrine and paracrine actions of osteonectin by tumor and tumor microenvironment causing prostate cancer malignancy and metastasis. ## Identification of additional transcriptional regulatory elements in the human osteonectin promoter region The GGA-box 1 in the human osteonectin (hON) promoter region is vital for maximal transcriptional activity, whereas the pyrimidine-rich spacer between GGA-boxes 1 and 2 exerts a downregulatory effect. Comparison of the bovine, mouse, and human osteonectin exon 1 DNA sequences revealed a notable multiple repeat of the sequence CCTG in all species, with a consistent cluster of 7–8 bases upstream from the start of exon 1. We therefore examined 2 hON- promoter–reporter constructs, p522E-Luc and p522H-Luc, to investigate the role of the CCTG sequence in hON promoter function. The promoter fragment in p522H-Luc is identical to that in pGL2-spdel, which has potential transcriptional activity in human cell lines. p522E-Luc has a 5′ region of hON promoter sequences similar to that of p522H-Luc, but the 3′ end extends to bp + 62, in which 4 CCTG units are included. Transfection of these constructs and pGL-3-TATA (which serves as a reference of promoter activity) into human bone stromal cell lines, including hFOB, HS27A, and osteosarcoma MG63 showed a marked increase in luciferase activity by p522E-Luc compared with p522H-Luc. Further 3′ extension of the promoter into bp +73 located at the intron 1 region (p522I-Luc) resulted in decreased luciferase activity, clearly demonstrating that the region between bp +39 and bp +62 is responsible for the additional upregulation of hON promoter activity in human bone cells, whereas the region between bp +63 and bp +73 contains a negative regulatory element. We next evaluated the transcriptional activity of the hON-522E promoter in various prostate cancer cell lines reflecting different stages of prostate cancer progression. The transfection data showed that hON-522E promoter activity was detected in all cell lines tested. However, comparison of the luciferase activity of these transfected prostate cancer cell lines and those with endogenous osteonectin RNA expression revealed that hON-522E promoter activity is relatively higher in AR-negative, more aggressive, and metastatic cell lines, including DU145, PC3 and PC3M cells as compared to AR-expressing LNCaP, C4-2 and C4-2B cell lines. To assess the potential utility of the hON-522E promoter in expressing a transgene in a tissue-specific manner *in vivo*, an adenoviral vector containing the hON-522E promoter or the ubiquitous cytomegalovirus (CMV) promoter driving the luciferase reporter gene (Ad-522E-Luc or Ad-CMV-Luc) was administered intravenously to male athymic mice either healthy or carrying orthotopic PC3M tumors. After 2 days of viral administration, major organs, including the liver, lung, kidney, spleen, intestine, heart, brain, colon, testes, prostate, and muscles, and PC3M tumor xenografts were harvested to measure the luciferase expression. We observed that in the normal organs, Ad-522E-Luc mediated luciferase transgene expression in the liver, spleen, and lung was significantly lower than that of Ad-CMV-Luc, with reduction rates of 16%, 41%, and \<1%, respectively. Consistent with a previous study revealing increased expression of osteonectin mRNA in pancreatic ductal epithelial cells, luciferase activity transduced by Ad-522E-Luc in the pancreas was 34 times higher than that induced by Ad-CMV-Luc. While the luciferase activity was barely detected in the normal mouse prostate, regardless of Ad vector administration, extremely high luciferase expression by Ad-522E-Luc was observed in PC3M prostate xenografts, with expression 5 times greater than that by Ad-CMV-Luc. These results suggest that the hON-522E promoter is suitable with respect to efficient and selective transgene expression for transcriptional targeting of AR-negative and metastatic prostate cancer cells. ## Evaluation of *in vitro* cytotoxicity of recombinant Ad-522E-TK in prostate cancer and bone stromal cells To assess the feasibility of hON promoter–directed co-targeting gene therapy for prostate cancer, we designed a replication-deficient adenoviral vector, Ad-522E-TK, carrying the hON-522E promoter–driven herpes simplex virus TK gene. The effectiveness of TK gene delivery in the prostate cancer and bone stromal cell lines was determined using the TK enzymatic activity assay after exposing these cells to Ad-522E-TK and normalized to the external control Ad-CMV-TK for viral infectivity. The cell lines tested, including prostate cancer and bone stromal cell lines, revealed successful TK gene transduction by Ad-522E-TK with an at least two-fold stronger activity in AR-negative prostate cancer cells DU145, PC3 and PC3M as well as bone stromal cells MG63 and HS27A in comparison with AR-positive LNCaP lineage cell lines, which was similar to the result of luciferase reporter activity by p522E-Luc. More strikingly, Ad-522E-TK–infected PC3M cells exhibited a higher rather than lower in TK activity than the cells infected with Ad-CMV-TK; this result was consistent with that of Ad-522E-Luc transgene activity in PC3M xenograft tumors. Upon additional prodrug ganciclovir (GCV) treatment, Ad-522E-TK markedly and dose-dependently decreased the growth of PC3M cells with a higher efficacy than that of Ad-CMV-Luc; Ad-522E-TK or GCV alone exerted no cytotoxic effects on PC3M cells. In addition, infection of MG63 and HS27A with Ad-522E-TK also induced significant cell death when combined with GCV. These results demonstrated the ability of Ad-522E-TK to target both prostate cancer and bone stromal cells. ## Antitumor effect of *in vivo* Ad-522E-TK treatment in combination with GCV To determine the therapeutic efficacy of the hON-522E-promoter–directed gene therapy in the treatment of CRPC human prostate cancer *in vivo*, we evaluated the antitumor effect of Ad-522E-TK combined with GCV in a PC3M subcutaneous xenograft model in nude mice. The PC3M xenograft was observed to be an extremely aggressive tumor that grew to form large tumors (\>1 cm diameter) in 5 weeks. The growth of PC3M tumors was significantly inhibited in animals treated with Ad-522E-TK combined with GCV (p \< 0.005). In controls, Ad-522E-TK alone nonsignificantly inhibited tumor growth (p \> 0.5), and GCV alone exerted no tumor regression effect as compared to the vehicle (PBS)-treated group. The treated mice did not reveal any gross change in weight. Histological analyses revealed healthy and packed tumor cells in either PBS-, GCV alone-, or Ad-522E-TK combined with PBS-treated control groups, whereas large necrotic regions were observed in tumors excised from animals treated with the combination ofAd-522E-TK and GCV. In addition, an extensive decrease in proliferative cancer cells (Ki-67) concomitant with intensely TUNEL-stained apoptotic cells (TUNEL) within the tumor area confirmed effective cancer cell killing by Ad-522E-TK combined with GCV. Although the current model cannot resemble the microenvironment of bone metastatic prostate cancer, these results as a proof-of-principle study demonstrate the efficacy of Ad-522E-TK plus GCV gene therapy for the treatment of CRPC prostate cancer. # Discussion Prostate cancer commonly metastasizes to bony sites, where cells acquire an aggressive, rapidly proliferating, androgen-independent phenotype. The ability of several non-collagenous matrix bone proteins to increase migration and invasion by prostate cancer cell lines has been examined, supporting a model in which bone-derived factors attract prostate cancer cells preferentially to such sites. In addition to bone stromal cells such as osteoblasts (hFOB) and bone marrow fibroblasts (HS27A), the results of the present study reveal an apparent upregulation of osteonectin mRNA expression in bone metastatic prostate cancer cells (C4-2B) as compared with their non-bone metastatic sublines (LNCaP and C4-2). Similarly, in the bone-derived prostate cancer PC3 lineage, the highly metastatic variant PC3M cells expressed greater level of osteonectin. This observation is fully consistent with the hypothesis that osseous metastatic prostate cancer cells are osteomimetic, allowing the cells to thrive in bone. Unlike osteocalcin promoter that has been proposed for prostate cancer/bone stroma co-targeting gene therapy based on the conventional osteoblastic reactions demonstrated in experimental models and clinical manifestations of prostate cancer skeletal metastasis, elevated osteonectin expression was also observed in prostate cancer epithelium and cancer-associated stroma in primary prostate tumors through RT-PCR and IHC analyses. Therefore, osteonectin-targeted therapy has a broad advantage in blocking paracrine and autocrine events in tumor microenvironment at both prostate and bone sites to prevent and cure bone metastasis in patients. Although information regarding the expression and potential function of osteonectin in human malignant tumors has increased considerably, little is known about the positive and negative regulatory elements in the human osteonectin promoter compared with those in cattle and mice. Two enhancers and a repressor element for the human promoter have been mapped between nucleotides 165 and 130, 51 and 120, and 130 and 121, respectively In the present study, we mapped a region containing a cluster of 4 CCTG repeats between bp +39 and +62 within exon 1 of the human osteonectin gene that is responsible for the additional upregulation in bone stromal cells. The mechanism whereby exon 1 confers tissue-specific gene expression in bone and in particularly in highly bone metastatic prostate cancer PC3M cells is unclear. This positive regulation is probably exerted directly at the transcriptional level through physical interaction with a transcriptional factor that exists in bone cells and prostate cancer cells that have acquired osteomimetic properties. The transcription factor that directly binds to the CCTG repeating sequences is yet to be investigated. In addition, the presence of the repeating CCTG units in the untranslated 5′ mRNA leader probably influences either mRNA translation or stability. In summary, we have defined and characterized a human osteonectin promoter (hON-522E) that contains only positive transcriptional regulatory elements and is highly active in AR-negative and metastatic prostate cancer cells. When combined with GCV, the recombinant Ad vector Ad-522E-TK was found to effectively induce PC3M cell death *in vitro* and slow the growth of pre-existing PC3M prostate tumors *in vivo*. In this study, only the antitumor effects of Ad-522E-TK through intratumoral administration were examined; the efficacy and safety concerns of systemic use of Ad-522E-TK were not addressed. However, the antitumor effects of osteonectin-promoter–mediated gene therapy for prostate cancer bone metastasis without other organ toxicity might be achieved by understanding the cis- and trans-acting factors in the hON promoter. Because of the heterogeneity of both primary and metastatic prostate tumors, hON-522E-mediated gene therapy may be applied as an adjuvant to AR-targeted therapeutics for treating metastatic CRPC. # Supporting Information We would like to acknowledge Mr. Chun A. Changou for his excellent technical support at TMU Core Facility. We also thank TMU Language Editing for editorial assistance. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: CLH SYS. Performed the experiments: JLC CYH. Analyzed the data: JLC SYS. Contributed reagents/materials/analysis tools: LWKC YHS KCC SDY YRL. Wrote the paper: CLH.

# Introduction Online activity has become an inherent part of modern society and a way of living among its members. The Internet provides a vast amount of information to its users as well as aid and assistance in times of need. During the current financial crisis and the subsequent economic and production crises, most of the developed as well as developing economies are being hit by an economic downturn that is tightly connected with growing unemployment. Job loss can be a very traumatizing experience with long lasting impact on those who experience it. Seeking a new job then becomes an integral part of everyday life. In the current digitalized era, job seeking does not restrict itself to job offices because seekers (as well as potential employers) increasingly turn to the Internet as a source of information and new possibilities. As such, job seekers leave a digital track of their activity. The analysis and examination of various patterns of online activity have become a fruitful branch of research in recent years with some exciting applications such as elections, investment allocation, private consumption and consumer behavior, future orientation, earnings announcements, spread of disease, and economics and finance. In terms of unemployment and its possible examination utilizing the online activity of Internet users, there has been some research done in the area that focuses primarily on Google engine search queries. The first study focusing on the possible connection between Google search activity and unemployment rates in Germany shows the usefulness of adding search query data into models. The subsequent research analyzed the connection between queries and claims for unemployment benefits in the USA and the unemployment rate itself has also been studied. A job search activity index based on Google search data has even been developed. Most of these studies focus on the US economy and its modeling, while other economies have been studied rather marginally. Here, we focus on the possible connection between job-related search queries using the Google search engine and the unemployment rate in countries of the so- called Visegrad Group (the Czech Republic, Hungary, Poland and Slovakia). Our contributions lay in the following. First, we focus on a set of countries that would be normally treated as marginal and that are thus not often studied. However, if claims are made for the utility of online search activity (and specifically Google searches), its efficiency should be shown not only for developed and well covered countries but also for the smaller ones, and the results might prove useful to all policy makers even in these types of regions. Second, we provide a careful step-by-step procedure for unemployment modeling, focusing not only on simple correlations but also on nowcasting with an out-of- sample analysis. Third, a cross-countries comparison is delivered that is rather unique given comparable studies that focus primarily on one specific country. # Results The unemployment rates have undergone a quite heterogenous evolution in the analyzed countries. In the Czech Republic, the rate ranged between 4% and 9% between the years 2004 and 2013. Initially, there was a significant downward trend from 2004 to 2008, when the rate dropped from 9% to 4%. As the recession hit the Czech Republic in 2008, the rate started to increase to reach a new maximum of 8.5% in 2010. Since that time, the unemployment rate has fluctuated between 7% and 8.5%. The Hungarian unemployment rate steadily rose from the year 2004 to 2010, at which date it reached a new maximum of nearly 12%. After that point, the rate fluctuated for almost 3 years between 10.5% and 12% and started declining in 2013. Unemployment in Poland experienced a steady decline from the astronomical rate of nearly 22% in the 2004 to 6% in 2009. However, as the recession hit Poland, the unemployment rate began rising again. With some minor fluctuations, it smoothly increased to the current level of approximately 10%. In Slovakia, the unemployment rate appears to have a similar pattern as that in the Czech Republic, although on a different scale. In 2004, Slovakia had an unemployment rate of almost 20%. This rate linearly decreased to 8% in 2009. With the recession, the unemployment rate quickly escalated to 16%, and it has fluctuated around that point since. The differences between countries are well illustrated in the descriptives statistics provided in. Mainly for Poland and Slovakia, we observe wide fluctuations in time which are mirrored in higher variance and range of the unemployment rates. Even though the rates are close to being symmetric and show no strong excessive kurtosis, normality is rejected for all but Slovakia. The evolution of Google searches is illustrated in. There are evident seasonal patterns in all four series. Hungary is characterized by a quite regularly increasing trend in Google searches, whereas Slovakia shows the opposite, and the remaining two analyzed series remain quite stable over time. Although there appears to be a connection between Google searches and the unemployment rates for the Czech Republic and Hungary that is visible to the naked eye, we can hardly claim any relationship without a proper analysis. Information about the data collection and selection of Google search terms is provided in the Methods section. ## Basic relationship As the initial step, we present the results of stationarity tests, which tell us whether we should analyze the original series or some of their transformations. In, we show the results of the ADF and KPSS tests (see the section for more details) for the original as well as the logarithmic series and their first differences. The outcome is quite straightforward, as we do not reject unit roots for either of the original series (or their logarithmic transformation for the Google searches; we do not examine the logarithmic transformation for the unemployment time series as these are already in the percentage representation). Further testing, which is not reported here, shows no cointegration relationship between unemployment and the search query series, so we need to proceed with the first differences of the series. For most of the cases, we support the stationarity of the first differences. In the analysis, we further proceed with the first differences of the unemployment rate and the first logarithmic differences of the Google searches. We opt for this specification because the combination of percentage representation and logarithmic transformation allows for a straightforward interpretation as an elasticity, i.e., as a proportional relationship. For the very basic relationship between the unemployment rate and the intensity of job-related searches on Google, we study the equation $$\begin{array}{r} {\Delta\text{UR}_{t} = \alpha_{0} + \alpha_{1}\Delta\log\left( \text{GI} \right)_{t} + \varepsilon_{t}} \\ \end{array}$$ where ΔUR_*t* and Δlog(GI)_*t* represent the first difference of an unemployment rate at time *t* and the first logarithmic difference of the Google searches at time *t*, respectively, for a given country, and *ε*_*t* is an error term. The elasticity between Google searches and the unemployment rate from is estimated at 0.686 (with the *p* = value of 0.008), 0.185 (0.125), 0.331 (0.216) and 0.606 (0.001) for the Czech Republic, Hungary, Poland and Slovakia, respectively, with the heteroskedasticity and autocorrelation consistent (HAC) standard errors. The proportional relationship thus varies across the analyzed countries, but it remains positive for all four countries and statistically significant for two out of the four (at the 1% significance level). Specifically, the relationship is very strong for the Czech Republic and Slovakia, with values above 0.6. This result shows that the changes in the unemployment rate are well projected into the online search queries for vacancies and job-related terms. Studying the connection between these two variables thus appears to be promising and worth further utilization and investigation. ## Nowcasting Macroeconomic time series, such as unemployment rates, have a special property that is not present for financial series or other series in natural sciences—they are available with a pronounced lag. This lag occurs due to data processing and collection, which usually take several months; even after this period, there are sometimes corrections to the reported values. This characteristic makes a series that is available immediately without any lag and that is strongly correlated with the variable of interest very useful for forecasting the present value of the variable without waiting several months. This type of forecasting of the present is usually referred to as “nowcasting”. In the previous section, we showed that Google searches for job-related terms are related to the unemployment rate, which makes search queries potentially useful for the nowcasting of unemployment. As a nowcasting model, we consider $$\begin{array}{r} {\Delta\text{UR}_{t} = \beta_{0} + \sum\limits_{i = 3}^{L}{\beta_{i}\Delta\text{UR}_{t - i}} + \sum\limits_{j = 0}^{L}{\gamma_{j}\Delta\log\left( \text{GI} \right)_{t - j}} + \varepsilon_{t}} \\ \end{array}$$ where the unemployment rate is assumed to be available with a three month lag (please refer to the Methods section for more details about the data collection). We again consider the differenced series due to the stationarity issues discussed above. As a base model, we use the model specification in without the Google terms so that the competing model is defined as $$\begin{array}{r} {\Delta\text{UR}_{t} = \delta_{0} + \sum\limits_{i = 3}^{L}{\delta_{i}\Delta\text{UR}_{t - i}} + \nu_{t}.} \\ \end{array}$$ For both models, we consider a maximum lag *L* which we set as *L* = 3, 6, 12. This way, we are able to comment on the quality of models with regards to the amount of information taken into consideration. The upper bound is set to 12 months as the unemployment series are usually strongly cyclical. The results of the nowcasting models are summarized in. In the table, we show the adjusted *R*² (${\bar{R}}^{2}$) as a measure of the models’ quality controlling for the number of explanatory variables. We observe that for all countries, the inclusion of the Google series strongly enhances the model. The most promising results are reported for the Czech Republic where the models improve strongly regardless the number of lags taken into consideration. For the other three countries, we observe that the base models with 3 and 6 lags are very weak, even reaching negative values of ${\bar{R}}^{2}$. A strong seasonal (annual) pattern in the unemployment rates is thus visible here. The Google series are thus evidently useful for the in-sample modeling of the series, which is supported by a statistical significance of the online searches for all countries and regardless the maximum lag used. However, it is the out-of-sample performance that eventually matters. We divide the analyzed period into two—a training (fitting) period and a nowcasting period. The model parameters are fitted on the data between 01/2004 and 12/2011 (96 observations). Nowcasting performance is then evaluated on the series between 01/2012 and 12/2013 (24 observations). The “Google model” is compared to the base model using the Diebold-Mariano test (see the section for more details). In, the resulting statistics are summarized. For the Czech Republic, the model using Google searches is on average outperforming the base model for each lag selection. For the maximum lag of 3 and 12 months, the difference is statistically significant. Similar results are reported for Hungary, for which the Google specification outperforms the base model for all lag selections as well. The difference is statistically significant for lags up to 3 and 6 here. For Poland, we find statistical significance only for the maximum lag of 6 months, and for Slovakia, the base model even outclasses the “Google model”. The results are thus quite diverse. # Discussion Data showing the online activity of Internet users has proven useful in various fields. Nowcasting the unemployment rate is one of these fields. Contrary to the prevailing trend in the literature focusing on well-developed (Western) countries, we have utilized job-related Google searches in the Visegrad Group countries, i.e., the Czech Republic, Hungary, Poland and Slovakia. Although data availability and Internet utilization might not be as widespread in this region as one would expect for developed countries, we have shown that, in fact, online searches provide a strong foundation for unemployment modeling. In summary, we have shown that the basic dynamics of Google searches for job- related terms closely follow the unemployment rates. Further, we have utilized this idea to successfully nowcast the unemployment rates using the current and lagged values of Google searches. Our findings indicate that the information left online by Internet users can be easily utilized even for small or medium countries such as those of the Visegrad Group. However, this is true mainly for the Czech Republic and Hungary but much less so for Poland and especially Slovakia. Even though one of the reasons might be a different level of Internet penetration in the regions, we speculate that such diversity is caused by different customs and specifically international mobility in the analyzed countries. The Czech and Hungarian nationals usually do not move for work inside their country and even less so internationally. However, this is not the case for the Polish and Slovak citizens which are willing to move for work abroad. This is also reflected in the Google searches. For the job-related searches in Poland, one of the topical keywords is “gumtree” which relates to the UK advertising website, which well reflects willingness of the Polish nationals to seek job abroad. In a similar way, the topical keywords for Slovakia include “práce” and “prace” which are the Czech equivalents to the Slovak “práca” and “praca”. This again mirrors frequent moves of the Slovaks to the Czech Republic in their search for work. The international mobility of the Polish and Slovaks thus has a strong influence on the informative value of the Google searches and their usefulness for the unemployment modeling. # Methods ## Data The monthly unemployment data for the Czech Republic, Hungary, Poland and Slovakia have been obtained from the Eurostat database (<http://ec.europa.eu/eurostat/>). The basis for the unemployment measurement among EU countries is the EU Labour Force Survey (EU LFS)—a continuous and harmonized household survey that, in accordance with EU legislation, is conducted in each member state. The monthly data from Eurostat are estimates based on the results of EU LFS. Because there are no legal obligations that the EU countries deliver monthly data, these data are often interpolated/extrapolated using national surveys or registered unemployment data. Eurostat defines an unemployed person as someone aged between 15 and 74 without work during the reference week who is available to start working within two weeks and who has actively sought employment at some time during the last four weeks. In our analysis, we use the general (both sexes, 15–74 years old), raw (not seasonally adjusted) unemployment rate. We use these data because we do not know the method used to make the seasonal adjustment and because the Google data are also not seasonally adjusted. The Google search queries data have been downloaded from the Google Trends webpage (<http://www.google.com/trends/>). As the languages of the studied countries differ, we have looked for various terms. As Czech, Polish and Slovakian are all Slavonic languages, the searched words are very similar or even the same. For Czech, we searched for “práce” and “prace” (i.e. both with and without diacritics), for Polish “praca” and for Slovakian “práca” and “praca” (again both with and without diacritics) but also the Czech “práce” and “prace”, which turn out to be very frequently searched for by Slovakians. For Hungarian, we used the terms “állás” and “munka”. These all are equivalents for the English “job” and “work”. Other related words have not passed throught the Google threshold or only incomplete series are available. As the Google Trends engine allows to compare up to five terms for a given setting (in our case a country and a time frame), we can use more series for each country. In the cases of more searched queries (the Czech Republic, Hungary and Slovakia), we sum the series together and rescale them to 100. This can be done as if multiple series with the same spatial and temporal characteristics are obtained from the engine, these share a common scale. The weekly series obtained from the Google Trends site have been transformed to monthly series on the basis of the number of days in the month. All series, both the unemployment rate and the Google searches, are studied between January 2004 and December 2013 (120 observations). The dataset is provided in the. ## Stationarity A stochastic process {*x*_*t*} is stationary if for every collection of time indices 1 ≤ *t*₁ \< *t*₂ \< *t*_*m*, the joint probability distribution of (*x*_*t*₁, *x*_*t*₂, …, *x*_{*t*_*m*}) is the same as the joint probability distribution of (*x*_{*t*_1+*h*}, *x*_{*t*_2+*h*}, …, *x*_{*t*_*m*+*h*}) for all integers *h* ≥ 1. To test for stationarity, we utilize the Augmented Dickey- Fuller (ADF) test and the KPSS test. The tests have opposite null hypotheses and thereby provide a complementary pair, which is commonly used for stationarity testing. In the ADF procedure, the OLS regression is run on $$\begin{array}{r} {\Delta x_{t} = \alpha_{0} + \theta x_{t - 1} + \gamma t + \Delta x_{t - 1} + \Delta x_{t - 2} + \cdots + \Delta x_{t - p} + \varepsilon_{t}} \\ \end{array}$$ to perform the test, where *α*₀ and *γt* are an intercept and a time trend, respectively, and *p* represents the lag order. The null hypothesis under which the series contains a unit root is found for $$\begin{array}{r} {H_{0}:\theta = 0} \\ \end{array}$$ against the alternative $$\begin{array}{r} {H_{A}:\theta < 0.} \\ \end{array}$$ The ADF test statistics are then computed as the usual *t*-statistics, which, however, follow a more complicated distribution under the null hypothesis. The null hypothesis of the KPSS test is opposite to that of the ADF test, i.e., the KPSS test has the null hypothesis of stationarity. The test is based on the OLS regression of the series {*x*_*t*}: $$\begin{array}{r} {x_{t} = \alpha_{0} + \gamma t + k\sum\limits_{i = 0}^{t}\xi_{i} + \varepsilon_{t}} \\ \end{array}$$ where *α*₀ and *γt* again represent an intercept and a time trend, respectively, and *ξ*_*i* are independent and identically distributed random variables with zero mean and a unit variance. The null hypothesis of stationarity is found for $$\begin{array}{r} {H_{0}:k = 0} \\ \end{array}$$ against the alternative $$\begin{array}{r} {H_{A}:k \neq 0.} \\ \end{array}$$ The KPSS test statistic is defined as $$\begin{array}{r} {KPSS = \frac{\sum_{t = 1}^{n}S_{t}^{2}}{n^{2}{\hat{\omega}}_{T}^{2}}} \\ \end{array}$$ where *S*_*t* is the partial sum of the residuals $$\begin{array}{r} {S_{t} = \sum\limits_{i = 1}^{t}{\hat{\varepsilon}}_{i}} \\ \end{array}$$ and ${\hat{\omega}}_{T}^{2}$ is an estimator of the spectral density at frequency zero. ## Nowcasting accuracy To compare the forecasting accuracy of the proposed models, we utilize the Diebold-Mariano test based on absolute errors. An absolute error is simply defined as *a*_*i* = ∣*f*_*i* − *y*_*i*∣ where *f*_*i* stands for a nowcast value and *y*_*i* is an observed real value. We do not use also popular squared errors here as these are usually applied to magnify higher errors. However, the nowcasting errors in our case are always lower than unity which makes the squared errors counterintuitive so that we avoid using them. Diebold and Mariano propose a test to compare the predictive accuracy of two competing forecasts. Let ${\{\varepsilon_{t}^{1}\}}_{t_{0}}^{T}$ and ${\{\varepsilon_{t}^{2}\}}_{t_{0}}^{T}$ be the sequences of forecast error losses from two competing forecasting measures by a particular loss function (absolute errors *a*_*i* in our case). The null and alternative hypotheses are then stated as $$\begin{array}{r} {H_{0}:\mathbb{E}\left\{ \varepsilon_{t}^{1} \right\}_{t_{0}}^{T} = \mathbb{E}\left\{ \varepsilon_{t}^{2} \right\}_{t_{0}}^{T}} \\ \end{array}$$ $$\begin{array}{r} {H_{A}:\mathbb{E}\left\{ \varepsilon_{t}^{1} \right\}_{t_{0}}^{T} > \mathbb{E}\left\{ \varepsilon_{t}^{2} \right\}_{t_{0}}^{T}.} \\ \end{array}$$ The Diebold-Mariano test assesses the accuracy based on the loss differential $$\begin{array}{r} {d_{t} = \left\{ \varepsilon_{t}^{1} \right\}_{t_{0}}^{T} - \left\{ \varepsilon_{t}^{2} \right\}_{t_{0}}^{T}} \\ \end{array}$$ which is equal to zero under the null hypothesis. The Diebold-Mariano statistic is then $$\begin{array}{r} {S = \frac{\overline{d}}{\sqrt{{\hat{LRV}}_{\overline{d}}/T}}} \\ \end{array}$$ where $\bar{d}$ is the mean loss differential and ${\hat{LRV}}_{\bar{d}}$ is a consistent estimate of the asymptotic (long-run) variance of $\sqrt{T}\bar{d}$ defined $$\begin{array}{r} {LRV_{\overline{d}} = \gamma_{0} + 2\sum\limits_{j = 1}^{\infty}\gamma_{j},\;\;\gamma_{j} = \text{cov}\left( d_{t},d_{t - j} \right).} \\ \end{array}$$ Under the null hypothesis, the testing statistic goes to a standard normal distribution so that $S\overset{A}{\sim}N(0,1)$. # Supporting Information Google data are registered trademarks of Google Inc., used with permission. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JP LK. Performed the experiments: JP LK. Analyzed the data: JP LK. Contributed reagents/materials/analysis tools: JP LK. Wrote the paper: JP LK.

# Introduction The incidence of lung cancer is increasing worldwide, particularly in developing countries. In China the death rate of lung cancer has been increasing from 7.1 to 30.8 per 100,000 during 1975–2005. People dying due to lung cancer accounted for 23% of total amount of cancer death in 2005. 80% of primary lung cancers are non-small cell lung carcinoma (NSCLC) which is characterized by a long asymptomatic latency and poor prognosis. Without an early diagnostic approach, over 40% of lung cancer patients develop metastasis at the time of diagnosis and survive for a short time period under a conventional chemotherapy. Only 15% of NSCLC patients can survive over 5 years. Thus, it is essential to identify biomarkers for early prediction of lung cancer. Cigarette smoking is a well known driving force for lung cancer development. The lifetime risk of developing lung cancer is 17.2% in male smokers and 11.6% in female smokers, which is much higher than that in nonsmokers with 1.3% in male and 1.4% in female. Although most lung cancers are associated with cigarette smoking, it is statistically estimated that 15% of them in males and 53% in females, accounting for about 25% of all lung cancers, are not attributable to cigarette smoking. Lung cancers arising in nonsmokers are more frequently adenocarcinomas, affect females disproportionately more than males, and have regional differences ranging from 10–15% in Europe and North America to 30–40% in Asian countries. Moreover, nonsmoker lung cancers have improved survival and are more sensitive to epidermal growth factor receptor (EGFR) tyrosine kinase inhibitor therapy. It might be due to that the activation of EGFR by gene mutations appears more often in nonsmoker lung cancers. Taken together, lung cancer in nonsmokers would probably be considered a separate cancer category. If so, it would rank as the seventh most common cause of cancer death worldwide. However, whether these clinical-pathological and molecular differences between lung cancer in nonsmokers and smokers are related to cigarette smoking is still unknown. The interest in cancer-associated changes in gene methylation has grown enormously in recent years with the speculation that the promoter methylation status may provide an early biomarker for tumorigenesis. Silencing of genes by aberrant promoter hypermethylation has been recognized as a key event in cancer initiation and progression. Highly sensitive assays such as methylation-specific PCR (MSP), which could detect one methylated allele in the presence of 10³–10⁴ unmethylated alleles, have been used to assess gene-promoter methylation in primary tumors, serum, plasma, sputum or specimens from the aerodigestive tract epithelium. Numerous studies have investigated the methylation statuses of specific genes in body fluids and tumor tissues of lung cancer patients, and identified more than 60 genes as being epigenetically silenced in lung tumors. The proof-of-concept studies suggested that gene- specific promoter methylation occurs as an early event in lung cancer. For example, hypermethylation of *p16^INK4α* (also known as cyclin- dependent kinase inhibitor 2A, *CDKN2A*) or O⁶-methylguanine-DNA methyltransferase (*MGMT*) was found in sputum samples 5 to 36 months prior to clinical diagnosis. Moreover, the frequency of *p16^INK4α* hypermethylation increased progressively from 17% in basal-cell hyperplasia to 24% in squamous metaplasia, and to 60% in squamous cell carcinomas. The correlation between gene methylation and recurrence of lung cancer has also been reported. The promoter hypermethylation of several genes including *p16^INK4α*, cadherin 13 (*CDH13*), Ras association (RalGDS/AF-6) domain family member 1 (*RASSF1A*) and adenomatous polyposis coli (*APC*) in NSCLC specimens is associated with early recurrence after surgery. Moreover, the reversibility of DNA methylation modification makes it possible that the expression of genes that have undergone epigenetic silencing becomes reactivated by the inhibitors of DNA methylation. Many agents such as 5-azacytidine and zebularine are currently being tested in clinical trials. Despite the significance of gene promoter methylation in predicting incidence or prognosis and in epigenetic therapy of lung cancer, the manner in which these epigenetic lesions accumulate during carcinogenesis is not completely understood. The missing links among environmental factors, DNA methylation changes, and lung cancer limit the applications of methylated genes as a biomarker for early detection of lung cancers. The correlation between cigarette smoking and aberrant gene methylation has been extensively studied, but the results are inconsistent and inconclusive. The present study mainly focused on *p16^INK4α* gene as it is the first gene identified in lung cancer, and is transcriptionally silenced predominantly through aberrant promoter hypermethylation. Here, we performed a literature- based systematic review and meta-analysis to quantitatively analyze the correlation between cigarette smoking and *p16^INK4α* gene methylation in NSCLC patients. # Methods ## Ethics Statement An ethics statement was not required in this study. ## Literature search We systematically searched for all published articles indexed in PubMed database from 1966 to June 18, 2011 with the Medical Subject Headings (MeSH) and corresponding free text: “smok\* AND (p16\* OR CDKN2A OR INK4A) AND (methylation OR epigene\*)”. We also manually searched the references of these publications in order to retrieve additional studies. Only those published as full-text articles and in English or in Chinese were included as candidates. ## Inclusion and exclusion criteria Studies were selected for analysis if they met the following criteria: 1) they were original epidemiological studies on the correlation between cigarette smoking and *p16^INK4α* methylation; 2) they were conducted in lung cancer patients; 3) the specimens used for methylation analysis must include surgically resected primary tumor samples, while other specimens such as sputum, serum, bronchial lavage samples and normal or non-malignant lung tissue may be used, but not essential; 4) *p16^INK4α* methylation status was examined using methylation-specific PCR (MSP) or quantitative MSP (QMSP); 5) the subjects in every study comprised nonsmokers and smokers (former smokers and/or current smokers), irrespective of minor discrepancies of the definition of nonsmokers over all studies. To avoid duplication of data, we carefully checked the author names, research institutions and procedures for enrolling participants. Where several publications reported from the same population data, only the most rounded study with more information was included. ## Data collection For each eligible study, we collected information regarding authors, year and source of publication, country of origin, inclusion criteria, exclusion criteria, histology of lung cancer, types of biological specimen, number of participants, participants' age and gender, smoking behaviour, *p16^INK4α* methylation frequencies in nonsmokers and smokers and the method for methylation detection. All included studies used nonsmokers as a control group, though some of them did not provide the definition of nonsmoker. In studies defining nonsmoker, there were three different definitions of nonsmoker: (1) daily cigarette consumption×years of smoking = 0; (2) less than 100 cigarettes in entire lifetime; (3) less than 20 pack-years. Since it is impossible to redefine nonsmoker based on a unified standard, we combined nonsmokers in our meta-analysis according to their original group in each individual study. Correspondingly, subjects except nonsmokers were smokers comprising former smokers and current smokers, and light, moderate or heavy smokers. Then data was integrated in 2×2 tables demonstrating the methylation/unmethylation of *p16^INK4α* gene according to the cigarette smoking status (smoker/nonsmoker). All data were extracted independently by two reviewers using a standard form. Minor discrepancies were resolved by the authors' discussion. ## Meta-analysis and statistical analysis The foremost analysis examined the differences in the frequency of *p16^INK4α* methylation in lung cancer tissues between smoker and nonsmoker patients. Summary OR was obtained across all studies. Heterogeneity was examined using the I² statistic, which represents the proportion of variation in the effect sizes that is attributable to true differences across studies rather than to a random error. Results without heterogeneity were pooled using the fixed-effect model, following the Mantel-Haenszel method. Otherwise, the random effect analysis with the method of DerSimonian and Larid was used. The meta-analyses were performed using Stata statistical software (version 10.0, Stata Corporation, College Station, Texas, USA). Results were shown in forest plots, where the sizes of the boxes for individual studies were inversely proportional to the variances of the log relative risks, and the horizontal lines represent 95% confidence interval (CI). The frequencies of *p16^INK4α* methylation in smokers and nonsmokers were compared by Wilcoxon signed rank test. The coefficients of Spearman's rank correlation were calculated between frequency of *p16^INK4α* methylation and sample size. These analyses were performed using SPSS for Windows (version 11.5, SPSS Inc., Chicago, IL, USA). # Results ## Study characteristics Following the inclusion and exclusion criteria described above, 19 studies, – were included in the analysis. The characteristics of these studies are summarized in. Of these 19 studies, eight defined the subtypes (adenocarcinoma or squamous cell carcinoma) of NSCLCs. Ten studies were conducted in Asia (2 in China, 6 in Japan and 2 in Korea), five were in USA, and the remaining four were in Australia, Greece, Chile, and multi-areas in Asia-Pacific regions (USA, Australia, Japan and Taiwan) respectively. In two studies, lung adenocarcinoma (AC) patients and squamous cell carcinoma (SCC) patients were analyzed separately, therefore they were treated as separate items in the meta-analysis. As for the primer sequences of MSP, 15 studies used the same primers designed by Herman et al. in 1996. The primer sequences for detecting methylated *p16^INK4α* gene were 5′-TTA TTA GAG GGT GGG GCG GAT CGC-3′ (sense) and 5′-GAC CCC GAA CCG CGA CCG TAA-3′ (antisense). The size of the PCR product for the methylated reaction was 150 bp. The primer sequences used for the unmethylated promoter were 5′-TTA TTA GAG GGT GGG GTG GAT TGT-3′ (sense) and 5′-CAA CCC CAA ACC ACA ACC ATA A-3′ (antisense). The size of the PCR product for the unmethylated reaction was 151 bp. ## Combined results and subgroup analyses In general, the frequencies of *p16^INK4α* methylation ranged from 19 to 83% (median 34%) in smoker patients, which were much higher than those in nonsmoker patients with range from 0 to 91% (median 20%). In the meta-analysis, 2802 NSCLC patients including 2037 smokers and 765 nonsmokers were included in pooling the overall correlation estimation. Under the fixed-effects model, the pooled odds ratio (OR) of *p16^INK4α* methylation in smoker patients, compared to nonsmoker patients, was 2.25 with 95% CI = 1.81–2.80. Of six studies only including AC patients, the subgroup analysis showed no significant difference in the association between cigarette smoking and *p16^INK4α* methylation when comparing AC patients (OR = 2.38, 95%CI = 1.75–3.23) and all NSCLC patients (OR = 2.25, 95% CI = 1.81–2.80). Because there was a borderline significant but moderate degree of heterogeneity among the 19 studies (I² = 33.1%, *P* = 0.072), we performed sensitivity analyses to identify potential sources of heterogeneity. Stratification by sample size showed a stronger association in larger-size (\>100 patients per study) studies (OR = 2.39, 95% CI = 1.88–3.05) than that in smaller-size studies (OR = 1.72, 95% CI = 1.04–2.84). Stratified analysis also revealed that the associations varied among the subjects from different regions. The association between cigarette smoking and *p16^INK4α* methylation tended to be stronger in 9 Asian studies (OR = 2.63, 95%CI = 1.92–3.61) compared to the 5 North American studies (OR = 1.62, 95%CI = 1.13–2.34). In addition, there was no inter-study heterogeneity in Asian studies (I² = 0, *P* = 0.709). Four of the 19 included studies, and another two excluded studies, had compared the frequencies of *p16^INK4α* methylation in adjacent noncancerous tissues or sputum specimens from NSCLC patients with or without smoking habits and no significant differences were found. In another 7 excluded studies, *p16^INK4α* methylation was examined in sputum, bronchial lavage samples or blood specimens from cancer-free subjects. We failed to find a link between cigarette smoking and the frequency of *p16^INK4α* methylation. These results suggested that cigarette smoking had no impact on *p16^INK4α* hypermethylation in the surrogate samples from NSCLC patients, and that the positive association between cigarette smoking and *p16^INK4α* hypermethylation was not present in health conditions. ## Publication bias To ensure the quality of this study, we performed a Begger's funnel plot and Egger's tests to eliminate the publication bias of included studies. As shown in, the shapes of the funnel plots showed a little asymmetry at the bottom. However, Egger's test, which provided statistical data of funnel plot symmetry, did not show any evidence of publication bias (t = 0, *P* = 0.998). # Discussion The present study, based on the accumulated evidences from 19 cross-sectional studies, indicates that cigarette smoking is positively related to *p16^INK4α* hypermethylation in tumor tissues from NSCLC patients. The frequency of *p16^INK4α* methylation in smoker lung cancer patients was 2.25 times higher than that in nonsmoker patients. The association appeared to be stronger in Asian patients and in studies with a larger number of subjects, but without a histology (AC or SCC) specificity. However, this positive correlation did not exist in adjacent noncancerous tissues from NSCLC patients and in biological specimens of ‘healthy’ subjects without cancer. Given the results that cigarette smoking leading to *p16^INK4α* hypermethylation was related to the stage progression of tumorigenesis, we speculated that *p16^INK4α* hypermethylation might be an early marker for cancer diagnosis, particularly in cigarette smoking patients. Recent epidemiological studies have revealed that molecular mechanisms underlying the development of lung cancers differed between nonsmokers and smokers. For instance, *EGFR* pathway is frequently activated by gene mutations in nonsmoker lung cancers, while mutations of *KRAS* often occur in smoker lung cancer. However, the mutations in either gene in lung adenocarcinomas are rarely seen although the biological consequences of KRAS and EGFR mutations share similarities in regulation of cell proliferation, survival and apoptosis. In the present study, we demonstrate that the frequency of *p16^INK4α* hypermethylation is slightly but significantly higher in smoker patients than that in nonsmoker patients. It is well known that *p16^INK4α* plays an essential role in the development of most human cancers for the reason that the p16/cyclinD1/CDK4/RB signaling pathway controls the cell cycle at the G1/S transition. Hypophosphorylated RB inhibits G1/S transition by binding to E2F1 transcription factor and exerts its tumor-suppressor function. Once hyperphosphorylated by the cyclinD1/CDK4 complex, RB releases E2F1, which results in transition from G1 to S phase. p16 prevents RB from phosphorylation by inhibition of CDK4, leading to a cell cycle arrest. Suppression of p16 expression allows unregulated phosphorylation of the RB protein and leads to uncontrolled cell cycle progression and cell division. *p16^INK4α* has a low frequency of mutations in lung cancer. Its inactivation is mainly through gene promoter hypermethylation. For example, Nakata et al. found that in tumors with *p16^INK4α* hypermethylation, 63.3% showed reduced expression; whereas, in tumors without *p16^INK4α* hypermethylation, only 33.7% showed reduced expression (*P* = 0.0002). The positive correlation between cigarette smoking and *p16^INK4α* hypermethylation demonstrates that cigarette smoking plays an important role in determining the molecular signatures involved in lung cancer development. The mechanism for cigarette smoking inducing gene-specific hypermethylation, e.g. *p16^INK4α*, remains unclear. *De novo* methylation uses S-adenosyl-methionine as a methyl donor and adds a methyl group to the cytosine ring to form methyl cytosine, which is catalyzed by DNA methyltransferases (DNMT) 1, 3a, or 3b. It is estimated that DNMT1 is responsible for about 90% of methyltransferase activity in mammalian cells. DNMT1 overexpression was found in many types of cancers including lung cancers, particularly in patients who were smokers. A recent study found that DNMT1 was highly expressed in tumor tissues in a dose response manner compared with the non neoplastic stroma tissues, not only in tobacco-specific carcinogen nicotine-derived nitrosamine ketone (NNK)-induced mouse lung cancer but also in human lung cancer associated with cigarette smoking. Moreover, it was demonstrated that NNK increased DNMT1 expression and activity by blocking its degradation related to ubiquitin- proteasome. AKT/GSK3β/βTrCP signaling is implicated in the accumulation of nuclear DNMT1, which leads to hypermethylation of *p16^INK4α*, fragile histidine triad gene (*FHIT*) and retinoic acid receptor β (*RARB*), and ultimately leads to tumorigenesis and poor prognosis. Although the direct interaction of DNMT1 to the *p16^INK4α* gene promoter is not yet characterized, these findings indicated that tobacco-induced DNMT1 overexpression might be responsible for maintaining the hypermethylation status of *p16^INK4α* gene. Lung cancer in nonsmokers is now a prominent public health concern. However, the major causes of them have yet not been identified. Environmental tobacco smoke (ETS), for example, second-hand smoke, has been recognized as a high risk factor. According to the report from International Agency for Research on Cancer (IARC), the risk for developing lung cancer from ETS exposure might reach 35% in men and 25% in women. Given that cigarette smoking has a cause-effect on *p16^INK4α* hypermethylation, ETS exposure may explain, at least partly, the variable percentage of *p16^INK4α* hypermethylation in nonsmoker patients. Other factors such as exposed to asbestos, chromium, arsenic, cadmium, silica, or nickel, or outdoor air pollutants, previous lung disease, and dietary factors have also been implicated in non smoking-related risk. But so far there is still a missing link between environmental factors, *p16^INK4α* hypermethylation, and lung cancer, which limits the use of gene-specific hypermethylation as a biomarker to detect lung cancer in early stage. Thus, the molecular mechanism underlying lung cancer, irrespective of tobacco-association, should be further elucidated. In this study, we observed that the frequency of *p16^INK4α* hypermethylation in NSCLC patients varied among different studies. The combined frequency in the present meta-analysis was less than 35%. The discrepancy and the relative low frequency might be due to the method used for detection of methylation, the variation in defining cigarette smokers, and insufficient information of clinical outcome. Although MSP is sufficiently sensitive, the conditions of PCR may affect the results to a large extent. The results seemed to be a little artificial particularly when PCR reaction was performed using both methylated and unmethylated primers. As for definition of cigarette smoker or nonsmoker, it lacked a consistent criterion followed by each investigation. In addition, current smokers and former smokers were not clearly distinguished, and the quantity of smoking was not calculated in the meta-analysis due to limited data. Moreover, insufficient clinical information such as the stage of NSCLC made it difficult to predict the prognosis based on the results provided. In conclusion, cigarette smoking is suggested to be positively related to *p16^INK4α* methylation in human NSCLC, highlighting the potential importance of *p16^INK4α* promoter methylation in early cancer diagnosis. Furthermore, it is well known that the risk for developing lung cancer in smokers is 8 to 13 times higher than that in nonsmokers, while the risk of *p16^INK4α* hypermethylation in lung cancer patients with smoking habits was only 2.2 times increased than that in nonsmoker patients, we speculate that many other aberrant epigenetic modifications, together with the genetic damage are involved in lung cancer development, which needs to be addressed in further investigation. # Supporting Information [^1]: Conceived and designed the experiments: W-QC WC. Performed the experiments: PY TL MJ Z-NH. Analyzed the data: BZ WZ S-XZ. Wrote the paper: BZ WZ Z-NH W-QC WC. [^2]: The authors have declared that no competing interests exist.

# I. Introduction Recent years have witnessed the rapid development of digital economy, which is new impetus to the economic transformation and upgrade. However, due to the differences between resource endowment and use efficiency, digital transformation objectively widens the digital divide. The digital divide is manifested in the disparities in Internet infrastructure access between different regions, the differences between the ability and frequency of individuals to use the Internet, and the differences in the returns brought about by digital technologies in different regions. Its existence further enlarges the endowment of information resources in different regions. As a factor of production, information continues to increase its participation in the economy, which makes the influence of digital divide on the income divide between regions keep growing. Recent literature rarely quantifies the digital divide and mostly analyzes the negative impact of the digital divide on society from a qualitative perspective. We discuss the influence and mechanism of digital divide on income divide in a theoretical and empirical way. Firstly, based on the practice of Xu Xiang et al., we introduce the production function including data capital, and compare the difference of growth rates between different types of capital under steady- state conditions, finding that data capital grows the fastest. The existence of digital divide makes the shares of data capital used by different regions vary greatly, which will eventually widen the income divide between regions by affecting the level of regional output. Secondly, relied on the panel data of prefecture-level cities from 2014 to 2018, we construct the digital divide index from three levels of "access divide", "use divide" and "efficiency divide", and utilize the two-stage spatial DID+IV model, so as to empirically test the impact of digital divide on income divide. The "Internet + government affairs" policy is introduced into the first-stage regression to construct the spatial DID model, and the exogenous digital divide index estimate is obtained, which is used as tool variable to deal with the endogenous problems. By introducing different mechanism variables from the three dimensions of "access divide", "use divide" and "efficiency divide" respectively, we verify the three ways that digital divide affects income distribution. The results of this paper remain robust under the regression of benchmark model, heterogeneity analysis of different regions, types of digital divide and robustness test. Firstly, the existence and expansion of digital divide will further widen the income divide between regions. Secondly, the digital divide in the eastern and coastal areas has more significant influence on income divide, and the new digital divide has more significant influence than the traditional one. Thirdly, the digital divide can affect income distribution through access divide, use divide and efficiency divide, resulting in asymmetric information, differentiation of human capital and sluggish upgrading of industrial structure. This paper is related to the following three literatures. The first is the study on digital divide, covering definition, measuring method and influence. The concept of digital divide was first proposed by Toffler and further enriched by Norris, Mossberger et al. and Van Deursen et al.. Digital divide has gradually expanded from simply divide in Internet infrastructure construction between regions to the income inequality brought by the Internet. Currently, the measurement methods of digital divide mainly include single-index measurement and multi-index measurement, the former one refers to the margin ratio method. The key of multi-index measurement lies in determining the weight to get the relevant indexes (Xiao and Wu,). Scholars have deeply analyzed the digital divide from the perspective of delaying the upgrading of industrial structure, influencing the relative sense of separation of micro-individuals and widening the income divide between urban and rural areas (Chen et al.,; Luo and Cha,; Peng et al.,; Zhang,; Kong et al.,). The second part is the study on income divide, containing measurement methods, development trends and influencing factors. The research on the measurement of income disparity began earlier. Sen et al. systematically expounded the conceptual framework and the practical problems of inequality measurement, and proposed an improved method of Gini coefficient to measure inequality. Shorrocks proposed generalized entropy index based on Thiel index. Scholars often used urban-rural income ratio, Gini coefficient and Thiel index to measure inter- regional income inequality (Zhao,; Chen and Li,). In addition, it is found that the income divide in China shows that: 1. In terms of distribution pattern, it shows typical spatial heterogeneity (Tian,; Li and Wang,). In terms of the development trend, the income divide continues to rise, but there is a convergence trend (Knight et al.,; Li and Zhu,). The existing researches mainly analyze the influencing factors of income divide at the macro level, such as democracy, financial development and globalization (Acemoglu et al.,; Jauch and Sebastian,; Dorn et al.,). The last part is the study on the influence of digital technology on income distribution. First of all, digital technology can improve the information asymmetry and reduce the income divide between regions by popularizing the Internet (He and Xu,). Secondly, digital technology can change the structure of factor allocation, Benzell et al. think that the digitization trend can reduce the marginal cost of labor and capital, but result in highly unequal rewards for talents lacking supply elasticity. In addition, Korinek et al. think that the use of artificial intelligence technology will lead to a further widening of the skill divide among workers, which further enlarges the inequality in income distribution. The possible marginal contribution in this paper lies in: First, we make exponential innovation to quantify the digital divide. We construct the digital divide index from three levels: "access divide", "use divide" and "efficiency divide", expanding the connotation of the existing indexes. While previous construction of indexes mostly focused on the differences in digital infrastructure, we focus on the "efficiency divide", or the difference in returns brought by digital technology. Secondly, we make method innovation to apply spatial DID + IV model. Based on the counter-fact frame, the exogenous digital divide index is obtained, which reduces the error of parameter estimation caused by endogeneity in the model. Thirdly, we make theoretical model innovation to introduce data capital-contained production function, and use the growth rate difference of different capital to reflect the widening effect of digital divide on income divide. The structure of this paper is as follows: The first part is the introduction; the second part is the theoretical modeling and mechanism analysis; the third part is the econometric model and variable explanation; the fourth part is the regression result and analysis; the fifth part is the conclusion and policy suggestion. # II. Theoretical mechanisms ## (I) Information asymmetry caused by access divides The access divide mainly refers to the difference in information infrastructure construction, which will hinder the normal information exchange between regions. Firstly, the access divide brings information cost effects. The region with complete information infrastructure construction has strong network effect, and the searching cost of digital commodity information is close to zero. However, the regions with low construction level are faced with information barriers, which makes it difficult to obtain effective information resources at low cost. The difference between information search and access costs will further widen the income divide between regions. Secondly, the access divide makes the supply and demand of information mismatched. The economic development of backward regions is slow and advanced knowledge and technology is required to be introduced to adjust the solidified industrial structure. However, the construction level of information infrastructure reduces the accessibility of information, and it is not easy for the backward regions to obtain accurate information suitable for local development level, which leads to the stagnation of their development level. While developed regions can receive abundant information resources from both domestic and foreign countries, resulting in information asymmetry among regions, which will widen the income divide between regions. Thirdly, the policy of promoting "new infrastructure" is mostly piloted in regions with relatively complete information infrastructure, while the policy, as a kind of institutional supply, will enlarge the existing infrastructure advantages of developed regions. Compared with traditional infrastructure, the core digital technology of "new infrastructure" is more extensive, innovative and obvious spillover effect, and the digital products formed by it have the characteristics of zero marginal cost or close to zero in production. The advanced regions will carry out digital reconstruction for the traditional infrastructure through "new infrastructure" to improve the multiplier effect of traditional infrastructure on economy, while the backward regions can’t enjoy the policy dividend, and it is quite hard to build complete information infrastructure, causing that construction divide of information infrastructure will eventually widen the income divide between regions. Additionally, featured by spatial spillover effect of new infrastructure policy, the advanced regions take the lead in building the new infrastructure construction, driving the development of the surrounding regions. However, due to the access divide, it is difficult for the surrounding regions to receive relevant knowledge and technology, which weakens the spillover effect of policies, and finally leads to the widening of the income divide between regions. ## (II) Effect of human capital differentiation caused by use divide The use divide mainly refers to the difference between people’s ability and frequency of using Internet, which will lead to obvious differentiation of human capital, one is digital human capital, and the other is traditional human capital. According to the new human capital theory, the ability factors (including cognitive and non-cognitive abilities) that are not easily replaced by technological factors are the key factors influencing income inequality. Firstly, compared with traditional human capital, digital human capital can promote laborers’ income through information channel effect, financing effect and social interaction effect, thus widening the income divide between the two kinds of human capital. Secondly, as for the inside of digital human capital, informatization and digitalization will produce technology-biased progress, bring polarization effect to labor market, increase relative demand for high-skilled labor force, and induce skill premium phenomenon. And the huge benefits brought by digital non-competition are highly concentrated in the hands of the executives and shareholders of the large digital enterprises, which belong to the high-skilled management level within the digital human capital. However, the workers of other executive levels can only share the surplus income together, thus divide is applied to generate a larger income divide within the digital human capital. Thirdly, Internet policy can indirectly influence regional innovation ability by promoting the inter-regional flow of innovation capital and innovators. Moreover, the developed regions can fully combine local advantages and policy dividends, which can produce huge siphon effect, further drain the innovative human capital in backward regions, and further widen the divide of innovation capacity between different regions. What’s more, innovation can widen the income divide through "productivity effect" mainly caused by that the developed regions can take advantage of innovation more effectively to improve local productivity, that is, the promotion on productivity of developed regions by innovation is greater. Besides, the implementation of Internet policy has spillover effect on the improvement of innovation level in surrounding regions. However, due to the existence of use divide, the surrounding regions may not possess the literacy and ability to accept relevant innovation knowledge, resulting in differentiation of innovation level among regions, and finally widening the income divide between regions. ## (III) Delayed effect of industrial upgrading brought by benefit divide The benefit divide mainly refers to the difference in returns brought by digital technology, which will slow down the industrial structure in backward regions and upgrade income divide between regions. Firstly, reduce industrial efficiency and delay the integration of informatization and industrialization. Hindering the integration of informatization and industrialization will not be conducive to the emergence of efficient technological means, mode of production and technological form, thereby reducing industrial efficiency. While the improvement of industrial efficiency is the main way to promote the upgrading of industrial structure, and the existence of benefit divide will delay the optimization and upgrading of industrial structure in backward regions, that not only enlarges the existing endowment difference, while boosting the probability of element mismatch, which will lead to the widening of income divide between regions. Secondly, the benefit divide is not conducive to the combination of industry and finance, and reduces the efficiency of capital allocation. The level of financial digitalization varies greatly in different regions. The traditional financial sector faces the dilemma such as "domain mismatch" and "scale mismatch". Local enterprises are still constrained on financing issues, which is not favorable to realizing the combination of industry and finance, debasing the speed of digital transformation and blocking the upgrading of industrial structure. Enterprises in developed regions can timely obtain sufficient funds for innovation activities such as R & D, so as to promote the upgrading of the overall industry, and thus obtaining higher profits. The ultimate benefit divide will widen the income divide between regions and enterprises by affecting the speed of financing. Thirdly, the policy for promoting the industrial digitalization transformation can advance the cross-border integration of industry, reconstruct the competition mode of industrial organization, and upgrade the industry power. Due to the fact that the industrial base in backward region is weak, so it is difficult to realize digital transformation in a short time by the assistance of policies, on the contrary, advanced regions can fully combine technology and policy dividend to steadily promote industry digitization which can bring huge multiplier benefit to economy, and become the main engine of driving economic development day by day. Differences in industry digitization levels between different regions will further widen the income divide. As an important part of digital economy, industrial digitalization has significant spatial spillover effect, which offers support in improving the total factor productivity of neighboring regions. As a result of the existence of benefit divide, it is difficult for neighboring regions to undertake digital industries in advanced regions, and the promotion of total factor productivity is limited, bringing about productivity differences among regions, and eventually inducing income inequality. # III. Empirical methodology ## (I) Model construction and data description ### 1. Model construction In order to verify the impact of the digital divide on income distribution, the following model is developed with the use of panel data from prefecture-level cities from 2014 to 2018: $$\mathit{theil}_{it} = \alpha_{0} + \beta_{1}gap_{it} + \beta_{2}\sum X_{it} + \eta_{i} + \xi_{it}$$ Where i represents region and t represents year; and *theil*_*it* represents the Thiel index for measuring income disparities between different regions. *gap*_*it* represents the digital divide index, *X*_*it* represents the control variables in the model, including the regional control variables: economic openness (*open*_*it*), urbanization level (*urbanize*_*it*), industrial level control variables: industrial structure deviation (*stru*\_*dev*_*it*), and individual level control variables: logarithm of per capita real GDP (ln*gdppc*_*it*), average years of education (*edu*_*it*), and dependency ratio with the elderly (*dependency*_*it*). Considering that the digital divide index is not an exogenous variable, and there is has a causal relationship between the digital divide index and income divide, we further establish a spatial DID+IV model to deal with the endogenous problem. Since the spillover effect between the control groups is not obvious, the spillover effect of the experimental group to the experimental group and the experimental group to the control group is considered only. In which *W*_*T*,*T* represents the indirect effect of the experimental group to the experimental group, and *W*_*NT*,*T* represents the indirect effect of the experimental group to the control group. Instrumental variables have both relevance and exogenous characteristics. On the aspect of relevance, the exogenous policy of "Internet + government affairs" which is closely related to the digital divide has been selected. By using the spatial DID model to carry on the regression and based on the counter-fact frame; it can be known the causal effect of exogenous policy on the digital divide, so the estimated digital divide index also has the exogenous characteristic, which can’t directly affect the income divide. The regression between the estimated value of digital divide index and income divide will result in a robust result, which reduces the deviations of model parameter estimation. <img src="info:doi/10.1371/journal.pone.0273334.e002" id="pone.0273334.e002g" /> g a p i t = α 0 \+ β 1 D I D i t \+ β 2 W T T D I D i t \+ β 3 W N T T D I D i t \+ β 4 ∑ X i t \+ η i \+ ξ i t <img src="info:doi/10.1371/journal.pone.0273334.e003" id="pone.0273334.e003g" /> t h e i l i t = α 0 \+ β 1 g a ^ p i t \+ β 2 ∑ X i t \+ η i \+ ξ i t ### 2. Data description We select the data of 280 prefecture-level cities in China from 2014 to 2018 as the research samples, and the data in the empirical study mainly come from *“China City Statistical Yearbook”*, *“Statistical Bulletin of National Economic Development”*, *“China Broadband Rate Download Report”*, RESSET Financial Research Database and EPS Database, etc. ## (II) Variable selection and index calculation ### 1. Exponent measure of digital divide This article chooses "access divide", "use divide" and "benefit divide" as the first-level indicators to construct digital divide index. The "access divide" refers to the inequality of Internet access, referring to *“China Digital Economic Development Index Report 2017”*, while Internet broadband user access rate and mobile phone penetration rate are selected to reflect the inequality of Internet infrastructure construction. The "use divide" refers to the inequality of Internet use capacity and frequency, referring to the construction of the measurement index systems of *“Digital Divide and Anti-Poverty Research”*, and the Internet penetration rate and broadband download rate are selected to reflect the divide in Internet use process. "Benefit divide" brings digital industrialization, industry digitization and digital government governance into the index system, reflecting the difference of returns brought by digital factors. Among them, digital industrialization includes the development level of information communication industry such as software and information technology service industry, the income of software and technical service, the income of software product, the number of information transmission and computer service employees, and the number of Taobao Village. The construction of digital government, an important part of digital governance, adopts the influence index measured by *"Influence of Government Affairs Weibo" and "Influence of Government Affairs WeChat"* published by People’s Daily to reflect the level of digital government governance. In the process of calculating industry digitization, refer to the measurement framework of China Digital Economic Development White Paper. Currently, the basic measurement method of industry digitalization widely used in academic circles is the perpetual inventory method (PIM), which uses the data of fixed assets investment to obtain the capital stock in different years, and then measures the industry digitalization level in different regions. ### 2. Income divides measurement The existing methods of measuring income divide mainly include Gini coefficient, Thiel index and urban-rural income ratio. The Gini coefficient is based on assumptions of different population distributions, and required to be derived using micro-individual income data. It is difficult to make panel data because the existing micro-database is investigated at intervals of several years. Therefore, we adopt the Thiel index and the urban-rural income ratio in the process of measuring the income divide. Thiel index applies the concept of entropy to measure income inequality, fully considers the population structure and income change, and complies with the current urban-rural dualistic structure, showing strong practical significance. The calculation method is as follows: $$theil_{it} = {\sum\limits_{j = 1}^{2}\left( \frac{Y_{ij,t}}{Y_{i,t}} \right)}\ln\left( {\frac{Y_{ij,t}}{Y_{i,t}}/\frac{N_{ij,t}}{N_{i,t}}} \right)$$ Where, *theil*_*it* refers to the Thiel index of region i and period t, reflecting the income divide between urban and rural regions. *Y*_*i*,*t* represents the total income of the cities and towns (villages) of Region i, *Y*_*i* represents the total income of the urban and rural regions of Region i. *N*_*ij* represents the total population of cities and towns (villages) of Region i, and *N*_*i* represents the total urban and rural population of Region i. ### 3. Intermediary variable Post and telecommunication service income (*post*): proxy variable as information asymmetry (Li,); In remote and poor regions, the transportation convenience is insufficient and the communication infrastructure is scarce, so the residents have insufficient ability to obtain information, receive education and medical care, and information development is weak. The foundation of the Internet lies in post and communication. It is necessary to improve the availability of post and telecommunications, so that more people may enjoy related services. Specific measures include increasing the distribution density of postal network points and improving the efficiency of the postal and telecommunications industry. The physical distance between increasing branches of post and telecommunications institutions and customers will be shortened, and the degree of information asymmetry will be alleviated. polarization effect of labor market: namely, the sum of proportion of high-skilled and low-skilled labor divided by the proportion of medium-skilled labor (*laborpolar*) as proxy variable of human capital structure; proportion of employed population of tertiary industry (*tratio*): proportion of employed population of tertiary industry in total labor force, as agent variable of industrial structure. ### 4. Control variables The definition and measurement of concerned control variables are as shown in. ### 5. Instrument variables It is generally believed that there is an endogenous problem between the digital divide and income distribution, and the digital divide will affect the income distribution. Additionally, the digital level of the regions with higher income level is high, while the digital development level of the regions with lower income level is slow, that is to say, there may be a causal relationship between the digital divide and income distribution. We will adopt the instrument variable method to solve the endogenous problem. Instrument variables must have the following two characteristics: relevance and externality. On the one hand, the instrument variable selected is closely related to the digital divide, on the other hand, it is irrelevant to the error term, that is, the instrument variable cannot directly affect income distribution. We selects the estimation term of X as instrument variable to be endogenously processed, therefore, we use the national policy "*Promoting the ‘Internet plus Government Services’ to carry out the pilot implementation plan of information benefiting the peopl*" issued by the National Development and Reform Commission of China in 2016 as an exogenous variable. The specific connotation of this policy is to select 80 pilot cities with both their urban and rural areas in China, building up the integrated online government service platform, realizing the standardization of government services by disclosing government service matters to all urban and rural residents within the jurisdiction. The policy is implemented at the prefecture-level city level, covering all urban and rural areas in 80 pilot cities, namely Shenzhen, Jinan, Yinchuan, Foshan, Weifang, Jiaxing, Suzhou, Yingtan, Benxi, Wuhu, Changchun, Urumqi, Guangzhou, Siping, Chengdu, Wenzhou, Weihai, Linfen, Xiamen, Fuyang, Guiyang, Dongguan, Fuzhou, Quanzhou, Chongqing, Dalian, Golmud, Xianyang, Luoyang, Daqing, Hangzhou, Xiaogan, Yiyang, Xi’an, Xinyu, Shenyang, Liaoyuan, Huaian, Qinhuangdao, Hohhot, Qingdao, Yuxi, Wuxi, Guilin, Shizuishan, Ningbo, Zhengzhou, Yuncheng, Yichang, Baiyin, Beijing, Wuzhou, Nanning, Shihezi, Shanghai, Jiyuan, Xiangtan, Shangrao, Neijiang, Chengde, Wuhai, Qitaihe, Lhasa, Tianjin, Lanzhou, Baoshan, Hefei, Shijiazhuang, Alar, Putian, Wuzhong, Karamay, Mianyang, Xiangyang, Wenshan, Changsha, Liaoyang, Yining, Harbin, Dunhuang. First, to bridge the access divide by building an Internet government affairs platform and alleviating the information asymmetry level of all urban and rural residents within the areas; Second, to reduce the use divide by increasing the inclusiveness of government services to reduce the degree of human capital differentiation; Third, to bridge the benefit divide by improving the intelligence of government services horizontally reduce the transaction costs of enterprises as market entities. Therefore, we selected 2016 as the event shock year, the urban and rural areas of pilot cities as the experimental group, and the urban and rural areas of non-pilot cities as the control group to establish a spatial DID model with two-stage regression. <img src="info:doi/10.1371/journal.pone.0273334.e007" id="pone.0273334.e007g" /> g a p i t = α \+ β 1 D I D i t \+ β 2 W T T D i t \+ β 3 W N T T D i t \+ β 4 e d u i t \+ β 5 s t r u \_ d e v i t \+ η i \+ ξ i t On the basis of policy externality, the spatial DID model constructs a counter- fact frame and the regression coefficient can reflect the causal effects of the policy of “"Internet + government affair”" on the digital divide. Therefore, the estimated value of digital divide index obtained is exogenous, and it can’t directly affect income divide. # IV. Benchmark regression results and analysis ## (I) Descriptive statistics The descriptive statistics has been shown in as follows: ## (II) Spatial DID model We use the policy of “*Internet plus Government Services*” which was proposed in 2016, so year 2016 was selected as the event impact year. Choosing pilot cities as experimental group and non-pilot cities as control group to establish the spatial DID model under bi-directional fixed effect, and this model can get the net effect of policy through the difference of two times, which can truly reflect the influence of policy. Traditional DID models assume that the potential outcome of any individual will not be affected by other individuals, which is called the Stable Unit Processing Value Hypothesis (STUTVA). However, in practice, the neighbor regions will also influence the local region, so using the traditional DID model may lead to error in parameter estimation. SDID will be more effective in estimating policy effects, so the model is further expanded and the specific expression of the model is as follows: $$gap_{it} = \phi_{i} + \theta_{t} + \mu\left( X_{it} \right) + \alpha DID_{it} + \beta_{1}W_{TT}DID_{it} + \beta_{2}W_{NTT}DID_{it} + \xi_{it}$$ In which, *W*_*T*,*T* represents the indirect effect of the experimental group to the experimental group, *W*_*NT*,*T* represents the indirect effect of the experimental group to the control group. *α* refers to the direct impact of Internet + government policy on the digital divide in this region, *β*₁ represents the spillover effect of experimental group to neighboring experimental groups, and *β*₂ represents the spillover effect of experimental group to neighboring control groups. ### 1. Parallel trend test Before the double difference model can be tested empirically, a parallel trend test must be carried out. The implication of this test is that the experimental group and the control group must have the same trend before and after the implementation of the policy. DID model allows for some differences between the control group and the processing group, provided that the difference does not change over time. The common trend test can be done either graphically or empirically. Considering the non-experimental period before 2016, the digital divide index is regressed to exogenous variable X, time trend variable, interaction term between time trend variable and group virtual variable as well as group virtual variable. The final result showed that the coefficient of interaction term was not significant, P value was 0.874, and the parallel trend test could be judged as qualified by combining graph and regression results. The parallel trend test shows that the policy of "Internet + government affair" is exogenous. ### 2. Results of model regression According to the regression results of SDID, the following conclusions can be drawn: ① DID coefficient is significantly negative, indicating that the direct spillover effect brought by Internet plus policy will narrow the digital divide between regions"Internet + government affairs" policy can accelerate the popularization of infrastructure and improve the overall level of digital government affairs. ② The coefficient of *W*_*NTT**DID* is significantly positive, indicating that the experimental group had a positive spillover effect on the adjacent control area. "Internet + government affairs" policy supports experimental group to vigorously build Internet government affairs platform, this kind of infrastructure is mostly used for the communication between vertical government systems, but less for communication and interaction with neighboring cities. The construction of government affairs platform has strong local characteristics, the spillover effect of knowledge and technology is not significant, the digital government affairs level of the experimental group and the adjacent control group cannot be improved in coordination, but the digital gap between them is widened. ③ The coefficient of *W*_*TT**DID* is also significantly positive, which indicates that the experimental group also has a positive spatial spillover effect on the adjacent experimental areas. The first situation is that there is a big difference in the maturity of digital government construction on the Internet platform between the two adjacent experimental groups, while the promotion effect of digital government level presented by the same policy is quite different. High-level areas can rely on the existing platforms to rapidly promote the construction of “Internet + government affairs”, while low-level areas need a certain period of time to promote and popularize facilities, finally widening the digital divide. The second situation is that the infrastructure and other aspects of the two neighboring experimental groups have little difference, but the local effect of "Internet + government affairs" policy is prominent, and it is difficult to coordinate the digital government construction synchronously. Therefore, the change of either party will widen the digital divide between the two regions. ④It can be found by comparing the three different effects that the direct effect of “Internet + government affairs” policy is the strongest, the second is the spillover effect between the experimental groups, and the last is the spillover effect of the experimental group to the control group. It shows that the influence of the policy of “Internet + government affairs” on the digital divide is transmitted through the direct effect path. Models (1)–(2) report mixed OLS and fixed-effect model regression results in turn. From the regression results, it can be seen that the digital divide is positively correlated with the income divide. In the mixed OLS model, every 1 unit increase in the digital divide will widen the income divide by 0.1251 units. In the fixed-effect model, every 1 unit increase in the digital divide will widen the income divide by 0.103 units. Considering that the digital divide index is not an exogenous variable, and there is a possibility of mutual causation between digital divide index and income divide, it is necessary to reduce the error of model parameter estimation through instrumental variable processing. Through the parallel trend test in the first stage regression of spatial DID + IV, the externality of “Internet + government affairs” policy is illustrated. The estimated value of digital divide index based on the counterfactual framework constructed by spatial DID is also exogenous. What’s more, this estimate value cannot directly affect the income disparity between regions, and satisfies the two conditions of correlation and externality of instrumental variables. Hence, the digital divide index estimated by spatial DID regression is selected as instrumental variable in this paper. Model (3) substitutes the estimated value of the digital divide index obtained from the spatial DID model into the regression, and presents the two-stage regression result of spatial DID + IV. Every 1 unit increase of digital divide will widen the income divide by 0.134 units. After adding the control variables, the F value of the first stage is 20.38, greater than 10, which is in accordance with the rule of thumb, indicating that there is no problem of weak instrumental variables. Then the over-identification test of the externality of the instrument variable was carried out, and the p value was 0.697. It is considered that the selected instrumental variable is exogenous and not related to the disturbance term, which further verifies the rationality of the instrumental variable. After dealing with endogenous issues, the conclusion remains that the digital divide widens income disparities. Among the control variables, there is a negative correlation between the average length of education and the income divide. The improvement of education level helps to upgrade the gradient of human capital, improves the efficiency of labor resource allocation and promotes the homogeneous convergence of income divide. The improvement of urbanization level narrows the income divide, and the steady progress of urbanization facilitates the free flow of factors, promotes the rural labor force to transfer from the low-return sector to the high-return sector and obtain structural dividends, thus narrowing the income divide between urban and rural areas. There is a positive correlation between regional opening level and income divide. China’s position in the division of labor system of global value chain has moved up gradually from a large trading country to a trading power. The improvement of regional openness puts forward higher requirements for human capital, reduces the marginal efficiency of labor force to a certain extent, and widens the wage divide of employees with different skill levels. # V. Further analysis ## (I) Robustness test Benchmark regression leads to the conclusion that the digital divide will widen income divide. In order to test the robustness of the above results and further demonstrate the impact of the digital divide on the income divide, the following three robustness tests will be carried out: First, change the measurement method of dependent variables, and replace the explanatory variables in simple OLS regression with urban-rural income ratio to verify again. The results of model (1) show that the digital divide still has a significant effect on the income divide after changing the indicator of income divide to urban-rural income ratio. Secondly, 5% of the samples with the highest and lowest digital divide index, i.e., Beijing, Guangzhou, Shanghai, Dongguan City, Hangzhou City, Foshan City, Suzhou City, Zhuhai City, Tongren City, Haozhou City, Qinzhou City, Dazhou City, Yulin City, Yongzhou City, Haidong City and Luohe City were removed, and then the robustness test was carried out. Model (2) presents the regression results of this method and it is found that the conclusion that the digital divide widens the income divide is still valid. Concurrently, different measurement methods are used to calculate the digital divide index, and the entropy weight method is replaced by the main component analysis method. Model (3) reports a regression report using principal component analysis to measure the digital divide index, which increases the income divide by 0.323 units for every 1 unit increase of the digital divide. Through the above three robustness tests, it is seen that the conclusion obtained in the benchmark regression model is valid, showing that the conclusion that the digital divide widens the income divide is robust. ### (II) Heterogeneity analysis China’s digital divide presents obvious spatial differentiation, we first divide the samples into groups based on regional distribution. From the results of heterogeneity analysis, it is demonstrated that the influence of digital divide on income divide is significant in the central and eastern regions, but not in the western regions, it is mainly because the form of digital divide changes with the development of digital economy, from traditional "access divide" gradually evolved into the "utilization divide" and "benefit divide". Due to the low digital level in the western region as a whole, the construction of digital infrastructure is slow. The income divide in different regions mainly comes from traditional factors of production, such as capital and labor, and the application of information technology belongs to a kind of skill-biased progress, so the influence of digital divide on income divide is not obvious in western region. The influence of digital divide on income divide is the most significant in eastern China because of the early start of digital economy and obvious difference in digital technology returns. According to the maritime standard HYT094-2018 code of coastal administrative area, the whole country is divided into coastal areas and inland regions, and the sub-sample regression is carried out. The regression results express that the digital divide in coastal regions has a significant impact on the income divide. An increase in the digital divide by 1 unit increases the income divide by 0.284 units. By virtue of the advantage of natural geographical location, coastal areas attract sufficient funds and advanced technology, and the digital economy develops rapidly. While developing, the difference of digital dividend returns is obvious, so the digital divide has a high impact on income distribution. However, due to the restriction of transportation and other conditions, the development of digital economy is at a relatively low level, and the influence of digital divide on income divide is not obvious. We divide the digital divide into "access divide", "use divide" and "benefit divide" in the process of index construction. The former two focus on the construction of digital infrastructure and are therefore unified into the traditional digital divide. The "benefit divide" is defined as a new digital divide. According to the regression results, the impact of the new digital divide on income distribution is higher than that of the traditional digital divide. Every 1 unit increase in the new digital divide will widen the income divide by 0.170 units. The spread of mobile technology gradually bridges the traditional digital divide, while the new digital divide has been widened due to the difference between the growth and iteration effects of the digital economy. The digital divide is manifested more prominently by the difference in economic benefits brought by the Internet, so the new digital divide has a more significant impact on the income divide. ## (III) Mechanism verification The digital divide will widen the income divide of Chinese residents. For the purpose of better explaining the relationship between the two, the transmission mechanism behind it will be further discussed from three perspectives of information asymmetry human capital differentiation and industrial structure upgrading. In order to test the information asymmetry effect caused by the access divide, we introduce the interaction between post and telecommunication service income (post) and digital divide index. The first column of the regression results reports the results of the model, and the interaction between the digital divide index and post and telecommunication revenue is significantly positive, expressing that narrowing the digital divide can perfect income distribution by reducing information asymmetry. In regions with large access divide, the phenomenon of information asymmetry is obvious. In the regions with low digital development level, the low availability and high search cost of information resources restrict the improvement of local income level. In order to test the effect of the human capital differentiation caused by using the divide, we introduce the interaction term between laborpolar variable and digital divide index. The third column of regression results reports the results of the model, where the interaction between the digital divide index and the share of the tertiary labor force is significantly positive. The labor market polarization and human capital differentiation brought by informatization and digitization will further widen the income divide between individuals. In order to test the delayed effect of industrial upgrading caused by the benefit divide, we introduce the interaction term of proportion of labor in tertiary industry (tratio) and digital divide index in the tertiary industry. The third column of the regression results reports the results of the model, where the interaction between the digital divide index and the share of the tertiary sector labor force is significantly negative, showing that by narrowing the digital divide, the optimization and upgrading of industrial structure can be promoted, thus reducing the income divide. The expansion of the digital divide will slow down the speed of industrial diffusion, which is mainly due to the high threshold of digital industry, which makes it difficult to spread to areas with low technological level. Also, it is not conducive to the free movement of labor force between different regions and industries. Some studies show that the application of information technology increases the income share of capital and technology and reduces the income share of labor, while the income of the people in the backward areas is mainly derived from labor, which exacerbates the income inequality between regions. Narrowing the digital divide can improve income distribution through two channels. The first is factor allocation effect, which can promote labor transfer to regions or industries with high rate of return, improve the skill level of laborers and reduce income divide. The second is knowledge spillover effect. The existence of digital divide affects the dissemination of knowledge technology and creates an invisible barrier. After breaking the barrier, the diffusion of digital knowledge and technology can be accelerated so that the development of regional integration can be realized. # VI. Conclusions This paper uses the panel data of 280 prefecture-level cities in China from 2014 to 2018 to explore the relationship between digital divide and income gap. Different from the previous studies, we focus on the benefit divide. In addition, it focuses on the mechanism and heterogeneity of income distribution influenced by the digital divide. It is found that: (1) The existence of digital divide widens the income divide between regions, and each increase in the digital divide by 1 unit widens 0.134 units. (2) The influence of "Internet + government affairs" policy on the digital divide has both direct and indirect effects, and the local effect of policy implementation on the digital divide is the most obvious. (3) Information asymmetry, differentiation of human capital and upgrading of industrial structure have intermediary effects, which play an important role among access, use and benefit divide respectively. (4) After considering the regional differences, it is learned that the digital divide in the eastern coastal areas has more significant impact on the income divide. After considering the differences of digital divide at different levels, it is found that the new digital divide has the most obvious effect on widening the income divide. Considering that the digital divide can affect income distribution through three channels: access, use and benefit divide, we put forward some suggestions for each mechanism. First of all, access divide will bring information asymmetry effect, so it is necessary to improve infrastructure construction to reduce information asymmetry effect. Digital divide in essence, is a kind of information divide. The key to bridging the digital divide is to broaden the availability of information, widen the channel of information transmission and optimize the information communication environment. Secondly, it is also necessary to strengthen the construction of digital access capacity in backward areas through new infrastructure construction, promote the development of information infrastructure towards inclusion, and empower the indexation growth of economy. Thirdly, access divide will bring about the effect of differentiation of human capital, so it is necessary to elevate the digital degree of human capital, perfect the non-digital policy system, and consummate the human capital from "quantity" to "quality" by the influence of digital technology spillover in the backward area, and lowerdown the income inequality between digital-intensive area and non-digital-intensive area through technology complementation. Moreover, the improvement of non-digital policy embodiment can reduce the damage brought by high monopoly to the fair competition market environment. Finally, the benefit divide will bring the sluggish effect of industrial upgrading, so the suggestion of adjusting industrial structure and releasing "digital" dividend is put forward. We will establish a mechanism for inter-regional talent flow, enhance the spillover effect of knowledge and technology, and coordinate inter-regional innovation capacity. Meanwhile, we will actively promote the deep integration of digital economy and real economy based on Internet of Things, 5G, artificial intelligence and other new technologies, accelerate the combination of data elements and traditional production factors, exert the resultant force of industrial digitalization and digital industrialization to boost the rationalization and upgrading of industrial structure and build a new development pattern. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Reconstruction of critical-size bone deficiencies remains a major challenge in orthopedics. The bone tissue engineering technique provides a new approach to this problem. The seeding and subsequent *in vitro* culture fundamentally affects the osteogenic activity of tissue-engineered bone grafts because they determine the initial density and spatial distribution of seeded cells in the scaffold as well as their subsequent behaviors (e.g. proliferation, differentiation, migration). Many factors can affect the efficiency of seeding and the outcome of the subsequent *in vitro* culture, including in the technique employed for seeding and the hydrodynamic condition provided for subsequent regeneration. Currently, cells are seeded primarily by static or hydrodynamic methods. In the static method, a suspension containing seeded cells is dispensed on a scaffold, followed by a period of rest to allow the cells to enter the scaffold. With this method, the initial cell density (the number of cells which attached in 3D scaffold when tissue engineering bone were preparation and without culturing in vivo or in vitro) in the scaffold can be increased by increasing the cell concentration of the suspension within a certain range, though at the expense of seeding efficiency (i.e. the percentage of cells that entered the scaffold), but cannot be further increased beyond a plateau level. In comparison, in the hydrodynamic seeding method, cells are allowed to adhere to the scaffold in a dynamic fluid flow created by a bioreactor. With this method, the cell agglomeration accelerates with the cell density in the seeding suspension, thus facilitating the adherence of cells to the scaffold, increasing the speed and density of cell seeding, and improving the spatial distribution of cells in the scaffold. In addition to seeding, hydrodynamic conditions can also substantially affect the subsequent *in vitro* culture of cell-scaffold constructs. A dynamic fluid flow was found to positively affect the behavior of seeded cells, such as proliferation, differentiation, and migration. However, dynamic fluid flow may also result in cell detachment and shear-induced damage, and thus, loss in cell utilization. A number of studies have separately exploited the advantages associated with a higher initial cell density or hydrodynamic culture. Zhao *et al* increased the initial density of human umbilical cord mesenchymal stem seeded cells in injectable bone tissue engineering constructs by using hydrogel microbeads. Ericka *et al* seeded chondrocytes onto polyglycolid acid scaffolds under hydrodynamic conditions, and obtained intermediate initial cell densities and sustained subsequent proliferation. The optimal tissue engineering technique should combine methods to increase the initial cell density and create an appropriate hydrodynamic environment to accelerate the *in vitro* maturation of the cell-scaffold constructs into clinically applicable grafts. Here, we investigate whether a combination of fibrin glue-assisted seeding and hydrodynamic culture in rotating wall vessel bioreactor can substantially improve the seeding efficiency and subsequent proliferation and osteoblastic differentiation. We further determined if these improvements translated into enhanced osteogenic activity in a nude mice subcutaneous implantation model. This study aims to understand the effects of the key factors of tissue engineering preparation methods, including initial cell density and hydrodynamic culture methods, in an attempt to provide experimental basis for improvement the osteogenesis performance of bone tissue engineering. # Materials and Methods ## Ethics statement Nude mice (6 weeks old) were purchased from the Laboratory Animal Center of our university. The animal experiment was approved by the ethics committee of Third Military Medical University and conducted in conformity with the ‘Guiding Principles for Research Involving Animals and Human Beings’ as adopted by The American Physiological Society. ## Isolation and characterization of hMSCs Human mesenchymal stem cells (hMSCs) derived from bone marrow of the iliac crests of young healthy volunteers were provided from Tissue Engineering Research and Development Center of The Third Military Medical University. hMSCs were isolated by density gradient centrifugation in Percoll (density = 1.077 g/mL; Sigma, USA) and adherent culture. The cells were expanded by in vitro culture and confirmed to be hMSCs by a previously published method. To induce osteogenic differentiation, the hMSCs were cultured in DMEM/F12 medium supplemented with 10% FBS, 0.1 µM dexamethasone, 10 mM glycerophophate, and 50 µM ascorbic acid. After 12 days, the induced cells were detected by immunochemistry using antibodies (1∶100) for ALP, osteocalcin and collagen type I (Sigma, USA) according to product instructions. At the 12th day after osteoinduction, the ALP activity of induced hMSCs was measured and the osteocalcin concentration in the culture medium of induced hMSCs was detected by an enzyme-linked immunosorbent assay (ELISA) kit (BioSino, Beijing, China). ## Scaffold preparation Cubes of human demineralized cancellous bone matrix (DBM, 4 mm×4 mm×4 mm) were obtained from the tissue bank of our university and used as the scaffolds in this study. The porosity of DBM is 70% and the pore size is 300–800 µm. The DBM was prepared by a series process, as previously described by Tan et al.. ## Construction of implants and Grouping Four kinds of bone substitutes were constructed based on different approach of seeding and culture. Hydrodynamic seeding and hydrodynamic culture (group A): Fifty DBM scaffolds and 5.0×10⁷ MSCs were added into the high-aspect ratio vessel of a rotary cell culture system (Synthecon RCCS-1, Houston, TX, USA). The vessel was filled with 50 ml of DME/F12 culture medium (Hyclone, Logan, UT, USA) and degassed. The rotation speed was adjusted daily (18–24 rpm) to ensure that the rotating trajectories of the scaffolds would not collide with the vessel wall or converge to the center. The rotary culture system was incubated in an atmosphere of 5% CO₂ and 100% relative humidity at 37°C, with daily adjustment of rotation speed and a change of medium every 48 h. Hydrogel-assisted seeding and hydrodynamic culture (group B): The fibrin glue (25 mg/ml, Tissucol, Baxter, Austria) was prepared by mixing fibrinogen and thrombin within 90 min under sterile condition. The hMSCs were mixed with a fibrin glue to form a suspension containing 2×10⁷ cells/ml. The suspension was added dropwise onto the DBM scaffolds (0.05 ml/scaffold). Then 0.05 ml of catalytic agent was sprayed on each scaffold, and was allowed to stand for 5 min for the glue to gelate. Then the scaffolds were added into the high-aspect ratio vessel of a rotary cell culture system, and were cultured with the same condition of group A. Static seeding and static culture (group C, control group): The hMSCs were dissociated in DMEM/F12 medium into a suspension containing 2×10⁷ cells/ml and added dropwise onto the DBM scaffolds (0.05 ml/scaffold). The seeded scaffolds were statically cultured in 24-well plates without additional medium for 2 h, then turned over to reduce the downward flow of liquid and the cell loss, and cultured in the same conditions for another 2 h. Then, each well was filled with 1 ml DMEM/F12 and incubated in an atmosphere of 5% CO² and 100% relative humidity at 37°C, with a change of medium every 48 h. Hydrogel-assisted seeding and static culture (group D): DMB scaffolds were seeded by the same hydrogel-assisted method as in the group B and were cultured statically in 24-well plates (Corning 3524, USA) containing DME/F12 medium (1 ml/well) and incubated in an atmosphere of 5% CO₂ and 100% relative humidity at 37°C, with a change of medium every 48 h. ## Evaluation of seeding efficiency Twenty four hours after seeding, the seeding efficiency of each group was analyzed, which was defined as the ratio of the number of cells existing in the scaffold to the number of cells added originally to the scaffold. After transferring bone substitutes to other wells, non-adherent cells in the well were collected by rinsing repeatedly with DMEM/F12 medium. Then, the cells attached to well bottom were digested with 0.25% trypsin plus 0.01% EDTA and collected. The cells in the supernatant and those adhering to the bottom of the well were separately counted by hemocytometer. The sum of these two portion of cells was recorded as ‘remaining cell number’ in each well. The seeding efficiency was calculated by: (initial cell number-remaining cell number)/initial cell number. ## Cell viability The viabilities of cells in scaffolds were assayed at various time points (8 h, 16 h, 24 h, 48 h, and 3–14 d after seeding). The cell-scaffold constructs were removed from their medium, rinsed with phosphate buffered saline (PBS), and placed in a 96-well culture plate. Cell viability was determined as reported. Cell counting kit-8 (CCK-8, 20 µL/well, Dojindo Chemical Institute, Kumamoto, Japan) was added to each well, followed by further culture for 3 h (5% CO 2, 37°C, 100% relative humidity). Then, the constructs were removed, and the optical density of each well at 450 nm was measured with an ELISA reader (reference wavelength: 655 nm), with cell-free DMB scaffolds as the controls. ## ALP activity The ALP activities were measured at various time points (2, 4, 6, 8, 10, 12, 14 and 16 d) after seeding. The cell-scaffold constructs were rinsed twice with PBS and then lysed with 0.2% Triton X-100 (Sigma, USA). The lysate was centrifuged at 600 g for 5 min and the supernatant was collected and incubated for 15 min (5% CO₂, 37°C, 100% relative humidity). The absorbance at 405 nm was measured on a microplate reader and converted into the ALP activity against a standard curve, which was established based on the reaction of 10 ml of a p-nitrophenyl solution (Wako) and 200 ml of substrate buffer for 30 min. ALP activities were expressed as U values. ## Scanning electron microscopy (SEM) All specimens per group were was treated by a series of procedures for SEM observation after 14-day culture, including incubating in 3% (w/v) glutaraldehyde solution for 48 h at 4°C, washing over-night with 0.1 M PBS solution, decalcifying with 0.5 mol/L EDTA for 7 days, washing over-night with 0.1 M PBS solution, incubating in 2.5% glutaraldehyde solution for 48 h at 4°C, and post-fixing and staining with 1% osmium tetroxide for 2 h at 4°C, dehydrating in a graded series of ethanol, coating with an ultrathin gold layer. Then the specimens are observed under SEM (1000-B, AMRAY, Bedford, MA, USA) to assess investigate morphological features of the attached cells and ECM fibers formed on scaffolds. ## In vivo osteogenesis The *in vivo* osteogenetic activities of scaffolds and cell-scaffold constructs were evaluated with a subcutaneous implantation model in 24 nude mice. Each mouse received four implants in its back, termed implants I–IV. Implant I was a cell-free DBM scaffold and placed at the left rostral position. Implant II was a cell-scaffold construct seeded with 2×10⁷/ml hMSCs by the hydrogel- assisted method followed by dynamic culture for 12 d; this was placed at the right rostral position. Implant III was a construct seeded with 1×10⁸/ml hMSCs by the hydrogel-assisted method; it was placed at the left caudal position immediately after seeding without *in vitro* culture. Our pretest found that during the 12-day dynamic culture (as in Implant II), cell proliferation in the scaffold increased by 10 fold (data not shown). Therefore, to ensure an equal cell density before *in vivo* implantation, the initial cell density for implant III was 9 times greater than implant II. Implant IV was a construct seeded with 2×10⁷/ml hMSCs by the hydrogel-assisted method followed by static flask culture for 12 d; it was placed at the right caudal position. The mice were killed by decapitation 4, 8, and 12 weeks after implantation (n = 8, for each time point). Each mouse was placed in a supine position and examined by X-ray radiography. The implants were retrieved, stripped of soft tissues, and weighted wet. They were also analyzed by dual-energy X-ray absorptiometry (Challenge, DMS, Montpellier, France) for bone mineral densities. ## Histological analysis Histological analysis of the implants was undertaken at 12 weeks after surgery. Specimens from the bone defect sites were fixed with 4% paraformalclehyde for 48 hours, decalcified with 0.5 mol/L EDTA for 2 weeks, dehydrated with gradient ethanol solutions for 2 days, vitrified with dimethylbenzene, embedded in paraffin, and cut to yield 6 µm thick sections. The sections were stained with haematoxylin and eosin (H&E) for histological evaluation and examined under light microscope. ## Statistical analyses Data were expressed as mean ± standard deviation. Data were analyzed by one-way analyses of variance (ANOVA) and Student–Newman–Keuls (SNK) post hoc tests using SPSS 12.0 (SPSS, Chicago, IL, USA). A p-value of less than 0.05 was considered statistically significant. # Results ## Cell culture and characterization The hMSCs tended to form calcium nodus after 12 d conditional culture. The hMSCs also stained immunohistochemically positive for ALP, osteocalcin and collagen type I after 12-day osteogenic induction. The ALP activity of hMSCs and osteocalcin concentration in the culture medium were significantly higher in induced group than that in control group (P\<0. 01). ## Microscopy of cell-scaffold constructs In group A (dynamic seeding followed by dynamic culture), after culture for 8 h, a small number of cells were observed on the surface of the pores in the scaffolds. Then, the cells gradually increased in number and became uniformly distributed on the surface of the pores. After culturing for 5 days, the cells started to produce ECM, the cells and ECM were both uniformly distributed. In group B (hydrogel-assisted seeding followed by static culture), the cell-laden gel filled most pores in scaffold after seeding immediately and the resulting constructs remained almost unchanged during subsequent culture. In group C (static seeding followed by static culture, control group), the seeding suspension rapidly penetrated the DBM scaffolds after seeding and reached the bottom of the wells. Two hours after seeding, a large number of spindle-like cells appeared at the well bottom. After 5 days of subsequent culture, the cells in the scaffolds started to produce extracellular matrix (ECM) resembling spider webs. The cells and ECM were distributed non-uniformly at this time point. In group D (hydrogel-assisted seeding followed by dynamic culture), the seeded cells were carried by the fibrin gel and filled most pores in the scaffolds after seeding. The cells were identified by their low refractivity. A small amount of fibrin gel was found on the bottom of the wells. The cell number increased with culture time, and was higher than group C at the same time points. The number of attached cells and density of ECM fibers in the interior of the scaffold after 14-day culture are significantly different among four groups. The number of attached cells and density of ECM fibers in group B is most and the cells and ECM were uniformly distributed in the scaffold. The cells in the scaffolds and ECM fibers presented in group C are minimal and non-uniformly scattered in the pores of the scaffolds. Group D had more cells attached to the pore and ECM than group A. ## Seeding efficiency The seeding efficiency was 32.10±0.72% in group A and 25.24±1.56% in group C, compared with 82.14±1.09% in group B and 81.53±1.37% in group D. Groups B and D were similar (p = 0.396), and the differences between all other group pairs were highly significant (all p\<0.01). ## Cell proliferation and osteoblastic differentiation In group A, the cell number remained stable during the first 3 days, followed by a continuous increase till day 14. In group B, the cell number was stable during the first 2 days, then started to increase, and attained a plateau on day 12. In group C (control group), the cell number remained stable during the first 4 days in culture, then started to increase, and plateaued on day 8. In comparison, the cell number in group D decreased slightly between 8–24 h in culture, then decreased more rapidly between days 1–8, and remained stable thereafter. The ALP activities in all groups increased from day 2 to day 6. The activities in groups A and B remained stable thereafter. In comparison, the activities in groups C and D continued to increase, although at lower levels and slopes. ## SEM of cell-scaffold constructs SEM revealed similar results to optical microscopy. The number of cells and density of ECM fibers in group B is most, then group D, followed by group A and group C. ## X-ray radiography The radiographic densities of all implants increased from week 4 to week 12. Implant II(hydrogel-assisted seeding of 2×10⁷/ml hMSCs, followed by dynamic culture for 12 d) showed substantially higher density than the other implants, and implant I (cell-free DBM scaffold) had the lowest densities at both time points. At week 12, implant I showed a slightly higher density compared with the host soft tissue, while implant II clearly showed increased density indicating calcification. The implants III (hydrogel-assisted seeding of 1×10⁸/ml hMSCs without further *in vitro* culture) and IV (hydrogel- assisted seeding 2×10⁷/ml MSCs followed by static culture for 12 d) also showed signs of calcification, but substantially weaker than that in implant II. ## Wet weight and bone mineral density Twelve weeks after implantation, implant I showed a significantly lower wet weight compared with other implants (all p\<0.01). Moreover, the wet weight of implant II was statistically higher than the implants III (p = 0.008) and D (p = 0.004). Implants III and IV were similar (p = 0.770). Twelve weeks after implantation, implant II showed a significantly higher bone mineral density than all other implants (all p\<0.01). The bone mineral density of implant I was significantly lower than the other implants (all p\<0.01). Implants III and IV were similar (p = 0.741). ## Histology of retrieved implants Twelve weeks after implantation, implant I showed partial degradation of DBM scaffold and replacement by fibrous connective tissues around the periphery. Implant II showed relatively mature bone trabeculae but no chondroid tissues. Implant III showed less mature bone trabeculae than implant II, in addition to chondroid structures in a few locations. Implant IV showed new bone trabeculae that were less mature than those formed in implants II and III; transformation of chondroid tissue to immature bony tissue was also locally observed. # Discussion In the present study, we evaluated the effects of seeding methods on seeding efficiency and initial cell density for constructing tissue-engineered bone. Compared with other synthetic bone substitutes, tissue-engineered grafts generally have superior osteogenic activities because of the incorporation of seeded cells. Various factors can influence the osteoblastic differentiation of marrow stromal cells in tissue engineering scaffolds during cultivation, including the density and spatial distribution of the seeded cells in the scaffolds. Seeded cells are commonly seeded in scaffolds by static infiltration. Although convenient, this method can attain only limited cell density. Under the action of gravity force, the seeded cells are easily detached from the scaffold and became concentrated at its bottom side, thus resulting in loss of cells. Various methods have been used to promote cell penetration and minimize cell detachment, such as the use of negative pressure and magnetic field. Although effective to varying degrees, these methods cannot substantially increase the initial cell density in the scaffold. Recent studies found that RWVBs can produce a simulated microgravity environment to allow cells to diffuse and become uniformly distributed in the interior of scaffolds. Hydrogels have been combined with seeded cells to construct grafts for the repair of cartilage as well as bone. Hydorgels alone, however, are not satisfactory for constructing bone grafts because of their poor strength and limited bone conductivity. Despite this major disadvantage, hydrogels may improve the adhesion between seeded cells and the scaffold. In this study, we compared the seeding efficiency and initial cell density resulting from three seeding methods: fibrin hydrogel-assisted seeding, hydrodynamic seeding (simulated microgravity in RWVB), and the simple static infiltration. Microscopy, cell counting, and viability assays showed that fibrin hydrogel-assisted seeding generated a significantly higher seeding efficiency and initial cell density than the other two methods. The improvement can increase the utilization of seeded cells and is expected to increase the osteogenic activity of the resulting grafts. Fibrin glue has been clinically confirmed to be safe, biocompatible, and fully absorbable within two weeks. A recent clinical study used fibrin as a carrier for chondrocytes to treat cartilage defects and obtained positive results. The fibrin glue used in this study was a mixture of fibrinogen, thrombin, factor XIII, and calcium salt. Fibrinogen is a major plasma protein (350 kDa) that stimulates proliferative signals by serving as a scaffold to support the binding of growth factors and to promote the cellular responses of adhesion, proliferation, and migration during wound healing. Thrombin is an enzyme that converts soluble fibrinogen into insoluble fibrin between 10 and 60 seconds and acts as a tissue adhesive. Factor XIII, which exists in the fibrinogen component of the glue, cross links and stabilises the clot's fibrin monomers. These glue contents in mixture formed an efficient cross-linking network that could capture MSCs rapidly and promote the cell attachment and proliferation. Therefore, higher seeding efficiency was obtained in fibrin hydrogel-assisted seeding groups. We further identified the effect of hydrodynamic culture on cell proliferation and differentiation *in vitro*. There is still no consensus on whether tissue- engineered bone grafts need to be cultured *in vitro* before implantation. Many studies have suggested that *in vitro* culture can allow the seeded cells to stably adhere on the scaffold and, thereby, prevent their detachment, migration, or death resulting from changes of microenvironment. Wang et al, however, suggested that the *in vivo* condition should be optimal for the growth, differentiation, and function of cells. In contrast, *in vitro* cultured constructs may be structurally unstable, mechanically weak, and subject to changes in tissue structure and type. In an attempt to combine the advantages of pre-implantation culture and *in vivo* microenvironment, some studies also explored ectopic implantation to engineer mature, vascularized bone grafts. These “*in vivo* engineered” grafts were found to have superior osteogenic activities, but the technique involves a long *in vivo* culture and additional damage to the patient. Recent development of bioreactor techniques has made it possible to better simulate the *in vivo* microenvironment, promote mass exchange, and create appropriate mechanical stimuli. These improvements may be used to produce more mature and bioactive tissue-engineered grafts. In tissue engineering of grafts, the supply of nutrients and removal of metabolic wastes is more difficult than in conventional cell culture. The mass transport in the common static culture method relies on the concentration gradient and is thus inefficient. As a result, cells typically do not survive well in the center of the graft and in some cases even undergo necrosis to form voids. This has severely limited the size of grafts that can be obtained by tissue engineering. An appropriately designed bioreactor may provide hydrodynamic conditions to promote mass transfer, stimulate stem cells to differentiate into osteoblasts, and thus overcome this disadvantage. In this study, we found that when comparing static and hydrogel-assisted seeding, the statically cultured cell-scaffold constructs achieved lower plateau values. In comparison, regardless of the initial cell densities, the dynamically cultured constructs showed continued increase in cell density and became approximately two times higher than the statically cultured grafts. Furthermore, with a higher seeding efficiency and cell density by the hydrogel-assisted seeding, group B achieved plateau earlier than the group A. The ALP activities of the constructs followed the order of: group B\>group A\>group D\>group C, consistent with the trend of cell number between days 6–14. These findings suggest that hydrogel-assisted seeding followed by hydrodynamic culture can substantially increase the initial seed cell density in constructs, achieve a higher cell density earlier than static culture, and is the optimal one among the four methods studied here. The favourable effect of hydrodynamic culture may be attributed to three factors. First, the vortex in the bioreactor generated fluid flow in the construct, which enhanced mass transfer and improved the cell distribution. A computational analysis suggested that sufficient flow fluid can be generated in porous scaffolds despite being partially sealed with a material similar to fibrin. Second, the shear stress resulting from the fluid flow may have simulated the seeded cells to differentiate, mature, produce extracellular matrix, and calcify. Third, the hydrodynamic condition might promote cell-cell, and cell-matrix interaction and signal communication, which enhanced their autocrine/paracrine activities and maintained their differentiation. In this study, we also observed that osteogenic activity could be influenced by the initial cell number and in vitro culture methods. Ectopic osteogenesis in nude mice is a widely used method for evaluating the performance of bone substitutes. Moreover, subcutaneous implantation is a challenging model for the implants because of the lack of osteoblast progenitors in the implantation area. Twelve weeks after implantation into the subcutaneous pocket, implant I (cell- free DBM) was filled mainly by soft tissues and showed only slight increase in radiographic density, indicating its lack of osteogenic activity in this site. Implant II showed the highest osteogenic activity according to radiography, histology, wet weight, and bone mineral density. This implant was seeded by the hydrogel-assisted method (2×10⁷ cells/ml, 0.05 ml), followed by hydrodynamic culture for 12 days to achieve the plateau cell number and, hypothetically, the best osteogenic activity. Its superior performance confirmed that the combination of hydrogel-assisted seeding and hydrodynamic culture is a promising protocol for tissue-engineering bone grafts. Implant III showed an intermediate osteogenic activity between the implants I and II. This implant was seeded with the same number of hMSCs as implant II by the hydrogel-assisted method, and was immediately implanted without *in vitro* culture. Therefore, a comparison between implants III and II demonstrated that the *in vitro* culture increased the osteogenic activity of implants. The increase may be attributed to several aspects. The *in vitro* culture increased the number of seeded cells, and allowed the cells to adhere more stably to the scaffold and thus prevented their detachment after implantation. The cells might also rearrange in order to more effectively interact and communicate with each other. Additionally, the cells might produce extracellular matrix and osteogenic factors during the *in vitro* culture, which accelerated the subsequent osteogenesis in the subcutaneous pocket. Similarly, implant IV also showed lower osteogenic activity than implant II. Compared with implant II, implant IV was seeded with the same number of cells but statically cultured *in vitro* before implantation. Its inferior performance may be primarily attributed to its lower cell number as a result of the static culture, which lacked mechanical stimulation for the cells to proliferate and differentiate. In summary, both *in vitro* and *in vivo* results suggest that hydrogel-assisted seeding can significantly increase the seeding efficiency and the initial cell density in the cell-scaffold construct. A subsequent hydrodynamic *in vitro* culture can significantly increase the plateau cell density. Correspondingly, bone grafts produced by the combination of these two methods can achieve the highest osteogenic activity. These findings can have a significant bearing in clinical applications and in optimizing tissue engineering strategy. [^1]: The authors have declared that no competing interests exist. [^2]: Obtained permission for use of hMSCs: TYH. Conceived and designed the experiments: FL JZX. Performed the experiments: FL TYH ZHZ ZX. Analyzed the data: XHW JZX. Contributed reagents/materials/analysis tools: TYH. Wrote the paper: FL TYH.

# Introduction The biomechanical properties of the human thoracic spine are still largely unknown, since spinal research has mainly focused on the lumbar and cervical spine in the past. Especially the influence of the anterior rib cage structures on the stiffness and motion behavior of the thoracic spine is poorly investigated. Detailed knowledge of the thoracic spine biomechanics including the rib cage is required for improving surgical and conservative treatments of the thoracic spine, for instance in the case of scoliosis or hyperkyphosis. Numerical models of the thoracic spine and rib cage therefore provide a useful tool. In order to validate these models, kinematics data and stiffness properties of the thoracic spine including the rib cage are essential. Since the validation of numerical models is based on the design of the experimental test setup, the results of former in silico studies varied considerably. A clear definition of initial and boundary conditions is therefore of high importance in order to obtain long lasting and impactful data. An established method for characterization of spinal stiffness properties and data collection for the validation of numerical models is provided by the concept of stepwise reduction of single anatomical structures. This concept was mainly applied to single motion segments of the cervical and lumbar spine in the past. The first in vitro experiments quantifying the effect of the rib cage structures on thoracic spine stability were performed on canine specimens, in which the effects of the posterior elements, costovertebral joints, and intervertebral discs were investigated. Further in vitro studies using human thoracic spine specimens, which included the rib cage structures, determined the effect of discectomy, rib head resection, facetectomy, laminectomy, transversal and median sternotomy, costosternal release, sternal fracture, the anterior rib cage except of the costovertebral joints, and the superior rib cage. All these studies concluded that the single rib cage structures have a significant effect on thoracic spine stability. Nonetheless, there are still many open questions regarding the influence of the rib cage on the stability of the thoracic spine. For example, the monosegmental range of motion of the thoracic spine including the entire rib cage has never been investigated, nor has the effect of the rib cage on the biomechanics of the entire thoracic spine. Therefore, the objectives of the present study were: (1) To determine the stabilizing effect of specific rib cage structures (intercostal muscles, sternum, and costovertebral joints) on the entire human thoracic spine in order to calibrate and validate finite element models of the thoracic spine and rib cage system and to quantify the effect of standard surgical interventions (median longitudinal sternotomy, costovertebral release) on thoracic spine stability. (2) To examine the monosegmental ranges of motion of the thoracic spine in polysegmental testing with and without the rib cage. (3) To investigate the effect of the rib cage on coupled motions of the thoracic spine. Based on the previously published experiments, it was hypothesized that each specific structure has a significant effect on the stability of the thoracic spine. # Materials and methods ## Study design Six fresh frozen human thoracic spine specimens (C7-L1) including all rib cage structures with an average age of 56 years (range 50–65) were dissected from entire torsos (one male and five female). During preparation, specimens were scanned for ligament damage, bone fractures, and osteophytes. Bony, cartilaginous, and ligamentous structures as well as the intercostal muscles were left intact. The specimens were stored at -20°C and thawed at 6°C for 10 h prior to preparation and testing, which were both performed within 20 h to avoid disintegration. The cranial end of C7 and the caudal end of L1 were each embedded half in PMMA (Technovit 3040, Heraeus Kulzer, Werheim, Germany) for biomechanical testing. Prior to potting, screws were placed each in the cranial (C7) and caudal (L1) endplates to obtain a better fixation. Screws were also driven through the intervertebral disc between C7 and T1 to restrict the high mobility of this motion segment. The test procedure was designed as a stepwise reduction of the single specimen structures whereby biomechanical testing and motion analysis between each reduction step was performed. The intercostal muscles and the rib heads were removed using a scalpel. Sixth to eighth rib head removal simulated the concept of costovertebral release, which provides good results in the surgical treatment of scoliosis. Median longitudinal sternotomy and the removal of the anterior part of the rib cage up to the rib stumps were performed using an oscillating saw (OR-SY-518.01, Synthes®, Zuchwil, Switzerland). Median longitudinal sternotomy is a standard clinical intervention, which is performed in most cardiac surgeries such as cardiopulmonary bypass surgery or heart transplantation. The destabilizing effect of median sternotomy on the thoracic spine has already been shown in a previous study, where the stabilizing effect of sternal closure was investigated. The conduction of the study and the use of human specimens were approved by the ethical committee board of the University of Ulm, Germany (No. 302/14). The specimens were provided by the Anatomy Gifts Registry program (AGR, Hanover, Maryland, USA). ## Biomechanical testing The specimens were loaded quasi-statically on the cranial end (C7) with pure bending moments of ± 2 Nm in the three loading planes flexion/extension, lateral bending, and axial rotation using a custom-built spine tester. The loads were applied with a constant loading rate of 1.0°/s in flexion/extension and lateral bending as well as 0.5°/s in axial rotation to avoid viscoelastic superposition of the motion response. To reduce viscoelastic effects, 3.5 loading cycles were performed in each loading plane, of which the last loading cycle was chosen for data evaluation. The caudal end of the specimens (L1) was each fixed in an angle flange. Because the angle between the thoracic and lumbar spine is about 20°, the angle flange was positioned at 20° for four of the specimens with regular kyphosis and at 30° for two of the specimens with hyperkyphosis to reach a balanced, centered position within the spine tester. ## Motion analysis Simultaneously with mechanical loading, motion data was captured using the optical motion tracking system Vicon MX13 (Vicon Motion Systems Ltd., Oxford, UK) consisting of six cameras. Motion analysis was performed to calculate range of motion (ROM) and neutral zone (NZ) together with the applied moment for the single functional spinal units (FSUs) from T1 to T12, as well as the coupled motions of the FSUs. All single vertebrae were therefore each equipped with three optical markers. The motion analysis used in the test setup was determined to have an accuracy of 0.1 mm or 0.1°, based on preliminary tests. Motion data was evaluated using Nexus 1.8.5 (Vicon Motion Systems Ltd., Oxford, UK) and processed with MATLAB R2014b (MathWorks, Natick, MA, USA) and Microsoft Excel 2010 (Microsoft Corp., Redmond, WA, USA). ROM was defined as the maximum deflection of the single motion segments in one direction. NZ was defined as the intersection point of the hysteresis curve with the displacement axis at a bending moment of 0 Nm. NZ represents the laxity of a specimen and specifies the region in which no or only small moments are generated during an applied displacement. ## Statistical analysis Statistical tests were performed using the statistical analysis software SPSS 21 (IBM Corp., Armonk, NY, USA). ROM and NZ data were checked for normal distribution within the six specimen groups using Shapiro-Wilk normality test (p \> 0.1) and then tested for significance (p \< 0.05) using parametric Student’s t-test to evaluate differences between the six testing conditions. Because all ROM and NZ data were normally distributed, data were represented as mean with standard deviation. Holm-Bonferroni method was performed in order to account for the familywise error rate caused by multiple comparisons between the different groups. Finally, Cohen’s sample effect size (0.2 \< d \< 0.5: small effect, 0.5 \< d \< 0.8: mean effect, d \> 0.8: strong effect) and observed power (1-β \> 0.8) were evaluated to estimate the achieved statistical power of the results. Power analysis was performed post hoc using the Software G\*Power. Monosegmental ROM data were analyzed qualitatively and represented as median value to minimize the effect of outliers resulting from polysegmental testing. Coupled motions of the entire thoracic spine were also checked for normal distribution and represented as mean with standard deviation. # Results The rib cage and its single structures were found to have a strong effect on thoracic spine rigidity. Dissecting the intercostal muscles caused a significant increase of ROM and NZ in lateral bending (+ 22.4%, + 30.0%) and axial rotation (+ 22.6%, + 21.7%). Removing the anterior part of the rib cage up to rib stumps had the strongest effect on thoracic spine rigidity. Compared to the previous condition, ROM and NZ increased significantly in flexion/extension (+ 21.1%, + 25.3%), lateral bending (+ 9.8%, + 10.9%), and axial rotation (+ 30.1%, + 72.5%) without the anterior rib cage. Compared to the intact condition, the removal of the anterior part of the rib cage produced a significant increase in the ROM and NZ during flexion/extension (+ 52.2%, + 45.6%), lateral bending (+ 42.0%, + 54.0%), and axial rotation (+ 94.4%, + 187.8%), with a strong effect size (d \> 0.8) in each loading direction, as well as a high power (1-β \> 0.8) in axial rotation. Removing the entire rib cage increased the ROM significantly in all loading directions relative to the intact condition. ROM and NZ generally increased after each resection step except for the last resection step in flexion/extension. Performing a median sternotomy increased the ROM significantly in flexion/extension and axial rotation compared to its previous condition without intercostal muscles, as previously shown. Compared to the intact condition, ROM and NZ were significantly increased after a median sternotomy in flexion/extension (+ 25.0%, + 16.2%), lateral bending (+ 29.3%, + 38.8%), and axial rotation (+ 49.4%, + 66.9%). Removing the sixth to eighth right rib head increased the ROM significantly in axial rotation (+ 11.7%) relative to its previous condition with all costovertebral joints left intact. Removing the entire rib cage generally increased the monosegmental ROM, except for the lower two segments T10-T11 (in lateral bending) and T11-T12 (in all three directions). In flexion/extension and lateral bending, the highest flexibility was detected in T1-T2, decreasing in the inferior direction. In axial rotation, the flexibility of the FSUs was more evenly distributed over the entire thoracic spine, but increased much more in the superior half after removing the entire rib cage compared to its intact condition. Because of inverse motions of the functional spinal units, the monosegmental ROM were found to be negative in some cases, especially in flexion/extension and generally in the condition without the entire rib cage. Coupled motions of the thoracic spine with the intact rib cage were detected mainly in lateral bending, with a large coupled axial rotation in the opposite direction, as well as in axial rotation with a slight coupled lateral bending in the same direction. After removal of the entire rib cage, these coupled motions increased approximately proportionally to the increasing ROM of the main loading directions. # Discussion Detailed knowledge of the biomechanics of the thoracic spine and the effect of the rib cage on its stability is essential for the improvement of surgical interventions, as well as therapeutic treatments in the thoracic region. Numerical models of the thoracic spine including the rib cage may help to understand and optimize the effects of specific medical procedures, but need to be validated by experimental data. Several in vitro studies investigated the effect of specific rib cage structures on thoracic spine stability using multiple kinds of test setups and few resection steps. The purpose of the present study was therefore to investigate the stabilizing effect of the rib cage on the thoracic spine and to provide data in order to calibrate and validate finite element models of the thoracic spine including the rib cage. The results of the present study showed that all rib cage structures contribute to thoracic spine stability. The strong stabilizing effect of the rib cage becomes distinct especially when looking at the difference between the intact rib cage and the isolated spine, which has never been investigated before. These findings suggest that the rib cage is of prime importance regarding the biomechanical integrity of the spine, since the thoracic spine is additionally stabilized by the ribs and the costovertebral ligaments, in contrast to the cervical and lumbar spine, which are mainly stabilized by the adjacent musculature. Therefore, it can be assumed that the rib cage, in addition to its function as protector for the inner organs, respiration support, and framework for the insertion of muscles, has the role of a flexibility limiter for the spine in order to prevent damage caused by high bending moments during movements. The stabilizing effect of the intercostal muscles, also never previously investigated in vitro, was especially evident during lateral bending and axial rotation, when these muscles were stressed in tension. Intercostal muscles should therefore be left intact for biomechanical testing and should be included in computational models of the rib cage. Removing the anterior part of the rib cage up to the rib stumps had the strongest effect on thoracic spine stability of all resection steps performed in the present study, as well as any former in vitro studies. Watkins et al. detected the strongest effect in flexion/extension using a biaxial testing device without applying pure moments, whereas Mannen et al. observed the strongest effect in axial rotation, which corresponds to the results of the present study. The sternal complex, which connects most of the single ribs with each other anteriorly, seems therefore to be of high importance for the rigidity of the thoracic spine regarding the effect of the rib cage. The destabilizing effect of median sternotomy on the thoracic spine, especially in axial rotation, was also detected in the study of Brasiliense et al., although relatively smaller compared to the present study. This might be due to their test setup, in which only the upper four segments were tested, while relative motions between the ribs and vertebrae were restricted. Combining the findings of these two studies, sternal closure is recommended after surgical transection of the sternum from a biomechanical perspective to prevent spinal destabilization. The removal of the right sixth to eighth rib head only produced a significant effect in axial rotation compared to its previous condition with all costovertebral joints left intact. This agrees with the study of Feiertag et al., who also did not find a significant increase in the ROM after removal of the right ninth rib head under flexion/extension as well as lateral bending. In contrast, Oda et al. detected a significant increase in the ROM after right rib head removal in all loading planes using thoracic functional spinal units. However, since Oda et al. performed a radical discectomy before rib head removal, the results are not directly comparable. Nevertheless, partial costovertebral release seems to destabilize the thoracic spine at least in axial rotation. This might facilitate the surgical correction of scoliosis, whereby a spinal fixation is recommended to avoid spinal instability. Compared to the literature, monosegmental ROM of the present study was quite low in flexion/extension and lateral bending. This might indicate that higher bending moments than 2 Nm and monosegmental testing provide a more physiologic simulation of the in vivo situation in these two loading directions, especially in the lower segments where the intervertebral discs exhibit larger cross sections. Also the high NZ to ROM ratios in flexion/extension and lateral bending suggest that the flexibility of the thoracic spine is much higher in vivo. The results of the present study also indicate that the lower two pairs of ribs do not have a significant effect on thoracic spine stability, since the costovertebral joints of the lower two pairs of ribs are not located at the level of the intervertebral disc and therefore do not provide intersegmental stability. In the present study, lateral bending in one direction induced a considerable axial rotation in the opposite direction. This was also observed in the study of Brasiliense et al. in the upper third of the thoracic spine, whereas Willems et al. determined coupled axial rotation in same direction. The slight coupled lateral bending in the same direction during axial rotation observed in the present study is in accordance with the results of former in vivo studies. Coupled motions of the thoracic spine can be caused by several factors, for instance by the specific orientation of the facet joints, the morphology of the intervertebral discs, ligamentous structures, or the ribs. The different coupled motion behavior during lateral bending and axial rotation in our study is probably due to morphological characteristics of the thoracic spine and rib cage. This complex motion behavior should be analyzed separately by means of monosegmental testing of the thoracic spine including the ribs. Because of the alternating motion pattern, monosegmental coupled motions were not quantified in the present study, but it is known that coupled motions of thoracic spinal segments decrease in the inferior direction. The present study entailed some limitations. Statistical differences and observed power of the results were quite low, which can be traced to the small specimen number (n = 6). However, increasing the number of specimens to achieve higher power is also related with ethical and financial issues. Moreover, the present study primarily intended to provide data for numerical modeling. Statistical significance could have been also increased without using the Holm- Bonferroni method, which is critically discussed in research. The relatively high standard deviations of the results followed from the interindividual differences of the single specimens, which showed each a similar motion behavior in relative terms. Another limitation of the present study was the polysegmental testing, which is just qualitatively comparable to monosegmental testing, since the ROM is tending to be smaller in polysegmental testing. Monosegmental ROM was also sometimes found to show an inverse motion behavior in the present study and should therefore be tested separately in a monosegmental test setup with associated rib and sternal portion. Including a follower load in the test setup, which is sometimes stipulated in thoracic spine research, may improve the monosegmental motion behavior, but it would also increase the specimen stiffness and the number of boundary conditions in a numerical model validation. In the present study, pure bending moments of 2 Nm were applied, which correspond to former in vitro studies using a polysegmental test setup for the thoracic spine, whereas generally pure moments of 5 Nm are recommended for the thoracic region in spine research. Because of specimen length and limited testing device dimensions, loading of the specimens with 5 Nm was not feasible in flexion/extension and lateral bending. However, it was imperative to perform polysegmental testing when evaluating the effect of the entire rib cage on thoracic spine stability. Due to its in vitro nature, the results of the present study are not directly comparable to studies which were performed in vivo, although it was previously shown that the application of pure moments can simulate physiological loading conditions. In future studies, muscle forces should therefore be included in the experimental setup to simulate an even more physiological motion behavior of the thoracic spine and the rib cage. # Conclusions The rib cage and its single structures have a significant stabilizing effect on the thoracic spine. Therefore, the rib cage should be regarded in clinical scenarios, in vitro, as well as in numerical models of the thoracic spine. Using the biomechanical data of the present study, numerical models of the thoracic spine including the rib cage can be calibrated and validated. # Supporting information The authors thank Tobias Böckers, Ulrich Fassnacht, Ernst Voigt, and Michael Reinehr from the Institute of Anatomy and Cell Biology, School of Medicine/University of Ulm, for their support. They also want to thank David Volkheimer and Theodor Di Pauli von Treuheim for carefully editing the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** CL KA HJW. **Data curation:** CL. **Formal analysis:** CL. **Funding acquisition:** HJW. **Investigation:** CL. **Methodology:** CL NG. **Project administration:** CL HJW. **Resources:** HJW. **Software:** CL NG. **Supervision:** HJW. **Validation:** CL. **Visualization:** CL. **Writing – original draft:** CL. **Writing – review & editing:** CL HJW. [^3]: Current address: SpineServ GmbH & Co. KG, Ulm, Germany [^4]: Current address: University Hospital of Basel, Basel, Switzerland

# Introduction Polyploidy has played a major role in generating angiosperm biodiversity. Chromosome counts suggest that between 30 and 80% of angiosperm species are polyploid, while genomic studies of selected model and crop species reveal evidence of extensive ancient genome-wide multiplications. Indeed recent genomic investigations indicate that most, if not all, angiosperm species have undergone at least one genome duplication event in their evolutionary history, and several have evidence of multiple polyploidy-diploidisation-polyploidy cycles. Angiosperm genomes are astonishingly plastic in their ability to tolerate considerable karyotypic (e.g. chromosome number variation, translocations), genetic (mutations, retroelement transpositon, deletions) and epigenetic (DNA methylation, histone methylation/acetylation) variability. This tolerance enables polyploids to form and establish and has contributed significantly to their widespread occurrence. Large-scale genetic changes induced by polyploidy are thought to influence the transcriptome, metabolome and proteome, which can concomitantly alter the phenotype and ecology of the individuals. Most of the new genetic changes are probably maladaptive, but in a few rare instances individuals arise that are able to outcompete the parental diploids or colonize new niches. There are several examples of recent speciation via polyploidy that occurred within the last 150 years: *Spartina anglica*, *Senecio cambrensis* and *S. eboracensis*, *Cardamine schulzii*, *Tragopogon mirus* and *T. miscellus*. Studies on the genetic consequences of allopolyploidy in these *de novo* polyploid species reveal some significant differences. In *S. anglica* allopolyploidy induced few changes in genome structure, but there is epigenetic reprogramming, while in *Senecio*, and *Tragopogon* allopolyploids there are substantial genetic changes including loss of sequence (genomic DNA profiles) and perturbations to the transcriptome (cDNA profiles). However, little attention has been paid to the chromosomal content of any of these allopolyploids, and it is unknown what mechanisms drive the observed genetic changes. For example, genetic change could be driven by local mutation, small- scale deletion or insertion of a particular sequence, or via major chromosomal changes, including whole-arm transposition and chromosome losses or duplications. Because there is little or no understanding of chromosomal variation in these recently formed, natural polyploids, we have embarked on a characterization of karyotypic variation between individuals and populations of the two *Tragopogon* allopolyploids from North America. *Tragopogon mirus* and *T. miscellus* have proved to be excellent evolutionary model systems for understanding early allopolyploid formation. These allopolyploids are derived from three diploid progenitors (each 2*n* = 2*x* = 12), *T*. *pratensis*, *T*. *porrifolius* and *T. dubius*, the latter being shared by both allopolyploids (*T. mirus* derived from *T*. *dubius*×*T*. *porrifolius*, 2*n* = 4*x* = 24; *T. miscellus* derived from *T*. *dubius*×*T*. *pratensis*, 2*n* = 4*x* = 24). The diploid parents of both polyploids are in well separated clades and are not closely related based on ITS/ETS sequences, allozymes and other genetic markers. There are many reports of natural F₁ hybrids involving these diploids and we have also produced them in the glasshouse, but these hybrids are highly sterile suggesting minimal homeologue pairing at meiosis. In contrast the allopolyploids *T. mirus* and *T. miscellus* are fertile and expanded their ranges rapidly after their initial formations, in large part via multiple origins. The two allotetraploids now occupy a large geographic area of eastern Washington and adjacent Idaho, USA, and comprise many thousands of individuals in several populations. Molecular analyses have revealed that *T. mirus* has recurrently formed at least 13 times and *T. miscellus* possibly as many as 21 times \[13,16,18,19,20, Symonds et al. unpubl.,21\], reviewed in Soltis *et al.*. Furthermore, there are genetic differences between populations of each tetraploid, most of which likely reflect variation found in the diploids and inherited in the polyploid populations through recurrent formation, and others of which may have arisen through subsequent divergence of the polyploids. Examples of the latter include the divergence of rRNA gene copy number, sequence homogeneity and expression patterns. Here we show that population differences are also reflected in substantial variability in karyotypes among individuals, differences that appear correlated with irregular meiosis. # Results ## Genome structure of *Tragopogon* allotetraploids The chromosome sizes, numbers and centromere indices are similar between the parental diploid species, and three tandem repeats characterized by us do not generate differences in chromosomal distribution. For these reasons we are unable to identify the parental origin of the chromosomes in the derived allotetraploid species via morphology alone. We therefore used FISH with total genomic DNA probes (called Genomic In Situ Hybridization or GISH) to determine the genomic composition of karyotypes of *T. mirus* and *T. miscellus* individuals. We analysed 12 plants, nine of *T. mirus* from three populations and three of *T. miscellus*, two from one population and a third individual from a second population. Five of these plants, including representatives of each tetraploid, were chosen because we had previously observed, that they had particular 45S nuclear ribosomal DNA (45S rDNA) compositions and expression characteristics. GISH labelling enabled the genomic origin of the chromosomes to be determined by fluorescence colour: digoxingenin-labelled genomic DNA of *T. dubius* labelled chromosomes of *T. dubius* origin green or yellow (D-genome) and biotin-labelled genomic DNA of either *T. pratensis* (to *T. miscellus*) or *T. porrifolius* (to *T. mirus*) labelled the P-genome (either *T. pratensis* or *T. porrifolius*) orange or red. However, the distinction between genomes was only possible after electronic merger of the images because there was considerable cross-hybridisation of probes. The homologous group assigned to each chromosome (A–F) was determined using size, arm ratio, position of 5S and 45S rDNA, as described in Pires et al., and by using DAPI bright bands that were revealed after denaturation in some metaphases. Previous cytogenetic analyses revealed that both *T. mirus* and *T. miscellus* had 24 chromosomes; aneuploidy has not been previously reported in these taxa. This chromosome number was also observed here in all but one individual of *T. mirus*, which had 23 chromosomes. It was assumed that GISH would partition the chromosome sets into 12 chromosomes of *T. dubius* origin and 12 chromosomes of either *T. pratensis* or *T. porrifolius* origin depending on the tetraploid being analysed. However, only four had an entirely balanced additive karyotype of that expected from the chromosomes of the diploid parents ( and and and). We were surprised to observe that five of the eight *T. mirus* karyotypes and two of the three *T. miscellus* karyoypes with 2*n*  = 24 chromosomes had unbalanced genomic contributions. This was manifest by some chromosomes occurring in one or three copies (monosomic and trisomic, respectively; see). In addition, four plants, including individuals of both *T. mirus* and *T. miscellus*, carried intergenomic translocations (arrows). The *T. mirus* individuals 2603-33A and 33B are siblings from different flower heads of the same plant. Significantly, their karyotypes exhibit substantial differences; 2603-33B has an expected karyotype assuming complete additivity of parental chromosome sets, while 2603-33A is monosomic for chromosome F^du and trisomic for chromosome E^po. ## Meiotic and mitotic aberrations in synthetic reconstruction of *T. mirus* Meiosis in newly formed, synthetic allotetraploids (S₀ and S₁) was examined to determine if meiotic aberrations occurred in early allotetraploid generations and could be a source of genomic imbalance observed in root tip metaphases. Meiotic cells of developing anthers at diplotene were analysed. In many instances regular bivalent formation occurred, although frequently bivalents overlapped, making it difficult to be certain if they were in fact multivalents. However, in a number of cases resolution was sufficient to determine the presence of quadrivalents. Using GISH, the genomic origin of the chromosomes could be resolved, although the distinction was less clear than at metaphase. show GISH labelling of a quadrivalent with two chromosomes of *T. dubius* origin and two chromosomes of *T. porrifolius* origin physically linked via chiasma in a large, twisted ring. An analysis of root tip metaphases in 12 allotetraploids of the subsequent S₁ generation revealed one plant that was 2*n* = 23, the rest were 2*n* = 24, as expected. There were cytogenetic abnormalities in two additional plants. One of these plants (73-14) had unusually large 45S rDNA loci occurring on both copies of chromosome A^du ( A, B). This plant also had an rDNA locus on chromosome A^po of *T. porrifolius* origin, but the expected site on chromosome D^po was absent. The second plant (134-16-3) had a translocation of *T. porrifolius* origin to chromosome C^du. Since there is no missing chromosome segment of *T. porrifolius* origin, it must be assumed that this was a non-reciprocal translocation induced during the preceding meiosis ( C, D, E). ## 45S rDNA decondensation Previous cytogenetic analyses of diploid *Tragopogon* species revealed that *T. dubius* and *T. pratensis* carry a 45S rDNA locus on chromosome A^du and A^pr, respectively, while *T. porrifolius* has two loci on chromosomes A^po and D^po. To determine if these sites were inherited in the respective allotetraploid species we used FISH with digoxigenin-labelled probe pTa71 (yellow fluorescence,). In *T. miscellus*, 45S rDNA loci occur on chromosomes A^du and A^pr. At metaphase both chromosomes carried secondary constrictions. In *T. mirus*, 45S rDNA loci occur on chromosomes A^du, A^po, and D^po. These loci were also frequently identifiable after GISH labeling, without the use of the rDNA probe pTa71, appearing with a brighter green (to *T. dubius* origin loci) or yellow fluorescence. In all cells, both mitotic and meiotic, the number of 45S loci and their parental distribution were as expected, i.e. pairing and segregation of rDNA-carrying chromosomes is probably regular. This was not the case, however, for 5S rDNA-carrying chromosomes (5S probe, red fluorescence). We would expect *T. mirus* to carry 5S rDNA loci on chromosomes A^du, A^po and F^po. But here we observed one *T. mirus* individual (2602-03-10) trisomic for chromosome F^po and each of these chromosomes carried a single 5S rDNA locus. It was a surprise that the chromosomes carrying the 45S rDNA were balanced because at diplotene they were frequently in close proximity ( E, F), probably reflecting transcriptional activity and nucleolar function. The absence of aberrant numbers of 45S rDNA- carrying chromosomes at metaphase suggests that they paired regularly at meiosis. An analysis of 45S rDNA-carrying chromosomes in 14 diplotene nuclei of *T. mirus* revealed that chromosome A^du was always associated with the nucleolus. Often this chromosome was associated with a chromosome from the *T. porrifolius* genome (when identifiable it was chromosome D^po). Similar results were observed at metaphase , where chromosome A^du and sometimes chromosome D^po carried secondary constrictions (the exception is plant 2603-33A–below). In contrast, chromosome A^po was always condensed and unassociated with the nucleolus. Interestingly, in two individuals of population 2602 only one of the two homologues of chromosome A^du carried secondary constrictions. In plant 2603-33A, there was no secondary constriction on chromosome A^du and the 45S rDNA locus was unusually small. Instead, chromosome D^po had a secondary constriction. Similarly in two synthetic *T. mirus*, plants 134-16-3 and 73-14, there were secondary constrictions on A^du and A^po, with the locus on chromosome D^po missing in the latter individual. In natural *T. mirus* and *T. miscellus*, the presence of secondary constrictions correlated well with patterns of condensation/decondensation at interphase; for example, chromosomes A^po and D^po have no secondary constrictions in *T. mirus* 2601-4 (not shown) and exhibit four sites of condensed rDNA chromatin while in *T. mirus* (2602-0-3-10) chromosome A^po is without a secondary constriction and at interphase there are two highly condensed rDNA loci. ## Southern blot analysis of 5S rDNA loci Given the imbalance observed in the inheritance of individual chromosomes, we would expect the copy number of 5S rDNA units to vary among plants. For example, individual 2602-3-10, showing trisomy of the F^po chromosome, had seven 5S rDNA signals instead of the expected six. On the other hand, if there was F^po monosomy, we might expect a reduction in gene copy numbers. Using Southern blot hybridisation, we determined the parental *T. porrifolius*/*T. dubius* 5S rRNA gene copy number ratio in the progeny of a single *T. mirus* plant, 2602-0-3. Several progeny were studied, including those analysed by GISH (plants 2602-0-3-10 and 2602-0-3-4). To ascertain the parental origin of the 5S rDNA units, we digested genomic DNA with *Taq*I restriction enzyme. This enzyme takes advantage of a restriction site present in *T. porrifolius* units that is not found in *T. dubius* units. Consequently, after digestion the *T. porrifolius* 5S rDNA was digested mostly to monomers while there was little digestion of *T. dubius* units that remained at high molecular weight. In *T. mirus*, the 5S rDNA probe used for Southern hybridisation revealed signals from both parental gene families. However, the distribution of signal intensities between genes of *T. porrifolius* and genes of *T. dubius* origin differed between progeny. For example, the trisomic individual 2602-0-3-10 had a *T. porrifolius*/*T. dubius* 5S rDNA unit ratio close to 1.0, while the “normal” disomic plant 2602-0-3-4 had a ratio that was 0.82. Thus, trisomy of chromosome F^po resulted in an approximately 10% increase in the number of 5S genes in the genome. This value is in a good theoretical agreement with the expected increase in number of 5S arrays from six to seven (per diploid cell). Three other plants (2602-0-3-5, -7 and -8) had a 5S rDNA gene ratio corresponding to a trisomic genotype, suggesting meiotic aberrations were frequent in this lineage. On the other hand, two plants (2602-0-3-2 and -3) had a decreased number of *T. porrifolius* genes compared with the expectation (plant 2602-0-3-4, with balanced chromosomal sets). Perhaps these two individuals are monosomic for F^po. # Discussion ## Effectiveness of GISH Phylogenetic analyses of internal (ITS) and external transcribed spacer (ETS) sequences of nuclear 45S rDNA suggest that the diploid progenitors of both *T. mirus* and *T. miscellus* are distantly related, with *T. dubius* in one major clade of *Tragopogon* and both *T. pratensis* and the populations of *T. porrifolius* that served as parents in another major clade. Earlier allozyme studies also suggested that the three diploid progenitors were well differentiated genetically. Likewise, comparisons of cDNA-AFLP genetic markers between the diploid species reveal that *T. pratensis* and *T. dubius* share only between 30-40% of markers, again suggesting considerable genomic divergence. Recently, Markova et al. used GISH with genomic DNA probes from diploid species of *Silene* onto metaphases of related diploids to show that labeling strength was inversely correlated to genetic distance, i.e. there was strongest labeling to the most closely related species. Given these data and an apparently large genetic distance between *Tragopogon* diploids, we might expect GISH to work effectively on the derived allopolyploids. Instead, we observed that it worked rather weakly, with much cross-hybridisation of probes, and the genomic distinction was only resolvable after electronic merging of images. Indeed previously we had reported that GISH had not worked at all in *Tragopogon*. Our success here is due to improved quality of the electronics on the microscope's camera and improved facilities to electronically merge images using Openlab software, but it remains an enigma why GISH does not work more effectively in *T. mirus* and *T. miscellus*. ## rDNA inheritance and expression patterns We expect *T. miscellus* to inherit two 45S rDNA loci (carried on chromosome A^du from *T. dubius* and A^pr from *T. pratensis*), while *T. mirus* inherits three loci (carried on chromosomes A^du from *T. dubius* and A^po and D^po from *T. porrifolius*). The observations of only two 45S rDNA in synthetic *T. mirus* 73-14 can be explained in one of two ways. First, the enlargement of the rDNA locus on chromosome A^du and the locus loss on D^po occurred through an rDNA translocation/fusion involving the loci on D^po and A^du. However, given that the karyotype is balanced and the material is the S₁ generation, this scenario seems unlikely because it would suggest the union of two gametes carrying the same abnormality, or arising via restitution of segregating chromosomes in meiosis II. The second and alternative explanation is that the rDNA locus number may reflect the situation in the *T. porrifolius* parent used for the cross, although our past analyses of *T. porrifolius* never revealed such a polymophism. Unfortunately the precise *T. porrifolius* parent used in the cross is now deceased (the plants are annuals or biennials), preventing us from determining which of the competing hypotheses is correct. Previously we showed that natural *Tragopogon* allopolyploids also do not have fixed patterns of 45S rDNA inheritance, with some individuals showing a balanced distribution of rDNA sequences originating from each of the parental diploids and some showing biased inheritance, typically with the number of rDNA units from *T. dubius* being underrepresented. However, from molecular studies it was not clear whether non-Mendelian rDNA inheritance was caused by a decrease in copy number (elimination) or unit replacement (e.g. via homogenisation mechanisms). In one individual of *T. mirus*, 2603-33 (referred to here as 2603-33A), there was a considerable reduction in the number of *T. dubius* units (to ∼100 copies/diploid) from an expectation of ∼700 units (typical number for any given tetraploid population). An analysis of this individual's karyotype revealed that the loss in copy number can be accounted for by a reduction in the size of the rDNA locus on chromosome A^du. The sibling to this plant, 2603-33B, does not show this rDNA copy number deletion. Therefore, the locus size reduction in 2603-33A may have arisen from a deletion event occurring as a consequence of meiotic instability in the parent. Despite the relatively small size (perhaps only 100 copies/diploid of the 45S rDNA unit) of the *T. dubius* 45S rDNA locus, it usually dominates rRNA expression in leaf material of *T. mirus*. Even in the 2603-33A individual, with extremely reduced numbers of 45S rRNA genes of *T. dubius* origin, the locus accounts for 97% of the rRNA transcripts (see also). Nevertheless, in root-tip metaphases a secondary constriction is observed on chromosome D^po, suggesting some transcription of the *T. porrifolius* locus occurs in that tissue, perhaps to compensate for the reduced 45S rDNA gene copy number when there is a high demand for ribosomes in metabolically active meristematic cells. In plant 2603-33B, with substantially higher numbers of *T dubius* derived 45S rRNA genes, the units of *T. dubius* origin are transcribed, and those rDNA units of *T. porrifolius* origin are silent. This is reflected in the lack of a secondary constriction at the rDNA locus on chromosomes A^po and D^po in this individual. The presence of secondary constrictions correlates strongly with levels of decondensation at interphase and almost certainly reflects transcriptional activity at the preceding interphase. Two individuals of *T. mirus* (2602-4 and 2602-0-3-4) show different decondensation of the two A^du homologues, probably reflecting genetic or epigenetic differentiation between the two homologues. In population 2605 of *T. miscellus*, all individuals investigated show only partial dominance in the expression of rDNA units of *T. dubius* origin over those of *T. pratensis* origin. This is also seen in the occurrence of secondary constrictions on both rDNA-carrying chromosomes A^du and A^pr. An analysis of 45S rDNA evolution in *Nicotiana* polyploids indicates that parental loci are initially maintained in young polyploids, although the sequences within a locus may be subject to concerted evolution, and over time frames of \>1 million years individual loci are lost. In *Tragopogon* polyploids we do not know which of the different karyotype variants will survive selection and become fixed. Perhaps over longer evolutionary timescales interlocus homogenisation and new rDNA variants will occur and spread across all rDNA loci. Our study shows that more than one pathway can lead to non-Mendelian inheritance of rDNA units in allotetraploids: (i) elimination or amplification of repeats within an array can occur without changes in locus number (as in the case of 45S rDNA locus in 2601-33A); (ii) a change in the number of rDNA-bearing chromosomes (and loci) without any material change in the number of genes at each locus (as with the 5S rDNA locus–); or (iii) a combination of both, although this situation was not detected in this study. ## Meiotic irregularities The chromosome multiplication step of polyploidy is thought to establish species isolation barriers between the newly formed polyploid and its diploid parents, whilst providing a homologous partner for each of the chromosomes. However, analyses of newly synthesised allopolyploids reveal that early-generation individuals are often infertile, or have highly reduced fertility, due to problems with meiosis including irregular pairing of homologous chromosomes. Selection for fecundity in synthetic polyploids is associated with generation- by-generation increased fertility. Nevertheless even after many thousands of years of evolution, meiotic irregularities can still occur, as observed in *Triticum aestivum* (wheat), an allohexaploid where meiotic misdivision has been exploited in the formation of wheat aneuploid lines. An analysis of meiosis in the newly synthesised *Tragopogon* allotetraploids revealed the frequent occurrence of multivalents. Such aberrant pairing patterns may result in imbalanced chromosome contribution in subsequent generations as well as intergenomic translocations. Both abnormalities were observed in several of 12 synthetic (S₁ generation) *T. mirus* plants analysed at metaphase. One plant was 2*n* = 23 and lacked a *T. dubius* chromosome and one plant carried a non-reciprocal translocation from a *T. porrifolius* chromosome to chromosome C^du Multivalents were also observed at a low frequency in natural populations of *T. mirus* and *T. miscellus* and in the few F₂ plants resulting from diploid F₁ hybrids of *Tragopogon*. Meiotic abnormalities can cause unequal segregation of homeologous chromosomes and are a likely driver of the chromosomal imbalance between genomes observed in both *Tragopogon* allotetraploids. The two sibling plants of *T. mirus* 2603-33 had different karyoypes: one had a balanced karyotype of 12 chromosomes from each parental genome (2603-33B), while the other (2603-33A) was monosomic for chromosome F^du and trisomic for chromosome E^po. Given these data, it is likely that the parent plant had a balanced karyotype, and meiotic irregularities, probably arising through multivalent formation, gave rise to the imbalanced karyotype of plant 2603-33A. Multivalent pairing can arise either through (1) synapsis and recombination between homeologous chromosomes in meiosis I, or (2) synapsis between chromosomes carrying intergenomic translocations. Four out of 12 mitotic karyotypes in plants from natural populations had intergenomic translocations visible at mitotic metaphase following GISH , and additional smaller translocations, not resolvable by GISH, may also be present. Nevertheless, the quadrivalent indicated in (see arrow) does not show intergenomic translocations at the resolution obtained using GISH. Our analyses of plants from different populations, although sample sizes were small, showed that aberrant chromosome numbers were more prevalent for certain chromosome types. For example, chromosome A was not involved in any of the aneuploidy evetns detected, whereas chromosome F accounted for almost 40% of all subgenomic chromosome imbalances. There might be several explanations for these results. First the homology between F-type chromosomes may be larger than between other chromosomes. It is likely that the greater sequence and morphological similarities between homeologous chromosomes, the more likely there will be homeologous and multivalent pairing. In support of this, Nicolas *et al*. observed in *Brassica napus* haploid hybrids that chromosomes with the highest synteny had the highest frequency of homeologous pairing. Nevertheless, in *T. mirus*, the presence of a 5S rDNA locus on F^po but not on F^du indicates some divergence between the homeologues. Secondly, all chromosomes may be equally likely to form multivalents, but aberration in copy number of some chromosomes, e.g. homeologous group A, may not be favored by selection. The surprisingly high incidence of trisomy and monosomy in highly fertile plants with 2*n* = 24 suggests that “compensating trisomy” may be operating in natural populations of both *Tragopogon* allotetraploids. Compensating trisomy was first described by Blakeslee to refer to a situation in which the loss of a normal chromosome is compensated by the presence of the two arms in new translocated associations (secondary chromosomes). That concept was extended to include the replacement of the primary chromosome by two tertiary chromosomes, or by a secondary and a tertiary chromosome. Compensatory trisomy has been reported in *Datura stramonium* from progeny of a plant exposed to radium, and from crops, including from Poales, and tomato ; all were produced experimentally. The putative compensatory trisomy observed here may be an example from natural plants. The allotetraploids *T. mirus* and *T. miscellus* formed from their diploid progenitors within the last 80 years (perhaps even within the last 60 years) and given their biennial habit, the number of generations to present is likely to be less than 40. The long-term outcome of meiotic irregularities in these species is not easily predicted. On the one hand, the genomic imbalance will reduce fitness and perpetuate cycles of meiotic irregularities. This may lead to a cascade of reduced fitness, generation upon generation. Such a phenomenon was observed in synthetic *Brassica* allopolyploids maintained by single seed descent. In the wild, cryptic karyotypic instability manifest by aberrant ratios of homeologous chromosomes might ultimately lead to a slow reduction in fitness and ultimately extinction. Perhaps this accounts for the loss of some local populations of both *T. mirus* and *T. miscellus* and of some recently formed *Senecio* polyploid populations. Nonetheless, it is noteworthy that despite meiotic irregularities in the initial synthetic S₀ plants and subsequent S₁ generations, pollen fertility and seed set were generally high in the synthetic *Tragopogon* polyploid lines (Tate *et al*., in prep.). In addition, selection will favour the most fertile individuals, likely to be those with the most regular chromosome pairing. Thus, if the population can expand through early bottlenecks of reduced fertility, the derived populations are likely to be more fertile with regular bivalent pairing. Certainly in well-established allotetraploids (10⁴–10⁵ myrs old), e.g. of *Nicotiana* and *Triticum*, no major imbalances in chromosome numbers or the distribution of chromosomes to subgenomes are normally observed. ## Genomic instability The angiosperm genome is characterised by its plasticity to genetic change, including large-scale chromosome number changes, aneuploidy and polyploidy. However, it may be significant that in *Tragopogon* polyploids, all the imbalances in parental chromosome dosages between individuals occurred within a near-regular karyotype of 2*n* = 24 (one exception at 2*n* = 23). Given the frequency of plants showing genomic imbalance we might expect to find more plants with unexpected chromosome numbers. Perhaps there is selection against plants that deviate from 2*n* = 24 and against those that are nullisomic for a particular chromosome, or perhaps our sample size was too small to find a representative range of abnormalities. GISH data revealed that in two out of three individuals of *T. miscellus* and four of nine individuals of *T. mirus,* there is an imbalance in the parental contribution of the chromosomes. However, the imbalance observed resulted from monosomy or trisomy, and no individual was nullisomic for a particular chromosome. An analysis of 10 genes in *T. miscellus* using genomic and cDNA CAPS revealed that 65% of individuals displayed losses of one of the two homeologues and a further 5% of individuals showed silencing of one of the two homeologues in leaves. The missing alleles were interpreted as sequences that had been stochastically eliminated from the *T. miscellus* genome. The results here may suggest that chromosome loss or non-reciprocal translocations may contribute to the loss of alleles, although we did not find any example of a homozygous deletion for a chromosome segment. The differential expression of rDNA on the A^du chromosomes of two *T. mirus* individuals points to epigenetic or genetic heterozygosity between the homologous chromosomes. Large-scale genetic changes caused by parental genome imbalance will influence the inheritance of genetic markers, which will in turn influence the transcriptome, proteome and metabolome. Clearly genetic analyses of young or synthetic polyploids require in-depth cytogenetic studies to assess the contribution that chromosomal changes play in the inheritance of genetic markers. Unfortunately in recent years that work has seldom been standard practice. Such cytogenetic data are clearly needed even if chromosome counts appear regular \[see also 43\], since a deeper analysis of the genome and chromosome substructure can reveal substantial chromosome dosage deviation from expectation. # Materials and Methods ## Plant material Seeds of *Tragopogon* were collected from natural populations in Idaho (ID) and Washington (WA) (USA) and planted either in a greenhouse at the Department of Botany, University of Florida, or in field plots at the Institute of Biophysics, Academy of Sciences of the Czech Republic, Brno. Root tips from young, healthy, vigorously growing plants were harvested and placed in ice cold and saturated aqueous 2mM 8-hydroxyquinoline (Sigma-Aldrich Company Ltd, Poole, Dorset, UK). After 60 min incubation on ice, the roots were fixed in ethanol-acetic acid (3:1) at room temperature overnight (several washes) and stored in the same solution at −20°C until use. Developing anthers of 1 mm length or less, that contain diplotene nuclei, were excised from young buds of length 1 cm or less. Root tip and meiotic material were fixed in freshly prepared 3:1 ethanol: glacial acetic acid for two days. Root tip material was then transferred to 90% ethanol at −20°C for long-term storage. ## Making synthetic Tragopogon allotetraploids Descriptions of the methods used to generate synthetic polyploid lines are given in detail in a separate paper reporting the formation and availability of these lines for research (Tate *et al*. in prep). Briefly, numerous repeated *T. dubius*×*T. porrifolius* crosses were made, and seeds from successful crosses germinated on moist filter paper. The chromosome number was doubled to resynthesise polyploids that closely resemble *T. mirus* by placing seedlings with fully emerged cotyledons in 0.1% or 0.25% colchicine solution overnight. After washing with water for two-three days, the seedlings were transferred to 2.5” pots with soil (and grown in the glasshouse at the University of Florida under standard conditions). Plants coded N197-3.132-1 and N197-4.98-1 were used for meiotic analyses (S₀ generation) and root tips of selfed progeny (S₁ generation) analysed in root tip mitosis, metaphases of plants coded 134-16-3 and 73-14 are shown. ## Feulgen staining of meiotic cells For meiotic squashes, small developing inflorescences (1 cm or less in length) were collected from the greenhouse and fixed in 3:1 (as above, for root tips). Individual anthers were then removed and macerated in 60% glacial acetic acid, stained in aceto-orcein, spread under a coverslip by warming over a naked flame, and observed using a Zeiss Photomicroscope III. At least 15 meiotic cells per plant were scored. ## Preparing cell spreads for in situ hybridisation Chromosome preparations for FISH, using either cloned probes or total genomic DNA probes (called GISH) were made using modifications of established methods. Briefly, root tips or developing anthers were digested in 0.3% (w/v) cellulase Onozuka R-10 (Apollo Scientific Ltd, Stockport, Cheshire, UK), 0.3% (w/v) pectolyase Y23 (MP Biomedicals, Solon, Ohio, USA) and 0.3% (w/v) drieselase (Sigma-Aldrich Company Ltd., Poole, Dorset, UK) for 28 min and transferred to 1% citrate buffer for 2 h. The meristematic cells behind the root cap were isolated in a drop of 60% acetic acid and squashed onto a glass slide. For meiotic preparation, fixed anthers were dissected and meiotic cells gently dispersed into a drop of 60% acetic acid, a coverslip applied and gently warmed. Coverslips were removed following freezing with dry ice. ## *In situ* hybridisation Fluorescence *in situ* hybridisation followed standard protocols,, Leitch *et al*., Telgmann-Rauber *et al*.. Genomic DNA from *T*. *dubius*, *T*. *pratensis*, and *T*. *porrifolius* for labelling by GISH was extracted using DNeasy Plant mini kit (Qiagen Ltd, Crawley, West Sussex, UK) following manufacturer's instructions. Genomic DNA was labelled with biotin-16 dUTP (Sigma-Aldrich Company Ltd., Poole, Dorset, UK) or digoxigenin-11-dUTP (GE Healthcare, Chalfont St Giles, Buckinghamshire, UK) using nick translation following standard protocols. The probe against 5S ribosomal DNA (rDNA) was prepared by amplifying the gene using primers described in Fulnecek et al. and biotin-16-dUTP-labelling protocol as described in Leitch et al.. The probe against 45S rDNA was the clone pTa71, which includes the 18- 26S rDNA subunit isolated from *Triticum aestivum*, which was labelled with digoxignenin-11-dUTP as described in Leitch et al.. Briefly, slides were denatured in 70% (v/v) formamide in 2× SSC (0.3 M sodium chloride, 0.03 M sodium citrate) at 70°C for 2 min and the hybridisation mix added (4 µg.ml⁻¹ labeled probes and 50% (v/v) formamide, 10% (w/v) dextran sulphate, 0.1% (w/v) sodium dodecyl sulphate in 2× SSC). *In situ* hybridisation was carried out overnight at 37°C, after which the slides were given a stringent wash (20% (v/v) formamide in 0.1× SSC at 42°C). Sites of probe hybridisation were detected using 20 µg.ml⁻¹ fluorescein-conjugated anti-digoxigenin IgG (GE Healthcare, Chalfont St Giles, Buckinghamshire, UK) and 5 µg.ml⁻¹ Cy3-conjugated avidin (Roche Pharmaceuticals, Lewes, East Sussex, UK) in 4× SSC containing 0.2% (v/v) Tween 20 and 5% (w/v) bovine serum albumin. Chromosomes were counterstained with 2 µg/ml DAPI (4′,6-diamidino-2-phenylindole, Sigma Aldrich Company Ltd., in 4× SSC) and stabilised in Vectashield medium (Vector Laboratories Ltd, Peterborough, UK) prior to data acquisition using either: 1) Leica DMRA2 epifluorescent microscope fitted with an Orca ER camera and Open Lab software® (Improvision, Coventry, UK) or; 2) Olympus BX61 epiflurescent microscope using Olympus Microsuite 5 software® (Olympus America Inc, Center Valley, PA, USA). The images were analysed with Adobe Photoshop® version 7 and treated for colour contrast and uniform brightness only. At least 5 mitotic or meiotic cells per plant were scored with each probe used. ## Southern hybridisation *Taq*I restriction enzyme digestion was carried out on genomic DNA extracted from leaves using standard protocols with modifications as in. After fractionation in 1% agarose by gel electrophoresis, the DNA was transferred to a Hybond N+ membrane (GE Healthcare, Chalfont St Giles, Buckinghamshire, UK). The ³²P-labelled 5S rDNA probe was from a 120-bp *Xba*I/*Eco*RI fragment of 5S rDNA cloned from *N. tabacum*. Probe hybridisation was conducted under high-stringency conditions in the Church-Gilbert hybridisation buffer at 65 °C overnight. The radioactivity signals were quantified by phosphorimager scanning (Storm, GE Healthcare, UK). We thank Mr. R. Joseph for assistance. [^1]: Conceived and designed the experiments: DS PS AK AL. Performed the experiments: KYL DS JT RM HS ZX AL. Analyzed the data: KYL PS RM HS AK CP ZX AL. Contributed reagents/materials/analysis tools: DS PS AK CP AL. Wrote the paper: DS PS JT AK CP AL. [^2]: The authors have declared that no competing interests exist.

# Introduction Leptin is a secreted adipocytokine that regulates energy expenditure. Centrally, leptin acts on hypothalamic circuits to inhibit food intake. Peripherally, it stimulates immune and barrier cells to promote activation, proliferation, resistance to cell death, and wound repair. Leptin is known to engage the cytokine receptor homology-2 (CRH2) domain in the extracellular region of the leptin receptor (LepR) with high affinity. Upon leptin binding and clustering of activated receptor complexes, Janus Kinase 2 (JAK2) phosphorylates intracellular tyrosine residues, which recruit transcription factors: Src Homology Phosphatase 2 (SHP2), Signal Transduscer and Activator of Transcription 5 (STAT5), and STAT3. SHP2 positively regulates the Extracellular Signal-Regulated Kinase/c-fos (ERK/c-fos) pathway. Activated and dimerized STAT3 translocates to the nucleus where it initiates a transcriptional program, which includes the upregulation of SOCS3, a feedback inhibitor of leptin signaling. The Q223R encoding SNP (rs1137101) is exceedingly common and widely distributed. According to data from the 1000 genomes project, the overall allelic frequency is 41% A, encoding glutamine, and 59% G, encoding arginine. Interestingly, frequencies vary significantly by region and ethnicity. Numerous studies have found modest or strong association of this variant with obesity and adiposity, multiple forms of cancer, peritonitis, adverse drug reactions and susceptibility to enteric – and respiratory infections after controlling for ethnicity, age, sex, and environmental factors. However, studies in separate populations have found no significant association for several of these conditions. In 2011, a meta-analysis of published findings on the Q223R variant suggested that heterogeneity in association between this mutation and body weight in human populations may be attributable to differences in study design and power. LepR contains two extracellular CRH domains separated by an immunoglobulin-like domain. Two fibronectin type 3 (FN III) domains separate the CRH2 domain from a single transmembrane region and a long intracellular signaling region containing a membrane-proximal box 1 motif. The LepR Q223R mutation results from an A to G transversion encoding a glutamine to arginine substitution in the N-terminal CRH1 domain. It has been shown that the CRH2 domain is both necessary and sufficient for leptin binding; while CRH1 is dispensable for a high affinity interaction. To test the hypothesis that the Q223R amino acid replacement affects ligand- binding kinetics of the leptin receptor, we measured the leptin/LepR interaction using SPR technology on a BiaCore T200 platform. To measure leptin association and dissociation for both the mutant and wild type receptors, the extracellular domain of each receptor was expressed as an Fc-fused chimera and immobilized to a dextran matrix coated with anti-IgG antibody. Increasing concentrations of leptin were then injected into the system. The rates of binding and dissociation were monitored at each leptin concentration in real-time as changes in the refractory index of the capture surface measured in response units (RUs). The ability to monitor these interactions in real-time allowed for the determination of binding kinetic constants by global fitting to a model of 1∶1 ligand to receptor binding. # Materials and Methods ## Tissue Culture, Transfection, and Concentration HEK293T/17 cells were maintained in a humidified incubator at 37°C with 5% CO₂ and cultured in HEPES buffered Dulbecco's Modified Eagle's Medium-F12 (Gibco) supplemented with 10% heat-inactivated fetal bovine serum (FBS). pMET7 constructs encoding the WT or Q223R extracellular domain of the murine LepR, C-terminally fused to the Fc region of murine IgG1, a FLAG-tag, and a 6x His tag were obtained from the Cytokine Receptor Lab at Ghent University. Sequences of the WT and Q223R constructs were confirmed by Sanger sequencing with T7 primers provided by Genewiz, inc. When HEK293T/17 cells were between 70 and 90% confluent, transfection using Lipofectamine 2000 (Invitrogen) was performed per the product instructions using WT, Q223R, or empty vector plasmid DNA. 12 hours after transfection, monolayers were washed and growth medium was replaced with serum free Optimem (Gibco) supplemented with 2 mM sodium butyrate. 48 to 72 hours later, supernatants were harvested and cleared of cellular debris by centrifugation. Cleared supernatants were concentrated 20x in Amicon Ultra 15 ml centrifugal filter devices with a nominal molecular weight limit of 100 kDa at 3000×*g* for 15 minutes. Buffer exchange to 1x PBS was performed following the instructions of the filtration device manufacturers (Millipore) before BiaCore analysis. ## Immunoblotting Supernatants were heated to 95°C in 4 x SDS-PAGE sample buffer for 5 minutes and separated by molecular weight using SDS-PAGE in Mini-PROTEAN TGX Precast Gels (4–20%). Replicate gels were either stained using Coomassie blue or proteins were transferred to polyvinylidene difluoride (PVDF) membranes by standard wet transfer methods. PVDF membranes were blocked with 5% milk in tris-buffered saline−0.05% Tween 20 (TBST) for 1 hr. Blots were either probed in two steps using anti-murine LepR (R&D) and a horseradish peroxidase (HRP)-conjugated secondary antibody (Sigma Aldrich) or in one step using HRP-conjugated anti- mouse IgG1 (Sigma Aldrich). Ab specific HRP-conjugated anti-mouse IgG1 was used as a control for non-specific binding to residual IgG components in the expression medium. Blots were washed 3 times for 5 minutes in TBST between probing steps and antibody bound proteins were visualized with ECL reagents (Pierce). Stained gels were imaged using a typhoon fluorescent imager (GE). ## Surface Plasmon Resonance The BiaCore biosensor T200, CM5 biosensor chips, N-hydroxysuccinimide (NHS), N-ethyl-N-(3-diethylaminopropyl)carbodimid (EDC), ethanolamine-HCl, and HBS-EP buffer were obtained from BiaCore AB (GE). Rabbit polyclonal anti-mouse IgG1 was purchased from GE Healthcare. Recombinant murine leptin (MW 16 kDa) was obtained from R&D Systems Europe, Ltd. Anti-mouse IgG1 was immobilized on two channels of a carboxymethyl dextran chip (CM5) by amine coupling to the dextran matrix, activated with NHS/EDC for 5 minutes. The antibody was injected at 200 µg/ml in sodium acetate buffer (10 mM, pH 4.0). Approximately 2000 resonance units (RU) of antibody were coupled in each channel for each experiment. Blocking of the activated surface was achieved with a 5 minute injection of 1 M ethanolamine (pH 8.5). A typical run at 25°C involved the injection of 100-200 µl of Amicon- concentrated extracellular murine leptin receptor Fc (LepR_ec-Fc) chimera to achieve 100–300 RU of immobilized receptor. The surface was equilibrated for 6 minutes prior to leptin or HBS-EP buffer injection. A range of leptin concentrations (1.25–60 nM) was tested for each receptor chimera and their association was monitored for 600 seconds. The dissociation phase was monitored for 1000 seconds before regeneration of the surface using 10 mM glycine, pH 2.0. The flow rate was set to 30 µl/min. To monitor nonspecific binding, an antibody-coated channel lacking bound LepR_ec-Fc was run with leptin and HBS-EP buffer. Sensorgrams were generated from reference- subtracted leptin binding data for each receptor. ## Statistics and Data Analysis Kinetic parameters were derived using BIA evaluation software 3.1 (BiaCore AB). A non-linear least squares analysis model for 1∶1 binding was applied to fit data from association and dissociation phases simultaneously and globally across all leptin concentrations tested. Means and SEM were calculated from two separate experiments at the indicated concentrations of leptin. A student's t-test was performed on rates of association and dissociation for each receptor. # Results ## Expression and concentration of LepR_ec-Fc chimeras Q223R and WT murine LepR_ec-Fc chimeras were expressed in the supernatants of adherent HEK293T/17 cells. Each preparation was concentrated by ultrafiltration with a nominal molecular mass limit of 100 kDa. Because concentrated preparations used in BiaCore experiments were not affinity purified, we modified the BiaCore immobilization step to use anti-murine IgG1 as the capture molecule instead of direct coupling. Murine leptin (the analyte) was obtained from R&D and resuspended in phosphate buffered saline at 1 mg/ml before dilution in BiaCore running buffer to the appropriate concentrations. shows analysis of expression (a) and concentration steps (b,c) in the preparation of the LepR_ec-Fc chimeras. ## Binding kinetic analysis of WT and Q223R LepR_ec-Fc with recombinant leptin Kinetic parameters were derived from global Langmuir fitting for 1∶1 binding kinetics to sensorgrams of leptin-LepR binding at 1.25, 2.5, 5, 6.25, 10, and 60 nM murine leptin. Values for the on rate (k_a), off rate (k_d), and dissociation constant (K_D) were derived from global fits from two separate experiments using independently prepared protein for the WT and Q223R chimera receptors. shows representative sensorgrams depicting leptin binding and dissociation with reference surface background subtracted alone and overlaid with globally fitted curves to the 1∶1 model of leptin binding for WT **and** Q223R. Binding and dissociation are shown in response units (RU). Because the preparations containing the WT receptor contained less protein than the Q223R preparations, less of the WT chimera was immobilized before leptin binding and a lower maximum concentration was bound during the kinetic runs. However, sufficient receptor was immobilized to derive kinetic constants for each run and differences of this magnitude are unlikely to alter kinetic determinations. Two subnanomolar affinity interactions were detected and monitored in our system to derive kinetic constants for each receptor. Standard errors were calculated and student's t-tests performed on data from two separate experiments for the three parameters. Derived constants are summarized in. No significant difference was measured and we concluded that the Q223R amino acid replacement does not significantly alter 1∶1 binding kinetics of murine leptin binding to the murine leptin receptor. # Discussion The most important finding of this work is that the common Q223R encoding SNP in the LepR extracellular domain does not affect the rates at which leptin binds to and dissociates from its receptor. We investigated whether the Q223R substitution could alter ligand binding using surface plasmon resonance (SPR). Because it is both common and non-conservative, the Q223R amino acid replacement has been among the most studied variants of LepR. However, the functional consequences of this variant remain poorly defined. This mutation occurs in the CRH1 domain, which is dispensable for high affinity leptin binding, but may have roles in downstream signaling, receptor trafficking, or surface expression. Stratigopoulos et al. observed no difference in adiposity between WT and Q223R isogenic mice fed a high fat diet. A recent study found no significant association between this LepR polymorphism and obesity in humans. However, the homozygous Q223R encoding genotype did correlate strongly with increased serum cholesterol and low density lipoprotein (LDL) in both obese and non-obese subjects. In addition, our earlier work has demonstrated a dramatic association between the human Q223R LepR mutation and amebiasis, a common cause of diarrheal morbidity and mortality among children globally. Furthermore, patients carrying the Q223R mutation are at increased risk of *Clostridium difficile* colitis. SPR was used to measure, in real-time, the interaction between murine leptin and the extracellular domain of murine LepR with or without the Q223R mutation. Until recently, the two primary models of leptin binding have suggested leptin:LepR stoichiometries of 2∶2 or 2∶4. However, a single particle electron microscopy study by Mancour et al. provided strong evidence for a ligand-induced native 2∶2 quaternary structure for leptin and its receptor. We sought to address our primary question in a minimalist fashion, focusing on the interaction between one extracellular domain of the leptin receptor and one leptin molecule. Future studies will seek to investigate the role of the Q223R mutation in oligomerization of the extracellular domain, which could explain mechanistically attenuated leptin signaling observed in cells transfected with this variant. Using SPR, Mistrik et al. measured values similar, but not identical, to our own for k_a, k_d, and K_D with murine leptin and murine extracellulary LepR bivalently fused to an Fc domain (1.5×10⁶ M⁻¹s⁻¹, 7×10⁶ s⁻¹, and 0.5 nM respectively). Alternative approaches using radioligand binding have also yielded dissociation constants in the sub-nanomolar range. However, Mancour et al. measured a markedly different K_D between murine leptin and the extracellular domain of murine LepR (17 nM) using isothermal titration calorimetry. Studies specifically comparing SPR to ITC have yielded comparable dissociation constants for multiple molecular interactions. However, a systemic comparison of techniques will be required to satisfactorily address discrepancies in parameters derived by these two technologies for the leptin receptor. One limitation of SPR is the challenge of modeling molecular interactions with low dissociation rates. The difference between the K_D value that we measure and that determined using SPR by Mistrik et al. is due largely to differences in measured dissociation rates (k_d). Even after 1000 seconds of monitoring in our system, only a fraction of bound ligand had dissociated from both mutant and WT receptor. This observation could in part be due to rebinding of leptin during the dissociation phase, a common confounding variable in SPR. It is also possible that a small contribution of mass transport in our study complicated accurate derivation of kinetic constants. Despite these limitations, the key finding of our work was internally consistent and indeed supports the conclusion that no significant difference in binding kinetics exists between Q223R and WT LepR. It remains possible that this mutation alters cell-signaling networks by affecting surface expression, oligomerization kinetics, or receptor turnover. In fact, we have shown previously using a STAT3 driven luciferase reporter assay that cells transfected with the Q223R receptor exhibit lower signaling via STAT3 relative to cells expressing WT LepR. Furthermore, Zabeau et al., showed that deletion of the entire CRH1 domain altered optimal downstream signaling via JAK2. The current study demonstrates clearly that these signaling phenotypes are unlikely to arise from kinetics of receptor-ligand binding and represents a step toward understanding how a mutation in the distal CRH1 domain of the leptin receptor affects human health and disease. We deeply appreciate the administrative and technical assistance provided by Dr. Daniel Conrad at Virginia Commonwealth University [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HPV CN LZ JT WAP CM. Performed the experiments: HPV CN. Analyzed the data: HPV CM WAP. Contributed reagents/materials/analysis tools: LZ JT. Wrote the paper: HPV CM. Provided revision and essential analysis during the writing stage: CN LZ JT WAP.

# Introduction *Toxoplasma gondii* and *Neospora caninum* are closely related coccidians that belong to a diverse group of parasitic protozoans of the phylum Apicomplexa. *T*. *gondii* is a zoonotic intracellular parasite capable of infecting any warm-blooded vertebrate. Estimated to infect around one-third of the human population, *T*. *gondii* causes neonatal mortality in humans and veterinary hosts and opportunistic infections in immuno-compromised individuals. Apart from being a significant pathogen in its own right, *T*. *gondii* has gained more attention as a laboratory model to study Apicomplexan biology because it can be easily propagated and is experimentally tractable. *N*. *caninum*, though similar to *T*. *gondii* in morphology, has a narrow host range of primarily cattle and dogs and is the leading cause of abortion in cattle resulting in significant economic losses. The difference in host range of these two closely related species (having highly conserved gene content) has been of particular interest in studying molecular mechanisms behind host specificity in Apicomplexans. Whole genome sequencing has been performed for 17 *T*. *gondii* strains till date that are publicly available in ToxoDb (<http://www.toxodb.org>). Of these, three prototypic strains- GT1 (Type I), ME49 (Type II) and VEG (Type III) serve as annotated reference genomes and are widely used as a basis for *Toxoplasma* research worldwide. The genome of *N*. *caninum* Liverpool strain was sequenced and annotated recently. Both *T*. *gondii* and *N*. *caninum* genomes are approximately 62 million basepairs (Mb), divided into 14 chromosomes and contain around 7,500 predicted protein-coding genes. Primarily, automated gene prediction tools such as TigrScan, HMMGlimmer, TwinScan, SNAP and AUGUSTUS have been used for genome annotation in coccidian genomes. These *ab initio* gene predictors use a variety of genome level information such as codon frequencies, GC content and a small set of curated gene structures as a ‘training set’ to predict coding sequences and exon-intron boundaries. However, such predictors suffer from several limitations that compromise the accuracy of the predicted gene models. Genome annotation is the fundamental resource that links the genome to relevant biological information ranging from sequence information such as coding sequences, gene exon-intron boundaries and untranslated regions to functional information such as transcriptional start sites (TSSs), coding and non-coding RNA transcripts, and alternative splicing. Therefore, it is imperative to improve the quality of genome annotation because gene models form the basis of future research in an organism and errors in them can compromise downstream analyses. With the advent of high-throughput RNA sequencing and proteomic technologies, several annotation pipelines such as JIGSAW, AUGUSTUS and PASA have been developed that use several lines of experimental evidence in addition to *ab initio* gene predictions to significantly improve accuracy of predicted gene models. In fact, during the course of this work, parallel efforts using such computational tools have significantly improved the assembly and annotation quality of *T*. *gondii* ME49 genome. In a previous comparative analysis of the *N*. *caninum* and *T*. *gondii* genomes, we have manually validated the organism–specific genes using tachyzoite stage transcriptomes (RNA-seq). However, there was no systematic effort to manually curate the predicted gene models using the RNA-seq datasets. Systematic manual curation of the gene models through careful inspection of the predicted gene structures in the light of several layers of experimental evidence (e.g. strand-specific RNA-seq, mass-spectrometry) leads to highly accurate gene models. Towards this goal of enriching and improving the genome annotation of these two Apicomplexan parasites, we resequenced the genomes of *Toxoplasma gondii* VEG strain (*Tg*VEG) and *Neospora caninum* LIV strain (*Nc*LIV) to close gaps and correct erroneous base-calls in the genome assembly. We performed strand-specific RNA sequencing (ssRNA-seq) and shotgun proteomics to get transcript and protein level evidences for curating gene models and predicting untranslated regions. Further, we annotated a set of putative natural antisense transcripts (cis-NATs) and long intergenic non-coding RNAs (lincRNAs) based on ssRNA-seq. # Materials and Methods ## Parasite cultivation Tachyzoites of *Toxoplasma gondii* VEG and *Neospora caninum* LIV strains were maintained in confluent layers of African Green monkey kidney cells (Vero) (ECACC, Salisbury, UK). The tachyzoites were harvested 3 or 4 days post- infection as described previously. ## Genome Sequencing Genomic DNA was sheared using Covaris-E series. The sheared fragments were made into libraries following Truseq DNA LT kit protocol and size-selected in agarose gel for 200–300 bp insert size. Sequencing of the DNA libraries was done on an Illumina HiSeq 2000 with paired-end 100bp read chemistry. ## Genome improvement Illumina reads were re-assembled with *T*. *gondii* VEG and *N*. *caninum* LIV genomes (ToxoDb v.8.0) as references. Sequence gaps were filled using Gapfiller (v1.10) run for two 7 iterations with *k-mer* size set to 31. Single base errors and indels were corrected using iCORN run for 5 iterations. The assemblies were evaluated by re-mapping reads onto the original and new assemblies and running REAPR to compare the number of perfectly mapped reads. Gene annotations were transferred to the new assemblies using RATT with “Transfer type” parameter as “Strain” (contains predefined nucmer parameters set for determining synteny between strains with 95–99% genome similarity). The *Tg*VEG and *Nc*LIV reads alignment bam files (ENA run IDs- ERR701180 and ERR701181) are available from the European Nucleotide Archive (ENA; <http://www.ebi.ac.uk/ena/>). ## Strand-specific RNA sequencing Total RNA was extracted from day 3 and day 4 tachyzoite stage parasites using TRIzol method as described previously in. Strand-specific RNA sequencing was performed from total RNA using TruSeq Stranded mRNA Sample Prep Kit LT according to manufacturer's instructions. Briefly, polyA+ RNA was purified from total RNA using oligo-dT dynabead selection. 1st strand cDNA was synthesized using randomly primed oligos followed by 2nd strand synthesis where dUTPs were incorporated to achieve strand-specificity. The cDNA was adapter-ligated and the libraries amplified by PCR. Libraries were sequenced on an Illumina HiSeq 2000 with paired-end 100bp read chemistry. The raw reads are available in ENA (accession numbers- ERR690605, ERR690606, ERR690607 and ERR690608). Raw reads were aligned onto the genomes using TopHat version 2.0.9 and transcripts were assembled using Cufflinks version 2.1.1. ## Whole cell proteome analysis Tachyzoite proteins were denatured with SDS, resolved in a 1-D polyacrylamide gel and in-gel tryptic digested as described in. LC-MS/MS analysis of the digested peptides was carried out by Proxeon EASY-nLC unit (Bruker Daltonik, Germany) through a capillary column (0.1 × 150 mm, with C18 AQ of 3 μm particles and 200 pore size, Michrom42 BioResources) with a 60 min gradient of 2−40% solvent B (solvent A: 100% H₂O; solvent B: 90% ACN, 0.1% formic acid). This system was connected to a LTQ-Orbitrap Velos (Thermo Scientific, Germany). Elution occurred at a flow rate of 500 nl/min and ionization was performed using an applied voltage of 1.5 kV to the emitter. Data were acquired using Xcalibur software in data- dependent mode to facilitate automatic switching between MS and MS2. The precursor ion scan MS spectra (m/z 300−1600) were acquired in the Orbitrap with resolution R = 60,000 at m/z 400 with the number of accumulated ions being 1 × 10⁵. The ten most intense ions were isolated and fragmented in the linear ion trap (number of accumulated ions; 10⁵). The resulting fragment ions were recorded in the Orbitrap with resolution R 15,000 at m/z 400. The lock mass option enabled accurate mass measurements in both MS and MS/MS mode. The poly-dimethylcyclosiloxane ions generated in the electrospray process from ambient air (protonated (Si(CH₃)₂O)₆, m/z 445.120025) were used for internal recalibration in real time. In data-dependent LC-MS/MS experiments dynamic exclusion was used with 45 sec exclusion duration. Tandem mass spectra were extracted by Xcalibur software. All MS/MS samples were analyzed using Mascot (Matrix Science, London, UK; version 2.4.0) against a concatenated organism-specific target-decoy database of i) protein-coding sequences; ii) protein sequences of all possible ORFs longer than 30 amino acids in all six reading frames of the genome and iii) protein-coding sequences of *Macaca mullata* (Rhesus monkey) to eliminate host cell contaminant peptides. Scaffold (version Scaffold_3.6.2, Proteome Software Inc., Portland, OR) was used to validate MS/MS based peptide and protein identifications. Peptide identifications were accepted if they could be established at greater than 95.0% probability as specified by the Peptide Prophet algorithm. Protein identifications were accepted if they could be established at greater than 99.0% probability and contained at least 1 uniquely identified peptide. Protein probabilities were assigned by Protein Prophet algorithm. The identified peptides were mapped onto the genomes using TBLASTN (exact matches). Identification and mapping were repeated again after manual curation to determine the final set of protein identifications. ## Manual curation, UTRs and long non-coding RNA prediction Gene models and different tracks of evidence were visualized and curated in Artemis genome visualization tool and Artemis Comparison Tool. We curated the gene models in both the genomes based on three tracks of evidence in the following order of importance—1) Transcript evidence from ssRNA-seq, 2) Peptide evidence from mass spectrometry based shotgun proteomics and 3) protein level sequence similarity (TBLASTX) between *T*. *gondii* and *N*. *caninum* genomes. Our manual curated gene models can be broadly classified into five categories- 1. Corrected gene models- We visually compared each gene model in both the genomes with the mapped ssRNA-seq paired-end reads. We modified only those gene models that had clear transcript-level evidence of exonic and intronic regions. We corrected exon-intron boundaries (defined by AT and GT/GC start and end sites respectively), added exons that were missing in transcript- level exonic regions, deleted exons that were present in transcript-level intronic regions and extended or truncated gene boundaries. These corrections also included exons that were wrongly predicted based on transcript reads on the opposite strand. Since ssRNA-seq cannot resolve translation start sites, we encountered cases where i) more than one frame of translation was possible for a gene, and ii) multiple transcriptional start sites were possible due to more than one start codon in the same reading frame. We were able to resolve these ambiguities in genes with peptide level evidence. By mapping peptide sequences, we were able to pinpoint the correct frame of translation and regions where genes should be extended to an upstream start codon. In other cases, the foremost start codon within the transcript was chosen as the translational start site. 2. Merged gene models- If a continuous transcript spanned across two or more predicted gene models with clear evidence of splice junctions between them, we identified them as incomplete gene models and merged them into a single gene. 3. Split gene models- In some gene models, it was evident that there were two or more separate neighboring transcripts from distinct lack of reads or splice junctions between them. We split and corrected them into separate genes. 4. New gene models- We predicted new genes in ORFs that had clear RNA-seq evidence and some of the highly expressed new genes also showed peptide evidence. 5. Deleted gene models- We considered a gene as spurious if it lacked RNA- seq evidence and overlapped with another RNA-seq supported curated gene, in the same or opposite strand. We did not delete non-overlapping genes that lacked RNA-seq evidence, as our RNA-seq data is only from tachyzoite stage and these genes might be expressed in other life stages and hence cannot be confidently called "spurious". We also compared the gene models of orthologs in both the genomes during the curation process. This particularly helped in correcting genes with low or incomplete RNA-seq evidence and improved our confidence in newly created genes that had an ortholog in *T*. *gondii* or *N*. *caninum*. To assess annotation quality, amino acid sequences of the publicly available set and our curated set of gene models were scanned for known protein domains using InterProScan v4.6. Full-length transcripts were assembled using Cufflinks version 2.1.1 with default parameters. UTRs were predicted by comparing these transcripts and manually curated gene models using in-house Perl scripts. For each gene, the regions between start/stop of the transcript model and the start/stop of the manually curated gene model were inferred as UTRs. Polyadenylated non-coding RNA transcripts were predicted by adapting the protocol described in and the non-coding potential of the transcripts were calculated using Coding Potential Calculator. We assembled novel transcripts from the mapped RNA-seq reads using Cufflinks version 2.1.1 and compared with the manually annotated gene models using cuffcompare. Putative lincRNA transcripts were defined as those transcripts i) \>200 base pairs long; ii) with coding potential score (calculated by CPC) lesser than −1; and iii) not overlapping with coding regions or UTRs in the same strand. Of these, transcripts that overlapped with gene models in the opposite strand with overlap region at least 100 bps were classified as anti-sense non-coding RNA (ancRNA) while the rest were classified as lincRNA. Candidates satisfying these criteria were then manually checked and only those that had uniform and distinct transcript evidence were retained. The improved sequences and annotations of *T*. *gondii* VEG and *N*. *caninum* LIV can be found at ENA *(*Accession numbers between LN714489 and LN714514 for *Tg*VEG and between LN714474 and LN714488 for *Nc*LIV sequence and annotation). # Results and Discussion ## Genome improvement and manual evaluation of gene models We improved the genome sequences by closing 111 and 39 gaps respectively in *Tg*VEG and *Nc*LIV nuclear genomes. We added 25,781 and 13,574 nucleotides in *Tg*VEG and *Nc*LIV genome assemblies respectively. Closure of gaps within genes and at exon-intron boundaries improved RNA-seq read mapping at these loci. This helped us identify incorrect splice sites and frame-shifts giving rise to erroneous gene models. Then we manually verified the gene models of more than 7,000 genes each in *Tg*VEG and *Nc*LIV genomes, with RNA-seq and peptide evidence. We corrected 2,497 and 2,495 genes in *Tg*VEG and *Nc*LIV respectively, accounting for almost one-third of the genes. 401 *Tg*VEG and 90 *Nc*LIV genes were merged into a single gene, and 120 *Tg*VEG and 135 *Nc*LIV genes were split into two or three genes. A total of 560 and 200 new genes were created in *Tg*VEG and *Nc*LIV respectively while 603 and 35 genes were found spurious and deleted. ## Mass spectrometry and proteogenomic annotation Controlling the target-decoy false discovery rate (FDR) at 5% and filtering for proteins with at least one unique peptide, we identified a total of 2,142 *Tg*VEG proteins with an estimated target-decoy FDR of 0.1% and 0.5% at the peptide and protein level respectively. Similarly, we identified 1,956 proteins in *Nc*LIV, albeit with a higher estimated peptide and protein FDR of 0.4% and 4.9%. Though our criterion for protein identification was at least one uniquely identified peptide, we observed that only 10% of the proteins identified were based on single peptide evidence with the rest having at least two or more peptides assigned to them. Also, we observed a good representation of *Tg*VEG-*Nc*LIV orthologs in the dataset, with 1,513 orthologous genes having peptide evidence in both the organisms, which proved useful for effective proteogenomic annotation. We mapped 16,737 and 14,777 unique identified peptides (exact matches) onto *Tg*VEG and *Nc*LIV genomes respectively and these consisted of either peptides mapping to protein-coding sequences or peptides mapping only to six-frame translated ORFs and not to annotated protein-coding regions (called “orphan” peptides). We were able to correct 13 *Tg*VEG and 30 *Nc*LIV gene models and create 16 *Tg*VEG and 25 *Nc*LIV new genes based on these “orphan” peptides that are also supported by RNA-seq evidence. After our manual curation, only 22 *Tg*VEG and 33 *Nc*LIV “orphan” peptides still remained unaccounted for in the genomes. Of these, 8 *Tg*VEG and 16 *Nc*LIV peptides were mapping to intergenic regions while the rest emerged from intronic regions, being possible products of alternate splicing events. Our proteomics data provides mass spectrometry evidence for 51 additional *T*. *gondii* proteins that are currently available in ToxoDb v8.0. The importance of integrating proteome data in genome annotation of *T*. *gondii* was clearly shown in a multi-platform global proteomic analysis, where 15% of the peptides mapped to alternate gene models and marked discrepancies in the annotated gene models such as wrong splice junctions and wrong frames of translation. We corrected similar discrepancies in our curation process reducing the final “orphan” peptides to just 0.001% of the \~15,000 total peptides in each genome. Peptide evidence was crucial to determine correct reading frames, translational start sites and to confirm correct splice junctions, particularly under circumstances where RNA-seq evidence was unclear. ## Annotation quality assessment We selected Pfam domain content and confidence scores of the domain hits as crude indicators of improvement in annotation quality. Tools such as InterProScan can identify known domains present in a protein and the percentage of proteins in an annotated genome with domains can indicate the annotation quality. We observed a 22% and 12% increase in domain content in *T*. *gondii* and *N*. *caninum* respectively after our manual curation. We discovered 362 and 177 new domains respectively from Pfam database v26.0 alone, of which 202 and 35 domains were present in newly created genes. Apart from the number of domain hits, the quality of these hits, reflected by their e-value scores, can also judge annotation quality. A correct gene model is expected to have better sequence alignment with a known functional domain than an incorrect model, gaining a lower e-value. In our set of corrected gene models, we recorded a greater than 10-fold decrease in e-values of Pfam matches in 223 and 349 domains in *T*. *gondii* VEG and *N*. *caninum* LIV respectively. In a small number of cases, we observed an increase in e-values because our curation led to insertions within the existing Pfam domains causing it to split into two or more domains with higher e-values. The general trend of increased confidence scores of the Pfam domains and the gain of new functional domains in the genomes together indicate a significant improvement in the annotation quality. The percentage of protein-coding genes having functional domains stand as of now at 77.89% and 78.82% in *T*. *gondii* and *N*. *caninum* which is comparable to 77.15% in the recently improved annotation of Type II *T*. *gondii* ME49 (ToxoDb v8.0). ## *T*. *gondii* and *N*. *caninum* have characteristic long untranslated regions in transcripts Based on assembled transcripts from our ssRNA-seq datasets, we annotated 4,711 5’UTRs and 4,523 3’UTRs in *Tg*VEG, with 4,232 *Tg*VEG genes gaining both 5’ and 3’UTR annotations. Similarly, we annotated 4,704 5’UTRs and 4,638 3’UTRs in *Nc*LIV, with 4,333 *Nc*LIV genes gaining both 5’ and 3’UTR annotations. We compared the UTR lengths with those in other model eukaryotes, namely *Schizosaccharomyces pombe* (genes were retrieved from PomBase, <http://www.pombase.org/>), *Arabidopsis thaliana*, *Caenorhabditis elegans*, *Drosophila melanogaster* and *Homo sapiens* (RefSeq gene annotations of all four model organisms were retrieved from UCSC Genome Browser, <http://www.genome.ucsc.edu/>). We observed that the 5’UTRs are strikingly longer in *T*. *gondii* and *N*. *caninum* (810–860 nucleotides) compared to other model eukaryotes compared in the study. Mann-Whitney test proved a highly significant increase (all p-values less than 2.22 x 10–15, Bonferroni corrected, 10 tests) in 5’UTR lengths in *T*. *gondii* and *N*. *caninum* against the model eukaryotes. 3’UTRs were similar to those in humans but longer than those in other eukaryotes. We also compared the UTR lengths separately for genes conserved and not conserved across these eukaryotes to determine whether the differences in length distributions were arising mainly from species-specific genes. We found that the UTRs were uniformly longer in *T*. *gondii* and *N*. *caninum* than in other eukaryotes independent of gene conservation. Next, we checked whether UTR sizes differed between coccidian genes (orthologs in *T*. *gondii* and *N*. *caninum*) and non-coccidian genes (orthologs in *T*. *gondii*, *N*. *caninum* and *P*. *falciparum* (PlasmoDb v9.0)). No significant variation was seen between UTR lengths in genes restricted to coccidians and the set of genes that were also shared with *Plasmodium*. To identify gene families that were enriched with long UTRs, known and predicted host-interaction genes such as rhoptry, microneme and dense granule genes, and apiAP2 family of transcription factors (apiAP2 Tfs) were retrieved from. Comparing the UTR lengths of these gene classes with the median UTR size of all genes, we observed that apiAP2 transcription factors have significantly longer 5’ and 3’ UTRs in both *T*. *gondii* and *N*. *caninum*. Of the 68 apiAP2 Tfs in each organism identified by, 46 TgVEG and 48 NcLIV genes had either or both of their UTRs annotated. Out of these, 37 *Tg*VEG and 41 *Nc*LIV apiAP2 Tfs had longer UTRs than the average UTR size in the respective genomes, and of which a majority of them were orthologs (28 apiAP2 Tfs ortholog pairs). In contrast, microneme proteins in *N*. *caninum* had significantly shorter 5’UTRs than rest of the genes. Next, we analysed size variation and sequence conservation between *T*. *gondii* and *N*. *caninum* UTRs to find whether the high level of conservation between their coding regions also extended to their UTRs. We observed that the UTR sizes were reasonably correlated between *Tg*VEG-*Nc*LIV orthologs, though less correlated as compared to the intron and coding region sizes. Kendall tau correlations were 0.54 and 0.38 (p-value less than 2.2 x 10^–16) for 5’UTRs and 3’UTRs respectively while coding regions and introns had high size correlations of 0.95 and 0.87. For sequence conservation, we selected neighboring genes with annotated UTRs and with at least 1,000 basepairs between the adjacent UTRs. UTR sequences and 500 basepair flanking intergenic sequences were compared between *T*. *gondii* and *N*. *caninum* and the percent identity (defined as the number of identical positions in the alignment per base-pair of the shorter sequence of the pair) was calculated. The UTRs were generally more conserved between the species compared to the intergenic regions with a median percent identity of \~54% compared to \~40% in intergenic regions. UTRs play an important role in post-transcriptional regulation of gene expression and modulate mRNA transport from the nucleus, their subcellular localization, translation efficiency and stability in eukaryotes. Apicomplexan parasites have complex life cycles that consist of both proliferative and latent life stages, which enable the parasite to propagate through diverse vector and host environments. The parasites regulate the transition between these life- stages with high precision by regulating gene expression at multiple levels—epigenetic modifications, transcription and translation. The discovery of Pumilio (Puf) RNA-binding proteins in *Plasmodium* and life stage-specific expression of elF4A RNA helicase in *Toxoplasma* provides clear evidence that UTRs play a key role in post-transcriptional regulation in Apicomplexans. Puf proteins bind to 3’UTRs of specific gene transcripts to repress their translation and affect their stability. On the other hand, elF4A unwinds the secondary structure of 5’UTRs initiating translation of the mRNA transcripts in *Toxoplasma gondii*. Cis-acting elements in 5’ and 3’ UTRs play a key role in transcripts that undergo DDX-6 class RNA helicase DOZI mediated translational repression in *Plasmodium* female gametocytes. Recently, both polysome profiling and ribosome profiling in *Plasmodium* have shown increased occupancy of RNA- binding proteins and ribosomes in 5’UTR regions but how this affects translation efficiency is still unclear. To further elucidate these regulatory mechanisms, annotation of UTRs in the genome is essential to discover binding motifs of RNA- binding proteins, target sites of regulatory non-coding RNAs and repetitive elements that affect translation efficiency. With the help of ssRNA-seq, we were able to annotate UTRs for over two-thirds of the genes and we found the 5’UTRs to be strikingly large in the parasites with a median size of 810–860 nucleotides, almost 4 times higher than other model eukaryotes, included in the study. The previous estimation of 5’UTR size in *Toxoplasma* was 288 nucleotides based on expressed sequence tags (EST) analysis and 120–140 nucleotides based on transcriptional start site sequencing (TSS-seq). But these estimates were based on only a subset of genes. The EST data covered only 814 annotated genes and the estimate from TSS-seq was based on a subset of genes with signal peptides for higher reliability. In fact, when all the genes were taken into account along with the predicted TSS sites, the authors estimated a higher 5’UTR size in the range of 580–600 nts. But they considered this estimate as unreliable, reasoning that the 5’ ends of the ORFs could be incorrect because of wrong gene annotations in *Toxoplasma* shown by previous studies. Since our UTR dataset was generated after manually curating the gene models and from strand-specific RNA- seq in two closely related organisms, we believe that the large size of 5’UTRs are not due to incorrect gene models but signify true biological features of these protein-coding genes. A recent study reported similar findings of long 5’UTRs in the range of 607 to 1040 nts in *Plasmodium*. Therefore, long 5’ UTRs could be a common feature among Apicomplexan parasites playing a key role in their tightly regulated life cycles. Unusually long 5’UTRs in a subset of genes in *T*. *gondii* and *N*. *caninum* such as apiAP2 Tfs could have a possible role in gene regulation in these parasites and further studies are required to test this hypothesis. We also mapped modified histone marks (H4ac, H3K9ac and H3K4me3) that are known activation marks, previously identified in *T*. *gondii* RH tachyzoites, onto *Tg*VEG and compared the regions to our annotated 5’UTRs. Out of the 52 high quality histone marks, 40 overlapped with the 5’UTRs, usually starting at a few hundred bases upstream of the 5’UTR and spanning the whole length or part of the 5’UTR. The other 12 peak regions were near the 5’ end of gene models whose 5’UTRs were not annotated in our dataset due to lack of clear RNA-seq evidence. Taken together, we provide a high-confidence set of UTR annotations in both the parasite genomes. ## Putative non-coding RNA in *Toxoplasma gondii* and *Neospora caninum* From a total of 1,270 and 1,940 polyadenylated, intergenic, non-coding transcripts assembled in *Tg*VEG and *Nc*LIV respectively, we identified 688 putative lincRNA and 209 putative ancRNA in *Tg*VEG; 881 lincRNA and 300 ancRNA in *Nc*LIV satisfying the criteria mentioned in the Methods. We observed similar average lengths of 1,310 and 1,369 nts for lincRNA in *Tg*VEG and *Nc*LIV respectively. The *Tg*VEG ancRNA were on an average 2,711 nts long with a mean overlap length of 1,056 nts with the sense coding gene while the *Nc*LIV ancRNA were 2,527 nts long with a mean overlap length of 924 nts. Both lincRNA (\~72%) and ancRNA (\~80%) species were predominantly unspliced transcripts in both the organisms, similar to previously reported in *Toxoplasma gondii*. We also observe a strong anti-correlation (R-square value of \~0.77) between the expression of anti-sense non-coding RNA and sense coding mRNA in the sense- antisense transcript pairs. Comparing ancRNA-mRNA pairs at 144 loci in *Tg*VEG and 213 loci in *Nc*LIV, we see that as the expression of a coding mRNA decreases, the expression of its corresponding antisense non-coding RNA transcript increases. Our results in *T*. *gondii* and *N*. *caninum* substantiate similar findings from previous studies using SAGE data in *T*. *gondii* and *P*. *falciparum*. This points to interesting regulatory functions for these ancRNA transcripts in Apicomplexan parasites that remain to be elucidated. Long non-coding RNA have been shown to have roles in regulating transcription, post-transcriptional regulation and chromatin remodeling. Though their precise cellular functions remain to be elucidated, many studies have discovered thousands of lncRNA genes in fruit fly, zebrafish and roundworm. lncRNA in parasites have not been yet investigated in depth though they have been shown to have specific functions in *Oxytricha trifallax* and *Leishmania*, and many putative lncRNAs have been discovered in *Plasmodium falciparum*. Here we have identified a set of putative lincRNA and ancRNA transcripts in *T*. *gondii* and *N*. *caninum*. It is worth noting that since we used poly(A)+ RNA- seq libraries, our dataset lacks non-polyadenylated lncRNA populations. lincRNA in *T*. *gondii* and *N*. *caninum* have a mean length of \~1,340 bps, are predominantly single exon transcripts and have similar A/U content as untranslated and intergenic regions. We also provide additional RNA-seq evidence for inverse correlation existing between anti-sense and sense transcript abundances found by based on SAGE and extend the same phenomenon to *N*. *caninum*. # Conclusions High-throughput sequencing has enabled researchers to sequence the whole genome of several important parasites to gain unparalleled insights into their life cycle, biological mechanisms and evolution. The quality of genome annotation greatly influences any downstream study, where faulty or incomplete gene models alter the gene sequence and affect analyses based on them. Hence, efforts towards improving the annotation of the genome are beneficial for any genome- based research. Here, we have identified and manually corrected publicly available gene models in *T*. *gondii* VEG and *N*. *caninum* LIV genomes, which amounted to over one-third of the total predicted protein-coding genes. Apart from roughly 468 *Tg*VEG (6.2% of the genes) and 845 *Nc*LIV (11.6%) genes that had no RNA-seq evidence, the correctness of each gene model has been manually verified with transcript and peptide evidence. While we observe a significant increase in annotation quality after manual curation, the genomes may still contain a small fraction of inaccurate gene models that could not be resolved properly by our supporting datasets. Annotation is a continuous process and more lines of evidence in the future would help in improving the genome quality further. For example, transcriptome data across all life-stages rather than just the tachyzoite stage used here could provide better transcript evidence for curating stage-specific expressed genes or alternative splicing events. Similarly long range genomic reads generated from third generation sequencing technologies will help to provide much improved genome assemblies. Apart from accurate annotation of protein-coding genes, generating a catalog of other genomic features and potential regulatory factors such as non-coding RNA and untranslated regions makes a genome annotation more comprehensive. UTRs are gaining prominence in post-transcriptional regulatory processes and their role in Apicomplexan parasites with heavily regulated life cycles need to be further elucidated. We have created a comprehensive catalog of UTRs in *T*. *gondii* and *N*. *caninum* based on strand-specific RNA-seq. Both the parasites have uniquely long UTRs compared to other eukaryotes indicating UTRs may play key regulatory roles at least in a subset of genes in these parasites. This is further implied by the fact that apiAP2 transcription factors, a major family of transcriptional regulators have significantly longer UTRs than the rest of the genes. Long non-coding RNAs are yet another interesting class of transcripts with potential roles in transcriptional regulation in eukaryotes and we have identified a putative set of lncRNAs in both the parasites. With description of these catalogs and manually curated protein-coding genes, we have vastly improved the genome annotation of *Toxoplasma gondii* VEG and *Neospora caninum* LIV for future use by the research community. # Supporting Information We would like to thank the personnel in the KAUST Biosciences Core Laboratory for their support in generating the raw Illumina sequencing and mass spectrometry data. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: A. Pain. Performed the experiments: AR A. Panigrahi SJV. Analyzed the data: AR TM RN TBM EM TDO. Contributed reagents/materials/analysis tools: JW. Wrote the paper: AR TM RN TBM EM AsP SJV TDO JW A. Pain.

# Introduction Moral judgments and intuitions feature prominently in everyday life. They saturate the news, social media, and ordinary conversation and argument. Increasingly they also feature prominently in the academic literature. Morality has become a major interdisciplinary research focus, explored intensively within the fields of economics, philosophy, and evolutionary biology among others. Within the psychological study of morality, two key intellectual developments have been the ‘intuitionist turn’ and the emergence of pluralist accounts. The former development reflects the growing recognition of the importance of emotion and intuition in moral judgment, in contrast to the rationalism of earlier approaches such as Kohlberg’s. The latter development stemmed from cross- cultural research that broadened the scope of morality beyond individualistic considerations of harm and fairness to include values associated with sociality and spirituality. The idea that morality is not monolithic has had its most influential realization in the form of Moral Foundations Theory (MFT). When first propounding MFT, Haidt and Joseph aimed to categorize the automatic and intuitive emotional reactions that commonly occur in moral evaluation across cultures, and they identified five psychological systems (or foundations): Harm, Fairness, Ingroup, Authority, and Purity. Each moral foundation has a distinct set of associated concerns, vices, and virtues. The Harm foundation includes concerns of cruelty, the suffering of others, and the virtues of compassion, caring, and kindness. Fairness covers issues of injustice, unfair treatment, reciprocity, equality, cheating, and individual rights. The Ingroup foundation is concerned with loyalty and obligations for group membership, self-sacrifice, and betrayal. Authority refers to social order, an obligation to conform to hierarchical relationships, and obedience and respect for authority and tradition. The Purity foundation is sensitive to contagion, both physical and spiritual, and encompasses concerns of sanctity, self-control, and the virtues of innocence and wholesomeness. Graham, Haidt, and Nosek proposed that the foundations form two higher-order clusters of moral foundations. Harm and Fairness, the two *Individualizing* foundations, focus on the rights of the autonomous individual. Ingroup, Authority, and Purity, the three *Binding* foundations, have a collective focus on group cohesion. Many studies have used the MFT framework to explore the ideological political divide between liberal and conservative moralities or moral values. For example, Graham et al. found that political liberals endorsed the Individualizing Harm and Fairness foundations, whereas conservatives endorsed all five foundations (including higher endorsement of the three Binding foundations when compared to liberals). To date, MFT has been applied to many different research questions and contexts. For example, numerous studies have explored the individual difference correlates of endorsement of the foundations, how endorsement varies as a function of relational context, or how foundation-specific moral judgments are communicated. The breadth and quantity of this work shows MFT’s generativity as an account of the moral domain. One largely unexplored dimension of the moral foundations is historical. Almost all previous research has examined use of the foundations at one point in time, and only one study has investigated temporal changes. Garten and colleagues examined shifts in moral foundation language use in US political speeches from 1988 to 2012. Garten et al. investigated the moral foundations-related words that Democrats and Republicans used in the vicinity of the word ‘gay’. They found that Republicans were significantly more likely to use Purity-related words than were Democrats and showed a significant increase in Purity language in the period from 1996–2004 relative to the previous eight years. A broader and more fundamental question would be to ask whether and how the foundations have changed in their cultural salience over historical time. Large scale, culture-level questions of this sort are ideally suited to ‘big data’ methodologies in moral psychology, such as explorations of vast linguistic corpora that have been dubbed examples of ‘culturomics’. A popular tool for such analyses is Google NGram Viewer, which allows users to gather word frequencies from the Google Books corpus of digitized books. These frequencies represent the proportion of any given input word within the corpus in any given year, thereby allowing rises and falls in relative frequency–an index of cultural salience or popularity–to be tracked over long periods of time. Psychological researchers have employed the tool to explore historical shifts from third person to first person pronouns, and changing concepts related to age stereotypes and happiness. Greenfield demonstrated an ‘us’ to ‘me’ shift in word frequencies in English- language books published from 1800 to 2000 and revealed other patterned changes indicative of a movement from collectivist rural values to values that are more individualistic and urban. Zeng and Greenfield replicated this shift in Chinese- language books, finding that words linked to materialism and individualism increased between 1970 and 2008, whereas words associated with collectivist values decreased. Findings such as these appear to reveal important macro-level changes in cultural norms and values. However, none of them to date have directly examined changes in morality itself. The one exception is a pair of studies by Kesebir and Kesebir that examined changes in moral virtues over time by recording the frequencies of moral excellence- and virtue-related words in the American corpus of Google Books using the NGram Viewer. Their first study documented a decrease in general virtue-related words (e.g., morality, conscience, character). Their second study examined fifty virtues of a more specific nature (e.g., honesty, kindness, trustworthiness), revealing a significant decline in 74% of those virtues over the course of the 20^th century. These studies reveal important historical trends in the cultural salience of some moral concepts but are limited in their attention to positive (virtue) concepts alone, their focus on specific terms rather than broader patterns linked to theoretical accounts of morality, and their emphasis on linear change trajectories (i.e., increases or decreases) rather than more complex, nonlinear patterns. Building on past work, the present study investigated historical shifts in the cultural salience of multiple domains of morality, as revealed by changes in the relative frequency of large sets of moral terms within the Google NGram database of English language books. The five moral foundations and general morality were each represented by large, validated sets of individual words, and patterns of aggregate change within each set were examined for the years 1900 to 2007. The study’s aim was primarily descriptive and exploratory, seeking to characterize the potentially complex and nonlinear patterns of change in the moral domain across the 20^th century, patterns that have yet to be mapped. The analysis aimed to link these cultural trajectories tentatively to broader historical conditions, and to divide the 20^th century into periods based on distinctive configurations of particular moral domains. Although the study was primarily exploratory, it was guided by three research questions. First, we asked whether general morality (i.e., words referring directly to good and bad, ethics and evil) had declined in salience over the course of the 20^th century consistent with the findings of Kesebir and Kesebir. Second, we asked whether moral foundations associated with social cohesion and order (Binding foundations) tended to decline in relative frequency, whereas those associated with individual welfare and rights (Individualizing foundations, respectively) tended to increase. This expectation is consistent with past findings of rises in indicators of individualism and falls in indicators of collectivist values over the 20^th century, and also with the pronouncements of historians and social theorists. Prominent historian Eric Hobsbawm, for example, argued that “the cultural revolution of the latest twentieth century can thus best be understood as the triumph of the individual over society, or rather, the breaking of the threads which in the past had woven human beings into social textures”. Third, at the level of individual moral foundations, the study asked whether harm-based morality has become more salient since the rights revolutions of the 1960s, as Haslam proposed in his work on ‘concept creep’. This work argues that the meanings of harm-related psychological concepts such as abuse, bullying, and trauma have broadened in response to rising cultural sensitivity to harm over the past half- century, a rise that would be expected to register in the growing salience of harm-based morality. # Method Six sets of terms (whole words or word stems) representing distinct forms of moral discourse were drawn from the Moral Foundations Dictionary, created by Graham et al. (2009) to be used with the Linguistic Inquiry and Word Count program (LIWC). Five dictionaries corresponded to the moral foundations (one for each foundation: Harm, Fairness, Ingroup, Authority, Purity) and one to “general morality”. The moral foundations dictionaries divided terms into positive “virtue” terms and negative “vice” terms, each representing the distinctive content of valued and disvalued concepts within the respective foundations, designed to measure the different ways that people discuss their foundation- specific intuitions. Dictionary development involved both an expansive and a contractive phase. In the expansive phase six researchers generated associations, synonyms, and antonyms for the core concepts for each of the foundations: harm and care, fairness and reciprocity, ingroup and loyalty, authority and respect, and purity and sanctity. Words that were only distantly related or had primary meanings without moral connotations (e.g., ‘just’) were deleted during the contractive phase. The dictionaries were validated by four raters scoring passages containing a subset of the words for contextual relevance. Several terms from the moral foundations dictionaries (20 of 295; 6.8%) appeared in more than one dictionary. These were removed so that all moral foundations dictionaries contained only non-overlapping terms (mean dictionary size = 55 words; range 39 \[Ingroup\] to 80 \[Purity\]). The general morality dictionary contained a broader set of 41 morally saturated terms. Twelve of these terms also appeared in one of the moral foundations dictionaries, but these overlapping terms were retained in the general morality set because its temporal trajectory was not to be compared directly to those of the moral foundations. The final sample of 304 unique terms, organized in the six sets, is listed in. Relative frequencies in the Google Books database of each of the 304 terms for the years 1900 to 2007 were drawn from Google NGram Viewer. The default corpus, “English 2012”, was used, representing books in the English language published in any country. Information on the distribution of country of publication and on changes in that distribution across the study period is not available. Most books were drawn from university libraries and the database includes half a trillion English words. This database yields frequencies up to 2008, but that year’s data were found to show substantial discontinuities with preceding trends and were therefore excluded. Relative frequencies were calculated with ±3 year smoothing to reduce the jaggedness of the time series, as the goal of the analysis was to identify changes in frequency of words over extended periods rather than short-term fluctuations. NGram does not directly generate frequencies for word stems, which composed approximately half (51.0%) of the 304 terms. For each of these terms, NGram was used to select the three most frequently used whole words within the study period (based on mean relative frequency over the 108 years) and the relative frequencies of these three terms were aggregated as detailed below. The relative frequency data generated by NGram were scaled prior to further analysis. The year in which each word achieved its highest relative frequency received a score of 100, and all other years were scaled so that their score represented a percentage of that peak value. In the case of terms that were word stems rather than whole words, the frequencies of their three highest-frequency whole word representatives were summed and these summed scores were then re- scaled by the same method. This process ensured that each word stem received a single series of scores scaled in the same way as the whole word terms. By this means, a dataset was constructed containing 304 terms in each of 108 years, where values represented each term’s relative frequency as a percentage of its highest elevation during that period. The scaled terms in each set were then averaged for each year and their trajectories (time series) were plotted. These average values have a straightforward interpretation: if the value is 10 points higher in one year than another, then the average term in the set was 10% of the maximum relative frequency higher in the former than in the latter. Other methods of scaling terms (e.g., standardizing individual terms prior to averaging them) would not allow such ease of interpretation. The resulting averaged plots therefore represent broad and systematic tendencies, aggregated over large numbers of terms, for conceptually related forms of moral language to vary in cultural salience over historical time. # Results Prior to presenting our descriptive data analyses, we conducted preliminary checks on the internal consistency of the six sets of moral terms across the 108 years, and on the temporal coherence of their time series. Spearman Brown reliabilities based on random split-half correlations indicated excellent internal consistency for five of the sets (Authority =.93, harm =.90, Ingroup =.98, Purity =.99, General morality =.99). However, the Fairness set comprehensively lacked consistency, generating a negative split-half correlation of.82. This finding does not imply that the Fairness set would necessarily lack consistency when assessing language use cross-sectionally, just that its words do not exhibit consistent patterns of historical change. Analysis of auto- correlations of each of the six time series demonstrated that their temporal variation was highly predictable rather than random. Auto-correlations at a four-year lag–the shortest lag outside the ±3 year smoothing band, which artifactually inflates auto-correlations at shorter lags–were lowest for Fairness (0.60) and otherwise ranged from 0.70 for Authority, to 0.81 for Harm and 0.90 for both Ingroup and General Morality. In sum, with the important exception of Fairness, the sets of moral terms demonstrated internal and temporal coherence. Descriptive statistics, correlations with year, and intercorrelations among the six sets of moral terms over the 108 years are presented in (all correlations are Pearson correlations with a critical value of *p* \<.01 in view of the large number of correlations computed). The correlations with year reveal broad linear temporal trends in each set, and the intercorrelations show the degree to which the time series for different sets have similar or dissimilar overall trajectories. The standard deviations indicate that the Fairness, Authority, and Harm foundations had relatively low variability over time, whereas the Ingroup and Purity foundations and the General Morality dictionary had more substantial variability. The six sets of moral terms had markedly different directions of change over the 20^th century. The Ingroup foundation demonstrated an overall rise, Harm and Fairness were relatively stable, whereas Authority and especially Purity and General Morality declined. Interestingly, the temporal intercorrelations among the moral foundations were not consistent with the typical cross-sectional pattern of associations among the Individualizing (Harm & Fairness) and Binding (Ingroup, Authority, and Purity) foundations. Despite both being Individualizing foundations, Harm and Fairness were negatively associated (potentially indicating substitution of one for the other over time), and although Authority and Purity were positively associated with one another they were both negatively associated with Ingroup morality. Ingroup was positively associated with the Individualizing Fairness foundation, and Harm was positively associated with the Binding Purity foundation. Although their interpretation is complicated by the measurement deficiencies of the Fairness term set, these findings indicate that historical changes in the cultural salience of the moral foundations are not structured by more general changes in Binding and Individualizing morality. Once again, such findings do not necessarily call into question the coherence of the Binding and Individualizing groupings proposed by moral foundations theorists. These groupings may capture the covariation structure of individual differences in morality assessed at one point in time but not the structure of temporal changes in culture-level morality. To answer our first research question, we examined the trajectory of the General Morality set, which contained an assortment of morally-freighted terms (e.g., bad, character, ethic\*, evil, good, immoral\*, moral\*, principle\*, righteous\*, value\*, wrong\*). reveals a steep decrease from 1900 to around 1980, consistent with the general decline in virtue language previously identified by Kesebir and Kesebir. However, this decline is followed by an equally steep rise thereafter which was not identified in that earlier work. In effect, moral content in the Google Books database declined steadily throughout the 20^th century until an inflection point in around 1980. The subsequent rebound of moral language may point to the reinvigoration of social conservatism in the Anglophone world at around this time, led by such figures as Ronald Reagan and Margaret Thatcher, and manifest in the ongoing ‘culture wars’ and rising political polarization. Our second research question concerned historical changes in broad Individualizing and Binding moral dimensions, represented by groupings of the five moral foundations. However, the lack of internal consistency of the Fairness foundation term set and the lack of consistent positive associations among the foundations within each grouping meant that this question could not be meaningfully addressed. In short, one of the two Individualizing foundations had serious measurement limitations and the Binding foundations did not form an internally coherent group. The second research question was therefore not addressed further and we proceeded to examine the time series for the five moral foundations independently. Time series for each moral foundation are presented in. The five trajectories are very distinctive. Harm shows a gentle decline until about 1960, punctuated by noticeable short-term rises around the two World Wars, and then rises steeply from about 1980. This pattern is consistent with the expectation based on the third research question. Fairness, Harm’s fellow Individualizing foundation, has a relatively flat trajectory but begins to decline in the late 1970s, so that the two foundations trend in opposite directions in the last three decades of the study period. The problematic measurement properties of the Fairness term set make this discrepancy risky to interpret. Ingroup shows a relatively steady increase over the entire century. Authority declines until about 1950 then shows a striking rise to the late 1960s and an equally striking fall thereafter. The failure of Ingroup and Authority to rise in the vicinity of the World Wars as did Harm is perhaps surprising, as concern with loyalty and obedience might be expected to elevate during times of major conflict. Purity declines sharply from 1900 to about 1980 and then rises in a pattern that mirrors the trajectory of General Morality. To formalize these descriptions of the shape of the six moral time series, we conducted a series of curve-fitting regression analyses. Three models were run for each series: the first included only a linear effect of time, the second added a quadratic effect, and the third added a cubic effect. Significance tests of these models are suspect due to the smoothing of the time series, but the explained variance added by each effect offers a guide to interpreting the patterns described above. The General Morality and Purity series are both dominated by strong linear declines combined with smaller curvilinear (quadratic) effects representing their post-1970s rebounds. The Ingroup series shows a strong linear rise with no curvilinear trends. Harm and Fairness are dominated by curvilinear (quadratic) trends, the former concave and rising late in the 20^th century and the latter convex and falling in that period. Authority shows the most complicated curve, with a dominant cubic term to capture the circumscribed rise and fall of the foundation between 1950 and 1980. In sum, the moral time series are characterized by reliable curvilinear patterns, and linear, quadratic, and cubic terms collectively account for a very high proportion of their variance (*M* = 81.8%). To explore further the patterns observed for the five moral foundations, separate indices were computed for aggregates of their positive and negative terms (“virtues” and “vices”). Correlations between virtues and vices within foundations were strong for Purity (*r* =.99), Harm (*r* =.71), and Authority (*r* =.67), but relatively weak or negative for Fairness (*r* =.44, *p* \<.001) and Ingroup (*r* =.25, *p* \<.01). Inspection of the 10 time series qualified several findings concerning the five foundations. First, the rise and fall of Authority on either side of the late 1960s is especially marked for its vice terms (e.g., defy\*, disobe\*, dissent\*, rebel\*). Second, Fairness vices (e.g., unequal\*, unfair\*, unjust\*) display a pattern that is very similar to that of the Authority vices but was largely masked when the former were merged with Fairness virtues. They rise steeply from around 1955 to 1970 and then drop sharply. Finally, Ingroup virtues and vices reveal dissimilar time series. Ingroup vice terms (e.g., enem\*, foreign\*, immigra\*, terroris\*) are relatively static or in gentle decline from 1900 to the mid-1960s and then rise steeply. Ingroup virtue terms (e.g., communal, family, group, nation), in contrast, rise steeply from 1900 to around 1970 and then decline somewhat. # Discussion The temporal patterns revealed by our analysis are not consistent with a simple narrative of linear rises or falls in the cultural salience of morality through the 20^th century. Although the fundamental moral terms collected in the General Morality dictionary showed a steep decline, compatible with a broad reduction in the cultural salience of morality, that decline was not inexorable, reversing sharply from about 1980. According to the Google Books database, at least, the culture manifest in English language books has substantially de- moralized over the past century, but also re-moralized in the more recent past. This reversal of an otherwise striking decline in moral language, a nonlinear effect revealed as an aggregate pattern across 41 morally-saturated terms, is an important qualification to earlier findings that individual virtue-related moral terms have tended to decline in relative frequency. A more complex narrative of an individualistic morality gradually replacing a morality based on social order and cohesion is also incomplete. Despite being opposed in theory, the Binding and Individualizing foundations did not form two coherent groupings. Neither of the Individualizing foundations (Harm and Fairness) displayed a rising linear trend between 1900 and 2007, contrary to previous findings on other indicators of individualism such as rising first- person pronoun use and materialist values, although Harm’s nonlinear trend showed a suggestive rise late in the 20^th century, consistent with prediction and Haslam’s argument concerning ‘concept creep’. The Harm and Fairness time series were also negatively correlated, a finding inconsistent with Individualizing foundations being a coherent set, at least where patterns of historical change are concerned. However, the measurement deficiencies of the Fairness term set make it difficult to confidently interpret the negative correlation. The three Binding foundations also lacked coherence as a group, and they had no consistent tendency to decline over the 20^th century. Purity demonstrated such a decline, reflecting a reduction–at least until around 1980 –in the salience of sanctity-based morality, but Ingroup morality showed a steady rise, and Authority displayed a complex nonlinear trajectory. Thus, our findings do not offer support for the view that Individualizing and Binding morality have shown a wholesale rise and fall, respectively. The most that could be claimed in that regard is that one form of Individualizing morality (Harm) and one form of Binding morality (Purity) display this pattern, and the former rise was only evident from about 1980 onward. Thus, a key finding of our study is that the five moral foundations have unique, often nonlinear trajectories that cannot be aggregated into broader groupings. Each form of morality captured by a foundation appears to have its own irreducible historical patterning. Although our finding that the Individualizing and Binding foundations lacked empirical coherence as groupings might be taken as evidence against the validity of the groupings themselves, we would caution against that inference. Although Individualizing and Binding may not accurately capture the covariation structure of historical changes in morality at the cultural level, they may still validly describe the structure of covariation in morality assessed at one point in time, and/or assessed at the level of individual differences. There is ample evidence that Individualizing and Binding foundations cohere empirically in other research contexts, and are related to additional factors such as political orientation in divergent ways. Consequently we do not see our findings as posing a fundamental challenge to moral foundations theory. The complex patterns revealed by the five moral foundations can be very tentatively integrated into five historical periods. The first period, covering 1900 to WWI and corresponding to the final stages of the Belle Époque, is one of liberalization, illustrated by reductions in the Purity and Authority foundations. The second, interwar period, which includes two convulsive conflicts and the Great Depression, interrupts this liberalizing process. The Binding foundations, particularly Purity and Ingroup, rise and then fall during this period, suggesting a preoccupation with protecting the cohesion and sanctity of group-based identities. The third period, from around the end of WWII to around 1968, can be interpreted as a gathering crisis of authority. Three foundations rise in parallel: Authority and Fairness (both especially for vice terms), and Ingroup (especially virtue terms). These linked increases signify a growing cultural preoccupation with disobedience, rebellion, injustice, and inequality, coupled with an increasing focus on loyalty to groups. The fourth period, from around 1968 to around 1980, represents a second liberalization. The Binding foundations (especially Authority and Purity) decline steeply and Harm begins a steady rise that points to a growing concern with suffering, care, and protection of the vulnerable. The fifth period, from around 1980 to the end of the study period in 2007, involves a relatively sudden shift in the salience of moral concepts. The inflection point corresponds roughly to the beginning of more than a decade of uninterrupted conservative rule in the Anglophone heartlands of the USA and the UK. Moral content increasingly saturates the database and the Binding foundations–especially Purity but also Ingroup and Authority–reverse their previous decline. The Individualizing foundations, led by Harm, continue their increase, so that both individualist and social order and cohesion-based moralities rise in parallel, suggesting a broader re-moralization. This positively correlated increase of normally antagonistic moralities of the political left and right may point to increasing moral polarization and conflict. Historical speculation of this sort can only be put forward with great care. Caution is required when drawing inferences regarding cultural trends from Google Ngram data. Although the corpus drawn from Google Books is the largest currently available there are several limitations in its sampling. First, the corpus reflects what has been published, not necessarily what is widely popular in the general public, and it is not specific to a single national ‘culture’. Second, some researchers have suggested that the Google Ngram English corpus has included progressively more scientific literature throughout the 20^th century. This does not necessarily undermine the capacity of the data presented in the present study to illuminate broad cultural trends, and the fact that some forms of moral language increased over the course of the 20^th century arguably conflicts with the expectation that moral language would decline with the increased inclusion of relatively neutral scientific reports. Third, because the dataset represents books published in English during a century in which major Anglophone countries rose and fell in their relative cultural influence (e.g., the rise of the USA relative to the UK), it is possible that some of the historical trends observed in our study are confounded by national differences. Fourth and finally, some research has suggested there is a time lag of up to a decade between exogenous events and their effects in literature. Future research could seek to replicate our findings with other corpora that potentially have different temporal relationships to exogenous variables, such as those that aggregate new media reports rather than books. Although Moral Foundations and the accompany Moral Foundations dictionary is a widely used and understood taxonomy of moral priorities, there are other approaches to categorization in the moral domain. Further research using dictionaries based on alternative moral schema such as Janoff-Bulman and Carnes’ Model of Moral Motives might reinforce the historical patterns suggested by our data or illuminate other trajectories of moral language. # Conclusions The present study adds to an emerging body of quantitative research on historical changes in human culture. It extends previous work with its thorough and systematic attention to the multiple dimensions of morality. The dynamic changes in the salience of morality through the 20^th century that it finds are complex, resisting simple linear narratives of uninterrupted rises or falls. There does appear to have been a progressive reduction in the cultural salience of morality in general since the beginning of the last century, but there has also been a vigorous rebound since the early 1980s. At a more fine- grained level, different moral foundations have markedly distinct trajectories, which correlate with major societal conflicts and developments. Understanding historical variations in moral judgments and values may help to illuminate social challenges in the present and those yet to arise. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction With the popularization of seedling nurseries, transplantation has become a very important activity in greenhouses. Manual transplantation in greenhouses is labor-intensive and costly, and growers have become increasingly interested in using automatic seedling transplanters. Since the first robotic transplanter was designed in 1987, much progress has been achieved. Research on automatic transplanters has mainly focused on mechanical engineering, machine vision and control engineering. Design of an end-effector is key for transplanting plug seedlings. Classified by the picking action of fingers, end- effectors are divided into two complementary forms, the clamp form and slide form. Usually fingers of clamp end-effectors are naked and fingers of slide end-effectors have sleeves. There main difference between these two forms is the picking action. For grasping a plug seedling, these two forms have differernt actions. Clamp end-effectors need to approach the seedling from the top, and keep going down until the fingers stab into the substrate and reach a certain depth, then the fingers clamp to the center of root plugs. The verticality of clamp end-effector fingers would be changed after clampping. Slide end-effectors with fingers retracting into corresponding sleeves also need to approach the seedling from the top, but stop going down when the sleeves reach the surface of the root plugs, then the fingers extend out of sleeves and stab into the substrate. The verticality of slide end-effector fingers would not be changed during the transplanting process. Compared with slide end-effectors, clamp end-effectors have better performance in grasping and holding plugs but have problems in releasing plugs. Sometimes, root plugs strongly adhere to the fingers and cannot be released successfully by gravity. On the contrary, slide end-effectors release plugs more effectively by the relative motion of needles and their sleeves but do not perform very well in grasping and holding plugs. To combine these advantages, an end-effector driven by a linear pneumatic cylinder was designed. This end-effector could clamp the plug with four needles simultaneously when the needles are stabbing into the nutritional substrate. Plug seedlings are the working object of end-effectors, and the cohesion to the root plug has a significant impact on transplantation results. One of the most important morphological quality criteria for plug seedlings is to have a firm and cohesive root plug, which is good for transplantation. The density and uniformity of the root system mainly ensure high cohesion to the root plug, as well as the moisture content, the substrate bulk density and the volume proportion of substrate ingredients (peat: vermiculite: perlite). In recent years, researchers have realized that innovations of end-effectors were not enough to promote automatic transplantation and began to combine the research with root plugs. Usually, the extrusion force that end-effector put on root plug and the adhesive force between the root plug and plug cell’s inwall were used to estimate the root plug. The extrusion force has a great responsibility to the transplantation performance. Root plug would be cracked if the extrusion force was too large. In contrast, plug seedling couldn’t be lifted out of the plug cell if the extrusion force was too small to overcome the adhesive force. The adhesive force was related to the conditions of growth media and material characteristics of plug tray. However, published studies separated the studies of the root plug from the transplanting process. Plug seedlings for the extrusion test were carefully moved out from plug trays manually. For the adhesive test, plug seedlings were usually grasped at the stems and pulled out from the plug cell. The relationship between these two forces and the transplantation performance of the same plug seedling could not be analyzed well. There were no existing systems to measure these two forces directly or indirectly during the end-effector’s transplantation movements. To solve this problem, a force measurement system was installed on the designed end-effector with a tension transducer and pressure transducer. During the practical transplanting process, the two forces could be recorded and calculated directly through the transducers. The present study aims to design an efficient end-effector for automatic transplanters in greenhouses and analyze the extrusion and adhesive forces during the transplanting process. Root system is not a agronomic factor, and the other three are agronomic factors that could be controled during nursery. The density and uniformity of the root system are also influenced by the three agronomic factors. In this manuscript, we just want to anslyze the agronomic factors. The specific objectives of this study are: (1) to design the structure of end-effector, (2) to establish a force measurement system to record extrusion and adhesive forces based on the designed end-effector, (3) to verify the adaptability and efficiency of the designed end-effector through a combinational experiment on the moisture content of substrate, the substrate bulk density and the volume proportion of substrate ingredients, and (4) to analyze the extrusion and adhesive forces during the transplanting process. # Materials and methods ## Structure design of the end-effector A good end-effector of transplantation, besides being reliable, should be able to effectively penetrate, grasp, hold, and release a wide variety of growth media with minimum damage to the roots. Other guidelines were that the end- effector should be structured simply and powered by low-cost actuators; accommodate a wide variety of plugs in different sizes of containers; and avoid plug aerial portion and operate on the root zone portion. However, most existing end-effectors were clamp forms or slide forms. Clamp end-effectors have problems in releasing plugs, and slide end-effectors have problems in grasping and holding plugs. In contrast, slide end-effectors performed well in releasing plugs and clamp end-effectors performed well in grasping and holding plugs. Based on these guidelines, an effective end-effector combined the advantages of clamp forms and slide forms was shown in. There are four symmetrically slanted needles, and each needle slides within a sleeve. The needles are actuated by a double-acting pneumatic cylinder. The depth into growth media and the verticality of needles could be adjusted easily for different plug trays and cell sizes. The depth could be adjusted by changing the position of the Upper Cross Plane, which is fixed to a Thread Rod by Hexagon Nuts 2 and 3. The verticality of needles could be adjusted through the Rotation Part. Screwing Hexagon Nut 1 changes the position of the Middle Cross Plane, thus the Rotation Part could rotate and change the verticality. is the planar kinematic sketch of the designed end-effector. The red curves are the trajectories of the two needle cusps in the diagonal direction. shows the process for transplanting one plug seedling. At the beginning of a transplanting cycle, the end-effector, with its needles retracted , would approach a plug from the top and extend the needles to penetrate into the growth media. After the Thread Rod knocked against the Bolt, the needles kept penetrating and started clamping the root zone portion of the plug. Then, the plug would be lifted out of the cell and transported to a growing tray. After getting the position right above the empty cell that needed to be filled, the end-effector would lower the plug into an empty cell in the growing tray and retract the needles to release the plug. The transplanting cycle is completed by leaving the growing tray. With the relative motion between needles and corresponding sleeves, the sleeves would push the plug into the cell in an upright position. This end-effector combined the advantages of clamp form and slide form without increasing the time consumption. ## Force measurement system Based on the designed end-effector, a force measurement system was established to record the extrusion and adhesive forces. As shown in, a tension transducer (HBM, U9C/100N) was installed between the end-effector and mechanical arm, and a pressure transducer (HBM, C9C/100N) was fixed on the bolt of end-effector. The output signals were amplified by two transmitters (LONGTEC and TR200H). The amplification factor was about 100. Then, the amplified signals were collected and analyzed by a portable data collector and analyzer (AVANT MI-7008), with the signal size of 8192 and the sampling frequency of 2560. All the data were stored in a computer for processing and analysis. displays the forces in the measurement system, and four equations can be obtained as follows: $$F_{1} = F_{L} + \left( {G_{e} + G_{s}} \right)$$ $$F_{2} = F_{s} + 4F_{n1}^{\prime}\text{sin}\alpha$$ $$F_{n1} = F_{n1}^{\prime}$$ $$F_{k}L_{1} = F_{n1}L_{2}$$ where *F*_*1* is the tension on U9C transducer when the end-effector is lifting a plug seedling, *F*_*L* is the adhesive force between the root plug and plug cell inwall, *G*_*e* is the total gravity of end-effector, *G*_*s* is the gravity of a plug seedling, *F*_*2* is the stress on the C9C transducer when the end-effector is clamping a root plug, *F*_*s* is the total resistance when the end- effector is clamping without a root plug, *F*_*n1* ($F_{n1}^{\prime}$) is the acting force between the Rotation Part and Short Pin, *F*_*k* is the extrusion force on a root plug, *L*_*1* is the distance between the acting point of *F*_*k* (1/3 of the stabbing depth from the needle cusp) and the rotation center (*N*_*1*), *L*_*2* is the distance between the acting point of *F*_*n1* and the rotation center. The applied force *F*_*k* and *F*_*n1* act as torques. Then, the adhesive force *F*_*L* and the extrusion force *F*_*k* are calculated as follows: $$F_{L} = F_{1} - \left( {G_{e} + G_{s}} \right)$$ $$F_{k} = \frac{L_{2}\left( {F_{2} - F_{s}} \right)}{4L_{1}\text{sin}\alpha}$$ ## Transducer calibration U9C and C9C transducers designed for measuring static and dynamic forces were used in the measurement system. The nominal forces were the same as 100 N, and the limit forces were 200% of nominal forces (200 N). The relative reproducibility and repeatability errors without rotation were smaller than 0.2% of nominal forces (0.2 N). The nominal temperature range was -10°C to +70°C. After the transmitters and signal collecting device configured as mentioned above, the transducers were calibrated on a standard electric universal testing machine (CMT4204). And the accuracy was 0.1 N. Force values and the corresponding voltage values are shown in. shows the liner fit for the U9C and C9C transducers. The linear fitting equations are as follows: $$F_{U} = - 12.26V_{U} - 14.34$$ $$F_{C} = 18.25V_{C} - 2.43$$ where *F*_*U* is the tension on the U9C transducer, *V*_*U* is the voltage measured by the U9C transducer, *F*_*C* is the stress on the C9C transducer, and *V*_*C* is the voltage measured by the C9C transducer. ## Performance tests of the designed end-effector The interaction experiment is an effective measurement to assay the comprehensive effect of multiple factors. To test the adaptability and efficiency of the designed end-effector, an interacting effect trial on the moisture content of the substrate, substrate bulk density and volume proportion of substrate ingredients was designed. Each factor had three levels. shows the factors and levels of the interaction experiment. To meet the requirements of seedling nurseries, the substrate porosity were greater than 85%, and all the water potentials were over -10 kPa. On March 4, 2016, cucumber seeds (JinPei 99F1, ZYANHN) were sowed in nine 6×12 trays (overall dimension 280×540*mm*) and nursed in the greenhouse at Zhejiang University. On the 14th day after sowing, the test was carried out on a transplanting platform (shown). The nine trays were divided into three groups and numbered individually. Trays in the same group had the same volume proportion (peat: vermiculite: perlite), but the average substrate bulk density in each cell increased from the former one in numbers. A total of 504 g, 612 g and 720 g nutritional substrate were evenly spread in the three trays of each group (7 g, 8.5 g and 10 g substrate per cell). The volume of the plug cell was 38.2 ml. The first group included Trays 1 to 3, which were filled with a peat: vermiculite: perlite mixture (6:3:1, v/v/v). The second group included Trays 4 to 6, which were filled with a peat: vermiculite: perlite mixture (6:2:2, v/v/v). The third group included Trays 7 to 9, which were filled with a peat: vermiculite: perlite mixture (7:2:1, v/v/v). All the three volume proportions are frequently used in seedling nurseries. Each tray was divided into three regions evenly along the long side of the tray. All of the trays were watered plentifully (water began to drip from the hole at the bottom) every morning till the last time. Region 1 and Region 2 were last watered four days and two days before the test beginning day respectively, and Region 3 was last watered in the morning of the test beginning day. The relative moisture content of the three regions in the same tray was different when the test was carried out (78%, 80%, and 82%). Ten plug seedlings in each region were randomly chosen for transplanting, and the results of transplantation were recorded, along with the signals of transducers. The working pressure of the pneumatic cylinder (SMC CQ2A20-40DM) was set as 0.44 MPa. The diameters of the needles were 2 mm. The stabbing depth of the needles was 40 mm, and the needles stabbed into the substrate at a 4° vertical angle. The distance between the two needle cusps in the diagonal direction was 36 mm before clamping and then contracted to 16 mm after clamping. The end-effector lifted plug seedlings at a speed of 0.1 m/s and moved at a speed of 1 m/s as set. The average time cost for transplanting one plug seedling is 2.8 s. # Results and discussion ## Transplanting performance of end-effector lists the treatment and seedling germination rate. All the germination rates are over 90%. Some germination rates are as high as 98.6% (Tray 1, Tray 5 and Tray 8), which means that each tray of those only have one empty cell. shows the transplanting results of the plug seedlings, which are classified in three kinds. Success results show that the plug seedlings were successfully lifted from the high-density trays, transferred to the low-density trays and released in the empty cell. If any step of the three procedures failed, the transplanting results would be identified as a fail. Sometimes the root plugs of successfully transplanted seedlings were cracked during transplantation, and these results were identified as harm. The success rate *Rs* and harm rate *Rh* are calculated as follows: $$Rs = \frac{\text{s}}{n} \times 100\%$$ $$Rh = \frac{h}{n} \times 100\%$$ where *s* is the number of successfully transplanted seedlings, *n* is total number of transplanted seedling, and *h* is the number of harmfully transplanted seedlings. Results from indicated that all the 270 plug seedlings were successfully transplanted and most plug seedlings were kept intact after transplantation (as shown). The total transplantation success rate of the end-effector was 100%, and the root plug harm rate was below 17%. There were 50 plug seedlings harmed, most harmed root plugs lost less than 1/5 of substrate or cracked slightly (as shown), and only 11 root plugs lost more than 1/5 of substrate (as shown). The number of harmed seedlings decreased with the decrease of moisture content. Too much water would loosen the root plug. Compared with the other two volume proportions, plug seedlings in trays filled with a peat, including vermiculite: perlite mixture (6:3:1, v/v/v), were less harmed after transplantation. A moderate increase of vermiculite is beneficial for the growth of root systems, which ensure high cohesion of the root plug. High bulk density does not make any better. The results indicated that a medium bulk density (8.5 g/38.2 ml) is good to form firm and cohesive root plugs for automatic transplantation. The harm rate is as low as 7%. shows the detailed statistics of harmed root plugs in the experiment. All plug seedlings in Tray 2 and Tray 7 were intact after transplantation. All plug seedlings in the low moisture content region (78%) of Tray 5 and Tray 8 were also intact after transplantation. The results in Tables and indicate that the designed end-effector is very efficient. ## Analysis of the extrusion and adhesive forces shows the typical transplanting process signals of the C9C and U9C transducers. The signals correspond to Step B to Step D of the transplanting process in. The signal collection started when the end-effector met the surface of root plugs. The end-effector clamped root plugs (the signal curves from Point b to Point c) and maintenance of the root plugs (the signal curves from Point c to Point c’). The plug seedlings were lifted and separated from the cell (the signal curves from Point c’ to Point d’). After separating, the end-effector continued rising to the height as set before (the signal curves from Point d’ to Point d). According to the linear fitting Eqs and, the tension on the U9C transducer (*F*_*U*) and the stress on the C9C transducer (*F*_*V*) could be easily obtained. The minimum tension value of the U9C transducer from Point c’ to Point d’ ($F_{Uc\prime d\prime}^{min}$) was used as *F*_*1*. The stable tension value of the U9C transducer from Point d’ to Point d (*F*_*Vd*′*d*) was used as the total gravity of end- effector and plug seedling (*G*_*e* + *G*_*s*). Before the experiment, the end-effector ran five times without plug seedlings. The average stress value of the C9C transducer from Point c to Point c’ ($F_{Vcc\prime}^{\prime}$) was used as the total resistance *F*_*s*. When the end-effector was working with plug seedlings, the average stress value of the C9C transducer from Point c to Point c’ (*F*_*Vcc*′) was *F*_*2*. For the experiment, sin *α* = 2/18, *L*_*1* = 136 mm, and *L*_*2* = 18 mm. The adhesive force *F*_*L* and the extrusion force *F*_*k* are calculated from Eqs and. displays the extrusion force *F*_*k*, the adhesive force *F*_*L* and the value of *F*_*K*/*F*_*L* in the combinational effect trial. Each result was the average value of 10 plug seedlings. Considering the results in, in each region, if there had any root plug that lost more than 1/5 of substrate or more than one root plug that lost less than 1/5 of substrate or cracked slightly, the transplanting results of this region would be considered poor. The transplanting results would not be considered good when the value of *F*_*K*/*F*_*L* is smaller than 5.99 or larger than 8.67. For a plug seedling at one condition, the adhesive force *F*_*L* is also a certain value, but the extrusion force *F*_*k* could be changed by adjusting the needles of the end- effector. The root plug would be broken if the extrusion force is too large. On the other hand, too small of an extrusion force could not hold the plug firmly. To guarantee a better transplantation performance, the end-effector was adjusted to make the extrusion force *F*_*k* in a reasonable range (5.99–8.67 times the adhesive force *F*_*L*). ## Significance test of the three factors based on ANOVA ### Significance test on the extrusion force ANOVA is a useful method when the objective is an assessment of the impact of certain controllable factors on a specific response. Analysis of variance for extrusion forces based on ANOVA are presented in. The results showed that all the three variable factors and their interactions had significant effects on the extrusion force (P \< 0.05). presents the independent impact of the three variable factors. There were no significant differences between 6:3:1 and 6:2:2 volume proportion of substrate ingredients with respect to extrusion force. The extrusion forces increased when the moisture content of growth substrate decreased. Root plugs would be soft and easily deformed if they contained too much water. Thus, the extrusion force decreased. The extrusion forces increased with the bulk density first and then decreased. indicates that root plugs were easily cracked when the bulk density was too large. When the root plugs cracked, the extrusion forces decreased. The mean extrusion force under 7:2:1 volume proportion of substrate ingredients was greater than the other two, which indicated that the increase of peat content enhances the extrusion force. ### Significance test on the adhesive force Analysis of variance for adhesive forces based on ANOVA are presented in. Results show that all the three factors had significant effects on adhesive force individually (P \< 0.05). There was a significant interaction between volume proportion and bulk density (P \< 0.05). presents the independent impact of the three variable factors on adhesive force. There were no significant differences between 6:3:1 and 7:2:1 volume proportion of substrate ingredients with respect to adhesive force. The adhesive forces increased when the moisture content of growth substrate decreased. A reduction of moisture would increase the adhesion between root plug and tray cell. The adhesive forces increased with the increase of bulk density. The expansion of root plugs would be stronger when the bulk density increased. Thus, the friction between root plug and tray cell increased. The mean adhesive force under 6:2:2 volume proportion of substrate ingredients were greater than the other two. This indicated that the increase of peat content could increase the adhesive force. # Conclusion The newly designed end-effector with four needles was proven to be an efficient end-effector. This end-effector clamped the plug at the same time with the four needles penetrating into substrate and without increasing time consumption. The total transplantation success rate of the end-effector was 100%, and the total harm rate to root plug was 17%. Most harmed root plugs lost less than 1/5 of substrate or cracked slightly. The designed end-effector performed well in all of the combinations of factors. Making the bulk density moderate before nursery and keeping the growth media at a low moisture content before transplantation is better for keeping the root plug intact. The combinational trial showed that all the three variable factors and their interactions had significant effects on the extrusion force. All the three factors had significant effects on adhesive force. The relationship between these two forces and the transplantation performance was established. To guarantee better transplantation results, the end-effector should be adjusted to make the extrusion force *F*_*k* in a reasonable range (5.99–8.67 times of the adhesive force *F*_*L*) before transplantation. This could provide a scientific basis for end-effector application during transplantation. # Supporting information The authors gratefully acknowledge the help provided by Lijun Zhang, Weinan Shi, Xiaolei Zhang, Gai Zhu, Guangyuan Liu and Yefeng Yang. Thank you for their help in the execution of experiment and manuscript submitting. [^1]: The authors have declared that no competing interests exist. [^2]: **Conceptualization:** ZJ HJ. **Data curation:** ZJ YH. **Formal analysis:** ZJ YH. **Funding acquisition:** HJ. **Investigation:** ZJ YH. **Methodology:** ZJ HJ. **Project administration:** ZJ HJ. **Resources:** ZJ HJ YH JT. **Software:** ZJ YH. **Supervision:** HJ. **Validation:** ZJ HJ. **Visualization:** ZJ. **Writing – original draft:** ZJ. **Writing – review & editing:** ZJ JT.

# 1. Introduction The BRI was initiated by the Chinese president Xi Jinping in 2013, and it was completely introduced during the visit of Kazakhstan and Indonesia. In the background of the history of China’s ancient overland and maritime Silk Road, the BRI consists of the Silk Road Economic Belt and the 21st Century Maritime Silk Road, geographically crossing Asia, Europe and Africa. The BRI comprises of over 50 different economies, which cover 80% of globe population and its the estimated cost is over \$21.1 trillion US dollars. Since the BRI was proposed in 2013, major achievements have been made in infrastructure construction, international trade, cross-border investment and sustainable growth. The World Bank report (<https://www.worldbank.org/en/topic/regional- integration/publication/belt-and-road-economics-opportunities-and-risks-of- transport-corridors>) noted that BRI is expected to increase real income by 1.2–3.4% in BRI economies and 0.7%-2.9% globally, with investments that could lift more than 7.6 million people out of extreme poverty. In particular, BRI transport projects have significantly reduced trade costs, which are expected to increase trade in BRI economies by 2.8% to 9.7%, increase in world trade by 1.7% to 6.2%, and total foreign direct investment in BRI economies by 4.97%. The BRI has strengthened cooperation in trade, investment, infrastructure construction, institutional and cultural exchange among Asia, Europe and Africa, and formed a new economic network. It aims to build a community of shared future for mankind, and it is an important platform for international multilateral cooperation and regional integration cooperation. Ultimately, it will make the world economy and society more open, inclusive, balanced and benefit-sharing. However, there are still significant political, cultural, social, economic and institutional differences among the BRI economies, especially in terms of economic growth and their policies. Such differences are not only between China and BRI economies, but also within BRI economies. Thus, these transnational differences are a set of differences between many stakeholders. In general, the business cycle synchronization directly reflects the co-movement of output growth and fluctuations of various countries, and reveals the coordination and difference of transnational economic policies and the policy coordination of various countries to deal with shocks. There are abundant researches on business cycle synchronization, but most of them focus on its measurement and determinants. Through comparison, it is found that these studies only focus on the business cycle synchronization between one country and the other, which cannot fully reveal the interaction synchronization among multiple countries. In fact, the real economic system has the basic characteristics of social network, and the overall economic activity is caused by the interaction of many separated and different economic subjects, including division of labor, cooperation, trade and other multiple interactions. Therefore, revealing the complex interdependence between economies is key to understanding the business cycle synchronization. However, in the real quantitative analysis, the traditional statistic analysis method cannot reveal the complex linkages of business cycles synchronization. Fortunately, the complex network analysis method investigates and studies economic and social phenomena and structures from the perspective of “linkages”, and can accurately reveal some intricate social and economic linkages between nodes such as countries. At present, complex network analysis methods are widely used in some academic subjects such as international trade, international investment, finance, macroeconomic volatility, industrial economy and regional economy. With the in-depth application of network analysis methods, complex network analysis has become an emerging trend to study the business cycles synchronization between countries, and it is also an ideal tool to reveal the topological characteristics of the BCSN. Base on the concept of graph theory and complex network theory, countries can be regarded as nodes, the business cycle synchronization linkages of each country can be regarded as the connecting edges of nodes, and thus the BCSN is formed. In empirical studies, pairwise correlation coefficient matrices can be used to represent the business cycle synchronization between different economies, and the topological structure of the BCSN can be studied by using complex network analysis method. Through further literature sorting, there were some studies analyzed the business cycle synchronization between China and other BRI economies, and investigated the topological characteristics of BCSN between China and other ASEAN member states under the BRI. Although these studies provide some reference for this paper, the research on the topological characteristics of the BCSN of BRI economies is still not profound enough. Then, what are the topological features of the BCSN of BRI economies? How does China fit into the network? At the same time, what are the differences in the topological features of the BCSN of BRI economies before and after the implementation of the BRI? In order to answer all these questions, this paper uses the instantaneous quasi- correlation method to measure the level of business cycle synchronization for 53 BRI sample countries in 2000–2019, and construct pairwise business cycle synchronization matrix, and then uses complex network analysis method to construct the BCSN, and empirically studies the whole and individual topological characteristics of the BCSN in order to enrich the existing research. In the end, this paper will reveal China’s influence in the network and compare the differences in the topological characteristics of the BCSN before and after the BRI was implemented. Based on the above synchronization matrix and network analysis method, this paper can find the characteristics of heterogeneity and diversity of the output synchronization linkages among various economies, revealing the interdependence and interactions of output linkages and provide a new research perspective for the study of BRI business cycle synchronization. The rest of this paper is arranged as follows. Section two shows a literature review related to this paper. Section three illustrates the methodology explanation and data description. Section four introduces the overall topological characteristics of BCSN. Section five reveals the individual topological characteristics of BCSN. And the last section is conclusions and discussions. # 2. Related literature The literature closely related to this paper mainly includes the following two categories: one is the measurement and research on the international business cycle synchronization; the other is on the BCSN. The measurement and research on international business cycle synchronization mainly includes static method and dynamic method. For the static methods, existing studies mainly adopt simple correlation coefficient method and dynamic factor model. It should be noted that although the static method can intuitively judge the level of the synchronization, it cannot reveal the dynamic characteristics of the business cycle synchronization. Therefore, the follow-up measurement research has gradually shifted from the traditional static methods to the dynamic methods. For dynamic methods, existing studies mainly focus on Markov regime switching model, GARCH model, concordance index, difference method and instantaneous quasi-correlation method. In addition, some scholars used the Method of Hodrick-Prescott (HP) filter to de-trend the original output data series and further calculate the business cycle synchronization. However, Hamilton (2018) believes that HP filtering method introduces unreal dynamic linkages that are not based on original data, and its results are affected by the size of smoothing parameter, so it cannot truly reflect the level of business cycle synchronization. Compared with other dynamic methods, the instantaneous quasi-correlation method is more efficient and could carry out dynamic calculation and analysis, which avoids problems caused by artificial parameter setting and distortion of original data. Therefore, it has been widely used in practical calculation and analysis. As for the research on the BCSN, the existing research is mainly conducted from the perspective of network analysis. Diebold and Yilmaz (2013) investigated the linkages between actual outputs of G7 countries from 1962 to 2010 by network analysis, and found that the indicator of density could measure the pairwise output fluctuation linkages between different countries. At the same time, global connectedness will change as the business cycle changes. Gomez et al. (2013) adopted correlation coefficient matrix and network analysis method to systematically investigate the co-movement of business cycles synchronization in various countries since 1950, and believed that the dynamic changes of interdependence among countries was mainly driven by the co-movement of regional economic growth rather than co-movement of world economy. Caraiani (2013) found that compared with the Granger causality method, using correlation coefficients to construct a directed business cycle synchronization network can reflect the relative influence of countries in the world economy more reasonably, and further empirical findings showed that the United States finally occupies the core position of the business cycle synchronization network of G7 and OECD. Papadimitriou et al. (2014) used Pearson correlation coefficient and minimum dominating set to make an empirical study on the BCSN structure of 22 EU sample member states, and found that after the adoption of the common currency euro, the output of member states had a higher correlation, and the BCSN density of EU was increasing. Xi et al. (2014) constructed the BCSN of G7 based on the pairwise maximum entropy model, and found that the network presented a clustering hierarchy and nearly accounted for almost half of the entire structure of the interactions within the G7 system. Antonakakis et al. (2015) used sign concordance index and threshold-minimum dominating set method to investigate topological characteristics of BCSN among 27 countries during 1875–2013 and find that there are obvious differences in node degree of different countries in different periods. Gomez et al. (2017) used correlation coefficient and minimum spanning tree technique to construct the BCSN of EU, analyzed the business synchronization linkages and accessibility among member states, and found that there was no obvious core-periphery structure inside the network. Ductor and Leva-Leon (2016) adopted the social network analysis method and the indicator of betweenness centrality to evaluate the relative influence of various countries for global BCSN. It is found that a country’s is more influential in the network tends to increase when the economy is in recession, but becomes less influential when the economy is expanding. With Pearson correlation coefficient, rolling window, and threshold-minimum dominating set methods, Papadimitriou et al. (2016) selected some indicators such as the total number of edges, network density, the number of dominant and isolated nodes and node degree to empirically investigate the topological characteristics of European BCSN during 1986–2011. Based on the dynamic network analysis, Matesanz and Ortega (2016) used similarity index and minimum spanning tree (MST) technique to construct the European BCSN from 1950 to 2013, and found that the correlations and connectivity of the network increased significantly in 2009. With the help of correlation coefficient index used by Cerqueira (2013), Belke et al. (2017) investigated the business cycle synchronization of the European Monetary Union, focusing on the core-periphery mode of the business cycle within the Union after the economic crisis. Leiva-Leon (2017) established American inter-state BCSN by Markov Regime Switching framework, and investigated the its evolution model with indicators of multidimensional scaling (MDS) and closeness centrality, and found that the network has an obvious core-periphery structure. Sebestyén and Iloskics (2020) employed the pairwise Granger causality between national outputs to construct the global shock contagion network, and found that it has a relative long path length and stronger transmission, and the degree distribution tends to be asymmetric. For the research on the BSCN of BRI economies, Huang and Yao (2018) used CM (Cerqueira & Martins) synchronization index to measure the business cycle synchronization between China and BRI economies and find that it shows a certain “decoupling” trend, and there are obvious differences in different stages and different development aspects. Cui et al. (2020) used dynamic correlation coefficient to calculate and found that China and Southeast and Central Asian countries, as well as Mongolia, Nepal, Pakistan, Sri Lanka and other countries have a high business cycle synchronization level. Du et al. (2020) found that the BRI strengthened the business cycle synchronization linkages between China and ASEAN countries, while the network density and clustering coefficient also increased to some extent. Based on all the related literature above, the existing researches have made great progress in the measurement and of topological characteristics BCSN, which provides theoretical and methodological support for this study. However, the research on the BCSN of BRI is still in the preliminary exploratory stage, and its topological characteristics have not been clearly explained. Therefore, compared with previous studies, this paper may have three main contributions. Firstly, this paper systematically investigates the business cycle synchronization linkages among BRI economies by using the instantaneous quasi- correlation method, so as to reveal the relevant stylized facts. Different from existing studies, Huang and Yao (2018) and Cui et al. (2020) only analyzed the unidimensional business cycle synchronization between China and other BRI economies, but did not further reveal the pairwise multidimensional synchronization linkages among other countries. Secondly, 53 sample countries were selected to fully reveal the business cycle synchronization linkages among the BRI economies. Different from Du et al. (2020), which only studies the local BCSN between China and ASEAN, this paper selects more sample economies for analysis, which can enrich the existing researches. Thirdly, this paper not only analyzes the topological characteristics of BCSN from the perspectives of overall and individual characteristics, but also compares the topological characteristics of the two phases before (2000–2013) and after (2014–2019) the BRI, in order to reflect the impact of the BRI on the structural evolution of BCSN. To sum up, based on pairwise correlation matrix and network analysis method, this paper can better analyze the business cycle synchronization of BRI economies, and fully reveal the topological characteristics of the BCSN. # 3. Research methods and data description ## 3.1 Measurement of business cycle synchronization According to Duval et al. (2016), this paper uses instantaneous quasi- correlation method to measure the BRI business cycle synchronization. The advantage of this method is that it does not adhere to the limitation of the range of the traditional correlation coefficient -1 to 1 and considers calculation mode of difference method. Besides, it is needless to consider the window period and the selection of filtering method, which can be used for dynamic calculation analysis, and thus is widely used. The calculation formula of this method is as follows: $$S_{ijt} = \frac{\left( {y_{it} - {\overline{y}}_{i}} \right)\left( {y_{jt} - {\overline{y}}_{j}} \right)}{\theta_{i}\theta_{j}}$$ Where *y*_*it* and *y*_*jt* is the real GDP growth rate of country *i* and country *j* in year *t*, and ${\overline{y}}_{i}$ and ${\overline{y}}_{j}$ is the mean of real GDP growth rate, *θ*_*i* and *θ*_*j* is the standard deviation of real GDP growth rate of the two countries, and the real GDP growth rate is measured by the logarithmic difference method. could measure the inter-temporal business cycle synchronization between the two economies in the BRI, and finally form a symmetric instantaneous quasi-correlation coefficient matrix. # 3.2 Complex network analysis methods BCSN Construction. From the perspective of complex network, the network is a set composed of multiple nodes (social actors) and edges between nodes (linkages between actors). In fact, the international BCSN is a complex economic network, not just a simple collection of multiple economies, but also the linkages of business circle synchronization among them should be considered. According to complex network theory and graph theory, BRI economies are regarded as nodes, and the business cycle synchronization (i.e. the value of instantaneous quasi-correlation between economies are regarded as edges between nodes). Therefore, according to Antonakakis et al. (2015) and Papadimitriou et al. (2016), the BCSN of BRI could be expressed as: *G* = {*V*, *E*}, where *V* = {*v*₁, *v*₂,…, *v*_*n*} represents the node set composed of the BRI economies, *E* = {*e*₁, *e*₂,…, *e*_*n*} represents edges set composed of business cycle synchronization of BRI economies. Taken into account of the symmetry characteristics of the instantaneous quasi-correlation matrix, a undirected and unweighted BCSN composed of *N* nodes can be constructed, in which the number of nodes *N* is 53 sample economies. In order to do the network structure analysis, the mean value ($\sum_{i}^{N}\sum_{j}^{N}S_{ij}/N\left( {N - 1} \right)$) of the weighted synchronization matrix is usually taken as the threshold value. The above undirected and weighted BCSN could be unweighted to obtain the adjacency matrix ***A*** composed of 0 and 1. Specifically, the mean of the undirected and unweighted instantaneous quasi-correlation matrix is firstly calculated, and then the value of the instantaneous quasi-correlation and the mean between different economies in the matrix is compared. Finally, if the value greater than the mean is set to 1, and the actual effective synchronization linkage is indicated; if the value is set to 0, representing the invalid synchronization linkage. At this point, the non-diagonal elements ***a***_***i*,*j*** (*i* ≠ *j*) of adjacency matrix ***A*** can be defined as: $$\mathbf{a}_{\mathbf{i},\mathbf{j}} = \left\{ \begin{matrix} {1,\mspace{360mu}\text{if}\mspace{360mu}\text{S}_{\text{ij}} \geq \sum_{i}^{N}\sum_{j}^{N}S_{ij}/N\left( {N - 1} \right)} \\ {0,\mspace{360mu}\text{if}\mspace{360mu}\text{S}_{\text{ij}} < \sum_{i}^{N}\sum_{j}^{N}S_{ij}/N\left( {N - 1} \right)} \\ \end{matrix} \right.$$ Therefore, in an undirected and unweighted BCSN, the edges between different economies (nodes) reflects the validity of the business cycle synchronization linkage. For the same node (*i* = *j*), the value of the main diagonal in the corresponding matrix is zero. Next, by referring to Qiu and Liu (2021), this paper investigates the topological characteristics of the BCSN of BRI economies from the overall and individual sides. Indicators of overall characteristics of network structure. In this paper, indicators such as density, network efficiency, clustering coefficient, average path length and condensed subgroup are selected to analyze the overall topological characteristics of BCSN evolution of the BRI economies. Network density (DS) describes the degree of business cycle synchronization between the BRI economies, which can be calculated by the ratio of the actual effective linkage number *M* to the theoretical maximum linkage *N*(*N* − 1) number of the BCSN in year *t*. At this point, the network density can be calculated by DS_t = 2M_t/N(N − 1). Network efficiency (NE) reflects the accessibility of economic fluctuation correlation among the BRI economies, and is usually expressed as the reciprocal of distance between all nodes in year *t*, that is h_ijt. The corresponding formula is $\text{NE}_{\text{t}} = \sum_{\text{i} = 1}^{\text{N}}\sum_{\text{j} = 1}^{\text{N}}\text{h}_{\text{ijt}}/\text{N}\left( {\text{N} - 1} \right)$. Clustering coefficient (CC) reflects the overall degree of interconnection among all economies in the network and their close neighbors, so as to describe the degree of clustering among some nodes. When the degree of the node i in the year *t* is D_it, and the number of edges between it and all its neighboring nodes is E_it, so the formula of the CC is CC_t = E_it/N\[D_it(D_it − 1)\]. Average Path Length (AL) represents the mean of connected edges that the shortest path length between all potentially connected nodes in the network passes through. It reflects transmission efficiency of the business cycle synchronization linkage between nodes. Assuming that the shortest path length between node i and j in the network is d_ijt, the formula is AD_t = ∑_i∑_j d_ijt/(N³ − N) − 1/N. The Quadratic Assignment Procedure (QAP) correlation analysis is a non- parametric test method for calculating the correlations between matrices of different variables based on random matrix permutation. By analyzing the correlation and significance level of business cycle synchronization matrix at different times, it can reflect the dynamic evolutionary characteristic of network structures. The cohesive subgroups reflect the composition of subgroups and the tightness of node linkages in the network, which is correspondent to subgroup density matrix reflecting the tightness of the business cycle synchronization correlation between each subgroup and its external subgroup. According to the block model theory, network roles can be generally divided into four categories: main beneficiary subgroup, net spillover subgroup, broker subgroup and two-way spillover subgroup. Indicators of individual characteristics of network structure. In this paper, indicators such as eigenvector centrality, betweenness centrality, coreness are selected to analyze the individual topological characteristics of BCSN evolution of the BRI economies. Eigenvector centrality (EC) reflects the relative influence of nodes in the BCSN. It considers structure type of the network, which represents the weighted average sum of all direct and undirected connections of each node, that is, its value is affected by the centrality of neighboring nodes. Betweenness centrality (BC) measures the controlling ability of a node over the BCSN, that is to describe the “hub” role played by an economy in transmitting economic fluctuation influence in the network. Assuming that the number of shortest paths between node j and k in the network at year t is ${\overline{\text{N}}}_{\text{jkt}}$. and the corresponding total number of shortest paths is N_jkt, so the formula is $\text{BC}_{\text{it}} = \sum_{\text{j} \neq \text{k}:\text{j},\text{k} \neq \text{i}}{\overline{\text{N}}}_{\text{jkt}}/\text{N}_{\text{jkt}}$. Coreness (COR) reflects the core status of a node in the BCSN, and reveals the special structure of the business cycle synchronization between the core and peripheral nodes in the network through core-periphery analysis. # 3.3 Data description and source To the choice of sample, we refer to the standards from China’s BRI website (<https://eng.yidaiyilu.gov.cn/index.htm>). Comprehensively considering the availability and consistency of data, 53 sample BRI economies from 2000 to 2019 are selected as the research objects in this paper. More details of 53 sample BRI economies are available in. The real GDP data used in this paper is expressed as GDP constant 2010 US\$, and the data comes from the World Development Indicators (WDI) database in the World Bank. Based on Eqs and, we collate the matrices data of undirected and unweighted quasi-correlation coefficient for different period. Details of the relevant data can be found in. # 4. BCSN structure: Overall characteristics analysis ## 4.1 Analysis of network structure evolution As can be seen from, the network density and network efficiency of the BCSN from 2000 to 2019 have obvious stage characteristics, showing a fluctuating downward trend, indicating that the business cycle synchronization linkages between economies is still in a relative weak connection state. From the perspective of different periods, compared with 2000–2013, the value of DS and NE decreased during 2014–2019. Therefore, since the inception of the BRI in 2013, the degree of output synchronization linkages and the influence of corresponding economic linkages of BRI economies is weakened. It is shown that the real and effective synchronization relationship between countries involving in BRI from 2000 to 2012 was obviously adversely affected by the 2008 global financial crisis, and the accessibility of economic fluctuation and risk correlation was also rapidly declining, which also indicated that countries adopted relatively active anti-business cycle prevention policies to resist the impact of the financial crisis. After the BRI was put forward in 2013, the closeness of BCSN has been strengthened, and the accessibility of corresponding economic links has also been improved, indicating that BRI aimed at building a mutually beneficial and win-win “community of shared future” had a positive impact on enhancing economic and trade cooperation and links among the BRI economics. Furthermore, the clustering coefficient decreased on the whole, and reached its peak during 2008–2010, indicating that the synchronization linkages among some economies inside the network had obvious clustering characteristics. Meanwhile, the shock of the global financial crisis in 2008 made the clustering characteristic more obvious. In addition, the average path length increased slightly on the whole, proving that the path length required for the transmission of the synchronization linkage is relatively short, that is, the influence of economic output and fluctuations of any country in the network only need to be transmitted once to reach other countries. In particular, compared with 2000–2013, the mean value of clustering coefficient during 2014–2019 decreased slightly, while the mean path length increased significantly, indicating that the clustering degree of output synchronization was weakened after the inception of the BRI. Further analysis revealed that the BCSN of the BRI economies always show a large clustering coefficient and a small average path length, which has the typical “small world” topological characteristic. From, correlation coefficient of business cycle synchronization matrix of BRI economies are basically near zero value, and most of them fail the significance test in 2000–2013, showing obvious weak related or unrelated characteristics. Hence, the BCSN of BRI economies does not have typical evolutionary characteristics of self-stability before the inception of the BRI. However, during 2014–2019, although the correlation of synchronization matrix is still low, it has gradually increased and passed the significance test. It is indicated that the positive correlation of synchronization matrix was dynamically strengthened, network structure gradually highlighted the progressive evolutionary characteristics as the proposal and practice of the BRI. Meanwhile, the correlation coefficient of the synchronization matrix during 2000–2013 and 2014–2019 is 0.060, which fails to pass the significance test. Therefore, there is no significant correlation for the BCSN of BRI economies before and after the inception of the BRI. ## 4.2 Structure analysis of network subgroup In order to better reveal the subgroup structure changes of BCSN of the BRI economics before and after the inception of the BRI, shows the change of density matrix of the BCSN and the role of subgroups during 2014–2019, and structure diagrams of the two network subgroups are obtained with the support of VOSviewer software (Figs). Furthermore, from the evolution of network structure, subgroup 3 and Subgroup 4 were the main parts of network structure during 2000–2003, while subgroup 1 (two-way overflow subgroup) and subgroup 2 (broker subgroup) were the main parts of network structure during 2014–2019. This shows that the BRI has strengthened the business cycle synchronization linkages of BRI economies, and formed a spreading correlated structure in space with China, Singapore, India, Saudi Arabia, Turkey, Russia and Kazakhstan as main nodes. The number of nodes in subgroup 1 has increased, mainly including countries such as China, Singapore and Turkey, and two other isolated nodes (Nepal and Yemen). After the inception of the BRI, the synchronization density of nodes inside subgroup 1 was relatively large and had the largest increase. Secondly, the subgroup 1 connection degree with subgroup 3 increased, while the one with Subgroup 2 and subgroup 4 decreased significantly. The number of internal received relations of the subgroup 1 is 222, the number of external sent relations is 205, the corresponding expected relation ratio is 0.308, and the actual relation ratio is 0.520, indicating that the Subgroup 1 is a two-way spillover group. Subgroup 2 mainly contains India, Israel, Russia and other countries, and the number of its internal nodes has increased significantly. In addition to the decline of synchronization degree of internal connection, the synchronization degree of external connection has increased significantly. The number of internal received and external sent relations of subgroup 2 is 290 and 244 respectively, and the expected relation ratio and actual relation ration is 0.365 and 0.543 respectively, which has typical broker group characteristics. It can be seen that after the inception of the BRI, the internal connection degree of subgroup 2 decreased, but its external connection degree increased significantly, and it took an important mediating effect in the whole network. The number of nodes inside subgroup 3 decreased significantly, and a relatively isolated subgroup containing only Indonesia and Malaysia was finally formed. At this time, synchronization linkages inside subgroup 3 became closer after the inception of the BRI. And the synchronization degree of subgroup 3 with subgroup 1 and subgroup 2 increased while the one with subgroup 4 decreased. The number of linkages received inside and sent out from subgroup 3 is 2 and 17 respectively and the expected relation ratio and actual relation ratio is 0.019 and 0.105, respectively, showing obvious outward spillover characteristics and belonging to the net spillover group. The number of nodes inside subgroup 4 has decreased mainly including countries like Vietnam, the Philippines, Pakistan, Poland and Romania. Synchronization linkages inside subgroup 4 became looser after the inception of the BRI, while the linkage degree with other three subgroups increased. The number of linkages received inside and sent out from subgroup 4 is 142 and 60, and the expected relation ratio and actual relation ratio is 0.250 and 0.703, respectively. It is a typical beneficiary in the BCSN and belongs to the beneficiary group. # 5. BCSN structure: Individual characteristics analysis In order to better reveal the evolution of individual characteristics of BCSN structure before and after the inception of BRI, this section will mainly show the individual characteristic values and their changes and corresponding rankings in main BRI economies from 2014 to 2019, and separately analyze the results and rankings of China’s individual characteristic in 2000–2019. ## 5.1 Node centrality analysis shows that, on the whole, the value of EC and BC of main BRI economies are still low. There are obvious transnational differences within different regions. This shows that the influence of main BRI economies in the BCSN still needs to be improved, and they have not effectively played the role of “hub”. After comparison, it is found that after the inception of the BRI, the relative influence of East Asia and Central Asia in the network is more prominent, while central and Eastern Europe and Central Asia make a better mediating effect. At the same time, some economies, such as China, Russia, Ukraine, Kazakhstan and Tajikistan, do not match the influence and mediating effect in the network, but perform better than other economies as a whole. For different geographic regions, EC of East Asian ranked first, but its CC ranked fourth. Among it, China’s EC improved while CC had an obvious decrease. It shows that East Asia has a certain influence, but the core hub role is not prominent. Further combined with, during the sample period, China’s relative influence in the network has significantly increased, and it has more influence after the inception of the BRI, while its CC is relatively low. Although the overall level of centrality in Southeast Asia has improved, its ranking is still relatively low. Singapore’s relative influence is stronger than Thailand’s, but Thailand’s hub role is more obvious. Both EC and CC of South Asia declined and ranked the last. Meanwhile, the centrality levels of Pakistan and India inside the region were low, which did not exert a certain influence and mediating effect on the regional and external business synchronization. Central Asia, which is deeply inland and geographically close to China, has the second highest level of centrality, in which Kazakhstan has a higher relative influence and Tajikistan plays a more prominent mediating role. The centrality levels of West Asia and North Africa have improved, while the relative influence of the node countries Saudi Arabia and Egypt has declined, and they have not played an obvious hub role. There are significant differences in the EC and CC of Central and Eastern Europe (CEE), with their relative influence reduced and far inferior to their hub role and there are also some significant differences in node countries inside the CEE. ## 5.2 Node coreness analysis It can be seen from that, on the whole, the BCSN of BRI economies is characterized by the coexistence of “multi-core” and “multi-periphery” and consist of multiple levels of cores, semi cores, and peripheries. Similar to the results of node centrality, there are significant transnational differences in coreness degree within different regions. Specifically, East Asia, West Asia and North Africa have a higher coreness level and occupy the relative core position of the network, while Central Asia and Southeast Asia are in the intermediate zone between the core and the periphery of the network, CEE and South Asia are in the periphery of the network. Combined with, it can be seen that China’s coreness level fluctuated from 0.068 to 0.168 during the sample period. The corresponding ranking reached first, which was similar to the EC calculated in the last section. It is shown that after the inception of the BRI, China is accelerating its integration into the BCSN which has a greater impact on output changes in other economies, and eventually occupies the core position of the network. This may be due to China’s strong driving force for economic growth and higher level of opening-up policy, as well as its systematic and reliable economic and financial risk prevention policy, and the formation of a good and close relationship of coordinated economic development with the BRI economies. From the perspective of different regions, although the coreness level of Mongolia is obviously worse than that of China, the gap between the two is small, contributing to the top ranking of East Asia as a whole. Singapore and Thailand achieved synchronized growth in coreness and tied for seventh place, but other countries in the region did not achieve high coreness level, resulting in the overall lagging behind. In South Asia, Pakistan and India are significantly lower in the coreness level and ranking, and at the periphery of the network. The coreness level of West Asia and North Africa has been improved and ranks second, but the core influence of Egypt and Saudi Arabia in the regional and external business synchronization linkage is not prominent. CEE ranked last in terms of coreness level, with Ukraine significantly higher than Russia. Central Asia has improved its coreness level, and Kazakhstan’s core position is obviously better than Tajikistan. # 6. Conclusion and discussion With the support of instantaneous quasi-correlation and complex network analysis method, this paper empirically analyzed the topological characteristics of BCSN of BRI economies from 2000 to 2019. The main research conclusions are as follows: First, in the sample period, the business cycle synchronization linkage of BRI economies is still relative weak, the network density and network efficiency has decreased after the inception of the BRI. Second, the BCSN of BRI economies always show a large clustering coefficient and a short average path length, which presents a typical structural characteristic of “small world”. Meanwhile, the clustering degree of output synchronization linkage of BRI economies is weakened after the inception of the BRI. Third, on the whole, the BCSN of BRI economies does not have the characteristic of gradual evolution. But since the inception of the BRI, the evolution of the BCSN of BRI economies shows a self-stability characteristic. Fourth, the BCSN structure is composed of four subgroups. The synchronization linkage level inside subgroups is obviously higher than the one outside subgroups. After the inception of the BRI, the BCSN structure of BRI economies mainly consist of two-way spillover subgroup and broker subgroup. Fifth, the individual characteristics of the BRI economies are obviously different, the relative influence of a country in the network does not fully show its hub role. After the inception of the BRI, the function of China and other important nodes, such as Southeast Asia, Central Asia and CEE, has been enhanced. From the perspective of different regions, East Asia plays a relatively big role in the network, CEE and Central Asia take the most prominent mediating effect. Sixth, China’s EC and coreness level have increased and ranked top during the sample period, but its CC level is still low. After the inception of the BRI, China’s relative influence in the entire network has increased significantly, but does not show much mediating effect. Combined with instantaneous quasi-correlation and complex network analysis method, this paper reveals the structural characteristics of the BCSN of BRI economies from the overall and individual aspects. It should be noted that there are still some deficiencies in this paper, which need to be improved by follow- up research. Specifically, future research can be improved from the following three aspects. First of all, besides the instantaneous quasi-correlation method, other methods can be used to measure the business cycle synchronization, such as Markov switching model, dynamic correlation coefficient method, GARCH Model, concordance index and difference method, to ensure the robustness of empirical results. In addition, combined with cutting-edge complex network analysis methods, it is considered to make a weighted business cycle synchronization matrix, and further investigate the characteristics of the BCSN of BRI economies such as robustness and vulnerability, so as to enrich the existing research. Finally, Multiple Regression Quadratic Assignment Procedure (MRQAP) could be used to empirically study the driving factors of the structure evolution of the BCSN of BRI economies, which may provide some policy suggestions for strengthening the business cycle synchronization linkage of BRI economies. # Supporting information We are grateful to anonymous reviewers for constructive and insightful comments. All remaining errors are the responsibility of the authors alone. 10.1371/journal.pone.0270333.r001 Decision Letter 0 Nishikawa Takashi Academic Editor 2022 Takashi Nishikawa 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. 4 Apr 2022 PONE-D-21-31756 Topological Characteristics of International Business Cycle Synchronization: A Network Analysis of the BRI Economies PLOS ONE Dear Dr. Sichao, 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. Two reviewers have evaluated your manuscript and provided the comments below (and in the attached PDF copy of the manuscript). While the reviewers indicated interest in the topic, they raised a number of significant issues that need to be addressed in order for the manuscript to be suitable for publication. Please revise to address all comments. The revised manuscript will be re-evaluated by these reviewers, if they are available. Please submit your revised manuscript by May 19 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, Takashi Nishikawa, Ph.D. 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\. Please ensure that you include a title page within your main document. We do appreciate that you have a title page document uploaded as a separate file, however, as per our author guidelines (<http://journals.plos.org/plosone/s/submission-guidelines#loc-title-page>) we do require this to be part of the manuscript file itself and not uploaded separately. Could you therefore please include the title page into the beginning of your manuscript file itself, listing all authors and affiliations. 3\. In your Data Availability statement, you have not specified where the minimal data set underlying the results described in your manuscript can be found. PLOS defines a study's minimal data set as the underlying data used to reach the conclusions drawn in the manuscript and any additional data required to replicate the reported study findings in their entirety. All PLOS journals require that the minimal data set be made fully available. For more information about our data policy, please see <http://journals.plos.org/plosone/s/data- availability>. Upon re-submitting your revised manuscript, please upload your study’s minimal underlying data set as either Supporting Information files or to a stable, public repository and include the relevant URLs, DOIs, or accession numbers within your revised cover letter. For a list of acceptable repositories, please see <http://journals.plos.org/plosone/s/data-availability#loc-recommended- repositories>. Any potentially identifying patient information must be fully anonymized. Important: If there are ethical or legal restrictions to sharing your data publicly, please explain these restrictions in detail. Please see our guidelines for more information on what we consider unacceptable restrictions to publicly sharing data: <http://journals.plos.org/plosone/s/data-availability#loc- unacceptable-data-access-restrictions>. Note that it is not acceptable for the authors to be the sole named individuals responsible for ensuring data access. We will update your Data Availability statement to reflect the information you provide in your cover letter. \[Note: HTML markup is below. Please do not edit.\] 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: Partly Reviewer \#2: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No 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: 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: 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: Please see my detailed comments in your PDF manuscript as attached. \- Recalling the inception of the BRI in 2013, data analysis should be made: \- Remove redundant description of the BRI. \- Check several typos. Change "BRI economics" to BRI economies. \- Implcation should be rewritten. Reviewer \#2: General comments The paper is interesting, its structure is correct, and the topic is clearly a hot one. Results are fine unless not surprising. However, from my point of view, the following main issues make its publication not recommended at its current state. 1\. Introduction is extremely long and not very informative. At the end of it, we still do not know what the “BRI” is (who is financing the initiative, what for, what the achievements are, what kind of projects have been implemented, and so on) 2\. The paper, from my point of view, is very poorly written and difficult to understand. There are plenty of examples, I will show just a couple of them: • Page 8, “With the continuous advancement of economic globalization and international division of labor, the correlation trend of output periodicity and economic scale among countries presents a certain nonlinear and multi-threaded complex relationship, and then forms an interactive BCSN in space” • Page 9, “are regarded as nodes in this paper, and the quasi-correlation coefficients between the actual economic scales of economics are regarded as the connecting edges between nodes” 3\. Figures throughout the paper are plotted in an extremely low quality and they are no informative 4\. The paper shows an important number of “speculative” statements, especially when it comes to policy advice. For instance, page 13, the authors claim: “It is worth noting that in 2019, network density and efficiency saw a significant decline, which was mainly caused by the unilateral trade protectionism and continuous trade frictions among western countries led by the US in recent years, which worsened the previous positive global economic and trade environment”. However, the authors do not analyse the variables affecting output synchronicity and consequently they cannot state such connection. In page 4 the authors claim that their paper will provide policy inspiration to enhance economic synchronicity among BRI countries, however this is not possible with the analysis the authors provide. 5\. References regarding network analysis applied to synchronization is quite poor. The authors just cite 3 or 4 papers dealing with this issue but many more can be found and should be referenced in the paper. In the same line, the authors have not deeply revised the economic literature on business cycle synchronization. 6\. I do not understand Table 1. In it the Networks density measure is provided. This measure is shown for different years. However, if I am correct, the edges of the network represent correlation degree while nodes represent countries. Therefore, what is the meaning of this measure for year 2000, which is the first year of analysis? Or the same measure for 2019? This is not clear to me \*\*\*\*\*\*\*\*\*\* 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.0270333.r002 Author response to Decision Letter 0 13 May 2022 Responses to Reviewer \#1: Dear reviewer: Thank you for your comments on our manuscript. Those comments are valuable and very helpful for revising and improving our paper. We have studied the comments carefully and have made corrections which we hope to meet with approval finally. Below we address each issue the reviewers raised and describe corresponding changes in the revised manuscript. For ease of reference, changes were made by using blue text in the revised manuscript. Reviewers’ comments are repeated in full using italics, while our responses are typed in standard. 1.General Comment: Please see my detailed comments in your PDF manuscript as attached. Response: Thank you for your valuable time and efforts in this review process, and thanks for your excellent comments that could improve the quality of our manuscript. Following you detailed comments in PDF manuscript, we significantly improved the quality of this manuscript based on your constructive comments. Here we submitted a new version, which has been further revised according to your suggestions. In summary, major modifications include: First, we have carefully rewritten the abstract according to the research objectives, research methods, data, and the framework of the main findings Second, in the first introduction part, we introduced BRI in a more concise way, restated the research questions in this paper, and expounded the research content more accurately. Third, in the second part of related literature, we reorganized relevant literature, made some improvements in the measurement of business cycle synchronization, the research on the structural characteristic of the BCSN, and summarized the shortcomings of existing research and the contribution of this paper Fourth, in the third part of research methods and data description, we mainly improved the description of complex network analysis methods, especially the construction process of BCSN, and added the calculation formula of structural characteristic indicators. Fifth, in the fourth part of the overall characteristics analysis, we deleted the redundant results of weighted clustering coefficient (WCC), added measuring results and the analysis before and after the inception of the BRI, redrew the diagram of network subgroup structure with VOSviewer software, and elaborated overall structural characteristics of the BCSN from the network structure evolution and network subgroup structure.  Sixth, in the part of individual characteristics analysis, we deleted the redundant node strength (NS) results, improved the interpretation of the measurement results before and after the inception of the BRI, and illustrated the individual structural characteristics of the BCSN from node centrality and coreness degree. Seventh, in the conclusion and discussion part, we rerefined the research conclusions, deleted the original policy recommendations, and added the discussion on future research prospects. Eighth, we tried our best to correct punctuation marks, misspellings, grammar, references and other details in the paper, and deleted the “speculative” statements, in order to improve the quality of this paper and meet relevant requirements. 2.Major comments Comment 1: Recalling the inception of the BRI in 2013, data analysis should be made. Response 1: Thanks for your constructive comments. Following your suggestions, we had added the data analysis to recalling the inception of the BRI. The detailed revision is shown on pages 1-2 and below: The BRI comprises of over 50 different economies, which cover 80% of globe population and its the estimated cost is over \$21.1 trillion US dollars (Klinger 2019). Since the BRI was proposed in 2013, major achievements have been made in infrastructure construction, international trade, cross-border investment and sustainable growth. One World Bank report\[ For a more details of the achievements of the BRI, please refer to the World Bank report Belt and Road Economics:  Opportunities and Risks of Transport Communities, <https://www.worldbank.org/en/topic/regional- integration/publication/belt-and-road-economics-opportunities-and-risks-of- transport-corridors>.\] noted that BRI is expected to increase real income by 1.2-3.4% in BRI economies and 0.7%-2.9% globally, with investments that could lift more than 7.6 million people out of extreme poverty. In particular, BRI transport projects have significantly reduced trade costs, which are expected to increase trade in BRI economies by 2.8% to 9.7%, increase in world trade by 1.7% to 6.2%, and total foreign direct investment in BRI economies by 4.97% (World Bank, 2019).  Comment 2: Remove redundant description of the BRI. Response 2: Thanks for your point. Following your suggestions and detailed comments in PDF manuscript, we had removed redundant description of the BRI in the second paragraph of the first page. The detailed revision is shown on pages 1-2 and below: The BRI was initiated by the Chinese president Xi Jinping in 2013, and it was completely introduced during the visit of Kazakhstan and Indonesia. In the background of the history of China's ancient overland and maritime Silk Road, the BRI consists of the Silk Road Economic Belt and the 21st Century Maritime Silk Road, geographically crossing Asia, Europe and Africa. The BRI comprises of over 50 different economies, which cover 80% of globe population and its the estimated cost is over \$21.1 trillion US dollars (Klinger 2019). Since the BRI was proposed in 2013, major achievements have been made in infrastructure construction, international trade, cross-border investment and sustainable growth. One World Bank report\[ For a more details of the achievements of the BRI, please refer to the World Bank report Belt and Road Economics:  Opportunities and Risks of Transport Communities, <https://www.worldbank.org/en/topic/regional- integration/publication/belt-and-road-economics-opportunities-and-risks-of- transport-corridors>.\] noted that BRI is expected to increase real income by 1.2-3.4% in BRI economies and 0.7%-2.9% globally, with investments that could lift more than 7.6 million people out of extreme poverty. In particular, BRI transport projects have significantly reduced trade costs, which are expected to increase trade in BRI economies by 2.8% to 9.7%, increase in world trade by 1.7% to 6.2%, and total foreign direct investment in BRI economies by 4.97% (World Bank, 2019). The BRI has strengthened cooperation in trade, investment, infrastructure construction, institutional and cultural exchange among Asia, Europe and Africa, and formed a new economic network (Anwar et al., 2021). It aims to build a community of shared future for mankind, and it is an important platform for international multilateral cooperation and regional integration cooperation. Ultimately, it will make the world economy and society more open, inclusive, balanced and benefit-sharing.\[ More details about the BRI could check the website of Belt and Road Portal, <https://eng.yidaiyilu.gov.cn/index.htm>. \] However, there are still significant political, cultural, social, economic and institutional differences among the BRI economies, especially in terms of economic growth and their policies.\[ Such differences are not only between China and BRI economies, but also within BRI economies.  Thus, these transnational differences are a collection of differences between many parties.  \] Comment 3: Check several typos. Change “BRI economics” to “BRI economies”. Response 3: We appreciated this important reminder, and Thanks for spotting this typo. Following your suggestions, we changed “BRI economics” to “BRI economies” in this paper, and carefully checked the similar misspellings. Comment 4: Implication should be rewritten. Response 4: Thanks for your excellent comments that could improve the quality of our manuscript. Following your suggestions and detailed comments in PDF manuscript, we had rewritten the implication. Based on your comments, we make the following amendments: First, we changed the title of the last part of the paper (page 22) from “Conclusion” to “Conclusion and Discussion”. Second, we summarized the conclusions of this paper and discussed the structural characteristics of BCSN before and after the inception of the BRI. Third, this paper only revealed the structural characteristics of BCSN of BRI economies and did not make regression test on the driving factors of the BCSN, so this paper could not make corresponding policy suggestions. So, we deleted the policy implications in the introduction part and the policy recommendations in the last paragraph (page 26), and added the discussion of future research directions. The related revision is shown on pages 22-24 and below: With the support of instantaneous quasi-correlation and complex network analysis method, this article empirically analyzed the Topological characteristics of BCSN of BRI economies from 2000 to 2019. The main research conclusions are as follows: First, in the sample period, the business cycle synchronization linkage of BRI economies is still relative weak, the network density and network efficiency has decreased after the inception of the BRI. Second, the BCSN of BRI economies always show a large clustering coefficient and a short average path length, which presents a typical structural characteristic of “small world”. Meanwhile, the clustering degree of output synchronization linkage of BRI economies is weakened after the inception of the BRI. Third, on the whole, the BCSN of BRI economies does not have the characteristic of gradual evolution. But since the inception of the BRI，the evolution of the BCSN of BRI economies shows a self-stability characteristic. Fourth, the BCSN structure is composed of four subgroups. The synchronization linkage level inside subgroups is obviously higher than the one outside subgroups. After the inception of the BRI, the BCSN structure of BRI economies mainly consist of two-way spillover subgroup and broker subgroup. Fifth, the individual characteristics of the BRI economies are obviously different, the relative influence of a country in the network does not fully show its hub role. After the inception of the BRI, the function of China and other important nodes, such as Southeast Asia, Central Asia and CEE, has been enhanced. From the perspective of different regions, East Asia plays a relatively big role in the network, CEE and Central Asia take the most prominent mediating effect. Sixth, China's EC and coreness level have increased and ranked top during the sample period, but its CC level is still low. After the inception of the BRI, China's relative influence in the entire network has increased significantly, but does not show much mediating effect. Combined with instantaneous quasi-correlation and complex network analysis method, this paper reveals the structural characteristics of the BCSN of BRI economies from the overall and individual aspects. It should be noted that there are still some deficiencies in this paper, which need to be improved by follow- up research. Specifically, future research can be improved from the following three aspects. First of all, besides the instantaneous quasi-correlation method, other methods can be used to measure the business cycle synchronization, such as Markov switching model, dynamic correlation coefficient method, GARCH Model, concordance index and difference method, to ensure the robustness of empirical results. In addition, combined with cutting-edge complex network analysis methods, it is considered to make a weighted business cycle synchronization matrix, and further investigate the characteristics of the BCSN of BRI economies such as robustness and vulnerability, so as to enrich the existing research. Finally, Multiple Regression Quadratic Assignment Procedure (MRQAP) could be used to empirically study the driving factors of the structure evolution of the BCSN of BRI economies, which may provide some policy suggestions for strengthening the business cycle synchronization linkage of BRI economies. \[References\] Anwar M. A., S. Nasreen, and A. K. Tiwari. Forestation, Renewable Energy and Environmental Quality: Empirical Evidence From Belt and Road Initiative Economies\[J\]. Journal of Environmental Management, 2021, 291(8):112684-112683. Klinger J. Environment, Development, and Security Politics in the Production of Belt and Road Spaces\[J\]. Territory Politics Governance, 2019,8(5):657-675. World Bank. Belt and Road Economics: Opportunities and Risks of Transport Corridors\[R\]. Washington, DC: World Bank. 2019. Responses to Reviewer \#2: Dear reviewer: Thank you for your comments on our manuscript. Those comments are valuable and very helpful for revising and improving our paper. We have studied the comments carefully and have made corrections which we hope to meet with approval finally. Below we address each issue the reviewers raised and describe corresponding changes in the revised manuscript. For ease of reference, changes were made by using blue text in the revised manuscript. Reviewers’ comments are repeated in full using italics, while our responses are typed in standard. 1.General Comment: The paper is interesting, its structure is correct, and the topic is clearly a hot one. Results are fine unless not surprising. However, from my point of view, the following main issues make its publication not recommended at its current state. Response: Thank you for your valuable time and efforts in this review process, and Thanks for your excellent comments that could improve the quality of our manuscript. We significantly improved the quality of this manuscript based on your constructive comments. Here we submitted a new version, which has been further revised according to your suggestions. In summary, major modifications include: First, in the first introduction part, we introduced BRI in a more concise way, restated the research questions in this paper, and expounded the research content more accurately. Second, in the second part of related literature, we reorganized relevant literature, made some improvements in the measurement of business cycle synchronization, the research on the structural characteristic of the BCSN, and summarized the shortcomings of existing research and the contribution of this paper. Third, in the third part of research methods and data description, we mainly improved the description of complex network analysis methods, especially the construction process of BCSN, and added the calculation formula of structural characteristic indicators. Fourth, in the fourth part of the overall characteristics analysis, we deleted the redundant results of weighted clustering coefficient (WCC), added measuring results and the analysis before and after the inception of the BRI, redrew the diagram of network subgroup structure with VOSviewer software, and elaborated overall structural characteristics of the BCSN from the network structure evolution and network subgroup structure.  Fifth, in the part of individual characteristics analysis, we deleted the redundant node strength (NS) results, improved the interpretation of the measurement results before and after the inception of the BRI, and illustrated the individual structural characteristics of the BCSN from node centrality and coreness degree. Sixth, in the conclusion and discussion part, we refined the research conclusions, deleted the original policy recommendations, and added the discussion on future research prospects. Seventh, we tried our best to correct punctuation marks, misspellings, grammar, references and other details in the paper, and deleted the “speculative” statements, in order to improve the quality of this paper and meet relevant requirements. 2.Major comments Comment 1: Introduction is extremely long and not very informative. At the end of it, we still do not know what the “BRI” is (who is financing the initiative, what for, what the achievements are, what kind of projects have been implemented, and so on) Response 1: Thanks for your constructive comments that could improve the quality of our manuscript. Following your suggestions, we have simplified and adjusted the content of the introduction. To be specific, first of all, we have rewritten the introduction of BRI to highlight the origin, coverage, achievements, goals and challenges of BRI. Secondly, based on the current challenges BRI is facing, this paper introduces relevant research on business cycle synchronization, and reveals its key issues. Then, in order to reveal the complex interdependence among economies, we introduce the complex network analysis method and its application in analyzing the business cycle synchronization linkage. Finally, taking BRI economies as the research object, we elaborate the problems to be solved in this paper and introduce specific research methods. For more detailed modifications, please refer to the introduction part (page 1-4). More detialed introduction about BRI could refer to the website of Belt and Road Portal (<https://eng.yidaiyilu.gov.cn/index.htm>). In addition, in order to facilitate reviewer's understanding, we make a separate summary here. About the origin of BRI, based on the history of China's ancient Silk Road, Chinese President Xi Jinping proposed the Silk Road Economic Belt and the 21st Century Maritime Silk Road during his visit to Kazakhstan and Indonesia in 2013, thus forming the whole Belt and Road Initiative. As the initiator and leader of BRI, China hoped to strengthen economic cooperation among BRI economies through this initiative and jointly realize sustainable economic and social development. About the coverage of BRI, the Belt and Road is composed of more than 50 different economies, covering 80% of the global population with an economic scale of over 21 trillion US dollars (Klinger, 2019). BRI is an effort to create jointly-built trade routes that emulate the ancient Silk Road and promote regional cooperation in Asia, Europe, and Africa. About the development achievements of BRI, it has strengthened cooperation in trade, investment, infrastructure construction, institutional and cultural exchange among Asia, Europe and Africa, and formed a new economic network (Anwar et al., 2021). Since the inception of the BRI in 2013,a lot of achievements in infrastructure development, international trade, cross-border investment and sustainable growth have been made. One World Bank report\[ For a more details of the achievements of the BRI, please refer to the World Bank report Belt and Road Economics:  Opportunities and Risks of Transport Communities,  [https://www.worl dbank.org/en/topic/regional-  integration/publication/belt-and-road-economics- opportunities-and-risks-of-transport-corridors](https://www.worldbank.org/en/top ic/regional-  integration/publication/belt-and-road-economics-opportunities-and- risks-of-transport-corridors).  \] noted that BRI is expected to increase real income by 1.2-3.4% in BRI economies and 0.7%-2.9% globally, with investments that could lift more than 7.6 million people out of extreme poverty. In particular, BRI transport projects have significantly reduced trade costs, which are expected to increase trade in BRI economies by 2.8% to 9.7%, increase in world trade by 1.7% to 6.2%, and total foreign direct investment in BRI economies by 4.97%(World Bank, 2019). Based on policy coordination, facilities connectivity, unimpeded trade, financial integration and people-to-people bond, some great achievements have been made since the inception of the BRI in 2013. For policy coordination, at presen, China has signed more than 200 agreements on BRI cooperation with 149 countries and 32 international organizations, and successfully hosted two BRI Forums for international cooperation. For facilities connectivity, infrastructure construction cooperation between China and BRI economies in railways, ports, aviation, energy and communications has strengthened the infrastructure quality of BRI member states. For unimpeded trade, trade and investment among BRI economies have grown considerably, and BRI members have actively participated in the China International Import Expo. For financial integration, the Silk Road Fund and the Asian Infrastructure Investment Bank were set to provide reliable financial support for BRI construction projects. For people-to-people bond, China and BRI economies have deepened cooperation in cultural exchanges, scientific and technological innovation, and conducted a series of humanitarian assistance in medical care, poverty alleviation and food supply. About development goals of BRI, it aims to build a community of shared future, and it is an important platform for international multilateral cooperation and regional integration cooperation. Ultimately, it will make the world economy and society more open, inclusive, balanced and benefit-sharing.\[ More details about the BRI could check the website of Belt and Road Portal,<https://eng.yidaiyilu.gov.cn/index.htm>.\] About the challenge BRI is facing, there are still significant political, cultural, social, economic and institutional differences among the BRI economies, especially in terms of economic growth and their policies\[ Such differences are not only between China and BRI economies, but also within BRI economies.  Thus, these transnational differences are a collection of differences between many parties.  \]. Just because of the difference in economic base and growth, it poses certain challenges to promote BRI development. Comment 2: The paper, from my point of view, is very poorly written and difficult to understand. There are plenty of examples, I will show just a couple of them: • Page 8, “With the continuous advancement of economic globalization and international division of labor, the correlation trend of output periodicity and economic scale among countries presents a certain nonlinear and multi-threaded complex relationship, and then forms an interactive BCSN in space” • Page 9, “are regarded as nodes in this paper, and the quasi-correlation coefficients between the actual economic scales of economics are regarded as the connecting edges between nodes” Response 2: Thanks for your excellent comments, and we appreciated this important reminder. Following your suggestions, we tried our best to improve the English expression and grammar in this paper in order to meet the relevant requirements as far as possible. As the example sentence on page 8 you mentioned, after careful consideration, we have deleted it directly. According to Antonakakis et al.(2015) and Papadimitriou et al.(2016), we constructed BCSN directly using complex network analysis methods without additional explanation. As you mentioned on page 9, after careful analysis, we have changed the sentence as: “According to complex network theory and graph theory, BRI economies are regarded as nodes, and the business cycle synchronization (i.e. the value of instantaneous quasi-correlation) between economies are regarded as edges between nodes.” Comment 3: Figures throughout the paper are plotted in an extremely low quality and they are no informative. Response 3: Thanks for your excellent point. Following your suggestions, we had deleted the original figures, and redrew it by VOSviewer software. The new figures are shown on pages 16 and below: Fig. 1. Network Subgroups during 2000-2013. Source: Drawn by the authors from VOSviewer software. Fig. 2. Network Subgroups during 2014-2019. Source: Drawn by the authors from VOSviewer software. Comment 4: The paper shows an important number of “speculative” statements, especially when it comes to policy advice. For instance, page 13, the authors claim: “It is worth noting that in 2019, network density and efficiency saw a significant decline, which was mainly caused by the unilateral trade protectionism and continuous trade frictions among western countries led by the US in recent years, which worsened the previous positive global economic and trade environment”. However, the authors do not analyse the variables affecting output synchronicity and consequently they cannot state such connection. In page 4 the authors claim that their paper will provide policy inspiration to enhance economic synchronicity among BRI countries, however this is not possible with the analysis the authors provide. Response 4: Thanks for your constructive comments. It should be noted that the purpose of this study is to reveal the structural characteristics of the BCSN, and the driving factors of BCSN are not further investigated through regression test. Therefore, this paper is indeed unable to explain the results of BCSN structural characteristics, let alone make corresponding policy suggestions based on the existing results. In view of this, as you mentioned, we have deleted the “speculative” statements on page 4 and page 13 mentioned by experts, and re-examined and revised the “speculative” statements in the part of introduction, structural characteristics analysis and policy suggestions in the paper. In addition, it is worth noting that we have changed the title of the last part of the paper (page 22) from “Conclusion” to “Conclusion and Discussion”. And then, we summarized the conclusions of this paper and discussed the structural characteristics of BCSN before and after the inception of the BRI. In the end, we deleted the policy recommendations in the last paragraph (page 23-24), and added the discussion of future research directions. The related revision is shown on pages 22-24 and below: With the support of instantaneous quasi-correlation and complex network analysis method, this article empirically analyzed the Topological characteristics of BCSN of BRI economies from 2000 to 2019. The main research conclusions are as follows: First, in the sample period, the business cycle synchronization linkage of BRI economies is still relative weak, the network density and network efficiency has decreased after the inception of the BRI. Second, the BCSN of BRI economies always show a large clustering coefficient and a short average path length, which presents a typical structural characteristic of “small world”. Meanwhile, the clustering degree of output synchronization linkage of BRI economies is weakened after the inception of the BRI. Third, on the whole, the BCSN of BRI economies does not have the characteristic of gradual evolution. But since the inception of the BRI，the evolution of the BCSN of BRI economies shows a self-stability characteristic. Fourth, the BCSN structure is composed of four subgroups. The synchronization linkage level inside subgroups is obviously higher than the one outside subgroups. After the inception of the BRI, the BCSN structure of BRI economies mainly consist of two-way spillover subgroup and broker subgroup. Fifth, the individual characteristics of the BRI economies are obviously different, the relative influence of a country in the network does not fully show its hub role. After the inception of the BRI, the function of China and other important nodes, such as Southeast Asia, Central Asia and CEE, has been enhanced. From the perspective of different regions, East Asia plays a relatively big role in the network, CEE and Central Asia take the most prominent mediating effect. Sixth, China's EC and coreness level have increased and ranked top during the sample period, but its CC level is still low. After the inception of the BRI, China's relative influence in the entire network has increased significantly,but does not show much mediating effect. Combined with instantaneous quasi-correlation and complex network analysis method, this paper reveals the structural characteristics of the BCSN of BRI economies from the overall and individual aspects. It should be noted that there are still some deficiencies in this paper, which need to be improved by follow- up research. Specifically, future research can be improved from the following three aspects. First of all, besides the instantaneous quasi-correlation method, other methods can be used to measure the business cycle synchronization, such as Markov switching model, dynamic correlation coefficient method, GARCH Model, concordance index and difference method, to ensure the robustness of empirical results. In addition, combined with cutting-edge complex network analysis methods, it is considered to make a weighted business cycle synchronization matrix, and further investigate the characteristics of the BCSN of BRI economies such as robustness and vulnerability, so as to enrich the existing research. Finally, Multiple Regression Quadratic Assignment Procedure (MRQAP) could be used to empirically study the driving factors of the structure evolution of the BCSN of BRI economies, which may provide some policy suggestions for strengthening the business cycle synchronization linkage of BRI economies. Comment 5: References regarding network analysis applied to synchronization is quite poor. The authors just cite 3 or 4 papers dealing with this issue but many more can be found and should be referenced in the paper. In the same line, the authors have not deeply revised the economic literature on business cycle synchronization. Response 5: Thanks for your excellent comments that could improve the quality of our manuscript. Following your suggestions, after carefully re-reading and sorting out relevant literature, literature closely related to this paper mainly fall into the following two categories: One is the measurement research on the international business cycle synchronization (literature from economics view), the other is the related research on the BCSN (literature about network analysis ). Therefore, we readjusted and improved the related literature part, adding not only the economic literature on the measurement of business cycle synchronization, but also the literature on the structural characteristics of the BCSN. Specific modifications can be found in the related literature part on pages 5-8 in the paper. In order to facilitate the review by reviewers, we presented the revised related literature part as: The literature closely related to this paper mainly includes the following two categories: one is the measurement and research on the international business cycle synchronization; the other is on the BCSN. The measurement and research on international business cycle synchronization mainly includes static method and dynamic method.  For the static methods, existing studies mainly adopt simple correlation coefficient method (Giovanni and Levchenko, 2010; Papadimitriou et al., 2014) and dynamic factor model (Del Negro and Otrok, 2008; Kose et al., 2012). It should be noted that although the static method can intuitively judge the level of the synchronization, it cannot reveal the dynamic characteristics of the business cycle synchronization. Therefore, the follow-up measurement research has gradually shifted from the traditional static methods to the dynamic methods. For dynamic methods, existing studies mainly focus on Markov Regime Switching Model (Hamilton and Owyang, 2012; Leiva,Leon, 2014; Ductor and Leva-Leon, 2016), GARCH model (Savva et al., 2010; Antonakakis,2012), Concordance Index (Harding and Pagan, 2006; Cerqueira and Martins, 2009; Cerqueira, 2013), Difference Method (Kalemli-Ozcan et al., 2013;  Pyun and An, 2016) and instantaneous quasi- correlation method (Abiad et al., 2013;  Duval et al., 2016). In addition, some scholars used the Method of Hodrick-Prescott (HP) filter to de-trend the original output data series and further calculate the business cycle synchronization (Ng, 2010; Huang and Zhu, 2015). However, Hamilton (2018) believes that HP filtering method introduces unreal dynamic linkages that are not based on original data, and its results are affected by the size of smoothing parameter, so it cannot truly reflect the level of business cycle synchronization. Compared with other dynamic methods, the instantaneous quasi- correlation method is more efficient and could carry out dynamic calculation and analysis, which avoids problems caused by artificial parameter setting and distortion of original data. Therefore, it has been widely used in practical calculation and analysis (Abiad et al., 2013;  Yao and Tang, 2020). As for the research on the BCSN, the existing research is mainly conducted from the perspective of network analysis. Diebold and Yilmaz(2013) investigated the linkages between actual outputs of G7 countries from 1962 to 2010 by network analysis, and found that the indicator of connectedness(density) could measure the pairwise output fluctuation linkages between different countries. At the same time, global connectedness will change as the business cycle changes. Gomez et al.(2013) adopted correlation coefficient matrix and network analysis method to systematically investigate the co-movement of business cycles synchronization in various countries since 1950, and believed that the dynamic changes of interdependence among countries was mainly driven by the co-movement of regional economic growth rather than co-movement of world economy. Caraiani (2013) found that compared with the Granger causality method, using correlation coefficients to construct a directed business cycle synchronization network can reflect the relative influence of countries in the world economy more reasonably, and further empirical findings showed that the United States finally occupies the core position of the business cycle synchronization network of G7 and OECD. Papadimitriou et al. (2014) used Pearson correlation coefficient and minimum dominating set to make an empirical study on the BCSN structure of 22 EU sample member states, and found that after the adoption of the common currency euro, the output of member states had a higher correlation, and the BCSN density of EU was increasing. Xi et al. (2014) constructed the BCSN of G7 based on the pairwise Maximum Entropy Model, and found that the network presented a clustering hierarchy and nearly accounted for almost half of the entire structure of the interactions within the G7 system. Antonakakis et al. (2015) used sign concordance index and threshold-minimum dominating set method to investigate topological characteristcs of BCSN among 27 countries during 1875-2013 and find that there are obvious differences in node degree of different countries in different periods. Gomez et al. (2017) used correlation coefficient and minimum spanning tree technique to construct the BCSN of EU, analyzed the business synchronization linkages and accessibility among member states, and found that there was no obvious core-periphery structure inside the network. Ductor and Leva-Leon (2016) adopted the social network analysis method and the indicator of betweenness centrality to evaluate the relative influence of various countries for global BCSN. It is found that a country's is more influential in the network tends to increase when the economy is in recession, but becomes less influential when the economy is expanding. With Pearson correlation coefficient, rolling window, threshold-minimum dominating set and other methods, Papadimitriou et al. (2016) selected some indicators such as the total number of peripheries, network density, the number of dominant and isolated nodes and node degree to empirically investigate the topological characteristics of European BCSN during 1986-2011. Matesanz and Ortega(2016) used similarity index and Minimum Spanning Tree technique (MST) to construct the European BCSN from 1950 to 2013. Based on the dynamic network analysis, the correlations and connectivity of the network increased significantly in 2009. In addition, differences in window size, filtering method and similarity measure also affect the characteristics of the BCSN. With the help of correlation coefficient index used by Cerqueira (2013), Belke et al. (2017) investigated the the business cycle synchronization of the European Monetary Union, focusing on the core-periphery mode of the business cycle within the Union after the economic crisis. Leiva-Leon (2017) established American inter-state BCSN by Markov Regime Switching framework and investigated the its evolution model with indicators of MDS (multidimensional scaling) and closeness centrality. It is found that the network has an obvious core-periphery structure. Sebestyén and Iloskics (2020) employed the pairwise Granger causality between national outputs to construct the global shock contagion network, and found that it has a relative long path length and stronger transmission, and the degree distribution tends to be asymmetric. For the research on the BSCN of BRI economies, Huang and Yao (2018) used CM (Cerqueira & Martins) synchronization index to measure the business cycle synchronization between China and BRI economies and find that it shows a certain “decoupling” trend, and there are obvious differences in different stages and different development aspects. Cui et al. (2020) used dynamic correlation coefficient to calculate and found that China and Southeast and Central Asian countries, as well as Mongolia, Nepal, Pakistan, Sri Lanka and other countries have a high business cycle synchronization level. Du et al. (2020) found that the BRI strengthened the business cycle synchronization linkages between China and ASEAN countries, while the network density and clustering coefficient also increased to some extent. Comment 6: I do not understand Table 1. In it the Networks density measure is provided. This measure is shown for different years. However, if I am correct, the edges of the network represent correlation degree while nodes represent countries. Therefore, what is the meaning of this measure for year 2000, which is the first year of analysis? Or the same measure for 2019? This is not clear to me. Response 6: Thanks for your point. As you proposed, we used the quasi- correlation degree and nodes to represent the periphery of the network and the BRI economies respectively. In the complex network analysis part of this paper, we gave the meaning of the structural characteristics indicators, and then calculate the results of the annual network structural characteristics. At this point, we intended to show the results of overall structural characteristics on the BCSN of BRI economies in some years in Table 1, including density (DS), network efficiency (NE), clustering coefficient (CC), average distance (AD) and other indicators. Before explaining this problem, we have added the formula and definition of each indicator for your understanding. In this paper, density (DS) describes the degree of business cycle synchronization between the BRI economies. For better understanding, we added its formula as ，where represents the actual number of effective linkages in the BCSN in year. represents the nodes in the network. It should be noted that the DS is calculated by annual unweighted and undirected BCSN based on a symmetric Instantaneous Quasi-correlation matrix. At this time, for the BCSN in year, the actual number of effective linkages in the network equals to the actual effective business cycle synchronization linkages, while in theory, the maximum linkage number equals to =53\*52=2756. Furthermore, when the instantaneous quasi-correlation between two economies is greater than the average among all economies, it is indicated that the business cycle synchronization linkages between the two economies is actually effective, and it will be recorded as an actual effective linkage number. Since the output growth and growth rate of BRI economies vary significantly every year, the size of the instantaneous quasi- correlation among economies will also change, leading to changes in the actual number of effective linkages in the network. Meanwhile, the constant number of nodes in the network is 53, so theoretically the maximum linkage number does not change.   To sum up, no matter in 2000, 2019 or any other year, DS is calculated in the same way with the same meaning, that is, the ratio of the actual effective number of linkages in the network to the theoretical maximum number of linkages in year t. \[References\] Abiad A., D. Furceri S. Kalemli-Ozcan, and A. Pescatori. Dancing Together? Spillovers, Common Shocks, and the Role of Financial and Trade Linkages\[M\]. Washington: International Monetary Fund,World Economic Outlook 2013. Antonakakis N. Business Cycle Synchronization During US Recessions Since the Beginning of the 1870s\[J\]. Economics Letters, 2012,117(2):467-472. Anwar M. A., S. Nasreen, and A. K. Tiwari. Forestation, Renewable Energy and Environmental Quality: Empirical Evidence From Belt and Road Initiative Economies\[J\]. Journal of Environmental Management, 2021, 291(8):112684-112683. Belke A., Domnick C., Gros D. Business Cycle Synchronization in the EMU: Core vs. Periphery\[J\]. Open Economies Review, 2017, 28(5):1-30. Caraiani P. Using Complex Networks to Characterize International Business Cycles\[J\]. PLOS ONE, 2013,8(3), e58109:1-13. Cui, Q. Y., Y. Zhang, and S. Wang. International Macroeconomic Policy Coordination Under the BRI: Mechanism Basis and China’s Role\[J\]. Economists, 2020(8):49-58. Del Negro, M., and C. Otrok. Dynamic Factor Models with Time-Varying Parameters: Measuring Changes in International Business Cycles\[C\]. Federal Reserve Bank of New York Staff Reports, No. 326. 2008. Diebold F. X., Yilmaz K. Measuring the Dynamics of Global Business Cycle Connectedness \[C\]. PIER Working Paper No.13-070, 2013. Du, J. Z., C. Z., Wang, and J. L. Wang. An Analysis of Business Cycle Synchronization Between China and ASEAN Countries in the Context of the Belt and Road Initiative\[J\]. Journal of Finance and Economics Theory, 2020(4):23-31. Ductor L, Leiva-Leon D. Dynamics of Global Business Cycles Interdependence\[J\]. Journal of International Economics, 2016, 102:110-127. Duval R, Li N, Saraf R, Seneviratnea D. Value-added trade and business cycle synchronization\[J\]. Journal of International Economics, 2016,99:251-262． Giovanni D, J., Levchenko G. Putting the Parts Together: Trade, Vertical Linkages and Business Cycle Comovement\[J\]. American Economic Journal Macroeconomics, 2010, 2(2):95-124. Gomez D. M., Torgler B., Ortega G. J. Measuring Global Economic Interdependence: A Hierarchical Network Approach\[J\]. The World Economy, 2013,36(12):1632-1648. Gomez D. M, Ferrari H. J, Torgler B, Guillermo J. O. Synchronization and Diversity in Business Cycles: A Network Analysis of the European Union\[J\]. Applied Economics, 2017,49(10):1-15. Hamilton J. D. Why You Should Never Use the Hodrick-Prescott Filter\[J\]. Review of Economics and Statistics, 2018,100(5):831-843. Hamilton J. D. M. Owyang. The Propagation of Regional Recessions\[J\]. Review of Economics and Statistics, 2012, 94(4):935-947. Harding D., Pagan A. Synchronization of Cycles\[J\]. Journal of Econometrics, 2006,132(1):59-79. Huang Z. L., Yao T. T. Business Cycle Synchronization and its Transmission Mechanism between China and the Countries along the BRI\[J\]. Statistical Research, 2018,35(9):40-53. Huang Z. L., Zhu B. H. Real Business Cycle and Taxation Policy Effects in China\[J\]. Economic Research Journal, 2015,50(3):4-17+114. Kalemli-Ozcan S., Papaioannou E., Perri F. Global Banks and Crisis Transmission\[J\]. Journal of International Economics, 2013, 89:495-510. Klinger J. Environment, Development, and Security Politics in the Production of Belt and Road Spaces\[J\]. Territory Politics Governance, 2019,8(5):657-675. Kose M. A., Otrok C., and E. Prasad. Global Business Cycles: Convergence or Decoupling?\[J\]. International Economic Review, 2012,53(2):511-538. Leiva-Leon D. Measuring Business Cycles Intra-Synchronization in US: A Regime- switching Interdependence Framework\[J\]. Oxford Bulletin of Economics & Statistics, 2017,79(4):513-545. Matesanz D., Ortega G. J. On Business Cycles Synchronization in Europe: A Note on Network Analysis\[J\]. Physica A: Statistical Mechanics and its Applications. 2016; 462:287-296. Ng E. Y. Production Fragmentation and Business-cycle Co-movement\[J\]. Journal of International Economics, 2010, 82(1):1-14. Papadimitriou T, Gogas P, Sarantitis G A. European Business Cycle Synchronization: A Complex Network Perspective, Network Models in Economics and Finance\[M\]. 2014, in Chapter 13:265-275. Papadimitriou T., Gogas P., Sarantitis G. A. Convergence of European Business Cycles: A Complex Networks Approach\[J\]. Computational Economics, 2016,47:97-119. Pyun J. H., An B. J. Capital and Credit Market Integration and Real Economic Contagion During the Global Financial Crisis\[J\]. Journal of International Money & Finance, 2016, 67:172-193. Savva C. S., Neanidis K. C., Osborn D. R. Business Cycle Synchronization of the Euro Area with the New and Negotiating Member Countries\[J\]. International Journal of Finance & Economics, 2010, 15(3):288-306. Sebestyén T., and Z. Iloskics. Do Economic Shocks Spread Randomly?: A Topological Study of the Global Contagion Network\[J\]. PLOS ONE, 2020,15:e0238626. Xi N, Muneepeerakul R, Azaele S, Wang Y G. Maximum Entropy Model for Business Cycle Synchronization\[J\]. Physica A: Statistical Mechanics and its Applications, 2014,413:189-194. Yao W., and A. D. Tang. Financial Integration and Business Cycle Synchronization\[J\]. China Journal of Economics 2020,7(2):61-85. 10.1371/journal.pone.0270333.r003 Decision Letter 1 Nishikawa Takashi Academic Editor 2022 Takashi Nishikawa 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 Jun 2022 PONE-D-21-31756R1Topological Characteristics of International Business Cycle Synchronization: A Network Analysis of the BRI EconomiesPLOS ONE Dear Dr. Sichao, Thank you for submitting your manuscript to PLOS ONE. We have received two reviewer reports. Both reviewers now recommend publication of the manuscript. However, Reviewer \#1 provided a number of minor comments in the attached file. Thus, before formally accepting your manuscript for publication, we encourage you to consider these comments carefully, revise the manuscript accordingly, and resubmit. Please submit your revised manuscript by Jul 18 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 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, Takashi Nishikawa, Ph.D. 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: (No Response) 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 \#1: (No Response) Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\*  3\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: (No Response) Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\*  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: (No Response) Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\*  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 Response) Reviewer \#2: (No Response) \*\*\*\*\*\*\*\*\*\*  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: Dear Authors Please note that the changes in the revised version must be highlighted in red color so that reviwers can easily recongize them without time consuming work. It is professional way. Unfortunately, I can't see them in the revised version. I would like to decline review not because of the quality of your paper but because of formality issue. 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 \#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.0270333.r004 Author response to Decision Letter 1 5 Jun 2022 Responses to Reviewer \#1: Dear reviewer: Thank you for your comments on our manuscript. Those comments are valuable and very helpful for revising and improving our paper. We have studied the comments carefully and have made corrections which we hope to meet with approval finally. Below we response each issue that the reviewers raised, and describe corresponding changes by using red text in the revised manuscript. Following you detailed comments in PDF manuscript, we improved the quality of this manuscript based on your constructive comments. Here we submitted a new version, which has been further revised according to your suggestions. In summary, major modifications include: According to your advice, we have revised the abstract on page 1, footnote of introduction on page 3 and the format of references on page 25-29. In addition, we also checked the content of other parts of this paper and corrected the errors found. In the end, it should also be noted that in the last paragraph of page 10 and the formula (2) on page 11, due to the mistake in the last revised manuscript, we misused the row mean of the matrix as the threshold value for data processing. In fact, we should use the mean of the matrix as the threshold value instead, and the undirected and weighted BCSN could be unweighted to obtain the adjacency matrix A composed of 0 and 1. In view of this, we have modified the relevant explanations in the last paragraph of page 10 and the formula (2) on page 11. The above modifications do not affect the reliability and validity of the empirical results of this paper. 10.1371/journal.pone.0270333.r005 Decision Letter 2 Nishikawa Takashi Academic Editor 2022 Takashi Nishikawa 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. 9 Jun 2022 Topological Characteristics of International Business Cycle Synchronization: A Network Analysis of the BRI Economies PONE-D-21-31756R2 Dear Dr. Sichao, 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, Takashi Nishikawa, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Thank you for answering the final question regarding the threshold used to process the quasi-correlation matrices. Reviewers' comments: 10.1371/journal.pone.0270333.r006 Acceptance letter Nishikawa Takashi Academic Editor 2022 Takashi Nishikawa 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 Jun 2022 PONE-D-21-31756R2 Topological Characteristics of International Business Cycle Synchronization: A Network Analysis of the BRI Economies Dear Dr. Sichao: 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. Takashi Nishikawa Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Waxy maize (*Zea mays* L. *sinensis* Kulesh), also known as sticky maize, is one sub-type of maize that was first discovered in Southwestern China and then prevalently grown in other Asian countries. The endosperm of waxy maize has a high content of amylopectin (nearly 100%), and is thus characterized by high viscosity, easy digestion, and good light transmittance. These excellent characters and fresh harvest make waxy maize widely used in frozen food processing, paper-making and livestock feeding industries. However, due to the limited levels and types of essential amino acids, especially lysine, the nutritional value of waxy maize is relatively low. Generally, the lysine content in maize grain should be more than 0.5% (\>51 mg per gram of protein) to meet human and livestock requirements, but waxy maize has a lysine content of only 0.24–0.34%. By introgression of *opaque2* (*o2*) and *opaque2* modifier (*o2m*) alleles into elite maize inbred lines with marker-assisted selection (MAS) technique, the genetically modified *opaque2* maize, also known as quality protein maize (QPM), shows an improved lysine content of approximately 0.4%. Therefore, it is of importance to breed a novel waxy maize line with high lysine content by introgressing the *o2* and *o2m* traits with MAS. Accumulation of starch and storage proteins occurs in the developing endosperm of maize, the quality of which is contributed to by the action of the *Waxy1* (*Wx1*) and *Opaque2* (*O2*) genes. The single copy 3.8-kb *Wx1* contains 14 exons and is mapped on the short arm of chromosome 9. Previous studies have shown that transposable elements Ac/Ds and En/Spm, deletion mutation and mutagenic ethylmethane sulfonate (EMS) mutagenesis account for the pre-mRNA splicing or translation errors and result in a low expression level of *Wx1*. As a result, the granule-bound starch synthase I (GBSS I) activity of the *wx1* mutant has a decreased activity (5–95%) in amylose synthesis, leading to the low level of amylose but high level of amylopectin in maize endosperm and pollen. O2 is a transcriptional factor that contains a basic leucine-zipper (bZIP) motif. It is specifically expressed in the developing endosperm and directly regulates the expression of 22-kDa α-zeins. The substantial reduction of α-zeins is concomitant with increased accumulation of non-zeins, consequently accounting for the increased contents of lysine and tryptophan in maize mutants. In addition, a large number of studies have shown that O2 also has pleiotropic effects on the expression of non-storage proteins including ribosome- inactivating protein b-32 (RIP), cytosolic pyruvate phosphate dikinase 1 (cyPPDK1), lysine ketoglutarate reductase-saccaropine dehydrogenase (LKR-SDH), acetohydroxyacid synthase (AHAS) and Opaque2 heterodimerizing protein 1 (OHP1). Thus O2 as a regulator plays a crucial role in maize endosperm development by influencing the storage protein and nitrogen/carbon metabolism. Although the individual functions of *Wx1* and *O2* are well known, how these genes interact to maintain the starch-protein balance is yet unknown. It has been reported that *o2* mutation can alter the transcriptional patterns of *Wx1* in varying degrees, but no evidence revealed the regulatory effect of *o2* on the expression of *wx1*. By backcrossing of *o2* and *o16* traits with MAS, the quality and lysine content of waxy maize have been successfully improved. However, the molecular mechanism underlying the ameliorated amino acid composition of maize endosperm and specific kernel phenotype is yet unknown. In *o2* mutants, genes associated with glycolytic pathway, endoplasmic reticulum (ER) stress responses and amino acid synthesis demonstrated differences in transcript profiles, thus offering an unbiased hypothesis that some novel mechanisms play a vital role in the modification of waxy maize endosperm. To disclose the extensive changes of endosperm metabolism when introgressed *o2* into waxy maize Zhao OP-6/*O2O2*, we constructed a set of near-isogenic lines (NILs) with QPM as the donor, and performed submicroscopic observations of the endosperm structure and biochemical analysis of the protein bodies and nutrient contents. Further proteomic analysis of immature seeds identified several specific proteins involved in metabolic pathways, such as synthesis of starch and protein, composition of amino acids and carbohydrate metabolism. Combined analysis of the results gives us an intriguing insight into the effect of integrating *o2* and *wx1* on the metabolism of maize developing endosperm. # Materials and Methods ## Plant materials QPM CA339 derived from pool33 (Centro Internacional de Mejoramientode Maizy Trigo, Mexico DF, Mexico) was used as the non-recurrent parent (donor). The elite waxy inbred line Zhao OP-6 (Maize Research Center of Institute of Crop Sciences, Chinese Academy of Agricultural Sciences, Beijing, China) was used as a recurrent parent (receptor). The *o2-wx* NILs, Zhao OP-6/*o2o2*, were selected from the two parents by using MAS technique. Co-dominant SSR markers phi057 and phi027 were used to select heterozygous (*O2WX/o2wx*) individuals, while the dominant SSR marker phi112 was used to detect transposable element *rbg* at BC₆F₁. All heterozygous genotypes were converted to the recurrent parent Zhao OP-6 through six backcrossing cycles, followed by three rounds of self-pollination. Theoretically, the *o2-wx* NILs had 99% of the recurrent parent genome, and were phenotypically uniform and genetically homogeneous after six generations of backcrossing. All of the maize materials were grown in adjacent plots in the experimental station (N40°36′ and E116°34′) of Chinese Academy of Agricultural Sciences during the summer of 2013. A minimum of three well-filled ears of each genotype were sampled at 18 day after pollination (DAP), when the conversion of importing sucrose and amino acids into starch and storage proteins reached a high level. Ears were picked up at approximately noon, and kernels were collected from the centre of each ear. Embryos and surrounding pericarps were dissected, frozen immediately in liquid nitrogen, and stored at -80°C before use. Mature kernels were harvested after physiological maturity and dried in a green house. To avoid biological variations, equal numbers of well-filled ears (≥ 3) were pooled and treated as one sample, and each experiment had two or more replicates. ## DNA extraction, PCR amplification, electrophoresis and genotype analysis Seedling leaves of parents and offsprings were collected and used for DNA extraction with CTAB method. The integrity and quality of DNA were detected by electrophoresis in 1% agarose gel, and the DNA concentration was adjusted to \~100 ng/μL. Primers specific for phi057, phi027, phi112 and 100 SSR markers were adopted from the maize genome database MaizeGDB ([http://www.maizegdb.org](http://www.maizegdb.org/)) and synthesized by AuGCT (Beijing, China). PCR amplification and product analysis were performed as reported previously. To establish a genotype database of each individual, band patterns A, B, H, and U of the 100 SSR markers distributed on ten maize chromosomes were adopted. Pattern A indicated the origin of recurrent parent Zhao OP-6, B for non- recurrent parent CA339, H for heterozygous genotype, and U for unidentified genotype. According to the statistical analysis, the genetic background recovery rate of BC₆F₃ individuals was calculated with the formula G (*g*) = \[L + X (*g*)\] / (2L), in which G indicates the number of backcross generations, X indicates the number of molecular markers with the same pattern as parent Zhao OP-6, and L is the number of polymorphic SSR markers. The formula E \[G(*g*)\] = 1 - (1/2) ^*g*+1 was used to calculate the theoretical genetic background recovery rate, and G referred to the number of backcross generations. ## Abundance analysis of *o2* and *wx* transcripts Developing maize endosperm was collected from the center of three well-filled ears at 18 DAP. Total RNA was individually extracted with TRIZol kit (Thermo Fisher) and pooled together at the same amount. After DNA removal with RNase- free DNase I (Sigma-Aldrich), the cDNA was synthesized using an oligo d(T) primer kit (Promega). Two pairs of specific primers that spanned the exons of *o2* and *wx1* (O2RT-F: 5′-TCAGGAATAATCCAGTGCAGAA-3′ and O2RT-R: 5′-TCGACGTTAGCGTCGTTGTA-3′, and WXRT-F: 5′-TGTAGCTGCTTGCTTGTGCT-3′ and WXRT-R: 5′-CACCGAACAGCAGGGATTAT-3′) and one primer pair specific for the glyceraldehyde-3-posphate dehydrogenase (GAPDH) gene as the reference (GAPDH-F: 5′-CCCTTCATCACCACGGACTAC-3′ and GAPDH-R: 5′-AACCTTCTTGGCACCACCCT-3′) were designed for the analysis of transcript abundance. The PCR amplification products were separated on 1.5% agarose gels for analysis. ## Kernel characteristics and structure observation The characteristics and appearance of intact kernels were collected by a SONY α700 camera (mode A). Endosperm hardness was graded from 1 to 5 following the CIMMYT universal standards, i.e., grade 1 (completely vitreous) to grade 5 (completely opaque), with a 25% difference between grades. The kernel density was calculated by dividing the weight of 50 kernels by the volume (displacement of 95% ethanol in a cylinder). Mature kernels were peeled with a blade at the peripheral region, spray-coated with platinum, and observed under a Hitachi scanning electron microscope (SEM, S3400N). Developing kernels at 21 DAP were prepared as below: kernels (including partial pericarp) (1 mm × 3 mm × 1 mm) were sequentially fixed in 2.5% (w/v) glutaraldehyde and paraformaldehyde, followed by post-fixation in osmium tetraoxide. After dehydration in a gradient of ethanol (75−100%), samples were transferred to a propylene oxide solution and gradually embedded in paraffin. Sections of samples were prepared by a diamond knife microtome and observed under a Hitachi H7600 transmission electron microscope (TEM). Submicroscopic structure analysis was performed at the Institute of Food Science and Technology of Chinese Academy of Agricultural Sciences. ## Biochemical characterization and protein quantification Mature kernels were dried at 65°C to constant weight and pulverized to a fine powder for the analysis of lysine and crude protein contents (at least two replicates and 40 g powder per sample). Crude protein content was measured with the Kjeldahl method according to the Chinese National Standard GB2905-82 (Nitrogen-to-protein conversion factor, K = 6.25). To determine the contents of 17 free amino acids (FAAs), all samples were pretreated following the Chinese National Standard GB7649-87, and analyzed by the S433D full-automatic amino acid analyzer (Beckman Coulter). The content of kernel crude oil was determined by using the Bruker Minispec MQ20 NMR Analyzer. Total starch was extracted, the content of amylose was measured by using the AUTOPOL III Polarimeter (Rudolph), and the percentage of amylopectin was calculated. Zein and non-zein proteins were extracted according to the previous study with some modifications. For each sample, a total of 15 mature kernels were collected from the centre of three well-filled ears (5 kernels from each ear), mixed, and soaked in distilled water for 6 h. Pericarps were then removed without damaging endosperm, and endosperm was further ground into a fine powder in liquid nitrogen. Powdered endosperm (50 mg per sample) was transferred to a 2 mL Eppendorf tube and incubated in 0.4 mL of extraction buffer (70% ethanol, 2% 2-mercaptoethanol, and 1% SDS) at 37°C with agitation for 2\~3 h. After 10-min centrifugation (12,000 rpm) at room temperature (Centrifuge 5424R, Eppendorf), the supernatant comprised the zein fraction, and the sediment consisted of nonzein proteins. Extracted protein were measured using a bicinchoninic acid protein assay kit (Solarbio) according to the instructions. In order to identify the change of zein composition, aliquots of each supernatant (100 μL) were transferred to a fresh tube and dried at 50°C until the liquid evaporated absolutely. Distilled water (100 μL) was then added to dissolve the pellets. The same amounts of samples based on protein concentration were loaded onto a polyacrylamide gel (4% stacking gel and 15% separation gel), followed by staining with Coomassie Brilliant Blue R250 (Bio-Rad). Each sample had two replicates. ## 2D SDS-PAGE Developing kernels at 18 DAP were used for protein extraction and 2-D SDS-PAGE analysis. The pooled endosperm powder (50 mg) was subjected to protein extraction using the acetone precipitation method and dried in a lyophilizer at -20°C. The freeze-dried samples were then resuspended in IPG lysate buffer (0.2 M urea, 5% CHAPS, 50 mM thiocarbamide, 0.7% DTT, and 40% Bio-lyte) and incubated at 4°C for 1 h with frequent shaking. After removal of the insoluble fraction (12,000 rpm, 4°C and 15 min), soluble proteins were quantified by using a 2-D Quant Kit (GE Healthcare) with bovine serum albumin as a standard. IPG rehydration buffer (0.2 M urea, 5% CHAPS, 0.05 M thiocarbamide, 0.7% DTT, 40% Bio-lyte, and 1% bromophenol blue) was then added to adjust the final volume to 350\~400 μL. The first dimension separation was performed using the Ettan IPG- phor-II (GE Healthcare). Aliquots of protein extract (\~600 μg) were separated on 18 cm Immobiline Dry Strips (pH 5–8, Bio-Rad), with three technical replicates per sample. After rehydration for 14 h at 50 V, isoelectric focusing (IEF) was performed at 18°C following the procedures as shown below: 50 mA per strip and 250 V STEP for 1 h, 500 V STEP for 1 h, 2000 V GRAD for 30 min, 2000 V STEP for 1 h, 5000 V GRAD for 30 min, 5000 V STEP for 1 h, 8000 V GRAD for 2 h, 30000 V for focus and finally 500 V STEP for 10 h. The strips were then immersed in 8 mL of two types of equilibration buffer for 15 min with agitation. The immobilized proteins were subjected to second dimension separation on 12% SDS- PAGE gel at 16°C using the Mini-Protein II vertical gel apparatus (Bio-Rad). The conditions were: 1 W for approximately 30 min until the blue line reached the separation gel, and 10 W until the electrophoresis was finished. Proteins were visualized with Coomassie brilliant blue R250. Each sample had three technical replicates. Gels were compared, and spot intensities were quantified using the ImageMaster 2D Platinum 5.0 analysis software (Bio-Rad). Student’s *t*-test was used to determine the spot intensity changes between samples after normalization, and those with more than 1.4-fold changes were defined as significant difference (*p* \< 0.05). ## Identification of protein spots by mass spectrometry Protein spots that consistently appeared in three replicates and showed significant difference between samples were excised manually for mass spectrometry (MS) analysis. Briefly, gel pieces were destained in 25 mM NH₄HCO₃ and 50% (v/v) acetonitrile for 3 to 15 min at room temperature, freeze-dried, and rehydrated in 15 ng/μL sequencing-grade trypsin (Promega) at 4°C for 1 h. NH₄HCO₃ (25 mM) was then added to completely cover the gel pieces. After 16 h incubation at 37°C, the gel pieces were transferred to fresh Eppendorf tubes containing 5% trifluoroacetic acid and incubated at 37°C for another 1 h. Acetonitrile (50%) and trifluoroacetic acid (2.5%) were then added, and the mixtures were incubated at 37°C for 1 h. The freeze-dried peptides were analyzed by a Matrix Assisted Laser Desorption Ionization Time of Flight Mass Spectrometry (ABI-4800, AB Scienx) with the positive ion reflector mode and the sweep range of 900 to 4000 Da. Peaks derived from the mass spectra with a value of S/N \> 10 were searched against the NCBI database by using the Mascot search engine. Only the protein spots with a score of over 70 (*p* \< 0.05) were considered to be a putative protein and identified at Lab Assistant Biotechnology Company (Beijing, China). # Results ## Polymorphism of SSR markers at target loci and MAS of *o2-wx* NILs Three polymorphic markers were tested in this study. As shown in, two markers (phi057 and phi112) specific for the *o2* allele were found to be polymorphic between the waxy maize Zhao OP-6/*O2O2* and the QPM CA339/*o2o2*, and phi027 of the *wx1* allele had polymorphism between the parents. Therefore, these polymorphic markers can be used for the MAS of corresponding target alleles: co- dominant markers phi057 and phi027 for the selection of heterozygous genotype *O2WX/o2wx* and recessive homozygous genotype *o2wx/o2wx*, while phi112 for the genotypes *O2O2* and *O2o2*. Genotype *O2WX/o2wx* was selected from the progeny and backcrossed to the recurrent parent Zhao OP-6/*O2O2*. After six backcrosses, recurrent parents and corresponding heterozygous plants had similar agronomic traits including stature, leaf, seed and flower. Finally, the recessive *o2wx*/*o2wx* plants were selected following two to three generations of self- pollination of *O2WX/o2wx* plants. The electrophoretic patterns of eight *o2-wx* NILs are shown in. The band pattern of Zhao OP-6/*o2o2* was consistent with that of CA339/*o2o2* and Zhao OP-6/*O2O2*, indicating that the *o2* allele has been successfully introgressed into the genetic background of Zhao OP-6/*O2O2*. Transcript abundance analysis indicated that the expression levels of *wx1* in Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2* were identically low, while *O2* showed different expression levels. The transcript abundance of *o2* in Zhao OP-6/*o2o2* was similar to that in CA339. To calculate the recovery rate of original genetic background of Zhao OP-6/*o2o2* individuals, 100 SSR markers were selected for polymorphism analysis. Of them, 54 were found to be polymorphic between parent plants. In the BC₆F₃ generation, the average recovery rate of selected individuals was 91.5%, 7.7% lower than the theoretical value. ## Kernel characteristics and submicroscopic structure Under normal and transmitted light, the kernels of Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2* were completely vitreous indicated that the hardness of the two lines were 1.. In addition, no significant difference was detected in the hundred-kernel weight and kernels density. Under SEM, the starch granules of Zhao OP-6/*o2o2* were compact and embedded in a dense proteinaceous matrix, while those of Zhao OP-6/*O2O2* and CA339 had relative smooth, loosely packed starch granules with little contact with protein bodies. The dense packing of protein bodies around starch grains may account for the more vitreous endosperm. Immature endosperm cells of Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2* at 21 DAP showed similar micro- and ultra-structures, in which protein bodies were regularly shaped, well separated from each other, and evenly surrounded by starch granules. Notably, smaller protein bodies accumulated in Zhao OP-6/*o2o2* endosperm cells at the volume of one third of that of the parents, but at higher amount. Moreover, numerous small and cisternal ERs were dilated in the endosperm cells of recurrent parent Zhao OP-6/*O2O2* and accumulated on the peripheral cell walls with polygonal ring or vesicle-like structures. These structures may originate from the endo-membrane system. ## Changes of zeins and FAA composition in *o2-wx* NILs endosperm To investigate the potential biochemical differences between Zhao OP-6/*o2o2* and the parents, we studied the major characters and components of mature kernels. No significant difference was found in the contents of moisture and oil between Zhao OP-6/*O2O2* and Zhao OP-6/*o2o2*, although these lines showed difference from CA339 in moisture and oil contents and hardness. However, in comparison to parent plants, Zhao OP-6/*o2o2* showed reduced FAAs but increased total starch contents. Moreover, the percentage of amylopectin of Zhao OP-6/*o2o2* was 98.5%, higher than that of Zhao OP-6/*O2O2* (97.7%, *p* \< 0.01). The qualitative and quantitative differences of zein, non-zein proteins and FAAs of different maize lines were also assessed and compared. A shown in, there is no significant difference in the total protein contents of Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2*. However, decreased amounts of zeins and increased amounts of non-zeins were found in Zhao OP-6/*o2o2*. SDS-PAGE indicated that Zhao OP-6/*O2O2* had five major polypeptides in endosperm, i.e. 27-kDa γ-zein, 22-kDa α-zein, 19-kDa α-zein, 16-kDa γ-zein, and 15-kDa β-zein. Four of them (except for the 27-kDa γ-zein) were also found in Zhao OP-6/*o2o2* but with decreased amounts, especially the 22-kDa α-zein and 15-kDa β-zein. The reduced accumulation of 19-kDa α-zein had also been reported in a previous study. These two α-zeins showed considerably reduced accumulation in the donor parent CA339. Notably, in comparison with Zhao OP-6/*O2O2*, the content of 27-kDa γ-zein was significantly increased in Zhao OP-6/*o2o2*, which is similar to CA339. The unchanged level of 16-kDa γ-zein ruled out the possibility that expression of zein gene was generally affected in Zhao OP-6/*o2o2*. Comparison of kernel FAA composition also revealed differences between Zhao OP-6/*o2o2* and the parents. In total, Zhao OP-6/*o2o2* had 16.7% less FAAs than recurrent parent Zhao OP-6/*O2O2*. In the mature kernels of Zhao OP-6/*o2o2*, the contents of lysine and glycine were most significantly increased by 51.6% and 26.9%, respectively, while threonine, cysteine, and methionine derived from the aspartic acid pathway had no change in amounts. Though the two glutamic acid-derived amino acids, histidine and arginine, showed no difference between Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2*, the content of glutamic acid decreased in Zhao OP-6/*o2o2*. Besides, the contents of leucine, serine and alanine were significantly decreased by 47.8%, 24.9% and 38.1%, respectively. And Zhao OP-6/*o2o2* had slightly decreased contents of isoleucine, tyrosine, and phenylalanine. ## Proteomic comparison of parent plants and *o2-wx* NILs 2-D SDS-PAGE was used to compare the proteins in maize endosperms. As results, the molecular weights of maize endosperm proteins ranged from 10 to 130 kDa with the pH gradient from 5 to 8. A total of 40 protein spots displayed significant abundance differences or showed altered accumulations at the protein level. Except for unidentified protein spots (due to weak spectra or unsuccessful database searches), 25 protein spots were identified by MS. A comparison of protein abundance between Zhao OP-6/*o2o2* and parent plants showed that fewer protein species (10 of 25 protein spots) were up-regulated in Zhao OP-6/*o2o2*. Of them, proteins involved in the maintenance and folding of proteins in the ER and defense to biotic and abiotic stresses were enriched, including the 17.5 kDa class II heat shock protein (spot 2), 17.4 kDa class I heat shock protein (spot 3), nucleoside diphosphate kinase 1 (NDPK1, spots 6 and 7) and 17.0 kDa class II heat shock protein (spot 61). Moreover, the 14 kDa zinc-binding protein (ZBP14, spot 8), trypsin/factor XIIA inhibitor (TPA, spot 10) and glucose-1-phosphate adenylyltransferase (ADPase, spots 34 and 51) showed increased abundances in Zhao OP-6/*o2o2*, which were probably related to the improved zinc site binding and starch synthesis. On the other hand, the down-regulated proteins in Zhao OP-6/*o2o2* might be involved in different metabolic pathways. The cyPPDK1-related proteins (spots 25, 27, 29, 39, and 40) were predominant with similar fold changes and accounted for 24% of the differential proteins. The significant reduction in the accumulation of sucrose synthase 1 (SH1, spots 46 and 47), a key enzyme in glycometabolism, indicated the changes of sugar metabolism in Zhao OP-6/*o2o2*. The α-glucan phosphorylase 1 (PHS1, spot 30) and 1,4-α-glucan-branching IIb (SBE IIb, spot 67) that are involved in starch metabolism were both down-regulated in Zhao OP-6/*o2o2*. Although several proteins involved in protein folding and plant defense were enriched in Zhao OP-6/*o2o2*, the expression of Chaperonin 60 (CPN60, spot 68) and RIP (spot 16) was completely suppressed. The suppression of RIP might be associated with the low expression levels of EF2-related proteins (spots 31, 44, and 45). S-adenosyl-L-homocysteine hydrolase (SAHH, spot 62) that takes part in methylation by regulating methyl transferase reactions was also down-regulated in Zhao OP-6/*o2o2*. The differential expression of these proteins in Zhao OP-6/*o2o2* and parent plants may explain the principle biochemical and morphological variations of *o2-wx* NILs. # Discussion In the present study, MAS technique was used to produce a set of *o2-wx* NILs by introgressing *o2* allele into waxy maize line Zhao OP-6/*O2O2*. Recurrent parents and the *o2-wx* NILs (Zhao OP-6/*o2o2*) had similar agronomic traits including stature, leaf, seed and flower. However, some specific fragments of CA339 were also found in Zhao OP-6/*o2o2* by calculating the recovery rate of original genetic background. The result indicated that it is impossible to produce NILs harboring single recessive allele *o2-wx* only based on recurrent phenotypes and MAS using a single marker. Therefore, a larger number of markers are indispensable to create single allele NILs by MAS for foreground and background selections. Endosperm of *o2* mutants is usually soft, fragile, rich in water, and susceptible to pests. However, some endosperm modifier alleles having capacity of changing opaque endosperm into vitreous one can mitigate these weaknesses. In the present study, Zhao OP-6/*o2o2* was found to share similar kernel phenotype to the recurrent parent, such as vitreousness, hardness, density and hundred- kernel weight, suggesting that the endosperm modifier *o2m* allele may play a crucial role in the waxy genetic background. Moreover, the *wx1* allele has been identified to be positively correlated with the vitreousness of maize endosperm and thus may represent another endosperm modifier allele. SEM analysis indicated that starch grains, protein bodies, and viscous cytoplasm in combination formed a more compact matrix in the vitreous region of Zhao OP-6/*o2o2* than the parents. Normally, maize starch granules do not form cohesive contacts with one another. However, there are physical connections between starch granules of Zhao OP-6/*o2o2*, which may account for the hard, vitreous phenotype. Meanwhile, accumulation of protein bodies contributes to a more vitreous and harder endosperm of maize. TEM further revealed the configuration and distribution of protein bodies. In Zhao OP-6/*o2o2* endosperms at 21 DAP, smaller and regularly shaped protein bodies were found to prominently aggregate into clumps. Interestingly, though Zhao OP-6/*o2o2* endosperm cells accumulated smaller protein bodies at one third of the parent volume, the number of the protein bodies was higher than that of the parents. This observation corresponds to the putative role of increased 27-kDa γ-zein in CA339 endosperm. The 27-kDa γ-zein is located on the periphery of protein bodies and is widely cross-linked by disulfide bonds and covalent linkages between protein bodies; these interactions provide a mechanism for cementing protein bodies around starch grains. Moreover, the 27-kDa γ-zein has been demonstrated to play a key role in endosperm modification by RNAi and γ-radiation; its accumulation contributes to a rigid matrix vitreous endosperm as well. Notably, dilated rough ER was observed surrounding the protein bodies in Zhao OP-6 endosperm cells, indicating that ER stress is probably induced in plant cells by a variety of factors, including agents affecting calcium homeostasis, abiotic and biotic stresses, and inhibitors of glycosylation. ER stress usually occurs in *o2* mutants, and leads to a non-homeostasis environment for protein folding and formation of disulfide bonds in ER lumen. However, ER stress was rarely found in Zhao OP-6/*o2o2*. The reason might be that several proteins involved in the maintenance and folding of proteins in the ER are up-regulated due to the alteration of protein body structure in ER. The major pathways regulated by O2 are amino acid biosynthesis and starch- protein balance. Typically, FAAs constitute about 1 to 3% of the non-protein nitrogen in wild-type maize endosperm whereas exhibit increased levels in *o2* mutants. In the present study, the contents of FAAs in Zhao OP-6/*o2o2*, especially leucine, serine and alanine derived from the glycolytic intermediates and most abundant in the early endosperm development, appeared to have a genetic predisposition for down-accumulation. This result is against the previous study, indicating that the *wx1* as an *opaque2* modifier may influence the amino acid composition in maize endosperm. Reduction of these abundant amino acids, especially leucine, is largely responsible for the relative increase of minor amino acids, especially lysine and glycine, as the endosperm matures. The increased level of lysine in Zhao OP-6/*o2o2* largely depended on the reduction of α-zein synthesis which excludes lysine while several non-zein proteins which accounts for most of the higher percentage of lysine are associated with varying degrees of increased accumulation. In addition, O2 has the ability to regulate the activity of cyPPDK1 and AHAS. AHAS catalyzes the first step of branched amino acid synthesis, and cyPPDK1 is a key regulator of the glycolytic pathway derived from pyruvate; both enzymes were also down-regulated by *o2* and finally accounted for the reduction of leucine concentration. Pyramiding of recessive alleles, especially double-recessive or triple-recessive mutations, has specific genetic effects on ameliorating the quantity and quality of starch, sugar, oil, and protein in endosperm. A previous study reported that the double-recessive *o2* and *wx1* mutations had no effect on the grain starch content. However, in the present study, the total starch content of Zhao OP-6/*o2o2* kernels was significantly higher than the recurrent parent Zhao OP-6. Interestingly, the amylopectin content of Zhao OP-6/*o2o2* was also very significantly higher than that in Zhao OP-6/*O2O2*. These results demonstrated that the altered starch structure by completely suppressing the expression level of *wx1* in *o2-wx* NILs may elucidate the role of GBSS I in production of amylopectin. Compared with Zhao OP-6/*O2O2* and soft *o2* genotypes, amylopectin in Zhao OP-6/*o2o2* with reduced intermediate-length α-1,4-linked glucose chains is associated with increased swelling in water and formation of tight contacts between starch granules. In addition, several starch biosynthesis genes, especially ADPase and SBE IIb, showed altered expression levels between Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2* based on 2D SDS-PAGE. ADPase that catalyzes the first committed step of starch synthesis was up-regulated, while SBE IIb that catalyzes the formation of α-1,6-linked glucan and is required for amylopectin synthesis at the surface of the starch granule was down-regulated. Previous studies indicated that the deficiency of SBE Ia has no impact on endosperm starch structure, whereas SBE IIb is closely related to the reduced level of amylopectin. But there is no evidence to verify the predominant function of SBE IIb over SBE Ia, and SBE Ia might have effect on amylopectin structure only in the absence of SBE IIb. In *vitro* assay, SBE Ia preferentially produces longer polymers (\> 16) while SBE IIb produced shorter ones (\< 12). Thus the increased amylopectin level may largely depend on the SBE Ia activity. In addition, expression changes of one or more starch biosynthesis enzymes probably results in the starch modification of Zhao OP-6/*o2o2*, or altering the expression or mutation of one starch biosynthetic enzyme has multiple effects on enzyme activities. As highlighted before, the development of maize endosperm is complex, which is driven by qualitative and quantitative coordinate expression of endosperm protein asset of different genotypes. To better clarify the role of O2 in endosperm protein expression and to investigate their possible interactions in waxy genetic background, 2D SDS-PAGE analysis of 18 DAP endosperms was performed. Of 25 differentially expressed protein spots, those putative cyPPDK1 spots represents an exception that showed similar down-regulation (up to 3-fold or higher). Although cyPPDK1 isoforms had slight changes in p*I*s and molecular weights, they were all controlled by O2 as shown before. A close examination of the expression patterns of proteins involved in sugar and starch metabolism shows that Zhao OP-6/*o2o2* has perturbations in sucrose metabolism and environmental response. Sucrose has been shown to act as signals to trigger changes in the expression of a broad range of post-transcriptional modifications. Therefore, the differential expression of SH1, a key enzyme in sucrose synthesis pathway, indicated that genetic backgrounds of different lines or different environment responses may indirectly affect the expression of SH1 in *o2* mutants. Moreover, the expression levels of several proteins involved in maintenance and folding of proteins in ER showed cross-occurrence between Zhao OP-6/*o2o2* and Zhao OP-6/*O2O2*. For example, the small cytoplasmic chaperones (heat shock protein) were significantly up-regulated in Zhao OP-6/*o2o2*, whereas the expression of Chaperonin 60 was down regulated. These proteins are related to the protein unfolding, and their interaction is likely related to protein body distribution and structure in the ER. Stress response and defense- related proteins including NDPK1 and RIP also showed differential expression in Zhao OP-6/*o2o2*. NDPK1 is a master regulator of diverse pathways in the cell and plays a important role in plant defense responses, while RIP has a defensive role against pathogens and viruses and represents a well-known target of O2 regulation. In Zhao OP-6/*o2o2*, the expression level of NDPK1 was increased by about 2-fold, while RIP was completely suppressed. The up-regulation of NDPK1 is more likely related to pleiotropic responses instead of resistance to pests since evidences have shown that *o2* mutants are much more sensitive to pests. The decrease in RIP abundance may be associated with the low expression of EF2; as a result, the efficiency of translation was suffocated. Regulation of protein expression is mainly controlled by signal transducers, which is central to a myriad of biological processes at the molecular level. In the present study, two proteins, SAHH and ZBP14, were identified to involve in the signal transduction. SAHH is an NAD⁺-dependent tetrameric enzyme which catalyzes the breakdown of *S*-adenosyl homocysteine to homocysteine and adenosine. The decreased accumulation of SAHH in Zhao OP-6/*o2o2* may affect cell growth and regulation of gene expression. ZBP14 as one of the key enzymes in cellular signal transduction has a central role in the control of many cellular processes, and its activity is modified by activators such as diacylglycerol. Although further studies are required to provide convictive proofs that O2 regulates the differential expression of proteins directly or indirectly, the interaction of these proteins is a strong indicator to facilitate the construction of specific phenotype and influence the functions of endosperm cells. # Supporting Information The authors thank Dr. Zhiguo Tian for kindly providing waxy maize inbred line Zhao OP-6. We also thank the reviewers for their hard work on reviewing our manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MSL SHZ. Performed the experiments: ZQZ. Analyzed the data: ZQZ LYS. Contributed reagents/materials/analysis tools: ZQZ LYS XXZ NY RPX HZ XHL JFW ZFH DGZ HJY. Wrote the paper: ZQZ LYS MSL.

# Introduction Intensive endurance training is very demanding for the human organism and its regulatory systems. Among them, cardiovascular control by the autonomic nervous system (ANS) presents alterations varying with the training versus recovery balance. Heart rate variability (HRV) represents sinus node modulation by the sympathetic and parasympathetic branches of the ANS. HRV analysis is validated as a noninvasive method to study individual functional adaptations occurring to a given training stimulus in athletes. HRV analysis is proposed as a valuable tool to study the athlete’s training versus recovery equilibrium and to detect early overreaching state that can decrease the athlete’s performance level. To our knowledge, most studies performed on this topic have focused on male athletes or mixed populations, and very few have studied female athletes. This omission is significant because the neural control of circulation differs with sex, especially before menopause. HRV markers of sympathetic activity response after an orthostatic challenge test have been reported to be higher in male than in female athletes presented with a similar training load. Moreover, HRV analysis in athletes has been focused mainly on the changes observed after acute post-exercise or throughout a training season in order to prevent fatigue and overtraining. HRV parameters changes induced by repeated days of endurance exercise have been scarcely studied, and only in male athletes. Yet, it seems that the daily monitoring of HRV throughout a multi-day sporting event would be of interest for coaches to anticipate fatigue and to guide the athlete for his best final performance. Therefore, the two objectives of this study were to describe a comprehensive characterization of resting heart rate (HR) and HRV changes in well-trained female cyclists during and following a multi-stage cycling event, and to propose an adapted method to follow fatigue and recovery in these athletes throughout this kind of sports event. We hypothesized that cyclists performing a Tour de France would incur a substantial HR increase and HRV decrease at rest along the days which would be worsened throughout the stages. # Materials and methods ## Population Ten healthy and well-trained (regional or national level) female cyclists coming from 4 countries (Belgium, France, Spain, Ukraine) completed the 21 stages of the men’s 2017 Tour de France, but one day before each stage of the official race. The event was performed without competition spirit. The cyclists completed the flat stages as a group, and each performed the time-trial and mountain stages at her own pace. All participants gave informed written consent to participate in this study, which received the approval of the Rennes University Hospital Ethics Committee (Number 2013-A01524-41) and was conducted in accordance with the Declaration of Helsinki. ## Pre-participation medical evaluation All athletes had a medical examination before the event, including a physical exam, a resting electrocardiogram (ECG), a transthoracic echocardiogram, and an incremental maximal cardiopulmonary exercise test on an electronically braked cycle ergometer (Excalibur Sport, Lode, The Netherlands). The exercise protocol started with a warm-up period (100 W for 5 min and 150 W for 1 min) followed by a step load-increase of 25 W/min until exhaustion. None of the athletes ingested any contraceptive and their menstrual cycle was not controlled. ## Description of the cycling event The cycling event consisted of 21 stages and two rest days. Due to the length of the stages, cyclists fed and hydrated continuously on the bike, according their needs. So individual food and fluid intake could not be controlled nor recorded during the stage nor between stages. ## HRV study Baseline HRV values were established from the average of four days before the first stage. HRV was monitored each day during the event and on the first day following completion of the Tour, then once per week during the four weeks post- Tour. Cyclists refrained from any intense efforts during both the pre- and post- Tour periods. ### Recording RR samples The RR interval samples were recorded with a sampling rate of 1000 Hz with a HR monitor (Polar V800, Kempele, Finland). Recordings were performed right after the cyclist woke up in the morning following an overnight fast, in a quiet, semi-darkened room, temperature range of 22–25°C.. The RR samples were collected during two successive 7- minute periods, in supine and standing positions. All athletes were familiarized with the monitor use. ### HRV analysis The RR data recorded was downloaded via Polar FlowSync software for Mac version 2.6.4 (Polar, Kempele, Finland) and exported for analysis with the Kubios HRV Standard software v3.0.0 2 (Biosignal Analysis and Medical Imaging Group at the Department of Applied Physics, Kuopio, Finland). For the analysis, the last 5 minutes window for each position was used. All the ectopic beats were filtered with the artifact correction option of the software. A very low threshold was applied when needed. Both time and frequency domain analyses were performed. The root-mean-square difference of successive normal RR intervals (RMSSD), which reflects HR vagal modulation, was calculated. The high (HF:0.15–0.40 Hz) and low (LF:0.04–0.15Hz) frequency domains were analyzed. The HF band reflects vagal modulation while the LF band indicates both sympathetic and parasympathetic influences. RMSSD, HFnu, LFnu (normal units) absolute values and their difference between supine and standing positions were calculated. The normalized (or normalized unit) spectral indices are defined by the developers of the Kubios HRV Standard software v3.0.0 2 as HFnu = HF / (LF + HF) and LFnu = LF / (LF + HF) (Biosignal Analysis and Medical Imaging Group at the Department of Applied Physics, Kuopio, Finland) in accordance with the recommendations. ## Load of exercise analysis Each individual exercise daily load was calculated using the training impulse score (TRIMPS) method. HR and GPS data were monitored continuously during each stage with the HR monitor. HR was divided into five zones, i.e. 50–60%, 60–70%, 70–80%, 80–90% and 90–100% of the individual maximal HR. Work load quantification was derived from the duration spent within the five HR zones. ## Rate of perceived exertion The individual rate of perceived exertion (RPE) was evaluated with the Borg CR-10 scale within 30 minutes after the end of each stage. ## Statistical analysis Data are presented as mean ± SD. All analyses were performed using SPSS v.21 for Mac and STATISTICA v.7.1 for Windows. Normal Gaussian distribution of the data was verified by the Shapiro-Wilk test. Analysis of variance for repeated measurements was used. A Tukey’s post-hoc test was used to identify where the differences lie. In addition to the day-to-day effects on HR and HRV, the Tour was divided into 3 periods: period 1 (stages 1–9), period 2 (stages 10–15) and period 3 (stages 16–21), with rest days following stages 9 and 15. Pearson’s product moment coefficient was calculated to assess the relationships among RPE, TRIMPS, the mileage of the stages, and the HR and HRV parameters. Significance was set at *P*\<0.05. # Results All cyclists successfully finished the Tour de France. shows the demographic characteristics and the result of the cardiopulmonary exercise test of the subjects. HR, HRV, and workload values are presented for each stage (including resting days) of the cycling event in ; and for each period in. In order to see the evolution of the subjects along the cycling event, presents HR, HRV, and workload values for each stage (including resting days). To have a greater vision of the subjects’ evolution, illustrates HR and HRV parameters by periods. ## Heart rate evolution Supine HR during period 1 increased in comparison with its basal value after stage 2 (flat). This increase persisted until stage 9 with no difference among the stages. During period 2, HR increased after the first medium mountain stage (12^th stage) and then decreased to its basal value. During period 3, HR was higher than its basal value only after the 17^th (high- mountain) and 19^th (flat) stages. For each of the three periods the global supine HR was higher than the basal value. HR was higher during period 1 than during periods 2 and 3, but no difference was observed between periods 2 and 3. The HR value returned to its basal values after each rest day and during the recovery period. The standing HR values presented no difference during the cycling event, except after stage 2. No differences were observed among the three periods. Positive correlations were observed between supine HR and the distance of the stage (r = 0.46, P = 0.037), TRIMPS (r = 0.60, P = 0.004), and RPE (r = 0.46, P = 0.038). The higher the workload, the higher supine HR was after a recovery night. ## HFnu and LFnu evolution Supine HFnu was decreased compared to its basal value from stage 2 until the first rest day, after which it went back to the basal value. Then, the HFnu value decreased from the 19^th stage and returned to its basal value after the 21^st stage. Lastly, the HFnu value was very close to its basal value after one week of recovery. As expected, LFnu mirrored the evolution of HFnu, with an increase of LFnu in compared to basal and recovery values. The changes in daily LFnu were more frequently different than the HFnu ones and overall LFnu was higher than HFnu during the entire cycling event. The opposite was observed only before the event start and during the recovery. No differences were observed between mean supine HFnu and LFnu values by period. The daily standing-supine differences for HFnu and LFnu were lower than their basal values during the entire Tour, except after the second rest day. No differences were observed after one week of recovery. The mean period analysis of standing- supine difference of these parameters showed the same significant differences as the daily values compared to baseline during all three periods, except for the second rest day. The standing HFnu values were lower and LFnu values were higher than supine, with no variation during the cycling event (Figs and). ## RMSSD evolution The supine RMSSD global evolution showed the same trend as HFnu, even though more stages showed RMSSD declines. The same was observed with the periods analysis, with a more marked RMSSD decrease during all the periods in comparison to the basal and recovery periods. No significant variation was noted for the standing RMSSD. The daily RMSSD standing-supine difference was lower than the basal value during the whole cycling period, except after the second rest day and after one week of recovery. There was a significant correlation between supine RMSSD and distance of the stages (r = -0.45, *P* = 0.0001). The mean period analysis of RMSSD standing-supine show a reduction compared to baseline but with no differences among periods. ## Rate of perceived exertion evolution There was a more marked decrease after the first rest day than after the second one. The RPE evolution was positively related to TRIMPS (r = 0.61, *P* = 0.003) and to the distance of the stages (r = 0.91, *P* = 0.0001). # Discussion To our knowledge, this study is the first to investigate the day-to-day resting HR and HRV responses in well-trained female cyclists during a multi-stage cycling event. We observed variable responses of HR and HRV parameters in relation to the distance of the stage throughout the Tour which partially confirms our hypothesis. The difference between standing and supine HR and HRV has a practical application for monitoring fatigue in a female cyclist. ## Evolution of HR and HRV indices during the cycling event As proposed, we studied the HR standing/supine difference and HRV indices responses to the orthostatic stress that reflects the adaptation of the sinus node. Because significant changes concerned mainly resting values with no change for standing values (Figs), the discussion will concern resting HRV indices. Globally we observed an inversion of the ratio LFnu/HFnu with a value higher than 1 throughout the event, except for one value after the second rest day, in comparison to a LFnu/HFnu lower than 1 during the pre-race and post-race periods. During the first period (stages 1–9) of the event, both HR and HRV changes show a biphasic pattern. Indeed, in comparison to the basal value, period 1 showed three parts. No change was observed after stages 1 and 2, then after stages 3 to 5 a statistically significant decrease of HR response was noted. Concerning HRV indices, we observed a significant increase in the LFnu/HFnu ratio explained by a significant increase in sympathetic (LFnu) input associated with a significant decrease in parasympathetic (HFnu) input. Lastly, in stages 6 to 9 there was a less marked decrease of HR associated with a progressive increase in HFnu and a decrease in LFnu (Figs). The decrease of HR standing/supine difference observed during stages 3 to 9 was due to both an increase of resting HR and a decrease of standing HR. Throughout the second and third periods the HR as HRV responses to orthostatic stress were attenuated, with similar changes as seen during stages 6–9. The decrease of HR responses during these two periods was due to a lesser increase in standing HR than during the basal period, with fewer changes in supine HR. It must be noted that despite a very low level of mean TRIMPS during the last two stages of the Tour (20–21), 124 au and 276 au, respectively, we observed a marked decrease of HR response to orthostatic stress due to a low increase in HR and a marked increase in LFnu/HFnu when standing. This discrepancy between TRIMPS and HR response could be interpreted as a fatigue sign. The biphasic response observed during period 1 was not linked to a workload difference (mean TRIMPS 1205 au and 1397 au, respectively, for stages 3–5 and 6–9; NS). The HR and HRV response we observed could be explained by the change in the cardiac preload reported after a few days of intense training, due to delayed hormonal responses. A similar biphasic response observed in left ventricular function has been reported after a four-day simulated multi-stage cycling. The higher workload recorded during the first period compared to the second and third periods could explain the differences observed in HR and HRV responses. The beneficial effect of a rest day for the responses to orthostatic stress appears clearly in Figs and, for one day as well as during the recovery period post-Tour. Concerning the effects of the rest days proposed during the race, to our knowledge only one study investigated the effects on HRV indices of rest days (days 10 and 17) during the Vuelta a España performed by male cyclists. Similar to the present study, no difference was observed between rest days and pre-race HRV indices. Our results show that after one week of post-race recovery, all HR parameters and HRV indices were similar to the pre-event basal values. Supine RMSSD values, another parasympathetic parameter, mimics the globally responses of HFnu (Figs). After the two hardest stages of the event (HM stages 9 and 12, with 181 and 214 km), an acute decrease of supine and standing RMSSD values was observed. To summarize, in comparison to the basal value, the day-to-day HR and HRV responses to an orthostatic active test vary with the duration of the cycling event. The temporary marked alterations of HR responses and of autonomic balance (LFnu/HFnu) observed during the first stages do not seem to be due to a real fatigue state but to acute stress stimuli (subjects not used to these distances). During the next stages both a stable and modest decrease of HR responses and an increase of LFnu/HFnu were observed. We did not observe a pattern of fatigue accumulation with the repetition of stages. However, at the end of the event (two last stages) marked alteration of the parameters studied was again observed. These results are in accordance with the most common ‘fatigue’ pattern described in athletes during intensive training. This is confirmed by the quick and complete recovery after one week. Another observation is that supine HR response correlated with the distance of the stages, TRIMPS, and RPE. ## Practical applications from the study Several studies, reviews, and meta-analyses are in favor of the value of the HRV follow-up in male and female athletes to guide their training to prevent overreaching and overtraining. Our study corroborates that HRV is a useful non- invasive tool that can help to program training and identify fatigue in endurance cyclists. From a practical point of view, the results of this study can help to propose the more useful validated parameters for a daily practice during a multistage cycling event. Globally, given their lack of sensitivity the contribution of the isolated variations of the HR and HRV standing parameters used in this study appear as the less contributive (Figs). Concerning the HR survey, the difference between standing and supine HR after an active orthostatic test showed a good value and the variations of resting supine HR appear as a very simple tool. Concerning the HRV parameters, the use of two indices (RMSSD and HFnu) reflecting the response of the sinus node to the parasympathetic stimuli does not seem useful. The association of the two spectral indices LFnu and HFnu in supine position appears to be the most informative. ## Limitations of the study This study presents some limitations. First, the number of athletes studied was small. This limitation was mainly due to the specific and heavy daily logistic associated with this project that was not an official competition. Second, the cycling event was not a competitive one and our results need to be confirmed in competition because of the specific somatic stress linked to competition. Third, the daily recovery included only one night’s rest without a scientific adapted nutrition or massage session. A more scientific recovery might alter the observed results. # Conclusions The characterization of the HRV response during the entire Tour de France adds new information concerning the fluctuating sympathovagal response of an ultra- endurance event. Despite incomplete recovery, the extent of cardiac suppression with each successive stage, and its physical load, is not summative or augmented. Just one week is enough to restore baseline values. These results suggest that well-trained female cyclists need only one week to recover from an effort such as the Tour de France circuit. The authors gratefully acknowledge the enthusiastic participation of the subjects in this study and the generous technical support of Johan Cassirame and Armel Cretual. We also thank Dave Tanner for his help in English proofreading. [^1]: The authors have declared that no competing interests exist.

# Introduction *Bacillus anthracis* is a Gram-positive, spore forming bacterium that causes anthrax. It is considered a high threat bioterrorism agent because of the relative ease with which its highly resilient spores can be weaponized and dispersed. Renewed interest in developing rapid diagnostic tools and effective medical countermeasures against anthrax was generated subsequent to the anthrax letter attacks in 2001. We sought to determine whether carbohydrates located on the *B*. *anthracis* (Ba*)* vegetative cell have structural or immunochemical properties suitable for the development of new and improved medical counter- measures, such as improved diagnostics, vaccines and targeted immunotherapies. *B*. *anthracis* is a member of the *B*. *cereus* (*Bc*) group of related species that also includes *B*. *thuringiensis*. These species show high DNA sequence similarities, but display an array of pathogen and non-pathogen associated phenotypes. The considerable variability in their infection- associated properties and host specificity has primarily been attributed to differences in the plasmid content. We have previously reported that the structure and composition of HF-released secondary cell wall polysaccharides (HF-SCWPs) from *Ba* vegetative cells are not dependent on plasmid content, are specific for *Ba* and antigenic. For example, antisera raised against live and killed *Ba* spores react with the HF- SCWP from *Ba*, but not with the HF-SCWPs from *Bc* ATCC 10987, a dairy isolate, or from the type strain *Bc* ATCC 14579. In an immunological survey with a polyclonal antiserum raised against *Ba* HF-SCWPs, we demonstrated qualitative serological cross-reactivity with cells and isolated cell walls from *Bc* strains capable of causing severe or fatal illness in humans (e.g. *Bc* G9241 and 03BB87 causing severe pneumonia) but not with closely related *Bc* cells and cell walls from non-infection-associated strains, such as *Bc* ATCC 10987 and *Bc* ATCC 14579. This observation suggested a degree of antigenic structure conservation between SCWPs from infection-associated *Ba* and *Bc* strains. The repeating unit structures for HF-SCWPs from all currently examined Bc group members are summarized in. We investigated the SCWP structures of two recently isolated *Bc* strains, strain *Bc* CI from Côte d’Ivoire and strain *Bc* CA from Cameroon designated as *B*. *cereus* biovar *anthracis*, that caused “anthrax-like” disease in great apes. These specific *B*. *cereus* strains have recently been identified as the causative agents of severe outbreaks of anthrax-like disease among a diversity of great apes in the African tropical rain forest environment (see Hoffmann C, Zimmermann F, *et al*., *Nature* 548, 82–99 (2017). We compared these structures to those from infection-associated and non-infection-associated *Bc* group members. Our findings show that the HF-SCWPs of both great ape strains contained the identical HexNAc trisaccharide backbone core structure and Gal modifications known from *Ba* HF-SCWP. In addition, their SCWP contains the additional Gal- substitution at 50% of ManNAc residues originally identified in human infection- associated *Bc* strains G9241/03BB87/03BB102. Interestingly, both great ape- derived strains contained an additional, unique disaccharide substitution at some of the ManNAc residues that varied in percentage between them. The data from strains *Bc* CI and CA support the hypothesis of a conserved structural core motif of SCWPs in virulent, infection-associated *Bc* group strains. We also describe a comparison of the immunoreactivity of these SCWPs with both *Ba* SCWP polyclonal antiserum (pAb) and monoclonal antibody (mAb) EAII-6G6-2-3 specific for *Ba*. The results corroborate the structural data and further support the existence of an antigenically distinct and structurally conserved motif in the SCWPs of disease-causing members of the *Bc* group of related species. # Methods ## Bacterial strains and clinical isolates The strains and isolates used in this study are listed in. ## Use and care of animals All animal procedures were carried out in strict accordance with recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Specifically, procedures for antisera production in rabbits were approved by the Institutional Animal Care and Use Committee (IACUC) of the Centers for Disease Control and Prevention (protocol \#1420SAIRABC-A1). The rabbits were housed in stainless steel cages. Room temperature was maintained in a range from 20 to 23°C; the light/dark cycle was approximately 12 hours each per d. Water was available to the animals *ad libitum* and environmental enrichment was provided e. g. with treats and toys. Rabbits were ultimately euthanized by exsanguination under anesthesia. There was no direct use of non-human primates *per se* in our current study; all *Bacillus* strains used in this study originally isolated from the great apes were previously described by Klee *et al*.. ## Isolation of cell walls from *Bc* CA and *Bc* CI vegetative cells Cell walls were prepared from cultured cells as previously described following the modified procedure as described. The procedure yielded approximately 15–35 mg of cell walls per liter of cultured cells. ## Isolation and purification of SCWP from *Bc* CA and *Bc* CI The SCWPs examined in this study, and their parent strains, are summarized in and (adapted from and unpublished results). The SCWPs, bound to peptidoglycan via phosphate, were released from the cell walls by treatment with 48% aqueous hydrofluoric acid (HF) and purified as previously described. Briefly: the cell wall preparations from the *Bc* isolates from great apes (strains *Bc* CA and *Bc* CI) were treated with 48% HF at 4°C for 48 h. The reaction mixtures were rotated slowly in an end-over end mixer to assist solubilization of the cell wall material. The reaction mixtures were then adjusted to pH 6.5 by rapidly transferring the mixtures directly into an ammonium hydroxide solution (20% v/v) pre-cooled to -20°C. The solutions were then dialyzed at 4°C vs. deionized water for 3 d using 1,000 MWCO dialysis tubing (Spectrapor). Dialysates were concentrated by centrifugal vacuum-evaporation then subjected to ultracentrifugation (100,000 x *g*; 4°C, 4 h) to remove HF-insoluble material, yielding a supernatant containing the solubilized HF-SCWPs (crude SCWP preparation). The HF-insoluble residue consisted of peptidoglycan fragments and protein (amino acids) as determined by GC-MS analysis of the heptafluorobutyrate esters and the TMS-methyl glycosides. Crude HF-SCWP was dissolved in deionized water, filtered (0.45 μm cellulose acetate) (Mettler-Toledo Rainin LLC, Oakland, CA, USA), then purified by size exclusion chromatography on Superose-12 FPLC column (using “AKTA System”, Amersham) in 50 mM ammonium acetate pH 7.8 (refer to Supporting Information). The eluent was monitored by UV absorbance (210 nm/260 nm/280 nm) and refractive index, and fractions containing the SCWP were identified by glycosyl analysis. The HF-SCWP fractions were dialyzed (MWCO 2,000 Da, Spectra/Por, Spectrum Labs, Rancho Dominguez, CA USA) against water, lyophilized and used for NMR, linkage, and immunological analyses. Typical yields were 5 mg of purified SCWP per 40 mg of cell wall material. ## Glycosyl chemical analyses Glycosyl linkage analysis was performed according to a modification of the method of Ciucanu and Kerek as described previously. The resulting partially methylated alditol acetates (PMAA) were dissolved in dichloromethane and analyzed by GC-MS using a DB-1 or SP-2330 capillary column as described. The carbohydrate compositions of the SCWP and derived fractions were determined by preparing the TMS-methyl glycosides with GC-MS (electron impact) analysis using a 30-m DB-5 capillary column (J&W Scientific, (Agilent Technologies Inc.) Folsom, CA USA). The absolute configuration of glycosyl residues was determined by preparing the diastereomeric TMS-(α)-2-butyl glycoside derivatives, with GC- MS analysis on a DB-1 column with comparisons to authentic D-Gal, D-GlcNAc, and D-ManNAc derivatives. ## Reduction of the reducing-end residue of *Bc* SCWPs Portions of the *Bc* CA and CI HF-SCWPs (3–4 mg) were dissolved in water (200 μl) and mixed with 0.5 mg sodium borodeuteride (NaBD₄) in water (200 μl) and reduced for 1 h at room temp. The reaction was neutralized by adding methanol containing dilute acetic acid, then evaporated several times from methanol in a centrifugal vacuum-evaporator (Speed Vac, Thermo-Fisher Scientific), before dialyzing (1000 MWCO membrane) for 1 d vs. deionized water. The reduced-HF-SCWP (red-HF-SCWP) was rechromatographed on Superose-12, isolated as described above, and subjected to NMR and linkage analyses as for the native materials. ## Nuclear magnetic resonance analyses ¹H spectra and all two-dimensional homo- and heteronuclear spectra of the great ape derived HF-SCWPs were recorded at 25°C on a Varian Inova 600 MHz or Varian 800 MHz spectrometer, each equipped with a 3-mm cryogenic probe, using Varian software (Varian Medical Systems, Palo Alto, CA, USA). The SCWPs released by HF treatment, and purified by Superose-12 chromatography were exchanged with D₂O (99%, 3 repetitions) and dissolved in 200 μl D₂O (“100%”) yielding clear solutions at approximately 7 mg/ml in 3 mm tubes. The ¹H-spectra/axes were referenced to 2,2-dimethyl-2-silapentane-5-sulfonate sodium salt (0.1% DSS in D₂0) at δ_H 0.00 ppm, yielding the HOD resonance at 4.78 ppm. Certain samples, available only in small amounts (e.g., *Ba* CDC684 HF-SCWP, 400 μg) were dissolved in 90 μl of D₂O and analyzed in 3 mm Shigemi tubes (Shigemi, Inc) using the cryogenic probe. The ¹³C axis in 2D analyses was then referenced to the calibrated proton axis, using the Varian software command “setref1(‘c13)”, an accepted method which uses the referenced proton dimension to calculate the 2^nd dimension (¹³C) according to a mathematical relationship. This procedure yields most reproducible and consistent referencing in biomolecular NMR, where the natural abundance of ¹³C and the % of DSS is too low to allow reliable detection of the DSS carbon signal, see *e*.*g*., Wishart, D. S., Bigam, C. G., Yao, J., Abildgaard, F., Dyson, H. J., Oldfield, E., Markley, J. L., and Sykes, B. D., "¹H, ¹³C and ¹⁵N Chemical Shift Referencing in Biomolecular NMR," *J*. *Biomol*. *NMR 6*, 135–140 (1995). ¹H-¹H COSY data were recorded in the absolute value mode with a 6.10 ppm spectral width and a matrix size of 1024 x 4096 complex points with 8 scans per increment. ¹H-¹H zTOCSY was recorded with a mixing time of 150 ms and a matrix of 400 x 4096 at 32 scans per increment. Nuclear Overhauser effect analysis was performed by phase-sensitive ¹H-¹H NOESY experiments, collected with a 150 ms mixing time and matrix size identical to that used for zTOCSY, with 64 scans per increment. Carbon chemical shifts and carbon-proton one-bond correlations were determined with a gradient-selected ¹H-¹³C HSQC collected in the ¹H-detection mode. The acquisition time was 0.2 s, and the matrix size was typically 256 x 2,000 complex points, at 88 to 144 scans/increment with multiplicity distinction for primary/secondary protons. The coupling constant (j1ch) was set to 150 Hz and the carbon spectral width was 110 ppm, transmitter offset 65 ppm. Gradient ¹H-¹³C-H2BC spectra were acquired with a data matrix of 1600 (¹H) x 140 (¹³C) complex points with acquisition times of 0.2 s (¹H) and 22 ms (¹³C) respectively. Phase-sensitive ¹H-¹³C HMBC spectra were acquired with 256 x 2048 complex points at 96 to 144 scans/increment. The carbon and proton sweep widths were identical to HSQC values, and the acquisition time (t2) was 0.27 s. The HMBC transfer delay was set to 50 ms (10 Hz), and 1-bond J-filter was 150 Hz. Anomeric configurations of the glycosyl linkages were assigned from carbon- proton coupling constants (*J*_C1,H1) measured for the native SCWP by ¹H-¹³C HSQC analysis without ¹³C decoupling. The acquisition time was 0.6 s, collecting 128 scans per increment; the j1ch was set to 160 Hz and carbon sweep width was 80–120 ppm. ## Molecular modeling Using the GLYCAM Webtool (39), and the associated Glycam06 force field parameters for carbohydrates we generated several minimal-energy conformers for oligosaccharide structural segments for representative HF-SCWPs. Both wire-frame and space filling renditions of these SCWP segments are shown, as described in Results section. ## Conjugation of HF-SCWPs to protein carriers To facilitate immobilization to polystyrene microtiter immunoassay plates and to make polyclonal antibody, the HF-SCWP was conjugated to either bovine serum albumin (BSA; Sigma, St. Louis, Mo) or keyhole limpet hemocyanin (KLH; Sigma) as described by Leoff *et al*., 2009. ## Antibody preparation Monoclonal antibody (mAb) EAII-6G6-2-3 specific for *Ba* was obtained from T. Abshire at the United States Army Medical Research Institute for Infectious Diseases (Fort Detrick, MD). Polyclonal anti-Ames HF-SCWP-KLH antiserum was generated in rabbits. Briefly, each of two female (2.0–3.5 kg) New Zealand White rabbits (Myrtle’s Rabbitry, Thompson Station, TN) was inoculated intramuscularly at two sites in the dorsal hind quarters. Each rabbit was immunized at 0, 14, 28, and 42 d with a total of 500 μg (primary vaccination) or 250 μg (all other vaccinations) of the *Ba* Ames HF-SCWP-KLH conjugate (1.0 mL total volume using Sigma Adjuvant System (Sigma, St. Louis, MO)). Serum used in this study was collected 14 d after the final immunization. All animal work was conducted with CDC IACUC approvals. ## Immunosorbent assay analyses Monoclonal (mAb) and polyclonal (pAb) immuno-analyses were done using HF-SCWP conjugated to BSA (Sigma, St. Louis, MO) in electrochemiluminescent (ECL) immunosorbent assays (MSD; Meso Scale Discovery, Gaithersburg, MD). Each SCWP- BSA conjugate was adjusted to a concentration of 2 μg/mL total carbohydrate content in 0.01 M PBS pH 7.4. 25 μL of the concentration adjusted SCWP-BSA conjugates and 25 μL of 1 μg/mL of BSA protein (negative control) were incubated at +4°C for 16–18 h. For use, plates were washed, aspirated, and 150 μL of blocker (0.01 M PBS pH 7.4, 0.5% Tween-20, 5% skim milk \[BD, Franklin Lakes, NJ\]) was added to each well (60 min; 20–25°C). Plates were washed three times with wash buffer (0.01 M PBS, pH 7.4, 0.1% Tween-20) on an automated plate washer (BioTek ELx405, Winooski, VT). Monoclonal antibody EAII-6G6-2-3 was evaluated over a range of 93.6 μg/mL to 0.0731 μg/mL in the blocker buffer (60 min; 20–25°C). After washing, bound EAII-6G6-2-3 was detected with MSD sulfo-tagged goat anti-mouse IgM detection antibody at 2.5 μg/mL and incubated at room temperature for 1 h with shaking. The conjugate was detected using MSD Read Buffer and plates were analyzed using a MSD SECTOR imager for emission at 620 nm. Rabbit pAb was detected with 2.5 μg/mL of sulfo-tagged goat anti-rabbit IgM (MSD, Gaithersburg, MD). To assess the reactivity of each HF-SCWP sample to polyclonal anti-Ames SCWP-KLH, antiserum was added in two fold serial dilutions starting at 1/100 for each strain to generate a full dilution curve. Response curves were fitted to a 4 parameter logistic (4-PL) nonlinear regression model for quantitative analysis. The quantitative endpoints for serological analysis were the effective concentration of mAb EAII-6G6-2-3 generating a 50% response (EC₅₀) and the effective dilution of rabbit pAb generating a 50% response (ED₅₀). The EC₅₀ and ED₅₀ values were determined as the reciprocal of the dilution corresponding to the inflection point (‘c’ parameter) of the 4 parameter logistical (4-PL) model to fit a dose- response curve to the data as described. Reportable values were the geometric mean (GM) ED₅₀ and EC₅₀ values from three independent experiments. GM reactivity for each of the individual strains were compared using Student’s t-test with a P = 0.05 level of significance. Low EC₅₀ values and high ED₅₀ values indicated a high level of reactivity against specific HF-SCWP. # Results ## Nuclear magnetic resonance (NMR) spectroscopy of the HF-SCWPs from *Bc* great ape isolates and comparison to human infection-associated *Bc* HF-SCWPs We have previously reported detailed NMR analysis of *Ba* derived SCWPs, and human infection-associated *Bc* derived HF-SCWPs. For comparison, the NMR parameters from one of the previously examined human infection-associated HF- SCWPs (from strain *Bc* G9241) are reported in, along with those from the great ape isolate *B*. *cereus* Cameroon (*Bc* CA). Our initial ¹H-NMR analysis of the HF-SCWPs isolated from the *Bc* CA and *Bc* CI strains indicated that their structures were probably similar to those of the human infection- associated *Bc* G9241 and *Bc* 03BB87 strains, although with some interesting and significant differences in their ¹H spectra. A number of points were noted in these proton spectra: (i) HF-SCWPs from both great ape isolates appeared to contain all of the NMR anomeric signals as found in the *Bc* G9241 HF-SCWP, however the ratios of these resonances were altered in both great ape isolates, compared to the *Bc* G9241; (ii) in addition to these “identical” components, both of the great ape isolates displayed two additional “new” signals in the anomeric region, not previously observed in the *Bc* G9241 and *Bc* 03BB87 human infection-associated strains. These new components were designated residues **J** and **K**, and their NMR parameters are listed in, along with the NMR parameters of all other repeating unit components. ## Comparison HF-SCWP from *Ba* strains and infection-associated *Bc* strains All of the *Ba* strains examined to date, including *Ba* Ames, *Ba* Pasteur, *Ba* Sterne 34F2, *Ba* 7702, and *Ba* CDC684, as well as examined strains of human infection-associated *Bc* G9241 and *Bc* 03BB87 share an identical SCWP glycosyl backbone which consists of a linear HexNAc trisaccharide repeating unit: →4)-β-D-Man*p*NAc-(1→4)-β-D-Glc*p*NAc-(1→6)-α-D-Glc*p*NAc-(1→, designated here as: →4)**B**β(1→4)**C**β(1→6)**A**α(1→ to conform with previous nomenclature (26). With the exception of *Ba* strain CDC684, these polysaccharide backbones are highly galactosylated with α- and β-Gal residues, and the galactosylation pattern (location, linkage, anomeric configuration) is species/strain specific. HF-SCWPs from all examined *Ba* strains share the same overall galactosylation pattern yielding the hexasaccharide repeating unit shown in. Thus, the ratio of Gal substitution to HexNAc backbone residues is 1:1, which represents a high density of side chain attachment for a polysaccharide. In these *Ba* derived HF-SCWPs, a small percentage of the repeating units typically lack one or more of the Gal substituents, contributing to signal heterogeneity in their NMR spectra. This heterogeneity in Gal side residue attachment is observed in all the examined *Ba* and *Bc* derived HF-SCWPs. In the human infection-associated *Bc* derived HF-SCWPs, a single additional Gal substitution is introduced at 50% of the backbone ManNAc residues, yielding the repeating structure and spectrum shown in. These *Bc* derived HF-SCWPs essentially consist of two types of repeating units, one *with*, and one *without* the Gal-substituted ManNAc. The introduction of this additional Gal residue modifies the magnetic environment of adjacent residues, such that their anomeric signals split, yielding twin signals, where one member of each pair arises from repeats which *lack* the Gal-ManNAc, and the other member signal arises from repeats which *carry* the Gal-substituted ManNAc. Other ring positions are affected to lesser extents. The additional Gal attachment on these *Bc* derived HF-SCWPs results in polysaccharides having a side residue/backbone ratio of \> 1:1. An extensive NMR comparison of the *Ba*-derived and human infection-associated *Bc*-derived HF-SCWPs was previously described and representative HSQC spectra are included here for reference (Supporting Information). In our present study, the NMR anomeric region of the great ape derived *Bc* strain CA HF-SCWP is compared with *Ba* and human infection-associated *Bc* derived HF-SCWPs in. The “new” signals are indicated, and their associated spin systems and residue identification are described in the following sections. Signals originally appearing in the *Ba* samples are again split into “twin signals”, but the splitting ratio is not 50/50, reflecting the fact that the Gal-substitution on ManNAc is not 50% in the great ape-derived SCWPs. This feature, a lower percentage Gal-substitution at the backbone ManNAc residue, is observed to an even greater degree with the great ape *Bc* strain CI HF-SCWP. The ¹H-anomeric regions of *Bc* CI, *Bc* CA, the human infection- associated strain *Bc* G9241, and *Ba* Sterne 34F2 are all compared in, demonstrating their relationship and differences. The ¹H-¹H TOCSY spectra of the great ape *Bc* CA HF-SCWP and the previously analyzed human infection-associated *Bc* G9241 HF-SCWP are compared in, demonstrating the similarity of the conserved residues, and the presence of new glycosyl systems in the great ape isolate. The ¹H-¹H TOCSY spectra of SCWPs from the two great ape isolates *Bc* CA and *Bc* CI are compared in demonstrating their overall similarity. ## Identification of new Gal substituents on SCWPs of great ape strains *Bc* CA and *Bc* CI In, we compare the regions of the ¹H-NMR spectra where the new, additional signals (**J** and **K**) are observed in the great ape derived HF- SCWPs compared to *Bc* G9241. This region also reveals the percentage of ManNAc residues (**B′**) that are substituted with an α-Gal residue (**G**), originally identified in the human infection-associated *Bc* G9241 HF-SCWP. In, the HF-SCWP from *Bc* G9241 demonstrates the near 50% substitution of the ManNAc residue (**B** +**B′**) with this Gal(α1→3) substituent (**G**). In comparison, the HF- SCWPs from both *Bc* strains CA and CI appeared to show a lowered ratio of ManNAc substitution by residue **G**, as calculated by integration of the anomeric signals for residues **G**, **J**, and (**B**+**B′**+**B′′**). Thus, area integration suggested that substitution of ManNAc by **G** occurs at a lower percentage in the great ape derived SCWPs, compared to the *Bc* G9241. We also initially found that one of the new anomeric signals (assigned residue **K**, H1 δ_H 5.13), actually had the same chemical shift as a low- abundance resonance previously observed in all *Ba* and *Bc* HF-SCWPs, and identified as the free reducing end (GlcNAc, α-anomer) of the polysaccharides. The presence of this free α-reducing end thus interfered with area integration and spectral characterization of the new residue **K** in the great ape isolates. To remove this interference, the *Bc* CA and *Bc* CI HF-SCWPs were reduced at the reducing end with sodium borodeuteride, yielding the reduced-HF- SCWPs (red-HF-SCWP). In, the effect of reduction on the α-reducing end signal is indicated by a decrease in the relative signal area at **K+α**, in the *Bc* CA red-HF-SCWP. Although reduction was incomplete, the area of **K+α** was decreased substantially to the point where it was roughly equivalent to the area of residue **J**, due to removal of mostthe α-reducing end contribution. To further clarify the effect of reduction on the SCWPs, the ¹H-¹³C-HSQC spectra of the anomeric regions, before and after reduction, were examined. These HSQC spectra demonstrate that the NaBD₄ treatment was effective in removing both the α- and β- reducing end anomeric signals and it did not appear to alter the repeating anomeric signals of the HF-SCWPs in any manner. A comparison of the TOCSY spectra of this region before and after reduction is also provided. Having removed interference by the α-reducing end, the identification and location of the new residues **K** and **J** in the red-HF-SCWP samples were investigated. The *J*_C1,H1 coupling constants for residues **J** and **K** were 173.0 Hz and 172.0 Hz, indicating that both residues were probably of α-anomeric form. ¹H-¹H-COSY and ¹H-¹³C HSQC analysis identified H2 for each residue, indicating that neither was a 2-amino sugar. ¹H-¹H COSY, and TOCSY analysis revealed weak magnetic transfer beyond position 4 for both residues, consistent with the *galacto*-configuration. New, hydroxy-methylene protons, not observed in the *Bc* G9241 spectrum, were identified in the HSQC spectra of *Bc* CA (and *Bc* CI) for residue **J** at δ 3.63/3.54/63.18 (H6/C6,), and TOCSY analysis demonstrated connectivity to **J** H5/C5 (δ_H/δ_C 3.73/72.80, another new signal not observed in the *Bc* G9241 HSQC), and subsequently to H4/C4 (δ_H/δ_C 4.18/65.84) of residue **J**. Additional HSQC and TOCSY spectra comparing relevant regions from the great ape *Bc* CA and human infection-associated *Bc* G9241 HF-SCWPs are shown in and **Figs**. The **K** system was defined with the assistance of HSQC and H2BC one and two- bond couplings for assignment of positions 6 and 5. H2BC revealed methylene protons (δ_H 3.73/3.73) which coupled to C6 (δ_C 61.6) and C5 (δ_C 71.52), with assignment of H5 (δ_H 4.09) from a new HSQC signal, not present in the Bc G9241 spectrum. Partial TOCSY spectra defining residues **J** and **K** are shown in. The *galacto*-configuration was assigned to both **J** and **K** on the basis of weak scalar coupling at H4, the relative downfield shift at H4, and intra-residue NOEs between H3-H4 and H5-H4 in addition to H3-H5. As described below, GC-MS analysis of the TMS-methyl glycosides and partially methylated alditol acetates detected elevated levels of Gal in the great ape HF-SCWPs, compared to the *Bc* G9241, and failed to detect any other “new” hexoses. ## Location of additional α-Gal substituents (residues J and K) on the great ape isolates *Bc* CA and *Bc* CI HF-SCWPs In the SCWPs from human infection-associated strains *Bc* G9241 and *Bc* 03BB87, the attachment of the α-Gal (here assigned residue **G**) to the 3-position of 50% of the ManNAc residues was clearly demonstrated by inter-residue NOEs and trans-glycosidic HMBC connectivities. This substitution of ManNAc at the 3-position results in a unique signature at ManNAc **B′** H3/C3 (δ_H/δ_C 4.24/80.28) which differs from that of the unsubstituted ManNAc residue **B**, and is unique in the HSQC spectrum. For both great ape derived HF-SCWPs, a certain percentage of the backbone ManNAc residues were found to be substituted at O3 by an α-Gal residue (**G**) (compare residues **B** and **B′**, the latter having H3/C3 δ_H/δ_C 4.24/80.09). Inspection of the ¹H-¹H NOESY spectrum for *Bc* CA red-HF-SCWP revealed an inter-residue NOE between **G**1 (δ_H 5.05) and **B**′3 (δ_H 4.24), indicating that Gal is linked α1→3 to ManNAc in these great ape isolates as well. Most important, further inspection of the NOESY spectra indicated that the new residue **J** anomeric proton (δ_H 5.13) showed only two strong NOEs, to a proton at δ_H 3.91 (identified as H2 from COSY) and an NOE to a new proton at δ_H 4.27, subsequently identified as H3 of a new variant of the ManNAc residue (assigned **B′′**). These results indicated that the α-Gal residue **J** is also linked to the 3-position of a percentage of ManNAc residues, however, the residue **J** spin system differed from that of residue **G**. Inspection of the **J** spin system defined above suggested that it may itself be substituted at H3/C3 (δ_H/δ_C 4.18/77.20) by an unknown residue. This substituent was readily identified as the new α-Gal residue **K**, having a strong inter-residue NOE between **K**1 (δ_H 5.13) and **J**3 (δ_H 4.18), **,** and NOEs to protons **J**2 (δ_H 3.91) and (intra-residue) to **K**2 (δ_H 3.85). The resulting glycosidic sequence, **K**α1→3**J**α1→3**B′′**, was substantiated by a trans-glycosidic ¹H-¹³C HMBC connectivity between **J**H1 (δ_H 5.11) and **B′′**C3 (δ_C 80.09) (shown). The linkage **G**α(1→3)**B′** was also confirmed by a trans-glycosidic HMBC between **G**H1 (δ_H 5.05) and **B′**C3 (δ_C 80.09), as previously observed for the human infection-associated *Bc* strains G9241 and 03BB87. Additional confirmation for intra-residue assignments defining the new residues in the *Bc* CA red-HF-SCWP were confirmed by 2-bond ¹H-¹³C H2BC connectivities observed for **J**H1/**J**C2 (δ_H 5.11/δ_C 67.03); **K**H1/**K**C2 (δ_H 5.13/δ_C 68.90), and **K**H2/**K**C1 (δ_H 3.85/δ_C 95.33) *inter alia*. shown in. To further clarify the nature of the new residues and their linkages, chemical linkage (methylation) analysis was performed and the resulting PMAA derivatives were compared for great ape derived HF-SCWPs and the previously analyzed human infection-associated *Bc* G9241 HF-SCWP. Only the great ape derived HF-SCWP yielded low levels of a 3-linked Gal PMAA derivative, not previously observed with any of the human pathogen derived strains. Together these results indicated that HF-SCWPs from great ape derived CA and CI strains exhibit a nearly identical structure as that from *Bc* G9241, with one interesting and rather unusual difference. All contained a portion of their backbone ManNAc residues substituted with α-Gal at the 3-position (a feature that is lacking in the *Ba* derived HF-SCWPs). However, the great ape isolates exhibited a further modification, in which a percentage of the ManNAc residues carried a Galα1→3Galα1→3 disaccharide substituent. summarizes the relative percentage of these substituents, estimated from the NMR signal areas of anomeric resonances, for these human infection-associated and great ape-isolated *Bc* strains. ## SCWP reactivity with polyclonal anti-Ames SCWP-KLH antibody For comparative analysis, strains were grouped based on their structural similarity in the aminoglycosyl backbone and its galactosyl modifications. The HF-SCWPs from all examined isolates reacted with the polyclonal (pAb) anti-Ames HF-SCWP antiserum. The highest reactivity was observed with the HF-SCWPs from strains *Ba* Sterne 34F2 and *Ba* 7702 (**,**). There were no significant differences in the reactivity of the pAb serum with HF-SCWPs from *Ba* Sterne 34F2, *Ba* 7702 and infection-associated *Bc* 03BB87 and *Bc* G9241 strains (p = 0.17). There was, however, an approximately 35% reduction in pAb binding to the HF-SCWPs from the great ape disease causing strains *Bc* CI and *Bc* CA compared to the infection-associated *Ba* and *Bc* strains (p \<0.05). There was an approximately 70% reduction in pAb binding to the HF-SCWP from the avirulent *Ba* CDC684 strain and the non-infection-associated strain *Bc* ATCC 10987 as well as an approximately 50% reduction of pAb binding to the HF-SCWP from strain *Bc* ATCC 14579 when compared to the infection-associated *Ba* and *Bc* strains (p \<0.05). The ED₅₀ values obtained with HF-SCWPs from the two non- infection-associated strains *Bc* ATCC 14579 and 10987 were statistically different from each other (p \<0.05). However, while there was a significant difference between the HF-SCWP reactivities from the non-infection- associated type strain *Bc* ATCC 14579 and *Ba* CDC684 (p \<0.05), there was no significant reactivity difference between dairy isolate *Bc ATCC* 10987 and the highly attenuated, Gal-substitution negative *Ba* CDC684 strain (p = 0.24). The HF-SCWP pAb reactivities separated into distinct groups. HF-SCWPs from the strains *Ba* Sterne 34F2 and *Ba* 7702 (Structural Group 1 ED₅₀ values: 1746 and 1786, respectively; p = 0.4) had similar immuno-reactivities to the human infection-associated strains *Bc* G9241 and *Bc* 03BB87 (Structural Group 2 ED₅₀ values: 1844 and 1878, respectively; p = 0.42). The immuno-reactivities of these two groups were not significantly different from each other (p = 0.17,). HF-SCWPs from the *Bc* CI (ED₅₀ value: 1143) and *Bc* CA (ED₅₀ value = 1212) were not statistically different from each other (p = 0.32). The lowest reactivities were observed with the HF-SCWP from *Ba* CDC684 (ED₅₀ value: 517), and from the non-infection- associated strains *Bc* ATCC 10987 and *Bc* ATCC 14579 (ED₅₀ values: 547 and 842, respectively). ## HF-SCWP reactivity with monoclonal antibody EAII-6G6 There was a hierarchy of mAb reactivity (EC₅₀ values) with the examined HF-SCWPs. The HF-SCWPs from *Ba* Sterne 34F2 and *Ba* 7702 strains had the highest reactivities (EC₅₀ values of 0.56 and 0.56 μg/mL, respectively), and were not statistically different from each other (p = 0.48). This reactivity was significantly lower (p \< 0.05) for all *Bc* strains and isolates evaluated (e.g. the dairy-isolate strain *Bc* ATCC 10987 had an EC₅₀ from 141.89 μg/mL and the *Bc* ATCC 14579 type strain of 223.45 μg/mL). However, all infection-associated *Bc* strains displayed distinctly higher reactivity. For example, the human infection-associated strains *Bc* G9241 and *Bc* 03BB87, Structural Group 2, had approximately 10- to 30-fold lower EC₅₀ values and the great ape isolates, Structural Group 3, approximately 2- to 3-fold lower EC₅₀ values compared to the non- infection-associated *Bc* strains (p \< 0.05). The variance in the EC₅₀ values between Structural Group 2 (*Bc* G9241 and *Bc* 03BB87, p = 0.13) and between Structural Group 3 (*Bc* CA and *Bc* CI strains, p = 0.32) were not statistically significant. The HF-SCWP from the highly attenuated *Ba* CDC684 was of particular interest. This SCWP comprised the core HexNAc trisaccharide backbone but was devoid of α- and β-Gal substitutions. This HF- SCWP displayed a significantly higher EC₅₀ (lower reactivity) compared to HF-SCWP from *Ba* Sterne 34F2 and *Ba* 7707 (p \< 0.05), and from the infection-associated *Bc* G9241, *Bc* 03BB87 and *Bc* CA and CI (p \< 0.05). The reactivity of the *Ba* CDC684 isolate was not significantly different to that of the non-infection-associated and structurally unrelated HF-SCWP from *Bc* ATCC 14579 and ATCC 10987 (p = 0.18 and 0.21, respectively;). Essentially, the HF-SCWP from *Ba* CDC684, the great ape *Bc* CA and *Bc* CI isolates, and the *Bc* ATCC 14579 and ATCC 10987 environmental isolates may be considered non- reactive with the mAb although the two great ape isolates did show a slightly elevated reactivity compared to the environmental isolates. Molecular models, generated using the Glycam Webtool with the corresponding force field parameters for carbohydrates yielded several minimal-energy conformers allowing us to compare the epitope presentation for representative SCWP segments from a strong- mAb binder (*Ba* Ames/Sterne), an intermediate mAb binder (human-isolate *Bc* G9241), and a non-binder (*Bc* ATCC 10987). Several minimal energy conformers were generated including *gt-gt* and *gg-gg* forms; only the *gt-gt* conformers are shown. The possible role of these conformers, including the relative importance of individual Gal residues in EAII-6G6 recognition is discussed in and in the Discussion. Comparisons of HF-SCWP structures indicate that mAb reactivity can distinguish between the polysaccharide backbones of *Ba* and *Bc* based on the galactosylation pattern of the →4-β-ManNAc-(1→4)-β-GlcNAc-(1→6)-α-GlcNAc-1→ backbone. The absence of Gal residues such as in the *Ba* CDC684 SCWP results in a dramatic loss of binding activity with an EC₅₀ value of 356 μg/mL compared to a strain with full galactosylation such as *Ba* Sterne 34F2 with EC₅₀ 0.56 μg/mL. The addition of one α-Gal to O3 of the ManNAc residue in *Bc* G9241 SCWP resulted in a statistically significant reduction in binding (EC₅₀ 13.6 μg/mL). The addition of a second α-Gal, i.e., as the disaccharide substituent \[Gal(α1→3)Gal(α1→\], to O3 of the ManNAc residue (refer to **Figs and ,** or) in the SCWP of the *Bc* CI and *Bc* CA isolates resulted in further reductions in binding (EC₅₀ 49.94 and 64.34 μg/mL, respectively) of the monoclonal antibody when compared to a monosaccharide α-Gal substitution. # Discussion Structural investigations of SCWPs from *Ba* Ames, *Ba* Sterne 34F2, *Ba* 7702, *Ba* Pasteur and *Ba* CDC684, and from infection-associated *Bc* G9241 and *Bc* 03BB87 showed that their SCWPs were comprised of a linear aminoglycosyl (Hex*p*NAc) trisaccharide backbone repeating unit that contains the common structural core motif: →4)-β-D-Man*p*NAc-(1→4)-β-D-Glc*p*NAc-(1→6)-α-D- Glc*p*NAc-(1→. With the exception of the highly attenuated *Ba* CDC684, the GlcNAc residues are highly galactosylated with α- and β-Gal. The SCWPs from infection-associated *Bc* strains had an additional α-Gal substitution on the ManNAc backbone residue. Since this aminoglycosyl backbone motif is not present in the non-infection-associated strains *Bc* ATCC 14579 and ATCC 10987, we hypothesized that SCWP molecular structural features (e.g., secondary conformational arrangement and optimal presentation of the Gal substituents) could have an important functional role in pathogen development, including such aspects as wall assembly and cell division/cell chain length. Note that although the HF-SCWP are antigenic, they are not toxic factors and are not considered to directly contribute to *Ba*/*Bc* virulence, however, SCWP may have an indirect role as a result of their known ability to anchor S-layer proteins and direct localization of S-layer associated proteins (e.g. BSL O) to cell septa. Previous studies by Ezzell *et al*. with mAb EAII-6G6-2-3 and *Ba* derived cell wall fractions showed that its reactivity was inhibited by free monosaccharide galactose and lactose (i.e. Galβ1→4Glc), but not by GlcNAc. This observation suggested to us that the epitope for this antibody may be localized in the SCWP, with particular emphasis on the presentation and arrangement of Gal residues carried on the HexNAc backbone. Modifications in SCWP structures of the *Bc* group strains may therefore be correlated to EAII-6G6-2-3 binding. The recently isolated and phylogenetically well-defined *Bc* CA *and Bc* CI causing anthrax- like symptoms in great apes were evaluated to further increase the scope of SCWP structural analysis in infection-associated *Bc*. The results presented here reveal that the SCWPs isolated from great ape infection- associated *Bc* CA and *Bc* CI isolates had SCWP primary structures almost identical to the SCWP from human infection-associated strain *Bc* G9241 except that a percentage of the backbone ManNAc residues also carry a Galα1→3Galα1→3 disaccharide substituent. Apart from this disaccharide, the SCWP structures from the *Bc* CA and CI strains contain the structural features previously identified in infection-associated *Ba* and *Bc* strains, i.e. the conserved aminoglycosyl trisaccharide backbone that is substituted with strain- specific galactosylation modifications at conserved locations. The structural modifications discovered in the SCWPs from the great ape infection- associated *Bc* CA and CI strains compared to the SCWP from *Ba* raised the question whether these modifications affect SCWP immuno-reactivity compared to the those of SCWPs from other infection-associated and non- infection-associated *Bc* group strains. The data from this study indicate that the pAb antiserum did not have sufficient specificity to distinguish between the HF-SCWPs from the *Ba* Ames and human infection-associated isolates *Bc* G9241 and *Bc* 03BB87. The HF-SCWPs from both *Bc* CA and *Bc* CI had similar reactivity, with approximately a 35% reduction in binding of the pAb relative to HF-SCWPs from the *Ba* Ames and human infection- associated isolates *Bc* G9241 and *Bc* 03BB87. There was a 50% reduction in pAb reactivity to the non- infection-associated strain *Bc* ATCC 10987, and an approximately 80% reduction in reactivity to HF-SCWP from the non-infection-associated strain *Bc* ATCC 14579 and the nongalactosylated aminoglycosyl repeating motif from *Ba* CDC684. These data suggest that the conserved aminoglycosyl repeating unit backbone, common only to the *Ba* strains and the pathogenic *Bc* isolates, is a predominant antigenic determinant for the pAb, and that their strain specific, α- and β-Gal substitutions at GlcNAc and at ManNAc, confer additional, sometimes unpredictable, polyclonal response. This is emphasized by the fact that a significant response, well above base line, was observed between the polycolonal sera and the HF-SCWP from *Ba* CDC684 (which lacks all Gal substitutions and consists only of the conserved *Ba* aminoglycosyl backbone. The polyclonal serum was also somewhat reactive with HF-SCWPs having a structurally different HexNAc backbone, such as those from the nonpathogenic strains *Bc* ATCC 10987 and the type strain *Bc* 14579. These non-infection- associated environmental isolates have HexNAc backbone structures that differ from those shared by the pathogenic *Ba* and *Bc* isolates, specifically, \[→4)-β-D-Man*p*NAc-(1→4)-β-D- Glc*p*NAc-(1→6)-α-D-Gal*p*NAc-(1→\], and both nonpathogenic isolates displayed reduced pAb immunoreactivity. Thus the pAb was able to distinguish between the HF-SCWPs from non-infective isolates and the clinical isolates, although the distinction was not impressive and the pAb response appears to be rather unpredictable, using any straightforward molecular model. It is not clear, for example, why the HF-SCWP from the *Bc* ATCC 10987 and the HF-SCWP from *Ba* CDC684, having different backbone structures, showed very similar immunoreactivity with this polyclonal antiserum. The similar immunoreactivity may be a consequence of both SCWPs containing a conserved →4)-β-D- Man*p*NAc-(1→4)-α-D-Glc*p*NAc-(1→ structural element. Likely, it typifies the broad, non-specific nature of polyclonal sera compared to a mAb, and further investigation is required to define the exact epitope(s) involved in pAb binding. In contrast to the polyclonal reactivity, the mAb EAII-6G6 had significantly greater ability to discriminate between subtleties in SCWP structur~~e~~. There was approximately a 100-fold reduction in binding of the mAb to HF-SCWPs from both *Bc* CA and *Bc* CI, relative to HF-SCWPs from the *Ba* Sterne 34F2 or *Ba* 7702, and a 20-fold reduction in reactivity compared to human infection- associated isolates *Bc* G9241 and *Bc* 03BB87. Small differences in mAb reactivity between isolates *Bc* G9241 and *Bc* 03BB87 were also detected, however, these differences could simply result from experimental variability, for instance, differences in purity of the two preparations. All of the 2-D NMR signals arising from each of these two *Bc* human infection-associated SCWPs were essentially indistinguishable, suggesting they are of identical structure. However low levels of insoluble contaminants (e.g. peptidoglycan fragments), and low levels of a contaminating glycogen polymer vary slightly among different preparations, and their presence could likely account for the slight displacement of these two binding curves. Reactivity of the mAb to the non- galactosylated aminoglycosyl repeating motif from *Ba* CDC684 was reduced more than 600-fold compared to *Ba* Sterne 34F2 or *Ba* 7702. Reactivity to the differing HF-SCWP structures from *Bc* ATCC 10987 and *Bc* ATCC 14579 was reduced by 250- and 400-fold, respectively. Thus, the observed reactivity of EAII-6G6 with the examined HF-SCWPs suggests that four main structural groups can be distinguished with this mAb: 1) non-reactive (*Ba* CDC684 and the non- pathogenic *Bc* ATCC 10987 and *Bc* ATCC 14579, 2) low reactive (both great ape isolates (*Bc* CA and *Bc* CI), 3) intermediate reactivity (human infection- associated *Bc* G9241 and *Bc* 03BB87), and 4) highly reactive (*Ba* Ames, Pasteur, Sterne, and related examined *Ba* strains). The structural basis of this reactivity is suggested to reside in the presentation of multiple Gal residues in the proper 3-dimensional conformation. This 3-dimensional arrangement (secondary structure) is optimal only in the *Ba* derived strains, (with the exception of *Ba* CDC684 which lacks galactose). The introduction of an additional Gal residue, at O-3 of the backbone ManNAc residue (i.e., for *Bc* G9241 and *Bc* 03BB87), disrupts this secondary structure of the polysaccharide sufficiently, such that EAII-6G6 binding is reduced. The attachment of *further* additional Gal, (i.e., in the form of Galα1→3Galα1→ disaccharide to ManNAc in both great ape isolates) causes further deviation from the optimal secondary structure necessary for mAb recognition, yielding the weak binding observed with this structural group. One possibility is that approach of the mAb to the multiple Gal epitope presented by the SCWP is simply hindered by the presence of the new disaccharide (and monosaccharide) Gal substituents, at ManNAc in these strains, as visualized by molecular models. Alternatively, the presence of these new substituents could also result in a change in the overall secondary conformation presented by the SCWP, compared to that of *Ba* Ames, Pasteur, Sterne, and related *Ba* strains. A combination of these factors may also account for the observed binding differences. It is interesting to note that HF- SCWPs derived from the nonpathogenic strains (e.g., *Bc* ATCC 10987 and *Bc* ATCC 14579), which do not contain the conserved HexNAc trisaccharide backbone of the infection-associated strains, elicit binding essentially at baseline levels, similar to BSA. The *Ba* CDC684 HF-SCWP, which *does* contain the pathogenicity- associated HexNAc backbone but lacks all Gal substituents, also shows this same lack of reactivity with the EAII-6G6 mAb. Thus, both the pathogenicity- associated HexNAc trisaccharide backbone, together with the optimal pattern of α/β- attached Gal residues (found only in *Ba* strains), is required to yield the maximum binding of EAII-6G6. Molecular models generated using the Glycam web tool with standard Glycam06 force field parameters yielded several low energy conformers for SCWPs from three representative strains, *Ba* (Ames), *Bc* G9241, and *Bc* ATCC 10987 (strong, intermediate, and non-binding) as compared in. These models reveal how additional Gal substituents located at ManNAc could perturb mAb binding and accessibility to the correct Gal epitope. For instance, shows identical views of a minimal energy *gt-gt* conformer for an eleven (or twelve) residue segment for both the *Ba* Sterne 34F2 and the *Bc* G9241 HF- SCWPs. The backbone segments depicted are of the sequence: **C-A-B-C-A**, with ManNAc (residue **B**), depicted in the middle of the segment since this ManNAc residue is the focal point of the structural difference. Of interest is the close proximity of the α-Gal (**E**) and β-Gal (**F**) substituents at the 6-linked GlcNAc (residue **A**) to each other, as well as their close proximity to the α-Gal monosaccharide substituent (residue **G**) at ManNAc in the *Bc* G9241 segment. We propose that accessibility of these two Gal residues (**E** and **F**) to the mAb could be reduced by the presence of a monosaccharide Gal substituent at ManNAc (as in the human isolate *Bc* G9241), and further reduced by the presence of a Galα1→3Gal disaccharide at ManNAc, resulting in the weaker binding for the great ape isolates. The particular low energy conformer depicted in also suggests the possibility that exposed portions of the two closely located Gal residues (**E** and **F**) might contribute, perhaps equally, to the EAII-6G6 epitope. Note also that the SCWP from *Bc* ATCC 10987 does indeed carry a β-Gal substituent, yet is essentially non-reactive with the mAb EAII-6G6 (**,**). This observation is consistent with the hypothesis that the *Ba*-conserved “pathogenicity motif” backbone is essential in presenting the Gal substituents to the mAb, by providing the proper conformational scaffolding. Although not proven here, we suggest the possibility that the combined arrangement of both Gal residues **E** and **F** may contribute to the EAII-6G6 epitope. The HexNAc backbone itself probably does not interact directly with the mAb, (note the non-binding observed with the *Ba* CDC684 SCWP), however, this pathogenicity-specific backbone is necessary to provide the proper scaffolding for presentation of the multiple Gal epitopes which do interact with the mAb. The models provide a visual assist to our central hypothesis, that the correct number of Gal residues, carried on the proper HexNAc backbone at specific locations, are required for optimal mAb EAII-6G6 recognition. Removal or addition of Gals or any change in the Gal attachment pattern, from that displayed in *Ba* strains, results in variable degrees of attenuation of mAb binding. The aminogycosyl backbone is also likely to be of importance for the SCWP immunoreactivity with the mAb, as shown with the nearly abrogated reactivity of SCWP samples from the non-infection-associated *Bc* isolates. In contrast to the conserved core repeating trisaccharide backbone of the infection-associated *Ba* and *Bc* strains, the HF-SCWPs of *Bc* ATCC 10987 and the *Bc* type strain ATCC 14579 possessed entirely different backbone repeating structures. Presumably due to the differences of the aminoglycosyl backbones of these strains, they presented significantly higher EC₅₀ values when compared to the *Ba* Sterne 34F2 and *Ba* 7702 strains. The EAII-6G6-2-3 antibody has been used in diagnostic tests and to probe for the distribution of the EAII-6G6-2-3 epitope in SCWPs of *Bc* group strains. De *et al*. analyzed 230 *Ba* isolates with a two component direct fluorescent antibody (DFA) assay using fluorescein labeled EAII-6G6-2-3 against the cell wall polysaccharide antigen and a monoclonal antibody against the capsule antigen. Of the 230 *Ba* isolates tested, 228 were positive for reactivity with EAII-6G6. The assay specificity for EAII-6G6-2-3 was only 78.6%, and almost 100% of the false-positive isolates were *Bc* or *Bacillus thuringiensis* strains. Ten of 23 (43%) *B*. *cereus* strains were reported positive for the cell wall antigen. Further DFA analysis of *B*. *cereus* strains G9241 and 03BB87 were also positive for cell wall polysaccharide antigen, whereas that of *Bc* ATCC 14579 was negative. We suggest that the mAb antibody is recognizing a specific epitope in the cell wall polysaccharide antigen, which is impacted by number and position of galactose modifications. While we have shown that removal of all Gal residues (*Ba* CDC684) causes loss of binding and the addition of Gal to ManNAc reduces binding (*Bc* G9241/03BB87, and *Bc* CA and CI), we do not yet know the specific contributions of each of the three Gal residues in the *Ba* structure that provides optimal binding. Our minimal energy models suggest a preliminary possibility that Gal residues **E** and **F** may together contribute to the EAII-6G6 epitope. A detailed analysis of the minimal *Ba* SCWP epitope and Gal positions required for EAII-6G6-2-3 binding is being investigated by preparing synthetic structural analogs. Further, while we know that the terminal repeat unit is pyruvylated, O-acetylated, and de-*N*-acetylated, we do not know the extent of their contribution, if any, to the mAb-epitope. However, all of our data suggest that these non-carbohydrate substituents are not key elements of the EAII-6G6 epitope. For instance, HF-SCWPs from all of the examined strains, (all pathogenic *Ba* strains, the *Ba* CDC684 non galactosylated strain, the human isolates *Bc* G9241 and *Bc* 03BB87, the great ape isolates *Bc* CA and CI, and both ATCC environmental strains), carry a pyruvate residue at identical location on the terminal repeating unit as reported here (**,** and **Figs**) and discussed in (26), however, only the galactosylated strains *containing the pathogenicity-associated HexNAc backbone* demonstrate significant EAII-6G6 binding. It has been established that cell wall polysaccharides can be virulence factors in both Gram-positive and Gram-negative pathogens and are used as diagnostic targets or as carbohydrate-based vaccine antigens. For instance, the galactomannan cell wall polysaccharide of *Aspergillus fumigatus*, an opportunistic fungal pathogen, is an antigen that is the basis for a diagnostic test against the tetra-β-(1→5) galactofuran chain. As such, cell wall polysaccharides amongst the closely related *B*. *cereus* group may be useful in development of rapid diagnostic tests for inhalation anthrax. Monoclonal antibody EAII-6G6-2-3 has been used and previously published as a differential diagnostic reagent for *B*. *anthracis*, and it is still used at CDC for immunohistochemistry tests in tissue. We demonstrated that the ability of mAb EAII-6G6-2-3 to distinguish between disease causing strains of the *Bc* group is based on the number and position of galactosyl modifications around a secondary cell wall polysaccharide backbone. So far, in *Bacillus* species capable of causing infection this polysaccharide was comprised of →4)-β-D- Man*p*NAc-(1→4)-β-D-Glc*p*NAc-(1→6)-α-D-Glc*p*NAc-(1→ and we propose that this aminoglycosyl core constitutes a pathogenicity-related motif associated with *Bacillus* infectivity. In addition, a putative association between pathogenicity and the galactosyl modifications is emerging. Based on the reported *in vivo* activities of the various strains evaluated here, the presence and relative positions of the modifications may be associated with relative infectious potency. For example, all the strains containing the →4)-β-D-Man*p*NAc-(1→4)-β-D-Glc*p*NAc-(1 →6)-α-D-Glc*p*NAc-(1→ core motif are associated with human disease or capable of causing fatal infection in animal models. However, the *Ba* CDC684 strain lacking galactosyl residues is highly attenuated in a guinea pig model of infection whereas the galactosylated *Ba* Sterne 34F2 and *Ba* 7702, although lacking the poly-D-glutamic acid capsule virulence factor, retain a significant level of virulence in mouse and rabbit models of anthrax. The accumulating data are increasingly indicative of the aminoglycosyl core motif as a marker of pathogenicity and it is tempting to speculate that the core motif in its glycosylated form may be a putative pathogen associated molecular pattern that enhances the capability of the organism to cause infection. # Supporting information The authors thank Teresa G. Abshire (USAMRIID/Frederick, MD, USA) for providing *Bc* CDC684 cell materials, John Glushka (CCRC/UGA, GA, USA) for assistance with NMR analyses, and Jarad Schiffer (CDC, Atlanta, GA, USA) for his help in interpreting the immunological data. The findings and conclusions in this report are those of the author(s) and do not necessarily represent the views of the Centers for Disease Control and Prevention. Patent Pending–University of Georgia Research Foundation, Inc. and US Centers for Disease Control and Prevention. ATCC American Type Culture Collection *Bc* CA *B*. *cereus strain Cameroon (CA)* *Bc* CI *B*. *cereus strain Côte d*’Ivoire (CI) BcG9241 *B*. *cereus strain G9241* Ba684 *B*. *anthracis CDC 684* BSL *Bacillus S-layer associated protein* COSY ¹H-¹H correlation spectroscopy DSS 2,2-dimethyl-2-silapentane-5-sulfonate, sodium salt gt-gt *gauche-trans* GC-MS gas-liquid chromatography-mass spectrometry HF hydrofluoric acid H2BC heteronuclear 2-bond correlation spectroscopy HF-SCWP SCWP released from peptidoglycan by HF treatment HMBC heteronuclear multiple bond coherence spectroscopy HSQC heteronuclear single quantum coherence spectroscopy NOESY nuclear Overhauser effect spectroscopy PG peptidoglycan PMAA partially methylated alditol acetate Pyr pyruvate red-HF-SCWP HF-released SCWP treated with borodeuteride to reduce the reducing end SCWP secondary cell wall polysaccharide SEC size exclusion chromatography SLH surface layer homology TMS trimethylsilyl TOCSY total correlation spectroscopy [^1]: The authors have declared that no competing interests exist.

# Introduction Rheumatoid arthritis is a chronic, systemic autoimmune disease of unknown etiology, affecting between 0.3 and 1% of the global population. It affects the joints, connective tissues, muscle, tendons, and fibrous tissue leading to reduced quality of life, disability, and early mortality for sufferers. Established treatment approaches, focused on sequential monotherapies and step- up combination therapies, have underserved many patients. Recently, it has become clear that addressing the underlying inflammatory processes early and prompt disease treatment with biological disease-modifying anti-rheumatic drugs (bDMARDs) such as TNF inhibitors (TNFi), is more successful in terms of limiting radiological progression and minimising loss of mobility and function. However, whilst targeting TNF in RA patients represents a significant advance in treatment options, approximately 30–40% of treatment-naïve patients in clinical trials do not respond adequately to current drug treatment. Although these reagents are highly beneficial in the majority of patients, their relatively high cost and potential safety liabilities are problematic in prescribing for broad patient populations. Several factors have been identified in susceptibility to RA, including 2 to 3-fold increased risk in women and 1.3 to 2.4-fold higher risk among smokers. An increase in risk is also seen in individuals positive for anti-citrullinated peptide antibodies. Despite extensive efforts, there are currently no accepted molecular or genetic biomarkers qualified for use in RA diagnosis and therapy. Therefore, current treatment paradigms operate without guarantee of patient benefit. Until recently, the search for genetic markers associated with RA and anti-TNF response has focused on genes involved in RA susceptibility and disease pathways and genes involved in TNF production and signalling. Over the last decade, a pool of DNA variants proposed as predictors of response to anti-TNFs has been uncovered following pharmacogenetic analyses of RA cohorts. For example, the common *TNF* single nucleotide polymorphism (SNP) -308G\>A (rs1800629), proposed to be associated with RA treatment outcome, has been rigorously investigated with often conflicting results. More recently, genome-wide array and sequencing approaches have afforded researchers increased scope to search in an unbiased manner for clinically relevant biomarkers associated with RA and response to TNF blockers. However, the first round of these analyses, focused on European or Caucasian patients undergoing therapy with etanercept, infliximab and adalimumab, have generated further ambiguous and conflicting findings. The largest genome-wide association study to date, consisting of 2706 patients from 13 European cohorts, including those which had previously indicated SNP associations with RA, showed no association in a meta-analysis, possibly due to data or population heterogeneity. However, eight candidate loci did show replication in Dutch patients from the DREAM cohort (n = 882), with consistent suggestive associations in two further cohorts in a subsequent meta-analysis. Although none of these markers reached genome-wide significance, the directionality of association for three SNPs (rs1568885, rs1813443 and rs4411591) was consistent in all four analyses and may merit further study. Subsequent genome-wide studies and meta-analyses have had limited success at replicating most candidate SNPs, however associations with *GFRA1*, *MED15*, *PTPRC*, and the *PDE3A*-*SLCO1C1* region have been successfully replicated in at least one independent cohort. Recent evidence suggests that some these SNP associations may be drug-specific, which could explain the inability of some studies to replicate them. Integrated analysis of genomic and transcriptomic datasets has confirmed the therapy-specific nature of TNFi response. In addition, substantial phenotypic heterogeneity between RA cohorts may make replication of TNFi SNP associations more difficult. Therefore, machine learning approaches capable of abstracting clinical and genetic predictors have been developed in an effort to classify patients according to their likelihood of responding to treatments. Despite the relatively high heritability of TNFi response, inclusion of SNPs previously associated with TNFi response provides limited improvement in the accuracy of machine learning models. To date, no GWAS has been reported for certolizumab pegol. We report the first genome-wide study of a North American cohort of RA patients with moderate to severe RA, with the objective of discovering DNA variants associated with CZP response. # Materials and methods ## Patients The study protocol was approved by the Institutional Review Boards at Duke and Columbia University and written informed consent for participation was obtained at study entry for all subjects. North American patients (≥18 years) were selected from the REALISTIC phase IIIb study (NCT00717236). All participants were diagnosed with adult onset RA as defined by 1987 American College of Rheumatology (ACR) criteria for at least 3 months. In addition, they had either not tolerated or responded unsatisfactorily to at least one DMARD prior to study entry. Patients were assessed using a 28 joint count and were defined as having active disease if they had at least five tender and at least four swollen joints, C-reactive protein (CRP) ≥10 mg/l and/or ≥ 28mm/hour erythrocyte sedimentation rate (ESR) and had moderate to severe RA at study entry. ## GWAS of common SNPs ### Genotyping A total of 2,372,361 SNPs in 413 patients were genotyped using the Illumina Omni2.5M array platform. SNP genotypes were called using the Illumina GenomeStudio Software package Version 2011.1 with genotyping module version 1.9.4 according to the manufacturer’s instructions. *Quality control (QC)*. A series of QC checks were carried out to ensure sample integrity. Biological sex was assessed for concordance between the genetically- inferred and self-reported gender. Duplicate samples and cryptic relatedness were identified based on genetic data using PLINK. Samples found to be related at the level of first cousins or closer (pairwise estimate of identity-by- descent exceeding 0.125) were considered excessively related and one sample from each pair was removed at random. Quantitative estimates of ancestry (principal components analysis (PCA) implemented in EIGENSTRAT software) were performed using LD pruned SNPs to estimate genomic ancestry and compared with self- reported ethnicity. Data quality was also evaluated at the marker level to remove low quality genotypes. Two subjects missing \>10% of genotype calls and one subject with an anomalously low inbreeding coefficient (F statistic outside of 6 standard deviations from the mean) were excluded. A total of 8629 poorly performing markers missing \>10% of genotype calls and 3144 heterozygous haploid genotypes were removed. In addition, rare variants with a minor allele frequency (MAF) of \<1% were excluded. No SNPs were found to deviate from Hardy-Weinberg equilibrium when a Bonferonni-adjusted threshold of 2.91x10^-8 was applied. Following QC, data were available for 1,685,541 SNPs in 360 individuals of primarily European ancestry. Of these, 302 individuals treated with CZP were included in the statistical analysis. Disease activity Score (DAS28 ESR) data were not available for all individuals and where data was missing for individual outcome measures, those patients were excluded from subsequent analysis. Only a minor difference in ACR20 response (45.7% vs 48.3%) was observed between weeks 6 and 12, therefore only week 6 (ACR20_W6) was used for the analysis. ### Imputation After QC, non-ambiguous SNPs, including rare variants with MAF\<1%, were used as input for imputation using haplotypes from the 1000 Genomes phase1 integrated reference panel. Prephasing of haplotypes was performed using SHAPEIT software. Imputation of missing and un-genotyped SNPs was then performed on 5Mb segments from each autosome and chromosome X using IMPUTE2 according to recommended best practices. High confidence SNPs with greater than 90% imputation confidence were then merged using GTOOL and converted back to PLINK format for association analysis. ### Statistical analysis Four primary outcomes were tested for association within the study: ACR20 and decrease in DAS28 ESR \>1.2 at week 6 (both as dichotomous variables), and change in DAS28 ESR at week 6 and 12 (continuous variables). Testing for association between individual outcome measures and each SNP were carried out using logistic or linear regression, after correcting for gender and genetic ancestry using the top 3 principal component axes with significant Tracy-Widom statistics, explaining 2.78% of the observed variance. ### Analysis of candidate SNPs Using a targeted approach harnessing known GWAS hits in autoimmune disease and genes involved in the TNF pathway, SNPs with higher prior probability of association with CZP response were tested for enrichment of association beyond that expected under the null hypothesis. Of 1618 SNPs previously reported to be associated with auto-immune diseases (SNPs with p\< 5x10^-8 obtained by searching GWAS Central for the MeSH term “autoimmune disease”), 72 were specifically associated with RA. These SNPs have a higher probability of association with the CZP outcomes than SNPs selected randomly. A total of 1403 out of 1618 SNPs were present on the Omni2.5M chip or were available in the imputed dataset. Of the 72 RA-associated SNPs, 64 were available for testing. To examine the role of genetic variation in genes involved in TNF signaling, a further hypothesis driven approach was carried out using 13303 common SNPs located within 112 genes that were annotated as participating in TNF signaling (hsa04668) according to KEGG pathway ontologies. Significance scores for the subset of 13303 TNF-signaling SNPs were compared against a uniform distribution of p-values that would be expected under the null hypothesis using quantile- quantile plots. A review of the literature was used to identify 94 SNPs previously reported to be associated with TNF inhibitor response. Of these, 86 SNPs were present in the Omni2.5M array or imputed dataset. ## Whole-exome sequencing ### Patient selection Extreme non-responders (hereafter referred to as “non-responders”) were defined as ACR20 failures at week 6 and 12, with maximum reduction in DAS 28 ESR at week 12 compared with baseline (BL) of 0.33 and maximum reduction in DAS 28 ESR at week 6 compared with BL of 0.6, to eliminate secondary loss of response. Extreme responders (hereafter referred to as “responders”) were defined as individuals with a positive ACR70 score at 12 weeks. Using these criteria, a total of 81 non-responders and 40 responders were identified (summary shown). ### Sequencing and quality control Whole-exome sequencing (WES) was performed using the Roche Nimblegen Human Exon v2 capture kit and KAPA library prep kit according to the manufacturer’s instructions. Sequencing was performed using the 100bp paired-end read protocol on the Illumina HiSeq 2000 instrument. After quality filtering the raw sequence data using CASAVA (Illumina, Inc., San Diego, CA, USA), adapter sequences were trimmed and the sequencing reads were aligned to the reference human genome (NCBI37/hg19) using BWA software. Duplicate reads were then removed from the dataset using Picard (Broad Institute, Boston, MA, USA). Variant calling was performed using GATK with local re-alignment around insertion/deletion variants (indels) and base quality recalibration for single-nucleotide variants (SNVs). Similarly sequenced control samples available at the Duke Center for Human Genome Variation (CHGV) served as the comparison group for a subset of the exome sequencing analyses. ### Assesment of population structure As the majority of subjects were self-described Caucasians, we sought to limit our analyses to that ancestry group in order to minimize the influence of population stratification between cases and controls. Preliminary ancestry predictions were perfomed during bioinfomatic processing using a panel of 12840 high coverage SNPs, based on principal components generated by EIGENSTRAT for cases, controls, and set of 4289 samples with pre-evaluated genetic ancestries. All cases and controls were initially assigned to one of six geographic ancestry groups (Caucasian, African, Hispanic, East Asian, South Asian or Middle Eastern) using a multinomial logistic regression model which provided a probability estimate of a sample belonging to each of the six groups. After restricting the cases and control samples to those predicted to be of Caucasian ancestry, all samples were further required to have EIGENSTRAT PC scores within 6 standard deviations of the mean on PCs 1 and 2. From an initial 121 CZP-treated patients and 2654 CHGV controls, ancestry pruning resulted in 74 CZP-treated patients (19 responders, 55 non-responders) and 1546 controls of European ancestry. After removal of ancestry outliers, no major differences in population structure were observed between case-control groups. Therefore no additional adjustment for genetic ancestry was included in the WES association analysis. The first two principal components accounted for 0.3% of the observed variance. Study characteristics after ancestry pruning are shown in. ### Variant quality filtering and association testing Variants were filtered to include only those located in CCDS genes, with a GATK variant quality score in the 99.9 percentile and in the following functional categories: stop-gained, stop-lost, nonsynonymous coding, essential splice site, and coding region indels. Genotypes with ≤3 reads were considered missing for any individual. Variants were excluded if they were missing in ≥10% of cases or controls, or if they showed significant deviation from Hardy-Weinberg equilibrium in controls. In addition, 6044 variants previously identified as being artifacts based on comparison of CHGV controls with publically available data from the Exome Variant Server (EVS; <http://evs.gs.washington.edu/EVS/>) were removed from the dataset. Individual variants were tested for association using Fisher’s exact test. In addition, gene-based collapsing tests were performed following the method of Li and Leal, with adjustment for differences in coverage at the exon level between cases and controls. Variants qualified for inclusion in these tests if they had a minor allele frequency (MAF) \<1% in both the study sample (combined cases and controls) and EVS, and fell into the functional categories above. There were three major comparisons in the study: responders vs non-responders, responders vs controls + non-responders, and non-responders vs controls + responders. The within-cohort comparison of responders and non-responders benefits from the clarity of the two alternative phenotypes; however, since the patients in these groups were selected from the extremes of the outcome distribution, it is reasonable to also compare each extreme to a population control sample, as the misclassification rate in controls is unlikely to be high and the sample sizes available are much larger. Statistical significance was assessed using Bonferroni correction for the number of tested variants or genes (in the collapsing analysis). # Results We have carried out a two-stage genetic analysis seeking to identify DNA variants associated with response to CZP in a North American cohort with moderate to severe RA. ## Patient population The patient population was derived from the REALISTIC phase IIIb study and the subjects used for the GWAS are summarized in. 79.5% of participants in the study were women, with average age at entry of 56 years. Mean duration of disease was 9.1 years with 22% of the population having disease duration of \<2 years. Median CRP level at entry was 9.0 mg/l and median ESR was 36.0 mm/h. Mean DAS 28 ESR at entry was 6.4. 60% of the population for which data were available was positive for anti-citrullinated peptide antibodies and 69% positive for rheumatoid factor. Previous exposure to anti-TNF agents had occurred in 47% of patients, while 72% were on methotrexate, 62% on steroids and 23% on statins at baseline. A small number of patients were receiving other DMARDs on study entry (lefluonamide n = 21; azathioprine n = 1). ## GWAS After Bonferroni correction for multiple testing, using categorical or continuous outcome variables (ACR20 and DAS28), we were unable to demonstrate genome-wide significant associations between any genotyped SNP and therapeutic response to CZP for any of the outcome measures. For three of the measures, ACR20 and DAS28 ESR reduction of \>1.2 at week 6 (dichotomous variables), and change in DAS28 ESR at week 6 (a continuous variable), there was no evidence of SNPs associated with outcome. For the change in DAS28 ESR at week 12 outcome measure, two SNPs approaching genome-wide significance were observed (rs12287315 and rs35355083). Neither of these SNPs were located within genes; rs12287315 (p = 5.78x10^-8; β = -1.08) is located 76 kb upstream of the spondin-1 (*SPON1*) gene and rs35355083 (p = 1.57x10^-7; β = -1.14) is located in an intergenic region 110 kb upstream of the semaphorin 4G (*SEMA4G*) gene. Imputation of ungenotyped SNPs in a 10Mb window around each lead SNP identified two additional SNPs in these regions with stronger associations with DAS28 ESR at week 12. These included rs78675205 (p = 2.61x10^-8; β = -1.21) on chromosome 10, located downstream of the Paired Box 2 (*PAX2*) gene, and rs72873110 (p = 9.61x10^-9; β = -1.13) located on chromosome 11, approximately 3kb upstream of the *SPON1* coding region. While the significance scores for these two variants exceeded the original threshold for statistical significance, they were not statistically significant when the association analysis was performed using all common SNPs from a genome-wide imputation and no additional associations were detected outside of these two regions. ## Candidate SNP and pathway analysis We supplemented our genome-wide search for variants with a targeted analysis focused on SNPs with a higher prior probability of association, through previous documentation of association with auto-immune disease, as well as a sub-group of 64 SNPs associated specifically with RA. Similarly, we generated a candidate list of 13303 common variants observed in 112 genes reported to be involved in TNF biology. No enrichment of association beyond that expected under the null hypothesis was observed using these approaches (– Tables). We further tested whether SNPs previously associated with response to TNF inhibitors showed evidence of association in the current study. As shown in, the previously reported associations with TNF inhibitor response were not observed in our cohort. ## Exome sequencing analysis In a complementary analysis, we screened for rare variants by WES of 19 extreme responders and 55 non-responders. Additionally, patients from these extremes of the response distribution were compared with a large sample of more than 1500 ethnically matched control samples sequenced in the same laboratory using identical data processing methods. shows the sample size and the number of variants tested in each comparison. For all comparisons, the single-variant tests did not reveal any variant showing significant association with responder or non-responder status. The p-value distributions were non-uniform and tended to be less significant than expected under the null, probably owing to both the small sample sizes and the relatively low allele frequency distribution of the variants tested. For gene-based collapsing tests, the results were similar, with no individual gene showing significant association with therapeutic response after correcting for multiple testing. The top ten SNPs and genes most significantly associated with CZP response are shown in Tables and. Interestingly, among the top ten most highly associated results in the gene- based analysis was the potassium channel, subfamily K, member 5 (*KCNK5)* gene, which is known to be a critical factor in T cell activation. In addition, *KCNK5* expression has previously been reported to be strongly correlated with RA disease severity and up-regulation of *KCNK5* expression was found to be predictive of treatment failure in RA patients receiving the anti-IL-6 therapy tocilizumab, which inhibits RA inflammation through a different biological pathway to that targeted by anti-TNF therapies. In our collapsing analysis, non- synonymous variants in the *KCNK5* gene were present in 16% (3/16) of CZP extreme responders, while no patients (0/55) from the CZP non-responder group were found to have any predicted functional variants in the *KCNK5* gene. # Discussion We have completed an exploratory pharmacogenetic analysis of a cohort of 302 RA patients of Western European descent treated with CZP. A two-stage analytical approach was employed, using high content beadchip genotyping analysis to examine the role of common human genetic variation, followed by a second stage consisting of WES of extreme responders and non-responders, to investigate the association of rare genetic variants with CZP treatment response. In the first stage of the analysis examining common variation, we were unable to demonstrate robust associations with any of the clinical outcome measures. Two novel candidates (rs35355083 near *SEMA4G* and rs12287315 near *SPON1*), showed strong trends toward association with change in DAS28 ESR at week 12, but did not meet genome-wide statistical significance. These variants have not been associated with response to TNF inhibitors previously, however a low-frequency variant located in the *FAR1*-*SPON1* intergenic region was previously reported to be associated with Etanercept response. Imputation of ungenotyped SNPs near these suggestive associations revealed two additional variants with stronger associations with DAS28 ESR at week 12. While the p-values for these associations exceeded the original threshold for statistical significance, they were not found to be significant when the association analysis was performed using all common variants from a genome-wide imputation. Semaphorins are a diverse group of proteins involved in regulation of cell movement and migration via interaction with cognate plexin or neuropilin receptors. Recent studies have indicated that they play a role in many aspects of the immune system, including innate immunity and cell trafficking. In the mouse, SEMA4G is required for cerebellar development, however, a clear role in auto-immune disease has yet to emerge. Spondin 1(SPON1) is an extracellular matrix protein involved in regulation of neuronal outgrowth and inhibition of angiogenesis. It has been shown to bind to the extracellular domain of amyloid precursor protein and impairs cleavage by the beta secretase BACE 1, reducing beta-amyloid production. Its expression has been reported to be elevated in osteoarthritis lesions in both humans and rodents, and its activation of TGF-beta *in situ* may promote osteoarthritis pathogenesis. The involvement of these genes in immune signaling, as well as the reported role of spondin 1 in cartilage homeostasis, suggests that these associations may be biologically relevant. However, their failure to show significant association with CZP response after correction for multiple testing should prompt caution with regard to interpretation of the findings. As with all such candidates, replication of these SNPs in additional cohorts and biological validation will be required. Whilst the genotyping stage of the analysis was intended to screen for common variants associated with CZP response, for complex diseases such as RA, such approaches typically require large cohorts and/or large effect sizes. One particularly compelling reason to search for rare variants in the study of therapeutic response to drugs is that drug response may be correlated with the underlying genetic cause of disease, and it is expected that much of the risk for human disease will be found in the rare-variant spectrum. To address the risk of overlooking low frequency variants that contribute to CZP response, we implemented a WES approach in the second stage of the study. This approach enables identification of rare variants with larger effect sizes, despite the small sample sizes available. In this case, sequencing focused on 74 patients from the REALISTIC trial with extreme responses to CZP treatment; including 55 non-responders and 19 super-responders. Neither the single-variant tests nor the gene-based collapsing method identified any gene or variant that was significantly associated with CZP response. A modest, but non-significant, enrichment of variation was identified in the *KCNK5* gene in our collapsing analysis. The previous implication of this gene in RA disease progression, as well as the specific observation that increases in *KCNK5* expression are correlated with the failure of antibody-based therapies targeting similar inflammatory pathways involved in RA disease pathology, may suggest that additional investigation of *KCNK5* as a modifier of RA therapies is merited. With the caveat that *KCNK5* and the two SNPs in *SEMA4G* and *SPON1* may be worthy of follow up, we have failed to demonstrate the presence of any variants with a clinically robust association with response to the anti-TNF agent CZP. While the lack of statistically significant associations could result from the limited size of our study, this finding is also consistent with many previous candidate gene and genome-wide efforts to discover genetic determinants of response to anti-TNF agents. Researchers have either been unable to discover or replicate initial findings in clinically relevant cohorts, or the statistical performance of the candidates has not been sufficiently compelling to justify follow up or clinical implementation. A *post hoc* analysis of the estimated statistical power for the GWAS component of our study indicated it had over 80% power to detect common variants with a MAF of 25% and a relative risk of 2.5, but was only sufficiently powered to detect more rare variants with a MAF of 5% when the relative risk was over 5. Similarly, the WES portion of our study was likely to be underpowered to detect enrichments of rare variants with moderate effect sizes (RR\<6.0;). RA is a heterogeneous disease with complex genetic and environmental etiology. Although we have focused our analysis on a largely homogenous Caucasian cohort of 302 patients, the study population comprised patients with differing clinical characteristics at entry. We have accommodated variability in disease severity and duration, and use of other medications in the analysis but were not able to adjust for the variability in biology underpinning disease heterogeneity and response/non-response. In addition, the development of antibodies against TNF inhibitors has recently been recognized a significant contributor to treatment failure. Our study did not account for the induction of antibodies against CZP. It therefore remains a possibility that heritable biomarkers predictive of risk do exist for sub-populations within diverse RA populations such as this one, but that their penetrance is diluted to an extent whereby they are not detectable. Further complexity is added by the routine use of composite outcome measures such as DAS28 and ACR20, the primary endpoints used in this clinical study. They are comprised of a mixture of objective physical measurements, laboratory biomarker data, and subjective patient assessments. Robust associations may be discernible with a more refined focus on a single objective endpoint, although it should be noted that this may be misleading in a mixed patient population. The goal of personalizing RA therapy is to identify patients where the likelihood of a beneficial outcome is optimized and the risk of adverse events is reduced or eliminated. Many studies with conflicting and ambiguous outcomes have been published investigating the pharmacogenomics of etanercept, infliximab and adalimumab in RA populations, and to date there have been no predictive anti-TNF genomic biomarkers established for clinical use. This report, the first pharmacogenomic analysis of CZP response, is in accord with these earlier efforts, and has been unable to demonstrate compelling evidence for the existence of genomic biomarkers of value in targeting this anti-TNF to a responsive patient sub-population. In diseases such as RA, where there are complex, polygenic contributions to pathology and response to therapeutic intervention, it remains a possibility that combinations of clinical, environmental, and molecular markers will hold more promise for managing therapy. # Supporting information [^1]: The specific roles of these authors are articulated in the ‘author contributions’ section. This does not alter our adherence to PLOS ONE policies on sharing data and materials.

# Introduction According to the UNAIDS 2013 report, worldwide 3.3 million children under 15 years were living with HIV in 2012. Over 90% of these infected children live in sub-Saharan Africa. It was estimated that during the same year 260,000 new HIV infections occurred among children, most of them as a result of mother-to-child transmission of HIV (MTCT). Mother-to-child transmission rates declined overall from an estimated 26% in 2009 to 17% in 2012, and coverage of prevention of mother -to-child transmission (PMTCT) services increased to 65%. The number of children younger than 15 years receiving ART increased to 34%. However, these summary estimates comprise enormous disparities between the different countries. In 2012, the coverage of PMTCT services in the Democratic Republic of Congo (DRC) was estimated to be only 13%, and only 9% of children younger than 15 years in need of ART (according to the WHO 2010 guidelines) were receiving ART. Evidence shows that early administration of ART in HIV-infected infants reduces mortality and HIV progression by 75%. However, achieving this depends on several factors, including the early diagnosis of HIV infection, adequate quality of antiretroviral (ARV) drug prescription, good support for children on ART and maintaining/retaining these HIV-infected children in care and ART programs. Patients' retention in care programs remains a major challenge in sub- Saharan Africa. It is also one of the key indicators to assess the success of ART programs. Retention rates among children vary from 71 to 95% at one year, and between 62 and 93% at 2 years in sub-Saharan Africa. To maximize retention in ART programs, a number of factors need to be considered: structural factors such as transportation, availability and accessibility of services, service providers' attitude, caregiver's HIV status and religious and other beliefs (traditional healers), and patient-related factors. The objective of this study was to determine retention rates of HIV-infected children on ART in three different programs and sites in the DRC, and to assess risk factors for attrition. Based on the study findings, we propose recommendations to improve retention in pediatric HIV care and treatment programs. # Methods ## Study Sites Three facilities, known to offer care for HIV infected children and with a sufficient number of patients were asked and agreed to participate in the study. Two of these facilities are located in Kinshasa, the capital of the DRC (an Amo- Congo health facility and the Centre Hospitalier Monkole). The third facility (HEAL Africa) is located in Goma, the capital city in the North Kivu province. Amo-Congo is a Congolese non-governmental organization (NGO) that began in 1993 with providing support for orphans and vulnerable children in Kinshasa. Over the years this NGO expanded its activities to several provinces and included voluntary counseling and testing, and provision of HIV care to adults and children. In 2005 Amo-Congo started, with financial support from Global Fund for AIDS Tuberculosis and Malaria (GFATM) with the provision of ART in their Ambulatory Treatment Centers (ATCs), which are standalone clinics that offer counseling and testing, and HIV care. This study was conducted at the ATC in Kasa-vubu in Kinshasa. The “Centre Hospitalier Monkole”, the Monkole Hospital, is a private health hospital (55 beds), in the Mount Ngafula I health zone in Kinshasa which started its activities in 1991. Their HIV unit started activities in 2005 with financial support from GFATM through the United Nations Development Programme (UNDP), and organizes HIV activities such as HIV prevention activities, ART and management of opportunistic infections. Both the Amo-Congo and Monkole health facility serve an urban population. HEAL Africa Hospital was started in 1998 and is located in Goma/North- Kivu, and serves as a center of excellence for the province. It has 155 beds and works with 4 services/departments including surgery, internal medicine, pediatrics and obstetrics and gynecology. The HIV/AIDS pediatric care service of this facility is called AIDS Children Program (CAP), and was integrated into HEAL Africa hospital in 2005 (with financial support from UNICEF, Global Strategies and the Clinton Foundation). This facility serves populations from peri-rural and rural areas, and supports community organization in the rural area. ## Study type and inclusion criteria From October 2012 to August 2013, we conducted a retrospective cohort study of HIV-infected children on ART at these three health facilities in the DRC. Children older than 6 months and younger than 15 years at the moment of data collection who had initiated ART at least three months prior to data collection were included in the study. ## Data collection We conducted a retrospective chart review of 720 medical charts. In each of the three research sites a random sample of medical charts was selected from a sampling frame that consisted of all children ever started on ART at that site between September 2007 and May 2012. The sampling frame was prepared by the site supervisor; the list of the medical files to be selected was generated by the data analyst using a computerized random number generator (‘sample’ function in Stata). Randomly selected children that were not eligible for the study were replaced by the next child in the sampling frame. Only one child per family was included in the study (the first child from the family on the list). A specifically designed study form was used for data abstraction, which was carried out by a team of two nurses. All the members of the study team were trained in ethics (for confidentiality issues) and data abstraction. ## Data analysis and sample size A random sample of 250 ART-treated children in each site was needed to allow the estimation of the retention rate to be measured with a precision of 5%, assuming the retention rate was at least 80%. In all three study sites, patients received ARV medication according to the WHO 2010 guidelines. Once on ART, HIV-positive children were required to come to the clinic every three months. The child or the caregiver/guardian had to go to the pharmacy every month to collect the ARV drugs. A child or caregiver visiting the clinic within the three months preceding the data collection was considered to be retained in the program. Attrition was defined as death or loss to follow-up. Lost to follow up stands for a person under treatment and missing the scheduled visits beyond 90 days. Multivariable Cox regression was used to assess the association between retention and other variables in the models. The following baseline variables were assessed as potential risk factors for attrition: age, gender, clinical stage, CD4 count at ART start, presence of opportunistic infection, hemoglobin, weight (z-score weight-for-age), height (z-score height-for-age) and education level as children's characteristics; age, education level, gender, religion, profession, civil status and relation with the child as caregivers' characteristics. Variables with p-value ≤0.10 during univariable analysis were considered for inclusion in the multivariable model. Z-scores for weight-for-age and height-for-age were generated using the 2006 WHO growth standards and the zscore06 module in Stata (<http://www.ifpri.org/staffprofile/jefleroy>). The complex model was reduced by stepwise backwards elimination approach. Study site was included in the model as a fixed effect. All associations were tested using likelihood-ratio tests. For variables with less than 10% missing values conditional imputation was used and for variables with at least 10% missing data the missing indicator method (in which a separate category is generated for patients with data missing for a certain predictor). As HEAL Africa Hospital had the highest number of children in care, this site was considered as the reference site. Epi info^TM 7.0.9.34 was used to enter data from all the sites by a trained data manager in Kinshasa. The data analysis was conducted using Epi info and Stata 11. ## Ethics statement The study protocol was approved by the Institutional Review Board (IRB) of the Kinshasa School of Public Health at the University of Kinshasa. All study members were trained in human subject research and good clinical practices. Due to the nature of the study it was not possible and feasible to obtain informed consent from the patients whose medical charts were abstracted for the study. Data were extracted from routine records, and were analyzed anonymously. # Results The number of HIV infected children under ART did not reach 250 as was reported by the Amo-Congo and Monkole facilities. Therefore in total only 720 medical files of children on ART were available at the three study sites. Applying the inclusion criteria, 198 patient files were excluded: 77 patients were above 15 years at the moment of data collection, in 58 files the date of ART initiation was missing and 63 patients started ART within 3 months prior to data collection. After this process, 522 patient files were selected and included for the analysis. The details regarding patient flow are presented in. ## Children's characteristics From 2007 to 20012, 522 HIV positive children were included in the analysis. The mean age of the children was 4.7 years (standard deviation \[SD\]: 3.3); 80% were in WHO clinical stage 3 or 4 at the start of ART. The first line ART regimen provided to 65% of the children was Zidovudine + Lamivudine + Nevirapine (AZT/3TC/NVP). The proportion of children who did not yet attend school was 43%; 27% attended primary school. ## Care givers' characteristics In 51%, the children's caregivers were direct parents (either mother of father). Among the caregivers, 51% were male and the mean age of the caregivers was 42 years (SD: 13). Those who had a formal job represented 53% of the caregivers, 66% had at least the secondary school education. 42% of the caregivers were married, 62% belonged either to the catholic or protestant church, and 37% to the independent church. ## Retention The overall retention probabilities are presented in. Overall retention rates were 88% (95% CI \[95% confidence interval\]: 85%–91%) at 6 months, 85% (95% CI: 81%–87%) at one year, 79% (95%CI: 75%–83%) at two years and 74% (95% CI: 70%–78%) at 3 years. The overall median time in the ART program was 3 years (IQR \[inter quartile range\]: 1.3–5.0), with site-specific median times of 2.5 years (IQR: 1.3–4.6) at AMO-Congo Kasavubu, 1.5 years (IQR: 1.2–3.5) at the Monkole facility and 3.5 (IQR: 1.5–5.0) years at the HEAL Africa facility in Goma. Site- specific retention rates were 88%, 67% and 92% at 6 months; 84%, 59% and 90% at 12 months and 76%, 56% and 86% at 24 months, respectively for Amo-Congo/Kasa- vubu, Monkole facility and HEAL Africa. Out of the 522 HIV infected children, 109 (20.8%) were lost from the ART program: 22 (20%) had died, 87 (80%) were lost to follow-up and, 4 (4%) were transferred to other ART programs. ## Risk factors for attrition The following variables were considered in the multivariable model: study site, the child's age, CD4 count, education level and uptake of Cotrimoxazole; type, gender and religion of caregiver. In the final model four variables remained associated with children's attrition from ART programs: study site, the child's age and CD4 count at ART start and religion of the caregiver. Patients enrolled in the Monkole facility (aHR: 4.77, 95% CI: 2.90–7.85) were likely to experience higher risk of attrition compared to the patients enrolled in care at the HEAL Africa hospital. Children ≤2 years age (aHR: 2.16, 95% CI: 1.48–3.16) experienced higher attrition compared to older children. Children with a CD4 cell count ≤200/µL (aHR: 2.70, 95% CI: 1.40–5.21) and those with a CD4 cell count between \>200–\<350/µL (aHR: 2.87, 95% CI: 1.36–6.07) were likely to experience higher risk of attrition compared to children with CD4 count ≥350/µL. Children whose caregivers were affiliated to the independent church (aHR: 1.86; 95% CI: 1.19–2.90) were more likely to experience attrition compared to those whose caregivers were affiliated to the catholic and protestant church. ## Risk factors for loss to follow up and death In multivariate analysis, four variables remained associated with loss to follow up: study site, the child's age and CD4 count at ART start and religion of the caregiver. Children enrolled in the Monkole facility (aHR: 3.41, 95% CI: 1.98–5.86) were likely to experience higher risk of loss to follow up compared to the patients enrolled in care at the HEAL Africa hospital. Children ≤2 years age (aHR: 1.89, 95% CI: 1.25–2.85.) experienced higher risk of loss to follow up compared to older children. Children with a CD4 cell count ≤200/µL (aHR: 3.13, 95% CI: 1.46–6.68) and those with a CD4 cell count between \>200–\<350/µL (aHR: 3.80, 95% CI: 1.64–8.77) were likely to experience higher risk of loss to follow up compared to children with CD4 count ≥350/µL. Children whose caregivers were affiliated to the independent church (aHR: 1.79; 95% CI: 1.10–2.90) were more likely to experience attrition compared to those whose caregivers were affiliated to the catholic and protestant church. Two variables remained associated with death: study site and the child's age. Patients enrolled in the Monkole facility (aHR: 7.49, 95% CI: 2.84–19.77) were likely to experience higher risk of death compared to the patients enrolled in care at the HEAL Africa hospital. Children ≤2 years age (aHR: 2.78 95% CI: 1.19–6.51) experienced higher risk of death compared to older children. # Discussion Retention rates varied widely across the study facilities: between 93% and 67% at 6 months; 90% and 59% at 12 months and 86% to 56% at 24 months. The highest retention rate was found in Goma, which can be described as a mixture of a peri- rural and rural setting. The community support organization active in this rural area may have played a role in retaining ART patients as was also observed in other settings. Our findings corroborate also with other studies that report higher retention rates in rural areas compared to urban areas,.Retention rates among children on ART in our study were high compared to another study conducted among adults in the DRC, where retention rates of 81%, 75%, 65% and 57%, respectively at 6 months, one year, two years and three years were found. Somehow these better results among children highlight the low quality of care level for patients with HIV infection in the DRC in general. In the DRC the coverage in HIV care services is amongst the lowest in the world. PMTCT interventions remain the entry point for pediatric care treatment in DRC, but PMTCT coverage is very low (13%). Also coverage of family planning in DRC is very low, only 6% of women of childbearing age have access to family planning. The ART coverage in adults is less than 20%, with huge disparities across the country and alarming figures in the rural areas. ART coverage is even more dramatic in children: 9% in 2012. As a consequence of this low coverage of family planning, PMTCT and ART the number of HIV infected children is increasing. Advocacy and increased commitment from government and international organizations is needed to increase services for family planning, PMTCT and ART coverage. HIV services should be implemented in all health facilities allowing nurses to provide the ARV. Retention rates were significantly higher in HIV specialized facilities like HEAL Africa and Amo-Congo/Kasa-vubu compared to Monkole which is a general hospital and which offers a packet of integrated health services. In the DRC, most of public facilities are general health facilities that do not pay special attention to a complex health issue such as HIV. Poor quality of health services is associated with high attrition rates. Patients with a CD4 cell count \<350/µL were more likely to experience attrition compared to those with higher CD4 counts. We should be cautious about the interpretation of these findings. More than 10% of data on CD4 cell count was missing and the CD4 cell count percentage (to determine the eligibility for ART in children) was not available for most of the children. We speculate that children with a CD4 cell count \<350/µL were at higher risk of mortality and lost to follow up as reported in other studies where attrition was associated with advanced disease stage. Younger children (≤2 years) were also more likely to experience attrition. These findings corroborate with others studies conducted in developing countries. The children whose caregivers were affiliated to the independent churches were more likely to experience attrition compared to those whose caregivers were affiliated to the catholic church or protestant church. These independent churches often promise miracle healing. Therefore patients affiliated to these churches are more at risk for attrition. These findings corroborate with others studies where independent church members appear to have lower uptake of health services. Given the impact of religion on people in the DRC, collaboration with religious leaders could help to convince caregivers about the importance not to stop the ART. Specific training sessions for religious leaders about this topic should be considered. Reinforcement of community interventions such as the “Mentor Mothers (MM)” and “Community ART groups (CAG)” is critical. MM approach consists of using HIV positive peer mothers to support, educate, and empower other pregnant women, new mothers and their partners. CAG are peer support groups that provide psychosocial support to HIV-positive women, men, and children in their own community. Each CAG comprises between 6 to 12 members and keeps linkage with the health facility in order to receive ARVs and clinical support as needed. CAG members organize themselves to meet and provide peer support without involvement of the health providers. These support groups also play a key role in defaulter tracing and encouraging attendance at health facility appointments. The CAG approach helps to reduce the frequency of drugs' collection from the pharmacy and therefore the cost for traveling to the clinic. One patient (volunteer) travels to the clinic to collect ARVs for other members of the CAG. The CAG strategy was shown to improve significantly patients' retention and adherence. Our study has several limitations. As in most retrospective studies, some important information was not available in the medical charts e.g. weight, height and CD4%. Moreover, we were unable to assess the influence of the caregivers HIV status because this information was only available in a few of them. Other studies have shown that an HIV positive status of the caregiver may have a negative impact on the children's retention. Our study was performed in three non-public facilities. We speculate that the support provided to these facilities had a positive impact on the quality of data and services. Results may be worse in public facilities in the DRC where the support is scarce. In conclusion, attrition remains a challenge for pediatric HIV positive patients in ART programs in the DRC. In addition, the low coverage of pediatric treatment and the delayed start of ART exacerbate the situation of pediatric HIV/AIDS. Whilst moving towards the new 2013 WHO guidelines, serious commitment, in terms of government's political will and more support from the international donors, is needed to increase access to ART and retention in ART programs. We thank the General Direction of Cooperation and Development (GDCD) from the Belgium Technical Cooperation for the funding tis study. We recognize the significant contribution of Kinshasa School of Public Health (KSPH) and the Institute of Tropical Medicine at Antwerp. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: JDD AT RC MK WB. Performed the experiments: JDD WB CL MK. Analyzed the data: JDD OK RC CL. Contributed reagents/materials/analysis tools: JDD OK CL RC. Wrote the paper: JDD OK RC CL WB AT MK WB. Training for ethics: JDD AT.

# Introduction The genus *Premna* L. was traditionally placed in the subfamily Viticoideae Briq. of Verbenaceae, but is now transferred to the subfamily Premnoideae Bo Li, R.G. Olmstead & P.D. Cantino in Lamiaceae. With more than 100 species recognized (WCSP, <http://wcsp.science.kew.org>), *Premna* is one of the largest woody genera in the mint family, and mainly distributed in Old World tropical and subtropical regions from Africa to China, throughout southern Asia, to northern Australia, and various islands in the Pacific and Indian Oceans. In the past 30 years, a number of regional revisions of the genus have been published, such as for Australia, China, India, Malaysia, New Caledonia, Sri Lanka, Thailand, and Vietnam, and many new taxa were discovered and described. During an academic visit to Vietnam in 2016, Dr. Bo Li examined *Premna* specimens deposited in the herbarium of the Institute of Ecology and Biological Resources of the Vietnam Academy of Science and Technology (HN) in Hanoi and sorted out several unidentified specimens from those of *P*. *stenobotrys* Merr. These specimens were all collected from K’Bang District of Gia Lai Province, one of the five provinces in the Central Highlands of Vietnam, and had been collected between 1978 and 2011 (see for examples). Although the collections were determined to be morphologically identical to each other, they did not match any species currently recognized in *Premna*. Superficially, the unidentified taxon resembles *P*. *stenobotrys*, but apparently differs in an indumentum of dense golden hairs throughout. After careful dissection of few flowers taken from *L*.*K*. *Bien 830* (HN) and examination of digital vouchers of *VK 4457* (HN) provided by Dr. Do Van Hai, we found that this plant bears an atypical structure on its calyx tube: a semi-globose fleshy appendage. After consulting several Lamiaceae experts and a few taxonomists who are familiar with the flora of tropical Asia, we were told that such a calyx structure has not been observed in Lamiaceae, and its corolla is not similar to any known *Premna* species. In Lamiaceae, a slightly similar calyx appendage can be found in *Vitex gamosepala* Griff. of the subfamily Viticoideae or an outgrowth on the calyx forming a shield is common in *Scutellaria* L. species of the subfamily Scutellarioideae, but the structure of these differs from that of the unidentified taxon significantly. In the absence of any mature fruits or material suitable for DNA extraction, further analysis was not possible. To solve this problem, in April 2017, living plants and mature fruits of the plant were recollected successfully in Gia Lai. This provided an opportunity to infer the phylogenetic position of the new taxon in the context of the phylogeny of the Lamiaceae. Finally, its membership within *Premna* was supported by both morphological and molecular characters, and we can confirm that it represents a hitherto undescribed species after comparing the plant with congeneric taxa. It is named as *Premna vietnamensis*, and here reported. # Materials and methods ## Ethics statements Collection of this new species was conducted in compliance with existing regulations for plants defined as non-commercial, as determined by local government offices. Our field investigations did not take place in any natural conservation area and no specific permissions were required for visiting the location of the new species. In addition, our field studies did not involve any endangered or protected species. ## Nomenclature The electronic version of this article in Portable Document Format (PDF) in a work with an ISSN or ISBN will represent a published work according to the International Code of Nomenclature for Algae, Fungi, and Plants, and hence the new names contained in the electronic publication of a PLOS article are effectively published under that Code from the electronic edition alone, so there is no longer any need to provide printed copies. In addition, the new name contained in this work has been submitted to IPNI, from where it will be made available to the Global Names Index. The IPNI LSID can be resolved and the associated information viewed through any standard web browser by appending the LSID contained in this publication to the prefix <http://ipni.org/>. The online version of this work is archived and available from the following digital repositories: PubMed Central and LOCKSS. ## Phylogenetic analysis ### Taxon sampling and molecular data set In order to infer the subfamilial placement of the new species in Lamiaceae, chloroplast DNA sequences *matK*, *rbcL*, and *trnL-F* were generated from its three representatives, and added to a combined data set which was simplified from the data set D155 in Li et al.. The final data set included 64 ingroup accessions in Lamiaceae covering all 12 currently recognized subfamilies and five outgroup accessions representing Mazaceae, Paulowniaceae, Phrymaceae, and Orobanchaceae which are the closest relatives to Lamiaceae in Lamiales. The voucher information and GenBank accession numbers are provided in. ### DNA extraction, amplification and sequencing Total genomic DNA was extracted by using DNEasy^® Plant Mini Kit (QIAGEN^®, Valencia, California, USA) following the manufacturer’s specifications. The DNA extracts were dissolved in TE buffer and preserved at −20°C for further use. A single PCR Polymerase chain reaction (PCR) system and protocol was used for amplification of the three chloroplast regions. Primer pairs, PCR reaction, and amplification program followed Li et al.. PCR products were visualized using agarose gel electrophoresis and sequencing was served by the Sangon Biotech (Shanghai) Co., Ltd., using the same primers for PCR amplifications. ### Alignment and phylogenetic analyses Sequences were assembled and edited using Sequencher v.4.5 (Gene Codes Corporation, Ann Arbor, Michigan, U.S.A.). Sequence alignment was initially conducted in Clustal X v.2.0. and adjusted manually in BioEdit Sequence Alignment Editor v.7.0.0. The combined data set was analyzed using ML and Bayesian inference (BI) methods with gaps treated as simple indels determined by the program Gapcoder. ML analyses were performed on the web server RAxML Black Box. Before each submission, the “Maximum likelihood search” and “Estimate proportion of invariable sites” options were selected, with a total of 1000 bootstrap replicates performed. BI analysis was executed using MrBayes version 3.2.2. The best substitution types (Nst) and rate distribution models (rates) were determined by the Akaike information criterion (AIC) using Model Test v.3.7 with the hierarchical likelihood ratio tests. Four Markov chain Monte Carlo (MCMC) chains were run, each beginning with a random tree and sampling one tree every 100 generations for 20 000 000 generations. Convergence was assessed using the standard deviation of split frequencies, with values *\<* 0.01 interpreted as indicating good convergence. The first 25% of samples were discarded as burn-in, and the post-burn-in samples summarized as a 50% majority-rule consensus tree. ## Morphological observations The morphological description of the new species was based on observation and measurement of fresh and dried specimens, as well as materials preserved in FAA solution (formalin: acetic acid: alcohol = 18: 1: 1). Morphological variation was measured using a ruler and a micrometer. Morphological comparisons within *Premna* were based on specimens or their digital photos held at A, BM, CDBI, E, HN, HNU, IBSC, K, KUN, L, MO, NAS, NY, P, PE, S, US and WH (institutional acronyms follow the Index Herbarium), and on field photographs of some Asian species. High resolution images of the type specimens of *P*. *stenobotrys* (held at A, barcode nos. A00099807, A00099808) were consulted on JSTOR Global Plants (<http://about.jstor.org/>, accessed 10 October 2017), and one of the isotypes deposited at HNU was examined under a stereo dissecting microscope. # Results and discussion ## Phylogenetic position of the new species Based on the combined data set of three chloroplast DNA sequences (*matK*, *rbcL*, and *trnL-F*), BI and ML analyses yielded highly congruent topologies, thus the 50% majority-rule consensus tree from the BI analysis was here presented. Twelve highly supported clades were obtained, and correspond to the 12 subfamilies recognized in Li et al., Harley et al., and Li & Olmstead, viz., Ajugoideae, Callicarpoideae, Cymarioideae, Lamioideae, Nepetoideae, Peronematoideae, Premnoideae, Prostantheroideae, Scutellarioideae, Symphorematoideae, Tectonoideae and Viticoideae. Relationships among the subfamilies are mostly consistent with those presented in Li et al.. Significantly, the three representatives of *P*. *vietnamensis* formed a highly supported clade (posterior probability, PP = 1.00, ML bootstrap, BS = 100) and was deeply nested within the genus *Premna* which received moderate to high supports (PP = 1.00, BS = 85). Although the placement of the new species within *Premna* was confirmed, it is not possible to infer the phylogenetically closest species to *P*. *vietnamensis* because the sample of *Premna* included was not representative of the species diversity within the genus. ## Morphological comparisons In *Premna*, *P*. *vietnamensis* is morphologically and geographically most similar to *P*. *stenobotrys*. They are both woody climbers having ovate-oblong to elliptic leaves and congested thyrses, and their inflorescences extend as they become infructescences. *Premna stenobotrys* occurs in a similar habitat as *P*. *vietnamensis*, frequently climbing on many kinds of shrubs along roadside and evergreen broad-leaved forest edges. Furthermore, though the type specimens of *P*. *stenobotrys* were collected from northern Vietnam, it is widely distributed from central to northern regions of the country based on specimen records, as well as our own field observations, and it was also reported from Thailand. However, *P*. *vietnamensis* can be immediately distinguished from *P*. *stenobotrys* by its indumentum of dense, spreading, golden-brown hairs throughout (vs. nearly glabrous or easily glabrescent), significantly shorter petioles (2.0–6.0 mm vs. 15–50 cm), greenish yellow corolla (vs. red), and exserted stamens (vs. included). The indumentum, corolla and stamens of *P*. *vietnamensis* particularly resemble *P*. *fulva* Craib, a species widely distributed in Thailand, Laos, Vietnam and southwest China, but *P*. *fulva* differs from *P*. *vietnamensis* in having ovate to suborbicular leaves and flat-topped corymboses. Additionally, the fruits of *P*. *vietnamensis* are very similar to those of *P*. *clavata* de Kok, a newly described and remarkable species from Malaysia. The endocarp of both species is noticeably warty, and their fruits have only one well developed seed. On the basis of this fruit type, *P*. *vietnamensis* obviously belongs to the “*P*. *trichostoma*-group”, but apparently differs from every species in the same group. Based on the above molecular and morphological evidences, we can confirm that the plant discussed here represents a hitherto undescribed species, thus formally describe it as below. ## Taxonomic treatment ***Premna vietnamensis*** Bo Li, ***sp*. *nov***. \[urn:lsid:ipni.org:names:77177680–1\] (Figs, and –) **Type**:—VIETNAM. Gia Lai Province, K’Bang District, Dong Commune, 14°13′03.5′′N, 108°36′27.9′′E, alt. 650 m, along roadside No DT669 and edges of primary evergreen broad-leaved forest on sandstone mountains, 28 April 2017, *D*. *V*. *Hai & N*. *S*. *Khang HK 28042017–01* (fl. and fr.) (holotype IBSC; isotypes HN, IBSC, JXAU). ### Diagnosis *Premna vietnamensis* is characterized by its calyx tube bearing a semi-globose fleshy appendage, which has not been reported in *Premna* to date. It is morphologically and geographically most similar to *P*. *stenobotrys*, but differs from it by an indument of dense golden hairs on the branchlets, petioles, and peduncles; much shorter petioles (2.0–6.0 mm vs. 15–50 mm), and by its yellow-greenish corolla and exserted stamens. ### Description Woody climbers, up to 5 m in length. Branchlets, petioles, leaf blades and inflorescences densely covered with long, spreading, golden-brown hairs. Branches grey, terete, sparsely dusty tomentose, with an interpetiolar ridge between petioles. Branchlets reddish brown, spreading, terete, without bracts at the base. Leaves simple, opposite, almost distichous, or slightly decussate, ovate-oblong to elliptic, rarely cordate-ovate, papery, (6.5–) 9.0–14 (–15.5) × 3.5–5.5 cm, apex acute, base cordate, margin serrate in the distal half of the leaf lamina; veins 4–6 pairs, abaxially raised and adaxially obviously compressed, secondary veins curved and jointed near margin; petiole short, 2.0–6.0 mm long, slightly inflated. Inflorescences terminal, a congested thyrse, subglobose to pyramid-shaped, 1.5–3.0 cm long, 2.0–2.5 cm wide, but infructescence elongated to 2.5–4.5 cm, peduncle 1.0–2.5 cm long; bracts lanceolate, ca. 1.0 mm long, easily deciduous; bracteoles subulate, tiny. Calyx campanulate, 2.0–3.0 mm long, green when young but purplish to brownish black when mature, obviously 2-lipped; upper lip emarginate to irregularly 2–3 lobed, covering a semi-globose fleshy appendage behind the lobes, lower lip 2 lobed, lobes slightly fleshy, apex rounded. Corolla greenish yellow, 2-lipped, 4.0–4.5 mm long; tube externally glabrous, internally densely white villose around throat; upper lip 1-lobed, upward reflexed, entire, oblong-obovate, apex rounded; lower lip 3-lobed, middle lobe oblong-obovate, straight or slightly recurved, lateral lobes rounded to obovate, strongly backward reflexed. Stamens 4, subequal, filaments yellowish to greenish white, glabrous, upward reflexed, slightly exserted; anther white. Ovary globose, 1.0–1.5 mm long, glabrous; style yellowish to greenish white, slender, upward reflexed. Fruits drupaceous, clavoid to nearly globose, 5.0–6.0 × 4.5–5.5 mm; exocarp membranous, green when young, from yellowish green, purplish green to bluish brown when mature, sarcocarp succulent, endocarp hard, warty; one pyrene with four seeds, only one fully develops. ### Etymology The specific epithet of this new species, “*vietnamensis*”, is derived from the name of the country, Vietnam, where this distinct species was originally collected and has not been observed in other countries so far. ### Phenology Flower buds were observed in late February to early March. Flowering was observed from mid-March to early May and fruiting from late April to late May. ### Distribution and habitat *Premna vietnamensis* is currently known only from two localities in K’Bang District of Gia Lai Province, Central Highlands of Vietnam Vietnam: Dong Commune and Song Lang Commune, which are ca. 45 km apart. The plant grows well along roadside and edges of primary evergreen broad-leaved forest at elevations around 600–650 m. a.s.l., and climbs up to 5 m on many kinds of shrubs. ### Additional specimens examined (paratypes) VIETNAM. Gia Lai Province, K’Bang District, Dong Commune, 14°13′05.5′′N, 108°36′27.8′′E, alt. 645 m, along roadside No DT669 and edges of primary evergreen broad-leaved forest on sandstone mountains, 28 April 2017, *D*. *V*. *Hai & N*. *S*. *Khang HK 28042017–02* (fl. and fr.) (HN, IBSC, JXAU); Dong Commune, 14°12′16.3′′N, 108°36′17.8′′E, alt. 615 m, edges of primary evergreen broad-leaved forest on sandstone mountains, 28 April 2017, *D*. *V*. *Hai & N*. *S*. *Khang HK 28042017–03* (fl. and fr.) (HN, IBSC, JXAU); Dong Commune, 14°12′16′′N, 108°36′18′′E, 9 March 2011, *T*. *T*. *Bach et al*. *VK 4457* (fl.) (HN); Son Lang Commune, 25 April 1978, *T*. *D*. *Ly 586* (fl.) (HN); Son Lang Commune, 26 April 1978, *L*. *K*. *Bien 830* (HN); Son Lang Commune, 26 April 1978, *L*. *K*. *Bien 833* (fl.) (HN). ### Conservation status Based on our experiences on specimen examination and field observation, we consider *P*. *vietnamensis* as a rare species endemic to Gia Lai Province of Vietnam. It has been observed only in two separate sites in K’Bang District, although in this province it is very common to find similar habitats. Due to insufficient field surveys so far, we have not outlined the precise distribution boundaries of the species, and not ascertained its population status. Therefore, the information currently obtained is too inadequate to assess the species’ risk of extinction, whether directly or indirectly. In accordance with the IUCN Red List Categories, we propose to temporarily list the species as a taxon under the Data Deficient (DD) category. Further field surveys in Gia Lai Province of Vietnam are needed to gain more information on its abundance and/or distribution. # Supporting information We are grateful to the curators of CDBI, IBSC, HN, HNU, K, KUN, HITBC and PE herbaria for permission to access their collections, to Prof. Philip D. Cantino, Prof. Richard G. Olmstead, and Prof. David J. Mabberley for their comments when making identification of the new species, to Prof. Timothy M. Utteridge for sharing the image of *Vitex gamosepala* Griff., to Ms. Yun-Xiao Liu for the line-drawing illustration, to Mr. Hoang Son for his assistance while performing the field surveys. [^1]: The authors have declared that no competing interests exist.

# Introduction Two-dimensional (2D) substrates, such as tissue culture polystyrene and the surface of tissue analogs, make an enormous contribution to modern *in vitro* cell studies; however, traditional 2D platforms can not accurately mimic the complex 3D architecture of the extracellular matrix (ECM) where native cells reside. In 2D culture, the monolayer cells experience homogenous concentration of nutrients and growth factors which induce unnatural cell environments and cell-cell interactions, yielding a flat and stretched morphology. Recent studies have shown that the morphological differences of cells cultured in 2D and 3D can exhibit several striking differences in subtle cellular processes such as proliferation, apoptosis, differentiation, gene expression, migration, and drug sensitivities. On the other hand, the biological *in vivo* 3D systems, such as animal models, are expensive and time-consuming. Therefore, advanced *in vitro* 3D model systems are needed to fill the gap between the inaccurate 2D systems and the animal models, mimicking the complexity of the ECM and the physiological relevance of an *in vivo* biological system. In the last few decades, hydrogel scaffolds, cross-linked networks that possess high water contents, have attracted more and more attention in an attempt to mimic *in vivo* conditions for cell culture. The reticulated structure of cross- linked polymer chains with high water contents introduces a number of desirable cellular microenvironment characteristics: 3D spatial support for cell growth; porosities for cell migration; and facile transportation of oxygen, nutrients, waste, and soluble factors. Hydrogels can be formed from a range of natural sources and synthetic materials. Natural gels derived from ECM components and other biological sources such as collagen, fibrin, hyaluronic acid, chitosan, and alginate are biocompatible and inherit bioactivities that promote cell survival, proliferation, differentiation, and cellular function of many cell types. However, natural hydrogels have varying biochemical presentations and material properties that are difficult to control, which increases the risk and complexity of cellular study in this culture system. On the other hand, synthetic gels are highly reproducible with consistent composition and predictable manipulation of properties. However, synthetic polymers such as polyactide and polyglycolide have too large fiber diameter and porous size, which present poor scaffold structure and mechanical properties to accurately mimic the the full complexity of natural environment of cell growth. With the rapid development of rationally designed peptides as biological materials, peptide based hydrogel was considered as one of the most promising material for 3D cell cutlure because of its amino acid composition and the structural and mechanical similarity to natural ECM. In addition, for *in vitro* 3D cell culture, cell encapsulation and isolation are two critical steps to introduce 3D spatial support for cell growth and recover embedded cells from scaffold matrix for downstream studies respectively. For a convenient, effective, and safe encapsulation, cells should be added simultaneously with the initialization of hydrogelation. Therefore, mild and cyto-compatible hydrogel-forming conditions are preferred, to ensure that cells survive comfortably during gel formation. However, the solubility of gel transformation of current peptide/protein hydrogels (i.e., puramatrix gel, hydromatix peptide hydrogel, and matrigel) is triggered by adjusting pH or temperature. The undesirable low pH or cold temperature of the pre-gel solutions may cause the cell death when they are directly mixed. Hydrogel preparation procedure become complex when changing cell medium for pH balance and chilling experimental tools are required. Cells kept in sucrose solution or a gel-forming buffer struggle with lack of nutrients up to several hours during gel formation before cell medium can be added. Moreover, isolating cell from hydrogel matrix is another challenge to 3D cell culture. For most cases, changing the environmental factors back to extreme conditions or adding undesirable buffer for hydrogel degradation are required to initial the solubility of gel transformation before cells can be separated out. This process threatens the survival of cultured cells and may cause the failure for the whole downstream studies. Therefore, it is necessary to develop a hydrogel which not only presents a convenient and effective process for cell encapsulation, but also provides easy and safe cell isolation for further cell physiological and pathophysiological studies. In this study, a solution of a newly identified peptide called h9e was prepared at neutral pH and mixed with Minimum Essential Medium (MEM, with 10% FBS) at room temperature. After mixing, h9e peptides self-assemble into a hydrogel matrix with a final peptide concentration as low as 1 mM (0.17%). Without introducing any additional gel forming buffer or adjusting environmental pH or temperature, the peptide provides a convenient and mild hydrogel forming process and allows cells to be surrounded by their culture medium during cell encapsulation. More interestingly, the mechanical strength of this hydrogel matrix exhibits special deformability and reassembly capability, which allow the gel-soluble transformation through repeated pipetting. A breast cancer cell line, MCF-7, was selected as a model to grow in 3D culture of the h9e-MEM hydrogels. Studies of cell morphology, viability and proliferation showed that cells exhibited 3D cyto-architecture in the hydrogel matrix and kept high bioactivities for further studies after isolation. Cisplatin, an anti-cancer drug, was used to examine its efficacy on MCF-7 cells in hydrogel matrix. Overall, the results show a strong support that the h9e peptide is a promising 3D cell culture material for drug testing. # Materials and Methods ## Materials N,N-Dimethylformamide (DMF), Trifluoroacetic acid (TFA), piperidine, N,N-Diisopropylethylamine (DIEA), Triisopropylsilane (TIS), paraformaldehyde, glutaraldehyde, cis-Diamminedichloro-platinum, 0.4% Trypan blue, insulin, and anti-actin antibodies were purchased from Sigma-Aldrich (Milwaukee, WI). N-Methylpyrrolidinone (NMP), anhydrous ether, dichloromethane (DCM), and sodium bicarbonate were purchased from Fisher Scientific (Pittsburgh, PA). Rink Amide MBHA Resin, 2-(1H-Benzotriazole-1-yl)-1,1,3,3-tetramethyluronium Hexafluorophosphate (HBTU), and all protected amino acids were purchased from EMD Biosciences (San Diego, CA). N-hydroxybenzotriazole (HOBT) was purchased from CEM (Matthews, NC). Sodium pyruvate, non-essential amino acids, MEM with Eagle’s salts, TrypLE Express trypsin solution, 4′, 6-diamidino-2-phenylindole (DAPI), and Hoechst 33342 fluorescent dye (10 mg/mL) were purchased from Invitrogen (Carlsbad, CA). Fetal bovine serum (FBS) was purchased from Atlanta Biologicals (Lawrenceville, CA). Anti-survivin, anti-Ki67, and anti-caspase-3 antibodies were obtained from Santa Cruz Biotechnologies (Santa Cruz, CA). Anti- cleaved caspase-3 and HRP-linked anti-rabbit/mouse antibodies were purchased from Cell Signaling Technology (Danvers, MA). ## Cell Culture The MCF-7 human breast cancer cell line was purchased from American Type Cell Culture (ATCC) (Manassas, MA). The cells were grown in MEM medium with 10% (v/v) FBS, 1 mM sodium pyruvate, 17.9 mM sodium bicarbonate, 0.01 mg/mL insulin, and 1% (v/v) non-essential amino acid. Cells were maintained in T-75 cm² tissue culture flasks (Greiner Bio-one, Begium) at 37°C with 5% (v/v) CO₂. Cells were routinely passed with trypsin and medium was changed every other day. ## Peptide Synthesis and Hydrogelation The h9e peptide was synthesized according to a previously published protocol. Briefly, peptides were synthesized on an automated CEM Liberty microwave peptide synthesizer (CEM Corporation, Matthews, NC) according to the base-labile 9-fluorenylmethoxycarbonyl (Fmoc) strategy with Rink amide resin and Fmoc- protected amino acids. After final N-terminal Fmoc group deprotection, the resin-bound peptides were side-chain-deprotected and cleaved using TFA/TIS/water (95/2.5/2.5 v/v). Peptides were precipitated and washed three times with anhydrous ether, dissolved in acetonitrile and deionized water (50/50 v/v), then freeze-dried. Molecular weight and purity of the synthesized peptides were confirmed by matrix-assisted laser desorption/ionization time-of-flight mass spectroscopy and high-performance liquid chromatography. Lyophilized peptide was added to 100 mM sodium bicarbonate and completely dissolved by magnetic stirring for 3 hours with a final peptide concentration of 10 mM. For hydrogelation, peptide solution was added into MEM with 10% FBS and the mixture was hand-shaken for about 10 seconds. The peptide hydrogel formed within 15 minutes at room temperature with final peptide concentration of 1, 2, and 3 mM. ## Rheological Tests The storage and loss moduli (G′ and G″, respectively) of h9e hydrogels were determined on a C-VOR 150 rheometer system (Malvern instruments, Malvern, Worcestershire WR141XZ, United Kingdom) with a 20-mm diameter parallel plate geometry and 500 µm gap size. To mimic cell physiological conditions, all rheological tests were performed at 37°C unless otherwise specified. The peptide and MEM mixture was placed on the measuring system immediately after mixing for a gel-forming rate test. Single frequency (1 Hz) and steady shear strain (1%) were selected for a 1hour test. To determine the hydrogel reassembly capability, the peptide and MEM mixture was incubated at room temperature overnight for hydrogelation, then transferred to a lower measuring plate for 10 minutes, single-frequency test (1 Hz, 1% strain) for stabilization. The hydrogel was broken using 1 Hz frequency and 500% shear strain for 1 minute. Afterward, resetting the instrument parameters took 1 minute, and the hydrogel moduli during the reassembly period were measured under 1 Hz frequency and 1% shear strain for 1 hour. The amplitude sweep test (strain from 1 to 500%, 1 Hz frequency) was conducted multiple times to determine hydrogel reassembly capability after every time it was destroyed. Four testing circles were applied in this measurement and the hydrogel recovery time between every two circles was 1, 5, and 10 minutes. Furthermore, to test the response to different environmental temperatures, the peptide hydrogel was measured under a temperature profile test with steady oscillatory frequency (1 Hz) and strain (1%). The temperature was adjusted from 4°C to 50°C for two testing circles. For each circle, the instrument’s heating or cooling processes took 5 minutes, then another 5 minutes to arrive at the setting temperature (4°C or 50°C). To determine the G′ and G″ of hydrogel during cell isolation, 3 mM peptide hydrogel was diluted 15 times with MEM. After thorough mixing, the diluted solution was tested under 1 Hz frequency and 1% shear strain at 4°C for 1 hour. ## Culturing Cells in h9e Hydrogel The h9e solution was sterilized by UV irradiation (254 nm) for 30 minutes before cell encapsulation. MCF-7 cells were detached from the flask using trypsin and pelleted by centrifugation at 2,000 rpm for 5 minutes. The supernatant was removed and pelleted cells were re-suspended in MEM. The number of cells was counted using a cellometer Auto 2000 (Nexcelom Bioscience, Lawrence, MA). MEM containing MCF-7 cells were thoroughly mixed with h9e peptide solution and the mixture was seeded into each well of a 12-well culture plate (Becton Dickinson Labware, Franklin Lakes, NJ). The culture plate was placed in a 37°C incubator (Thermo Scientific, Asheville, NC) for approximately 30 minutes. After the complete hydrogelation, 1 mL of MEM was carefully added to the top of the hydrogel to prevent drying during long-term incubation. The medium on the top of the hydrogel was changed every 2 days. ## Cell Isolation from h9e Hydrogel To isolate the cells from the hydrogel matrix, 2 mL MEM was added to each hydrogel cell culture in the 12-well plate. The hydrogel with embedded cells was thoroughly mixed with MEM using a pipette and transferred to a centrifuge tube. An additional 4 mL MEM was added to wash the cell culture well and collected in the centrifuge tube. One mL trypsin solution was added to the well, which allowed cells to detach from the well. The plate was incubated at 37°C for 5 minutes and the solution was collected into the centrifuge tube again. The centrifuge tube was placed on ice. Another 6 mL MEM was added to the centrifuge tube for a final volume of 15 mL. The solution was thoroughly mixed and cells were pelleted by centrifugation at 2,000 rpm for 5 minutes at 4°C. The supernatant was removed and the cell pellet was collected. ## Cell Distribution Assay MCF-7 cells (1 × 10⁶ cells/well) were cultured in 3 mM h9e/MEM hydrogel for 5 days. The cell clusters were isolated from the gel matrix and re- suspended in 2 mL MEM. Ten µL (10 mg/mL) Hoechst 33342 fluorescent dye was added to stain the cell nuclei. The cell solution was incubated at 37°C for 30 minutes. Labeled cells were washed for 3 times with MEM medium and re-capsulated within the 3 mM h9e/MEM hydrogel matrix. The cell/gel constructs were observed on a LSM 700 confocal microscope (Zeiss, Jena, Germany). ## Scanning Electron Microscopy (SEM) The nanofiber network of hydrogel scaffolds as well as surface characters of the 3D cultured cells were observed under SEM. The hydrogel samples were dehydrated with increasing concentrations of ethanol from 50% (v/v) to 100% (v/v) at 5% per step and 15 minutes for each step. The ethanol was then removed by a critical point dryer (Samdri-790B, Tousimis Research Corp., Rockville, MD). The hydrogel samples with cells were fixed in a 2% paraformaldehyde and 2% glutaraldehyde mixture for 30 minutes before dehydration and critical point drying. Samples were then sputter-coated (Desk II Sputter/etch Unit, Denton Vaccum, Moorestown, NJ) 3 times (12 seconds each time) with 100% Pt. The SEM observation was carried out with an FEI, Nova NanoSEM 430 (Hillsboro, ON) at 5 kV and through a lens detector. ## Cell Morphology MCF-7 cells (2.8×10⁵ cells/well) were cultured in 2D monolayers or 3D hydrogels (1, 2 and 3 mM h9e/MEM hydrogel) for 1, 3, 5, and 7 days. The morphological characters of the cells were determined using an inverted light microscope (Nikon Eclipse TE2000-u, Kanagawa, Japan) at days 1, 3, 5, and 7. ## Cell Viability and Cell Proliferation MCF-7 cells (2.8×10⁵ cells/well) were cultured in 2D monolayers or 3D hydrogels (1, 2 and 3 mM h9e/MEM hydrogel). Cells in 2D culture were harvested and cells in 3D culture were isolated at days 1, 3, 5, and 7. A cell suspension was mixed with 0.4% trypan blue dye at 1∶1 (v/v) ratio. Cell viability and the number of viable cells were examined by Cellometer Auto 2000 (Nexcelom Bioscience, Lawrence, MA). ## Treating Cells with Drugs MCF-7 cells (0.5×10⁶ cells/well) were cultured in 3 mM h9e/MEM hydrogel for 5 days. After 5 days incubation, cell spheroids were isolated and treated with 40 µM cisplatin via three methods: top surface, top-bottom transwell, and pre-mixed with medium and peptide. For the top surface method, isolated cell spheroids were re-capsulated within the 3 mM h9e/MEM hydrogel matrix and cultured in 12-well plates. One mL of MEM medium containing 40 µM cisplatin was placed on top of the hydrogel in each well. For the transwell method, isolated cell clusters were re-capsulated within the 3 mM h9e/MEM hydrogel matrix and cultured on a transwell insert. MEM medium containing 40 µM cisplatin was placed on top and bottom chambers. For the pre-mixed method, isolated cell clusters were resuspended in MEM medium containing 3 mM h9e peptide and 40 µM cisplatin. The mixture was added in 12-well plates (1 mL of mixture/well) for hydrogelation. One mL of MEM was carefully added to the top of the hydrogel to prevent drying during long-term incubation. Medium was changed every 2 days in all the methods. Cell viability was examined at days 0, 1, 3, 5, and 10 by using the trypan blue excision method. ## Immunofluorescence and Confocal Microscopy MCF-7 cells were cultured on coverslips for 2D monolayer culture and with hydrogels for 3D culture. Cells were treated with cisplatin according to each condition, 2D or 3D cultures. Cell clusters in hydrogel culture were isolated. Cells grown on the coverslips or isolated cell clusters were fixed in 4% formaldehyde in PBS for 30 minutes. Cells were washed with PBS and then treated with 0.1% Triton-X 100 for 8 minutes for 2D monolayer culture and 1 hour for the 3D hydrogel culture. Cells were washed with PBS and blocked with 3% bovine serum albumin (BSA) in PBS for 1 hour at room temperature. After blocking, cells were incubated with primary antibodies overnight at 4°C. For the clusters of cell from 3D hydrogel culture, exposure to primary antibodies occurred in 1.5 mL tubes on an orbital rocker. Cells were incubated with Alexa-conjugated secondary antibodies for 1–6 hours at room temperature. DAPI was used to stain nuclei. The coverslips were mounted and sealed on histological glass slides with prolong- antifade reagent (Invitrogen, Camarillo, CA). To avoid flattening of cell clusters, a spacer was used to create a space between the coverslip and microscope slide. Image was captured using a confocal microscope (Carl Zeiss LSM 700 META, Narashige, MN). For the clusters, optical slices were obtained at adequate intervals on the Z axis (between 0.5 and 1 µm). Z-stack images were reconstituted in ZEN 2010 software. ## Western Blot Analysis MCF-7 cells, cultured in 2D monolayer and 3D hydrogel, were treated with cisplatin. Cells grown in 2D were trypsinized and pelleted by centrifugation at 2,000 rpm for 5 minutes. Cells grown in 3D hydrogels were isolated as described above. After washing with PBS, cells were incubated in lysis buffer (Cell Signaling Technology, Danver, MA) for 5 minutes. Cell lysates were sonicated using Vibra-Cell sonicator (Sonics & Materials Inc, Danbury, CT) and then centrifuged at 13,000 rpm for 30 minutes at 4°C. After centrifugation, supernatants were collected as whole cell extracts. Thirty ug of samples were separated by 4–20% gradient SDS-PAGE for 35 minutes at 200 V, and transferred to nitrocellulose membranes (Midwest Scientific, Saint Louis, MO). Membranes were blocked with 5% milk for 30 minutes and immunoblotted against protein of interest. Immunoreactions using chemiluminescence were visualized by FluorChem E Imaging Instrument (ProteinSimple, Santa Clara, CA). Intensities of the bands were digitized using Un-Scan-It software. ## Statistical Analysis Two-sample t-test was used to analyze the difference among the various treatments. P-value of less than 0.05 was considered statistically significant. # Results and Discussion ## Peptide Hydrogelation in MEM To initiate gel formation, 100 µl of 10 mM h9e peptide solution (pH 7–8) was added to 900 µL MEM medium to form 1 mL mixture with 1 mM (0.17% w/v) peptide concentration. The nanoscale morphology of the hydrogel matrix is presented by the SEM image. Peptide hydrogelation induced directly by mixing neutral pH peptide solution with MEM not only avoids the complex chemical gel cross-linking processes, but also utilizes a medium commonly used in biological and medical research, providing a physiological condition to cell encapsulation. Direct loading of drugs, proteins, or cells during gel formation is one of the most convenient and effective ways for encapsulation. To ensure homogenous distribution of embedded molecules, peptides should assemble as a nanofiber network in a relatively short period with reasonable strength, holding the suspended molecules before their precipitation. To determine the peptide gel formation rate, hydrogels were prepared with three concentrations, 1 mM (0.17% w/v), 2 mM (0.34% w/v), and 3 mM (0.51% w/v), in MEM. The storage modulus of the solution was measured at 37°C immediately after thorough mixing. shows the h9e peptide hydrogel formations with stable storage modules around 100, 400, and 700 Pa, respectively. The gel formation rates increase with peptide concentrations (inset of), and all three hydrogels reached a self-supporting strength (close or above 100 Pa) within 15 minutes. SEM images indicate that the hydrogel architecture is built by entanglement of 20 nm width nanofibers; however, the lower-concentration hydrogel (1 mM) shows a relatively looser matrix structure compared with the compact structure of the higher concentration hydrogel (3 mM ). This visual evidence further supports the strength differences of different concentration hydrogels. ## Dynamic Rheological Study of h9e Hydrogel The deformability and reassembly ability of MEM-induced h9e hydrogel were assessed by a dynamic rheological test. Peptide hydrogels with 1–3 mM were stored at room temperature overnight, then transferred to measuring system and stabilized for 10 minutes. By shear-thinning at 500% strain for 1 minute, all three hydrogels were converted to liquid state, showing a G′ lower than 0.2 Pa. After shear-thinning stopped, instrument parameters were reset after 1-minute waiting time and the hydrogel recovery was monitored using 1% shear strain for 1 hour. The data in the G′ of hydrogel recovery during this 1 hour test. To determine whether the hydrogel could maintain the reassembly capability even after shear-thinning many times, the hydrogel was measured under an amplitude sweep test conducted multiple times. Four testing circles were applied in this measurement and shear strain was increased from 1% to 500% within 5 minutes for each circle. For hydrogel recovery, the waiting time of 1, 5, and 10 minutes were applied, respectively. suggests that although the hydrogel architecture was completely broken into liquid form at the end of each circle, quick reassembly persisted even after shear-thinning multiple times. The results also showed that the percentage of recovery G′ increased with waiting time and that the hydrogel reassembly rate related to hydrogel concentrations. For example, with waiting time of 1 minute, about 73%, 76%, and 88% of the gel strengths were recovered for 1, 2, and 3 mM hydrogel, respectively, but after three shear-thinning circles and 10 minutes of waiting, 83% (1 mM), 84% (2 mM), and 100% (3 mM) of the gel strength were recovered. The higher reassembly rate is most likely caused by the more compact matrix structure of hydrogel due to higher peptide concentration (3 mM). In the solution with high concentration of peptide, some non-covalent gel network cross-links remain intact and the broken nanofiber groups are close to each other, making rebuilding the cross-links easy. Based on these rheological properties, the MEM-induced h9e hydrogel could be delivered via pipetting multiple times without permanently destroying the hydrogel architecture. This special shear-thinning and recovery properity of the hydrogel also provides an alternating method for cell isolation from hydrogel matrix through a mechanical shearing and dilution. For biological study, temperatures between 4°C and 37°C are commonly applied for many standard operational procedures *in vitro*; therefore, the response of the hydrogel materials to the temperature variations has a large impact on their practical applications. The rheological temperature profile test was performed to address this challenge. The temperature was adjusted from 4°C to 50°C for two testing circles. presents that the G′ of hydrogels moves along with temperature and performs 2–3 times higher at 50°C than that at 4°C. This thermal response is reversible according to the hydrogel heating and cooling circles. The results provide evidence for using h9e peptide hydrogel as a 3D cell culture. The hydrogel matrix is stiffened for cell encapsulation when it remains at 37°C, but is weakened at 4°C for cell isolation using standard centrifuge method. ## Cell Distribution and 3D Cells Culture in h9e Hydrogel The 3D cell culture matrices were prepared by directly adding peptide solution into MCF-7 cell suspension. The cells were encapsulated within the hydrogel scaffold during the peptide hydrogelation at 37°C. Cells were expected to distribute homogeneous and grow in 3D within the hydrogel architecture. To examine cell distribution, 5-day-incubated cell clusters in hydrogel were stained with a nucleic acid fluorescent dye (Hoechst 33342), encapsulated in h9e/MEM hydrogel, and observed using a LSCM with z-stack imaging. The homogeneous cells distribution was proved in both orthotropic and 3D view of LSCM z-stacks images, indicating that the peptide assembly into a nanofiber network is fast, and the gel is strong enough to suspend the cells. Cell clusters in round shapes were observed in the XY axis view of the orthotropic image , and multiple nuclei were presented in each cluster meaning that each cluster consists of many cells. This phenomenon was further confirmed in the inverted microscope image of MCF-7 cells in 3D hydrogel. The tumor-like clusters formation was observed, suggesting that a single seeded cell can form 3D cluster in hydrogel after multiple cell divisions. In contrast, the MCF-7 cells cultured in 2D environment attached to the bottom of the plastic plate and presented a flat shaped morphology. SEM images reveal the interaction of cells and the surrounding hydrogel matrix. is a protruding cell cluster from the nanofiber scaffold. Mutiple layers of nanofiber network attaching on the cell surface indicates that the 3D cell cluster was formed by multiple cell divisions of a single seeded cell. The expanding cells force the surrounding hydrogel matrix to collapse into layers of nanofiber network and attach to the cell surface. Such interface character of the cell surface and the surrounding hydrogel matrix suggests that MCF-7 cells in 3D culture were supported by the hydrogel matrix. The hydrogel matrix provides a scaffold environment on which cells could stably settle, but from which they also are able to extrude into three dimensions when they divide. ## Cell Isolation from the Hydrogel Matrix Isolation of cultured cells from the hydrogel matrix becomes a critical step for further cell physiological and pathophysiological studies. According to the rheological study above, hydrogel architecture can be destroyed under shear stress. Therefore, the hydrogel matrix can be converted to liquid by simply shearing procedure. As mentioned before, the fast reassembly property of this hydrogel is correlated to the peptide concentration. To assure enough time for cell isolation, the sheared hydrogel with 3D cultured cells need to be diluted and maintained at a low environmental temperature. In practice, a 3 mM peptide hydrogel was diluted 15 times with MEM and thoroughly mixed using a pipette, yielding a finial peptide concentration of 0.2 mM. The cells/h9e peptide/MEM mixture was converted to a liquid form. After 5 minutes of centrifugation at 4°C, the cells were clearly separated from the h9e peptide/MEM solution into a small pellet, which was observed at the bottom of the centrifuge tube. To confirm that hydrogel solution is in low-viscosity liquid form during the isolation, after dilution and pipetting, the G′ and G″ of this solution were determined by a 1 Hz single frequency rheological study with 1% shear strain. The experimental temperature was set at 4°C, the same as that used for cell centrifugation. indicates that after dilutions, the hydrogel was converted to Newtonian liquid showing that both G′ and G″ were around 0 Pa for the first 200 seconds. The G′ then increased to around 9 Pa within the first 10 minutes; however, it quickly dropped to 5 Pa and stayed at the same level as G″ for the remaining test. This result suggests that although the peptide has the tendency to reassemble, its solid strength is too weak to resist a slight disturbance, even shear stress from 1% strain. The diluted solution become stable as a low- viscosity liquid and would not reform as a hydrogel within the time domain for centrifuge isolation process. Besides, the shear stress encountered during centrifugation is much higher than 1% shear strain, which keeps the hydrogel solution in low-viscosity liquid form and makes cell isolation from this solution even easier. H9e peptide hydrogel provides a convenient method of isolating cells from the 3D hydrogel matrix. The next question is what percentage of cells in the hydrogel matrix could be collected from this method and whether they are viable after extraction. To address this question, 1.5×10⁵ cells were seeded in 3 mM h9e hydrogel and isolated using the above method after 30 minutes of incubation. Cell viability before encapsulation and after isolation was compared. No significant difference was found in viable cell number and cell viability before encapsulation and after isolation. These results suggest that cell isolation method for the h9e peptide hydrogel is safe and effective for biological studies. ## Comparisons of Cell Morphology, Viability, and Proliferation in 2D Monolayer and 3D h9e Hydrogel Cells were cultured in 2D monolayer (without h9e peptide) and in 1–3 mM 3D h9e hydrogel matrices for 1, 3, 5, and 7 days. Differences in cell morphology were first observed under 20X magnification of light microscopy. The morphological observation shows that, after the first 24-hour incubation, cells in the 2D plate attached to the the surface of plastic flask as shown in Day 1 of. Cells in the 3D architectures were floating in the gel matrix with a similar sphere shape in all three peptide hydrogel concentrations (Day 1 of). Over 3 days, cells in the 2D plate stretched as a monolayer; by day 5, without any attaching space on the plastic bottom, cells started to grow on top of the attaching layer and kept spreading and overlaying each other by day 7 (Day 3–7 of). In contrast, cells in the 3D matrix grew as rounded clusters over time. Compared with cells in 2 and 3 mM peptide hydrogel, which compacted more tightly and showed a cluster with multiple nuclei, some individual cells were localized on the edge of larger clusters of cells in 1 mM peptide hydrogel at days 3 and 5 (Day 3 and 5 of). This indicates that cells migrate easier in a loose nanofiber matrix, but are encapsulated tightly in a highly dense nanofiber network. In addition to the growth of the individual cluster, the neighboring clusters were merging into larger clusters, and the cells in all 3 concentrations of h9e peptide hydrogels became similar in morphology by day 7 (Day 7 of). With the cell isolation capability, the proliferation and viability of MCF-7 cells cultured in the 3D hydrogel matrix were determined. Cell viability was determined using Trypan blue excision method. Cells in the 2D culture proliferate rapidly by doubling in number every other day over 7 days of observation. Like many reported 3D cell cultures, cells in h9e hydrogel have a lower proliferation rate than those in the 2D culture. Cells in 1 mM hydrogel proliferate faster than those in 3 mM hydrogel by day 3, which may be due to the cell migration providing more space for cell division. By day 5, cells in both 1 and 2 mM hydrogel matrix had similar cell numbers and remained in a steady state by day 7. Cells in 3 mM peptide hydrogel proliferate even slower and remained in a relatively steady state from day 3. Cell proliferation rate of 2D and 3D cultures with various concentrations of h9e peptide has shown to be different. Interestingly, cell viability with various concentrations of h9e peptide in 2D or 3D cell culture stayed constant. ## Effects of Cisplatin, an Anti-cancer Drug, on MCF-7 Cells in 3D h9e Hydrogel The application of the h9e hydrogel as a 3D cell culture system for drug testing was examined using anti-cancer drug, cisplatin. To test drug diffusion in hydrogel, 5-day 3D cultures were treated with 40 µM cisplatin via three methods as described in : top surface, top-bottom transwell, and pre-mixed with medium and h9e peptide prior to hydrogelation. The results show that cisplatin can cause a decrease of cell viability through all three methods, indicating that all the methods can deliver cisplatin throughout the hydrogel matrix. At Day 3 after treatment, a significant 20% decrease of cell viability is observed in cells treated with 40 µM cisplatin via all the methods. However, at Day 10, 40 µM cisplatin can cause 40%, 60% and 85% decrease in cell viability via top surface, top-bottom transwell and pre-mixed with hydrogel method, respectively (Figure8A), suggesting that cisplatin is more effective when pre-mixed with hydrogel. All the results demonstrate that h9e peptide hydrogel permits the nutrients and drugs to diffuse throughout the matrix. shows cell responding time to cisplatin treatment in 3D hydrogel. Furthermore, cultured cells in hydrogel was tested with multiple doses of cisplatin. Cells were cultured in hydrogels for 5 days to form tumor-like clusters, and then treated with cisplatin via pre-mixed method as previously described. Cells were treated for 5 days and the IC₅₀ of cisplatin was determined. The results show that cisplatin decreases the viablity of cells in hydrogel in a dose-dependent manner. The IC₅₀ of cisplatin in MCF-7 cells under hydrogel condition is 31.25 µM, which is consistent with the literature for cisplatin, suggesting that hydrogel has the appropriate matrix for drug testing. The effects of cisplatin on cells in h9e hydrogel was further studied using confocal microscopy assay. Cells were cultured in 3 mM hydrogel for 5 days and then treated with 30 µM cisplatin for 48 hours via pre-mixed method. Seven-day cell cultures in hydrogel matrix without treatments were used as controls. After treatment, cells were isolated from the hydrogel and immunostained with antibodies against actin, Ki67, survivin, and cleaved caspase-3. DAPI was used to stain nuclei. Images of cells were captured with confocal fluorescence microscope (Carl Zeiss LSM 700 META). To obtain the 3D images, Z-stack images were taken (0.5–1 µm slices) and reconstituted in ZEN 2010 software. In this study, the morphology of 3D cell clusters in hydrogel matrix was first obseved by comparing to cells cultrued in 2D. Consistent with previous results , 7-day culture in h9e hydrogel allows MCF-7 cells to form tumor-like clusters with the diameter of \>50 µm. The cells in the hydrogel displays a rounded, clustered morphology distinct from the spread morphology of the cell in 2D monolayer. Immunofluorescence stain against actin showed to be localized in the cytoplatsmic region in both 2D and 3D cultures; however, unlike the elongated expression of actin in 2D cell culture, actin in 3D cell culture was expressed around the parimeter of the cell. Compared to the 3D clusters without cisplatin treatment, nuclear fragmentation was observed in all cisplatin-treated clusters. Actin was used as a biomarker in examing the effect of cisplatin on cytoskeleton. In the cisplatin-untreated clusters, actin was detected around the cell perimeter with an intact organization (-a). However, disruption of spheroids was observed in cisplatin-treated samples (-a). Ki67, a proliferation biomarker, was highly expressed in the control spheroids (-b), suggesting that cells cultured in hydrogel have proliferative ability. In the cisplatin-treated spheroids, Ki67 staining was only observed in cells with unchanged nuclei (-b). Furthermore, the expressions of survivin, an inhibitor of apoptosis, and cleaved caspase-3, an activator of apoptosis were selected to examine the effect of cisplatin on apoptosis of 3D cell cultures in hydrogel. The expression of survivin was only detectd around very fewintact nuclei (-c), and cleaved caspase-3 was highly expressed in the cisplatin-treated spheroids (-d), indiating that cisplatin can induce apoptosis in cells cultured in hydrogel. Overall, the confocal microscopy results provide evidence that cisplatin can diffuse into h9e hydrogel and exert its effects on cytoskeleton structure, proliferation, and apoptosis of MCF-7 cells in 3D hydrogel culture. The ability to recovery cells from 3D culture is an important features of the h9e hydrogel for cell culture and drug testing. Recoved cells from the 3D matrix are critical for further analysis. To examine this unique property of h9e hydrogel, Western blot analysis was performed. MCF-7 cells were cultured in 2D monolayer and 3D hydrogels for 7 days without any treatments ( untreated), or treated with 30 µM cisplatin for 48 hours after 5-day culture ( treated). For the 3D clusters in hydrogel, cisplatin was added via pre-mixed method. Cells were harvested from 2D culture or isolated from the hydrogels and whole cell extracts were obtained for Western blot analysis. The expressions of apoptotic markers (survivin, procaspase-3 and cleaved caspase-3) were detected. The results showed that in untreated cells, expressions of survivin and procaspase-3 were detected in both 3D and 2D culture; however, no cleaved caspase-3 was detected in either 2D or 3D cultures through Western blot analysis ( untreated). Because cleavage of caspase-3 is a hallmark of apoptosis, this result indicates that the h9e peptide hydrogel is an advanced material for *in vitro* 3D cell culture without induction of apoptosis. Consistent with previous confocal imaging results, all markers were dectected in 3D clusters treated with cisplatin ( treated). These results indicate that cells cultured and treated in h9e peptide hydrogel system can be extracted and analyzed for the downstream proteomic analysis. ## Conclusions In this study, a self-assembly peptide hydrogel system for 3D cell culture was examined. MCF-7 cells were encapsulated homogeneously during the hydrogelation simply by adding peptide solution to cell suspension of MEM. At physiological pH, the formation of hydrogel with controlled time and gel strength can be modulated by changing the temperature or peptide concentration. The specific shear thinning and rapid recovery rheological properties of hydrogel allow us to deliver this hydrogel material via pipetting multiple times. Unlike the attached flat morphology of cell in a 2D monolayer culture, cells residing in h9e peptide hydrogel adopt clustered structures that are reminiscent of tumor-like structure. More advantageously, the hydrogel architecture can be disrupted by pipetting mixing and further converted into liquid form through dilution with MEM medium. This property provides a convenient and safe method to isolate 3D cultured cells effectively from the hydrogel matrix by centrifugation. Compared with the 2D system, the cell proliferation rate is much slower in the 3D system and relative to different concentrations of h9e peptide. Cells cultured in 3D hydrogel culture have high viability and no significant apoptosis is induced by the 3D culture. The anti-cancer drug, cisplatin, can diffuse throughout the h9e hydrogel matrix and the h9e hydrogel allows testing of cisplatin in a time- and dose- dependent manner. Cells can be recovered from the 3D hydrogel system and used for the proteomic analysis, targeting specific pathways of drug-mediated response. Overall, h9e peptide hydrogel provides a promising 3D cell culture system for drug testing and other biomedical applications. # Supporting Information We thank Dr. Dan Boyle (Department of Biology, Kansas State University) for SEM images, Mr. Kent Hampton (Department of Entomology, Kansas State University) for SEM sample preparation, the University of Kansas Mass Spectrometry Lab for HPLC and mass spectrometric analysis, the Confocal Microfluorometry and Microscopy Core of NIH P20RR017686, and NIH 1R15CA152922. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HH YD XSS TAN. Performed the experiments: HH YD. Analyzed the data: HH YD XSS TAN. Contributed reagents/materials/analysis tools: XSS TAN. Wrote the paper: HH YD XSS TAN.

# Introduction An important number of studies have shown a correlation between increased cardiovascular diseases with testosterone deficiency. It has been demonstrated that sex hormones regulate vascular function since sex hormones deprivation alters the release, function and cell signaling pathways of endothelial factors that could lead to vascular dysfunction. Thus, previous studies from our research group have shown that orchidectomy increased ROS production and prostanoids release, while the production of nitric oxide and the antioxidant capacity were reduced. The increased oxidative stress has been associated to the development of cardiovascular diseases. Thus, oxidative damage of cellular membranes and enzymes by reactive oxygen and nitrogen species (ROS/RNS) has been described in cardiovascular diseases. Oxidative modification of cholesterol from cell membranes by ROS/RNS leads to the formation of cholesterol oxidation products (COPs), also called oxysterols. COPs can also be formed by enzymatic oxidation and/or by absorption from the diet. Regardless of the source, COPs have the ability to induce disruption of fine structure, alteration of integrity, fluidity, and permeability, and loss of biomembrane functionality. Furthermore, oxysterols have been implicated in the pathogenesis of various diseases including cardiovascular diseases, cancer, neurological disorders, and aging. COPs are able to modify low-density lipoproteins (LDL) and high-density lipoproteins (HDL) into pro-atherogenic and pro-inflammatory forms. Oxysterols can also trigger pro-oxidative, pro-inflammatory and cytotoxic reactions in the different cell types of the vascular wall, including endothelial cells, smooth muscle cells, fibroblasts, monocytes and macrophages. Derived from such evidence, COPs have been proposed as potential biomarkers for non-invasive studies of oxidative stress *in vivo*. Different studies have demonstrated that consumption of Omega-3 polyunsaturated fatty acids (PUFAs) such as eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) may reduce the risk of developing cardiovascular diseases. The anti- thrombotic, anti-inflammatory and vasoprotector effects of PUFAs on the cardiovascular system have been reported. Antioxidant propierties of DHA have been also demonstrated since it decreased ROS formation and bound to free radicals preventing tissue oxidation. Since it was reported that the lack of sex hormones increases oxidative stress and vascular inflammation, deteriorating factors that could lead to the development of CVD, it is possible that formation of oxysterols in the arterial wall could be modified by the loss of gonadal function. On the other hand, Omega-3 PUFAs may act as a cardioprotective treatment in the oxysterol formation induced by orchidectomy. Therefore, the purpose of the present study was to determine the effects of orchidectomy on the formation of the main oxysterols recognized as products of cholesterol autoxidation (7α-hydroxycholesterol, 7β-hydroxycholesterol, 7-ketocholesterol, 5,6α-epoxycholesterol, 5,6β-epoxycholesterol, cholestanetriol and 25-hydroxycholesterol) in the arterial wall of the aorta and mesenteric artery, as well as the possible preventive effect of a DHA-supplemented diet on COPs formation. # Materials and methods ## Animals, diets and experimental groups Male Sprague-Dawley rats (18-week-old) were purchased from Envigo (Mexico City). Rats were housed in stainless steel cages in a temperature-controlled (23 ± 2°C) room, under 12-hour light/dark cycles and standard feeding with fodder and water *ad libitum*. After 4 weeks of adaptation, animals were fed a maintenance diet for rodents (2018S Teklad global 18% protein rodent diets, Envigo, Madison, WI, USA) supplemented with fat (5%). The controls-diet groups were supplemented with sunflower oil (5%) and the DHA groups with 4.7% MEG-3^® (Ocean Nutrition Canada Ltd., Dartmouth, NS, CA) and adjusted to 5% with sunflower oil. Nutrient content and energy distribution of each diet is summarized in. After 2 weeks on the diet, animals were divided into two groups: control and orchidectomized males. Male sex hormone deprivation was induced by orchidectomy at 24 weeks-old under injectable anesthesia with Ketamine-Xylazine (80 mg/kg ket plus 10 mg/kg xil; IP). Rats were treated with 2 mg/kg meloxicam SC (Metacam 5 mg/mL; Boehringer Ingelheim) and with 50 mg/kg ibuprofen immediately after surgery, via IP administration for 5 days. The observation of seminal vesicles atrophy confirmed successful surgery. Animals were maintained under the experimental diets for six more weeks. At the end of the treatment, rats were sacrificed by ether inhalation and decapitation. The mesenteric artery and the aorta were carefully dissected out, cleaned of connective tissue and placed in Krebs-Henseleit solution (KHS) containing, in mmol/L: NaCl 115, CaCl₂ 2.5, KCl 4.6, KH₂PO₄ 1.2, MgSO₄ 1.2, NaHCO₃ 25, glucose 11.1, Na₂ EDTA 0.03 at 4°C. The investigation was performed in compliance with the *Guide for the Care and Use of Laboratory Animals* published by the USA National Institutes of Health (NIH publication No. 85.23 revised 1985), and approved by the Ethical Committee of the *Universidad Autónoma de Madrid* according to directives of the Ministerio de Agricultura, Pesca y Alimentación of Spain (PROEX 202/16). ## Determination of cholesterol and COPs in arterial tissue samples ### Lipids extraction Two aortic and two mesenteric rings (*ca*. 3 mm length) from each animal (from the all experimental group of rats) were individually weighed, thawed and subjected to total lipid extraction, as described by Folch *et al*. with some modification. Briefly, arterial rings were homogenized in 1 mL of PBS at 4°C containing BHT (0.05%). Each arterial homogenate was mixed with 16 mL of the Folch solution (chloroform-methanol 2:1 v/v) and 0.01% BHT. Then, 3 mL of NaCl aqueous solution (0.73%) were added. The resulting mixture was separated by centrifugation. The upper phase was discarded and the lipids were collected from the organic layer (lower phase). A mixture of chloroform-methanol-0.73% NaCl (3:48:47 v/v/v, 7.5 mL) was added to the recovered phase and then left at 4°C for 4 h to allow separation. The lower phase was recovered and then filtered through anhydrous sodium sulfate. Subsequently, the extract was resuspended in chloroform with BHT (0.02%) and frozen at -20°C until used. ### Saponification Cold saponification was performed according to the method of Menéndez-Carreño *et al*. to remove glycerides, free fatty acids and water-soluble impurities, as well as to release the esterified COPs. For sample preparation, each lipidic extract from the arterial tissue samples was added with 13 μg of the internal standard, 19-hydroxycholesterol (19-OH) (disolved in *n*-hexane:isopropanol, 4:1 v/v) for COPs quantification and with 1 mg of 5α-colestane (disolved in *n*-hexane: isopropanol, 4:1 v/v) as an internal standard for the quantification of cholesterol. The mixture was then dried under stream of nitrogen and mixed with 3 mL of a KOH solution in methanol (4 mol/L or 4 N) with BHT (0.05%), wrapped with aluminium foil and subjected to orbital shaking (300 rpm) at room temperature for 18 h to produce saponification. For extraction of the unsaponifiable matter, 10 mL of dichloromethane and 5 mL of citric acid (0.1% in double-distilled water) were added to each sample, and mixed vigorously in a vortex. The diethyl ether fraction was then separated by centrifugation. The aqueous layer (supernatant) was discarded and the organic phase was washed with portions of 5 mL citric acid (0.1%, v/v) solution until clear. The organic phase from the samples was dried over anhydrous sodium sulfate. The organic solvent was evaporated with a rotary evaporator at 50°C to remove the dichloromethane. The unsaponifiable extract was dissolved in diethyl ether in a conical vial and dried under nitrogen flow for the subsequent quantification of cholesterol and COPs. ### Analysis of total cholesterol by gas chromatography For cholesterol determination, 10% of the unsaponificable matter solution was dried under nitrogen flow and subjected to silylation according to Sweeley *et al*.. A pyridine:hexamethyldisilazane:trimethylchlorosilane (5:2:1, v/v/v) mixture was added (0.3 mL) and heated at 40°C for 20 min. The mixture was then dried under a stream of nitrogen and dissolved in 1 mL of *n*-hexane. Analyses were performed on an Agilent 7820A gas chromatograph. 1 μL of sample was manually injected in split mode (1:1). Separation of the compounds was performed on a Perkin Elmer PE-5 capillary column (30 m x 0.32 mm d.i. x 1 μm film thickness coated with 5% -phenyl-methylpolysiloxane). The oven temperature was programmed from 280°C to 300°C at a rate of 10°C/min, and mantained for 30 min. The injection and detection port temperatures were set at 325°C. UHP nitrogen was used as the carrier gas at a rate of 1.5 mL/min. Total cholesterol from the arterial tissue lipidic extracts was quantified by the internal standard method, using 5α-cholestane. Identification of the cholesterol peak in the samples was carried out comparing the retention times with that from the standard. Quantification of cholesterol was performed using a calibration curve. ### Purification and derivatization of COPs The remaining 90% of the unsaponificable matter were dried under nitrogen stream and re-suspended in 300 μL of *n*-hexane:ethyl acetate (95:5, v/v) and purified by NH₂ SPE cartridges, previously equilibrated with 3 mL *n*-hexane deied with anhydrous sodium sulfate at the bottom. The cartridge was eluted with the following solvent sequence: 6 mL of *n*-hexane:ethyl acetate (95:5, v/v), 10 mL of *n*-hexane:ethyl acetate (90:10, v/v) and 10 mL of acetone. The COPs were eluted from the NH₂ SPE cartridge with the acetone fraction according to the method of Rose-Sallin *et al*.. The purified COPs were dried under nitrogen flow and silylated, according to the method of Sweeley *et al*., and dissolved in 50 μL of *n*-hexane. ### Analysis of COPs by gas chromatography One microliter of the silylated COPs was manually injected into the GC under the same conditions used for the determination of total cholesterol. Total COPs were quantified using 19-hydroxycholesterol (19-OH) as internal standard. The COPs identification in the samples was performed by comparison of the retention times with those of the COPs standards and the quantification was performed by calibration curves, as it was done for cholesterol. ## Reagents All analytical grade solvents and reagents were supplied by Tecsiquim (Mexico City) and from Sigma-Aldrich (Mexico City). For identification and quantification of each COP the internal standards utilized were 19-hydroxycholesterol (internal standard for the quantification of COPs), 5*α*-cholestane (internal quantification standard for cholesterol), Cholest-5-en-3*β*-ol (cholesterol), cholest-5-en-3β-ol-7-one (7-ketocholesterol), 5α,6α-epoxycholestane-3β-ol (5,6α-epoxycholesterol), 5β,6β-epoxycholestane-3β-ol (5,6β-epoxycholesterol), cholestane-3β,5α,6β-triol (cholestanetriol), cholest-5-en-3β,7β-diol (7β-hydroxycholesterol), and cholest-5-en-3β,25-diol (25-hydroxycholesterol). Cholest-5-en-3*α*,7*α*-diol (7*α*-hydroxycholesterol) standard was supplied by Steraloids (Newport, CT, USA). Solid-phase extraction (SPE-NH₂) cartridges (500 mg amino-propyl stationary phase/3 mL) were purchased from Phenomenex (Grace Discovery Sciences, Deerfield, IL, USA). The silylation mixture was prepared with dried pyridine, hexamethyldisilazane and trimethylchlorosilane, from Sigma. ## Statistical analysis All values are expressed as the mean ± SEM (Standard Error of the Mean) of the five animals used. For each value, the mean of the cholesterol and COPs quantification -obtained in each of the two arteries from the same animal- was calculated. Statistical analysis of COPs concentration was performed by one-way analysis of variance (ANOVA) followed by Tukey’s multiple comparison test. For body weight, statistical analysis was done using paired Student’s *t*-test. A *p* value of less than 0.05 was considered significant. Data were analyzed using the GraphPad Software (San diego, CA, USA). # Results ## Animal body weight Animals remained healthy and behaved similarly well throughout the experiment. Before the experimental diets (control or DHA) were administered, body weight was evaluated in the four groups of animals, showing no statistical differences among the groups. The average weight gain at the end of 8 weeks under the experimental diets was similar for all the experimental groups, which is consistent with our previously published data. ## Cholesterol content in aorta Orchidectomy significantly increased cholesterol levels in the aortic tissue compared to the control group fed with the control diet. The DHA-supplemented diet caused a significant decrease in cholesterol levels in the aorta from orchidectomized animals. Samples from control rats fed with the DHA-supplemented diet showed similar levels to those found in the control rats fed with the control diet. ## COPs content in aorta Total COPs levels significantly increased in the orchidectomized group, and decreased in orchidectomized animals fed the DHA-supplemented diet. DHA supplementation did not stastically modify total COPs content in control rats. Consistent with the above trend, orchidectomy caused a significant increase in the concentration of the studied COPs in aortic segments: 7α-hydroxycholesterol (7α-OH), 7β-hydroxycholesterol (7β-OH), 7-ketocholesterol (7-KC), 5,6β-epoxycholesterol (5,6β-E), 5,6α-epoxycholesterol (5,6α-E), 25- hydroxycholesterol (25-OH), and cholestanetriol (CT). The DHA-supplemented diet decreased COPs levels in the aortic tissue of orchidectomized rats, except 7β-OH which did not reach stastistical difference. However, the DHA-diet to control animals tended to increase the concentration of all analyzed COPs. The proportion in the concentrations of the analyzed COPs in relation to the total COPs was similar in the different groups included in this study. CT was the most abundant COP (*ca*. 30% from total COPs), followed by 7β-OH (*ca*. 20% from total COPs), 7α-OH (*ca*. 18%), 5,6β-E (*ca*. 12%). 7-KC, was higher in both orchidectomized groups (fed with control and DHA-suplemented diet) the ratio remained above the control groups by *ca*. 8–10%. The proportion of 5,6α-E was similar in all groups, being ca. 5% of total COPs. 25-OH could not be detected in the control rats fed the control diet and for the other groups this COP was detected in very small concentrations (≤1%). ## Cholesterol content in mesenteric artery The content of cholesterol in mesenteric artery was significantly increased by orchidectomy and the DHA-supplemented diet significantly decreased those levels. The DHA-diet to control rats did not statiscally modify the content of cholesterol in mesenteric artery tissue. ## COPs content in mesenteric artery Similar results were observed for COPs content in mesenteric artery respect to those found in aorta from these animals. Total COPs content increased significantly in samples from orchidectomized animals, compared to the rest of the groups. The control animals fed with the control diet as well as those fed with the DHA-supplemented diet showed similar COPs levels. Again, DHA supplementation decreased COPs content in orchidectomized animals. Orchidectomy significantly increased 7α-OH, 7β-OH, 7-keto, 5,6β-E, 5,6α-E, 25-OH, and CT content compared to the rest of the groups included in the study. As observed in, 25-OH content from control rats fed with the control diet was not detected. The same ocurred with 7-KC, which was only detected in minimal amounts in the orchidectomized group, whereas it was practically absent from the mesenteric artery tissue in the other groups. In the mesenteric arteries from the orchidectomized animals fed with the DHA supplemented diet the concentration of COPs decreased significantly, except for 7β-OH, similarly to that found in aortic tissue. DHA supplementation in control animals tended to decrease the concentration of all the analyzed COPs (7α-OH, 7β-OH, 7-KC, 5,6β-E, 5,6α-E, 25-OH, and CT), unlike the observed in aorta. Regarding the distribution of COPs content in mesenteric artery tissue, CT was the most abundant COP (*ca*. 30%) followed closely by the 7α-OH and 7β-OH, which showed similar proportions between the same groups, arround 20% of total COPs. Next in proportion is the 5,6β-E representing *ca*. 15% of total COPs, with a slightly higher proportion in the control group (19%). The 5,6α-E represents about 10% of the total COPs, dropping to 6% in the orchidectomized animals fed the DHA-supplemented diet. 25-OH was practically absent, since it could not be detected. In the case of 7-KC, it was only detected in the orchidectomized group at very low concentrations. # Discussion This study shows, for the first time, that the loss of gonadal function induced an increase in the formation of COPs in both aorta and mesenteric artery wall. The preventive action of a DHA-supplemented diet on the orchidectomy-induced COPs formation is also demonstrated. These results are in agreement with data previously reported on both the effects of orchidectomy and the beneficial role of a DHA-supplemented diet. Regarding the results specifically associated to orchidectomy, we have previously reported that the loss of gonadal function of rats induced an overproduction of ROS in aorta and mesenteric artery, as well as an increase in the serum content of cholesterol, LDL-cholesterol and triglycerides. It is known that the cholestesterol embeded on the lipid bilayer from the cell membranes is prone to oxidation by ROS which could explain the increased formation of COPs derived from autoxidation processes described in this work. However, oxidation during sample processing can not be ruled out, despite of the utilization of BHT (0.05%). Likewise, it is important to consider that conversion to COPs other than those analyzed may exist. It is well known that the preferential site of oxidation of cholesterol by highly reactive species is at C7 having a relatively weak carbon-hydrogen bond. Moreover, the unique cholesterol double bond between carbons 5 and 6 comprises a vulnerable site for oxidation by free radicals. The most abundant non-enzymatic cholesterol oxidation products present in most tissues are 7α-OH, 7β-OH, 7-KC, and 5,6α-E and 5,6β-E, respectively induced by ROS and RNS, which is in line with the findings from the current study, except for 7-KC, which was only detected in mesenteric arteries from orchidectomized rats. CT is one of the most abundant oxysterols, derived from cholesterol by oxidation via formation of 5,6α-E and 5,6β-E as intermediates. The very low detection leves of 25-OH may be attributed to its primarily enzymatic origin, since this oxysterol is generally not considered to be a significant autoxidation product of cholesterol. In addition, the reported levels of 25-OH in most tissues are extremely low. Determination of COPs content can vary depending on the extraction and/or measurement methods, although very low COPs levels under physiological conditions have been reported. In this sense, the 7-KC, 7α-OH and 5,6α-E content -referred to total cholesterol- reported in the present study were similar to those described in human aorta and coronary arteries. However, the relative7-KC, 7α-OH, 7β-OH, 5,6α-E and 5,6 β-E content found in our study are higher than those described in human aorta and carotid artery. On the other hand, COPs are known to trigger oxidative stress by increasing the generation of superoxide anion, and down-regulating the expression/activation of Nrf2 Consistent with these results, many studies have reported increased COPs levels in the membranes of cells subjected to oxidative stress, as it has been detected in patients with diabetes mellitus, hiperlipidemia, chronic inflammatory processes and chronic renal failure. In this regard, the oxysterol clearance mechanisms are probably less efficient in those situations, as it was observed in orchidectomized rats in wich the antioxidant activity was decreased. Zhang *et al*. demonstrated that castration decreased enzimatic antioxidant activity (SOD and glutathione peroxidase) and increased malondialdehyde levels, an indicator of lipid peroxidation. In contrast, other studies found that castration did not increase lipid peroxidation. These discrepancies could be attributed to differences in the animal model used, since the maintenance period of gonadectomy determines the induced alterations. However, it is important to emphasize that the experimental model used in this study is the same as that used in our two previous studies and provides relevant information about the key role of the COPs in the orchidectomy-induced modifications of factors, additional to ROS, which regulate vascular function. It has been mentioned above that different cardiovascular risk factors, such as hypertriglyceridemia, hypertension, diabetes, obesity and overweight, have been associated to COPs content in human serum. In agreement with these results, COPs induced increased blood pressure, serum triacylgycerols as well as body fat index in Wistar rats. On the other hand, the induction of hypertension to rabbits by coarctation of the aorta showed an enhancement of COPs content in both plasma and aortic tissue. In addition, Ares *et al*. indicated that COPs promoted the stimulation of MAPK in human aortic smooth muscle cells, which could partially explain the activation of the MAPK sinaling pathway reported in hypertension after tyrosin kinase receptor transactivation. In this regard, our group has described the activation of MAPK and AKt pathways in mesenteric arteries of orchidedctomized rats, in which the transactivation of EGFR was involved. In this context, these results would indicate that the increase in the formation of COPs may participate in the orchidectomy-induced structural alterations that, in the long term, can lead to the development of hypertension, as it occurs in aged-orchidectomized rats. Apart from these effects, COPs alter prostaglandin synthesis and stimulate platelet aggregation, an important process facilitating atherosclerosis and thrombosis. Likewise, different COPs induced the expression of COX-2 and the release of prostaglandin E₂ and prostaglandin J₂, considered as pro-inflammatory events. Accordingly, it has been described that orchidectomy also leads to a pro-inflammatory environment by increasing prostanoid production in rat aorta and mesenteric arteries most likely as a consequence of increased oxidative stress produced during the experimental conditions. Nitric oxide is also a critical molecule in vascular function, and the inhibition of endothelial NO-synthase by COPs has been described, which is also in agreement with the decreased NO formation observed in aorta and mesenteric artery from orchidectomized rats. Although the decrease of NO production by COPs seems to be demonstrated, studies addressing the influence of COPs in the vascular function have shown contradictory results, since inhibition or no modification of endothelial-dependent vasodilation by 7-KC and 7α-OH-cholesterol have been reported. In our experimental model, orchidectomy did not alter the acetylcholine-induced response in aorta or mesenteric artery, since factor/mechanisms other than NO are simultaneously working to maintain a proper function. These results can be summarized in that orchidectomy induced an increase of the oxidative stress and anti-inflammatory mediators, and a deterioration of the lipid profile, which is accompanied by an increase in the formation of COPs. Interestingly, cholestanetriol -the most abundant COP in all samples, especially in arteries from orchidectomized rats- has been reported to be one of the most cytotoxic oxysterols in different cell lines such as rabbit aortic smooth muscle cells, mouse L cells, Chinese hamster V79 lung fibroblasts, human monocytic cells (U937), colonic adenocarcinoma cells (CaCo-2) and hepatoma liver cells (HepG2). However, it is important to point out the cytotoxic effects reported for different COPs other than cholestanetriol, also in different cell lines, related to vascular and nervous systems that contribute to the pathogenesis of cardiovascular and neurodegenerative diseases. It has been suggested that generation of some oxysterols can be reduced in the presence of antioxidants by scavenging ROS. For instance, supplementation of vitamin E to diabetic patients can decrease 7-KC and 7β-OH levels. Likewise, Uemura *et al*. found that apoptosis induced by 7β-OH and 7-KC in vascular endothelial cells was prevented by α-tocopherol. DHA exhibits potent anti- oxidant properties which attenuate ROS overproduction in endotelial cells and enhances the overall antioxidant status. Recently, the ability of DHA to prevent ROS overproduction and oxiapoptophagy induced by 7KC, 7β-OH, and 24-OH in oligodendrocytes was demonstrated. De Medina *et al*. found that DHA inhibited cholesterol-5,6-epoxide hydrolase activity, that catalyzes the conversion of 5,6α-E and 5,6β-E to its product CT, which exerts important citotoxic effects. Also, DHA showed protective effects on COPs-induced cell death. These results agree with our findings, since the DHA-supplemented diet to orchidectomized rats prevented the rise in COPs formation caused by orchidemtomy. Thus, the levels of COPs were near to those measured in control rats fed the control diet. Interestengly, the DHA-supplemented diet restored the increased contents of COPs induced by orchidectomy, similarly how it restores the lipid profile, and the redox and inflammatory status altered in this particular condition. These results are in agreement with the cardioprotective effects of DHA widely reported through its anti-inflammatory and anti-oxidant activities. It has been recently reported that DHA, one of the main fatty acids of the Mediterranean diet, attenuates the 7-KC-induced toxic effect on microgial cells. Based on all this information, human dietary habits deserve important consideration and comsumption of a Mediterranean diet could prevent the consequences of an increased oxidative stress that occurs in diverse physiopathological situations. However, something different seems to occur with the DHA-supplemented diet to the control animals, in which a tendency to increase COPs content was observed in the rat aorta. This result is in line with a previous report, in which the involvement of vasodilator prostanoids in the Ach-induced response in aorta from control rats was switched to vasoconstrictors after the DHA-suppelemented diet. Overall, these results indicate that DHA-supplementation exerts cardiopotective effects, especially in pathological conditions, which seems to be reasonable since the homeostatic mechanisms work to maintain proper function in healthy subjects without any nutritional supplement. In summary, we describe here the detrimental effects caused by the lack of sex hormones in the rise in lipid peroxidation, providing new information on the damage on lipid components responsible of maintaining membrane properties and cell signaling in different pathways involved in the vascular function of aorta and mesenteric artery, already published. The preventive effect of a DHA- supplemented diet on lipid peroxidation, which promotes the maintenance of vascular homestasis, is also shown. In this regard, it is important to note that the results observed in the mesenteric artery, together with those related to vasomotor function, are of special relevance since this vascular bed importantly contributes to the control of the peripheral vascular resistance and therefore to the regulation of blood pressure. The present findings reveal the interest in conducting clinical studies with DHA in patients with decreased levels of sex hormones (elderly patients and/or prostate cancer patients undergoing androgen deprivation therapy) who display cardiovascular risk factors. The authors thank Daniela Peña for the advice on experimental protocol and use of equipment and to Laura N. Bober, for the performance of surgeries. This study was supported by fellowships (to DMV and to MMR) from the CONACyT and grants (to MF) from Comunidad de Madrid (S2013/ABI 2783, “INSPIRA1-CM”) and Fondo Europeo de Desarrollo Regional. [^1]: The authors have declared that no competing interests exist.

# Introduction This study has three aims, establishing a more ecologically valid paradigm, controlling low-level adaptation, and comparing automatic change detection in older and younger participants. Regarding the first aim, in event-related potential (ERP) research, including research on visual mismatch negativity (vMMN), the most frequent practice of stimulus delivery is the presentation of stimuli alternating with empty inter-stimulus fields. A few exceptions in the vMMN literature are the studies by Besle et al., 2005 and Clery et al., 2012, where the infrequent and frequent events were different deformations of circles to ellipses. Other studies with non-empty fields investigated visual masking effects with zero or short stimulus onset asynchronies (SOA), but even in these studies there were blank intervals between the trials. In further ERP research two frequently used stimulation procedures also apply the continuous presence of stimuli. The first is contrast-reversal stimulation, which is useful in many fields, but only infrequently applied in studying information processing issues. The second common stimulus type contains moving patterns. In this case, the pattern remains visible but static during the inter-stimulus interval and the (standard or deviant) event is the abrupt start (of motion) of the pattern. However, in everyday conditions the appearance of visual events preceded and followed by empty fields is unusual. (A rare example is a lightning on a cloudy night.) In a more usual scenario, objects appear or disappear among other objects, or objects are continuously present, but some characteristics of the objects change (they can move, rotate, be occluded by other objects, etc.). The visual system develops over the lifespan through continuous interaction with the real (and not laboratory) world. Object representations are the main building blocks of the visual percept; and the identification of the same object under various conditions (e.g., temporarily occluded object) is an adaptive response of the visual system to the natural environment. Similarly, the mechanism underlying vMMN (see for a review) is a crucial part of the same visual system that results in object representations. Therefore, the theoretical connection between vMMN and object identification is a feasible option. Empirical evidence supports this assumption: vMMN’s sensitivity to object identification and categorization (for a review, see) is frequently reported in the vMMN literature; however in those studies the stimulus presentation is carried out using the traditional ERP research paradigm (appearance and disappearance of the whole object). In the present study our aim was to take a step towards a naturalistic scenario by simulating the occlusion of objects. Objects were continuously present on the screen, but from time to time a part of the object disappeared. The second aim was to set up a method for controlling low-level adaptation differences between the frequent and infrequent stimuli of the oddball sequences. In studies of vMMN unattended rare (deviant) stimuli are presented within sequences of unattended frequent (standard) ones, i.e., in passive oddball sequences. The deviant-related activity is considered to be an index of mismatch between an established (sensory) memory representation for the standard and the neural representation of an event different from the standard. Accordingly, vMMN is considered to be an additional activity elicited by the deviant, but not by the diminishing activity elicited by the frequently repeated standard (for reviews see). However, an unavoidable problem of the oddball sequence is that the ERP *difference* between the deviant-related and standard- related activity includes the putative activity increase to the deviant *plus* the activity decrease to the standard. The latter effect is termed as refractoriness, habituation, repetition suppression, or stimulus-specific adaptation (for a discussion on terminology see; we will use the term ‘stimulus- specific adaptation’ (SSA)). To separate the effects on the standard from the deviant-related effects, the most frequent method is the introduction of the equal probability control. In this procedure, ERPs to stimuli physically identical to the oddball deviant are presented within a sequence of various stimuli. Within the sequence the probability of each stimulus type is equal to the probability of the oddball deviant. Accordingly, in the absence of the standard the stimuli of the control sequence cannot elicit mismatch-related activity. The ERP difference between the oddball deviant and the physically identical control from the equal probability sequence is called “genuine MMN”. As an inherent problem of this procedure, the to-be-compared ERPs are selected from highly different sequences. Furthermore, in many cases it is difficult to vary properly the stimuli of the control sequence (e.g., in case of female vs. male photographs,; feature absence vs. feature presence in vMMN studies,). Our paradigm attempts to solve the low-level adaptation issue using the opposite logic. Rather than using an unadapted comparison from an equal probability sequence, during the ‘inter-stimulus period’ the objects were continuously present; therefore all stimulus features were adapted. Consequently, larger activity elicited by a rarely vanishing (disappearing) feature could not be attributed to low-level SSA. It is important to note that deviant-related activity is superimposed on *offset* responses. Offset and onset ERPs are different from each other, but offset responses are relatively infrequently investigated. Furthermore, the offset of the vanishing feature is followed by feature onset. Accordingly, the rarely vanishing feature is followed by the *rare appearance* of the missing feature. Consequently, in this case re- appearance of the stimuli may elicit onset-related vMMN. Offset stimuli are less salient than rapid onset stimuli. In a paradigm investigating the effects of unattended stimuli (cf. vMMN paradigm) the use of less salient events is especially insightful. In our study, we made use of disappearing sides of diamonds as offset stimuli; one pair of opposite sides was the standard, and the other pair was the deviant stimulus. As the third aim, in this study we investigated the automatic processing of rare stimuli in younger and older participants. In contrast to the auditory MMN (see for a review), investigations of age-related changes of vMMN are relatively rare, and the results are equivocal. In older participants deviant motion direction elicited smaller motion-related vMMN, but the sample size of the study was small (7 younger, 5 middle-aged and 9 older participants). Tales et al. presented task-irrelevant single or double bars (either as frequent or infrequent stimuli) within the sequence of a three stimulus (standard, deviant and target) oddball paradigm. Deviant-related negativity was smaller again in the older group; however, in a more recent study applying a similar method, Stothart et al. did not confirm this result. Recently we investigated age- related difference in temporal integration in a vMMN paradigm. In the control condition (without the requirement of integrating two consecutive stimuli) we obtained no age-related vMMN differences. Thus, two studies so far have reported age-related differences, but no such difference appeared in the other two studies. # Materials and methods 15 older (mean age: 66.4 years, SD: 3.1 years) and 15 younger women (mean age: 22.4 years, SD: 1.6 year) participated in the study. For all participants lower field checkerboard stimulation resulted in a posterior negativity within the 100–180 ms range (C2 or N1 component). All of them had (corrected to) normal vision (at least 5/5 in version of the Snellen charts). No one reported any neurological or psychiatric diseases. Both the older and younger participants were paid for their contributions. We ruled out dementia-related differences between the age-groups; full scale Wechsler IQ (measured by the Hungarian version of WAIS-IV;) of the older and younger group were 118.7 (SD: 15.3) and 97.6 (SD: 15.2), respectively. Written informed consent was obtained from the participants before the experimental procedure. The study was approved by the United Ethical Review Committee for Research in Psychology (Hungary). The experimental stimuli were presented to the participants using a 19 in. CRT monitor (Flatron 920B, 75 Hz refresh rate) from 1.4 m distance. The stimulation included both task-relevant and task-irrelevant stimuli. demonstrates both types of stimuli and the stimulus sequence. The task-relevant stimuli were two disks at the central area of the screen. One of the disks was red and served as a fixation point (0.11 degrees in visual angle). The second disk was green (0.22 degrees) and made horizontal pseudorandom motion around the fixation point. The task was to keep the green disk as close to the fixation point as possible with the left and right arrows of a keyboard. The error in the task was when the distance of the two disks exceeded 1.1 degrees (in case of an error, the color of the disk changed blue to provide online visual feedback). The participants’ performance (i.e. the sum of errors in one block) was reported on the screen at the end of each block. The task-irrelevant stimuli appeared around the task-relevant stimuli. The stimuli were homogenous patterns of six identical objects (in a 2 rows by 3 columns arrangement; luminance: 101.63 cd/m²) against middle grey background (luminance: 36.73 cd/m²). There were three types of stimulus patterns. In the first one, the objects were oblique squares (similar to a diamond) with diagonals. The diameter of one diamond was 1.65 degrees; the width of the lines was 0.055 degrees. Objects of the second and third patterns were similar to the first one except that the two opposite, parallel sides (45 or 135 degree) of the square were absent (together with the diagonals, the remaining objects were similar to oblique bow ties). The stimulus sequences were as follows. Every odd stimulus was a “diamond” pattern and every even stimulus was a “bow tie” pattern. It is important to mention that the screen was never blank. Bow ties and diamonds alternated with each other without inter-stimulus intervals. The two bow tie patterns (left and right tilted; the disappearing stimulus features were the 135-degree and the 45-degree orientated line segments, respectively) were presented in an oddball sequence. The probability of the deviant was 0.2; that is, the ratio of the two bow tie patterns was 1:4 in a sequence. The stimulus sequence resulted in four different events: rare disappearance, frequent disappearance, rare appearance and frequent appearance of stimulus feature. In a block there were 95 diamonds, 76 left tilted bow ties, and 19 right tilted bow ties. Based on the reverse control principle, we introduced an additional sequence, where the number of bow ties were interchanged (i.e. both bow tie patterns were either standard or deviant in separate sequences). There were 6 oddball and 6 reverse blocks, which resulted in 570 stimuli in each condition. The number of deviant stimuli was 114 per condition. The stimulus duration of all three patterns was 520 ms (with +/- 40 ms jitter in 13.3 ms steps). EEG was recorded with Neuroscan recording system (Synamps2 amplifier, EasyCap, Ag/AgCl electrodes, DC-200 HZ, sampling rate: 1000 Hz). We used 38 electrode locations in accordance with the extended 10–20 system. The ground electrode was attached to the forehead. The common reference electrode was on the nose tip. Both HEOG and VEOG were recorded with bipolar configuration between two electrodes (placed lateral to the outer canthi of the two eyes or above and below the left eye, respectively). The EEG signal was bandpass filtered offline with a non-causal Kaiser-windowed Finite Impulse Response filter (low pass filter parameters: 30 Hz of cutoff frequency, beta of 12.2653, a transition bandwidth of 10 Hz; high pass filter parameters: 0.1 Hz of cut off frequency, beta of 5.6533, a transition bandwidth of 0.2 Hz). Epochs of 500 ms (including 100 ms pre-stimulus interval serving as baseline) were extracted for all deviants and for those standards that immediately preceded the deviants. Epochs with larger than 100 μV voltage change were considered artifacts and rejected from the further processing. ERPs were calculated by averaging the extracted (and residual) epochs. According to the reverse control principle, epochs from both experimental (oddball and reverse) sequences were entered into the averaging process. Behavioral data were defined as the total time that the ball left the target area. Age-related differences were compared using a Mann-Whitney U test. VMMN was expected over the posterior locations within the 100–300 ms latency range. Accordingly, at the first step we calculated t-tests over this range at O1, Oz and O2 electrodes for the deviant *minus* standard difference potentials (difference from zero). Emergence of vMMN was considered if at least 10 consecutive t-values were significant (p\<0.05 or less) at least over two locations. In the second step we divided the range of the \~ 100 ms duration of the deviant-related negativity into five epochs of 20 ms each, and the average of these epochs were compared in ANOVA with factors of Age (older, younger) and Location (O1, Oz, O2). When appropriate, Greenhouse-Geisser calculation was applied. Effect size was calculated as partial eta square (η_p²). As post-hoc tests Tukey HSD was applied. # Results Tracking performance was worse in the older group compared to the younger adults, the group averages were 13.4 s and 2.9 s, respectively. According to the Mann-Whitney test, the difference was significant, U = 8, p\<0.001. As shows, deviant vanishing stimuli elicited a posterior negative wave, the vMMN. As shows, no such negativity emerged to the stimulus onset. In the younger group the vanish-related difference potentials were significant (t-tests, at least 10 consecutive significant values) in the 120–220 ms range at the Oz and O2 locations, and in the 120–202 ms range at the O1 locations. In the older group this range was 116–173 ms at the three locations. For the onset stimuli we obtained no significant range. Next, we segmented the significant range (i.e., vanishing stimuli) into five time ranges within the 120–219 ms period. Then we analyzed the difference potentials in the five ranges with six separate ANOVAs (one ANOVA for the whole time range and five ANOVAs for the five consecutive epochs) with the factors of Age and Location. contains the descriptive statistics of the five consecutive mean epochs. shows the scalp distributions of the vMMNs in the corresponding epochs. In a two-way ANOVA with factors of Age and Location we compared the amplitude of the difference potentials of the two groups for the whole 120–219 ms time period. The Age x Location interaction was significant, F(2,56) = 4.54, p\<0.05, η_p² = 0.14, ε = 0.76. The post hoc Tukey HSD test did not indicate significant differences, but according to the results of separate t-tests at the three locations the difference was significant at O2, t(28) = 2.11, p\<0.05, and approached significance at Oz (t(28) = 2.04, p = 0.05). shows the surface distribution of the difference potentials in the five time ranges in the younger and older groups. As the figure shows, in the three earlier ranges the data were fairly similar, but in the older group there was hardly any negative difference in the two later ranges. Results of the ANOVAs confirmed this observation. As for the first three epochs the only significant effect appears as an Age x Location interaction in the 140–159 ms range, F(2,56) = 7.01, p\<0.01, η_p² = 0.20, ε = 0.67. According to the Tukey HSD test, in the younger group the negativity was larger at O2 than at O1. In contrast with the earlier ranges, in the 180–199 and 200–219 ranges the main effect of Age was significant, F(1,28) = 5.58, p\<0.05, η_p² = 0.17 and F(1,28) = 4.50, p\<0.05, η_p² = 0.14, respectively. # Discussion In our study we introduced a new approach to investigate visual mismatch negativity (vMMN). The objects were continuously present, and stimulus presentation consisted of a part of the object disappearing repeatedly. This method has higher ecological validity compared to the usual paradigms. The main advantage of this technique is that it controls for low-level adaptation, and hence separates the standard- and deviant-related effects. As an important aspect of the present study, in comparison to frequently disappearing features of continuously presented objects, infrequently disappearing features elicited an additional posterior negativity, the vMMN component. In contrast, there was no ERP difference between the frequently and infrequently reappearing features. This result shows that low-level stimulus- specific adaptation (SSA) or refractoriness of elementary features does not necessarily contribute to the ERP difference elicited by frequent (standard) and infrequent (deviant) events. This is because in this paradigm the absence of well adapted stimulus features elicited vMMN, but the onset of a less adapted feature did not. In this respect the present paradigm is an alternative to the equal probability paradigm as a control for SSA, at least on the level of elementary (low level) stimulus features. The results reveal a new aspect of the vMMN’s nature. The prevalent approach of the refractoriness (cf.) issue claims that the early range of the deviant minus standard difference (before 200 ms) is inevitably due to the refractoriness effect. Our paradigm per definition ruled out the classical refractoriness effect. According to that logic, the difference wave should not have emerged after 200 ms post stimulus. In contrast, we obtained clear vMMNs in both age groups during the “refractoriness time window” in the vanish condition (and no vMMN in the appearance condition). This contradiction leads us to the reevaluation of the refractoriness or low- level adaptation concept in the vMMN research. The source of vMMN is uncertain, according to localization attempts, occipital and other posterior structures seems to be involved. More importantly, the vMMN latency range is far beyond the onset time of the striate cortex (i.e. the location of the first entry to the visual cortex). When presented with complex visual stimuli even the temporal cortex is active within 100 ms. Accordingly, vMMN activity of lower level visual areas, if such activity is present, results from feedback processes, in other words, it is a consequence of reentrant activity. Furthermore, even at the level of basic visual processing there is evidence of higher-level organization, evidenced by sensitivity to Gestalt-like grouping. Therefore, integrated visual representations may contribute to the emergence of the negativity. According to this interpretation vMMN emerged to the irregular disappearance of visual elements (deviants), but the reappearance of the elements (irrespective of the deviant or standard one) does not violate any rule, because in both cases (standard and deviant), it is the regular object. The nature of the experimental stimulus could be the solution to the contradiction between Kimura’s and the present results. In Kimura’s study, the event was the onset of a central bar. The relative weight of the more fundamental processes (i.e. refractoriness) is presumably lower (possibly below measurable level) during the processing of Gestalt-grouping or object formation and it is higher during the processing of a singleton. In sum, it is almost impossible to match up the results between vMMNs elicited by fundamentally different stimulus types. Furthermore, the present results corroborate the appearance of object-related vMMN, and vMMN to infrequent vs. frequent changes of visual objects. It should be noted that the Besle et al. and Clery et al. studies–in which continuous stimulus presentation was applied–did not control low-level adaptation, because both the standard and deviant features were different from the features of the inter-stimulus field (i.e. ellipses from circles). It is important to recognize the limits of the present paradigm. Most importantly, both the standard and deviant vanishing features were location- specific (45° lines were always at the top left and bottom right sides of the rectangles, and so were the 135° lines). Therefore it is possible that the vMMN of the present study is a sign of the sensitivity to change in frequent *versus* infrequent locations. This is a testable possibility for further research, and with the exception of a single study with highly different methods has not been investigated. vMMN to change in unusual location would be a notable result in the field. Furthermore, on a theoretical level the present results do not provide direct evidence of a mechanism underlying deviant-related negativity, particularly whether it is a predictive coding process. As a conservative interpretation, processing of the infrequent vanishing element was different from the processing of frequent ones. The other topic of the present study was the investigation of age-related aspects of automatic change detection. The results were straightforward. In the tracking task the younger group outperformed the older one. This is an expected result for visuomotor tasks requiring fast reactions. In the younger group vMMN terminated later than in the older group (\~220 ms vs. \~180 ms). In the 120–180 ms time period the amplitude of the difference potential was similar in the two age groups, and in this respect the results were similar to the Stothart et al. and Gaál et al. studies. Comparing the whole (120–219 ms) time period, at the right occipital location the amplitude was larger in the younger group, and this result corresponded to the results reported by Tales et al. and Lorenzo-López et al.. However, due to the different offset of the difference potentials in the two age groups, the whole period amplitude calculation is insufficient. As in the study of Lorenzo-López et al. shows, vMMN in the younger and middle aged groups were longer. Comparing the ERPs to deviant and standard, one cannot ignore the duration difference as Tales et al. did (; see their). The age-related difference in the earlier vs. later time period requires interpretation, even if the possible explanations are *post hoc*. Maintaining that the posterior negativity is an indicator of different processing of the infrequent and frequent events, the process is longer in the younger group. It is unlikely that the reason of the duration difference is the diminished processing speed in the younger group, because this possibility contradicts one of the most frequent observation of aging research, that of age-related cognitive slowing. Another possibility is the longer persistence of neural activity in the younger group. However, studies on visual masking and perceptual integration provided opposite results, i.e., higher persistence in elderly. We hypothesize that the longer negativity is an indicator of more extensive processing. Registration of an unusual event (“there is something unusual”) is supposed to be similar in the two age groups, but the registration of change may be followed by the identification of the particular change (“what is the new event?”). This level of processing emerges more frequently in the younger group. The description fits within the predictive coding framework (e.g.. In this framework mismatch components are considered error signals and are related to subsequent re-adjustment of mental representations. However, it is important to emphasize that the present results do not provide direct evidence for the operation of a predictive memory mechanism. In conclusion, infrequent disappearance, i.e., the associated change in the shape of objects elicits vMMN, but the reappearance of the object (even if it is an infrequent change at the level of the appearance of individual features) does not. Detection of infrequent change elicits more extended activity in younger participants, and the longer activity is considered to be an indicator of more profound automatic processing. # Supporting information The study was supported the by the National Research Found of Hungary (OTKA 104462 and 115457) and by the János Bolyai Research Scholarship of the Hungarian Academy of Sciences. We thank Dr. Szonya Durant and Dr. Dagmar Müller for the helpful comments. We thank for the technical assistance by Zsuzsa d’Albini and Zsuzsa Kovács. [^1]: The authors have declared that no competing interests exist.

# Introduction In vertebrate brains, sensory information is often represented in an orderly manner. Our body surface is, for example, represented in the well-known homunculus in the somatosensory cortex. Likewise, in the visual system, visual information reaching the retina is represented topographically in many of the target areas of the visual pathway: visual stimuli from neighboring locations in the visual field activate neighboring neurons within the visual processing area. The origin of the visual topographic maps within the brain is the retina. Its essential organization including photoreceptors, amacrine, horizontal, bipolar and retinal ganglion cells is similar in all vertebrates. It is also a common feature of all vertebrate retinae that the density of photoreceptors and retinal ganglion cells is not homogeneous. In many cases, there is at least one, sometimes even more regions of higher density. Often, a fovea is found with even more densely packed photoreceptors and retinal ganglion cells. As a rule, the dimensions of neuronal retinotopic maps depend on the density of retinal ganglion cells. The foveal area of many mammals, for example cats and primates, has a high density of retinal ganglion cells and, accordingly, the part of the retinotopic map representing the fovea is larger than parts representing similarly sized areas of the retinal periphery. This is called "foveal overrepresentation". Such overrepresentation is in many cases enhanced by differential wiring of foveal and nonfoveal photoreceptors and retinal ganglion cells: Within foveal regions, single or a small number of photoreceptors project to a given ganglion cell, while in the periphery, hundreds of photoreceptors can converge onto one ganglion cell. In mammals as well as in other vertebrates, it is not only one eye which is projecting to a given brain map. It has been frequently shown that information from both eyes is combined in retinotopic maps. In many cases, binocular neurons receive information from both eyes, and the percentage of binocular neurons is related to the overlap of the visual field of both eyes: if the visual field overlap is high like in primates or cats, there is a large binocular segment in the retinotopic map, and a comparatively small one in laterally eyed animals with a smaller binocular overlap like mice. The visual system of birds consists of three parallel projections originating in the retina-tectofugal pathway, thalamofugal pathway and accessory optic system (AOS). Visual information is transported completely to the contralateral hemisphere by the optic nerve. AOS is mainly involved in the processing of optic flow induced by self-motion. The tectofugal or collothalamic pathway has its first station in the mesencephalic optic tectum which is organized in up to 15 layers and contains a complete retinotopic map. From there the information is transferred in a non-topographic manner to the thalamic n. rotundus which has special subdivisions for color, luminance, and motion of visual stimuli. The first telencephalic station is the entopallium. Entopallial lesions in pigeon have demonstrated that the tectofugal pathway is important for pattern discrimination and complex processes like concept learning or discrimination of concepts (review:). At least for laterally eyed birds, the tectofugal pathway has been considered as the most important one because lesions of this pathway in general have massive effects in comparison with lesions of the thalamofugal or lemnothalamic pathway, which will be described in more detail below. It has to be noted, however, that there are as yet no lesion studies and no other information about its relative importance for frontally eyed birds. It is possible that in these birds the thalamofugal system, especially tuned to process binocular information (see below) may have become the more important projection, as it has been shown for the homologous geniculocortical system in mammals. The aim of the present study is to investigate functional aspects of the telencephalic station of the thalamofugal pathway, the visual wulst, which receives information from the retina, relayed by several thalamic nuclei (review:). In particular, we wanted to concentrate on the retinotopic representation of the visual field within this area. The experiments were performed in the zebra finch, a small songbird with laterally placed eyes and in this aspect representative of most avian tribes. The eyes of birds are, as mentioned before, in most aspects similar to those of mammals. Special features are the so called oil droplets which contribute to the high spectral sensitivity, the pecten, a big structure protruding into the vitreous which is thought to serve for nutritional purposes, and probably a much higher variation across species in photoreceptor and retinal ganglion cell densities. The visual wulst receives visual input from the retina relayed by the optic thalamic nuclei. It consists of three roughly parallel oriented layers. Visual input from the thalamus reaches the middle layer, the IHA (nucleus interstitialis hyperpallii apicale), and is transferred from this layer to the upper layer, HA (hyperpallium apicale) that sends projections back to the thalamus, and the lowest layer, HD (hyperpallium densocellulare) which is the source of intratelencephalic projections of the wulst. Developmental studies indicate that these layers cannot be compared with cortical layers although the area is said to be homologue to the visual cortex. Using optical imaging of intrinsic signals we have previously shown that the zebra finch visual wulst comprises several topographic maps which can be easiest detected from the IHA at a depth of around 500 μm; recording from deeper or more superficial layers revealed reduced signal-to-noise ratio. Information from both eyes reaches the visual wulst in spite of the fact that, in contrast to mammals, the optic nerves cross completely to the contralateral side in birds. Pettigrew and Konishi electrophysiologically recorded binocular neurons from the visual wulst of owls and described a number of neuronal features comparable with those of neurons of the mammalian visual cortex like precise retinotopic organization, direction and orientation selectivity and binocular disparity. The information from the ipsilateral eye was shown to originate from a secondary recrossing of thalamic efferents to the contralateral telencephalon by means of the dorsal supraoptic decussation. Such binocular processing within the visual wulst, however, is probably a specialty of birds with frontally placed eyes and a quite large binocular visual field. Although there is a substantial amount of recrossing fibers, the influence of the ipsilateral eye on the activation of the wulst has been shown to be marginal in birds with laterally placed eyes like the chicken and the zebra finch. It has been speculated that the size of the wulst is related to the amount of binocularity and the size of the binocular overlap. In previous electrophysiological experiments, we have found that in the zebra finch, ipsilaterally evoked activity in the visual wulst reduced activity induced by stimulating the contralateral eye.Birds with lateral eyes fixate objects within the environment by directing the fovea of one eye to it. In such a case, a suppression of the information from the nonfixating eye may make sense because binocular processing is not advantageous. Based on our previous report on the existence of several retinotopic maps in the zebra finch visual wulst, we are now aiming to describe the retinotopic visual field representations in more detail and ask the following questions: What is the extent of the visual field representation in the visual wulst? How far can the architecture of the wulst map be explained by the retinal ganglion cell distribution? Is there a foveal overrepresentation? Are there binocular interactions, i.e. does stimulation of the ipsilateral eye affect wulst activity induced by stimulating the contralateral eye? In this paper, we will focus on the most posterior map of the multiple visual field representations because it is strongly activated by visual stimuli and—in contrast to the more anterior maps—consistently visible in each imaging experiment. Instead of using optical imaging of intrinsic signals as in our previous mapping study (OIS), we here used autofluorescent flavoprotein imaging (AFI) because it is superior to the OIS method for the demonstration of retinotopic maps in the visual wulst of zebra finches as we have recently shown. In addition, we performed ophthalmoscopic measurements and counted retinal ganglion cells to provide a basis for the relation between retinal architecture and wulst retinotopic maps. # Materials and Methods 21 zebra finches of 100–110 days of age from the breeding stock of Bielefeld University were used for the imaging experiments. The retinae of additional three birds (age range 120–300 days) were used for the estimation of the ganglion cell density distribution and the measurements of retinal landmarks. ## Ethics statement All experimental procedures were performed according to the German Law on the Protection of Animals and permitted by the local government. All imaging experiments performed in Göttingen were approved by the „Niedersächsisches Landesamt für Verbraucherschutz und Lebensmittelsicherheit”(permission numbers 84–02.04.2011.A217 and 33.9-42502-04-14/1451), all retinal ganglion cells experiments by the "Landesamt für Natur-und Verbraucherschutz" (permission AZ 87–51.04.2010.A069, May 11, 2010, LANUV, NRW). ## Optical imaging Procedures were performed as described previously. Briefly, the birds were anesthetized by injecting 0.1 ml of 20% urethane intramuscularly, and then kept under infrared light to maintain the body temperature. The birds’ heads were fixed using the stereotaxic head-holder for small birds, modified to hold the head in a position where the beak was oriented horizontally using ear bars and a beak holder, a modification due to the fact that the original head angle of 45° was comfortable for preparing the stereotaxic atlas, but an angle of 0° is more appropriate to mimic the natural head position of a zebra finch sitting on a perch (Bischof, unpublished). A craniotomy was performed on the left hemisphere to expose the visual wulst, leaving the dura mater intact. The exposed brain area was then covered with warm agarose (2.5% in saline) and a glass cover slip. The contralateral eye was opened by retracting and fixating the lower eyelid in an open position by Histoacryl, keeping the nictitating membrane intact. If moistening by the nictitating membrane appeared to be not sufficient, silicon oil was applied to keep the sclera in good condition. Neuronal activity in the left visual wulst of zebra finches was captured using autofluorescent flavoprotein imaging. For data acquisition and analysis we used the Fourier imaging method introduced by Kalatsky and Stryker: To image sensory- driven activity, a temporally periodic stimulus was continuously presented to the bird and the brain’s response at the stimulus frequency was extracted by Fourier analysis. Optical images of visual wulst activation were obtained using a CCD camera (Dalsa 1M30) and a 130 x 55 mm lens with an aperture of 1.2 (Nikon, Tokyo, Japan), controlled by custom software. The acquired image covered 4.6x4.6 mm of the brain surface. The camera was focused below the wulst surface at a depth of 500 μm, and neuronal activity was captured using blue light (455 ±10 nm). Frames were acquired at a rate of 30 Hz, binned to 7.5 Hz and stored as 512x512 pixel images after spatial binning of the camera image. The approximate location of the center of the posterior map of the visual wulst is schematically shown in. ## Visual stimuli A high refresh rate monitor (Flatron LCD 295LM, 100 Hz, 46.5×30 cm), covering 75°×50° of the visual field, was placed at various angles around the bird at a distance of 30 cm from the eyes. In the experiments examining the extent of the lateral visual field represented in the visual wulst, we used up to 5 different monitor positions in the horizontal plane: 0° (in front of the bird), 30°, 60° (in a position that the fovea was directed to the center of the monitor), 90°, 120° with respect to the right (contralateral to the imaged hemisphere) eye. In case of stimulating the ipsilateral eye from the foveal direction, the monitor was positioned at 60° with respect to the left eye. For determining the frontal and binocular visual field representation, the monitor was placed at 0° with respect to the beak. Vertically, the monitor was positioned such that the horizontal line from the eye to the monitor was 7 cm above the lower rim of the monitor. In some experiments, we lowered the monitor by 15 cm. Visual stimuli consisted of vertical (azimuth) or horizontal (elevation) moving bars (4° wide), moving at a speed of 10°/sec. ## Data analysis Activity maps were calculated from the acquired frames by Fourier analysis to extract the signal at the stimulation frequency using custom software. While the phase component of the signal was used for the calculation of retinotopy, the amplitude component represents the intensity of neuronal activation, i.e. response magnitude expressed as fractional change in reflectance x10^-4. Retinotopic maps were colour-coded with respect to the position of the stimulus eliciting this activation. The combined information of the magnitude of neuronal activation and retinotopy is displayed in so-called polar maps. The colour codes used to display the retinotopic polar maps are shown along with each polar map. Please note that the colors in the wulst maps (Figs) represent different visual field positions for different monitor positions. Since blue always represents the center of the monitor, it corresponds to e.g. 0°, +30°, +60° etc. of the visual field for the azimuth maps, and to +15° for the elevation maps. To display visual field proportions in the retinotopically activated brain areas, iso-azimuth and iso-elevation lines were calculated using a Matlab routine and superimposed on the retinotopic polar maps. These calculations were preceded by a thresholding procedure which selected all pixels within the activity patch that showed at least 30% of the peak response, all other pixels were discarded. This procedure is illustrated in. Fig – gives examples of thresholded activity patches from birds 1 and 3, Fig – shows the calculated iso-azimuth/elevation lines, and Fig – illustrates the overlay of the thresholded activity map with the retinotopic visual field contours. To construct the entire wulst retinotopic map of individual birds from the “partial” maps obtained from the various monitor positions, we superimposed the iso-azimuth contour lines of all maps, and—in case of overlap—displayed only the lines from the central 20° of each monitor position. ## Foveal magnification Two parameters were used: First, we measured the size of the azimuth and elevation maps, combined the values from the two peripheral monitor positions (0° and 90°) and compared these values by t-test with the combined two”foveal” monitor positions (30° and 60°). We did this because our first impression was that the 30° and 60° segments were bigger than the peripheral ones. Second, we used a Matlab routine to evaluate the number of pixels within the activated region having the same phase value, from every 10° of the visual field starting from -5° in the ipsilateral visual field to +95° in the contralateral visual field. 30° and 60° maps were chosen for this measurement as they very well covered the area around the foveal representation and also the periphery. Since there is substantial overlap between the two monitor positions, there are more measurements/cases for each segment in the overlap region. Again, all the pixels with an amplitude lower than at least 30% of the maximum responsive pixel were discarded. ## Interaction of ipsi-and contralateral eye evoked responses We used two parameters for the determination of evoked activity differences between bilateral and contralateral stimulation: magnitude of the neuronal activity and average amplitude of the activity. Magnitude of the neuronal activity has been explained above. In addition, we calculated the average amplitude of the wulst activation in the region of interest using ImageJ (NIH, Rasband). ## Ophthalmoscopic measurement of retinal landmarks For the inspection of retinal landmarks we used a method described in detail by Nalbach et al who determined the position of retinal landmarks in the pigeon. In short, an ophthalmoscope constructed by adding additional lenses to an epi- illumination surgery microscope, was used for the inspection of the fundus of the eye. It was attached to a goniometer, which allowed the determination of azimuth and elevation of a given landmark with a precision of ± 0.1°. The values were taken relative to the optical axis of the eye which was determined by centering the corneal reflection of the epi-illumination in the pupil. Three zebra finches were used for this study. One week before the experiment started, the birds were anaesthetized by Equithesin (2mg/kg, i.m., 0.85 g chloral hydrate, 0.21 g pentobarbital, 0.42 g MgSO₄, 2.2 ml 100% ethanol, 8.6 ml propylene glycol, brought to a final volume of 20 ml with dH₂O), an anesthetic found to be ideal for birds. After applying a local anesthesia (xylocaine gel) to the scalp, a small incision was made on the scalp skin, and a plastic bar of 1 mm diameter and 3mm length was attached to the skull with dental cement. This bar served to fix the head of the anaesthetized bird in an adjustable clamp in a position that the observed eye was exactly located at the center of the goniometer. We measured the retinal landmarks in the left eye of all birds and also in the right eye of one of the birds. All measurements were superimposed (the right eye measurements were mirrored) by using the optical axis measurement and the beak direction. The angular measurements could be transformed into distances on the retina by taking into account the distance from the retina to the center of the pupil (4.1mm) and using conventional trigonometric formulae. This transformation is fully correct only near the optical axis but could be used to adjust the size relation between the retinal ganglion cell density map and our goniometric measurements. ## Spatial distribution of retinal ganglion cells After finishing the ophthalmoscopic measurements, an overdose of Nembutal was injected and the zebra finches were perfused with 0.9% saline followed by 4% formaldehyde. The head of the birds was then attached to the stereotaxic head- holder. The orientation of the eye was marked by inserting an insect pin into the eye bulb in dorsoventral direction. The eye was then removed from the skull, postfixed in 10% formaldehyde, and embedded in paraffin in a way that the insect pin remained in the exact vertical position. From the paraffin blocks, 20 μm horizontal sections were cut perpendicular to the insect pin tract. By this procedure, the sections were in the same plane as the horizon which helped to interpret the results. The sections were then mounted on Superfrost slides, stained with Giemsa-solution, and cover slipped with Depex (Serva Elecrophoresis GmbH). After drying, the sections were examined using a Zeiss Axioscope under a magnification of 400x. In every fifth section, the number of retinal ganglion cells/100μm was counted along the retinal ganglion cell layer of the entire retina. The number of neurons/100x100 μm was then estimated by multiplying with five. The data were stored in a two dimensional array and processed by Sigmaplot \[Systat Software Inc.\] to obtain a two dimensional density map. The number of cells/mm² was colour coded in steps of 2000. No attempt was made to correct the data for tissue shrinking or counting errors as we were interested in the detection of retinal areas showing enhanced ganglion cell densities, but not in the absolute numbers. We had also obtained landmarks (retina and pecten, see above) from ophthalmoscopic measurements that served to adjust the size and orientation of the obtained density map. ## Statistics One-way repeated measures ANOVA followed by Newman-Keuls multiple comparison tests was applied to compare the magnitude of activity of the retinotopic maps from the various monitor positions and to test foveal overrepresentation. Paired t-test was used to test the difference between the activity obtained using elevation and azimuth stimuli and also the difference between the activity evoked by the stimulation of contralateral eye and binocular stimulation. Values are expressed as mean±SEM. The levels of significance were set as \*: *p*\<0.05; \*\*: *p*\<0.01; \*\*\*: *p*\<0.001. # Results ## Properties of the representation of the contralateral visual field Extending our previous study about wulst retinotopic maps, we here focused on imaging the extent of the visual field representation in the visual wulst. To this end, we placed the stimulus monitor at five different locations from the frontal to the lateral visual field, and visualized wulst activation resulting from visual stimulation with moving bars (4° wide white bar). Both horizontal bars moving vertically (elevation stimulus) and vertical bars moving horizontally (azimuth stimulus) were used for determining retinotopic maps. We always recorded wulst activity from the left hemisphere and visually stimulated the right (contralateral) eye; a total of five zebra finches were used for this subproject. Figs and show examples of retinotopic and activity maps in two different zebra finches. We used a vertical bar moving from left to right and from right to left (azimuth stimulus) to examine the fronto-lateral extent of the visual field represented in the visual wulst. Figs, and illustrate the azimuth polar and activity maps of birds 1/2 when the stimulus monitor was positioned at 0°, i.e. in front of the bird. Similarly, Figs, , and represent the retinotopic and the corresponding activity maps obtained when the stimulus monitor was positioned at 30°, 60° and 90° respectively. The vertical extent of the visual field representation was examined with horizontal bars moving up and down. Figs – and – show the respective elevation and activity maps. The colour coding of the retinotopic maps relates to the position of the stimulus in the visual field that elicited the wulst activity and is illustrated above the maps (see for further explanation). The magnitude of visual wulst activation is grey-scale coded with high activation represented in black; it is additionally given as a number in the upper right corner of all activity maps. Taken together, the stimulated visual field extended from -30° to +130° laterally (see scale on top of Figs – and –), and from -13° to +37° in the vertical direction (see scale left to Figs). In all examined cases, there was a gradual shift in the location of the activated wulst region from anterior to posterior in the zebra finch brain when the visual stimulus was presented more laterally (Figs). This is most easily seen by comparing the location of the “blue” region in Fig, corresponding to the center of the stimulus monitor, in the different azimuth maps. These polar maps confirm the presence of retinotopic order in the zebra finch visual wulst and extend previous knowledge by demonstrating that the maps extend into the periphery of the visual field. Interestingly, there was a difference in the magnitude of the wulst activation (repeated measures ANOVA, n = 48, F = 4.132, R² = 0.4525, P = 0.0021) obtained from the various stimulus monitor positions. The strongest azimuth activity map was most of the time obtained when the stimulus monitor was at 60° (3.5±0.6) with respect to the eye (e.g. Fig). However, the 30° map (3.2±0.6, eg: Figs, and) did not differ substantially from the 60° map, but the maps induced by the 0° (2.5±0.5) and 90° (2.9±0.7) monitor positions were clearly less active (eg: Fig,). Nevertheless, we also had two examples in which the 90° azimuth map had higher activation than the 60° maps (for eg: compare Fig). The elevation maps were in general less consistent if compared with the azimuth maps. However, at least for the differences in the magnitude of activation measurements, the elevation map calculations confirmed those of the azimuth maps. Figs and – shows the elevation polar and activity maps. It was again the 60° map which showed the strongest activity in most of the cases (0°: 2.7±0.7, 30°: 3.4±0.6, 60°: 4. shows two examples of how we quantified the visual field extent in the visual wulst (for details see). Briefly, the thresholded activity maps were superimposed with iso-azimuth and iso-elevation lines. These analyses revealed that the azimuth visual field extended from about -5° in the ipsilateral visual field to 125° in the contralateral visual field, and the elevation visual field from about -10° below the horizon to +25° above the horizon. These dimensions of the wulst visual field representation can also be extracted qualitatively from the colour coded polar maps illustrated in Figs and. Fig and depict the entire representation of the visual field from birds1 and 2 (single maps illustrated in Figs) which were visually stimulated from -30° to 90°). In bird 1, visual stimuli from -5° to +125° medio-laterally and from 0° (horizon) to +25° above the beak induced measurable activity in the visual wulst. The visual field represented in the visual wulst of bird 2 extended from 20° to 110° laterally and from 0° to +25° above the horizon. Although there is some variation concerning the exact extent of the visual field, our experiments show that, if one takes the extreme cases with the argument that lower values are most probably due to non-optimal experimental conditions, the horizontal extension of the visual field reaches from -5° frontally up to 125° laterally, and vertically from 10° below to +25° above the horizon. From the reconstructed maps, we determined the size of the entire posterior retinotopically organized area in the visual field as about 1.1x1 mm; the center of this representation was located 4.1 mm anterior and 2.3 mm lateral to the Y point (meeting point of the cerebellum and the two hemispheres schematically shown) in the coordinate system used for the zebra finch brain atlas of Nixdorf and Bischof. Although there is pronounced interindividual variability in the details of the iso-azimuth and iso-elevation contour maps, they are consistent on another feature of the retinotopic maps: as clearly visible in, the spacing of the green iso-elevation lines is much wider than that of the red iso-azimuth lines, indicating that 10° of the visual field in the vertical plane is analyzed in about 2.5 to 3 times more neuronal space than 10° of visual field along the horizontal plane. Further comparison of the results from the different monitor positions revealed some interesting features of the intrinsic organization of the topographic representations. In mammals, as mentioned above, an overrepresentation of the foveal area has been found frequently. We therefore evaluated a potential foveal overrepresentation and also the interindividual variability of the maps between different birds by calculating the size of the areas between the azimuth lines of 10° distance along the topographic representation from -5° frontally to +95° along the horizon. illustrates large inter-individual differences in area sizes but no detectable foveal overrepresentation (indicated by an open blue arrow; (ANOVA, F (9, 44) = 0.7781, p = 0.6372). We also calculated the area of the activity patch induced in the visual wulst by peripheral (0°+90°) or foveal stimulation (30°+60°) (azimuth, peripheral: 0.6±0.2 mm², foveal: 0.8±0.1 mm²; elevation, peripheral: 0.7±0.2 mm², foveal: 0.9±0.2 mm²). Yet again there was no difference between the area sizes (repeated measures ANOVA, F (3, 6) = 0.8216, p = 0.4677). ## Retinal ganglion cell density measurements and retinal topography depicts the results of our retinal ganglion cell density measurements and the determination of the retinal landmarks by ophthalmoscopic investigation. Three zebra finches were used for this study. The calculated retinal ganglion cell density map is illustrated in ; where in the visual field these regions of higher ganglion densities were present is illustrated schematically in. It shows the position of the retinal landmarks appearing as if projected through the lens onto a virtual globe representing the visual space. This means that the measurements of the fovea appear in the direction where the fovea is looking. The pecten is located ventrally in the eye, so its position in the visual field is dorsally, indicating that the pecten occludes the view of a substantial part of the dorsal visual field. Small red circles, marked by “F” indicate the foveal high density region of, the yellow oval corresponds to the second high density region, and the area where the pecten is positioned is shown in green. shows a cross-section through the whole eye, including an arrow that marks the distance used for the transformation of visual angles into distances on the retina; and Fig – show examples of retinal cross-sections that were used for reconstructing the density map in A. In the cross sections, one can easily discern 4 layers containing cell bodies, with the ganglion cell layer (GCL) being the uppermost and relatively thin layer in each of the pictures. The density map shows how retinal ganglion cell density decreases from the fovea, which appears as a deep tapered depression in the retina, to more peripheral regions (Fig). The density map illustrated in is a partial reconstruction of the entire retina, containing the most important areas like fovea, horizon and pecten. Our measurements reveal a tiny region of high ganglion cell density at the fovea (F) with about 18000 ganglion cells/mm² (black), surrounded by regions with slightly smaller densities of 16000 cells/mm² (red), and 14000 cells/mm² (orange). High densities were also found at a second spot "looking" more frontally in the visual field with a maximum of 16000 cells/mm² (red, marked by a white arrow). The fovea and the second dense area are connected by a strip of slightly lower cell density (12000 cells/mm², yellow). This situation may be best described as a so- called visual streak. At the upper part of the map there is a white stripe indicating a region free of ganglion cells, corresponding to the pecten, an eye specialization found in sauropsids, which inserts in the retina and protrudes into the vitreous. The retinal ganglion cell density map was then slightly rotated and enlarged without distorting it to adapt it to our landmark measurements. We located the fovea (measurements, two each from the four analyzed retinae, at about 0° elevation (indicated by small filled red circles,), corresponding to the beak tip (when the beak is held horizontally), and 60° laterally, confirming earlier results. According to our landmark measurements, the second high density area was "looking" in a direction located about 0° to 45°. Obviously, the region of higher density was not totally parallel to the horizontal plane; a slight shift to more ventral regions was observed (yellow oval). The position of the pecten and its extension (measurements marked by green circles, extension by a green oval) indicated that it covers a substantial part of the dorsal visual field. We also included the measurements of the ora serrata, the peripheral rim of the retina (pink circle). However, because of the strong refraction at the periphery, these measurements are not entirely reliable. summarizes our results concerning visual field extent and retinal landmarks in a schematic representation of the visual field. The pink shaded area illustrates the extent of zebra finch visual field from where visual stimuli elicited measurable activation in the visual wulst in our imaging experiments. The visual field measured about 130° in the horizontal plane, extending from around -5° in the ipsilateral visual field to about +125° in the contralateral periphery (azimuth), and about 35° in the vertical direction, extending from -10° to about +25° (elevation) between 30°-60° azimuth. It was reduced to about 15°-20° in vertical extent both frontally and more laterally. shows a schematic and planar bird eye’s view of the fronto-lateral visual field represented in the visual wulst. Frontally, the binocular visual field measures about 10°, the monocular left and right visual field about 130°, while a caudal region of about 100° is most likely not represented in the visual wulst. ## Ipsilateral and contralateral input to the Visual wulst We also investigated whether the wulst gets activated by visual stimulation of the ipsilateral eye, and if so, whether the areas driven by the ipsilateral eye overlap with those activated by the contralateral eye. Given the fact that zebra finches have only a small binocular overlap, it would be difficult to comprehend if substantial regions of the wulst, which primarily represent very different parts of the visual environment would get activated by both eyes. If there is a frontal region where the visual fields of the two eyes are overlapping, afferent inputs from this visual field region could produce a binocular segment as found in mice, given that the afferents from both eyes converged on binocular neurons. We therefore examined the relation of contra- and ipsilaterally induced wulst activation in two experiments. First, the stimulus monitor was positioned at +60° (right visual field, right eye stimulation) or—60° (left visual field, left eye stimulation), and azimuth (Fig) or elevation stimuli (Fig) were presented. Activity was recorded in the left visual wulst. Interestingly, contralateral eye stimulation in this bird evoked multiple retinotopic maps after both azimuth and elevation stimuli (Fig). While in the azimuth map separation of the different maps is difficult, there were clearly 3 separate maps in the elevation map, with the anterior two maps being inverted with respect to the posterior map. Wulst activation induced by stimulating the ipsilateral eye was weaker and more restricted in the posterior two maps (compare Fig with and with). The pink color in indicates that ipsilateral activation was induced by stimuli at around 50° to 40° azimuth in the left visual field. The activity patches induced by contralateral azimuth stimuli did not differ from those induced by elevation stimuli. The same was true for activation induced by azimuth and elevation stimuli via the ipsilateral eye. Fig and illustrates the superimposition of thresholded contra- (red) and ipsilaterally (green) evoked activity regions. It can easily be seen that there are separate areas activated by the contra- and ipsilateral eye, but there was also substantial overlap in the middle and the most posterior activated regions, visible as yellow patches in Fig and. After having shown that visual stimuli positioned about 40 to50° laterally and -5-+25° vertically to the ipsilateral eye can activate the visual wulst, we next tested whether there is also a binocular wulst region activated from the frontal visual field, as present in mammals, and what effect binocular visual stimulation had on wulst activity. Six zebra finches were used for this experiment. The stimulus monitor was centered at 0°, in front of the beak. Both horizontal and vertical moving bars induced activity in approximately the same wulst location of the posterior map (compare Fig with, and with). Ipsilateral eye stimulation elicited only weak wulst activation in bird 5 (Fig) and no discernable activation in bird 6 (Fig). Interestingly, however, in both birds, binocular visual stimulation induced a wulst activation that was reduced compared to stimulation with the contralateral eye only (compare Fig / with /, and / with /). The average wulst activation after contralateral stimulation with an azimuth stimulus decreased from 1.98±0.27 (contra only) to 1.73±0.23 (contra+ipsi) (p\<0.05, t-test, n = 6) with binocular stimulation, and from 2.54±0.35 (contra only) to 2.24±0.39 (contra+ipsi) (p\<0.05, t-test, n = 6) for elevation. Thus, the ipsilateral eye exerted a measurable inhibitory influence on wulst activity when both eyes were stimulated together in the frontal visual field. As proposed for the whole wulst area based on previous electrophysiological results, contralaterally evoked activity was thus reduced if there was an additional input from the ipsilateral eye. As an additional parameter to test the relation between the contralaterally evoked activity and activity evoked by simultaneous stimulation of both eyes, we calculated the average wulst activity in a region including the activity patches evoked by both azimuth and elevation stimuli using the ImageJ software. Similar to the activity measurements described above, there was a significant difference between the two conditions (azimuth: contralateral: 73±8.4, contra+ipsi: 59.1±7.1, p\<0.05, t-test, n = 6, ; elevation: contralateral: 95.4±11.9, contra+ipsi: 78.4±12.9; p\<0.01, t-test, n = 6). # Discussion Our imaging experiments indicate that the posterior retinotopic map of the visual wulst, represents only part of the zebra finch visual field , extending horizontally from -5° to about +125° on the caudal border, its vertical extension spans from about -10° (below the horizon) to about +25°. Thus, visual stimuli from only a small strip along the horizon can visually stimulate the imaged wulst region. The overlap of the representations of the left and the right eye is around 10°, and our experiments with contra- and ipsilateral stimulation indicate that there could be a binocular segment characterized by binocular neurons as it has been demonstrated in mammals. While our cell density measurements show an increased number of retinal ganglion cells at around +60° laterally and along the horizon, our reconstructions of iso-elevation and iso- azimuth lines do not give clear indications of a foveal overrepresentation in the wulst. However, the iso-elevation lines are more widely spaced than the iso- azimuth lines throughout the entire wulst visual field representation. This means that a circular object within the visual field should activate an oval region in the wulst, with the longer axis aligned to the vertical axis. In the binocular visual field, information from the ipsilateral eye seems to inhibit that of the contralateral eye; whether this is also true for lateral parts cannot be derived from our present results. Visual stimuli in the lateral parts of the visual field around the fovea, evoked small activity patches from the ipsilateral eye, partially overlapping with the contralaterally evoked activity. ## Extent of the wulst visual field representation Our experiments demonstrated that the wulst visual field representation reached frontally slightly into the ipsilateral visual field (-5°) and was laterally limited at +125°. Vertically, it reached from +25° above the horizon down to -10° below the horizon. By using eye aperture studies, Bischof had determined the visual field of the zebra finch to extend from -12.5° to +150° laterally. The difference in visual field extent to the present imaging results may have several reasons. First, unanaesthetized birds which are able to move their eyes into the direction of measurement were used in the previous study, while the birds employed for imaging in the present study were anaesthetized by urethane which has been shown to cause negligible eye movements. We have confirmed the latter by recording eye movements during imaging experiments using an infrared camera (data not shown). According to Voss, awake zebra finches can move their eyes up to ±16.5° in any direction, which may be subtracted from the Bischof results to compare the data with the present data from birds without eye movements; by doing this, the results of both studies appear rather similar. Second, one cannot be sure whether measurements using the eye aperture give real information about the area which can be processed by the eye. However, our retinal measurements indicate that the ora serrata (pink line in Figs), the line which separates the receptive part of the retina from the non-receptive, is at about -15° at the frontal limit of the visual field and at about 150° at the caudal limit, just confirming the earlier measurements provided by Bischof. Next, the activity evoked by stimuli at the rim of the visual field may have been smaller than that evoked by the central parts and too small to be detected by our method and finally, it is possible that the stimuli used might not have activated enough wulst neurons. Keary et al. reported earlier that only stimuli presented at an elevation between 17° and 42° (a total elevation extent of 25°) could elicit activity in the visual wulst. Though the upper and lower limits of the elevation visual field is somewhat different from what we report here, the overall extent of the visual field around the foveal region is similar to the present results. The discrepancy in the upper and lower limits can be explained because the center of the stimulus monitor was taken as 0° in, while 0° elevation corresponds to the horizon in our study. According to our new imaging experiments, the vertical extent of the wulst visual field representation is rather small and corresponds to just a narrow strip along the horizon (-10 to +25°). This cannot be explained by measurement differences because the upper and lower limits of the avian visual field are much higher and lower than the limits measured here. Our experiments therefore unequivocally show that not the entire visual field is represented in the visual wulst map, a result which is in accordance with findings in the pigeon examining the projection from the retina to the thalamus, and also with lesion studies demonstrating that lesion of the thalamic nuclei affect lateral, but not frontal vision. One has to consider here that the visual wulst map is not the only brain region representing the visual field of the birds. It has been shown that the tectal map most likely represents the entire visual field. Similar differences between representations have also been observed in mammals. In the cat, the representation is largest in area 17 (V1) and smallest in Brodman area 21a. There are also areas with representations elongated along the horizontal meridian (e.g. VLS-ventral lateral suprasylvian area or PLLS—posterolateral lateral suprasylvian area, review:). The exact extent of the visual field representation of a particular area may thus be related to the special processing tasks of this area. What may be the function of a visual representation restricted to a rather narrow stripe along the horizon, as observed here for the visual wulst? It has been speculated that the existence of a visual streak—as we have found it in the retina and which contributes to the projections to the visual wulst map—could enhance the spatial resolution of this area, which may be advantageous for the perception of moving objects at the height of the horizon. It could also be speculated that this elongated area along the horizon has to do with the feeding behaviour of zebra finches. As reported by Bischof, zebra finches fixate a grain that they want to eat by the fovea, then make a turn of the beak towards the grain which most probably leads to an activation of the wulst area along the strip of higher retinal ganglion cell density and correspondingly the corridor detected by the wulst map, and then perform a ballistic movement with eyes closed towards the grain (see also below). A similar idea was raised by Moroney and Pettigrew for kingfishers, which fixate the prey with a central fovea and then make a head movement to transfer the image to the temporal fovea looking frontally, before starting to catch it. In both cases, a higher density of neurons may facilitate such a transfer. Already in 1958, Duijm proposed that the visual streak might be used "in the accurate establishment of the position of the eyes and therefore for compensatory eye movements and spatial orientation". Mathews also speculated about a role in navigation, serving as a baseline to which other objects in the visual field can be referred. The functions which have been attributed to the visual streak could also be the basis for an explanation of the visual wulst function if it was only the streak which is projecting to the visual wulst. This, however, is not the case. The visual field represented in the wulst map extends more dorsally than the streak and also reaches into the field lateral of the fovea. Thus, the enhanced density of ganglion cells along the horizon should, if the entire population of neurons was projecting to the visual wulst, result in a stronger representation within the retinotopic map. Especially, the fovea should show some overrepresentation. However, our wulst imaging experiments do not show this. According to our results, visual stimuli at 60° (foveal) as well as at 30° (roughly the second higher density region) indeed resulted in higher magnitudes of wulst activation which in turn may indicate that neurons within this area have a higher preference for the stimuli used in our experiments. But we did not observe indications of a spatial overrepresentation of foveal regions as it is present in the mammalian visual cortex. In macaque monkeys, the central 10% of the visual field are for example represented in 90% of the V1 area, in cats it is 50%. Such an overrepresentation was also found in other visual areas of the cat, but there are also examples of an underrepresentation. In the mouse, no overrepresentation of any part of the visual field were detected; however, in this case this may be due to the lack of a retinal fovea in this species. The findings in cats (over- and underrepresentation) indicate that an increased density of retinal ganglion cells does not necessarily cause overrepresentation of this region in all brain areas involved in visual scene analyses. Most likely, each area is “just” accessing the information which is necessary to perform a certain task. V1 of mammals is thought to be involved in stimulus identification, and such a task needs a high spatial resolution as provided by foveal overrepresentation. Following this reasoning, the lack of a distinct foveal overrepresentation in the zebra finch visual wulst may indicate that foveal processing and stimulus identification are not the main tasks of the visual wulst. Given some preference for the fronto-foveal region, it rather appears that the wulst processes information from the visual field just along the horizon, and is not especially equipped for object identification which would most probably make use of the higher ganglion cell density within the fovea. We therefore speculate that approaching objects might be processed by the visual wulst. This is a testable hypothesis, and it fits well to a concept which has been raised some time ago by Watanabe and Bischof. On the basis of lesion experiments, it was claimed that the visual wulst may process the "where" of an object, while the second visual pathway and its telencephalic target, the entopallium, is processing the "what", i.e. the identification of an object. Consistent with this idea, the optic tectum that projects to the entopallium comprises a foveal overrepresentation, in contrast to the visual wulst. Watanabe et al. have shown that the visual wulst is feeding into the hippocampus and is therefore a candidate for mediating visual information necessary for orientation in space. This hypothesis is also supported by studies demonstrating that the visual wulst participates in sun compass orientation of pigeons, and in earth magnetic field orientation (mediated by the thalamofugal visual system) of both European robins, zebra finches and garden warblers. The vertical limitation of the wulst visual field to a rather narrow stripe along the horizon is, however, not explained by this scenario. The same ideas which have been discussed above for an explanation of a visual streak may also apply here as well as the lesion based results concerning the visual wulst function. Both explanations, the idea of a role of the wulst for feeding behaviour and the more global one as a mediator of visual information for spatial orientation, are not mutually exclusive. We prefer the interpretation of a more global role of the visual wulst compared to just a feeding controller. An interesting feature of the wulst posterior retinotopic maps is the higher magnification in the elevation compared to the azimuth direction: Ten degrees of visual space in the elevation direction were represented in a two to three times longer distance on the wulst compared to the azimuth direction. Such anisotropy of maps has also been found in a variety of mammals, but in contrast to our results, the azimuth magnification factor was in all cases bigger than that of the elevation. Ng et al., however, reported that the activation of the visual wulst of pigeons is stronger if stimulated with vertically oriented stripes moving horizontally and explain it by an overrepresentation of vertical contours, which would fit to our results. Although all authors dealing with this question assume that overrepresentation of certain features are due to reflect natural visual scene statistics there is as yet no conclusive idea explaining the potential difference between birds and mammals. ## Interaction of both eyes? As mentioned above, the topographic maps within the mammalian visual cortex are partly receiving input from both eyes: within the so-called binocular segment, neurons respond to visual activation of both eyes. The extent of the binocular segment depends on the eye position of the animal. Mammals with laterally placed eyes and a small binocular overlap, i.e. a small frontal region where the visual fields of both eyes overlap, have a small binocular segment in their visual cortex compared with animals with more frontally positioned eyes with a larger binocular overlap. The basis of binocularity is a partial decussation of the optic nerves from both the eyes at the optic chiasm: in animals with a binocular zone, only ganglion cell axons from the nasal part of the retina cross to the contralateral hemisphere, whereas axons from temporal retina do not cross, but project ipsilaterally. The bigger the overlap of the visual field, the bigger is the binocular segment within the visual cortex. In contrast to mammals, in birds, the optic nerves cross completely to the contralateral side. Hence, ipsilateral stimulus processing and binocular interaction as shown in the visual cortex can be achieved only by a secondary re-crossing of fibers to the ipsilateral brain hemisphere. This actually occurs at both visual projections, the tectofugal and the thalamofugal pathway. As mentioned above, binocular interactions have been shown in owls which have quite frontally placed eyes and a substantial overlap of the two eye’s visual fields. In contrast, there is as yet no evidence for binocular processing in zebra finches. However, there are differences between results from anatomical studies showing a substantial projection originating in the ipsilateral eye, and electrophysiological studies indicating that this projection does not have much impact on wulst activation. By inactivating the contralateral wulst during recordings of ipsilateral stimulus responses, Bredenkötter and Bischof found indications of a mutual inhibition between the two hemispheres resulting in a suppression of the ipsilateral input that increased during development. They also demonstrated that inactivation of the contralateral eye did not enhance the wulst activation on ipsilateral stimulation. It is thus plausible to assume that the suppression of the ipsilateral input occurs between both hemispheres and is not a pure bottom- up effect. There is no direct connection between the wulst areas of both hemispheres. Therefore, the interaction may probably be accomplished via back projections from the wulst to the thalamic nuclei which also receive retinal input and project bilaterally to the visual wulst. Mutual inhibition, of both hemispheres as proposed here is comparable to interocular inhibitory effects in cats. The idea of a suppression of the less active eye to favour the more important information is thus not restricted to the avian visual system. Our results concerning the role of the ipsilateral input partly support the idea of such a suppression, but only for the "binocular segment" as discussed below. When stimuli were presented in the foveal region (60°), we observed a strong response from the contralateral eye—as expected. In contrast, stimulation of the ipsilateral eye elicited a much weaker and spatially much more restricted wulst activation. The fact that the area of ipsilateral activation was not totally overlapping with the area of activation evoked by the contralateral stimulus, casts doubts on the whole idea of a global mutual suppression of the contralaterally evoked activity by the ipsilateral input. Obviously, the suppression is more specific. However, we are not yet able to provide an idea about how this specificity may be expressed. Due to experimental restrictions, we were not able to examine the direct interaction of simultaneous ipsi- and contralateral eye stimulation with laterally placed stimuli. As mentioned above, the idea of a repression of activation of one eye by that of the other applies for our experiments with frontal stimulation. In these experiments, wulst activity was considerably reduced (Fig –) when both eyes were stimulated simultaneously, compared to trials in which only the contralateral eye was stimulated. Ipsilateral eye stimulation elicited only very weak wulst activation in some of the birds preventing a detailed analyses of map overlap. Schmidt and Bischof have demonstrated lower responses with bilateral stimulation compared with pure contralateral stimuli for the n. rotundus and the entopallium of the zebra finch: about 30% of the neurons recorded in the two areas showed this response pattern, in 40% of the cases, the ipsilateral stimulus enhanced the contralateral one, and there was no effect in the rest of the cases. Thus, the inhibitory interaction of both eyes seems to be a phenomenon observed in the entire visual system of zebra finches and is most probably not restricted to a potential binocular segment. Howarth et al. recently described "binocular" neurons in mice that did not respond to ipsilateral stimulation, but, in contrast to our case, ipsilateral stimuli enhanced responses to the contralateral eye if given simultaneously. Neurons with strongly reduced responses upon binocular stimulation were—however—also reported. Scholl et al. convincingly showed that neurons within the binocular segment are detecting disparities between the left and the right eye, and neuronal responses were either enhanced or diminished (compared to monocular stimulation with the same moving grating) by experimental variation of the disparities. Given the fact that such disparity neurons have been demonstrated in owls, it might be possible that the "binocular segment” in the zebra finch wulst also contains disparity neurons, although—according to behavioural studies—laterally eyed birds do not use the frontal eye field for distance estimation. Rather, as Voss and Bischof have claimed, the binocular overlap may serve to keep the internal representation of the visual scene intact because, if there were no overlap, a frontal "gap" in the visual field might arise with extensive lateral eye movements. Further experiments, showing the response characteristics of single neurons within the representation of the frontal visual field of laterally eyed birds, will be necessary to clarify the role of this part of the zebra finch visual wulst. Our sincere thanks are due to Hans-Ortwin Nalbach who helped with measuring the retinal landmarks, Edda Geissler for preparing the sections for the retinal ganglion cell measurements and also for the assessment of the densities, Joe Voss for processing the results shown in, Michael Siebrecht for providing the Matlab codes for the construction of the iso-azimuth and iso-elevation lines and Aju A Varghese for his technical advice in preparing some of the figures. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: NM SL HJB. Performed the experiments: NM. Analyzed the data: NM SL HJB. Wrote the paper: NM SL HJB.

# Introduction Dendrimers are a new class of nanoscopic containers and delivery devices. They have unique molecular architectures, spherical shapes, interior cavities, excellent monodispersity, and high density of surface functionalities. Dendrimers have high solubility in various organic solvents and water, allowing miscellaneous modifications on dendrimer surface and in-depth characterization by different techniques. These physicochemical properties render dendrimers the applications in host-guest systems. Tomalia *et al*. first reported the diffusion of aspirin molecules into the interior cavities of poly(amidoamine) dendrimers by NMR relaxation measurements. Meijer *et al.* synthesized a dendritic box by modification of poly(propylene imine) (PPI) dendrimer surface with 9-fluorenylmethyloxycarbonyl (FMOC)-protected amino acids and the dendritc box was able to encapsulate guest molecules such as Rose Bengal and *p*-nitrobenzoic acid. The release of *p*-nitrobenzoic acid from the dendritic box can be triggered by removing FMOC protection, while Rose Bengal cannot diffuse out of the dense shell due to its larger size than *p*-nitrobenzoic acid. In a separate study by the Meijer group, palmitic acid modified PPI dendrimers were successfully used for the extraction of anionic xanthene dyes from aqueous solution. These pioneer studies significantly promote the interests of scientific community in dendrimer-based host-guest systems. Our group systematically investigated PAMAM and PPI dendrimer-based host-guest systems using several NMR techniques in the past five years,. NMR techniques have proved to be powerful and informative tools in the characterization of dendrimers and dendrimer/guest complexes. Dendrimers have successive repeat units, well-defined branched structures, highly symmetric frameworks, and spherical shapes like proteins. Repeat unit means rapid chemical shift assignment of the dendrimer protons and much simplified NMR spectrum. Highly symmetric framework and well-defined branched structures allow us to predict the location and orientation of guest molecules within the interior pockets of dendrimers by nuclear Overhauser effect spectroscopy (NOESY). Spherical shape encourages us to characterize the size and morphology of dendrimers and dendrimer/guest complexes or aggregates using diffusion NMR techniques such as pulsed gradient spin echo (PGSE) and diffusion ordered spectroscopy (DOSY). Besides, the rapid technological development in high-field NMR spectroscopy and use of cryogenic probe system significantly increase the sensitivity in NMR analysis, and reduce the time and sample concentration needed in an NMR experiment. When a paramagnetic probe such as nitroxide spin label is in spatial proximity with a host molecule, magnetic dipolar interactions between the unpaired electrons of the paramagnetic probe and the nucleus of interest will result in an increase in nuclear relaxation rates, which is defined as paramagnetic relaxation enhancement (PRE) effect. In comparison with NOE, where the effect is limited to a short range within 5–6 Å, the PRE method can be used to get distance information in the range of 15–24 Å for nitroxide spin label, owing to the large magnetic moment of the unpaired electron. Distance between the spin label and the nucleus of interest can be determined from the increased *R*₂ relaxation rates, allowing quantitatively interpretation of the PRE effect. As a result, PRE NMR is becoming a sensitive and powerful tool to produce long-range site-specific constraints that can complement NOE constrains. In dendrimer-based host-guest systems, if a paramagnetic probe is encapsulated within the dendrimer interior pockets, the PRE effect will lead to decreased NMR signal intensities and increased peak linewidths of dendrimer protons close to the paramagnetic probe in a ¹H NMR spectrum, which is an important supplement to the NOESY techniques in the investigation of dendrimer-based host- guest systems. Previously, Newkome *et al*. investigated dendrimer structure by different NMR techniques using paramagnetic cobalt (II) as a probe. In this study, 2,2,6,6-tetramethylpiperidinyloxy (TEMPO) derivatives were used as paramagnetic guests. PAMAM dendrimers bearing hydroxyl, primary amine, and carboxylate groups were used as model hosts. Though the interactions of TEMPO derivatives and PAMAM dendrimers were previously investigated by Turro and Tomalia *et al*, they use an electron spin resonance (ESR) technique to analyze the dendrimer/TEMPO complexes. The major goal of this study is to reveal the host behaviors of PAMAM dendrimers with different surface functionalities by PRE NMR. # Experimental ## Materials Generation 4 (G4) ethylenediamine (EDA)-cored and primary amine-terminated PAMAM dendrimer (G4-NH₂), G4 EDA-cored and hydroxyl-terminated PAMAM dendrimer (G4-OH), and G3.5 EDA-cored and carboxylate-terminated PAMAM dendrimer (G3.5-COONa) were purchased from Dendritech (Midland, MI). 4-amino-TEMPO (TEMPO- NH₂) and 4-carboxy-TEMPO (TEMPO-COOH) were purchased from Sigma- Aldrich (St. Louis, MO). 4-hydroxy-TEMPO (TEMPO-OH) was obtained from Aladdin Chemistry Co. Ltd (Shanghai, China). Deuterium oxide (D₂O) was purchased from Beijing Chongxi High-Tech Incubator Co. Ltd. (Beijing, China). Dioxane and deuterated dimethyl sulfoxide (d6-DMSO) were purchased from Aladdin Chemistry Co. Ltd (Shanghai, China). Dendrimers were received in methanol and the solvents were distilled to obtain the products as white gels. All other chemicals were used as received without further purification. ## Sample preparation G4-NH₂, G4-OH, and G3.5-COONa were dissolved in D₂O at a concentration of 20 mg/mL. TEMPO-NH₂, TEMPO-COOH, and TEMPO-OH were dissolved in d6-DMSO at a concentration of 5 mg/mL. The stock solutions were stored at 4°C before NMR studies. The dendrimer/TEMPO complex solutions were prepared at different TEMPO/dendrimer molar ratios in 500 µL D₂O and d6-DMSO mixtures (V/V, 80/20). Dendrimer concentrations in the complex solutions are fixed at 1 mg/mL and the molar ratios of TEMPO/dendrimer are 0, 1, 2, 4, 8, 12, 16, 24, 32, 40, 48, and 64, respectively. Each sample contains a certain amount of dioxane which was used as internal standard. The samples were sonicated for 2 h before NMR analysis and the pH condition of each sample was measured immediately after the NMR experiments. ## NMR analysis ¹H NMR spectra of the PAMAM dendrimer and TEMPO complex solutions were obtained on a Varian 699.804 MHz NMR spectrometer at 298.2±0.1 K. The peak linewidths in ¹H NMR spectrum were measured using the VNMRJ software. # Results and Discussion ## Interactions of TEMPO-NH₂ and TEMPO-OH with three types of PAMAM dendrimers The host-guest chemistry of TEMPO-NH₂ and TEMPO-OH with PAMAM dendrimers bearing amine, hydroxyl, and carboxylate groups were investigated by ¹H NMR. G4-NH₂, G4-OH, and G3.5-COONa PAMAM dendrimers were used because these dendrimers have the same numbers of surface functionalities (64) and similar molecular size. Besides, these PAMAM dendrimers were reported to have interior hydrophobic pockets and open surface structures, which are excellent candidates in host-guest systems. As shown in, G4-NH₂ has four ¹H NMR peaks: H_{d, d'} at 3.26 ppm, H_{c, b'} at 2.79 ppm, H_b at 2.60 ppm, and H_a at 2.40 ppm, respectively. The peak at 3.72 ppm corresponds to the internal standard dioxane, while peaks at 3.32 ppm and 2.66 ppm arise from the residual methanol and DMSO, respectively. Both the chemical shift and linewidth of the G4-NH₂ peak were scarcely affected by the addition of TEMPO-OH. Previous studies have demonstrated that ionic binding of anionic guests on the surface of amine-terminated PAMAM dendrimer shifts the resonance signals of dendrimer surface (H_b' and H_d') to higher frequencies, and hydrophobic encapsulation of guests within dendrimer interior pockets shifts the signals of dendrimer interior (H_a∼d) to lower frequencies. The absence of H_a∼d signal shifts in indicates the absence of ionic binding and hydrophobic encapsulation. In addition, TEMPO-OH is a paramagnetic guest molecule. If TEMPO-OH molecules are encapsulated within the G4-NH₂ interior cavities via hydrophobic or ionic interactions, the PRE effect will lead to significant enhancement of linewidths of G4-NH₂ resonance signals. However, no significant variation in linewidth of G4-NH₂ signals during the addition of TEMPO-OH in suggests the absence of TEMPO-OH encapsulation within G4-NH₂. Similarly, when TEMPO-NH₂ was added into G4-NH₂, no significant change in the ¹H NMR spectrum of G4-NH₂ was observed. The only difference is the slight shifts of H_b' and H_d' to lower frequency when increasing TEMPO- NH₂/G4-NH₂ molar ratio. During the addition of TEMPO- NH₂ into G4-NH₂, the pH value of the dendrimer solution increases from 8.72 to 9.59 when the molar ratio of TEMPO- NH₂/G4-NH₂ reaches 64. Deprotonation of the NH₃⁺ groups (pKa∼10.0) on G4-NH₂ surface well explains the slight shifts of H_b' and H_d' signals in. The absence of G4-NH₂ linewidth variation during TEMPO-NH₂ titration also confirms the absence of TEMPO-NH₂ encapsulation within G4-NH₂. Similar phenomenons are observed between TEMPO-NH₂ (TEMPO-OH) and G4-OH (G3.5-COONa). These results suggest that no host-guest interaction occurs between TEMPO-NH₂ (TEMPO-OH) and the three types of PAMAM dendrimers. TEMPO-NH₂ (171 Da) and TEMPO-OH (172 Da) have dendrimer pocket- matched sizes. Our previous studies have reported successful encapsulations of much larger guests such as sodium dodecyl sulfate (288 Da), dexamethasone 21-phosphate (516 Da), and Congo red (696 Da) within G4-NH₂ PAMAM dendrimer. Therefore, the failure of TEMPO-NH₂ and TEMPO-OH encapsulation within PAMAM dendrimers should not be attributed to size limitations of the guests, but probably to the reason that PAMAM dendrimer pockets are not hydrophobic enough to encapsulate TEMPO-NH₂ or TEMPO- OH in d6-DMSO/D₂O (20%/80%, V/V) via hydrophobic interactions. ## Interactions of TEMPO-COOH with G4-NH₂ and G4-OH PAMAM dendrimers As shown in, signal intensities of all the G4-NH₂ resonance signals (H_a∼d, H_b' and H_d') are gradually decreased with increasing TEMPO-COOH/G4-NH₂ molar ratios. Also, the linewidths of the signals are much broader after the addition of TEMPO-COOH molecules. This phenomenon can be recognized as an evidence of TEMPO-COOH encapsulation within G4-NH₂. One may argue that decreased molecular mobility or increased molecular size also contribute to increased peak linewidth. In the system of G4-NH₂ and TEMPO-COOH, this effect is not significant as compared to the PRE effect. Take G4-NH₂ for an example, its diffusion coefficient is slightly changed after the addition of 8 or 16 molar ratios of TEMPO-COOH. When more TEMPO-COOH were added (molar ratio of 32 or 64), the peak intensities of G4-NH₂ are extremely low due to the PRE effect and the diffusion coefficients of G4-NH₂ are not available using pulsed gradient spin- echo (PGSE) NMR. In this case, 32 molar ratios of deoxycholate which was proved to interact with G4-NH₂ were used instead of TEMPO-COOH. Again, only a slight change (∼3%) on the diffusion coefficient of G4-NH₂ was observed before and after the addition of deoxycholate. These results suggest that the increased peak linewidths in are mainly caused by the PRE effect of TEMPO-COOH encapsulated within G4-NH₂ pockets rather than the increased dendrimer size after TEMPO-COOH addition. Besides variations in signal intensity and linewidth, shifts of H_b' and H_d' resonance signals to higher frequencies in were observed. This is due to ionic interactions between deprotonated TEMPO-COOH and cationic charged G4-NH₂ surface,. The pH value of TEMPO- COOH/G4-NH₂ solution decreases to 5.61 when the molar ratio of TEMPO- COOH and G4-NH₂ reaches 64. As the pKa values of surface amine and interior tertiary amine groups of PAMAM dendrimers are around 10.0 and 6.5 respectively. All the surface amine groups and partial interior tertiary amine groups of G4-NH₂ should be positively charged when the molar ratio of TEMPO-COOH and G4-NH₂ is above 40 (pH 6.81). In such a complex solution, TEMPO-COOH is negatively charged and can be either bound on the surface or encapsulated within the interior of G4-NH₂ via ionic interactions. To exclude the possibility that the decreased signal intensity and the broadened peak linewidth are caused by amine protonation rather than the PRE effect, we titrated G4-NH₂ with acetic acid. As shown in, almost no change in signal intensity and peak linewidth of H_a–d are observed. The changes in linewidths of H_b' and H_d' in can be explained by exchanges of NH₃⁺ and NH₂ on G4-NH₂ surface during the addition of acetic acid. We also used NOESY to prove the encapsulation of TEMPO-COOH within G4-NH₂. As TEMPO-COOH is a paramagnetic molecule, the PRE effect broadens both the G4-NH2 and the TEMPO-COOH peaks as shown in which makes it difficult to detect the cross-peaks of between G4-NH₂ and TEMP-COOH using a NOESY method. In this case, we treated the TEMPO-COOH with 12 N HCl to scavenge the TEMPO radical. The resulting guest was complexed with G4-NH₂ at a molar ratio of 32∶1 and the complex was analyzed by ¹H-¹H NOESY at a mixing time of 300 ms. As shown in, weak NOE cross-peaks between the methyl groups of the guest and the methylene protons (H_{a, c}) of G4-NH₂ were observed, indicating the encapsulation of the TEMPO-COOH derivative within G4-NH₂. This result indicates that the PRE NMR method can be used as an alternative method to NOESY in the investigation of inclusion structures. The PRE method using ¹H NMR to probe host-guest interaction is more facile and sensitive than the NOESY method. When G4-OH was titrated with TEMPO-COOH or acetic acid, the protonation of the interior tertiary amine groups of G4-OH also caused the broadening of resonance signals for adjacent methylene protons such as H_a–d. This makes it difficult to analyze the PRE effect in the system. However, the variations of linewidth for H_b' in the samples titrated with TEMPO-COOH are much higher than that in the samples titrated with acetic acid, indicating the presence of PRE effect in TEMPO-COOH/G4-OH. Since the surface of G4-OH is non- charged, no ionic interaction occurs between TEMPO-COOH and G4-OH surface. TEMPO-COOH molecules should be at least encapsulated within the outermost layer of G4-OH dendrimer. Interior methylene protons (H_a–d) of G4-OH dendrimer exhibit significant higher frequency shift, while surface methylene protons of G4-OH such as H_b' and H_d' are slightly shifted, suggesting quaternization of the interior tertiary amine groups of G4-OH (pH ranges from 8.11 to 5.10). The formation of inclusion complexes of TEMPO-COOH with G4-NH₂ or G4-OH should be driven by ionic interactions rather than hydrophobic interactions. This property is distinct from previous dendrimer/guest inclusion complexes such as PAMAM dendrimers with mycophenolic acid, sodium deoxycholate and sodium dodecyl sulfate, and phenylbutazone, in which the encapsulations are driven by hydrophobic interactions. ## Interaction of TEMPO-COOH with G3.5-COONa As shown in, titration of G3.5-COONa with TEMPO-COOH leads to higher frequency shifts of all the dendrimer protons. During this period, the pH value of the complex solution gradually decreases from 9.04 to 5.45, which means partial protonation of interior tertiary amine groups within G3.5-COONa. Interestingly, the signal intensities of peaks H_{a', b', c', d'} increase during the addition of TEMPO-COOH. Meanwhile, the linewidths of these peaks become lower. This result is not in accordance with those observed in TEMPO- COOH/G4-NH₂ and TEMPO-COOH/G4-OH. This phenomenon can be explained by the equilibrium equation in. Due to the presence of carboxylate groups on G3.5-COONa surface, the tertiary amine group located on the outermost layer has a much higher pKa value (8.0∼9.0) than the interior teriary amine groups (pKa∼6.5). Considering the initial G3.5-COONa solution has a pH value around 8.7, the exchange equilibrium between the anion (A) and the zwitterion (B) in caused the peaks of H_{a', b', c', d'} broaden in ¹H NMR spectrum. With the increase of TEMPO-COOH concentration, the pH value of the G3.5-COONa decreases to 6.0 and increased percent of the surface tertiary amines are protonated. The zwitterion (B) is the major form in the solution at high molar ratios of TEMPO-COOH and G3.5-COONa. Therefore, the peaks H_{a', b', c', d'} become narrow with the addition of TEMPO-COOH. This phenomenon is also observed when amino acids were titrated with acid. In the case of interior methylene protons such as H_a–d, the interior tertiary amine (pKa∼6.5) is not protonated at the beginning, and start to protonate after the addition of TEMPO, the exchange between protonated and non-protonated state caused the peaks H_a–d adjacent to these tertiary amine groups becoming broaden. Therefore, the linewidth of the peaks is controlled by exchange modulation of the tertiary amine groups rather than the PRE effect. To prove this speculation, we titrated G3.5-COONa dendrimer with acetic acid. As shown in, the same trend for linewidth of G3.5-COONa peaks was observed when this dendrimer was titrated with TEMPO-COOH and acetic acid. During this period, no PRE effect was observed between the protons (H_a', H_d') and the unpaired electron in the nitroxide radical. The failure of TEMPO-COO⁻ within G3.5-COONa cavities is probably due to the repulsion of TEMPO-COO⁻ ions and the anionic G3.5 surface. # Conclusions In this study, we investigated the host-guest chemistry of PAMAM dendrimers with amine, hydroxyl, and carboxylate surface functionalities and TEMPO radicals with amine, hydroxyl, and carboxyl groups using ¹H NMR titration experiments. PRE effects were observed in TEMPO-COOH/G4-NH₂ and TEMPO-COOH/G4-OH complexes. The inclusion complexes of TEMPO-COOH and G4-NH₂ (G4-OH) are driven by ionic interactions rather than hydrophobic interactions. In other dendrimer/TEMPO systems, only protonation or deprotonation of the surface functionalities and interior tertiary amine groups occur. The absence of TEMPO-COOH encapsulation within G3.5-COONa is due to the repulsion of TEMPO-COO- anion and the negatively charged G3.5-COONa surface. This study proved that PRE NMR is a powerful method in the investigation of dendrimer-based host-guest interactions. # Supporting Information [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YYC. Performed the experiments: FW NMS. Analyzed the data: YYC FW NMS. Contributed reagents/materials/analysis tools: YYC. Wrote the paper: YYC FW NMS.

# Introduction Serious concerns expressed during the last two decades regarding the use of chemicals in agriculture in terms of their adverse impact on the human health, environment and sustainable agricultural production. These concerns have led to the worldwide promotion of organic food production. The market for organic food products is growing rapidly worldwide, such foods meet certified organic standards for production, handling, processing, and marketing. With increasing awareness on safety consumption, healthy lifestyle and environmental protection, the competition within green organic food industry is becoming increasingly fierce. Developing the green industry is China’s national strategy to achieve a sustainable path and prevent further degradation of the environment. In 2021, the No.1 Central Document has put forward some recommendations for the development of green agricultural products and organic agricultural products through quality certifications of edible agricultural products. As the most important milestone in the industry and green environmental protection industry in the 21st century, forest green food industry has gradually become the mainstream direction of today’s green industry, which provides support for sustainable development of China’s forestry and has become a new economic growth point. Yunnan Province, which contains abundant forest resources in China, has issued a number of relevant policies in recent years to promote the development of forest green food industry and to enhance competitiveness of forest green food industry in Yunnan Province. Both in theoretical research and empirical research, it is shown that organic food industries in Europe, America and other countries had quite complete systems. Although there were differences in the degree of development, market operation and market management of organic food industry in developed countries are characterized by standardized operation, strict management and perfect industrial system of organic food market. The theoretical research on organic food was more in-depth and systematic, forming a relatively complete theoretical system in the aspects of organic food consumer behavior, market competition, policy support and so on. International studies have classified forest green food as organic food without independent analysis and certification. However, the research results of organic food theory had a good reference and guidance for the development of forest green food industry. The existing research has mainly carried out from four aspects: concept classification, consumption preference, policy support and industrial competitiveness evaluation. In terms of concept classification, the research on forest food began in the early 20th century. King published the book "four thousand years of farmers", which pointed out that human and animal manure and straw used in China’s agricultural planting are good organic fertilizers, and the fungal food cultivated with this fertilizer is the earliest organic food. Norikane used "big business" to describe the industrial status of forest food industry in the forest. Forest food was also known as "safe food" in Liberia. Forest food was an emerging multi-functional perennial mixed food production system. By solving the three main components of food security: supply, access and utilization, it is possible to promote food security and alleviate malnutrition in urban and suburban areas. Only forest food that has passed social certification, organic certification, product quality certification and international forest certification has qualified forest food. Forest food had an important contribution to the quality of human diet. According to the research, there were four use modes of forest food: forest food dependence, limited use of forest food, forest food supplement and professional consumers of forest food. Green Food was officially defined to include food grown under the conditions of strict supervision, control and regulation in production, processing, packing, storage and transportation. Forest green foods were a subset of forest foods and refer to as uncultivated foods from forested areas. Forest green food was mostly in wild states, rich in nutrition, delicious tastes, clean and pollution-free features which meet the growing demand for green and organic food. Green food adopted the whole-some quality control from field to table, while it required reasonable applications of inputs, including pesticide, fertilizer, veterinary drug and additive etc. to prevent any pollution of toxic and harmful matters in various stages of food processing so as to ensure environmental and product safety. Forest green food industry referred to as the industry that cultivates, processes, circulates, sells and provides related services for forest green food based on sustainable utilization of forest resources and good ecological environments, which included multiple categories of primary industry, secondary industry and tertiary industry of national economy. In terms of consumption preference, consumers of organic food were the guide of consumption opinions and should pay attention to the oral effect among consumers. Most consumers believed that organic food has health and environmental benefits, but lack of information, inappropriate marketing practices and high price are some of barriers that hinder the consumption of organic food. Different store business models and different types of families had different price sensitivity of organic food, which would form specific family purchase loyalty and affect each family’s willingness to pay for organic food. Because consumers were more rational when consuming organic food, they are more sensitive to the price of organic food, and were willing to pay more for expensive organic food. In fact, where there were only small visual and sensory differences between organic products and conventional products, consumers could only choose a product based on trust in the producers, so reputation was particularly important in the organic food system. Some organic food consumers have preference for food quality and safety, and sometimes offer suggestions for improvement in supply chains to promote the sales of organic food. From the policy support aspect in recent years, the sustainable development of forest food could drive the sustainable development of forest resources, which is essential in attainment of different sustainable development goals. The introduction of No.1 Central Document has provided a strategic direction for the development of forest green food industry. The development of forest green food industry is still at the primary stage, so there is was the need to seize opportunities at national macro policy level, and establish a forest green food development system with regional characteristics as soon as possible. Nigeria has also put forward many relevant policies aimed at accelerating the sustainable development, management and environmental protection of forests, but the expected results have not been achieved. In 2013, the national food security act of India marked a historic change in food security from a welfare based approach to a rights based approach. However, Bose’s research showed that the public distribution system of the act was not suitable for ensuring the food security of indigenous people and did not promote the development of forest food industry. In terms of industrial competitiveness evaluation, previous studies usually divided the development level and development potential of forestry industry into four types: "high level—high potential", "high level—low potential", "low level—high potential" and "low level—low potential". On this basis, the research results showed that the development potential of plantation industry was huge, and the production potential of forestry industry in most areas was greater than the sales potential. Industrial development potential was an important index to evaluate the degree of industrial development. It could improve the development potential of forestry industry from three aspects: brand construction, trust maintenance and service. Whether a domestic enterprise could give play to its competitive advantage was greatly affected by the domestic economic environment. Among them, production factors, demand, related industries and supporting industries, as well as its enterprise strategic combination and competition were the most influential and direct factors, in which opportunity and the government played an auxiliary role. With the continuous improvement of product quality standards, high-tech organic food had more advantages in market competition. Song analyzed the actual situation by calculating the index parameters in the diamond industry analysis system, and concluded Heilongjiang has comparative advantage in green organic food industry. Wang et al. established a competitive advantage evaluation model based on diamond system theory to evaluate the competitiveness of forest green food industry in Heilongjiang Province. Liu et al. analyzed the international competitiveness of China’s nut forest food by using international market share rate, trade competition index and revealed symmetric comparative advantage index. Raynolds believed that the competitive advantage of the organic food industry could be achieved through product innovation. At both domestic and international levels, scholars have made rich contributions on forest green food industry, but there is the need to further explore the topic from territorial perspectives. At present, domestic and foreign scholars have studied the competitiveness of forest green food industry from the national perspective, and studies based on scale of Yunnan Province are still needed. In view of this, the study looks at the connotation and characteristics of forest green food industry constructing competitiveness evaluation index system of forest green food industry from four aspects: production factors, demand conditions, enterprise strategy, structure and horizontal competition, and related and supporting industries, and made empirical analysis by using fuzzy evaluation method and AHP. The study focuses mainly on the competitiveness of forest green food industry in Yunnan Province by analyzing the competitiveness index of key factor level, general factor level and important factor level criteria factors. # Methods Firstly, the evaluation index system of forest green food industry competitiveness is first constructed. Fuzzy evaluation method and AHP are used for the empirical analysis of the competitiveness of forest green food industry in Yunnan Province. Fuzzy evaluation method has the characteristics of clear results and strong systematicness. It can better solve fuzzy and difficult to quantify problems, and is suitable for solving various uncertain problems. AHP can effectively analyze the non sequential relationship between the levels of the objective index system, and effectively measure the judgment and comparison of decision makers. Therefore, these two methods are selected to evaluate the competitiveness of forest green food industry in Yunnan Province clearly and objectively, and using the two methods at the same time can enhance the authenticity and reliability of the evaluation results. ## Study area The study is carried out in Yunnan province, China. Yunnan is located at 21°8’ - 29°15’ N and 97°31’ - 106°11’ E, and borders Guizhou province, Sichuan province, and Guangxi province. With a total area of 394,100 km², it contains 16 prefecture-level administrative divisions. The current resident population of Yunnan Province is 48.583 million. Yunnan is a low latitude region with high terrain in the northwest and low terrain in the southeast. It falls step by step from north to south. It is a mountainous plateau terrain, and the mountainous area accounts for 88.64% of the total area of the province. With gorgeous natural scenery, it is one of the important birthplaces of human civilization. Yunnan Province is extremely rich in forest food resources, with many kinds and wide distribution. According to the records of Yunnan statistical yearbook, in 2020, the forest area of the whole province reached 23.9265 million hm², the forest coverage rate of the whole province reached 62.4%, and the forestry output value reached 39.554 billion yuan. ## Data sources The competitive advantage evaluation system of forest green food industry includes four primary indicators: production factors, demand conditions, enterprise strategy, structure and horizontal competition, and related and supporting industries, and sixteen secondary indicators such as total area, investment amount and number of employees. The data of secondary indicators directly affect the measurement of the competitiveness of forest green food industry in Yunnan Province. The data used for the analysis span from 2016 to 2020. Secondary indicators include "total value", "qualitative value", "ratio value" and "quantitative value". Where, "total value" is the sum of annual data of the indicator; "qualitative value" refers to the score given by experts; "ratio value" is the ratio of indicators; "quantitative value" is the value of the indicator. "total value" indicators such as total area and investment amount and "quantitative value" indicators such as technical status are directly obtained from Yunnan Statistical Yearbook; natural endowment, food safety and other "qualitative value" indicators are obtained from the investigation of relevant experts; "ratio value" indicators such as per capital disposable income and market share are sorted out according to Yunnan statistical yearbook. ## Forest green food industry competitiveness model The index system of forest green food industry competitiveness evaluation is constructed by considering the indexes used in forest green food industry competitiveness and other related industries competitiveness evaluations which are frequently used. Finally, four indexes are selected which include production factors, demand conditions, enterprise strategy, structure and horizontal competition, and related and supporting industries. The comprehensive evaluation of competitive advantage is divided into general factor layer and specific factor layer. The general factor layer, namely the main criterion layer B, which is referred to as the first level index; the specific factor level which is referred to as sub criteria layer C, is also known as secondary indicators. There are two ways to determine the score of specific factor level. First, if the relevant statistical index value can be obtained directly, the score can be determined by referring to the corresponding score interval. Second, if it is difficult to obtain data through relevant indicators, experts will score them. Based on this the latter, the index system structure model of forest green food industry competitive advantage and identification index are constructed. Firstly, membership function is established. We use X = (*X*₁, *X*₂, *X*₃, *X*₄) to describe the competitive advantage of forest green food industry, and evaluate the industry’s competitive advantage in each factor. Suppose the evaluation set is Y = {*Y*₁, *Y*₂, *Y*₃}, where Y is the fuzzy comprehensive evaluation of the industry’s competitive advantage from the perspective of secondary index X_*ij*. The membership functions of the three grades are as follows: $$U_{Y_{1}}\ (\delta) = \left\{ \begin{matrix} {\frac{1}{15}\ (\delta - 85),85 \leq \delta \leq 100} \\ {0,\delta < 85} \\ \end{matrix} \right.$$ $$U_{Y_{2}}\ (\delta) = \left\{ \begin{matrix} {- \frac{1}{15}\ (\delta - 85),70 \leq \delta \leq 85} \\ {\frac{1}{10}\ (\delta - 60),\delta < 70} \\ \end{matrix} \right.$$ $$U_{Y_{3}}\ (\delta) = \left\{ \begin{matrix} {- \frac{1}{10}\ (\delta - 70),60 \leq \delta \leq 70} \\ {1,\delta < 60} \\ \end{matrix} \right.$$ Secondly, the fuzzy evaluation matrix is established. The main criteria layer fuzzy evaluation matrix is established to determine the fuzzy evaluation results of the evaluation object: $$\overline{A} = \left( a_{1},a_{2},a_{3} \right) = B \times \begin{bmatrix} A_{1} \\ \begin{matrix} A_{2} \\ A_{3} \\ \end{matrix} \\ A_{4} \\ \end{bmatrix},\ a_{j} = {\bigvee_{i = 1}^{3}{(B_{i}}}{\bigwedge_{i = 1}^{4}{a_{ij})}}$$ Normalize $\overline{A}$ and use fuzzy distribution method to evaluate industrial competitive advantage, and the final score of fuzzy evaluation of industrial competitive advantage is obtained. After analyzing the industrial competitive advantage, its strength is further assessed. On the one hand, it can objectively evaluate the competitive advantage of the industry. On the other hand, with more important observations, it can assess both constraints and successes of the industry and improve the competitive advantage through the evaluation of the intensity. Through the attributes of each evaluation index, it can appropriately adjust the business activities to further optimize and improve the competitive advantage of the industry. ## Empirical analysis of Yunnan Taking Yunnan Province as a sample, the reasons for testing and analyzing the competitive advantage model of forest green food industry are: firstly, the forest green food industry in Yunnan Province is developed earlier and the data is relatively abundant; secondly, the forest green food industry in Yunnan Province has developed rapidly and has representative characteristics of the whole country. ## Descriptive statistics The quantitative data classified as "total value" "ratio value" and "quantitative value" in the index system are from the statistical yearbook of Yunnan Province as shown in. The weight of each index of the impact assessment object and the evaluation value of "qualitative value" in the index system are all obtained through questionnaire administering involving forest green food experts in Yunnan Province. Through the questionnaire survey of experts, the relevant evaluation values are obtained, and the survey results sorted out by taking the average value. Using the formulas— to analyze the membership of the secondary factors. Then the fuzzy evaluation matrix R_i is: $$R_{1} = \begin{bmatrix} 0 & 0.33 & 0 \\ 0 & 0.67 & 0 \\ 0.33 & 0 & 0 \\ 0 & 0.67 & 0 \\ \end{bmatrix}\qquad R_{2} = \begin{bmatrix} 0.33 & 0 & 0 \\ 0 & 0.67 & 0 \\ 0 & 0.33 & 0 \\ 0.33 & 0 & 0 \\ \end{bmatrix}$$ $$R_{3} = \left\lbrack \begin{array}{lll} 0 & 0.67 & 0 \\ 0 & 0.67 & 0 \\ 0 & 0.33 & 0 \\ 0 & 0.33 & 0 \\ \end{array} \right\rbrack\qquad R_{4} = \begin{bmatrix} 0 & 0 & 0.5 \\ 0.33 & 0 & 0 \\ 0.33 & 0 & 0 \\ 0 & 0.33 & 0 \\ \end{bmatrix}$$ By *A*_*i* = *W*_*ij*×*R*_*i* can be calculated as follows: A₁ = (0.13,0.30,0); A₂ = (0.13,0.27,0); A₃ = (0,0.57,0); A₄ = (0.26,0.03,0.05) Then, the fuzzy relation matrix $\overline{A}$ of evaluation target is constructed by using formula: $$\overline{A} = (0.4,0.3,0.1,0.2) \times \begin{bmatrix} 0.13 & 0.30 & 0 \\ 0.13 & 0.27 & 0 \\ 0 & 0.57 & 0 \\ 0.26 & 0.03 & 0.05 \\ \end{bmatrix} = (0.15,0.26,0.01)$$ And normalize $\overline{A}$, $${\overline{A}}^{\prime} = \begin{bmatrix} \frac{0.15}{0.15 + 0.26 + 0.01} & \frac{0.26}{0.15 + 0.26 + 0.01} & \frac{0.01}{0.15 + 0.26 + 0.01} \\ \end{bmatrix} = \left\lbrack {0.35\mspace{14mu} 0.63\mspace{14mu} 0.02} \right\rbrack$$ Based on this result, there is a 35% certainty that Yunnan forest green food industry has strong competitive advantage, 63% certainty that its competitive advantage is average, and 2% certainty that its competitive advantage is weak. According to the principle of maximum membership degree, it can be considered that the competitive advantage of forest green food industry in Yunnan Province is at the general level. The determination of evaluation index weight directly reflects the role of indicators in the development of enterprises, and directly affects the evaluation results under certain conditions. Therefore, the determination of index weight is the key to scientific identification and evaluation of competitive advantage. AHP is used to determine the weight of each index. The judgment matrix is constructed based on expert opinions, and the sum product method is used to solve the comprehensive judgment matrix. The weight of each index is determined by combining expert experience and mathematical model (Tables). Each matrix passed the consistency test of CR\<0.1 fulfilling the condition for calculation of the weight of each index to the target layer. The index that can specifically measure the competitiveness of forest green food industry in Yunnan Province is been calculated. A study by Wang et al. measured the competitiveness of forest green food industry in Yunnan Province and obtained 85–100, 70–85 and 60–70 for high level, medium level and low level, respectively. ## Result analysis The influence degree of sub criteria factors on industrial competitive advantage is measured by the weight value, which is divided into three categories: key factors (\>0.1), important factors (0.03–0.1) and general factors (\<0.03). By calculating the product of the influence weight of each factor on the target and the dimensionless value of each factor, the competitiveness index of each factor is obtained, which is the competitiveness of the final evaluation target. The result of AHP shows that the comprehensive competitiveness index of forest green food industry in Yunnan Province is 83.98, which is at the middle level. It is similar to the evaluation result under the principle of maximum membership degree of fuzzy comprehensive evaluation, and the conclusion that 63% membership degree is used to judge the competitive advantage of forest green food industry in Yunnan Province is same as the result of AHP, all of them are in the medium level. # Conclusion and suggestion Taking Yunnan Province as the study area, the evaluation index system of forest green food industry competitiveness based on four dimensions which include production factors, demand conditions, enterprise strategy, structure and horizontal competition, and related and supporting industries, and makes recommendations for policy consideration. ## Conclusion The above empirical analysis of the fuzzy evaluation model and the AHP model for the competitive advantage of the forest green food industry in Yunnan Province showed that the industrial competitive advantage model based on the competitive advantage theory has strong pertinence and applicability in analyzing the competitive advantage of the growing forest green food industry. The model evaluation results showed that it is fully in line with the competitive advantage level of forest green food industry in Yunnan Province. After more than 20 years of efforts, the forest green food industry in Yunnan Province developed rapidly and was entering a new period of sustained and accelerated development. However, from the low export rate of forest green food in Yunnan Province, the competitive advantage of forest green food industry in Yunnan Province mainly came from its good ecological environment, large monitoring area and planting area, and extensive management, and these factors would no longer have a significant competitive advantage with the development of other regions, so it was difficult to sustain. At present, the overall competitiveness index of forest green food industry in Yunnan Province was 83.98, which indication that it has a strong competitive advantage in the country; among the influencing factors, the competitiveness indexes of natural endowment, total area and education level of the key factors were 27.90, 11.68 and 12.32, respectively, an indication of comparative advantage, while the competitiveness indexes of per capita disposable income, investment amount of the important factors and industrial environment of the general factors were 6.12, 5.03 and 0.56, respectively, showing that green food industry has no comparative advantage in those areas. Industry in Yunnan Province mainly depended on the critical factor, the sub criterion factors of general factors and important factors are at a disadvantage. There was no major inferior item in the sub criterion factors of production factor level. The sub criteria of the demand condition level, which was the second most important level, were not high as a whole, and individual projects had serious disadvantages. The sub criterion factors of the related supporting industries with the third importance level fluctuated greatly. Therefore, it is possible that the most important factor for Yunnan forest green food to gain competitive advantage is the competitiveness of key factors, but the influence of non-key factors could not be ignored. The next step to gain competitive advantages should start with the promotion of all important factors. Most of the important factors could be improved through deliberate policies, which was different from the key factors, which mainly relied on resource monopoly. Although general factors had an impact on competitive advantage, it was difficult to play a decisive role because of its low weight. ## Suggestion Though the forest green food industry in Yunnan Province has a strong competitive advantage, the per capita disposable income, the amount of investment and the competitiveness index of industrial environment in general factors were areas that green food industry has no comparative advantage. Based on this, regulations should be forward for consideration to encourage investments in the sector. ### Policy support Firstly, there is the need to formulate and implement targeted industrial policies. The competitive advantage of forest green food industry in Yunnan Province mainly comes from the advantages of natural resources, such as variety, yield and location. It is not difficult to maintain these advantages at present; the difficulty is to transform the existing advantages of key factors into long- term advantages and eliminate the instability of short-term advantages. It is necessary to expand the competitive advantage of all the important factors and make these advantages play a positive role in promoting the positive development of relevant competitive factors. These factors mainly depend on the government, system, science and technology. Governments at all levels should adopt targeted industrial policies according to their geographical location and product characteristics, promote the development of industries and enterprises, and improve the comprehensive competitiveness of the industry. Secondly, there is the need to strengthen the development of relevant supporting industries. The value of related industries and supporting industries lies not only in the ability to provide input for the leading industry with the lowest price, but also the geographical distance with the leading industry, which enables enterprises to transmit product information frequently and quickly, exchange innovation ideas, promote the technological upgrading of enterprises and create a benign and interactive "local economy". ### Enterprise development Firstly, using marketing strategy to stimulate demands. Because of the lack of general publicity of forest green food and the consumers’ lack of understanding of the value of forest green food, it is unable to form a stable concept of forest green food consumption, which limits the sectors’ growth. The public has a high awareness of the word "green food", but many people do not know about forest green food. Therefore, there is the need to meet the existing needs through objective and practical advertising and marketing strategies, and strengthen the guidance of consumers’ demands, so that the consumption awareness of forest green food is deeply rooted in the whole public. Secondly, foster the concept of cooperation to promote win-win development. In the process of market internationalization, whether the products enter the international market and continue to expand the market coverage and share are important indicators to test the overall level of industry and competitiveness of enterprises. In this process, it is very important to cultivate the concept of cooperation and competition, and promote the development strategy of win-win scenarios. At present, the low market concentration, small enterprise scale, less well-known brand and weak market competitiveness of China forest green food industry are directly related to the cultivation and development of cooperative competition concept. In order to make China’s forest green food enter the international market, it is necessary to innovate. Through cooperation and competition, optimize product structure and develop domestic and international markets. # Supporting information The data in this paper are from the statistical yearbook of Yunnan Province. Thank the statisticians for their work. [^1]: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

# Introduction Globalization and emerging markets create huge internationalization opportunities for domestic franchisors. Franchising is a promising and increasingly used organizational form to improve strategic, organizational, professional and client-related results in many sectors. Franchising is a business format which provides firms with an excellent opportunity to expand their business to international markets. The saturation of the domestic markets, especially in the developed countries, has forced firms to seek new markets. When establishing their internationalization strategy, firms can choose among different forms—direct investments, collaboration agreements, or franchising, among others. Franchising is one of the international expansion strategies most preferred by many firms. Franchising is a growth model for international franchisors to expand abroad that involves minimal financial risk and a quick go-to-market strategy. The attractiveness of franchising for international expansion is mainly due to its requiring a low level of capital investment and there being an adequate control of its distribution processes. Significant changes have taken place in recent decades in the economies of many industrialized countries through firms opening to new markets. This unstoppable process of internationalization has not only been carried out by large operators but also by medium and small retailers. Although the international franchising industry has become a major player in the development of countries, it has been uneven. The globalization prevailing nowadays means that there is an intensification of trade in local markets and greater levels of competiveness. This causes firms to have to seek new markets in order to be able to survive. In the specific case of Spain, joining the community project of the European Union has had an effect on the growing external Spanish business fabric. This increase in the number of Spanish firms in European markets has provided these firms with important experience. This has contributed to their presence in other markets, apart from those of Europe, such as those in the American and Asian continents, among others. In this way, the number of Spanish firms in the international markets of goods and services has progressively increased. Indeed, the international dimension has become a relevant part of these firms’ strategies. In the specific case of franchising, Spanish franchisors have used internationalization as a strategic growth option and as one of the best ways to increase the recognition and renown of their trade names and brands. The current situation of the economic crisis, which has existed in Spain for various years, has meant that the seeking of business opportunities in other countries has accelerated. According to the 2015 report of the Spanish Association of Franchisors, the figures of franchising in Spain shows that there were 1,199 trade names at the end of 2014, which had a total of 63,869 outlets of which 19,250 were own outlets and 44,619 franchised outlets. According to these data, franchising generated 159,822 jobs through franchised outlets and 89,092 jobs via own outlets in 2014. In the same year, franchising billed more than 25,879 million euros. The aforementioned figures show that franchising is strongly consolidated in Spain, as the majority of the networks have more than 25 outlets. Likewise, the massive presence of trade names in most sectors of activities means that there is now by a high degree of business competitiveness in franchising in Spain. This consolidation of the franchising system in Spain and the strong competitiveness of Spanish trade names, along with the lack of banking finance and the generalized drop in consumption due to the consequences of the economic crisis in Spain make a high percentage of them seek new markets to offset the weak economic results of the national market. According to the data of the, Spanish franchising exists in 113 countries and 148 Spanish trade names operate abroad, having 15,194 outlets. Comparing these data with those of previous years, we note the continuity of the growth trend. Theoretical frameworks show that firms tend to internationalize first to psychically close countries and then gradually move to more psychically distant ones. In fact, Portugal is the first destination of Spanish franchise networks in international markets according to data from the Spanish Association of Franchisors. Therefore, the study of international expansion of Spanish franchisors to Latin American countries is of great importance. Latin American countries have been an important part of this internationalization process of Spanish franchising firms. In fact, a total of 66 chains are operating in the Latin American market and have a total of 3,543 outlets. This is to a great extent due to the cultural link which exists between Spain and Latin American countries. The present work seeks to contribute to the study of international franchising by introducing several factors as the market potential and size, the per capita income and the unemployment level of the countries, among others, which have not been used in the previous literature. Also, the literature review concerning the international expansion of franchising indicates the lack of works which have exclusively centered on the Latin American market. This gap in the literature, together with the expansion of Spanish trade to the Latin American market justifies our work. Furthermore, in this study we use canonical correlation, a methodology that has not been used in the previous literature of international franchising. This specifically analyzes the factors which are external to Spanish trade names and that determine their inclination to operate in the Latin American market. The specific aims of this work are, therefore, on the one hand, to offer a general view of the current situation of Spanish franchisors in Latin American countries and, on the other hand, to analyze what the socio-economic or external factors of Spanish franchisors in this market are. To achieve these aims, this investigation begins with an analysis of the current situation of Spanish franchisors in foreign markets, a special emphasis being given to the Latin American market. The work continues with a review of the extant literature concerning the factors which determine the selection of markets. There is then a specific study of the socio-economic factors which determine the choice of markets for the concrete case of franchisors. Within this specific section we define the different hypotheses proposed in the work. Next, we describe our methodology, defining the variables which are the aim of the study and the canonical-correlation analysis. The results attained are shown and explained in the following section. The work finishes with the conclusions and business implications, as well as the limitations and future research lines. ## Statistics concerning the expansion of Spanish franchisors in Latin American countries The franchising system has existed for a long time in many countries and markets and has become one of the main expansion options for the businesses of many firms. The security of a tested business model is one of the main assets of this business activity. This has meant that over the last decades it has achieved spectacular levels of development in many countries. The franchising system began in Spain at the end of the 50s with the entry of some foreign brands, mainly from France—in the personal apparel sector—and the USA—in the fast-food sector. Nevertheless, it was not until the 80s that franchising really took off in Spain, becoming a successful business format via the huge entry of foreign chains and the appearance of numerous Spanish firms. Spain has thus gone from two trade names in 1960 to 47 in 1980, 750 in 2000 and 1,114 in 2015. The evolution of franchising in Spain has been characterized by two basic aspects: the consolidation of the system in Spain and the growing expansion of many franchisors toward other markets. Nowadays, franchising exists in practically all sectors of activities. This fact, along with the high percentage of trade names with more than 25 franchisees, shows that this business system has become solidly consolidated in Spain. Parallel to this consolidation, numerous trade names, convinced of their own strength and aware of the difficulties in the national market, have begun a process of expansion to other, international markets. All the experience attained throughout the years by Spanish trade names has led to the setting up of more professionalized projects which are better prepared to expand to other markets. In spite of the Spanish national market’s prevailing economic problems, Spanish franchising has continued with its expansion to other markets. Of the 148 Spanish chains which operate in some part of the world, 66 do so in the Latin American market. This is 44.6%, indicating the importance that this market has for Spanish franchising. Likewise, of the 15,194 Spanish outlets, 23.3% of them are located in Latin American countries. If we analyze the chains that exist in the Latin American market, we note that the *DIA* chain is the one which has the most franchised outlets in Latin America—856 units—followed by *MRW*, which has 601 units. Next are *Telepizza*, which has 344 units, and *No+Vello* with 343. The rest are far behind. A significant aspect instead of a massive presence of franchising in a broad variety of activities, it is restricted to a few sectors. The data in corroborate this. shows that the food sector has the greatest number of Spanish outlets in Latin American countries, all belonging to the *DIA* brand. The transport services sector also has a significant presence in Latin American countries. However, the same happens as in the food sector: there is a sole brand–*MRW*. Fashion is the sector with the strongest presence in Latin America and has 21 franchising chains. Beauty and cosmetics is another important sector in Latin America and has 11 franchising chains. In brief, it can be noted that in spite of the significant presence of Spanish franchising in Latin American countries, it still does not operate in many sectors of activities there. This is an important market niche for many Spanish trade names which are in the process of expanding their businesses toward other markets. Analyzing by countries, it is noted that Mexico stands out from the rest as to the number of chains operating there: 19.5% of the trade names in Latin America. Then come Colombia, the Dominican Republic and Venezuela, with each having 8%. As to outlets, Brazil stands out, having 22.7% of the total of franchised units, followed by Mexico which has 20.3%, Venezuela which has 19.8% and Argentina which has 15.6%. The rest of the countries have far less. Without doubt, Brazil is one of most significant emerging markets for Spanish franchising, chiefly due to its market’s potential and maturity, its growth rate, the closeness of its culture to Spain’s and the strong development of the Brazilian franchising system. In short, Spanish franchisors consider the Latin American market as one of the most important for expanding their businesses. However, despite this market’s potential, neither the number of chains nor the number of sectors of activities has attained high figures. This is why this market is a great opportunity for many Spanish chains which are currently analyzing their situation to open up toward foreign markets. ## Factors to be considered when choosing markets Choosing their markets has become a crucially relevant fact for those firms which have decided to develop their activity in foreign markets. Choosing international markets has thus become a fundamental decision. This decision requires information about the possible destination markets (for instance, government regulations, transaction-specific factors, global strategic variables and the dynamics and nature of market variables) and the valuing of this information is going to largely determine the degree of success or failure in the international market. Hence, an important aspect to take into consideration will be the analysis of the different country-markets where the firm can and must be present. The business literature has not been oblivious to this topic and various scientific works have dwelt on this matter. There are numerous relevant studies about determining attractive markets. There is a long time ago, and others which are more recent, such as, and. All this extant literature on the subject has reached similar conclusions. Nonetheless, some differences among the studies lie in both the indicators considered and the weightings assigned to these indicators. But, generically, all converge on the need to consider indicators related with social, political, and cultural factors, amongst others, in order to establish the potential attractiveness of foreign markets. There is a broad literature in this context and different theoretical frameworks have been proposed concerning the phases in the choice of markets, the methods to use in this process and the indicators which must be employed to value the appeal of an international market; among others. This work is going to center on this third approach. Despite the large number of previous studies on internationalization decisions, the academic attention paid to the expansion of franchisor to the Latin American region continues to be limited. In fact, most published works merely focus on the American or British chains. The aim of this work will be to find those factors which are relevant to the presence of the Spanish franchising system in the Latin American market. Regarding the indicators for choosing markets and to explain the presence of Spanish franchisors in Latin America, models which mean to rationalize the search process for a new international market and propose criteria for choosing it began to appear in the 70s. Some studies have been undertaken seeking to integrate previous works, suggesting as sets of criteria for choosing markets: the size of the product’s specific market and its growth, the availability and costs of the production factors, the level of economic development, and the country’s environment, the psychic distance, and the market’s competition and knowledge. present various types of information that exporters consider to be important when evaluating and choosing foreign markets. From the empirical research carried out among different sectors of activities, these authors conclude that all the elements analyzed coincide in placing at the top in the ranking of indicators the information on the destination country’s financial structure, as well as the information about the competition and the entry conditions. On the other hand, the information on market research, advertising and promotions, and demographic and socio-economic environments is the least valued. A very relevant work with a theoretical framework about the choice of markets is study. These authors carry out a comprehensive theoretical review of the literature on the information which can be useful when evaluating international markets. In their review they identify a total of 200 indicators, which they reduce (to achieve operability) through a process of interviews and focus groups with representatives of export-related governmental agencies, of international banking institutions and of private businesses with export experience. This process produces a theoretical framework with six dimensions that group together different indicators. Each of these dimensions and indicators coincides with the elements which the literature frequently cites as being important factors to consider when choosing a market and developing international strategies. This research also offers a ranking of the dimensions and indicators considered as being the most significant when selecting a market: the market’s potential, the country risk, and the cultural distance, among others. ## Socio-economic indicators relevant for franchisors when choosing markets In 2012, Baena proposed that more work is necessary to identify the key factors affecting to the international expansion of franchisors. In this work we are going to consider a set of variables, basically related with the destination country, although the literature indicates other factors connected with the franchisor or the franchisee that have also to be taken into account when studying the topic. Three different types of indicators can be grouped together from this set of factors and the points cited above. Factors directly linked to the trade name are also considered (size, international experience, access to resources, experience as a franchisor, etc.). In this sense, the work of determines that the international experience variables (number of years of internationally franchising), speed of internationalization (years from the origin of the franchisor to its international expansion) and the sector in which the franchisor operates are those that affect the franchisor’s strategy to compete in different country-markets simultaneously. Also, it must be noted that franchisors only use strategies of internationalization after they have accumulated local experience and knowledge. A second block of factors is those related with the franchisee itself and its personal characteristics (risk aversion, individualism, financial capacity, and so on). From the work of it is concluded that Spanish trade names prefer to enter markets where the franchisee has a high risk aversion (fear of failure in the business), as well as a low degree of individualism, understood as the competitive attitude of working individually vis-à-vis collectivisms. This evidence notes that franchising is configured as a form of business expansion which enables the minimizing of the business-associated risks. This is why, in those markets which have people with a strong entrepreneurial character and whose risk tolerance is therefore greater (high levels of individualism), the presence of franchising chains is less. A third block is those factors associated with the operation’s destination country. This is what this work analyzes. Of the different destination country factors to be taken into account in the internationalization of firms, most of the works analyzed evaluate both the cultural distance and the geographical distance. The growth of international expansion is continuing with both export and import oriented expansion being mostly within the same geographical area or those related with prior historical ties with the importing country. It seems that, all things considered, franchisors carefully consider both the geographical and cultural distance dimension. These are decisive in the international markets entry mode and in the choice of markets, as their differences affect, for instance, strategic decisions. Nevertheless, and as note, the cultural distance—considered from Hofstede’s indexes- are not going to be relevant in this work; that is to say, for the internationalization of franchisors toward Latin American markets. Cultural similarities and differences in the destination countries (the Latin American market) with respect to the domestic market (Spain) as to demographic, psychographic, lifestyles and values aspects mean that this variable is not relevant. The same occurs with the geographical distance, given that this is also similar between the origin markets and the destination markets. To expand a franchise toward the Venezuelan, Colombian market or any other of that continent is very different, in terms of geographical distance, from doing so in Portugal, France or Morocco, though Hoffman et. al.'s investigation concerning the internationalization of franchises results in marketing similarities seeming to present significant explanations with respect to the volume or flow of cross-border franchise activity between two nations. Yet there is a series of variables which are going to be decisive in the choosing of destination markets and which we are going to develop. The country risk variable, which encompasses macroeconomic aspects (inflation, depreciation the currency and interest rates) and political aspects (wars, expropriation, confiscation, currency convertibility, repatriation of incomes, tax increases, elimination of subsidies, limitations of the control and ownership of firms) has generally had little importance in the choice of the foreign markets entry mode, as well as in selecting the destination markets. Nonetheless, there seems to be a trend which indicates that firms are changing their way of valuing risks associated with the instability of political and economic conditions, mainly if the business which is being internationalized belongs to the services sector (as services firms are more workforce than capital intensive, in general, starting up activities in international markets does not tend to be accompanied by high investment, as takes place in manufacturing firms which do require much investment, mainly in installations and production technologies). Operating in international environments therefore means a series of risks related with the destination country. propose that a greater risk favors entry modes which involve a lower commitment of resources (such as, for example, franchising). In this way, the flexibility necessary to enable the firm to adapt to the fluctuations of the external conditions is achieved without incurring high fixed costs. Notwithstanding, once franchising has been established as an entry mode, due to these associated risks those countries are chosen where the risk of failure is less. Accordingly, the hypothesis proposed is: ***Hypothesis1**. The likelihood of being considered as a destination market in the internationalization of Spanish franchising firms decreases in countries with a high Country Risk Index*. The current and future performances of franchisor are affected by the economic criteria used in country selection. A relevant factor found in the literature for choosing appropriate markets is the Market Potential. One way, among others, of measuring this is through the country’s per capita income. Studies such as propose that the influence of share ownership on businesses is positively correlated with the destination’s economic development. Other studies on internationalization determine that a greater economic development of the country chosen for investment favors a greater involvement of the firms’ resources commitment. However, other research posits exactly the opposite. That is to say, that the markets with a greater potential are the most appropriate for using contractual modes, as according to them: a) in these markets there is greater competition and this decreases the return on the investment, b) knowledge transfer is easier and the adaptation costs less, and c) in these markets there is normally a greater legal protection in the contractual agreements. With respect to our study, international expansion offers firms different advantages. One of them is that through this strategy the trade names can enter the unsaturated new markets relatively quickly. This is because these markets mostly have lower levels of competition than those in the firm’s country of origin. Other works have also suggested that franchising chains prefer to enter markets which have sizable growth indexes that are sustained over time. This is due to their facilitating high levels of growth and income. We then propose the following hypothesis: ***Hypothesis 2**. The likelihood of a greater presence of franchisors as a business operation mode increase in countries with a high Per Capita Income*. The market potential variable can also be defined by other economic variables, such as the unemployment rate. The unemployment rate is widely recognized as a key indicator of labor market performance. As an economic indicator, the unemployment rate attracts a great deal of media attention, especially during recessions and tough economic times. This factor is related to the level of economic development and it is a key criterion in emerging markets. Hence, this vein suggests that there is a proportional indirect relation between the unemployment rate of the franchisee’s country of origin and the rate of expansion of franchising in that country. This is because the idea of being a franchisee becomes more attractive when the costs of the opportunity of adopting this decision decrease. In this way, the likelihood of the chain finding local partners interested in being franchisees increases and, therefore, the franchisor’s degree of internationalization in those countries is higher when the cost of the opportunity of being a franchisee decreases. Taking this argument to the extreme, those who are unemployed have a lower cost of opportunity than those who at that moment have a stable job, or who live in countries which have a solid economy that facilitates their access to a well- paid job. Hence: ***Hypothesis 3**. The likelihood of a greater presence of franchisors as a business operation mode increases in countries with a high level of unemployment*. Another of the factors which helps to understand a market’s potential is the population destiny or market size. Different studies note that the market size in terms of the population density of the investment’s destination country is a relevant factor when choosing a market. The market size affects productivity, given that large markets allow firms to exploit scale economies. In a period of globalization, international markets have become a substitute for national markets, especially for small countries or countries whose consumption rates are dropping. The empirical evidence shows that opening up trade is positively associated with growth. There is therefore a general feeling that trade has a positive effect on growth, particularly in countries with small markets. Internationalization can hence be considered as a substitute for internal demand to determine the market size for a country’s firms. In this sense, firms seek markets with a population density which enables them to aspire to a high market potential for their products. Countries such as China, India, Brazil, etc., have become, due to their high population, the goals of many firms. Franchising firms are no exception to this reality and those firms which seek the internationalization of their trade names target markets with high population levels. ***Hypothesis 4**. The likelihood of a greater presence of franchisors as a business operation mode increases in countries with a high density of population*. The last factor related with the destination country which is going to be considered in this work is the Global Competitiveness Index. This is established by the World Economic Forum. A country’s level of competitiveness is important because its elements are crucial for its growth, for its productivity and to boost the investment of both foreign and internal investors. A competitive country thus enables a more efficient and swift development. The competitiveness of countries contributes to the factors which determine growth, help to explain why the economies of some countries are more successful than others at increasing their levels of incomes and broadens the opportunities for their populations. Furthermore, this index offers business people the guidelines necessary to make investments and strategic decisions. Taking these arguments as a reference for this work, the following hypothesis can be established: ***Hypothesis 5**. The likelihood of a greater presence of franchisors as a business operation mode increases in countries with a high rate of competitiveness*. # Material and methods To be able to verify the hypotheses proposed in this work, we used data concerning the number of Spanish franchising chains which were operating in the Latin American market in 2014 according to the Tormo & Asociados consulting company. These data were obtained from a database published in the FRQ journal, edited by this company. The database had information concerning the number of Spanish franchising chains present in Latin American countries, as well as the number of outlets in each of them. From this information, the dependent variable to be used in our model is the Spanish chains’ degree of presence in Latin American countries. However, we must highlight that the use of the individual form of the two variables previously indicated can create a significant bias due to not totally reflecting the presence of Spanish franchisors in the different countries. For example, though few brands may be present in a country, they may a high number of franchised outlets. A country may also have numerous brands, yet have few franchised outlets. In such conditions it is difficult to determine what the degree or intensity of the presence of Spanish trade names in the different countries is. To avoid this problem, in this work we have used canonical-correlation analysis. Unlike multiple correlation analysis, where a single dependent variable is predicted from a set of independent variables, canonical correlation predicts multiple dependent variables from multiple independent variables. Canonical-correlation analysis is therefore applied to the study of the association between two groups of variables. One of the advantages of canonical-correlation analysis is that it does require the assumptions of multiple regression analysis, especially referring to the normality. In canonical-correlation analysis the matrix of the data is made up of *n* rows and *p* + *q* columns, split into two sub-matrixes *X* and *Y*, of columns *p* and *q*, respectively. The *n* rows represent the observations (in this case the countries), the first *p* columns *X* are the first group’s variables (independent) and the following *q* columns *Y* are the variables of the second group or dependent variables. So, canonical-correlation analysis seeks to relate the set of independent variables *(X*_*1*,*…*,*X*_*p**)*with the set of dependent variables *(Y*_*1*,*…*,*Y*_*q**)*. The aim is to find pairs of variables *U = u*_*1**X*_*1**+…+u*_*p**X*_*p*, *V = v*_*1**Y*_*1**+…+v*_*q**Y*_*q* with the maximum linear correlation between *U* and *V*. The variables *U* and *V* are called canonical variables. The correlation between them is known as a canonical correlation and the pair of variables (U, V) are called a canonical function. The dependent variables to be explained in this work are the number of chains and the number of franchised outlets, while the independent or explanatory variables are the country’s risk index, its per capita income, its unemployment level, its market potential and its global competitiveness index. The data of these variables corresponding to 2014 have been obtained for each of the Latin American countries. The data of the risk index variable have been obtained from the Organization for Economic Co-operation and Development (OECD). The OECD classifies countries taking into account various quantitative and qualitative indicators. Among the former are the record of international payments compliance and the economic and financial situations. Political risk is what is mainly valued in the qualitative aspects. The per capita income, the unemployment level and the market size variables have been obtained from the World Bank. The World Bank defines the per capita income as the gross domestic product (GDP) divided by the population in the middle of the year. The GDP is the sum of the gross added value of all the producers resident in the economy plus all the taxes on the products, minus all the subsidies not included in the value of the products. It is calculated without making deductions for the depreciation of manufactured goods or for the depletion and degradation of natural resources. It is expressed in US dollars. The unemployment level has been measured through the unemployment rate, which is defined as the proportion of the active population without work but looking for work and available to work. Yet it must be taken into account that the definitions of active population and unemployment differ according to the country. The market size has been measured via the country’s total population. The global competitiveness index has been obtained from the World Economic Forum (WEF). This index measures the ability of countries to provide their citizens with high levels of prosperity. In turn, this ability depends on how productively a country uses its productive resources. As a consequence, the index is a measure of the set of institutions, policies and factors which determine a country’s level of productivity. shows the data of each of the two models’ explanatory or independent variables in each of the Latin American countries. The data were analyzed via syntax with the IBM SPSS Statistics 20software. # Results The first result, which is obtained by applying the syntax in the IBM SPSS software, is the multivariate of significance tests. Each of the previous tests seeks to verify if the canonical correlations is equal to 0. As each p-value is below 0.05 there is an association between the two groups of variables. contains the information of the eigenvalues and the canonical correlations. Each eigenvalue comes from the square of the canonical correlation between each of the variables. We observe that the first canonical correlation is 0.8868. The first canonical function is obtained in such a way that it reflects the greatest possible correlation between the two sets of variables. The first canonical function corresponds to the first eigenvalue, which explains 92.06% of the total of the correlation’s variability between the two groups of variables, while the second canonical function only explains 7.94%. Thus, the relation between both sets of variables can be determined by the first canonical function, given that it explains a higher percentage of the correlation’s variability between both sets. The SPSS results also show the coefficients—both the gross coefficients and the standardized coefficients—for the dependent variables and for the independent variables. However, problems of multicollinearity make interpreting them difficult. This is why they must be explained based on the correlation between each dependent (independent) variable and the corresponding dependent (independent) canonical variable. Tables and show this information. It can be seen that the dependent variable which is most correlated with the factor of the international presence of franchisors in the Latin American countries is the number of franchised units. Its correlation is 0.9529, while the correlation between the number of chains and the factor of the international presence of the franchisor in the Latin American countries is due more to the franchised outlets than to the number of trade names, brands or chains in these countries. If we analyze the correlations between the independent variables and the socio- economic factors, it can be noted that the most correlated variable is the market size: 0.8719. The second most correlated value is the per capita income: 0.3106. However, the rest of the correlations are not significant, as their values are below 0.3 in absolute value. shows a summary of the confirmation or non-confirmation of the hypotheses proposed. It must be pointed out that the non-confirmation of the hypotheses means that there is no significant relation between the socio-economic factors and the presence of Spanish franchisors in Latin American countries. Thus, for example, the non-confirmation of hypothesis H1 does not imply that there is a greater presence of Spanish franchisors in the countries with a greater Country Risk Index, but that there is no significant relation between the Country Risk Index and the presence of Spanish franchisors in the Latin American countries. This is likewise the case with hypotheses H3 and H5. # Conclusions The previous results enable the deduction of a series of considerations concerning the socio-economic factors which determine the presence of Spanish franchisors in the Latin American market. As has been seen in this study, internationalization through the franchising system is a recognized and relevant fact, and Spanish franchisors are increasingly more present in different markets. The brand “Spain” is gaining in importance in the world and although everything “Spanish” has always been held in high esteem in Latin American countries, the strength of this brand image only confirms this fact. In short, a strong growth has taken place in recent years both in the number of different trade names and in the number of outlets, being present in all Latin American countries. As well as showing the presence of Spanish franchisors in these markets, the aim of this work has been to determine the relevant socio-economic factors in choosing markets. To do so, we carried out an analysis of the extant literature concerning the decisive factors in choosing markets and we have considered these five factors as being the most relevant (related with the destination market): Country Risk, Per Capita Income, Unemployment Level, Market Size and Global Competitiveness. From the study proposed it is deduced that the Spanish trade names which expand their businesses toward the Latin American market fundamentally base their decision on two socio-economic factors: firstly, on the destination market’s size and, secondly, on the destination country’s per capita income. Therefore, Spanish franchisors do seem to consider those markets which have a large market size as to the density of population and a considerable per capita income that enable them to attain high sales indexes. Most Latin American countries have significant growth rates. Brazil is the fifth world power in growth in terms of GDP, having had growth levels of around 8 to 9% in the last three years. Other countries on the continent have growth levels of around 5%. This means that the per capita income has increased and that consumption has been activated. Therefore, Spanish franchisors have wished to be present there. Countries such as China or India have been the destination target of many firms in recent years. This is due to their growth in terms of both GDP and population size. The same is happening for some Latin American countries such as Brazil, which has approximately 300 million inhabitants. A large population and a considerable per capita income are therefore relevant factors in firms’ choice of markets. On the other hand, it is inferred from this study that franchisors do not take into account other factors such as the destination’s Country Risk Index, its Unemployment Level or its Global Competitiveness Index. This may all be due to the franchising model itself as an entry mode abroad. This entry mode has an associated lower business risk in the event of failure, as the franchising firm’s levels of investment are relatively low. It is the franchisee who, by investing the capital, runs a greater risk. Hence, although the destination country which the firm wishes to opt for has a high Risk Index, the firm can consider it as a market target. This is because the company’s financial risk levels are low, while it gains in renown and image for being present there. As to the Unemployment Level, in countries where these levels are high, these unemployed people might see self-employment as a way out of this situation. Franchising hence may be considered as a good formula to do so. However, the Latin American culture itself and studies of entrepreneurship levels do not confirm this hypothesis. The other hypothesis not upheld is related with the levels of competitiveness. It appears that the very complexity of confirming the measurement—it is an index made up of twelve different pillars which cover a broad range of aspects—makes it not objective and direct enough to be measured and considered. To sum up, those Spanish firms which want to operate in the Latin American market should seek countries in which the market potential or size stands out among the socio-economic factors and there is a high per capita income. This work shows the socio-economic factors which Spanish trade names take into account when internationalizing their operations in the Latin American market. The results of this study show a series of socio-economic factors which influence the final choice of the destination country. Nevertheless, this decision is not only based on the socio-economic aspects of the destination country but also on the structure of the franchising firm itself and on the export experience toward other markets. Therefore, this study serves as a complement to other studies and helps franchisors with the difficult decision of choosing a destination for their internationalization. ## Limitations In spite of the importance of the topic analyzed, this work has a series of limitations. One refers to the geographical area to which it is restricted. A global study—not only centered on the Latin American market—could be useful to shed more light on the internationalization of franchising firms. This is both a limitation and a future research line. The methodology used is another limitation. The use of canonical-correlation analysis enables solving the multicollinearity problem which exists in a multiple linear regression model. Nonetheless, many authors consider that it is important to have a high sample size to be able to apply this type of analysis. In this case, only the data of Latin American countries have been taken into account, which is why the sample size has been restricted to 20 units. This is a significant limitation. ## Future research lines Regarding future research lines, it must be stressed that this work has analyzed a series of important socio-economic factors of Latin American countries to determine their influence in the internationalization of Spanish franchisors there. Despite this, it would be interesting to study other factors inherent to countries, apart from than those analyzed in this research. Thus, for example, an important factor could be the study of the level of maturity or consolidation of the franchising system in each country. That is to say, an interesting idea would be to study if the level of consolidation of franchising in Latin American countries affects the decision of Spanish franchisors to internationalize toward these countries. On the other hand, it would also be interesting to study how the internal variables of Spanish franchisors (size of the franchisor, years of franchising, sector, etc.) affect the choice of Latin American countries as destinations for their internationalization strategy. Finally, as has been emphasized before, it would also be interesting to replicate this study with markets, other than those of Latin America, to analyze the possible differences which may exist among them. # Supporting information This research was supported by the Andalusian Regional Government of Spain through the project P11-SEJ-7042. The authors wish to thank the anonymous reviewer for his helpful comments which have contributed significantly towards the improvement of this paper. [^1]: The authors have declared that no competing interests exist.

# Introduction Diabetes mellitus is a common disease that exerts tremendous impact on human health. It has been shown that patients with diabetes are also at a significantly higher risk of developing various types of cancer. Data has shown that approximately 80%of patients with pancreatic cancer suffer from hyperglycemia or diabetes. And high glucose (HG) has been considered as a subordinate cause, that can trigger direct and/or indirect mechanisms to promote cancer cell proliferation, migration and survival. However, the underlying mechanisms for this relationship are still not fully understood. Due to its clinical significance, increasing efforts have been made, trying to elucidate the link of carcinogenesis to the status of patients having high fasting glucose level, or being obese or diabetic, this is particularly important because an appropriate blood glucose level control could significant affect the occurrence and prognosis of cancer. On the other hand, mitochondria has been shown to play important roles in cancer cells, maintaining mitochondrial potential and oxidative equilibrium that are essential for apoptosis and cell viability. In fact, mitochondria is becoming an important therapeutic target for anticancer drug, such as mitocans, which can eventually cause cell death via interrupting mitochondrial integrity. Recently, studies have shown that GRIM-19, also named NDUFA13, acts as a cell death-regulatory protein that can be induced by the combination of interferon-beta and retinoic acid. GRIM-19 is also identified as one mitochondrial complex I subunit, which not only plays an important role in oxidative phosphorylation (OXPHOS) for ATP generation, but also is involved in the process of glycolysis, a key metabolic process for cancer. Thus, GRIM-19 has the ability to modulate cancer cell survival. Data has shown that a mono-allelic loss of GRIM-19 can promote carcinogenesis in mice and the tumor- derived mutations in GRIM-19 in human can also promote tumor growth in mice. Moreover, GRIM-19 exerts the pro-survival effects through its interactions with signal transducer and activator of transcription-3 (STAT3) which is an important member of the STAT family protein. In response to cytokines and growth factors, such as IL-6 and epidermal growth factor, STAT3 is activated through its phosphorylation at tyrosine 705 and forms homo- or hetero-dimers that translocate to the cell nucleus, acting as a transcription activator to regulate many cellular processes such as cell growth and apoptosis. Interestingly, data has also linked STAT3 to both normal and altered insulin signaling in the setting of diabetes. Our previous study has indicated that STAT3 signaling was involved in HG induced HepG2 cells proliferation. In fact, it has been well demonstrated that HG can exert toxic effects on normal organ cells, such as cardiac cells as well as cancer cells, for which activity of AMPK was shown to be closely involved. However, it has not been shown whether Grim-19 is involved in the HG/STAT3 signaling, especially whether the altered metabolism could link Grim-19 expression to AMPK activity. Therefore, in the present study using HeLa and H9C2 cell line, we sought to investigate: 1) the relationship between the expression level of GRIM-19 and activity of STAT3 signaling in the setting of HG; 2) how AMPK activity interact with Grim-19 in HG cultured cells and 3) how metformin, an AMPK agonist, modulates the expression of Grim-19. # Materials and Methods ## Cell lines and treatment conditions HeLa cells and H9C2 cells were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA). The HeLa cells and H9C2 cells were maintained in DMEM medium containing normal level (5.5 mM) or high (25 mM) level of glucose supplemented with 10% FBS, L-glutamine and antibiotics (100 units/ml penicillin, and 100 μg/ml streptomycin) at 37°C in the presence of 5% CO2. And the effect of IL-6 and the role of GRIM-19 in HeLa cells and H9C2 cells cultured in HG DMEM were investigated by plasmid that has been inserted either scramble or siRNA targeting GRIM-19. ## Reagents The MTT Cell Proliferation and Cytotoxicity Assay kit was purchased from Sangon Biotech (Shanghai, PR China), AG490 from Beyotime(Jiangsu, PR China) and IL-6 from eBioscience (eBioscience, CA, USA). Primary antibodies against β-actin and GAPDH were purchased from Santa Cruz (Shanghai, China), antibody targeting GRIM-19 from Abcam (ab3449, Shanghai, PR China), and antibody against p- STAT3 (#4113), total STAT3 (#12640), p- AMPKα (Thr172, \#2531), total AMPKα (#2532, p- Akt (Ser473, \#4058) and total AKT (#4685) were from Cell Signaling Technology (Danvers, MA, USA). And all the secondary antibodies were purchased from Biosynthesis Biotechnology (Beijing, PRChina); penicillin, streptomycin, DMEM medium, and fetal bovine serum (FBS) were obtained from GIBCO (Grand Island, NY, USA). ## MTT assays *in vitro* The effects of either HG and/or GRIM-19 over-expression on HeLa and H9C2 cells viability were determined by the nicotinamide,3-(4,5-dimetrylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) dye uptake method. Briefly, the cells were equally distributed in 96-well plates at a density of 1 x 10⁴ cells/well (counted by a hemocytometer). Cells were treated with different glucose level for 48 hours. Before incubation with MTT, DMEM (serum10%) medium was removed and the final concentration of MTT (Sigma, St. Louis, MO, USA) at 5 mg/ml in DMEM medium and incubation was for 6 h in the dark at 37°C. Then, the supernatant was removed from each well. The colored formazan crystal produced from MTT was dissolved in 150 μL DMSO and cell viability was determined spectrophotometrically at 490 nm (BioTek, Vermont, USA). ## Western blotting analysis Cells were harvested and washed with cold phosphate-buffered saline (PBS) and lysed in RIPA buffer (150 mM NaCl, 50 mM Tris-HCl, pH 7.2, 1% deoxycholic acid, 1% Triton X-100, 0.1% sodium dodecyl sulfate (SDS), 0.25 mM EDTA) with the protease inhibitor cocktail (Roche, Basel, Switzerland).The lysate was separated by SDS-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to NC membrane. The membrane was blocked with PBS containing 5% non-fat milk containing 0.05% Tween 20 followed by incubation with primary antibodies. Appropriate secondary antibodies were then used. The bound antibodies were visualized by enhanced chemiluminescene (Thermo, MA, USA). The band for each protein was then quantified by densitometry using Image J software (version 1.41, NIH, USA) and normalized to the expression of ß-actin or GAPDH for protein loading. ## Lactate acid quantification A commercially available kit was used (Cat# A019-2, JianCheng BioTech, Nanjing, PR China) to measured the lactate acid level in cell culture medium that was obtained at the end of each cell experiment. All the data were expressed as fold changes in comparison to controls. ## Reverse transcription PCR analysis After cells were transfected with GRIM-19 or siRNA-GRIM-19 for 48h, the cells were pre-treated with different levels of glucose for another 48 h. Total RNA was then isolated from the cells, using TRIzol® Reagent (Life Technologies, Rockville, MD, USA) following the manufacturer’s protocol. Complementary DNA was synthesized from 1 μg of total RNA using a cDNA Synthesis Kit (Takara Biotechnology Co., Ltd., China) following the manufacturer’s instructions. Primer sequences (Genecore, Shanghai, China) specific for CyclinB1, CyclinD1, Bcl-2, VEGF, STAT3 and GRIM-19 are shown in. β-actin was used for the loading control. After cDNA synthesis, the PCR reaction consisted of 32 cycles of denaturation at 95°C for 30 s, annealing at 56°C for 30 s, extension at 72°C for 30 s, and a further 5 min at 72°C in the last cycle. PCR products were separated by electrophoresis on a 2% agarose gel and visualized by staining with ethidium bromide. The expression was quantified densitometrically using the Gel Image Ver. 3.74 System (Tianon, Shanghai, China). ## Transient RNA interference and transfections Both GRIM-19 and STAT3 were knocked down using small siRNAs, with non-targeting siRNA (scramble RNA) used in parallel as a negative control respectively (GenePharma Co., Shanghai, PR China). Primary cultured cells were transfected with siRNA (targeting either GRIM-19 or STAT3) or scramble RNA using Lipofectamine 2,000 (Invitrogen, Carlsbad, CA) according to the manufacturer’s instructions. Forty-eight hours after transfection of mRNA silencing, HeLa or H9C2 cells were collected for protein expression analysis of either GRIM-19 or STAT3 to confirm the effects of siRNA. ## Plasmid, lentiviral construction and DNA infection The over-expression of GRIM-19 lentivirus was constructed (from Shanghai Sunbio Medical Biotechnology) which contained GFP and Flag. Human GRIM-19 sequence was amplified by RT-PCR from HeLa cells. WT-GRIM-19 with complete open reading frame was cloned into NOT I and EcoR V sites of mammalian expression vector Pflag- CMVTM-4. The pFLAG tag was added at the N-terminal of the GRIM-19 sequences in all constructs. Infection of plasmids to cells was performed using Lipofectamine 2,000 (Invitrogen, Carlsbad, CA) according to the manufacturer’s instructions. Forty-eight hours after infection, HeLa and H9C2 cells were analyzed for protein expression of GRIM-19 expression levels. ## Statistical analysis Dates are expressed as means ± SD for three or more independent experiments. Statistical significance was estimated by one- way ANOVA followed by Student–Newman–Keuls test for comparison of several groups. P \< 0.05 was considered statistically significant. # Results ## Enhanced HeLa cell proliferation induced by HG was associated with down-regulated GRIM-19 level MTT test demonstrated that HG treatment significantly increased proliferation of HeLa cells. HG treated HeLa cells had down-regulated GRIM-19 levels compared with those cultured with normal glucose medium, and this was associated with a significantly increase in STAT3 phosphorylation level at Tyrosine 705. Being consistent with these, RT-PCR also showed a decrease in mRNA expression level of GRIM-19 in HG treated cells, but not that of STAT3. Interestingly, HG resulted in a significant increase in lactate acid level in HeLa cells. Thus, we provided evidence that HG treated HeLA cells had enhanced proliferation, down-regulated expression of mitochondrial OXPHOS protein GRIM-19 and increased STAT3 phosphorylation level. ## GRIM-19 knock-down alone can activate STAT3 signaling and enhance HeLa cell proliferation To elucidate whether down-regulated GRIM-19 was responsible for the activation of STAT3 (phosphorylation level) and hence the proliferation activity, we transfected the HeLa cells with siRNA targeting GRIM-19 to down-regulate the GRIM-19 expression. The efficiency of transfection was confirmed by the mRNA level of Grim-19 using PCR. Interestingly, down-regulated GRIM-19 expression alone can increase the phosphorylation level of STAT3 when the HeLa cells were cultured in the normal glucose medium, indicating a causative relationship between decreased level of GRIM-19 and activation of STAT3 signaling. And the down-regulated GRIM-19 also resulted in an increase in HeLa cells proliferation as measured by MTT assay. In addition, down-regulated GRIM-19 also increased lactate acid levels of HeLa cells even cultured in normal glucose medium. ## Over-expression of GRIM-19 attenuated the STAT3 signaling activation induced by HG in HeLa cells We then went further to test whether over-expression of GRIM-19 can affect activated STAT3 signal induced by HG. GRIM-19 was infected into HeLa cells by lentivirus to over-express GRIM-19, showing dose-dependent increases in GRIM-19 expression which in turn attenuated increased STAT3 phosphorylation level induced by HG. However, GRIM-19 over-expression alone did not affect the phosphorylation level of AKT. The reversal of STAT3 signaling activation induced by HG was also evidenced by the decreases in the mRNA levels of STAT3 target genes, such as Cyclin B1, VEGF and Bcl-2, except for CyclinD1. Of note, over-expression of GRIM-19 was not associated with any changes in total STAT3 expression level. Ironically, over-expression of GRIM-19 did not reverse the increased lactate acid level induced by HG, instead, it actually increased the lactate level further compared with HG culture alone. ## Grim-19 was involved in IL-6/STAT3 signaling in HeLa cells We then tested whether over-expression of GRIM-19 also can suppress IL-6 activated STAT3 signaling, HeLa cells were cultured in normal glucose and exposed to IL-6 stimulation. Even though IL-6 alone did not result in any changes in Grim-19 expressions (data not shown), it significantly increased STAT3 phosphorylation levels (lane 3 vs.1) with no changes in the total STAT3 expression levels. And over-expression of Grim-19 significantly decreased phosphorylation level of STAT3 in the presence (lane 4 vs. 3) or the absence of IL-6 (lane 2 vs. 1.). However, this did not result in any changes in AKT phosphorylation levels. ## STAT3 is responsible for enhancing HeLa cells proliferation in the setting of HG culture We have demonstrated down-regulated Grim-19 enhanced whereas up-regulated Grim-19 inhibited STAT3 phosphorylation levels in HeLa cells. When AG490, a JAK2 specific inhibitor, was added to the HG culture medium, it had no effects on GRIM-19 expression (data not shown), however, significantly down-regulated HG activated STAT3 phosphorylation levels. Importantly, AG-490 also blocked HG induced cell proliferation. To further validate the role of STAT3, we showed that knock-down of STAT3 in HeLa cells by specific siRNA can significantly decrease cell proliferation caused by HG, without affecting GRIM-19 expression levels. ## Activation of GRIM-19/STAT3 signaling induced by HG also occurred in H9C2 cells To test how HG affects GRIM-19 expression in cardiac cells, we cultured H9C2 cardiac myoblast cells in HG medium. MTT test showed that HG treatment significantly increased H9C2 cell proliferation, which was also associated with down-regulated expression level of GRIM-19 and enhanced STAT3 phosphorylation, indicating that HG also activated GRIM19/STAT3 signal pathway to enhance H9C2 proliferation. Interestingly, however, HG cultured H9C2 did not show an increase in lactate acid production. Knock-down of STAT3 by transfecting H9C2 cells with siRNA targeting STAT3 did not significantly affect the GRIM-19 expression level, however, abolished the increased proliferation induced by HG. Silence of STAT3 in H9C2 resulted in an increase in lactate acid production. In contrast, over-expression of GRIM-19 in H9C2 cells by lentiviral infection can also attenuate increased STAT3 phosphorylation levels induced by HG, without affecting AKT phosphorylation levels. This also resulted in a significant decrease in H9C2 cells proliferation in the setting of HG, further suggesting that activation of GRIM-19/STAT3 signaling facilitated H9C2 proliferation. Over- expression of GRIM-19 dose-dependently increased lactate acid levels in H9C2. ## Roles of AMP- activated kinase in HG induced changes in GRIM-19 expression in HeLa and H9C2 cells To further test whether AMPKα activity was involved in HG induced down-regulated GRIM-19 expression, we then measured AMPKα activation to evaluate its interactions with GRIM-19 in the setting of HG. Intriguingly, HG cultured HeLa cells did not exhibit any change in AMPKα phosphorylation level, however, a significant decrease in phosphorylated AMPKα level was present in HG cultured H9C2 cells. Metformin administration only slightly increased phosphorylation level of AMPKα in HeLa cells in both normal glucose and HG culture condition. However, metformin resulted in a recovery of the down-regulated phosphorylted AMPKα2 level caused by HG culture in H9C2 cells, even though had no significant effects in normal glucose culture condition. In addition, metformin not only slightly increased GRIM-19 expression in normal glucose cultured HeLa and H9C2 cells, but also resulted in a complete recovery of GRIM-19 expression that was suppressed by HG. Interestingly, while metformin had no effects on STAT3 signaling in HeLa cells, however, it significantly abolished the increased STAT3 phosphorylation in HG cultured H9C2 cells. In addition, metformin increased lactate acid in both HeLa and H9C2 cells cultured either in normal glucose or in HG, showing time dependent effects in H9C2 cells which was associated with suppressed cell proliferation again in a dose-dependent fashion. # Discussion Our study demonstrates for the first time that HG can down-regulate the mitochondrial protein GRIM-19 expression in both HeLa and H9C2 cells, which in turn can activate STAT3 signaling, leading to enhanced phosphorylation level of STAT3 and cell proliferation capability of both HeLa cells and H9C2 cells. Importantly, GRIM-19 silencing alone with normal glucose culture resulted in similar effects as HG culture. In contrast, over-expression of GRIM-19 attenuated p-STAT3 activation induced by HG and resulted in a decrease in cell proliferation for both HeLa and H9C2 cells. In addition, side by side comparison was made between HeLa and H9C2 cells and a strikingly different expression pattern of changes was shown in AMPK activation,showing a decrease in AMPKα phosphorylation levels by HG culture in H9C2 cells, but no change in HeLa cells. In contrast, HG induced lactate secretion was only observed in HeLa cells, not in H9C2 cells. Metformin only had trivial effects on modulating AMPKα activities in HeLa cells, but significantly reversed down-regulated AMPKα phosphorylation levels induced by HG in H9C2 cells. Of note, metformin also reversed down- regulated GRIM-19 expression and up-regulated STAT3 activation both of which were caused by HG culture, however, only slightly increased GRIM-19 expression, yet had no effects on STAT3 phosphorylation levels in HeLa cells. STAT3 as an important transcription factor has long been recognized as a mediator for regulating its targeted genes in response to extracellular stimuli, thus playing key roles in various cellular biological processes such as cell growth and apoptosis. It has been well established that STAT3 can be activated in cancer cells, which makes itself an attractive target for anti-cancer therapy. Interestingly, activated STAT3 signaling has been involved in insulin resistance. STAT3 activation was induced in response to innate inflammation in the setting of high free fatty acid in diabetic setting. In our present study, we found that high glucose can up-regulate the p-STAT3 level in both HeLa and H9C2 cells, which was also observed in HepG2 cells in our previous work. In fact, it was GRIM-19 that was responsible for STAT3 activation induced by HG. The role for GRIM-19 in STAT3 activation was further confirmed by the data that down-regulated Grim-19 expression alone in normal glucose culture can mimic the effects of HG, and HG induced activation of STAT3 signaling can be blocked by GRIM-19 over-expression. Moreover, AG490 or siRNA targeting STAT3 can attenuate the enhanced cell proliferation in the setting of HG. Thus, our data strongly support the concept that Grim-19/STAT3 mediates HG induced cell proliferation for both HeLa and H9C2 cells. GRIM-19 as one important mitochondrial OXPHOS protein has already been shown to be able to gear metabolic states through modulating STAT3/HIF1α signal. Chen et al. reported that a down-regulated GRIM-19 expression can impair the mitochondrial complex I activity, leading to an increase in reactive oxygen species generation. In contrast, over-expression of GRIM-19 can markedly suppress inflammatory cytokine levels such as IL-6 and tumor necrosis factor-α in arthritic joints. In cancer cells, GRIM-19 expression was actually down- regulated in cancer cells, and its level was inversely correlated with phosphorylation level of STAT3, suggesting that a lower GRIM-19 level favors or promotes tumor growth via promoting STAT3 biological activity. At present stage, we still don’t know how down-regulated Grim-19 expression can activate STAT3 signaling. We can only assume that down-regulated Grim-19 expression resulted in impaired mitochondrial complex-I function which resulted in compromised mitochondrial function and activate the function of STAT3, favoring proliferation and/or survival of cells. Thus, a link between high glucose and Grim-19/STAT3 signaling has been observed. It should be noted that other parallel pathways, such as Shp-1 as an important phosphatase, could be independently involved in regulating STAT3 activation in the setting HG. It would be interestingly to test whether a higher STAT3 activity would be also observed in patients with a high glucose level such as cancer patients with diabetes or obesity. In fact, STAT3 activation in cancer cells is closely related to mitochondrial pathway. It has previously been shown that the mitochondrial uncoupling protein 2 regulates the ROS/Stat3 signaling pathway and responds to the chemotherapy in lung cancer cells. Interaction between STAT3 and Grim-19 is important for optimal function of mitochondrial complex I of electro transfer chain, which determines the appropriate amount of reactive oxygen species(ROS) to promote breast cancer growth. It was proposed that mitochondria might function as a central checkpoint by integrating various signals coming either from endogenous elements (such as ions, metabolites or even second messengers) and/or signaling proteins (such as kinases and phosphatases), or from exogenous factors (nutrients, oxygen). Thus, in the setting of HG, down-regulated Grim-19 could serve as a sensor and transmit the signaling of mitochondria to the nucleus through its interaction with STAT3, promoting the survival of cancer cells or proliferation of the cells such as H9C2. It will be also interesting to test if other components of mitochondria would function the same way as Grim-19.It certain warrants further studies that aim to elucidate whether this would be the same case for other types of cells. In our study, we also evaluated interaction between AMPK and GRIM-19, specifically whether AMPKα activity, an important sensor for energy status, would be involved in this HG induced down-regulated GRIM-19 expression. We have demonstrated strikingly different patterns of changes in AMPK activities between HeLa and H9C2 cells. We observed a low expression level of AMPKα in the HeLa cells, whereas in H9C2 cells AMPKα expression is relatively much higher, indicating different metabolic status between these two cell lines. This could explain why HG resulted in a significant decrease in AMPKα activity in H9C2 cells in comparison to much lower fold of changes in HeLa cells, highlighting the importance of AMPKα in modulating the metabolic regulation in cardiac cell line. Interestingly, metformin can abolish the changes induced by HG culture, i.e. reversal of down-regulated GRIM-19 and activated STAT3 signaling that were seen in HG cultured H9C2 cells, whereas in HeLa cells only slight increase in GRIM-19 was observed, without any changes in STAT3 phosphorylation status. Thus, our data not only confirm previous study that AMPK could regulate the activity of STAT3 in cardiac cells, but also elucidate that it was GRIM-19 which mediates activation of STAT3 signaling. ## Study limitations There are several limitation in this study we have to discuss. First, it has been well established that lactate functions as an important metabolic player in cancer cells. In our study, we observed HG culture elevated lactate secretion in HeLa cells compared with normal glucose culture, however, not H9C2 cells. Thus, lactate secretion induced by HG was closely correlated with enhanced cell proliferation, which was associated with down-regulated GRIM-19 expression. Meanwhile, it is difficult to interpret that up-regulated GRIM-19, on the other hand, also increased lactate secretion. Thus, relationship between lactate secretion and GRIM-19 expression cannot be clearly defined in our present study. Second, using metformin as a pharmacological tool, we also looked at the potential roles for AMPK in GRIM-19/STAT3 signaling pathway in the setting of HG culture, even though we observed a strikingly different pattern of changes in AMPK activity in HeLa and H9C2 cell lines in response to HG culture and metformin intervention, we need to be aware of that non-specific effects of metformin cannot be excluded. A genetic approach is certainly needed to further elucidate the role of LKB1/AMPK pathway. Third, our data indicated that AKT and STAT3 showed differential response to IL-6 stimulation, as AKT was not activated when cells were exposed to IL-6 stimulation which was strikingly different from the previous data. We reasoned that it could be due to the normal glucose cell culture was selected as a control, whereas the cells are usually cultured in HG medium in most of the experimental studies where the effects of IL-6 were investigated. Fourth, it remains unknown why metformin up-regulated GRIM-19 expression in both HeLa and H9C2 cells, which resulted in a normalization of STAT3 activation in H9C2 cells, however, failed to decrease STAT3 phosphorylation levels in HeLa cells. We reason that it could be due to different levels of AMPK activation and their responses to metformin therapy in these two cells levels. If this holds true, i.e., AMPK plays a key role in differentiating these two types of cells in responses to metformin, then it will be interesting to test whether knock-out or depletion of GRIM-19 affects metformin-mediated inhibition of p-STAT3 in H9C2 cells. Finally, it remain unknown how activated AMPK by metformin increase GRIM-19 expression in H9C2. Moreover, it is still not understood why AMPK level is relatively lower in HeLa cells than in H9C2. In fact the role of AMPK in cancer cells is still highly controversial. Further studies are certainly warranted to answer these important questions. In conclusion, the mitochondrial complex I protein GRIM-19 can mediate hyperglycemia-induced p-STAT3 signal in both HeLa and H9C2 cell culture. In H9C2 cells, GRIM-19 is involved in metformin enhanced AMPKα activation that is suppressed by HG, which in turn down-regulates STAT3 activity that is activated by HG. Further study designed to look further into GRIM-19 mediated mitochondrial pathway may cast new lights on the treatment of various HG- associated diseases, such as diabetes and tumors. This work was supported by the National Natural Science Foundation of China (#81070110 to M. Wei, \#81370346 to W. Zhu). GRIM-19 Gene associated with retinoid-IFN-induced mortality 19 HG High Glucose STAT3 Signal transducer and activator of transcription 3 MTT 3-(4, 5-dimetrylthiazol-2-yl)-2, 5-diphenyltetrazolium bromide AMPK AMP-activated protein kinase [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: YGL WZ MW. Performed the experiments: YGL BBH FL JWY ZFD. Analyzed the data: YGL BBH FL JWY ZFD GMN YWQ JBL MW. Wrote the paper: YGL WZ. [^3]: ‡ These authors contribute equally to the research work.

# Introduction Biomarkers in breast cancer contribute essentially to adjuvant and preoperative therapy assessment. Additionally to conventional prognostic factors as tumor- size, grading or nodal status, treatment decisions include the three established predictive biomarkers as estrogen (ER) and progesterone (PR) receptors and the HER2 status. Prognostic factors provide information on the likelihood of cancer progression in untreated patients, whereas predictive factors carry information on the probability of therapy response. Multigene assays have become more widely used to prognosticate breast cancer clinical course and assist in the decision making for or against adjuvant chemotherapy. The benefit of chemotherapy in addition to regular hormonal therapy remains a subject of dispute in hormone receptor positive early breast cancer,. Several tests were developed in the recent years measuring the expression profile of cancer-related genes and providing prognostic information on disease-free and overall survival. The Netherlands Cancer Institute in Amsterdam launched Mammaprint, a 70-gene assay in 2002. The genetic signature of Mammaprint predicted metastasis free survival and overall survival in a validation study on 295 breast cancer patients. Oncotype DX, a 21-gene assay was first tested in clinical trials in 2004. It is able to quantify the likelihood of distant recurrence and the probability of response to chemotherapy in early breast cancer. The Recurrence Score was validated in the NSABP B-14 trial in 2004 on 645 patients. The NSABP B-20 trial in 2006 analyzed 651 patients and validated the benefit of additional chemotherapy in patients with high Recurrence Score. Since the initiation of Mammaprint and Oncotype DX, additional multigene tests (e.g. Breast Cancer Index, Rotterdam, Invasiveness gene signature, PAM5) were developed, which are being either commercially available or currently under clinical investigation. These gene assays, either reverse transcription-quantitative real-time PCR (RT- qPCR)- or microarray-based increasingly meet clinical attention, as they represent potential additional tools to conventional pathological prognostic factors and to established international oncological guidelines. In 2011, a new 12-gene test, the EndoPredict assay was launched. It was validated independently in patients from two large randomized phase III trials (Austrian Breast and Colorectal Cancer Study Group (ABCSG)-6: n = 378, ABCSG-8: n = 1324). The EndoPredict (EP) risk score provided additional prognostic information to the risk of distant recurrence in hormone receptor positive, nodal negative breast cancer patients. The EPclin score which is the EP score combined with the clinico-pathological parameters tumor size and nodal status was the first RNA- based prognostic test for breast cancer to outperform all conventional clinic- pathologic risk factors alone or in combination with each other. The performance of the EndoPredict assay in decentralized testing using formalin-fixed, paraffin-embedded (FFPE) tumor tissue was successfully shown in seven European pathology institutions reaching 100% concordance between the different sites. In this retrospective study, we addressed to investigate the concordance of EndoPredict scores and the Oncotype DX Recurrence Scores in 34 hormone receptor positive breast cancer patients. # Materials and Methods ## Patients’ Characteristics 34 patients with invasive breast carcinoma were selected for this study (18 cases from the Institute of Surgical Pathology, University Hospital Zurich, Switzerland, 10 cases from the Institute of Pathology, University Hospital Heidelberg, Germany and 6 cases from the Pathology Institute Enge Zürich Switzerland). The time of diagnoses was between 2008–2012. All tumors were estrogen receptor and in the majority also progesterone receptor positive. On histology 28 tumors corresponded to invasive ductal carcinoma (82%), three to invasive lobular carcinoma (9%) and three tumors were diagnosed as a mixed invasive carcinoma with ductal, lobular and squamous components (9%). Clinico-pathological data of the tumors are summarized in. The study was designated and approved as a quality control study by the Review Board of the Institute of Surgical Pathology (project Nr. 285). The review board specifically waived from the need of an approval of the cantonal ethical committee. According to the Federal Swiss Law for research and as required by the ethical committee of Canton Zurich, no additional ethical committee approval was necessary, as the study was designated as a quality control study and all tissue samples were analyzed in a completely anonymized way. ## Immunohistochemistry for ER/PR/HER2 and Ki-67 Hormone receptor status (in all cases), HER2 status (in 16 cases, from Heidelberg and Pathology Enge) and proliferation fraction (in 33 cases) were determined during routine histological diagnostics using commercial antibodies following the manufactures’ recommendations on the Ventana Benchmark and Leica Bond autostainers. Primary antibodies were detected using the iVIEW DAB detection kit and the signal was enhanced using the amplification kit. Following markers and dilutions were used: HER2 (4B5 Ventana Basel Switzerland) (MIB-1 (Ki-67) (DAKO Denmark, Glostrup, dilution 1∶20), estrogen receptors (6F11, Ventana Basel, Switzerland, dispenser), progesterone receptor (1A6, Ventana, Basel Switzerland, dispenser) as described previously. Cut-off for positive ER/PR status was set as 1% of positively stained nuclei. HER2 immunohistochemstry was scored as described in the ASCO guidelines. ## Fluorescence in situ Hybridization (FISH) for HER2 HER2 status was determined within on the primary tumor using FISH only- methodology in 18 of 34 cases (cases from Zurich, University Hospital). All procedures for the FISH analyses were carried out by following the recommended protocol of the manufacturers using a dual fluorescence kit (PathVysion™, Vysis, Abbott AG, Diagnostic Division Baar, Switzerland).The reactions were evaluated using an Olympus computer guided fluorescence microscope (BX61, Olympus Schweiz AG, Volketswil, Switzerland). FISH testing was evaluated in reference to the ASCO guidelines. Cut-off for positive *HER2* status was set as *HER2*/*CEP17* ratio ≥2.2. ## Tissue Preparation for Oncotype DX Tests Upon request of the oncologists in charge, samples were submitted to Genomic Health (Redwood City, CA) for Oncotype DX testing for breast cancer prior to this study. For this assay, one representative paraffin block was chosen from the cases, containing the largest amount of invasive tumor cells on the hematoxyline & eosin (H&E) slides. The amount of tumor cells was at least 10% of the H&E slide. According to the pathology guidelines of Oncotype DX, 15 unstained serial slides of 4 micrometer thickness per tumor block were freshly cut from the paraffin blocks and submitted for the assay. Recurrence Score were assessed by Genomic Health in all patients. RNA-based ER, PR and *HER2* status were available in 33 of 34 patients. ## Tissue Preparation and RNA Isolation for EndoPredict Tests The same paraffin blocks assessed by Oncotype DX were used for EndoPredict. Slides and sections for the EndoPredict assay for this study contained immediately adjacent tissues to those previously submitted to Genomic Health. The amount of invasive carcinoma tissue was at least 10% of the whole section surface in each case. One H&E section and three adjacent serial unstained slides (4 µm) were cut from each paraffin block. On the H&E slide, the area of the invasive tumor cells was identified under light microscope and marked with ink. The same area was also marked on the unstained slides. Tumor tissue was scraped from the unstained slides into a plastic tube using a scalpel permitting the analysis of almost 100% of invasive tumor tissue by the EndoPredict test. Total RNA was extracted using a silica-coated magnetic bead-based method as previously described RNA was eluted with 100 µL elution buffer and subjected to DNase digestion as described to get DNA-free total RNA. Unstained slides and H&E sections for both Oncotype DX and EndoPredict analysis were prepared in an identical way in the Institute of Surgical Pathology, University Hospital Zurich. ## Performance of EndoPredict Test The EndoPredict assay (Sividon Diagnostics, Cologne, Germany) was performed as published previously. In brief, expression of 8 genes–of-interest (*AZGP1, BIRC5, DHCR7, IL6ST, MGP, RBBP8, STC2, UBE2C*) and three reference genes (*CALM2, OAZ1, RPL37A*) as well as the amount of residual genomic DNA (*HBB*) were assessed by one-step RT-qPCR using the SuperScript III PLATINUM One-Step Quantitative RT-PCR System with ROX (Invitrogen, Karlsruhe, Germany) according to manufacturer’s instructions in a VERSANT® kPCR Molecular System (Siemens Healthcare Diagnostics). Sequences of primers and FAM/TAMRA-labeled probes were published previously. EP and EPclin scores as well as classification into low or high risk of distant metastasis were calculated from analytical PCR results, tumor size and nodal status using a web-based implementation as described previously. RT-qPCR analyses and calculations of EP and EPclin scores were performed by laboratory scientists in Sividon Diagnostics blinded to the results from the Oncotype DX tests. The scores and risk groups for each patient were subsequently transferred for analysis to one pathologist (Z.V). Extraction of a sufficient amount of RNA and generation of a valid EP score was possible for all 34 study samples. ## Statistics Statistical analysis was performed using the Pearson’s correlation coefficient. Signifcance was defined as p\<0.05. # Results ## Oncotype DX Recurrence Score (RS) Results of the individual patients were provided by Genomic Health to the submitting clinicians (A.T, C.T) and to the pathologists (A.N, Z.V, P.S, F.F, A.H). Recurrence Score (RS) revealed low risk in 15 patients, intermediate risk in 10 patients and high risk in 9 patients. ## EndoPredict Test The EndoPredict test results in an EP risk score and an EPclin score. According to the **EP** risk score 11 patients were classified as low risk and 23 patients as high risk. The EPclin score (combining EP risk score with tumor size and nodal status) re- classified 8 of the 23 EP high risk patients into the low risk group resulting in 19 patients with low and 15 patients with high risk of distant metastasis. ## Correlation and Concordance between Recurrence Score and EP Risk Score Comparing EP risk scores with Recurrence Score, a moderate (yet significant) correlation was found as reflected by a Pearson coefficient of 0.65 (p\<0.01). Nine of 15 of samples classified as low risk by the Recurrence Score were also low risk by EP score (60%). Nine of nine RS high risk samples were also EP high risk (100%). By combining the Oncotype DX intermediate risk and high risk groups to one high risk group, the concordance of classification in low or high risk between both tests was found in 26 of 34 cases (76%). The results of RS and EP scores are summarized in **,** **and** **.** ## Correlation and Concordance between Recurrence Score and EPclin Score Comparing the combined molecular-clinicopathologic EPclin score with the Recurrence Score the correlation was substantially smaller in comparison with the correlation between RS and EP scores. Pearson coefficient was 0.45 (p = 0.01). Eleven of 15 samples classified as low risk by the Recurrence Score were also low risk by EPclin score (73%). Six of nine RS high risk samples were EPclin high risk (66%). Combining the Oncotype DX intermediate risk and high risk groups to one high risk group, the concordance of classification in low or high risk between both tests was detected in 22 of 34 cases (65%). The results of RS and EPclin score are summarized in **,** **and** **.** ## Correlation and Concordance of Ki-67 to EP Score, EPclin Score and RS We could find a statistically significant but moderate correlation between the two molecular scores and proliferation index. No significant correlation was observed between the EPclin score and Ki-67. (Pearson coefficient varied as follows: to EP: 0.55 (p\<0.0001), to EPclin: 0.24 (p = 0.16), to RS: 0.56 (p\<0.0001). Results of continuous Ki-67 values and risk classes are illustrated in. ## Comparison of ER/PR/HER2 Status with Conventional Morphology and Oncotype DX Assay We detected a *high concordance* in hormone receptor and HER2 status between conventional morphology and Oncotype DX testing. 33 of 33 patients were positive for ER with immunohistochemical (IHC) analysis and with Oncotype DX assay (100%). 28 of 33 patients had identical PR status with both methodologies (85%). Three patients had PR positive cells in approximately 20%–30% of the tumor cells on immunohistochemistry, which were assessed as negative with Oncotype DX. The re- analysis of the immunohistochemical PR reactions confirmed small amount of positively stained nuclei. 31 of 33 patients had matching HER2 status with both methods (94%). One patient had a negative HER2 status by FISH which was assessed as equivocal with Oncotype DX. Another patient had *HER2* amplification by FISH, which was negative by Oncotype DX. The FISH *HER2* reaction was re-analyzed again and the amplification status could be confirmed. Results of hormone receptor/HER2 status and RS are summarized in. # Discussion In this study, we investigated retrospectively the correlation between EndoPredict scores and Oncotype DX Recurrence Scores using 34 hormone receptor positive breast cancer samples. Importantly, EndoPredict score showed a significant but only moderate correlation with the Recurrence Scores obtained by Oncotype DX testing. We found a moderate concordance of results regarding classification into risk groups between the two assays (reaching 76%) (two tiered). A major discrepancy between the two gene signatures was detected in 6 of 34 patients (18%) (three tiered). The discrepancy and moderate correlation of the two molecular scores EP and RS might be due to differences in weighting of main biological motives covered by the genes included in the test algorithms such as proliferation or ER signaling. Some differences might be explained by the coverage of other motives, e.g. cell adhesion, invasion, or DNA repair. Interestingly, an even smaller agreement is achieved, if the molecular RS is compared to the combined molecular-clinico- pathological EPclin score as opposed to the molecular EP score. This lower agreement is likely caused by the fact that the EPclin score considers additional prognostic information that may not be reflected by the tumor’s RNA expression. Following the EPclin-based classification into low or high risk of metastasis would spare 19 of 34 (56%) patients a cytotoxic chemotherapy in the light of an estimated 10-years distant metastasis-free survival of 96% of EPclin low risk patients in the two clinical validation studies. Nevertheless, further prospective clinical trials are needed to validate these results. Another multi-gene test, Mammaprint, was previously been compared with Oncotype DX. In this analysis, a higher concordance (81%) between high and intermediate risk groups from the Oncotype DX and poor prognostic groups of Mammaprint tests were shown. Our study showed a weaker concordance of 76% between EP sore and high/intermediate risk groups assessed by Oncotype DX. Together, different multigene may result in different treatment recommendation for individual patients. One limitation of previous studies is the sample size. Further analyses with longer patient survival data are necessary for the re- validation of these results. Oncotype DX is a RT-qPCR based 21-gene assay using *RNA* from FFPE tissue, comprising 16 cancer genes primarily related to tumor proliferation. This test is performed in a central reference laboratory. The NSABP B14 clinical trial validated, that patients with low RS developed significantly lower distant metastases than those patients with high RS. Analysis on prognostic value of RS as to distant metastases in early hormone receptor positive breast cancer has been the subject of several further clinical studies since the NSABP B14 trial chemotherapy. The consecutive clinical trial, the NSABP B20 validated the predictive value of RS on additional chemotherapy on hormone receptor positive nodal negative breast cancer patients. Data on 651 enrolled patients revealed that patients with high RS exhibited improved response to chemotherapy. On the other hand, it was also shown that patients with low risk RS did not benefit from additional chemotherapy. Response on chemotherapy in intermediate risk RS is currently being investigated in the ongoing TAILORx clinical trial. The EndoPredict assay is an RT-qPCR-based 12-gene test using RNA from FFPE tissue, specifically validated in two clinical studies for recurrence prediction in hormone receptor positive, HER2 negative, nodal negative and positive breast cancer treated with adjuvant hormonal therapy alone. The EP score provided significant prognostic information in addition to conventional prognostic clinico-pathological parameters such as tumor size, nodal status, grading, quantitative ER and Ki-67 as well as Adjuvant!Online. Moreover, the combination of the molecular EP score with tumor size and nodal status to the comprehensive molecular-clinico-pathological EPclin score outperformed the established prognostic parameters in these two patients cohort. Recently, it could be shown in a proficiency testing program including seven different pathological institutes that EndoPredict can be reliably performed in a decentralized setting in molecular pathological laboratories without the requirement of a reference lab. In contrast to EndoPredict, Oncotype DX assay includes an RT-qPCR based determination of hormone receptor and HER2 gene amplification. This fact prompted several previous studies to compare expression profile of these parameters. Excellent correlation with 100% concordance has been reported by O’Connor et al. in a series on 80 breast cancer samples. We found high concordance in ER/PR/HER2 status between Oncotype DX assay and established FISH or IHC assays, also regarded as a “gold standard”. There were only two discrepant cases for the HER2 status and six cases for progesterone receptors with no discrepant cases for estrogen receptors. This observation is in line with occasional false negative HER2 results reported as part of the Recurrence Score. Importantly, Geradts et al. detected discrepancies in hormone receptor and HER2 status determined by the conventional assays (IHC and/or FISH) and RT- PCR methodologies. There was only 56 to 66% categorical concordance. A similar result was found in a further study showing 33% of HER2 IHC-positive samples to be HER2 negative in RT-qPCR whereas the concordance of both methods in HER2 IHC- negative samples was 95%. It is not clear at this time, which methodology is superior in respect of predictive power. This needs to be addressed in future prospective trials. In current clinical practice, such discrepancies in the most important predictive breast cancer biomarkers are significantly hampering the treatment decision making process. Interestingly, a strong correlation between morphological parameter (especially histological tumor grading) and Recurrences Score was established in a few previous studies. The classification into two- tiered (low and high) risk categories with EndoPredict assay can possibly yield in clearer separation of intermediate risk patients. Concordance between Recurrence score and other prognostic assays or clinico- pathological parameter is of interest in clinical decision making. Significant linear correlation between proliferation index (Ki-67) and Recurrence Score was established previously in hormone receptor positive breast cancer. These data recommend the potential use of more cost effective immunohistochemical assessment of proliferation fraction rather than ordering highly expensive Oncotype DX testing. Another study by Tang et al. found good independent prognostic information in tamoxifen treated patients when Recurrence Score and individual clinico-pathological parameter were analyzed together. Interestingly, combining Adjuvant! Online recommendation with Recurrence Score did not provide better prognostic benefit in their analysis. Recently, a good agreement of prognostic risk assignment between the gene expression-based “intrinsic” subtype test PAM50 and Oncotype DX was described. Determining predictive markers with routine pathology assessment and using standardized reproducible criteria for morphological parameter (as grading, tumor size) represent a much less expensive alternative to multigene expression assays. It has been suggested that routine pathology markers are probably as reliable as genetic signatures at the current time, especially if combined mathematically. We could detect significant but moderate correlation between continuous proliferation index (Ki-67) and RS and EP scores. This is at least partially due to the lack of standardization in assessing the Ki-67 index in breast cancer. A considerable percentage of women diagnosed with breast cancer are aware of the valuable information multigene tests may add to their immediate therapeutic options. The impact of Oncotype DX testing is clearly reflected on altered recommendations or therapy decision in view of RS, which reportedly varies from 19 to 44% of the studied patients,. In conclusion, our data show moderate concordance between EndoPredict Score and Oncotyope DX results on individual patients. In the light of previous clinical and analytical validation data EP bears the promise to be an additional tool for decentralized multigene testing by local pathology with the advantage of the inclusion of important clinic-pathological data as nodal status. Further clinical studies are needed to compare both tests with regard to prediction of early and late distant metastasis, chemotherapy benefit and clinical outcome. The authors thank Ralf Kronenwett, Kerstin Bohmann and Franziska Haufe (Sividon Diagnostics, Cologne, Germany) for conducting the EndoPredict assays on the 34 patients and providing laboratory details of the reactions for the manuscript. [^1]: Hereby the authors confirm that the Sividon Diagnostics, Cologne funded and performed the EndoPredict analysis on the 34 study patients upon request of the Institute of Surgical Pathology, University Hospital Zurich. The funding was restricted only to this study. The authors further confirm that this does not alter their adherence to all the PLOS ONE policies on sharing data and materials. [^2]: Conceived and designed the experiments: ZV HM. Performed the experiments: ZV AN. Analyzed the data: ZV AN PSchraml HM. Contributed reagents/materials/analysis tools: PSinn FF AH CT AT. Wrote the paper: ZV AN PSchraml HM.

# Introduction More than 3,000 human genes encode transmembrane proteins that have been classified into receptors, transporters, enzymes, and other proteins. Transmembrane proteins such as solute carrier (SLC) transporters play vital roles in import and export of nutrients, metabolites, and ions. Examples of SLC substrates are sugars, amino acids, vitamins, neurotransmitters, urea cycle metabolites, and metals including iron. The human SLC transporter families comprise more than 400 genes and 52 subfamilies under the SLC family clusters, including MFS (major facilitator superfamily) and APC (amino acid-polyamine- organocation) superfamily. The majority of SLC transporter proteins has 12-transmembrane segment (TMS) with diversities typically ranging from 9 to 14 TMS. Iron metabolism is coordinately regulated by iron import, export, and storage proteins localized in plasma membrane and intracellular compartments. For instance, human transferrin receptor-1 (TfR1) is an 85 KDa homodimer of a plasma single-transmembrane protein serving as a major iron importing receptor for transferrin bound to two molecules of ferric iron (Fe³⁺). Iron is also subject to import and export by SLC transmembrane transporters. One is SLC11A1 and A2; particularly SLC11A2 (Divalent Metal Transporter—DMT1 or NRAMP2) which is a ubiquitously expressed transporter of iron along with cadmium, cobalt, manganese, and zinc to a lesser extent. DMT1 plays a pivotal role in apical iron (Fe²⁺) absorption (after Fe³⁺ is reduced by ferric reductase DcytB,) into duodenal enterocytes. DMT1 also transports endosomal iron released from transferrin (and reduced by Steap3) into the cytoplasm for supply of ferrous iron (Fe²⁺) in various cell types. The DMT1 gene encodes four isoforms of mRNAs through alternative promoter and splicing, giving rise to at least four different DMT1 proteins. The four isoforms of human DMT1 are 61–65 KDa transmembrane proteins having 12-TMS. Another SLC protein in charge of iron export is SLC40A1 (Ferroportin 1—Fpn1 or IREG1) that is highly expressed in basolateral membrane of duodenal enterocytes and exports iron into blood stream, where Fe²⁺ is oxidized to Fe³⁺ by membrane-bound hephaestin or ceruloplasmin circulating in blood stream, transferring Fe³⁺ to transferrin for systemic delivery of iron (Fe³⁺). Fpn1 is also highly expressed in reticuloendothelial macrophages that release iron for recycling after erythrophagocytosis. Fpn1 is a 63 kDa multipass transmembrane protein predicted to contain 12-TMS. Hepcidin, a 25-amino acid peptide secreted from the liver in response to iron overload and inflammation, binds and induces internalization and degradation of Fpn1 leading to decrease in plasma iron levels and iron distribution to tissues. When Fe²⁺ levels are more than what cells need, the surplus of iron is stored into ferritin shell composed of 24 multi-subunits of tissue-specific ratios of H (heavy) and L (light) chains. Iron (Fe²⁺) is stored in ferritin shell as Fe³⁺ catalyzed by the ferroxidase activity of the ferritin H subunit. Western blotting is an important technique to investigate protein expression, posttranslational modifications (e.g. phosphorylation), and interactions with other proteins and macromolecules, in which we mostly use commercially available antibodies for detection of proteins of our interests. There are numerous monoclonal and polyclonal antibodies successfully used for western blotting and other applications (CiteAb, <https://www.citeab.com/>), while we still experience antibodies not working or far below our expectation of qualities for unknown reasons. To verify the quality and specificity of antibodies, we should take positive and negative control samples such as cell lysates prepared after transient transfection of expression plasmids or knockdown of endogenous mRNAs. In particular, detection of transmembrane proteins in western blotting has often been challenging due to difficulties in solubilization of those proteins with detergents and reducing agents. There are numerous numbers of transmembrane protein western blots so far published, including DMT1 and Fpn1 proteins (CiteAb). In this report, we initially judged our DMT1 and Fpn1 antibodies not working in our routine western blotting protocol; however, reassessment and modification of our protocol in a sample preparation step significantly improved the resolution of western blots. # Materials and methods ## Cell culture and maintenance 12 human cell lines used in this study and their culture media are listed in. All cell lines were cultured at 37°C in a humidified CO₂ incubator (5% CO₂, Sanyo model MCO-17AIC). Undifferentiated Caco-2 cells were used in this study. Adherent cells were passaged by trypsinization (0.25% Trypsin, 2.6 mM EDTA in calcium/magnesium free PBS) every 3–5 days when they were confluent. Suspension cells were passaged by 20\~30-fold dilution to fresh growth media every 4–6 days. ## Chemicals, and reagents All chemicals and reagents used in this study, their suppliers, and catalog numbers are listed in. Ferric ammonium citrate (FAC) and desferrioxamine (DFO) were dissolved in distilled water at 100mM and 25mM, respectively. Hemin was dissolved in 0.1M NaOH at 50mM. Human bioactive hepcidin was dissolved in water at 0.5mM, added to undifferentiated Caco-2 culture at 0.5uM, and incubated for 18–24 hr. ## Western blot and antibodies Whole cell lysates (WCLs) were prepared from sub-confluent or confluent cells grown in 60 mm culture plates (Greiner, 628160). In some experiments, cells were treated with FAC, DFO, or hemin for 18–24 hr before harvest. Cells were washed with phosphate buffered saline (PBS: 137mM NaCl, 27mM KCl, 15mM KH₂PO₄, 81mM Na₂HPO₄) and lysed in 200-400ul of PBS-based RIPA buffer (1X PBS, 1% NP40, 0.5% sodium deoxycholate, 0.1% SDS) containing 1X protease inhibitor cocktail (Calbiochem/Millipore). In, we also prepared HEK293 WCL in RIPA buffer sonicated for 10 sec. three times in ice-water (550 Sonic Dismembrator, Fisher) or WCL in transmembrane extraction reagent (Five Photon Biochemicals). We did not centrifuge WCLs to remove DNA or debris. Protein concentrations in WCLs were measured with 1X Bio-Rad protein assay dye reagent. 15–30 ug of WCLs were mixed with 2X SDS-PAGE (polyacrylamide gel electrophoresis) sample buffer containing 62.5mM Tris-HCl (pH 6.8), 25% Glycerol, 2% SDS, 0.01% bromophenol blue, and 5% β-mercaptoethanol. The volume of 2X SDS-PAGE sample buffer was equal or more than varied WCL volumes for an equal amount of total proteins. For preparation of heated and non-heated samples, one group was heated at 95°C for 5min in a dry bath incubator (11-718-2, Fisher) containing water in a heating block, while another group was left at room temperature. After briefly spinning down sample tubes, WCLs were loaded on 10% SDS-PAGE (10% acrylamide, 0.27% bisacrylamide, 375mM Tris-HCl pH 8.8, and 0.1% SDS) mini-gel (4x2.9 inches). The stacking gel was 5% acrylamide containing 0.13% bisacrylamide, 125mM Tris-HCl pH 6.8, and 0.1% SDS. Protein size markers (Bio-Rad Precision Plus Protein Standards) was loaded without heating. The electrophoresis was run at 12.5mA per mini-gel for approximately 1.5hr in the running buffer containing 25mM Tris, 192mM Glycine, and 0.1% SDS. Proteins separated on the gel were transferred to PVDF membrane (Thermo Scientific, 88518) in a mini trans-blot cell (Bio-Rad, 022711PM) at 300mA in a transfer buffer (25mM Tris pH 8.3, 192mM Glycine, and 20% methanol) at 4°C for 1.5–2hr. The PVDF membrane was blocked for 20-30min at room temperature in either 1% BSA or 5% skim milk dissolved in 0.1% Tween 20-containing Tris- buffered saline (TBS: 20mM Tris-HCl pH 7.6 and 137mM NaCl) and incubated with a primary antibody on a rocker platform (Speci-Mix, Thermolyne) at 4°C overnight. The primary antibodies for western blots listed in were diluted with either 5% skim milk in 0.1% Tween/TBS for anti-ferritin H (1:1,000), anti-TfR1 (1:20,000), anti-DMT1/SLC11A2 (1:1,000), anti-GAPDH (1:5,000), anti-Flag (1:1,500), or 1% BSA in 0.1% Tween/TBS for anti-Fpn1/SLC40A1 (1:1,000\~1,500). The PVDF membranes were washed with approximately 50 ml of 0.1% Tween/TBS for 10-20min, three times at room temperature, and incubated with a secondary antibody as listed in (5,000–8,000-fold diluted with the same dilution solution as a primary antibody, either 5% skim milk or 1% BSA in 0.1% Tween/TBS) for 1.5hr at room temperature. The PVDF membranes were washed with 0.1% Tween/TBS three times, incubated with ECL reagents at room temperature, and exposed to X-ray films (ProSignal Blotting Film, Genesee Scientific 30-507L or 30-810L). For successive incubation with another primary antibody, the membrane was soaked at room temperature for 1 hr in the stripping solution (pH 2.2 with HCl) containing 1.5% glycine, 0.1% SDS, and 1% Tween 20 (from Abcam’s western blot membrane stripping for restaining protocol). After washing membrane in 0.1% T/TBS for 15min three times, the procedure was repeated from blocking the membrane with 1% BSA or 5% skim milk in 0.1% T/TBS. ## Isolation of cytoplasmic and nuclear fractions Cells were washed with PBS, resuspended in 200-500ul hypotonic buffer (20mM Tris-HCl, pH 7.4, 10mM NaCl, 3mM MgCl₂) at 4°C for 15min, then mixed with 1/20 volume of 10% NP40, vortexed for 10 sec, and centrifuged at 12,000 rpm for 30 sec. The supernatant was recovered (cytoplasmic fraction), and the pellet was resuspended in 30-50ul of extraction buffer containing 100mM Tris, pH 7.4, 2mM Na₃VO₄, 100mM NaCl, 1% Triton X-100, 1mM EDTA, 10% glycerol, 1mM EGTA, 0.1% SDS, 1mM NaF, 0.5% deoxycholate, 20mM Na₄P₂O₇, and 1X proteinase inhibitor cocktail. The pellet suspension was vortexed for 10 sec, incubated on ice for more than 30min, vortexed for 30 sec, and centrifuged at 12,000 rpm for 10min at 4°C. The supernatant (nuclear fraction) was recovered. ## Plasmids, siRNA, and transfection pCMVSport6Fpn1 (MGC human SLC40A1 sequence-verified cDNA, clone ID: 5213437, MHS6278-202801506) was purchased from GE Dharmacon. pBluescriptR-DMT1 (SLC11A2 MGC human clone ID: 5300266, MHS1010-98075408) was purchased from Open Biosystems. The 2.5kb DMT1 cDNA was isolated after digestion with NcoI and HindIII followed by filling-in with DNA pol. Klenow fragment, and cloned into SmaI site of pCMV and pCMVFlag vectors. These plasmid DNAs, verified by DNA sequencing (Eurofins), were transiently transfected into HEK293 cells by electroporation (X-Cell, Bio-Rad) as described previously. Human Fpn1 siRNA was purchased from Sigma-Aldrich (`GAUGGAACUUGGUAUCCAU`\[dT\]\[dT\], and `AUGGAUACCAAGUUCCAUC`\[dT\]\[dT\]) and 5ul of 20uM solution was transfected into Caco-2 cells (grown in 2ml media/35 mm plate) with 200ul of Opti-MEM containing 6ul of Lipofectamine RNAiMax (Invitrogen) for 18hr, followed by treatment with 250 and 500uM FAC for additional 8hr in the presence of siRNA prior to harvest. Knockdown of Fpn1 in Caco-2 cells was first verified by qPCR as described previously using three sets of human Fpn1 primers. 1. Set1 forward: `5’-CTACTTGGGGAGATCGGATGT-3’` 2. Set1 reverse: `5’-CTGGGCCACTTTAAGTCTAGC-3’` 3. Set2 forward: `5’-TGGATGGGTTCTCACTTCCTG-3’` 4. Set2 reverse: `5’-GTCAATCCTTCGTATTGTGGCAT-3’` 5. Set3 QuantiTect Primer Assays for Hs_SLC40A1 (QT00094843, Qiagen) # Results ## Routine practice of heating samples in Fpn1 (SLC40A1) western blot According to the citation history for Fpn1 antibody in CiteAb, the rabbit polyclonal anti-Fpn1 antibody NBP1-21502 (Novus Biologicals) has been most successfully used in research article publications (<https://www.citeab.com/>). After preliminary testing of optimum dilutions of this primary antibody with either 5% skim milk or 1% BSA in Tween/TBS buffer, we decided to use 1,000–1,500-fold dilution in 1% BSA in Tween/TBS. We prepared WCL samples from several culture cell lines, heated at 95°C for 5min, loaded on a 10% acrylamide SDS-PAGE gel, and performed western blotting. The blot gave rise to multiple bands including the putative Fpn1 band\* (63 kDa) between 50 and 75 kDa size markers. The expression of Fpn1 can be increased by high iron levels through dissociation of IRE-binding protein 1 and 2 (IRP1 and IRP2) from the iron-responsive element (IRE) in 5’-UTR of Fpn1 mRNA. Additionally, another Fpn1 mRNA (termed FPN1B) transcribed from an alternative promoter lacks the IRE but has the same open reading frame, therefore FPN1B expression is not responsive to cellular iron status. To characterize the putative Fpn1 band further, we treated SW480 cells with ferric ammonium citrate (FAC) or cisplatin that directly inhibits IRP2 binding to IRE and performed western blotting. The putative 63 kDa faint band was not increased by FAC but slightly increased by cisplatin treatment in SW480 cells. In contrast, ferritin H was significantly increased by FAC and cisplatin because ferritin H mRNA also contains a functional 5’-IRE, suggesting that FAC and cisplatin treatments are working to drive 5’-IRE-dependent translational upregulation. Collectively, we demonstrate that the most popular anti-Fpn1 antibody gave rise to the band around 63 kDa but it did not show clear regulation through the IRP-IRE system in SW480 cells. Thus, without taking a positive control for western blotting, the identity of the 63 kDa band in remained undetermined. ## No solubilization problem but heating samples may compromise Fpn1 western blot To validate the putative Fpn1 band and anti-Fpn1 antibody further, we made WCL from HEK293 cells transiently transfected with a human Fpn1 expression plasmid and performed western blotting in a routine sample preparation and heating protocol. The overexpressed Fpn1 (untagged) gave rise to one specific band that was stuck on the top of the separation gel. This result indicates that the Fpn1 antibody (NBP1-21502, Novus Biologicals) can detect at least overexpressed human Fpn1 protein by western blotting, in which the Fpn1 protein appeared to be insolubilized or aggregated. To solve the potential solubilization problem first, we used three different cell lysis conditions; RIPA lysis buffer, RIPA plus sonication, and transmembrane extraction reagent (Five Photon Biochemicals). To solve the aggregation problem, we compared heated samples at 95°C to no heating (room temperature) of WCL from HEK293 transfected with pCMV-empty or -Fpn1 plasmid. As shown in, regardless of three different lysis conditions, transfected Fpn1 was stuck on the top of the gel when they were heated at 95°C prior to gel loading. In contrast, no heating samples gave rise to a specific Fpn1 band between 50 and 75kDa markers regardless of their different lysis conditions. Furthermore, HEK293 WCL in RIPA mixed with three different 2X SDSPAGE sample loading buffers containing 5% β-mercaptoethanol (the regular sample loading buffer, see), 4M urea, or both 5% β-mercaptoethanol and 4M urea did not improve the migration problem of overexpressed Fpn1 protein when they were heated. By contrast, no heating of these WCL samples showed the right size of Fpn1 band regardless of the use of different loading buffers. Taken together, we concluded that no heating of protein samples is the most critical factor for the right resolution and detection of transfected Fpn1 protein by western blotting with this anti-Fpn1 polyclonal antibody. ## Heating samples compromised DMT1 (SLC11A2) western blot Another key SLC protein handling cellular iron traffic is DMT1 (SLC11A2). According to the citation data in CiteAb, there are more than 10 different anti- DMT1 antibodies used for publications. Among them, we used a rabbit monoclonal anti-DMT1 antibody (D3V8G, Cell Signaling Technology). To test whether heating protein samples also affects the resolution of DMT1 western blots, we prepared WCLs in RIPA or cytoplasmic and nuclear fractions from HEK293 cells transiently transfected with human DMT1 (untagged) or Flag-DMT1. Samples were mixed with 2x SDSPAGE sample loading buffer (containing 5% β-mercaptoethanol), heated at 95°C for 5min or left at room temperature, and performed SDS-PAGE and western blotting with anti-DMT1 or Flag antibody. As observed in the Fpn1 western blots in, heating protein samples significantly impaired the DMT1 protein bands. In clear contrast, no heating of samples (left at room temperature until sample loading) showed good resolution of \~75kDa and additional larger and smaller bands detected with anti-Flag (top) as well as anti-DMT1 antibodies (, bottom). DMT1 mRNA isoforms contain 3’-UTR IRE that should be subject to mRNA destabilization when IRP1 and/or IRP2 dissociates from the IRE in iron overloaded conditions or by inactivation of IRP2 following cisplatin treatment. Indeed, endogenous DMT1 proteins were detected by western blotting in unheated WCLs isolated from SW480 cells treated with FAC (0.1–0.5 mM) or cisplatin (2–100 ug/ml), in which FAC or cisplatin treatment slightly decreased DMT1 protein levels. Collectively, no heating of protein samples prior to loading to SDS-PAGE gel is critical for detection of transfected and endogenous DMT1 proteins by western blotting. ## Effects of heating samples from 12 human cell lines on DMT1 and Fpn1 western blots To assess the effect of heating samples on detection of endogenous DMT1 and Fpn1 proteins in various cell types by western blotting, we prepared WCLs from 12 human cell lines (lysed in RIPA), mixed with 2x SDSPAGE sample loading buffer, heated at 95°C for 5min or not heated (left at room temperature), and performed SDSPAGE and Western blotting. Given the results of transfected Fpn1 and DMT1 in Figs and, only unheated WCLs from HEK293 cells transfected with empty vector, DMT1, and Fpn1 were loaded together as positive controls of western blots. As shown in, endogenous DMT1 proteins were detected in most of these human cell lines; however, heating samples at 95°C for 5min reproducibly impaired the major DMT1 band, some of which were stuck on the top of the gel (arrow on the right). Fpn1 western blots were more complicated because of multiple major bands in addition to no clear difference between heated and unheated samples. This was an unexpected result because transiently transfected Fpn1 protein in HEK293 cells was detected only when the samples were not heated prior to loading to SDSPAGE gel. The Fpn1 western blotting was working in because transiently transfected Fpn1 used as a positive control was detected in the range between 50 and 75kDa size markers. The endogenous protein showing the same migration as transfected Fpn1 was also detected in several cell types such as Caco-2 cells. As our Fpn1 western blots gave rise to multiple non-specific bands, we further verified the endogenous band migrating together with transfected Fpn1. First, transfection of Fpn1 siRNA into undifferentiated Caco-2 cells reduced Fpn1 mRNA level to 35.3% (p = 0.007 in unpaired t-test, FAC 0) and consistently diminished the endogenous protein band co-migrating with transfected Fpn1. Second, as hepcidin binds Fpn1 and induces Fpn1 internalization and degradation in Caco-2 cells, we treated Caco-2 cells (3 plates each) with bioactive hepcidin at 500 ng/ml for 24hr and performed Fpn1 western blotting, in which hepcidin-treated cells appeared to have decreased Fpn1 protein levels. Collectively, we concluded that the band co-migrating with transfected Fpn1 is the endogenous Fpn1 protein. ## Effects of heating samples for single-transmembrane TfR1 and cytoplasmic ferritin western blots We elucidated the effect of sample heating on western blots for two other key iron transport and storage proteins; single-transmembrane TfR1 and cytoplasmic ferritin. We prepared WCL in RIPA from Caco-2 cells treated with hemin, FAC, and an iron chelator deferoxamine (DFO) for 24hr. We compared heated samples with unheated samples for detection of each endogenous protein by western blotting. In Fpn1 western blotting, transiently transfected Fpn1 in HEK293 was reproducibly detected in unheated samples, whereas heated samples failed to show the right size of Fpn1 band between 50-75kDa protein size markers, rather showing very faint bands stuck on the top of the gel (arrow, right). FAC or hemin treatment in Caco-2 cells increased a band migrating similar to transfected Fpn1, in which the unheated group of samples showed slightly stronger bands compared to heated samples at 95°C for 5min. Conversely, we anticipated that the iron chelator DFO would inhibit translation of Fpn1 mRNA via the 5’-UTR IRE; however, we repeatedly failed to observe it (DFO compared to none). The endogenous DMT1 was consistently detected only when samples were not heated. In contrast, the H subunit of ferritin (ferritin H) was detected as increased by hemin or FAC treatment only in heated samples. As ferritin is composed of 24 subunits of ferritin H and L, heat denaturation of samples is necessary for ferritin western blot otherwise we often observed no ferritin bands or ferritin stuck on the top of the gel. Interestingly, western blot for an 85 KDa homodimer of a plasma single-transmembrane protein TfR1 showed that unheated samples gave much stronger bands in an iron-responsive manner (iron-mediated downregulation) than samples heated at 95°C for 5min. The unheated samples gave rise to an additional higher molecular weight band, probably a dimer of TfR1, in an iron-responsive manner (asterisk). The cytoplasmic protein GAPDH as a protein loading control was detected equally in both heated and unheated samples. Similar results were observed in HepG2 WCLs except for the endogenous Fpn1 band showing no clear differences between heated and unheated samples, both of which showed clear iron-induced Fpn1 expression following FAC treatment. # Discussion Western blotting techniques were improved and seemingly have established, which may make us often omit in publications describing critical steps and tips for detection of protein of interest in western blotting. This is partly because a standard protocol for sample preparation has a routine protein denaturation step by heating at 95–100°C in a sample loading buffer containing SDS and a reducing agent such as β-mercaptoethanol or dithiothreitol. This heating step is understood important for antibodies to find epitopes efficiently as well as to detect a specific protein in multiple protein binding complexes or a subunit of multimeric proteins as exemplified by the detection of ferritin in this study. However, detection of transmembrane proteins by western blotting seems to be more challenging and complicated. There are more than 3,000 human genes encoding transmembrane proteins, in which \~ 1,300 receptors including seven- transmembrane GPCR (G protein coupled receptor) and single-pass transmembrane receptors, \~ 800 transporters such as SLC, ABC (ATP binding cassette) transporters, and channels, and \~ 500 enzymes including transferases, hydrolases, and oxidoreductases. We sometimes find researcher’s troubleshooting discussions via internet medias about unsuccessful western blotting for detection of particular transmembrane proteins. Along with our initial failure to detect SLC family iron transporters in this study, we anticipate that routine heat denaturing protocols may not work or significantly loose band signals in western blotting to detect many more transmembrane and heat-sensitive proteins, perhaps due to heat-induced protein aggregations and/or conformational changes. In this study, we observed that non-heated samples solely solved the migration and resolution problems of DMT1 and overexpressed Fpn1 western blots, whereas three different cell lysis conditions or three different 2X sample loading buffers did not improve the situation. Despite the fact that both DMT1 and Fpn1 are 12-transmembrane proteins, heating samples at 95°C for 5min ruined DMT1 western blot (Figs,) whereas only partially impaired endogenous Fpn1 western blot (Figs and). However, heating samples completely impaired the migration of transfected Fpn1 (untagged) being stuck on the top of the separation gel. We do not have good explanations for these different outcomes in sample heating between DMT1 and Fpn1 12-TMS proteins and between transfected and endogenous Fpn1 proteins. However, these results alert us to be aware of potentially different effects of sample heating on western blotting of any other SLC transporters and transmembrane proteins. It is therefore important to find the optimum condition of sample preparation for each transmembrane protein western blotting. This was also the case in TfR1 western blotting, in which non-heated samples tended to give rise to stronger and consistent iron status-dependent TfR1 band despite showing the additional TfR1 dimer band. In contrast, heating samples was essential for detection of cytoplasmic ferritin H subunit protein in western blotting because ferritin is assembled of 24 subunits of H and L proteins. These results suggest that a single western blot loaded only either heated or non-heated protein samples cannot be used for successive reprobing with DMT1, Fpn1, TfR1, and ferritin antibodies. Another cytoplasmic protein GAPDH, often used as a protein loading control of western blotting, was not affected by heated or unheated protein samples. It should be noted that, although we observed no improvement for Fpn1 western blotting by switching from RIPA to RIPA plus sonication or commercially available membrane protein extraction lysis buffer, the lysis buffer components together with mechanical destruction of membranes may also significantly improve western blotting for some other transmembrane proteins more than the impact of heating protein samples. It may be a routine practice for laboratories and researchers primarily working on transmembrane proteins to avoid heat-denaturing protein samples prior to gel loading. Alternatively, they may have additional technical tips for particular transmembrane protein western blotting that we have not tested in this study. Indeed, no sample heating prior to gel loading for Fpn1 western blotting was noted in some earlier publications, while we observed relatively few attempts of DMT1 western blotting in earlier publications or no particular notes for western sample preparations except examples such as pre-incubation at 37°C for 15min by Nam et al. CiteAb records more than 50 citations for several applications with the rabbit polyclonal anti-Fpn1 antibody (NBP1-21502, Novus Biologicals) that we used in this study, in which more than 30 publications used this antibody for western blotting. Most of them did not clarify sample heating conditions or heated at 90–100°C, except 75°C for 10min or no heating. Even if protein samples might have been preheated in most of these publications, they might have observed the right size of the endogenous Fpn1 band as we observed in this study (Figs). CiteAb also lists more than 10 publications for DMT1 western blotting using Proteintech, Abcam, or Novus Biologicals anti-DMT1 antibody, only some of which noted that they treated samples with gentle heating at 37°C for 30min or 50°C for 15min. Anti-DMT1 antibodies that have different epitopes from our anti- DMT1 antibody (Cell Signaling Technology, near the N-terminus of human DMT1) may also affect the optimum conditions of protein sample preparation for better resolution of western blotting. # Conclusion In summary, only non-heated protein samples worked for western blotting with the anti-DMT1 (SLC11A2) monoclonal antibody having the epitope near the N-terminus. Non-heated protein samples gave rise to better resolution and/or stronger band signals in western blotting with anti-Fpn1 (SLC40A1) and anti-TfR1 polyclonal antibodies. Three different lysis buffers, three sample loading buffers, or sample sonication we tested did not improve the resolution of DMT1 and Fpn1 western blots. In contrast, only heated protein samples worked for ferritin western blotting with the anti-ferritin H monoclonal antibody, while sample heating did not affect western blotting for another cytoplasmic protein GAPDH frequently used for a protein loading control. Collectively, either heated or unheated samples most critically determined the quality and outcome of transmembrane protein western blotting we tested, suggesting that there should be many similar cases for western blotting of other transmembrane and heat- sensitive proteins. # Supporting information 10.1371/journal.pone.0235563.r001 Decision Letter 0 Missirlis Fanis Academic Editor 2020 Fanis Missirlis 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 May 2020 PONE-D-20-06977 Transmembrane Protein Western Blotting: Impact of Sample Preparation on Detection of SLC11A2 (DMT1) and SLC40A1 (ferroportin) PLOS ONE Dear Dr. Tsuji, 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 address all comments by reviewers, who are in general supportive of the publication of these results. There are two issues of concern, one related to repetitions/statistics, where I ask the authors for clarity above all. Likewise, there was some concern of the use of photoshop in some of the figures, please read carefully and adhere to PLOS One policy with respect to the images and state that you have done so in your reply. We would appreciate receiving your revised manuscript by Jun 07 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, Fanis Missirlis, Ph.D. 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\. Thank you for stating the following in the Acknowledgments Section of your manuscript: "Funding: This work was supported in part by P30ES025128 from the National Institute of Environmental Health Sciences to the Center for Human Health and the Environment (CHHE)." We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please provide an amended statement that declares \*all\* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at <u>[http://journals.plos.org/plosone/s/submit-now](about:blank)</u>.  Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement. Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." 3\. PLOS ONE now requires that authors provide the original uncropped and unadjusted images underlying all blot or gel results reported in a submission’s figures or Supporting Information files. This policy and the journal’s other requirements for blot/gel reporting and figure preparation are described in detail at <https://journals.plos.org/plosone/s/figures#loc-blot-and-gel- reporting-requirements> and <https://journals.plos.org/plosone/s/figures#loc- preparing-figures-from-image-files>. When you submit your revised manuscript, please ensure that your figures adhere fully to these guidelines and provide the original underlying images for all blot or gel data reported in your submission. See the following link for instructions on providing the original image data: <https://journals.plos.org/plosone/s/figures#loc-original-images-for-blots-and- gels>. In your cover letter, please note whether your blot/gel image data are in Supporting Information or posted at a public data repository, provide the repository URL if relevant, and provide specific details as to which raw blot/gel images, if any, are not available. Email us at <plosone@plos.org> if you have any questions. 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 more 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 sensitive information, data are owned by a third-party organization, etc.) 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. 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. \[Note: HTML markup is below. Please do not edit.\] 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: Partly Reviewer \#2: Yes Reviewer \#3: Yes \*\*\*\*\*\*\*\*\*\* 2\. Has the statistical analysis been performed appropriately and rigorously? Reviewer \#1: No Reviewer \#2: N/A Reviewer \#3: No \*\*\*\*\*\*\*\*\*\* 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: PLOS ONE Transmembrane Protein Western Blotting: Impact of Sample Preparation Overview: A study investigating the effect of heating vs non-heating of protein samples prior to loading and its subsequent effect on detection of SLC11A2, SLC40A1 and TfR1. Research groups have previously published SLC11A2 and SLC40A1 Western blots where the protein samples were heated prior to loading. Heating samples did not affect the bands in these studies 1,2. The author describes in the discussion, research groups have also previously published western blots where the samples were not heated prior to loading. Therefore, I am uncertain about the novelty of this study. Unfortunately, this was a very confusing paper to follow. The use of 12 cell lines to investigate the effect of heating may have been the reason for this confusion. My advice is to select a few cell lines which are involved in iron metabolism (i.e: Caco-2 and HepG2) and investigate what effect heating vs non- heating has on iron proteins from lysates obtained from these lines. Moreover, it would be useful to see if the effects observed are also present when other commercially available antibodies are used. Further comments Methods: Please include a section about cell culture maintenance. Please indicate how many days following transfection were the Caco-2 cells used? Did the authors verify knockdown using qPCR? Were Caco-2 cells undifferentiated/differentiated when lysed? Results: Please include figure captions. 1\. Kondaiah P, Aslam MF, Mashurabad PC, Sharp PA, Pullakhandam R. Zinc induces iron uptake and DMT1 expression in Caco-2 cells via a PI3K/IRP2 dependent mechanism. Biochemical Journal. 2019;476(11):1573-1583. 2\. Scheers NM, Almgren AB, Sandberg A-S. Proposing a Caco-2/HepG2 cell model for in vitro iron absorption studies. The Journal of nutritional biochemistry. 2014;25(7):710-715. Reviewer \#2: In the present work, Author Yoshiaki Tsuji investigated the impact of sample preparation for an optimal detection by western blotting of specific transmembrane proteins, SLC11A2 (DMT1) and SLC40A1 (ferroportin). This work is based on initial observation of the Author concerning a mediocre functioning of SLC11A2 and SLC40A1 antibodies in his routine western blotting protocol. In the MS Author nicely and adequately explains, step by step, how specific modifications in sample preparation lead to significant improved western blot resolution for membrane proteins. As far as technical issue for an optimal study of transmembrane protein is concerned, the MS deals about a subject of great interest and concerns proteins involved in iron metabolism, for which Author’s expertise is well established. That said I have two minor points to address to the Author: \- Discussion, first page. The sentence beginning with “Given researcher’s troubleshooting discussions…” should be rephrased because it is too long and complex in this present form. \- I understand protocols were tested on different cell lines. Were lysates from primary cell cultures been also tested? Reviewer \#3: The manuscript addresses the question how different sample preparation has major affects on the outcome of Western blotting of several proteins related to iron metabolism. This is a worthy study as it can save many researchers a lot of pain, and improve study-outcomes. Remarks The authors relate in the abstract to DMT1 as an iron exporter. In most cases it though functions as an importer, when viewing the cytosol of a cell as “in” and the endosome as a part of internalized “out”. Later in the text DMT1 is termed a “transporter”, which does it more justice. It would be much clearer to use Fpn and DMT1 throughout the paper, and not label the figures in one way and use in the text the SLC terms, sometimes in full and sometimes shortened e.g. SLC40. In the text, a reference to fig. 2A is missing. In Figure 3 B, the evidence for a decrease of DMT1 with increased FAC is weak and the decrease with increased Cisplatin is not convincing at all. Can you please quantify these graphs and also indicate how often each experiment was repeated. Indicate the number of times the experiments were repeated throughout the manuscript, best in the legends. There seems to be a discrepancy between Fpn results in fig. 6a and 4b. Any explanation? Figure legend 1 has a different format than the other legends, giving lot’s of info on blocking and washes. Could be omitted. Giving these details in materials and methods is enough. \*\*\*\*\*\*\*\*\*\* 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 \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] 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.0235563.r002 Author response to Decision Letter 0 17 Jun 2020 Response to Reviewers Manuscript ID#: PONE-D-20-06977 Manuscript Title: Transmembrane Protein Western Blotting: Impact of Sample Preparation on Detection of SLC11A2 (DMT1) and SLC40A1 (ferroportin) Corresponding Author: Dr. Yoshiaki Tsuji Reviewer \#1 Overview: A study investigating the effect of heating vs non-heating of protein samples prior to loading and its subsequent effect on detection of SLC11A2, SLC40A1 and TfR1. Research groups have previously published SLC11A2 and SLC40A1 Western blots where the protein samples were heated prior to loading. Heating samples did not affect the bands in these studies 1,2. The author describes in the discussion, research groups have also previously published western blots where the samples were not heated prior to loading. Therefore, I am uncertain about the novelty of this study. Unfortunately, this was a very confusing paper to follow. The use of 12 cell lines to investigate the effect of heating may have been the reason for this confusion. My advice is to select a few cell lines which are involved in iron metabolism (i.e: Caco-2 and HepG2) and investigate what effect heating vs non-heating has on iron proteins from lysates obtained from these lines. Moreover, it would be useful to see if the effects observed are also present when other commercially available antibodies are used. Response: Thank you for very constructive and helpful overview. First, we have never published SLC11A2 (DMT1) western blotting, retrospectively it was not successful because we routinely heated our samples and failed to detect convincing DMT1 bands. Second, yes this reviewer is right; we published SLC40A1 (Fpn) western blots using unheated samples in very low Fpn1-expressing cells (Cell Chem. Biol., 26:85-97, 2019). Third, as this reviewer pointed out, several papers showing DMT1 or Fpn1 western blots (in addition to the reviewer’s ref. 1 and 2) used heated samples (e.g. 100oC 10 min, 90oC 2 min, 75oC 10 min, 37oC 30 min), while most papers did not describe specific conditions for preparation of protein samples (discussed in this manuscript). I feel that some of published western blots for DMT1 (probably Fpn1 also) as well as some other SLC family members may not be accurate (particularly blots without indicating positions of protein size markers) or detected only partial quantities that did not aggregate even after heating, as seen in Figures 4 and 6. Please note that we have the following statement in the discussion (page 24) for PLOS ONE readers: “Indeed, no sample heating prior to gel loading for Fpn1 western blotting was noted in some earlier publications (22, 26, 28), while we observed relatively few attempts of DMT1 western blotting in earlier publications or no particular notes for western sample preparations except examples such as pre-incubation at 37oC for 15 min by Nam et al (44)”. Forth, all cell lines used in this study, including Caco-2 and HepG2, tightly regulate iron metabolism in response to environmental cues therefore have been used for investigation of the mechanisms of iron transport and storage. We hope that showing our results in all these cell lysates would help more researchers perform DMT1, Fpn1, or other transmembrane protein western blots with specific antibodies used in this study as well as their specific antibodies. We appreciate all of these reviewer’s important points. Comment 1.1 Methods: Please include a section about cell culture maintenance. Response: A new section of “Cell Culture and Maintenance” has been created in the Materials and Methods (page 6). Comment 1.2 Please indicate how many days following transfection were the Caco-2 cells used? Response: We have included the experimental timeline in the sub-section of Plasmids, siRNA, and Transfection in the Materials and Methods, page 12). Comment 1.3 Did the authors verify knockdown using qPCR? Response: Yes, we verified Fpn1 knockdown in untreated Caco-2 cells in the experiment shown Fig. 5A. The qPCR results (35.3%, p=0.007: Fpn1 mRNA with transfection of siFpn1 compared to siControl) were included in the main text (page 18) without showing the graph in order to stay focus on technical tips in transmembrane western blotting. The qPCR methods and primer sets have been included accordingly in the Materials and methods (page 11). Comment 1.4 Were Caco-2 cells undifferentiated/differentiated when lysed? Response: They were undifferentiated. In the revised manuscript, we have added a note that “undifferentiated Caco-2 cells were used in this study” in the “Cell Culture and Maintenance” section (page 6) along with in the main text (page 18). Thank you. Comment 1.5 Results: Please include figure captions. Response: Thank you, we have included figure captions in all figures. 1\. Kondaiah P, Aslam MF, Mashurabad PC, Sharp PA, Pullakhandam R. Zinc induces iron uptake and DMT1 expression in Caco-2 cells via a PI3K/IRP2 dependent mechanism. Biochemical Journal. 2019;476(11):1573-1583. 2\. Scheers NM, Almgren AB, Sandberg A-S. Proposing a Caco-2/HepG2 cell model for in vitro iron absorption studies. The Journal of nutritional biochemistry. 2014;25(7):710-715. Reviewer \#2 In the present work, Author Yoshiaki Tsuji investigated the impact of sample preparation for an optimal detection by western blotting of specific transmembrane proteins, SLC11A2 (DMT1) and SLC40A1 (ferroportin). This work is based on initial observation of the Author concerning a mediocre functioning of SLC11A2 and SLC40A1 antibodies in his routine western blotting protocol. In the MS Author nicely and adequately explains, step by step, how specific modifications in sample preparation lead to significant improved western blot resolution for membrane proteins. As far as technical issue for an optimal study of transmembrane protein is concerned, the MS deals about a subject of great interest and concerns proteins involved in iron metabolism, for which Author’s expertise is well established. That said I have two minor points to address to the Author: Comment 2.1: Discussion, first page. The sentence beginning with “Given researcher’s troubleshooting discussions…” should be rephrased because it is too long and complex in this present form. Response: Thank you, I totally agree. In the revised manuscript, it has been divided into two sentences and corrected to “We sometimes encounter researcher’s troubleshooting discussions via internet medias about unsuccessful western blotting for detection of particular transmembrane proteins. Along with our initial failure to detect SLC family iron transporters in this study, we anticipate that…..” (page 22) Comment 2.2: I understand protocols were tested on different cell lines. Were lysates from primary cell cultures been also tested? Response: I have not tested primary culture cells. I understand this reviewer’s point that inclusion of primary culture cells would be helpful to more researchers having the transmembrane western problem. I expect that this study will be basically applicable to primary cultures. Readers will first test heating and lysis buffer conditions in their primary culture samples as demonstrated in this study, which hopefully could provide them with solutions or a better idea of further trouble shooting. Reviewer \#3 The manuscript addresses the question how different sample preparation has major affects on the outcome of Western blotting of several proteins related to iron metabolism. This is a worthy study as it can save many researchers a lot of pain, and improve study-outcomes. Comment 3.1: The authors relate in the abstract to DMT1 as an iron exporter. In most cases it though functions as an importer, when viewing the cytosol of a cell as “in” and the endosome as a part of internalized “out”. Later in the text DMT1 is termed a “transporter”, which does it more justice. Response: We totally agree with the important point. The abstract has been revised as follows: transmembrane iron transporter proteins; SLC11A2 (divalent metal transporter 1, DMT1), SLC40A1 (ferroportin 1, Fpn1), and transferrin receptor-1 (TfR1), Comment 3.2: It would be much clearer to use Fpn and DMT1 throughout the paper, and not label the figures in one way and use in the text the SLC terms, sometimes in full and sometimes shortened e.g. SLC40. Response: This is another important issue we realized after reading the manuscript carefully. In the revised manuscript, we primarily use Fpn1 and DMT1 rather than SLC40A1 and SLC11A2, respectively. All changes have been yellow- highlighted. Thank you for very helpful suggestions. Comment 3.3: In the text, a reference to fig. 2A is missing. Response: We have confirmed a reference to Fig. 2A in the text (page 13). Comment 3.4: In Figure 3 B, the evidence for a decrease of DMT1 with increased FAC is weak and the decrease with increased Cisplatin is not convincing at all. Can you please quantify these graphs and also indicate how often each experiment was repeated. Indicate the number of times the experiments were repeated throughout the manuscript, best in the legends. Response: We agree with the point of the small DMT1 expression changes in Fig. 3B (and also Fig. 6B). As 1) the trend of slightly decreased DMT1 expression in response to iron treatment is consistent in Fig. 3B and 6B, and 2) this manuscript primarily focuses on technical tips in transmembrane western blotting, we have just tempered the original statement in the text by adding “slightly” decreased DMT1 protein levels (Fig. 3B) at page 16. The experimental repeats have been indicated in each figure legend. Comment 3.5: There seems to be a discrepancy between Fpn results in fig. 6a and 4b. Any explanation? Response: Thank you for paying attention to this point. Although there might be a slight difference between these Fpn1 western blots in Caco-2 cells, the trend was similar; namely, heating samples did not compromise Fpn1 western blotting except for transfected (overexpressed) Fpn1. Comment 3.6: Figure legend 1 has a different format than the other legends, giving lot’s of info on blocking and washes. Could be omitted. Giving these details in materials and methods is enough. Response: We agree with simplifying the figure legend 1. As suggested, we have dropped redundant information including blocking and washing procedures. 10.1371/journal.pone.0235563.r003 Decision Letter 1 Missirlis Fanis Academic Editor 2020 Fanis Missirlis 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 Jun 2020 Transmembrane Protein Western Blotting: Impact of Sample Preparation on Detection of SLC11A2 (DMT1) and SLC40A1 (ferroportin) PONE-D-20-06977R1 Dear Dr. Tsuji, 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, Fanis Missirlis, Ph.D. Academic Editor PLOS ONE Additional Editor Comments (optional): Reviewers' comments: 10.1371/journal.pone.0235563.r004 Acceptance letter Missirlis Fanis Academic Editor 2020 Fanis Missirlis 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 Jun 2020 PONE-D-20-06977R1 Transmembrane Protein Western Blotting: Impact of Sample Preparation on Detection of SLC11A2 (DMT1) and SLC40A1 (ferroportin) Dear Dr. Tsuji: 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. Fanis Missirlis Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Primary lymphocytes are activated *via* a multi-step mechanism that begins with weak adhesion and stimulation of the T-cell receptor (TCR) by antigens exposed on the surface of antigen-presenting cells (APCs). This direct interaction induces a dynamic process that leads to the formation of specialized membrane junctions and adhesion strengthening. The contact site between cells provides a highly organized immunological synapse, a multi-protein signaling apparatus for controlling gene expression. All signaling events must be coordinated in time and space to achieve accurate T-cell activation, and each of these activities is dependent on the actin cytoskeleton. Actin drives the process of cell polarization, maintains cell–cell contact and provides a scaffold for clustering, translocation and spatial segregation of proteins, key steps to amplify and sustain T-cell signaling. TCR interactions with CD8 protein on APCs results in increased concentration of the membrane-associated tyrosine phosphatase, CD45, in the central part of the cell-cell contact area. Afterwards, CD45 down regulates the activity of proximal lymphocyte-specific tyrosine protein kinase (Lck), thus modulating early antigen-independent signals leading to actin cytoskeleton rearrangements. Proteins that influence synapse structure, such as F-actin and CD45, are present in the cell-cell contact area not only during the early stage of IS formation but also during the multidimensional construction of a mature synapse. The polarization of actin towards cell contact area is accompanied by recruitment of talin which activates lymphocyte function-associated antigen 1 (LFA-1). LFA-1 participates in immune responses by forming a membrane-junction with Intercellular Adhesion Molecule (ICAM-1/2) when T-cells interact with APCs. Identification of new molecular interactions underlying regulation of the immune response may lead to finding novel strategies targeting the IS for therapy. Treatments targeting the synapse have helped to establish immunotherapy as a mainstream element in cancer treatment. It has been demonstrated that impaired actin polymerization results in CD4+ and CD8+ T cells from patients with chronic lymphocytic leukemia (CLL) exhibiting defective immunological synapse formation with APCs. T cell dependent immune responses protect the host from cancer, but also participate in destructive autoimmunity. The spectrin-based network plays a critical role in actin cytoskeleton organization and remodeling. Spectrins are multifunctional proteins involved in the regulation of cell morphology and mechanical properties. Numerous spectrin isoforms are expressed in nonerythroid cells, where the composition of the spectrin network is highly complex and performs different functions. Many data suggest that spectrins directly or indirectly may control the distribution and activity of several proteins engaged in cell-cell contact and signaling, including actin, CD45 and LFA1. It has been proposed that spectrin, as a protein involved in actin skeleton organization, may play an important role in T-cell activation and IS formation. In this study, we focus on the role of spectrin in T-cell activation, using confocal microscopy analysis combined with advanced electron microscopy techniques. We describe here the effect of spectrin depletion in T-cells early stages of immunological synapse formation i.e. on the dynamics of actin during lamellipodia and filopodia formation and distribution of LFA-1 and CD45 during cell-cell contact organization. Obtained data may suggest a regulatory function of nonerythroid spectrin in the formation of the contact area between the T-cell and an APC. # Materials and methods ## Antibodies Rabbit anti- αII and anti- βII spectrin primary antibodies were readily available from the Laboratory of Cytobiochemistry, University of Wroclaw, Poland \[eg.\]. Rabbit anti-human αII-spectrin (H-105), rabbit anti-human β-actin (I-19), goat anti-human CD45 (N-19), goat anti-human LFA-1 (N-18) were purchased from Santa Cruz Biotechnology, USA. Anti-goat and anti-rabbit Cy2-, Cy5- (Jackson Immunoresearch Laboratories, INC) or Alexa Fluor 568–conjugated (A-11057, Thermo Fisher Scientific) IgG and anti-rabbit IgG conjugated with HRP (Abcam) were used. F-actin was labeled with Alexa Fluor 568-phalloidin (A12380, Thermo Fisher Scientific). Immunogold labeling was performed using 10 nm gold-conjugated anti- rabbit secondary antibodies produced in goat (Electron Microscopy Sciences, USA). ## Cell cultures The Jurkat T-cell line (German Collection of Microorganisms and Cell Cultures, ATCC^® Number TIB-152), HuT 78 (gift from A. Miazek—Institute of Immunology and Experimental Therapy, Polish Academy of Sciences) and primary peripheral blood mononuclear cells (PBMCs) were cultured at 37°C, in 5% CO₂ and in a humidified atmosphere in RPMI 1640 medium supplemented with 10% FBS, 2 mM glutamine and 100 μg/ml penicillin and streptomycin (Sigma Aldrich, USA). ## White blood cell isolation from peripheral blood Peripheral blood from healthy volunteers was diluted three times with PBS. The Ethics Committee of the Medical University of Wroclaw approved the study protocol (no. KB-541/2011). Informed consent was obtained from all of the subjects (only adults were involved). White blood cells were obtained by density-gradient centrifugation at 800 g on Ficoll-Paque (density = 1.078; GE Healthcare) for 40 min at room temperature. After centrifugation, the fraction of white blood cells was harvested, counted, and cultured at 37°C, in 5% CO₂ and in a humidified atmosphere in RPMI 1640 medium supplemented with 10% fetal bovine serum, 2 mM glutamine and 100 μg/ml penicillin and streptomycin (Sigma Aldrich, USA). ## Spectrin knockdown Two strategies of silencing of the *SPTN1* gene expression were used: 1. Jurkat cells were transfected with short hairpin RNA plasmids (pGeneClip™ hMGFP, SureSilencing™ shRNA, SA Biosciences) expressing GFP and a non-targeting sequence used as control (non-relevant shRNA, named Nr- shRNA) or a sequence targeting the αII-spectrin gene, (αII-spectrin shRNA named Sp-shRNA) using the Cell Line Nuclofector Kits (Amaxa Biosystem) AMAXA, according to the manufacturer’s instructions. The transfection efficiency and cell viability were estimated by flow cytometry (FACSCalibur flow cytometer, BD Biosciences) at 24 h after transfection performed with either negative control shRNA labelled with pmax-GFP and after cell treatment with 5 μg/ml propidium iodide (PI). Three Sp-shRNA plasmids were tested 1, 2 and 3. Depletion efficiency of αII-spectrin expression was estimated by western blot 48 h after transfection. Other plasmid expressing the LifeAct peptide fused with fluoro-Ruby (Ruby-LifeAct, a red marker visualizing F-actin in living cells) were also used to transfection Jurkat cells. 2. Cells were transfected with lentiviral particles containing short hairpin RNA plasmids (shRNA, SA Biosciences) expressing a scrambled sequence, used as control (scrambled shRNA, named SC-shRNA), or a sequence targeting the αII-spectrin gene (αII-spectrin shRNA named Sp-shRNA), according to the manufacturer’s instructions (Santa Cruz Biotechnology, Inc. αII-spectrin shRNA (m) Lentiviral Particles: sc-36550-V). Depletion efficiency of αII-spectrin expression was estimated by western blot. The assessment of silencing effect for both strategies are presented in supplementary material. ## T-cell activation Activation of T-cells was performed according to two methods. In one approach, polystyrene magnetic beads coated with an optimized mixture of monoclonal antibodies directed against ε chain of CD3 complex and anti-CD28 expressed on the T-cell surface (Dynabeads R Human T-Activator CD3/CD28, Thermo Fisher Scientific) were used to simulate antigen presentation and to activate the T-cells. Magnetic beads washed in PBS containing 5% human serum were mixed with cells at a ratio of 1:1 and incubated in PBS buffer containing 5% human serum at 4°C for 30 min, then for 4 min at 37°C. Cells were cyto-spun onto coverslips (ø12 mm) by centrifugation for 10 min at 800 g. In the second method, coverslips (ø12 mm) were incubated with 0.01% poly-L- lysine for 24 h at room temperature (RT). At the indicated times, coverslips were washed and coated with anti-CD3 and anti-CD28 antibodies (Miltenyi Biotec) at a dilution of 1:500 by incubation for 1h at room temperature. After washing with PBS buffer, plates were used for seeding Jurkat T-cells (1.2x10⁵). ## Immunofluorescence and confocal microscopy Cells were fixed for 10 min in 4% paraformaldehyde (PFA), washed and blocked for 1 h in 1% FBS in PBS at room temperature. After this time, cells were permeabilized with 1% Triton X-100 and stained overnight at 4°C with anti- spectrin antibody, anti–LFA-1 antibody and/or anti–CD45 antibody. For visualization in the confocal microscope, the cells were exposed to the secondary antibodies mentioned above. The fluorescence staining of F-actin filaments in cell cultures with Alexa Fluor 568-phalloidin was performed for 40 min at RT. Samples were mounted and counterstained for nuclei with Roti-Mount FluorCare DAPI (1.5μg/ml, Carl Roth GmbH). Observations were performed using a Zeiss Laser Scanning Microscope 510 (LSM510; Carl Zeiss, Jena Germany). ZEN software and apochromat ×60 oil immersion objective lens (numerical aperture 1.4) were used for image acquisition. ## Live-cell analysis Jurkat T-cells were co-transfected with Ruby-Life Act and either SC-shRNA or Sp- shRNA. Observation of global cell dynamics, contact with Dynabeads coated with anti-CD3 and CD28 antibodies and lamellipodia formation were started at 48 h after transfection. Cells mixed with Dynabeads (1:1 number ratio) were plated in IQ4 slides (Biovalley) at a density of 1x 10⁵ cells/ml and incubated at 37°C in 5% CO₂. The different events (number and stability of lamellipodia) were followed in the Biostation system (Biostation IM, Nikon), every 5 min for 12 h. Ten fields were analyzed for each experimental condition. ## Static cell adhesion assays Adhesion assays were performed on culture dishes presenting an activated surface. Control and transfected cells growth in rich medium were plated in triplicate on 12-well plates (10⁶ cells per well) and incubated for 30 min at 37°C in 5% CO₂. After gentle washing in PBS buffer, adherent cells were counted using Image-ProPlus software. The results are expressed as the mean percentages of spectrin-depleted adherent cells versus control adherent cells. ## Scanning electron microscopy Cells were fixed by either 4% PFA or 2.5% glutaraldehyde in PBS during 30 min at 4°C. After fixation, the samples were washed three times for 10 min in distilled water. Cell dehydration was performed in 50% ethanol at RT for 10 min, then with increasing ethanol concentrations (60%, 70%, 80%, 90%, 98%). Samples were finally treated with 100% ethanol and then allowed to air-dry overnight. The coverslips were coated with gold using ScanCoat6 (Edwards) and observed via SEM using a SE1 detector, at 10 kV filament tension (SEM, Zeiss Evo LS 15). ## Immunogold labelling and transmission electron microscopy Cells were fixed in 3.2% paraformaldehyde and 0.5% glutaraldehyde for 10 min at RT and washed twice in Sorensen's buffer (pH = 7.3). Samples were then dehydrated in a graded series of ethanol, each one for 10 min, up to 70% of ethanol. Immunogold labeling was performed using LR-White resin (London Resin Company Ltd) mixed at 2:1 with 70% ethanol for 2x1h, followed by incubation in pure LR-White overnight at RT and in pure LR-White for 48 h at 50°C. Ultrathin sections (\~65 nm) obtained using an ultramicrotome were mounted on nickel grids. In order to inactivate aldehyde groups present after aldehyde fixation, grids were incubated in PBSG buffer (50 mM glycine in PBS, pH 7.4) for 20 min, then incubated with the blocking solution (Aurion blocking solution, EMS, USA) for 15 min at RT. After blocking, grids were washed twice with antibody incubation solution (0.2% Aurion blocking solution in PBS, pH = 7.4) and incubated with primary antibodies (rabbit anti-spectrin, 1:20 dilution) overnight at 4°C. After incubation, grids were immunolabeled with 10 nM gold- conjugated secondary antibodies (1:20 dilution) for 1h at RT. After washing in 0.2% Aurion blocking solution in PBS (6×10 min) and distilled water (2×5 min), the grids were post-stained in uranyl acetate and Reynolds reagent for 1 h. The samples were examined using a focused ion beam in an *AURIGA* workstation (FIB- SEM, AURIGA 60, Carl-Zeiss, Germany), 20kV of filament tension bright field. ## Statistical analysis Each experiment was performed at least in triplicate, and values represent means of at least three independent experiments. Significance was determined using the Student’s t-test, and a 95% confidence interval was set, a priori, as the desired level of statistical significance. # Results ## Spectrin redistributes to cell-cell contact sites during IS formation Stabilization of the contact area between an APC and T-cell requires actin polymerization at the site of the IS. Numerous actin-binding proteins such as PLC-γ1(Phosphoinositide phospholipase C), adaptor NCK (non-catalytic region of tyrosine kinase), ITK (interleukin-2-inducible T cell kinase), Arp2/3 complex (*A*ctin-*R*elated *P*roteins ARP2 and ARP3), WASp (Wiskott–Aldrich Syndrome protein) and Dyn2 (Dynamin-2) have been shown to redistribute to the IS. However, the behavior of spectrin has not been explored in this regard. Therefore, we assessed the distribution of spectrin in T-cells during cell-cell contact leading to IS formation. The activation of T-cell by APCs was simulated by using beads coated with antibodies directed against CD3 and CD28. In non- activated primary peripheral blood mononuclear cells (PBMCs), αII-spectrin is mainly observed in a diffuse distribution around the cell perimeter. Similarly, labeling with phalloidin shows a regular organization of the actin cytoskeleton at the plasma membrane. However, when primary lymphocytes within PBMC population are stimulated for 30 min with APC-simulating Dynabeads coated with anti-CD3 and CD28 antibodies, spectrin and actin are mainly detected at the cell-bead contact sites where they co-localize. Thus, activation of TCR causes the translocation of spectrin along with actin to the presumed IS. Identical results of αII-spectrin redistribution as for primary T cells were obtained in Jurkat T-cell line (data not shown), which supports the participation of this protein in IS organization. We used Jurkat T-cells as a model system in subsequent studies not only for convenience but also for several other reasons, the most important being the possibility of obtaining stable cell lines with partially silenced *SPTN1* gene expression. To more precisely assess the distribution of spectrin in naïve and stimulated Jurkat T-cells, we examined its localization at the ultrastructural level using immunogold labeling and transmission electron microscopy. As shown in, αII- spectrin immunolabelling (arrows point to the gold particles) in non-activated Jurkat T-cells localized mostly at the intracellular side of the plasma membrane. However, when cells were activated with anti-CD3 and CD28, immunolabeling is mainly observed at sites of cell-bead contact (arrows). These ultrastructural observations are analogous to the level of light-microscopical analyses. Both labelling methods reveal αII-spectrin accumulation at cell-bead contacts in both stimulated T cell populations used, i.e. in primary lymphocytes and Jurkat T cells. ## Knockdown of spectrin in T cells causes a defect in cell adhesion and lamellipodial networks A link between spectrin and adhesion complex formation has been established previously. Metral and co-workers reported that melanoma cells depleted of αII spectrin exhibit impaired adhesion and modified expression of integrins. In addition, spectrin has an important function in the adhesive properties of neurites and the surface expression of adhesion molecule LFA1 in human neuroblastoma cells. Presented above results revealed the presence of αII spectrin in Jurkat T-cells at sites of IS formation, suggesting possible role of spectrin in this process. One of the key events that occurs during IS formation is increased specific T- cell adhesion. To investigate the putative participation of spectrin in Jurkat T-cell adhesion, spectrin was depleted in these cells using two kinds of shRNA strategy (for details see section). In first, cells were transfected with control shRNA (SC-shRNA) or three different shRNAs targeting human αII-spectrin mRNAs. As shown by western blot analysis, 48 h after transfection, these different shRNAs efficiently silenced αII-spectrin expression by \~60%. The other strategy was obtaining stable spectrin depleted Jurkat cell line, cells were transduced with lentiviral vectors containing short hairpin RNAs expressing scrambled sequences (SC-shRNA), or αII-spectrin specific sequences (sp-shR,NA) and levels of αIIβII-spectrin were analyzed by Western blotting. As shown in cells transduced with αII-spectrin specific RNA show significantly lower levels of αII-spectrin than control cells, when normalized to cellular actin. Depletion of the αII-spectrin subunit leads to a decrease in the level of βII chain. Morphology and phenotype of these spectrin depleted cells obtained by two different strategies did not show noticeable differences, so even though most of experiments were performed on cells obtained *via* either strategy, most of the results are presented for stable spectrin-depleted Jurkat cell line unless indicated otherwise. To examine the effects of spectrin depletion on adhesion, control cells (wild type, WT), cells transduced with scrambled shRNA (SC) or with anti-spectrin shRNA (KD) were plated on polylysine-coated culture dishes and the number of cells that were adhered to the plate 30 minutes later was quantitated. As shown in only 40% of cells treated with anti-αII-spectrin shRNA adhered, indicating that loss of spectrin significantly decreased the adherence potential of Jurkat T-cells. The adherence of T-cells to the APCs is facilitated through the formation of lamellipodia. To assess the role of spectrin in the ability of T-cells to form lamellipodia and/or filopodia, we used scanning electron microscopy to analyze the actin lamellipodia at the leading-edge of control and spectrin-depleted T-cells bound to Dynabeads using scanning electron microscopy (SEM). Control Jurkat T-cells (WT) and cells treated with scrambled shRNA (SC) develop numerous lamellipodia and long finger-like projections called filopodia around the Dynabeads coated with anti-CD3 and anti-CD28 (arrowheads). In contrast, spectrin-depleted cells show much decreased number of lamellipodia and filopodia extensions. To confirm and quantitate this effect, lamellipodia formation was assessed by analyzing actin distribution in Jurkat T-cells plated on coverslips coated with antibodies against CD3 and CD28. Control Jurkat T-cells (WT) and cells transfected with scrambled shRNA (SC) were seeded on coverslips coated with CD3 and CD28 antibodies and cultured for 24 hours. Actin staining detected extensive cell spreading and lamellipodia extension, as shown by the thin threads containing actin which protruded from the stimulated cells (arrowheads). Importantly, depletion of spectrin from cells causes loss of lamellipodia, and induces alterations in cell morphology and shape. The cells appear much smaller, indicating that spectrin depletion interferes with cell spreading. This is in agreement with the decrease in cell-substrate adhesion of the spectrin-depleted cells, shown in. Similar studies were performed using cell line HuT 78 and comparable effect of spectrin depletion on actin redistribution during activation was observed. Namely, as shown in, activated spectrin-depleted HuT 78 cells did not form lamellipodia and actin did not accumulate in site of cell contact. Finally, the frequency of lamellipodia formation and their stability were analyzed in Jurkat T-cells by live imaging ( and Movies). Cells transfected (according method A) with scrambled shRNA (SC) or with anti- αII-spectrin shRNA were incubated with anti-CD3/CD28-coated Dynabeads and the number of lamellipodia and filopodia formed per 1 minute was quantified during the 12-hour observation period. As shown in, spectrin-depleted cells show more than a threefold decrease in formation of lamellipodia and filopodia compared to control cells. ## Depletion of spectrin in T-cells inhibits the IS formation T-cell activation depends on the activation of CD45 phosphatase and its accumulation in the cell-cell contact area, an event mediated by its interaction with the spectrin-ankyrin-actin network. Moreover, recent data have shown that adhesion molecules, such as LFA-1, initiate attachment of cells and facilitate the stabilization of the IS. According to recent reports, LFA-1 adhesion enhances actomyosin forces which, in turn, modulate actin assembly downstream of the TCR. Since CD45 and LFA-1 are known to relocate to the IS after T-cell—APC interaction and that we have shown that spectrin also redistributes to the IS, we further examined their possible co-recruitment. The distribution of actin, CD45 and LFA-1 in non-activated Jurkat T-cells revealed a regular pattern of plasma membrane localization (data not shown). In anti-CD3/CD28 activated cells, accumulation of CD45 and LFA-1 was observed at contact sites between the cells and the beads (, WT). Therefore we may conclude, that LFA-1 and CD45 as and αII-spectrin were present in cell-cell contact sites together with actin. Our results, showing the presence of spectrin at sites of IS in Jurkat T-cells, along with the defects in cell adhesion and lamellipodia formation that we observed in spectrin-depleted cells, raised the possibility that spectrin may participate in IS formation. Thus, we assessed IS formation by following the distribution of actin, LFA-1 and CD45 in control Jurkat T-cells (WT), cells treated with scrambled shRNA (SC) or with anti-αII-spectrin shRNA (KD) after challenging them with Dynabeads coated with anti-CD3 and anti-CD28. As shown in, cells transfected with scrambled shRNA (SC) show IS formation at the site of contact with the beads, as evidenced by the redistribution of actin (A), LFA-1 (B) and CD45 (C) into punctate structures tightly concentrated at the contact sites. In contrast, in spectrin-depleted cells, actin remains distributed all over the plasma membrane without any obvious concentration at the site of Dynabead contact. Similarly, LFA-1 is not recruited to contact sites and remains uniformly distributed all over the plasma membrane in spectrin-depleted cells. CD45 also shows a defect in redistribution and, although minimal clustering can be seen, the majority of CD45 remains diffusely distributed on the surface of spectrin-depleted cells. Thus, depletion of spectrin from Jurkat T-cells inhibits the formation of the IS. # Discussion Presented here data concerning spectrin involvement in the contact between T-cells and APCs, which is an initial stage of immunological synapse (IS) formation upon T-cell activation, imply a possibility of regulatory role of spectrin in the formation of the cell-cell contact. Cell-cell interactions of this type are a general mechanism to involve some proteins, which leads to the reorganization of the actin skeleton and the formation of specialized cell membrane domains associated with functions that are vital to the body's inflammatory and immune response. Spectrin, a protein of the membrane skeleton family remains in the dynamic equilibrium, both between the cytosol and the membrane and within the membranes (e.g. in the lateral plane of plasmalemma) in connection with the activity of cells, as observed, e.g. during platelet activation or during the formation of intercellular contacts between epithelial cells. Our data may indicate an involvement of spectrin in the process of building intercellular contacts, such as IS. Actin polymerization in T-cells is critical for several steps of the immune response, including TCR clustering and mature synapse formation. Here, using immune labelling of T-cells, fluorescent and electron microscopy, we observed that upon activation of both the primary T lymphocytes isolated from peripheral blood, in HuT 78 and Jurkat T-cell lines, spectrin and actin rapidly accumulate in the cell-bead contact area, accompanied by their loss from the cytoplasm what may be a sign of spectrin role in this process. Published data indicate that spectrin can directly or indirectly interact with proteins involved in cell adhesion. A defect in spectrin in fibroblast cells displayed impaired growth and spreading, a spiky morphology and sparse lamellipodia associated with actin cytoskeleton modifications, such as loss of stress fibres, and modified expression of several integrins. It seems that reduced levels of spectrin in T-cells may also affect cell adhesion as a first stage of IS formation. The results of static adhesion assays of Jurkat cells showed that loss of spectrin reduces the frequency of intercellular contact formation which manifested as significantly reduced number of adherent spectrin- depleted cells compared to control cells. Furthermore, data from confocal microscopy, scanning electron microscopy and live-cell observations support the supposition that the disruption of the spectrin skeleton organisation is associated with a loss of actin-rich lamellipodia in activated spectrin-depleted (KD) Jurkat T-cells. Spectrin was found to control the dynamics of actin skeleton indirectly by interaction of the αII subunit SH3 domain with two members of the Ena/VASP family: VASP and EVL, as well as with Tes, actin binding protein located at cell-cell contact. It also participates in the Rac activation for actin filament formation and spreading. The fact that T-cells isolated from Wiskott-Aldrich syndrome (WAS) patients show characteristic cytoskeletal defects and impaired function, highlights a direct interaction between αII spectrin and testin (Tes), a tumor suppressor localized along stress fibers and participating in cell adhesion. RNA interference knockdown of Tes leads to a loss of actin stress fibers. Moreover, Tes interacts with a variety of cytoskeleton proteins of focal adhesion engaged in IS (such as vinculin and talin). An equally alternative possibility is that lower level of the αII subunit results in a lower level of the spectrin tetramer, which is the functional form of spectrin and contains an actin-binding domain in its β-subunit. It was shown that αII-spectrin accumulates in specialized integrin clusters that initiate cell adhesion. An intermediate zone of focal adhesion complexes contains force-transduction proteins, such as talin and vinculin, and a stress- fiber interfacial zone that contains actin-associated proteins, such as VASP. We hypothesize that spectrin, through direct interactions with VASP, indirectly controls activation of talin and, in this way, may participate in the regulation of LFA-1 integrin clustering in the IS region. Binding of LFA-1 on T-cells to ICAM-1 on APCs has been shown to provide a second signal for T-cell activation. This finding may be supported by other data showing that spectrin interacts with many proteins engaged in LFA-1 activation. Furthermore, spectrin by interacting with ankyrin allows CD45 translocation to membrane microdomains during T-cell activation. Data obtained here revealed that, upon stimulation of T-cells, nonerythroid spectrin, CD45 and LFA-1 integrin localizes at the IS region of the plasma membrane indicating that spectrin may be involved in the early steps of IS formation and T-cell activation. Using shRNA approaches, we have observed that depletion of spectrin in T- cells results in a significant decrease of actin accumulation within the cell-cell contact area. Moreover, in spectrin-depleted T-cells, a pronounced decrease of CD45 and loss of LFA-1 accumulation in the IS area was also observed. The data presented here supports our hypothesis that spectrin may affect the ability of T-cells to form immunological synapses, possibly through interaction with CD45 phosphatase. In conclusion, our data may indicate a regulatory function of spectrin as a protein involved in first-stages of contact formation between T-cells and APCs. This seems to add another feature to this multifunctional protein of membrane skeleton which was not considered in the past # Supporting information We thank dr. Michał Majkowski for expert technical assistance on confocal microscopy observations. Special thanks go to Professor Elizabeth Sztul, University of Alabama at Birmingham, USA, for reading the manuscript. [^1]: The authors have declared that no competing interests exist.

# Introduction In forensic laboratories, autosomal short tandem repeats (STRs) are used in forensic parentage testing. At present, the commercial kits include 13 Combined DNA Index System (CODIS) STRs. The number of STRs in the kits determines the system efficiency. The combined power of exclusion and combined power of discrimination of kits should be above 0.9999 for duo and trio cases. For duo cases, the STR number was more than that of trio cases due to the lack of a biological mother or father. STR locus mutation reduces the combined paternity index (CPI) value of the case and may lead to inconclusive results. It is necessary to increase the number of STR loci to obtain an adequate CPI value for the cases. High CPI values could also result in false conclusions when the STR gene type of two unrelated people matched. X chromosome short tandem repeats (X-STR) and Y chromosome short tandem repeats (Y-STR) are inherited in a unique, non-Mendelian fashion. Y-STRs are passed down from father to son. A father’s X-STRs are passed down only to his daughter, and a mother’s X-STRs are inherited by her sons and daughters. X-STR and Y-STR analyses are important for supplementing autosomal STR kit results in forensic cases. In Chinese forensic laboratories, the CPI cutoff value of autosomal STRs for parentage testing is provided. In duo and trio cases, the alleged father or mother is confirmed when the CPI is ≥10000. The results are inconclusive when the CPI is 0.0001\<CPI\<10000, and the alleged father or mother is excluded when the CPI is\<0.0001. The Goldeneye 25A kit includes 23 autosomal STRs. It covers all the STRs of Goldeneye 20A, AmpFlSTR SinoFiler and AmpFlSTR Identifiler kits. In the present study, we evaluated the application of 4 autosomal STR kits and the sensitivity of the CPI cutoff value (CPI ≥10000) based on routine parentage testing of duos and trios in our laboratory. At the same time, 3 complex close relative cases were also analyzed, which required the addition of other autosomal STRs and X-STRs to confirm the conclusions. # Materials and methods ## Samples In this study, 1442 real trio parentage testing cases, 803 real duo cases (including 412 motherless and 391 fatherless cases) and 3 complex close relative kinship cases were analyzed to evaluate the use of the CPI cutoff value. All the cases were typed using the Goldeneye 25A Kit in our laboratory. Three complex relative kinship cases were also tested, adding up to more autosomal SRTs and X-STRs for real conclusions. For all the cases, the parents or guardians of the children signed the informed consent forms with our laboratory. At the same time, the study was approved by the Ethics Committee of Jinan Central Hospital affiliated with Shandong University, China. ## STR loci of each kit 1. Goldeneye 25A kit includes 23 STRs (D2S441, TPOX, D22S1045, D7S820, D1S1656, Penta E, D10S1248, D8S1179, D5S818, D19S433, D16S539, CSF1PO, Penta D, D3S1358, vWA, D2S1338, D18S51, D6S1043, D13S317, TH01, D12S391, D21S11, FGA); 2. Goldeneye 20A kit includes 19 STRs (TPOX, D7S820, Penta E, D8S1179, D5S818, D19S433, D16S539, CSF1PO, Penta D, D3S1358, vWA, D2S1338, D18S51, D6S1043, D13S317, TH01, D12S391, D21S11, FGA); 3. AmpFlSTR Identifiler kit includes 15 STRs (TPOX, D7S820, D8S1179, D5S818, D19S433, D16S539, CSF1PO, D3S1358, vWA, D2S1338, D18S51, D13S317, TH01, D21S11, FGA); 4. AmpFlSTR SinoFiler kit includes 15 STRs (D7S820, D8S1179, D5S818, D19S433, D16S539, CSF1PO, D3S1358, vWA, D2S1338, D18S51, D6S1043, D13S317, D12S391, D21S11, FGA); 5. Goldeneye 22NC kit includes 4 STRs found in the Goldeneye 25A Kit (D3S1358, D2S441, D1S1656, D10S1248) and 17 other autosomal STRs (D4S2366, D6S477, GATA198B05, D15S659, D8S1132, D3S3045, D14S608, D17S1290, D3S1744, D18S535, D13S325, D7S1517, D10S1435, D11S2368, D19S253, D7S3048, D5S2500); 6. Goldeneye 17X kit includes 16 X-STRs (DXS6795, DXS9902, DXS8378, HPRTB, GATA165B12, DXS7132, DXS7424, DXS6807, DXS6803, GATA172D05, DXS6800, DXS10134, GATA31E08, DXS10159, DXS6789, DXS6810). ## STR typing DNA was extracted using Chelex-100. Goldeneye 25A, Goldeneye 22NC and Goldeneye 17X kits were used for PCR according to the manufacturer’s instructions. Autosomal STR and X-STR were typed by the ABI PRISM 3500 Genetic Analyzer. The data were analyzed using GeneMapper ID-X 1.3 software. ## Statistical analysis The CPI values of duo and trio cases for four kits were calculated according to the Specifications of Parentage Testing in China; gene frequencies of 40 autosomal STRs from the Shandong Han population in China were used for CPI values of all cases. The CPI values of different groups (including trio, duo cases, and different kits) were compared using the Mann-Whitney U test; the sensitivity of the CPI cutoff value was compared using the chi-square test. The results were considered significant if P\<0.05. # Results ## Application of CPI cutoff value for parentage testing of trios As shown in, the sensitivity of the CPI cutoff value (CPI≥10000) was high for trio cases. The CPI values of all the trio cases were higher than 10000 typed by the four autosomal kits, and every case had a confirmed conclusion. CPI values from high to low were the Goldeneye 25A kit\> Goldeneye 20A kit\>AmpFlSTR SinoFiler kit\>AmpFlSTR Identifiler kit. There were significant differences between the four kits (*P*\<0.05). ## Application of CPI cutoff value for parentage testing of duos As shown in, for every trio case, we regarded it as two real duo cases: a motherless case and a fatherless case. We found that the CPI values of all duo cases were higher than 10000 when typed by the Goldeneye 25A kit. The CPI value of a portion of the duo cases was 0.0001\<CPI\<10000 typed by the other three kits, so the cases had inconclusive results. The rate of inconclusive results from low to high was in the order Goldeneye 20A kit\<AmpFlSTR SinoFiler kit\<AmpFlSTR Identifiler kit. The CPI values and rate of inconclusive results were significantly different between the three kits (*P*\<0.05). For the 412 motherless and 391 fatherless cases, the results were consistent with those of the duo cases separated from the trio cases. The CPI value and sensitivity of the CPI cutoff value were not significantly different between the motherless and fatherless cases (*P*\<0.05). ## Three complex close relative kinship trio cases As shown in, in the first trio case, the child was a boy and the alleged father was the child’s uncle. Typed by the Goldeneye 25A kit, only 2 STR loci had no alleles from his uncle, with a CPI = 3.6787x10². After adding up to 40 STR loci using the Goldeneye 22NC kit, 5 STR loci had no alleles from his uncle, with a CPI = 0.8825. Without the mother’s information, as a duo case, only 3 STRs of the child had no alleles from his uncle, with a CPI = 3.5647x10³. We could not exclude the possibility that his uncle was the biological father. The Y-STR kit was not used for the child and the child’s uncle, because the results of the Y-STR gene type were same for them. To obtain reliable conclusions, we added the STRs gene type of his biological father for comparison. The alleles at 40 STRs of the biological father matched that of the child (data not shown). The alleged fathers were excluded based on the results. In the second case, the child was a girl. The alleged mother was her aunt, who was typed using the Goldeneye 25A kit. As a trio case, only two STR loci had no alleles from her aunt, with a CPI = 5.2048 x10⁶. Adding up to 40 STRs, 5 STRs had no alleles from her aunt, with a CPI = 2.3643 x10⁴. As a duo case, all 40 STRs of child had alleles from her aunt, with a CPI = 1.1392 x10¹³. We added the X-STR kit test, and only one X-STR (DXS10134) had no alleles from her aunt. Adding a sample from her biological mother for comparison, the child’s alleles at 40 autosomal STRs and 16 X-STRs came from her biological mother (data not shown). The alleged mother was excluded based on the results. In the third case, we did not exclude the possibility that his aunt was his biological mother based on 40 autosomal STR loci. The client could not provide a sample of the biological mother. The child was a boy, so we added the X-STRs test. We found that the alleles at three X-STR loci of the child did not come from his aunt. Therefore, we ruled out the possibility that his aunt was his biological mother (Tables). # Discussion Autosomal STR analysis is the primary method for parentage testing in forensic laboratories. The STR gene type of a child is based on his father and mother. In a real trio case, the paternity index (PI) of every STR locus can be calculated according to the STR gene type of the child’s mother and alleged father. In a real duo case, the PI of every STR locus can be calculated between two samples that share at least one allele. For one case, the CPI of all kit STR loci can be determined by multiplying the PI of every locus. The results of parentage testing are directly decided by CPI values. The conclusion of the probability of paternity can be determined from the CPI cutoff value set by the Specification of Parentage Testing in China. In the present study, we calculated the CPI values of 1442 real trio cases for 4 kits, including the Goldeneye 25A, Goldeneye 20A, AmpFlSTR SinoFiler and AmpFlSTR Identifiler kit. All the CPI values of the trios were above 10000. The alleged father was the biological father without an STR mutation, and the CPI value was significantly different among the four kits (*P*\<0.05). The four kits were sufficient to draw conclusions in the trio cases. In all the duo cases, including motherless and fatherless cases, typed by the Goldeneye 25A kit, all the CPI values of the duos were higher than 10000, but some of the CPI values were 0.0001\<CPI\<10000 when tested by the other three kits. The system efficiency of the four kits from high to low were the Goldeneye 25A kit\>Goldeneye 20A kit\>AmpFlSTR SinoFiler kit\>AmpFlSTR Identifiler kit. The AmpFlSTR SinoFiler Kit is based on the AmpFlSTR Identifiler kit. In the AmpFlSTR SinoFiler kit, D6S1043 and D12S39 with a high power of exclusion substitute for the THO1 and TPOX loci, so the system efficiency of the AmpFlSTR SinoFiler kit is higher than the AmpFlSTR Identifiler kit. This result was consistent with another study. In the present study, the results showed that the four kits could meet the requirements of routine parentage testing of trios. The Goldeneye 25A kit was sufficient to draw conclusions in the duo cases, but the other three kits were insufficient for a portion of the duo cases. In order to resolve the duo cases with a low CPI value, we should add more autosomal STR loci to increase the CPI value. For complex close relative cases, the alleged father usually was the brother of the biological father; the alleged mother was the mother’s sister. In the first case, the alleged father was the child’s uncle, and the CPI of the trio was 0.0001\<CPI\<10000 using 23 or 40 STR loci. Y-STR analysis was useless, because they come from the same paternal line. To obtain a certain conclusion, we had to add 40 autosomal STR gene types of his biological father for comparison (data not shown). In the second case, the alleged mother was her aunt. The CPI value of the trio was higher than 10000, with 5 STR loci having no alleles from her aunt. Without a biological father, as an alleged mother and child case, the child’s alleles at 40 STRs came from the alleged mother, and the CPI value was 1.1392 X10¹³. Alleles at only one X-STR locus did not come from the alleged mother (Tables). Considering that alleles at 40 autosomal STRs and 16 X-STRs came from her biological mother (data not shown), we excluded the possibility that her aunt was his biological mother. In the third case, we did not exclude that his aunt was his mother when using 23 or 40 autosomal STRs, but we excluded his aunt using the X-STR gene type (Tables). In the second and third cases, two STR mutations were found after testing by the Goldeneye 25A kit, and the CPI\>10000. If we did not know the expected results of the identification, it would result in a false conclusion that the alleged mother was the child’s biological mother with loci mutation. Therefore, evaluating simple CPI values of duo or trio cases may lead to false conclusions. Mutations of the STR loci are relatively common in forensic cases. They reduce the CPI value of the case, and affect the conclusion of parentage testing with a low CPI value. More autosomal STRs should be added to confirm the mutation. One or two STR loci mismatches may be due to mutational events. Sometimes, we should consider the mutational events as error events. Considering the possibility that autosomal STR loci mutations may occur, it is necessary to increase the number of the required STR loci and supplement the samples of the triplet. In this way, the identification errors could be greatly decreased. In complex close relative cases, there is great similarity of STR gene type among the close relatives, because the alleged father or mother is in the child’s immediate family. In some cases, all the limited STR loci gene types of the child matched that of the alleged father or mother. In fact, they were not the child’s biological father or mother. Sometimes, high CPI values can lead to false conclusions. In complex close relative cases, the genetic background of the case should be considered, and more autosomal STRs, X-STRs and Y-STRs should be used for further confirmation. # Conclusion In summary, the four autosomal kits were adequate to draw conclusions in the trio cases, and the Goldeneye 25A kit was adequate to draw conclusions in the duo cases. The other three autosomal kits were not sufficient for satisfactory conclusions for all duo cases. For complex close relative kinship cases, more autosomal STR loci and other genetic markers are necessary. The CPI cutoff value (CPI≥10000) is satisfactory for all trio cases and most duo cases. In some complex close relative cases, high CPI values may result in false conclusions. # Supporting information We thank Professor Cheng-tao Li (Academy of Forensic Science, Ministry of Justice, Shanghai, China) for data analysis advice that was instrumental for this study. 10.1371/journal.pone.0225174.r001 Decision Letter 0 Rogobete Alexandru Academic Editor 2019 Alexandru Rogobete 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. 9 Oct 2019 PONE-D-19-24597 The Title: Application of CPI Cutoff Value Based on Parentage Testing of Duos and Trios Typed by Four Autosomal Kits PLOS ONE Dear Dr Zhang, 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 Nov 23 2019 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, Alexandru Rogobete, Ph.D., M.Sc., Clin Res 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 <h ttp://www.journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_mai n_body.pdf> and <http://www.journals.plos.org/plosone/s/file?id=ba62/PLOSOne_for matting_sample_title_authors_affiliations.pdf> 2.  Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. 3\. In your data availability statement you write, "All relevant data are within the paper and its Supporting Information files." Please ensure you have provided the individual data points used to create the figures and determine means, medians and variance measures presented in the results, tables and figures (<http://journals.plos.org/plosone/s/data-availability#loc-faqs-for-data- policy>). If these data cannot be publicly deposited or included in the supporting information, e.g. due to patient privacy or ownership by a third party, explain why and explain how researchers may access them. 4\. Please amend either the title on the online submission form (via Edit Submission) or the title in the manuscript so that they are identical. 5\. 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.' Please provide an amended Funding Statement that declares \*all\* the funding or sources of support received during this specific study (whether external or internal to your organization) as detailed online in our guide for authors at <http://journals.plos.org/plosone/s/submit-now>. Please state what role the funders took in the study.  If any authors received a salary from any of your funders, please state which authors and which funder. If the funders had no role, please state: "The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." Please include your amended statements within your cover letter; we will change the online submission form on your behalf. 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\. 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: I Don't Know 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: 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: The manuscript entitled "Application of CPI Cutoff Value Based on Parentage Testing of Duos and Trios Typed by Four Autosomal Kits" by Hong-mei Gao provides quality results. I also described some parts that could be improved where the sentences and paragraphs are even more difficult to follow. In the paragraph "STR loci mutations affect the conclusions of parentage testing with a low CPI value, and more autosomal STRs should be added to confirm the mutation" can you explain it's not clear written. When you use abbreviations, for the first time in your manuscript, you have to give details about that. I recommend that a native speaker of English review the manuscript to improve word choice, sentence structure, and grammar. The conclusion need to clear and specific, with 3 short conclusion. Thanks for the opportunity to read the manuscript. Reviewer \#2: The language and structure of the article are so unclear that the merit can't be assessed. Most of the sentences are overly long, complicated, as well as ungrammatical in some instances. Throughout the article the language is neither clear, nor concise. Typographical and spelling errors have also been encountered. The article gives relations to three cases that are not included in the study groups described in the methods. Therefore part of the article rather looks as a case report (in this case sample sizes are not large enough to produce robust results). Regarding the tables of this manuscript, they are not all placed directly after the paragrapgh where they were first cited (see Table 2). The References formatting does not respect the style of the journal (does not list the first six authors followed by et.al.). Some of the title used as references are rather outdated (over 20 years old). Laboratory protocols were not made fully availabe. Detailed protocols could enhance the reproductibility of the results. In part the Discussions only repeat the results without interpreting them. \*\*\*\*\*\*\*\*\*\* 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 \[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files to be viewed.\] 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.0225174.r002 Author response to Decision Letter 0 25 Oct 2019 We are sorry for some flaws and incorrections appeared in this paper and appreciate the hard-work of your technical staff for revising. Here are the answers for the comments raised by the reviewers in the paper and we made changes in the paper accordingly. Reviewer \#1: 1\. In the paragraph "STR loci mutations affect the conclusions of parentage testing with a low CPI value, and more autosomal STRs should be added to confirm the mutation" can you explain it's not clear written. Answer: We explained it in discussion. “One or two STR loci mismatches may be due to mutational events. Sometimes, we should consider the mutational events as error events. Considering the possibility that autosomal STR loci mutations may occur, it is necessary to increase the number of the required STR loci and supplement the samples of the triplet. In this way, the identification errors could be greatly decreased.” 2\. When you use abbreviations, for the first time in your manuscript, you have to give details about that. Answer: we revised the abbreviations in the paper, including X-STRs and Y-STRs. 3\. I recommend that a native speaker of English review the manuscript to improve word choice, sentence structure, and grammar. Answer: The paper was revised by a native speaker of English (U.S.A). 4\. The conclusion need to clear and specific, with 3 short conclusion. Answer: We revised the conclusion according the suggestion of the reviewer. Reviewer \#2: 1\. The language and structure of the article are so unclear that the merit can't be assessed. Most of the sentences are overly long, complicated, as well as ungrammatical in some instances. Throughout the article the language is neither clear, nor concise. Typographical and spelling errors have also been encountered. Answer: The paper was revised by a native speaker of English (U.S.A). 2\. The article gives relations to three cases that are not included in the study groups described in the methods. Therefore part of the article rather looks as a case report (in this case sample sizes are not large enough to produce robust results). Answer: \(1\) Three complex close relative kinship cases were analyzed to evaluate application of CPI Cutoff Value. At the same time, in order to explain the problem that high CPI value (CPI≥10000) will lead to false conclusions in a part of cases. \(2\) In our present study, we only had 1442 real trio cases, 803 real duo cases and 3 complex close relative cases typed using Goldeneye 25A kit. We will study more cases in the future. 3\. Regarding the tables of this manuscript, they are not all placed directly after the paragrapgh where they were first cited (see Table 2). Answer: we revised it in the results and discussion. 4\. The references formatting does not respect the style of the journal (does not list the first six authors followed by et.al.). Some of the title used as references are rather outdated (over 20 years old). Answer: we revised the references formatting according to the style of the journal. One reference is over 20 years old (NO. 8), because DNA was extracted according to the method. 5\. Laboratory protocols were not made fully availabe. Detailed protocols could enhance the reproductibility of the results. Answer: We added the laboratory protocols in materials and methods (samples). 6\. In part the discussions only repeat the results without interpreting them. Answer: we amended the part of the discussion. 10.1371/journal.pone.0225174.r003 Decision Letter 1 Rogobete Alexandru Academic Editor 2019 Alexandru Rogobete 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. 31 Oct 2019 Application of CPI Cutoff Value Based on Parentage Testing of Duos and Trios Typed by Four Autosomal Kits PONE-D-19-24597R1 Dear Dr. Zhang, 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, Alexandru Rogobete, Ph.D., M.Sc., Clin Res Academic Editor PLOS ONE Additional Editor Comments (optional): Dear Authors, I read your paper carefully and it looks much better. Thank you BR, Reviewers' comments: 10.1371/journal.pone.0225174.r004 Acceptance letter Rogobete Alexandru Academic Editor 2019 Alexandru Rogobete 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 Nov 2019 PONE-D-19-24597R1 Application of CPI Cutoff Value Based on Parentage Testing of Duos and Trios Typed by Four Autosomal Kits Dear Dr. Zhang: 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. Alexandru Rogobete Academic Editor PLOS ONE [^1]: The authors have declared that no competing interests exist.

# Introduction Many herbivores have an ability to manipulate their host plant morphologically and physiologically for their own benefit. Gall induction by various insects is a typical example of such manipulations –. Many researchers have been interested in the mechanism underlying the host plant manipulation by gall-inducing insects. Although previous studies have been succeeded to induce morphological changes in plant tissues by the artificial application of extract from gall- inducing insects, fundamental mechanism underlying the gall induction by insects is still unknown. This is due to the difficulty in constructing laboratory bioassay systems for most gall-inducing insects. Some previous studies strongly suggested the involvement of chemical stimuli secreted from insects, because the site of gall induction is different from the feeding sites of gall-inducing insects in some plant-gall inducer systems. In addition, a previous study reported plant regulators stimulating cell division and inducing neoplasm formation on pods of *Pisum sativum* L. (Fabaceae), which is derived from pea weevil *Bruchus pisorum* L. and *Callosobruchus maculates* F. (Coleoptera: Bruchidae). However, unlike gall induction by insects, the neoplasm induction by the pea weevil has negative effect to the inducer , and the event may not be generalized for gall induction by insects. The maize orange leafhopper *Cicadulina bipunctata* (Melichar) (Hemiptera: Cicadellidae) induces growth stunting and leaf galls characterized by the severe vein swelling on various Poaceae. Though previous studies attributed the symptoms to a leafhopper-transmitted virus, recent studies strongly suggest that some chemicals injected by *C. bipunctata* during feeding are responsible for the gall induction, ,. Similar to other gall-inducing insects, the gall induction of *C. bipunctata* is adaptive for the inducer because free amino acids and glucose contents increased significantly in galled leaves and it results in the faster development and higher survival rate of offspring growing on the plant. The leafhopper usually feeds on mature host leaves and galls are induced not on feeding sites but on distant, newly developing leaves at approximately one week after the initiation of feeding. Because the degrees of growth stunting and leaf-vein swelling were significantly correlated with infestation density and length, the leafhopper is considered to induce the symptom by a dose-dependent reaction. Galls are induced more severe when plants were attacked at younger stages. Both nymphs and adults of *C. bipunctata* have the ability to induce galls on host plants. A recent examination using barley chromosome disomic addition lines of wheat revealed that the degrees of growth stunting and leaf- vein swelling were not significantly correlated, implying that the two symptoms are independent phenomena even though both are initiated by the feeding of *C. bipunctata*. This leafhopper is an ideal study material to clarify the mechanism of gall induction by insects because of the following reasons: Mass-rearing techniques have already been established for this leafhopper ; feeding site and gall- inducing site are different in this leafhopper, which enables us to separate the chemical examination at the site of feeding and gall induction ; model plants such as rice *Oryza sativa* L. and wheat *Triticum aestivum* L. are readily available as hosts in various experiments, ; and comparative studies using non- galling leafhoppers as well as resistant plant varieties can be performed. In this paper, we focus on the physiological response of plants at the site of gall induction to elucidate the mechanism of morphological manipulation of plant leaves by the leafhopper. Using both susceptible and resistant varieties of maize to gall induction by the leafhopper, phytohormone dynamics in leaves that will exhibit the symptom of gall are comprehensively analyzed and water contents of leaves are also examined. # Materials and Methods ## Insect Stock Culture and Maize Varieties A stock culture of *C. bipunctata* originally collected from Kyokushi, Kumamoto, Japan (32.57° N, 130.50°E) in September 2000 was used for experiments. The leafhopper was reared on rice seedlings at 25°C under an LD 16∶8 h photocycle using the method described by a previous study. Two maize (*Zea mays* L.) varieties ‘3081’ and ‘30D44’ were used in the experiments. The former is a variety susceptible to feeding by *C. bipunctata*, and the latter has a high resistance to feeding by the leafhopper. The growth stunting and gall-inducing profile by *C. bipunctata* on the susceptible variety 3081 was examined in previous studies, : when the seedlings at the second leaf stage were infested by five males of *C. bipunctata* for eight days, the plant growth was significantly stunted and veins of the third leaf were severely swollen. However, the resistant variety 30D44 displays only weak symptoms of growth stunting and leaf vein swelling, even if seedlings were fed on by *C. bipunctata*. ## Feeding Experiments Using the method described in a previous study, maize seeds were individually sowed in plastic cups (220 mL) with soil and kept in the phytotron at 25°C under an LD 16∶ 8 h photocycle. All maize seedlings were covered by acrylic cylinders (4.5 cm diameter and 24.5 cm deep) with a nylon cloth for ventilation on the top. Seven days after sowing, seedlings at the second leaf stage were randomly separated into the following four categories: (1) five adult males of *C. bipunctata* were released on a seedling from the 7th to 15th days ( = leafhopper treatment for eight days); (2) no adult males of *C. bipunctata* were released ( = control); (3) five adult males of *C. bipunctata* were released on a seedling from the 7th to 11th days (early treatment for four days); and (4) five adult males of *C. bipunctata* were released on a seedling from the 11th to 15th days (late treatment for four days). In the experiment, we did not use females of *C. bipunctata* to avoid possible influences of their ovipositions into seedlings. Adults can be sexed easily based on the presence or absence of a black ovipositor on the ventral surface of the abdomen. In the categories (1) and (2), plants were dissected 48, 96, 144 or 192 hours after the release of leafhopper adults and the third leaf, which is expected to exhibit galls, were weighted and immediately freezed in liquid nitrogen for comprehensive phytohormone analysis. Similarly in the categories (3) and (4), the third leaf samples were prepared 144 or 192 hours after the release of leafhoppers. Six replications were conducted for respective categories and times. ## Quantification of Phytohormones Extraction and purification of IAA, isopentenyladenine (iP), *trans*-zeatin (tZ), GA₁, GA₄, ABA, jasmonic acid (JA), jasmonoyl isoleucine (JA-Ile) and salicylic acid (SA) were performed by solid-phase extraction. Stable isotope-labeled compounds used as internal standards were: d2-IAA (Sigma-Aldrich); d6-iP, d5-tZ, d2-GA₁, d2-GA₄ (Olchemim Ltd, Olomouc, Czech Republic); D6- ABA (Icon Isotopes, Summit, NJ, USA); d6-SA (Sigma-Aldrich); and d2-JA (Tokyo Kasei, Tokyo, Japan). 13C6-JA-Ile was synthesized as described in a previous study. For simultaneous measurement of phytohormones, approximately 100 mg fresh weight of the third maize leaf was lyophilized, ground with 10-mm zirconia beads, and extracted 2 times with a total of 10 volumes of 80% (v/v) methanol containing 1% (v/v) acetic acid with internal standards at 4°C overnight. Extracts were centrifuged at 4°C, 14,000 *g*, 10 min, and the supernatant was collected. The supernatant was evaporated to water containing 1% acetic acid, and applied to a pre-equilibrated Oasis HLB column cartridge (30 mg, 1 ml, Waters, Milford, MA, USA). After washing with 1 ml of water containing 1% (v : v) acetic acid, hormones were eluted with 2 ml of 80% (v : v) acetonitrile containing 1% (v : v) acetic acid, and then acetonitrile were evaporated in vacuo (e. g., less than 400 µl at 2 ml of 80% acetonitrile containing 1% AcOH) to give extract in acidic water. Extract were applied to a pre-equilibrated Oasis MCX column cartridge (30 mg, 1 ml, Waters). After washing the cartridges with 1 ml of water containing 1% acetic acid, the acidic and neutral fraction containing IAA, GA₁, GA₄, ABA, JA and JA-Ile was eluted with 2 ml of 80% acetonitrile. Two hundred microliters of this fraction was transferred, dried, and reconstituted with water containing 1% acetic acid for analysis of SA. The MCX cartridges were further washed with 1 ml of 5% (v : v) aqueous ammonia, and the basic fraction containing iP and tZ was eluted with 2 ml of 60% (v : v) acetonitrile containing 5% (v : v) aqueous ammonia. After removing acetonitrile in vacuo, acidic and neutral fractions were further applied to a pre-equilibrated Oasis WAX column cartridge (30 mg, 1 ml, Waters). After washing with 1 ml of 1% acetic acid and 2 ml of 80% acetonitrile, the acidic fraction containing IAA, GA₁, GA₄, ABA, JA and JA-Ile was eluted with 2 ml of 80% acetonitrile containing 1% (v : v) acetic acid. This fraction was dried and reconstituted with 1% acetic acid. Hormones were analysed by LC-electrospray ionization (ESI)-MS / MS (Agilent 6410) on a ZORBAX Eclipse XDB-C18 column (Agilent) as described in a previous study, and quantified using MassHunter v. B. 01. 02 spectrometer software (Agilent, Santa Clara, CA, USA). ## Leaf Water Content Feeding experiments of the abovementioned categories (1) and (4) were performed for examining water contents in leaves. Plants were dissected 192 hours after the release of leafhoppers and the third leaf was sampled. The leaves were dried for 24 hours at 50°C after the measurement of fresh weight. The water contents (% fresh weight) of the third leaf were calculated from the fresh and dry weights of each leaf. Ten replications were conducted for each category. ## Statistical Analyses The concentration of each phytohormone was examined in each variety by analysis of variance (ANOVA) between treatments and control at 48, 96, 144, and 192 hours after the release of leafhoppers). Leaf water contents were analyzed in each variety by ANOVA after all percentage data were arcsine-transformed. In multiple comparisons, treatment means were compared by Tukey’s honestly significant difference (HSD) test. # Results In the susceptible variety 3081, IAA concentration did not clearly change by the leafhopper treatment. Although in leafhopper treatment the concentration was significantly lower than in control at 96 and 144 hours after the release, it slightly decreased after removing the leafhopper (144 and 192 hours in early treatment) and increased in control at 192 hours. The concentration of iP was not significantly different between control and treatments in the susceptible variety but that of tZ was significantly higher in leafhopper treatment than in control at 96 and 192 hours. After removal of leafhoppers, tZ concentration slightly decreased (early treatment at 144 hours) and exposure to leafhoppers increased the tZ concentration (late treatment at 192 hours). Concentrations of GA₁ and GA₄ were significantly lower in leafhopper treatment than in control (Fig. 2AC). In late treatment, GA concentrations immediately decreased after the exposure to leafhoppers (at 144 hours) (Fig. 2AC). In early treatment, GA₁ concentration significantly increased at 192 hours ( = 96 hours after the removal of leafhoppers). ABA concentration was most clearly affected by leafhopper treatments. In the treatment, ABA concentration was ten to 30 times higher than in control. The concentration relatively decreased after the removal of leafhoppers (at 144 and 192 hours of early treatment) and started to increase soon after exposure to leafhoppers (at 144 and 192 hours of late treatment.). Concentrations of JA, JA-Ile, and SA were not significantly different among control and treatments. In the resistant variety 30D44, leafhopper treatments did not affect the phytohormone concentrations on the third leaf (Figs 1BDF, 2BDF and 3BDF), except for iP concentration at 96 hours. Particularly, increase in the ABA concentration was not detected at all in leafhopper treatments of 30D44. In the susceptible variety, water content at 192 hours after the release of leafhopper was significantly lower in treatment than in control, but no significant difference was detected between them in the resistant variety. # Discussion Previous studies have reported that exogenous auxins can induce gall-like tissues on plants, and identified IAA in gall-inducing insects. High concentrations of IAA in gall-inducing larvae of Tephritidae (Diptera) and Gelechiidae (Lepidoptera) associated with *Solidago altissima* (Asteraceae) were reported in some recent studies –. Moreover, sawfly larvae (Hymenoptera: Tenthredinidae) inducing leaf galls on *Salix japonica* (Salicaceae) have the ability to synthesize IAA. In the present study, IAA concentration in the third leaves tended to reduce in leafhopper treatment at 96 and 144 hours, but no significant differences were detected at 48 and 192 hours. In addition, late treatment did not exhibit differences in IAA concentration. No clear differences in the IAA concentration in this study may imply that IAA is not related to the gall induction by *C. bipunctata*. However, the results do not necessarily preclude the possibility that IAA has an important role in gall induction by *C. bipunctata*. Biosynthesis of IAA in plant tissues transformed by *Agrobacterium tumefaciens* or with IAA biosynthetic genes did not elevate IAA concentrations despite the altered phenotypes in these plants. No significant differences in IAA concentration may be caused by the rapid metabolism of IAA that results from feed-forward regulation of IAA-catabolizing enzymes induced by elevated IAA concentrations following either its transgenic production or endogenuous application. Further studies will be needed to conclude the role of IAA in the gall induction by *C. bipunctata.* In cytokinins, iP concentration was not significantly different in the susceptible variety but tZ concentration was significantly higher in the treatment at 96 and 192 hours. This indicates that feeding by *C. bipunctata* increases tZ concentration in extending leaves of plants fed on by the leafhopper. Selective elevation of tZ concentration may cause abnormal swelling of leaf veins in extending leaves through possible changes in auxin:cytokinin ratio, critically affecting cell division patterns and tissue differentiation of plants. In the study of gall-inducing sawfly, larvae contained tZ in their bodies at the concentration more than 1,000 times higher than in host willow leaves and larvae were strongly suggested to synthesize the phytohormone by themselves. Whether the increased tZ in extending maize leaves is originated from leafhoppers or endogenous to maize is worth examining in future studies. GA₁ and GA₄ concentrations both significantly decreased by the feeding of leafhoppers in the susceptible variety. As mentioned earlier, the growth of the seedlings fed on by *C. bipunctata* are significantly stunted in the susceptible variety but seldom in the resistant variety,. These phenomena are probably related to the growth stunting of maize induced by the leafhopper, because GAs are well known to be related to elongation growth of plants. Notably, ABA concentration was remarkably accumulated in extending leaves of susceptible variety after feeding by the leafhopper, but this phenomenon was not observed in the resistant variety at all. A similar pattern of the increase of ABA in gall tissue was reported in a recent study using a gall-inducing psyllid, but the role of ABA in gall induction has not yet been clarified. Although ABA is known to increase following various environmental stresses such as drying stress, it is seldom known to cause morphological changes in plant tissues. In this study, we also detected significant decrease in water content only in extending leaves of the susceptible varieties after feeding by leafhoppers. Increase in ABA might be a response of plants to drying stress initiated by the feeding of leafhoppers and gall induction. A study in China reported that galls induced on maize by *C. bipunctata* were abundant in relatively dried maize fields. Further studies are needed to clarify the relationship between soil water condition, or humidity, and the degree of gall induction by *C. bipunctata*. JA and SA are well-known to be related to induced resistance against herbivores and pathogens. In this study, JA and SA concentrations did not change significantly by leafhopper treatments both in susceptible and resistant varieties. A previous study revealed that the nymphal performance of *C. bipunctata* was higher in the susceptible variety of maize than the resistant variety. Based on our analysis, induced resistance via JA or SA pathways seem not to be activated both in susceptible and resistant varieties after feeding by the leafhopper. Therefore, the induced resistance involved in these phytohormones seems not to be related to the lower performance of *C. bipunctata* on the resistant variety. Our study clearly indicated that tZ and ABA concentrations increase but GA₁ and GA₄ concentration decrease in leaves to be galled by *C. bipunctata.* Further studies analyzing the up- or down-regulation of phytohormone responsive genes in the leaves will elucidate manipulation mechanism of plant tissue by the gall-inducing leafhopper. We thank Ms. N. Komatsu for her support in phytophormone analysis, and Ms. K. Abe and Ms. R. Yamada for rearing the insects for experiments. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: MT YJ KM MM YK. Performed the experiments: MT YJ KM YT KS. Analyzed the data: MT YJ KS. Wrote the paper: MT YJ KM MM YK.

# Introduction Gram negative nosocomial pathogen *Pseudomonas aeruginosa* causes a variety of infections including spontaneous bacterial peritonitis pyogenic liver abscess, sepsis and septic shock. Endotoxin, which is a hydrophobic glycolipid, is known to play a very imperative role in pathogenesis of *P. aeruginosa* mediated infections. It is well recognized that cell free endotoxin is significantly more biologically functional than cell bound endotoxin and antibiotics, particularly those that act as inhibitors of cell wall biosynthesis, induce enormous amount of endotoxin release during treatment. Plenty of experimental evidences from *in vitro, in vivo* and *ex vivo* models have advocated that antibiotics vary in their ability to trigger endotoxin release from gram-negative microbes. Further, *ex vivo* evaluation of whole mouse blood has established that there is a correlation between amount of endotoxin release following antibiotic exposure and pro-inflammatory cytokine production. Though liver is known to detoxify endotoxin but at the same time it also responds energetically to endotoxin leading to endotoxin induced inflammations. In liver, LBP (endotoxin binding protein) binds to endotoxin and activates CD14, toll-like receptor (TLR) 4 and MD2 surface receptor complex of macrophages, monocytes, hepatocytes and kupffer cells resulting in potent inflammatory response. Endotoxin binds with TLR4 receptor which is highly expressed in cells that respond to endotoxin, such as macrophages, monocytes, hepatocytes and kupffer cells and induces expression of inflammatory genes through TLR4/NF-kB signaling pathway. NF-κB family consists of five structurally related proteins known as Rel/NF-κB proteins; p50, p52, RelA, RelB, and c-Rel. Two signaling pathways are involved in the activation of NF-κB family. Canonical pathway (classical) and non-canonical pathway (Alternative). Canonical signaling pathway includes toll-like receptor super family which is helpful in recruitment of adaptor molecules such as TRAF (TNF Receptor Associated Factor) to cytoplasmic domain of the receptor. The canonical pathway induction involves RelA, RelB, c-Rel and p50 proteins to activate NF-κB. In the non-canonical pathway, ligand induced activation of NF-κB is due to activation of NFkB-2, leading to liberation of p52/RelB. Both these pathways activate transcription of array of different genes. TLR4 may have a role in non-canonical NF-κB signaling since its ligand (endotoxin) induces P100 processing in a B-cell line. Further NF-κB regulates the production of pro-inflammatory mediators, such as TNF-α, COX-2 and iNOS and IL-12 which are mainly responsible for endotoxin induced tissue injury. Till now antibiotic therapy is the most viable therapeutic choice which causes rapid killing of pathogen and quick recovery of infection. But it also leads to antibiotic induced endotoxin release which then interacts with humoral and cell mediated immune system to stimulate release of an array of inflammatory molecules leading to severe inflammation, fever, tissue injury and organ dysfunction. Hence, there is an urgent requirement for antibiotic-anti- inflammatory co therapy, selecting those antibiotics that will not only kill the pathogen instantly but also suppress the detrimental effects of endotoxin mediated inflammation. Current anti-inflammatory chemotherapy fails because of a number of side effects on cardiovascular, gastrointestinal and circulatory system. Therefore, therapy with no side effects might provide a hope for the suppression of inflammation induced by antibiotic mediated endotoxemia. Herbal plant like *Zingiber officinale* is a natural dietary spice with potent anti-inflammatory, antioxidative and anticancer properties. Zingerone \[4-(4-hydroxy-3-methoxyphenyl) butan-2-one\] is a stable active component of dry ginger rhizome and has been found to down regulate age related activation of proinflammatory enzymes ; protect human lymphocytes from radiation induced genetic damage and apoptosis reduce endotoxin induced acute lung injury in mice. To the best of our knowledge not many studies are available on its *in vivo* protective effect against hepatic inflammation induced by antibiotic mediated endotoxemia. Keeping this in perspective, the aim of the present study was to assess the protective effect of zingerone on endotoxin induced liver damage in terms of liver histology, serum endotoxin levels and malondialdehyde (MDA), myeloperoxidase (MPO), nitrogen intermediates (RNI) and pro-inflammatory cytokine levels in liver homogenate. Effect of zingerone on endotoxin induced mRNA expression of inflammatory markers (TLR4, RelA, NF-κB2, TNF-α, iNOS and COX-2) was also evaluated in detail following *P.aeruginosa* induced peritonitis in mouse model of liver infection. # Materials and Methods ## Ethical Statement The experimental protocols were approved by the Institutional Animal Ethics Committee (Approval ID: IAEC/96) of Panjab University, Chandigarh, India and performed in accordance with the guidelines of Committee for the Purpose of Control and Supervision of Experiments on Animals (CPCSEA), Government of India. All efforts were made to minimize the suffering of animals. ## Bacterial strain Standard strain *Pseudomonas aeruginosa* PAO1 was obtained from Dr. Barbara H. Iglewski, Department of Microbiology and Immunology, University of Rochester, New York, USA and maintained in nutrient agar stabs at 4°C. ## Drugs and chemicals Pure zingerone \[4-(4-hydroxy-3-methoxyphenyl) butan-2-one\] was obtained from Gogia Chemical Industries, India. Antibiotics were purchased from Himedia chemicals, India. All other reagents and chemicals used were of analytical grade. ## Antibiotic susceptibility of PAO1 Antibiotic susceptibility of PAO1 against ciprofloxacin, amikacin, gentamicin and cefotaxime was tested by the standard broth dilution method according to the guidelines of the National Committee for Clinical Laboratory Standard.MIC values for all the antibiotics were calculated. ## Screening of antibiotics against PAO1 in terms of bacterial killing and endotoxin release *in vitro* PAO1 was incubated at 37°C for 1.5, 3, 4.5 and 6 h in the presence of antibiotics (2 X MIC). Culture without antibiotics served as control to evaluate bacterial killing and endotoxin release. ## Bacterial killing and endotoxin release To qantitate bacterial number, samples were taken at different time intervals and serially diluted in phosphate buffer saline and spread plated to MacConkey agar plates. Colonies were counted after overnight incubation at 37°C. The amount of cell free endotoxin in these samples was determined after removing bacteria by passing cell free supernatant through 0.22 µm Millipore filters. 0.1 ml sample was incubated with 0.1 ml Limulus amebocyte lysate (LAL) (GenScript USA) at 37°C. Absorbance was measured at 545 nm spectrophotometrically. The endotoxin levels were calculated against a standard curve of pure endotoxin of *E. coli* as per manufacturer's instructions. ## Protective effect of zingerone on antibiotic mediated endotoxemia against *Pseudomonas aeruginosa* peritonitis in a murine model BALB/c mice of either sex (8–10 week-old; 20–30 g) were procured from Central Animal House Panjab University Chandigarh. Animals were allowed free access to food and water at all times and were maintained in a controlled temperature (20–25°C) and humid (50±5%) environment. A total of 6 groups having 16 mice in each group were used in duplicate. Mice were infected intraperitoneally with 500 µl of *P.aeruginosa* cells (10⁵ cfu/ml) to establish *P.aeruginosa* induced peritonitis, experimental model of liver infection. On the peak day of infection (5^th day) mice were administered with single dose of cefotaxime and amikacin intramuscularly. Cefotaxime at a concentration of 100 mg/kg body weight and amikacin at 75 mg/kg body weight of mice administered to achieve high serum concentration necessary for a rapid bactericidal action. In antibiotic-zingerone groups, mice were administered single dose of zingerone (100 mg/kg body weight) immediately after antibiotic administration. Zingerone dose selected was 100 mg/kg approximately corresponds to 1/10th of LD₅₀. PAO1 infected mice receiving normal saline served as control. After 0, 1.5, 3, 4.5, 6 h of antibiotic exposure, mice were sacrificed, blood was collected by retro-orbital puncture in two aliquots and serum was separated and liver was removed aseptically. Liver tissue homogenate and serum samples were stored at −60°C till analysis was carried out. ## Histopathological examination Liver tissue samples fixed in 10% buffered normal saline and dehydrated in 30–100% gradient ethanol. Paraffin wax blocks were prepared and 5 µ thin sections were stained with hematoxylin eosin and Masson's trichrome stain. Liver sections were examined for inflammatory response and liver fibrosis. ## Serum endotoxin levels LAL Endotoxin Assay Kit (GenScript USA Inc.) was used for detection of endotoxin levels in serum samples. Briefly, 0.1 ml serum was incubated with 0.1 ml Limulus amebocyte lysate (LAL) at 37°C. Absorbance was measured at 545 nm spectrophotometrically. ## Preparation of tissue homogenate Liver tissue was harvested, washed in ice cold physiological saline and homogenized in buffer using glass homogenizer to obtain 10% homogenate. The tissue homogenate was centrifuged at 12,000 X g for 10 minutes at 4°C and the supernatant was collected. ## Bacteriological examination To qantitate bacterial numbers, liver homogenate samples taken at different time intervals were serially diluted in phosphate buffer saline (PBS pH 7.2) and 0.1 ml from each dilution was spread plated on to MacConkey's agar plates. Colonies were counted after overnight incubation at 37°C. ## Biochemical analysis of liver homogenates for the production of inflammatory mediators ### Malondialdehyde (MDA) estimation Induction of pathology was evaluated on the basis of Malondialdehyde, the index of lipid per oxidation following the method of Anjaneyulu and Chopra.. Briefly, tissue homogenate was added to tris HCl followed by the addition of ice-cold trichloroacetic acid. Supernatant was taken and mixed with thiobarbituric acid. Tubes were covered and kept in a boiling water bath for 10 min. After cooling, absorbance was read at 532 nm. The level of lipid peroxide was expressed as nmoles of MDA formed/mg of protein. ## Reactive nitrogen intermediates (RNI) estimation Nitrite was estimated in the liver tissue of mice following the method of Rockett et al.. Briefly, samples were mixed with Griess reagent (Sigma Aldrich Chemicals Ltd., St Louis, MO, USA) followed by addition of trichloroacetic acid and incubated at room temperature. After centrifugation, the optical density of supernatant was read at 540 nm. ## Myeloperoxidase (MPO) estimation MPO activity was quantified by using the myeloperoxidase assay as described by Hang el al.. Briefly, tissue was homogenized in potassium phosphate with hexadecyl trimethyl ammonium bromide and EDTA. The homogenate was sonicated and centrifuged. Supernatant was mixed with o-dianisidine and absorbance was read at 490 nm at 0 min, 1 min 2 min at room temperature to determine change in absorbance per minute. It was calculated by using the formula: MPO activity (U/mg)  =  X/weight of the piece of tissue taken, where X = 10 X change in absorbance per min/volume of supernatant taken in the final concentration. ## Estimation of TNF-α, MIP-2 and IL-6 cytokines by ELISA Levels of pro-inflammatory cytokines (TNF-α, MIP-2 and IL-6) in liver homogenate were assessed by using ELISA kits (Peprotech USA) according to the manufacturer's instructions. Ninety-six-well microtiter plates (Falcon Corp., USA) were coated with 100 µl of a suitable capture antibody per well. The plates were coated with 100 µl of sample and incubated at room temperature for two hours. The plates were washed with wash buffer and incubated with streptavidin antibodies followed by incubation with biotinylated antibodies. Plates were incubated in the dark with TMB substrate after washing. Once adequate color developed ELISA plates were read at 405 nm using Microplate Manager 5.1 (BioRad Labs Ltd. USA). Cytokine levels were estimated by using the standard recombinant cytokine supplied along with the kits as a reference. ## Serum AST, ALT and ALP estimation Aspartate aminotransferase (AST), Alanine aminotransferase (ALT), and alkaline phosphatase (ALP) enzyme activities in serum were determined using ERBA test kits (ERBA Diagnostics, Mannheim, Germany) at 6 h interval in different groups. ## Therapeutic potential of zingerone on endotoxin induced hepatic inflammation in terms of mRNA expression of inflammatory markers (TLR4/RelA/NF-kB2/TNF- α/iNOS/COX-2) *in vivo* Group having 24 mice each (BALB/c 3–5 weeks old and weighing 20–30 gm) was put in duplicate. Purified endotoxin of *P.aeruginosa* PAO1 was administrated intraperitoneally (1 mg/kg body weight) and 6 mice each were sacrificed at 4, 8, 12 and 24 h. mRNA expression of the genes was evaluated in liver tissue using reverse transcription–polymerase chain reaction. To evaluate the therapeutic potential of zingerone in terms of production of mRNA of inflammatory genes, three groups of 6 mice each (BALB/c 3–5 weeks old and weighing 20–30 gm) in duplicate were used and were sacrificed at 8 h, as maximum mRNA expression was found at 8 h after LPS administration. In 1^st group endotoxin was administrated intraperitoneally (1 mg/kg body weight) and in 2^nd group the mice were administered one dose of zingerone (100 mg/ml) immediately after endotoxin treatment. Mice receiving normal saline served as controls. Level of mRNA expression of the genes was evaluated using reverse transcription–polymerase chain reaction. ## Reverse transcription–polymerase chain reaction (RT–PCR) Nucleotide sequence for genes was taken from NCBI data base. For each gene primers were designed using Primer 3 online tool. Primer sequences used for PCR amplification of c DNA are mentioned in Table. 1. Liver tissue was homogenized with Trizol Reagent (Invitrogen USA). The homogenate was centrifuged at 3000 X g at 48°C for 10 min. The supernatant was mixed with chloroform and precipitated with 75% ethanol. The total amount of RNA was determined using the spectrophotometric analyzer, Nano Drop 100 (Thermo scientific). RNA was reverse- transcribed into cDNA using a First-Strand cDNA Synthesis kit (Fermetas USA) with oligo (dT) primer. The cDNA was amplified with specific primers for TLR-4, RelA, NFκB2, TNF-α, iNOS, COX-2 and GAPDH as a control. Sample was incubated using a MJ Minin PCR Thermal Cycler (Biorad USA) and ran 34 times. The cycles lasted for 30 S at 94°C, for 60 S at 58°C, and for 60 s at 68°C for RelA, NF-κB1 and cycles lasted for 30 S at 94°C, for 60 S at 56°C, and for 60 s at 68°C for Cox-2, iNOS,TLR4,TNF-α. For control gene GAPDH the cycles lasted for 30 S at 94°C, for 60 S at 55°C. The final incubation was at 72°C for 5 min. Amplified PCR products were separated electrophoretically on a 1.0% agarose gel, and bands were visualized with ethidium bromide under ultraviolet transillumination. Densitometry of PCR product to determine relative mRNA expression was performed by Gel Doc Multi-Analyst (BioRad USA). ## Statistical analysis All experiments were performed in duplicate and repeated on different days. The effect of zingerone treatment on antibiotic induced endotoxemia and relative mRNA expression of genes in different treated groups with control was evaluated using two-way ANOVA test. p values were calculated and p\<0.05 was considered significant. Data was analyzed using Graph Prism 5.0 software. Values were expressed as mean + S.E.M. # Results ## Antibiotic susceptibility of PAO1 MIC values for ciprofloxacin, amikacin, gentamicin and cefotaxime against PAO1 were determined and found to be 0.3, 3.0, 30.0 and 25.0 µg/ml respectively. ## Effect of antibiotics on PAO1 in terms of bacterial killing and endotoxin release *in vitro* All antibiotics (2X MIC) showed decrease in viable counts and significant reduction was found at 6 h hour (p\<0.001). Ciprofloxacin showed highest bactericidal action as compared to rest of the antibiotics ( –A). Varied amount of cell free endotoxin was released on exposure to different antibiotics. Cefotaxime and amikacin were found to be efficient endotoxin releasing antibiotics and both the antibiotics significantly released high amount of endotoxin (p\<0.001) ( –B). On the basis of results from *in vitro* endotoxin release assay, cefotaxime and amikacin were selected for *in vivo* endotoxin release studies. Effect of zingerone was also evaluated for endotoxin release potential of antibiotics *invitro*. No significant effect was found (supplementary data) on the endotoxin levels indicating that zingerone did not interfere with the endotoxin release potential of antibiotics. ## Protective effect of zingerone on hepatic inflammation induced by antibiotic mediated endotoxemia in PAO1 infected BALB/c mice ### Liver histology Histological analysis of liver tissue obtained from antibiotic treated infected groups showed increased infiltration of neutrophilic granulocytes, necrosis of hepatocyte and hepatic portal inflammation along with hepatic portal haemorrhage and liver tissue fibrosis (-C,I and Fig2-D,J) as compared to infection (PAO1) control (-B,H). Mice without any infection did not show any inflammatory response (-A, G). Cefotaxime-zingerone (-E, K) as well as amikacin-zingerone (-F, L) treatment showed very less neutrophil infiltration, no necrosis and portal haemorrhage in the liver tissue. The findings were comparable to normal as observed in control group. ### Bacteriological examination Mean decrease in bacterial count was achieved in the liver of mice following infection with *P.aeruginosa* along with antibiotic treatment at different time intervals. After amikacin therapy, a steady decrease in bacterial count was observed from 7.6 log cfu (3 h) to 4.3 log cfu (6 h) ( -A). Similar trend was observed with cefotaxime and the viable counts were 9.4 log cfu (3 h) and 5.8 log cfu (6 h) ( -C). Simultaneous administration of zingerone along with amikacin and cefotaxime did not show any further decrease in viable count of bacteria at all time intervals except at 6 h when significant difference was observed (p\<0.05). ### Serum Endotoxin Levels Significantly high serum endotoxin levels were observed in PAO1 + Antibiotic group. With cefotaxime and amikacin, significant endotoxin release occurred between 3 to 4.5 h of exposure, reaching a maximum of 2.7 EU/ml and 1.88 EU/ml (p\<0.001) for (-B) cefotaxime and amikacin (p\<0.001) respectively (-D). Zingerone treatment significantly reduced the endotoxin levels at 3, 4.5 and 6 h. In cefotaxime and amikacin treated groups endotoxin levels were significantly reduced to 1.22 EU/ml and 0.72 EU/ml (p\<0.01) respectively at 6 h. ## Production of inflammatory mediators ### Malondialdehyde (MDA) estimation Liver homogenate of infected animals showed moderate amount of MDA but treatment with amikacin significantly increased MDA content and maximum increase was found at 6 h (45.66±3.4 nmoles/mg) (p\<0.001) ( A). Simultaneous treatment of amikacin with zingerone resulted in decrease in MDA content and significant decrease was found at 6 h (27.1±2.1 nmoles/mg) (p\<0.001) ( A). Similarly, cefotaxime increased MDA content significantly at all time intervals (p\<0.001) ( D). Simultaneous treatment of cefotaxime with zingerone decreased MDA content significantly at 4.5 h (p\<0.01) and at 6 h (p\<0.001) ( D). ### Myeloperoxidase (MPO) estimation Treatment with amikacin increased MPO content initially but significant increase was found at 4.5 h and 6 h (p\<0.001) ( B). Zingerone treatment slightly decreased MPO at 3 and 4.5 h but significant decrease was found at 6 h (0.66±0.16 U/mg nmoles/mg) (p\<0.01) ( B). Similarly, cefotaxime significantly increased MPO content at all time intervals (p\<0.001) ( E). Zingerone treatment reduced MPO content and significant decrease was observed at 4.5 h and 6.0 h (p\<0.01) ( E). ### Reactive nitrogen intermediates (RNI) estimation Infected mice showed moderate amount of RNI but treatment with amikacin significantly increased RNI content with maximum increase seen at 6 h (p\<0.001) ( C). Following treatment with zingerone, slight decrease in RNI content was found at 3 and 4.5 h but significant decrease was found at 6 h (p\<0.01) ( C). Likewise, cefotaxime significantly increased RNI content at 3 h, 4.5 h and maximum increase was found at 6 h (26.59±5.11 nmoles/mg) (p\<0.001) ( F). With zingerone treatment RNI content decreased at 1.5, 3.0 and 4.5 h interval but significant reduction was found at 6 h (16.9±1.8 nmoles/mg) (p\<0.01) ( F). ### Estimation of TNF-α, MIP-2 and IL-6 cytokines by ELISA Amikacin and cefotaxime treatment led to decrease in bacterial load but significant increase in TNF-α, MIP-2 and IL-6 proinflammatory cytokines production was observed. After amikacin therapy levels of TNF-α, MIP-2 and IL-6 were significantly increased at 3 h, 4.5 h and with maximum increase observed at 6 h (-D). Cefotaxime was found to be more effective in inducing production of proinflammatory cytokines. Significant increase of all the three cytokines was observed at 3 h, 4.5 h and 6 h (p\<0.001) (-A). Zingerone treated group showed decrease in the levels of proinflammatory cytokine at 1.5, 3, 4 h but significant difference was found only at 6 h. In amikacin + zingerone group, TNF-α levels were significantly decreased at 6 h (85 pg/mg) (p\<0.01) (-D). Zingerone treatment also decreased MIP-2 and IL-6 cytokine levels at 6 h (90 pg/mg) (p\<0.05) and (110 pg/mg) (p\<0.001) respectively (-E, F). Zingerone was also able to suppress cytokines production after cefotaxime exposure at 6 h. The levels of TNF- α, MIP-2 and IL-6 were found to be 105 pg/mg (p\<0.05), 135 pg/mg (p\<0.01) and 130 pg/mg (p\<0.01) respectively (-A,B,C). ### Serum AST, ALT and ALP levels Control group without infection showed normal AST, ALT and ALP levels in serum. Infection group showed elevated levels of these markers. Antibiotic treated groups showed comparatively high level of the tissue damage markers. Cefotaxime treatment showed highest level of these enzymes. Interestingly zingerone as co- therapy significantly reduced AST, ALT and ALP levels indicating protective effect of zingerone against antibiotic induced liver damage. ## Endotoxin induced liver inflammation in terms of mRNA expression of TLR4, RelA, NF-kB2, TNF- α, iNOS, COX-2 genes *in vivo* ### Time dependent expression studies of gene expression in liver tissue against purified endotoxin Endotoxin administration caused potential increase in TLR4/NF-κB dependent expression of genes. TLR4 mRNA expression increase was time dependent. It started increasing at 4 h and was found to be maximum at 8 h (\>7 folds) after which its expression declined (-A). Relative RelA mRNA expression was slightly increased at 4 h and maximum at 8 h (\>3 folds) (-B). Similarly, both NF-κB2 and COX-2 genes were expressed highest at 8 h (\>3 folds) and declined later (-C, F). Relative mRNA expression of proinflammatory cytokine TNF-α increased significantly at 4 h and reached its maximum level at 8 h (\>15 folds) (-D). iNOS gene expression was highest at 4 h (\>8 folds) and remained active up to 8 h (\>5 folds) decreasing thereafter leading to minimum level at 24 h ( B) (-E). Results indicated maximum expression of most of the genes at 8 h interval in endotoxin treated group ( A and B). At 12 h, expression level of all the genes started to decline and at 24 h, minimum expression was observed (Fig6). ### Effect of zingerone treatment on gene expression Maximum expression of inflammatory markers was observed at 8 h after endotoxin administration, therefore protective effect of zingerone in term of gene expression was evaluated at 8 h only. Results showed that in endotoxin induced animals, zingerone treatment could reduce the mRNA expression of TLR4 by \>2 fold (-A). Similarly, mRNA expression of RelA and NF- κB2 was also found to be inhibited significantly (\>1.5 folds and \>5 folds respectively) (-B, C). Relative mRNA expression level for TNF- α in zingerone treated group was significantly reduced (\>2 folds) as compared to endotoxin treated animals (-D). Specific inflammatory enzymes iNOS and COX-2 were found to be inhibited significantly (\>3 folds and \>5 folds respectively) (-E, F) in zingerone treated animals. Results showed that post endotoxin treatment with zingerone significantly reduced (p≤0.05) mRNA expression of all these inflammatory markers in mice. # Discussion Correlation between endotoxin release and corresponding type/dose of antibiotic is well known and many *in vitro* and *in vivo* studies are available on this aspect. Antibiotics rapidly kill the pathogen and release enormous amount of endotoxin in blood stream. Different classes of antibiotics targeting cell wall, protein synthesis, pathway of DNA metabolism differ in their potential to release cell free endotoxin. In the present study, endotoxin releasing potential of ciprofloxacin, amikacin, gentamicin and cefotaxime was studied in *P.aeruginosa* PAO1. Endotoxin release with ciprofloxacin was least and maximum with cefotaxime on treating *P.aeruginosa* cells *in vitro.* Ciprofloxacin acts on the A subunit of DNA gyrase, which inhibits DNA supercoiling, resulting in the inhibition of DNA replication without causing cell lysis. Amikacin and gentamicin that inhibit protein synthesis are also known to release low amounts of endotoxin as compared to beta lactam antibiotics. Whereas, cefotaxime (7-\[2-(2-amino-4-thiazolyl)-2-methoximino\]-acetamido cephalosporanate) has high affinity for penicillin-binding proteins (PBPs) and induces formation of filamentous cells leading to cell lysis. High endotoxin release in gram negative bacteria (*E.coli)* has also been linked to significantly high endotoxin level in plasma and IL-6 pro-inflammatory cytokines in serum. Since, cefotaxime and amikacin were found to release high amounts of endotoxin as compared to gentamicin and ciprofloxacin hence these two antibiotics were selected for *in vivo* studies. Immunostimulatory mechanism of *P. aeruginosa* in liver inflammation induced by antibiotic mediated endotoxemia is still not very well understood. Liver is responsible for detoxification of endotoxin from blood stream and is most susceptible to endotoxin mediated inflammatory damage. During infection and even during antibiotic treatment, liver becomes the primary target organ for endotoxin stimulation. Endotoxin-TLR4 mediated signalling pathway enhances production of inflammatory mediators following *P.aeruginosa* infection. Endotoxin-induced liver injury has been used as an experimental model to analyze the mechanism of endotoxin-induced liver inflammation using *E.coli* endotoxin. In the present study both cefotaxime and amikacin induced significant endotoxin release *in vivo.* To study this phenomenon *P. aeruginosa* induced peritonitis mouse model of liver infection was established. Animal group on peak day of infection were treated with high dose of either cefotaxime or amikacin. Liver inflammatory response was significantly high after 6 h of antibiotic administration and this was linked to high endotoxin release by antibiotics. This indicated that the high inflammatory response was induced by endotoxin release due to immediate lysis of bacteria and remained till the endotoxin was cleared from the organs and circulatory system completely. After 6 h inflammation was significantly reduced and infection treated completely in antibiotic treated group (data not shown). Biochemical analysis of liver homogenate for inflammatory mediators indicated elevated levels of MDA, MPO and RNI. Lipid peroxidation is well known marker for tissue destruction which indicates oxidative degradation of lipids and also indicative of inflammatory injury and tissue damage. Elevated MDA levels observed in this study indicated that the product of immediate lysis of bacteria caused stimulation of liver cells and generation of free radical damage that led to oxidative damage to cell membranes. Histopathological changes observed in tissue sections relate to reactive nitrogen intermediates (RNI) production, a potential source of free radical mediated inflammation or tissue damage. Since neutrophils are major effector cells in damaging the liver and an important source of free radicals, hence, enhanced MPO activity observed may have contributed to hepatocyte necrosis, proinflammatory cytokine production and hepatic inflammation. High myeloperoxidase activity is a marker of local and systemic inflammation, relating tissue destruction inflammatory response to bacterial antigens. Overzealous production of pro-inflammatory cytokines including TNF-α MIP-2 and IL-6 can result in shock, multi organ dysfunction, and even death. In the past, over expression of MIP-2 protein has been specifically linked with endotoxin mediated hepatic injury. Proinflammatory cytokines play a crucial role in endotoxin-induced liver injury leading to hepatotoxicity.TNF- α and IL-6 cytokine were found to be highly expressed in liver during inflammation as a result of endotoxemia. Following zingerone treatment proinflammatory cytokines also showed significantly low levels (p\<0.05). Anti-inflammatory activity of zingerone in this study, could be attributed to phenolic nature of zingerone which might have led to scavenging of free radicals. Methoxy group with phenolic hydroxyl group in zingerone facilitates proton release along with long chain ethyl methyl ketone group providing bulk stabilization to zingerone molecule. This may lead to cell penetration and scavenging of free radicals. Anti-inflammatory potential of zingerone treatment along with antibiotic therapy showed decrease in inflammatory response in terms of decreased neutrophilic granulocyte infiltration and no hepatic portal haemorrhage. Hepatic haemorrhage was also absent in zingerone treated liver tissue. Levels of Inflammatory mediators MDA, RNI and MPO in zingerone treated animals were also significantly reduced (p\<0.05). A significant body of evidence indicates that Injury by LPS particularly in liver involves LPS binding proteins (LBP) which activate the CD14/TLR4 receptor and in turn induce transduction of inflammatory signals resulting in the regulation of inflammatory mediator production. Inflammatory markers chosen for the study have been found to play significant role in LPS *in vivo* induced tissue injury through NF-κB. Time dependent expression of genes induced by LPS revealed that expression of some genes started early at a time interval of 4 h (iNOS, NF-κB2) and some at 8 h (TLR4,TNF-α, RelA, and COX-2). Level of expression was found to be variable but maximum expression was found at 8 h. In the present study, *P.aeruginosa* LPS significantly enhanced mRNA expression of TLR4 receptor leading to increase in the number of TLR4 receptors on the liver cell surface. Due to this, more binding of LPS to cells resulting in potent induction of inflammatory response was observed. Zingerone treatment significantly reduced the level of mRNA expression of TLR4 receptor indicating reduced number of TLR4 receptors and thereby less binding of LPS. This may have led to decreased inflammatory response after zingerone treatment. During gram- negative sepsis, LPS induced cells are triggered to produce large quantities of pro-inflammatory cytokines such as tumor necrosis factor alpha (TNF-a) in response to endotoxin. TNF-α is secreted by a variety of cells, including hepatocytes, kupffer cells mast cells and epidermal cells. However, mainly activating macrophages and natural killer cells, release potent biologically active substances which cause shock, fever, organ failure and other pathophysiological implications Workers have also found that TNF-α plays a crucial role in LPS-induced liver injury leading to hepatotoxicity. In the present study, LPS caused tremendous increase in TNF- α levels at 4 h and 8 h after LPS administration in liver tissue indicating that its production is mainly responsible for liver injury. Zingerone treated liver cells showed significantly low levels of TNF- α suggesting less hepatotoxicity and tissue inflammation. We also checked the mRNA expression levels for iNOS gene. Hyper expression of iNOS clearly indicated that oxidative damage to the liver is contributed by iNOS. iNOS expression is known to be enhanced by LPS leading to generation of nitric oxide radicals causing acute tissue injury. Zingerone treatment significantly suppressed the mRNA levels of iNOS gene suggesting its antioxidant activity. Another inflammatory enzyme COX-2 is also activated by LPS stimulus. Previous reports have shown a potential role of tyrosine kinase in LPS promoter region that contain 24 transcriptional factor- binding sites, including those for nuclear factor-κB (NFκB) family, that appears to be essential in the enhanced COX-2 gene expression seen in macrophages exposed to endotoxin. Cyclooxygenase-2 (COX-2) is an inducible enzyme of macrophages catalyzing the conversion of arachidonic acid to prostaglandins. Recent studies have suggested that increased levels of prostaglandins and cyclooxygenase activity and COX-2-derived bioactive lipids, including prostaglandin E2 (PGE2), are potent inflammatory mediators causing tissue injury. LPS induced very high mRNA expression of COX-2 (at 8 hour interval) and this probably may have led to increased production of prostaglandin E2 resulting in intense inflammation. Zingerone treatment significantly reduced mRNA expression of COX-2 which ultimately reduced the liver injury in treated animals. RelA, NF-κB2 are signaling molecules and regulate the expression of many inflammatory genes. Expression of these genes in the present study clearly indicated that these genes are involved in the signaling cascade and regulation of expression of inflammatory genes. Rel A and NF-κB2 gene expression was found to increase following LPS administration. Zingerone treatment significantly inhibited the expression level of these genes clearly indicating that zingerone was able to interfere with inter signaling pathways and suppress the hyper expression of important cell signaling molecules. Since, *P.aeruginosa* LPS showed maximum expression of all genes at 8 hour interval, this time period was chosen for observing the effect of zingerone on the expression of inflammatory markers. Expression of COX-2, TNF-α, iNOS, RelA, NF-κB2 and TLR4 was found to be highly suppressed by zingerone treatment at 8 h interval. Decrease in the mRNA expression levels in presence of zingerone indicated low amount of mRNA in the liver leading to decrease in protein levels with minimum LPS induced hepatotoxic effect. Zingerone has been found to be successful in reducing inflammation through multitargeted mechanism. In addition to free radical scavenging effects, reducing binding efficiently of LPS to LPS receptors and further interference with the activation of inflammatory signalling molecules. Results of the present study suggest that zingerone inhibited LPS-induced acute liver injury which was mediated via TLR4/NF-κB signaling pathway by suppressing the mRNA expression of inflammatory markers involved in this pathway. We hypothesize that zingerone may have altered the endotoxin receptor complex formation since ginger components particularly shogaols are known to inhibit TLR4 dimerization. Hence it may also have the potential to inhibit TLR4 dimerization or TLR4 and MD-2 complex formation. Both steps are necessary for the downstream signalling of the endotoxin induced expression of genes. The present study provides an insight on the impact of zingerone in suppressing inflammatory mediator production, reducing oxidative damage to liver tissue hence protecting liver from endotoxin induced injury. Understanding detailed mechanism of action of zingerone may lead to finding novel targets for suppression of LPS induced inflammation. # Conclusions Zingerone a nontoxic, inexpensive dietary natural compound with potent anti- inflammatory and pharmacological activities having no side effect showed hepatoprotective effect against endotoxin induced liver injury via scavenging free radicals and down regulating production of inflammatory mediators. This study opens different areas to venture zingerone as potential anti-inflammatory molecule for reducing endotoxin induced inflammation in *P. aeruginosa* infections as well as during antibiotic treatment. We acknowledge the INSPIRE programme of Department of Science and technology (DST) Govt of India. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: LK KH SC. Performed the experiments: LK. Analyzed the data: LK KH SC. Contributed reagents/materials/analysis tools: KH SC. Wrote the paper: LK KH SC.

# Introduction It has been estimated that one third of the world’s population is infected with *M*.*tuberculosis*. The lifetime risk of reactivation for tuberculosis (TB) among people with latent tuberculosis infection (LTBI) is estimated to be between 5 to 10%, with the majority developing the disease within the first five years after initial infection. The End TB Strategy thus asserts that expansion of preventive treatment for people at high risk is an essential component of the global strategy for TB prevention, care and control beyond 2015 towards the elimination of TB. The World Health Organization (WHO) similarly recommends active and systematic identification and treatment of people with LTBI for certain high–risk populations, mainly in high- and upper middle-income countries with an estimated TB incidence rate of less than 100 per 100,000. People incarcerated in correctional facilities are one such population, among who systematic testing and screening of LTBI “should be considered”, according to evidence of either low or very low quality. In Japan, there are currently 187 correctional facilities, including prisons and jails, with an average daily incarcerated population of slightly less than 60,000 persons. Health of the incarcerated population is under the jurisdiction of the Ministry of Justice, however, according to the Infectious Diseases Control Law of Japan, any physician who has diagnosed a case of TB or latent tuberculosis infection (LTBI) is required to notify the local public health center, and this applies to correctional physicians as well. It is then the responsibility of public health centers to enter patient information into the electronic national TB surveillance system, the Japan TB Surveillance (JTBS). The public health center also creates and maintains a case file in its local TB registry, with a more detailed information of the patients, collected either through interview with patients themselves, or communication with prison staff. The patient on the other hand may either be treated within his or her own facility, be sent to one of the prison hospitals, or at a non-prison hospital. When and if a patient is released while still on treatment, the correctional facility is advised to contact and coordinate with the public health center to arrange for post-release adherence support. However, although a guideline has been established regarding coordination between correctional facilities and public health centers in management of TB and LTBI among incarcerated populations, the actual practices still vary and are often influenced by personal relationships between the individual staff of public health centers and correctional facilities. Using the publicly available data, we had previous estimated that TB notification among the incarcerated population may be up to ten times higher than that of the general population, which was 13.9 per 100,000 in 2016. Yet TB screening upon entry to correctional facilities is not conducted routinely, and chest X-ray is only mandatory in the annual health-check. Active LTBI screening in correctional facilities is not conducted systematically in Japan. As for treatment, in Japan, the Guideline on Treatment of LTBI, published by the Prevention Committee and the Treatment Committee of the Japanese Society for Tuberculosis, recommends 6- or 9-months regimen by isoniazid as first option, followed by 4- or 6-months regimen by rifampicin, which is only recommended when the possibility of the use of isoniazid is ruled out. However, the current guideline does not make any particular mention of preventive therapy for people who are incarcerated in correctional facilities, who may be at a higher risk of interrupting the treatment–especially if they are released while still on treatment. No study has so far been published on LTBI treatment regimen in correctional facilities in Japan and little is known about the detailed practices. A project, funded by the Japan Society for the Promotion of Science is currently ongoing to determine the cost-effectiveness of screening for LTBI among the incarcerated population in Japan. Burden of LTBI and its treatment outcome among the incarcerated population is one of the key data needed to conduct the cost- effectiveness analysis–however, information regarding history of incarceration is not collected under the JTBS and thus it is not possible to determine the burden of LTBI or evaluate its treatment outcomes among the incarcerated population in Japan from the surveillance data. We thus conducted a survey of public health centers, whereby they were asked to retrieve information from their TB registries, with the purpose of determining the overview of the situation of LTBI, including treatment outcome, among the incarcerated population in Japan. # Method A two-step questionnaire survey, specifically designed for this study, was conducted with public health centers in Japan. The inclusion criteria for the first survey was all public health centers which have one or more correctional facilities under their jurisdiction. The survey was sent to all public health centers to investigate the number of LTBI notifications they received from correctional facilities between 1^st January 2015 and 31^st December 2016, and also to ask on management of LTBI in correctional facilities, especially regarding specific procedures, if any, for initiating preventive therapy for incarcerated persons. The inclusion criteria for the second survey was all those public health centers which have returned the first survey and which had notifications from the correctional facilities during the study period. The second survey included questions regarding demographic details of the patients, as well as clinical and treatment information and information regarding coordination with correctional facilities in ensuring continuation of care upon release. Public health centers which only reported receiving notifications for active TB were excluded. The data were entered into and managed using Excel (2013, Microsoft Corporation, US) and characteristics of the patients and treatment outcomes were analyzed, using R (version 3.1.3, R Development Core Team, Vienna, Austria). Characteristics of LTBI patients were compared with those of the newly incarcerated population in 2016, of which data was obtained from the Statistics of Correctional Facilities, compiled by the Ministry of Justice, Japan, and treatment outcomes were described. Proportions were compared using either chi-squared test or Fisher’s Exact test, as appropriate. Multiple regression was not conducted due to the small number of study population, and also large missing data for certain variables. No personal identifiers were collected, and thus informed consent was not obtained from the patients. The study protocol was reviewed and approved by the Institutional Review Board of the Research Institute of Tuberculosis, Japan Anti-Tuberculosis Association, Japan (reference number RIT/IRB 20–10). # Results ## Procedures for initiating LTBI treatment Flow chart of the study population is shown in. The first survey was sent to a total of 163 public health centers, out of which 133 (81.6%) responded. The 133 public health centers covered 153 out of a total of 187 correctional facilities in Japan. summarizes the types of correctional facilities which were covered in this study. Regarding the procedures for initiating LTBI treatment for incarcerated persons, 8 (6.0%) of the 133 public health centers actively guided the correctional facilities regarding LTBI treatment initiation based on a standardized procedure, while 10 (7.5%) responded that they entrusted the decision to correctional facilities but were aware of how the decision was being made. The remaining 115 public health centers either did not have any standardized procedures, or were unaware of how LTBI treatment was being initiated in the correctional facilities. However, of the 8 public health centers which responded that they had a standardized procedure only two took into account the remaining duration of the patient’s prison term. ## Characteristic of notified LTBI patients A total of 91 LTBI patients were notified from prisons and prison hospitals (i.e none from jails) in 2015 and 2016, to a total of 13 public health centers. However, 1 public health center failed to return the second survey, thus a detailed information of a total of 89 patients, notified to the remaining 12 public health centers, were available for analysis. 61.8% (n = 55) of the patients had been notified from one correctional facility, which had experienced a large outbreak and another 15.8% (n = 14) from another facility. The average number of notified case per facility, excluding the two facilities just mentioned, was 2 (range 1 to 5). 85 of the 89 patients had been notified as a result of contact investigation. Demographic characteristics of the 89 patients are summarized and compared with the newly incarcerated population in 2016, in. Information regarding treatment history, sources of infection and potential risk factors are summarized in. 82 (92.1%) were males, and 30 (33.7%) were aged 50–59 years. 83 (93.3%) were Japan-born. The proportion of males and of Japan-born were similar to (92.1% vs 90.2%, 93.3% vs 96.4%), however, the proportion of those aged 20–29 and 60 and above were smaller than, those of the newly incarcerated population in 2016 (2.2% vs 13.9%, 7.9% vs 18.3%). 86 (93.3%) were receiving treatment for the first time, however, two had history of previous treatment–one for active TB and another for LTBI. Sources of infection were known for 87 patients, of whom 82 were contacts of a TB patient from the same facility. Information regarding smoking, drug and alcohol abuse prior to incarceration was unknown for the majority of the patients. The most common comorbidity was hepatitis, reported in 16 (18.0%) patients, followed by hypertension (n = 11, 12.4%) and psychiatric disorders (n = 7, 7.9%). ## LTBI treatment status and outcome All the 89 patients had initiated LTBI treatment at their respective facilities. Information regarding treatment regimen was known for 82 patients, of whom all were treated with isoniazid monotherapy. Of the 82 patients, 63 were treated with 6-months regimen, while the treatment duration was not known for the remaining 19 patients. Treatment outcome was known for 88 patients, and their characteristics are compared by treatment outcome in. Of the 88 patients, 70 (80.5%) had completed treatment. Of the 18 who did not complete the treatment, 15 had been lost to follow-up upon release from the facilities while still on treatment, one had been moved to another facility, one had died while on treatment, and the other had refused to continue taking the medication. There were no statistically significant differences in sex, age groups and country of birth between those who completed and did not complete LTBI treatment. However, there was a significantly higher proportion of those who had been released while on treatment among those who did not complete the treatment, compared with those who did (42.1% vs 2.0%). summarizes the 37 patients who were released while on treatment, by their treatment outcome, and whether any sort of coordination between the public health centers and the correctional facilities took place, prior to the patient being released. In all but one case, public health center was able to receive the basic details, including the date of release and contact details of the patient after he or she has returned to the community, from the correctional facilities prior to the actual release. Nevertheless, the proportions of those who did and did not complete the treatment were relatively similar (58.3% vs 41.7%). In 12 cases, the nurse of public health centers was able to meet the patient in person to give patient education prior to the release. The proportion of those who completed the treatment was higher than those patients who did not receive the visit by the public health center nurse (75.0% vs 52.0%), though the difference was not statistically significant. # Discussion This is the first study to have evaluated the situation of LTBI and treatment outcome among incarcerated population in Japan. The characteristics of the study participants were quite similar to the general incarcerated population in terms of sex and country of birth. However, the proportion of those aged 20–29, and aged 60 and above were smaller among the study participants. It is reasonable to see a smaller proportion of those aged 60 and above in the study participants, as LTBI treatment is not actively pursued for older patients in Japan from concerns for adverse events. On the other hand, the study results could not provide a meaningful explanation as to why the proportion of those aged 20–29 was smaller among the study participants. Despite the small sample size, the study results also provided important insight and highlighted two critical policy issues, namely, the need to make it a standard practice to consider the patient’s prison term in initiating LTBI treatment, and the need for a better coordination to ensure continuation of care after release. The aforementioned guideline for treatment of LTBI in Japan recommends 9 month-isoniazid regimen over the 6-month regimen, and the 4- or 6-months regimen by rifampicin is only recommended when the possibility of the use of isoniazid is ruled out. Our results indicated that in fact, at least 71.6% (63/88) of the patients had received the 6-months regimen–however, the application of the 6-month regimen seemed not always be conditional upon the patient’s prison term as indicated from the results of the first survey. Thus, indeed, 42.0% (37/88) of the patients had been released while still on treatment, and the treatment completion rate was significantly lower among the latter, compared with those who completed the treatment within their prison term (57.9% vs 96.0%). A systematic review of isoniazid preventive therapy in correctional facilities has reported the median completion rate of 44%, ranging from 3% to 87%. Of the 18 studies included in the review, 7 reported on the completion rate within the correctional facilities, with the rate ranging from 31.6% to 87.0%. Compared with these results, the performance of LTBI treatment in the correctional facilities in Japan do appear to fare well. The completion rate was even higher than that of the general population, which has been reported to be around 70.0%, owing probably to the direct observed therapy being effectively implemented. Our results therefore suggest that incarcerated persons can and do benefit from LTBI treatment–which also has implications for discussions over LTBI screening–however, in order for the benefit to be optimized, best effort must be made to ensure completion, including considering the patient’s prison term prior to initiating the treatment and, which also leads to the second issue, coordinating to ensure continuation of care if and when patients are released while still on treatment. In the abovementioned systematic review, the completion rates in studies which have followed-up patients after release have been disappointingly low, ranging from 6% to 60%. In our study, though not statistically significant partially due to the small sample size, the proportion of those completing the treatment after release was higher in patients who received pre-release patient education by public health nurses than those who did not. However, the results of randomized controlled studies examining the effect of such intervention have been mixed. Two studies have reported that single pre-lease education and counselling did not improve post-release completion rates, while others, albeit poor outcomes, concluded that more intentional sessions did improve the completion rate by two- fold. Furthermore, several studies have suggested that post-treatment outcomes of infectious diseases, such as HIV, can be optimized with simultaneous treatment for substance abuse, which often co-exist with other health problems. In our study too, of the 16 patients whose history of substance abuse was known to the public health centers, 13 were recorded to be using illicit drugs prior to incarceration. A previous study has similarly reported high proportion of those with drug addiction among TB patients in the incarcerated population in Japan. Exchanging and sharing basic patient information between the relevant organizations is necessary but insufficient on its own, and a more concerted and organized effort is necessary to ensure continuation of care and support for patients after they are released. Finally, a note should be made of the potential for shorter regimens for LTBI treatment. In the early 2000s, a shorter regimen of 2-months rifampin and pyrazinamide (2RZ) had been recommended by the ATS/CDC and several studies were conducted in the correctional facilities. As compared with 6- or 12-months regimen by isoniazid, these studies showed significant improvement in the completion rate among those receiving shorter regimens. However, subsequent reports of severe hepatoxicity among HIV-negative recipients led to CDC amending the guideline for treatment for LTBI, preferring 9-months isoniazid regimen. In late 2011, the CDC recommended a new regimen of 3 months of weekly isoniazid and rifapentine (3HP). A systematic review with meta-analyses of efficacy and completion rates for shorter regimens, including the 3HP, concluded that these showed significant benefits in preventing active TB compared to placebos, and that regimens of 3 to 4 months were more likely to be associated with better completion. Rifapetine is currently not approved in Japan, however, discussion are ongoing regarding its use. This study is not without limitations. Due to the small sample size, we were limited to conducting minimal statistical analyses. Secondly, there was large missing data for several variables, which also precluded us from conducting additional analyses, including impact on treatment outcome. On the other hand, the status of data also revealed the difficulties public health centers continue to face in communicating with correctional facilities to collect the necessary patient information. Since the establishment of a guideline for public health centers on coordinating with correctional facilities in TB control in 2014, situation has certainly improved. However, efforts must further be accelerated to strengthen coordination and collaboration among the relevant organizations in controlling TB and LTBI in correctional facilities in Japan. # Conclusions LTBI treatment was being initiated without systematic consideration for the patients’ prison term. The treatment completion rate within jail was high, indicating the possibility that incarcerated population can benefit for LTBI treatment. On the other hand, the completion rate decreased significantly among those who had been released while still on treatment. In order to optimize the benefit, initiation of LTBI must carefully be considered upon the patient’s prison term, as well as coordination among the relevant organizations to ensure continuity of care after release. Discussions over possibility of introducing shorter regimens for LTBI treatment and its potential benefit for population with higher risk of interrupting the treatment, such as persons released from correctional facilities, should be encouraged. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction A C-arm is a fluoroscopic system comprising two units, an X-ray generator and a detector (image intensifier or flat panel) mounted in an arc-shaped gantry, together with a workstation used to visualize, store, and manipulate the images. Designed to acquire real-time planar images, C-arms have demonstrated to be a useful qualitative assessment tool to guide surgical procedures thanks to their open design, compactness and portability, which allow to set the C-shape around the patient lying in the bed. Another advantage is their low cost in comparison with other medical image modalities. However, there is an important drawback intrinsic to its 2D nature: the lack of depth information. In situations where a Computed Tomography (CT) system is not an option (either because unavailability or patient mobility restrictions), the use of a C-arm as a tomograph would raise the possibility of generating tomographic information, with the potential of improving diagnosis and likely surgical precision. So- called cone-beam CTs, based on advanced isocentric motorized C-arm systems (generally attached to a gantry), have been used in a broad range of image- guided procedures, including image-guided radiotherapy, mammography, dentistry imaging and general surgical procedures. When cost is a relevant issue, it could be helpful to obtain tomographic information using more basic C-arm systems. The straightforward use of a standard non-isocentric C-arm for tomography presents significant difficulties, apart from the geometrical non-idealities common to any Cone Beam Computed Tomography (CBCT). The first important problem to face is that the relative positions between source and detector may change for different projections. This derives from the mechanical strains in the arm due to the heavy loads at its ends. This fact, together with the non-circular orbit of C-arms determined by the slipping of the arm over the base, hinders the use of calibration methods widely used in CT systems that obtain global calibration parameters for all projections, such as the one proposed in. Other different methods have been proposed to generate the geometrical calibration of an imaging system projection by projection. The approach based on the so called “camera model” calibration, used for planar imaging, has shown instability and does not provide the center of rotation, rendering it not suitable for tomography systems. Cho et al. proposed a method specifically designed to obtain calibration parameters of a cone-beam CT system individually for each projection, based on the acquisition of a simple phantom with two circular patterns. A second problem is that the utility of the geometrical parameters obtained with any periodical calibration procedure depends on the mechanical stability of the imaging system. For C-arms aimed for planar imaging, the exact position of the source and detector elements tends not be repeatable for consecutive rotations due to low mechanical precision, thus making it impossible to achieve the accuracy required for obtaining good quality 3D reconstructions. Finally, due to physical movement limitations, only a few projections can be acquired, typically covering a much smaller angular span than the one used in conventional CBCT, where the system rotates around the patient through 360 degrees (full angular span), generally acquiring more than 360 projections. The reconstruction of these limited data implies an extremely ill-posed inverse problem where the use of conventional methods, such as the one proposed by Feldkamp, Davis and Kress (FDK), results in severe artifacts in the image (streaks and/or shape distortion). In this work, we present a method based on the use of patient surface information obtained with a 3D surface scanner that enables adapting a standard C-arm, originally designed for planar imaging, to be used as a tomograph. The key parts of the novel method are a geometrical calibration that compensates the high mechanical inaccuracy that prevent any accurate repetition of source- detector positions between acquisitions together with an advanced image reconstruction algorithm able to deal with very limited angular span and non- circular trajectories. # Proposed calibration-reconstruction protocol shows the workflow of the proposed method. A geometrical calibration of the system is obtained periodically, as it is customary in standard CT scanners, with an algorithm based on a method specifically designed to obtain individual calibration parameters for each projection angle proposed by Cho et al.. The result of the calibration is the set of parameters values shown in the right panel of for each projection angle: detector rotation (skew), inclination angles (pitch and roll), piercing point location (projection of the center of the calibration phantom), SDD (source to detector distance) and source and detector position. For each study, together with the projection data acquired at different source- detector positions we also obtain the surface information of the sample using a 3D surface scanner. We obtain a preliminary reconstruction of the sample from the projection data with the FDK-based method proposed in using the geometric calibration data. In parallel, we obtain a binary mask of the sample, and perform a 3D rigid registration based on mutual information with the preliminary reconstruction. The generation of the tomographic image is then achieved in two steps, using the previously registered mask of the sample. First, we refine the system calibration to calculate accurate positions of the source and detector in the current acquisition, by comparing the projections of the registered mask with the acquired projection data. Second, using the refined calibration, the image is reconstructed from the projection data also taking advantage of the available registered mask of the sample. The key steps of the proposed method are detailed in the following sections. ## Geometric calibration The algorithm we use for the geometric calibration is based on the work by Cho et al. As a thorough description can be found in the paper by Cho et al, in this section we only describe the key points and details specific for our implementation. The calibration phantom is a Polymethyl methacrylate cylinder with two circular patterns formed by ball bearings symmetrically embedded in the cylinder wall. Projections of the phantom are obtained at the angular positions expected to be used in the specific trajectory during the acquisition of the sample. For each projection, the algorithm completes 6 steps: 1) identification of center of mass positions of ball bearings, 2) ellipse fitting, 3) determination of *v*-*h* offset, 4) determination of skew angle (*η*), 5) determination of converging point and inclination angles (*θ* and *φ*), and 6) determination of source and detector positions. An initial segmentation of the markers is achieved in a semi-automatic way using a calibration program with an interactive graphic user interface, based on a global thresholding segmentation followed by morphological operations to remove artifacts. An initial threshold is estimated by Otsu’s method, but can be interactively adjusted by the user with a slider. Once the segmented markers are well identified, they are automatically classified into two sets (corresponding to each ring), according to their relative position with respect to the image center. To overcome the problem of overlapping markers, the tool allows the user to interactively correct the marker positions (place, remove or reposition), showing the updated ellipses overlaid to the image in real time. Following Cho et al, ellipses are parametrized as: $$a{(h - h_{0})}^{2} + b{(v - v_{0})}^{2} + 2c(h - h_{0})(v - v_{0}) = 1$$ where (*h*_*0*, *v*_*0*) are the coordinates of the ellipse center and *a*, *b*, and *c* are parameters that describe its shape. To relate these parameters with the coordinates of the centers of mass of the balls previously obtained with the calibration program (*h*_*b*, *v*_*b*), we use the polynomial description of the ellipse: $$p_{0}h_{b}^{2} + v_{b}^{2} - 2p_{1}h_{b} - 2p_{2}v_{b} + 2p_{3}h_{b}v_{b} + p_{4} = 0$$ where *p*_*i* are obtained solving the following system: $$\begin{pmatrix} h_{0}^{2} & {- 2h_{0}} & {- 2v_{0}} & {2h_{0}v_{0}} & 1 \\ & & \vdots & & \\ h_{7}^{2} & {- 2h_{7}} & {- 2v_{7}} & {2h_{7}v_{7}} & 1 \\ \end{pmatrix}\begin{pmatrix} p_{0} \\ \vdots \\ p_{7} \\ \end{pmatrix} = \begin{pmatrix} {- v_{0}^{2}} \\ \vdots \\ {- v_{7}^{2}} \\ \end{pmatrix}$$ Finally, we calculate ellipse parameters in using the relations described by Noo et al.: $$h_{0} = \frac{\left( {p_{1} - p_{2}p_{3}} \right)}{\left( {p_{0} - p_{3}^{2}} \right)},\text{v}_{0} = \frac{\left( {p_{0}p_{2} - p_{1}p_{3}} \right)}{\left( {p_{0} - p_{3}^{2}} \right)},a = \frac{p_{0}}{\left( {p_{0}h_{0}^{2} + v_{0}^{2} + 2p_{3}h_{0}v_{0} - p_{4}} \right)},b = \frac{a}{p_{0}},c = p_{3}b$$ We then calculate the horizontal and vertical displacements and the skew of the detector panel, as well as the source and detector positions following the Cho et al. method. Regarding inclination angles, if both pitch angle, *θ*, and roll, *φ*, are zero, both ellipses have the same shape and their long axis are parallel, as shown in (top, left). When *θ* is different from zero, one ellipse is lengthier than the other, as illustrated in the right panel of. The lines tangent to both ellipses converge to the *pitch converging point*, *P*_*θ*. Since the calculation of both inclination angles in Cho et al. is based on *P*_*θ* (equation 18 in) pitch angle *θ* is assumed to be different from zero. To remove this restriction, we propose the following modification to the Cho’s method to estimate both effects separatelly. The pitch angle is obtained from the geometric relationships: $$\theta = \text{arcsin}\left\lbrack \frac{Z_{S}\text{cos}(\varphi)}{v_{\theta}} \right\rbrack,Z_{S} = \left\lbrack \frac{2 \cdot R \cdot L_{1} \cdot L_{2}}{H \cdot \left( {L_{2} - L_{1}} \right)} \right\rbrack$$ where *Z*_*S* is the source distance, *L*_*1* and *L*_*2* are the distances between the two ellipses (right), and *R* and *H* are the radius and the distance between the two circular patterns in the real phantom. The roll angle is calculated by considering that when *φ* is different from zero, the long axis of both ellipses converge to the roll converging point *P*_*φ*(*h*_*φ*, *v*_*φ*), as shown in the left- bottom panel of. Roll angle, *φ*, is obtained using the Nelder-Mead simplex method to minimize the cost function *C*_*φ*: $$C_{\varphi} = \text{sin}(\varphi) + \text{c}_{1} \cdot \beth_{1}/2 \cdot (a_{1}) + \text{c}_{2} \cdot \beth_{2}/2 \cdot (a_{2})$$ where *ℶ*_*1* and *ℶ*_*2* are intermediate parameters for each ellipse given by: $$\beth_{k} = \frac{T \cdot a{}_{}^{}\sqrt{a_{k}}}{\sqrt{a_{k} \cdot b_{k} + a_{k}^{2} \cdot b_{k} \cdot \left( Z_{S}^{\prime} \right)^{2} - c_{k}^{2}}},k = 1,2$$ with *T* given by $$T = h_{\varphi} \cdot \text{sin}(\varphi) \cdot \text{cos}(\varphi)$$ ## Adaptive refinement of geometrical calibration The first algorithm of our proposed method, *adaptive refinement of geometrical calibration*, tackles the problem of low repeatability of the exact positions of the source and detector due to the high mechanical inaccuracy of these systems. This is achieved through the refinement of geometrical parameters in the detector plane: horizontal offset, *Δh*, vertical offset, *Δv*, and skew, *ΔS*. shows a graphical workflow of the algorithm. First, we generate a preliminary reconstruction, *reco*_*pre*, which is used to orientate and adjust the mask to the field of view of the C-arm, by means of a 3D registration based on mutual information. The registered mask is then projected using the initial system calibration, *calib*_*ini*. The misalignment between the projections of the mask, *mask*_*proj*, and the acquired data, *data*, reflect the errors in the values *H*, *V*, and *S* of *calib*_*ini*. Therefore, the parameters of a 2D registration between *mask*_*proj* and *data* at each projection angle, *ΔH*_*reg*, *ΔV*_*reg* and *ΔS*_*reg*, are used to generate refined values *U*_*corr*, *V*_*corr*, and *S*_*corr* as: $$H_{corr} = H + \Delta H_{reg} + O_{h} - \left( {O_{h} \cdot \text{cos}\left( {\Delta S_{reg}} \right) - O_{v} \cdot \text{sin}\left( {\Delta S_{reg}} \right)} \right)$$ $$V_{corr} = V + \Delta V_{reg} + O_{v} - \left( {O_{h} \cdot \text{sin}\left( {\Delta S_{reg}} \right) + O_{v} \cdot \text{cos}\left( {\Delta S_{reg}} \right)} \right)$$ $$S_{corr} = S - \Delta S_{reg}$$ where *O*_*h* and *O*_*v* are the coordinates of the detector center. ## Surface-constrained image reconstruction Image reconstruction of the limited data is done with a Surface-Constrained Method for Limited Data tomography (SCoLD) developed by our group, using projection-backprojection kernels implemented in GPU to reduce computational burden. The reconstruction algorithm follows the idea of the Total Variation (TV) minimization subject to a support constraint, which contains the *a priori* sample mask information, formulated as: $$\underset{u}{\text{min}}TV(u),s.t.\left\| {Au - f} \right\|_{2}^{2} < \sigma^{2},u \geq 0,u \in \Omega$$ where *u* is the reconstructed image, *Ω* the subspace that corresponds with the surface support of the sample, *A* is the system matrix, *f* represents the acquired data and *σ*² is the image noise. The L₁-constrained optimization problem, shown in, is efficiently solved using a Split Bregman formulation and expressed as the following unconstrained problems, which are sequentially solved at each iteration *k*: $$\begin{array}{r} {\left( {u^{k + 1},d_{x}{}^{k + 1},d_{y}{}^{k + 1}} \right) = \underset{u,d}{\text{min}}\left\| \left( {d_{x},d_{y}} \right) \right\|_{1} + \frac{\mu}{2}\left\| {Au - f^{k}} \right\|_{2}^{2} + \frac{\lambda}{2}\left\| {d_{x}^{k} - \nabla_{x}u - b_{x}^{k}} \right\|_{2}^{2} +} \\ {+ \frac{\lambda}{2}\left\| {d_{y}^{k} - \nabla_{y}u - b_{y}^{k}} \right\|_{2}^{2} + \frac{\gamma}{2}\left\| {v - u - b_{v}^{k}} \right\|} \\ \end{array}$$ $$b_{x}^{k + 1} = b_{x}^{k} + \nabla_{x}u^{k + 1} - d_{x}^{k + 1},b_{y}^{k + 1} = b_{y}^{k} + \nabla_{y}u^{k + 1} - d_{y}^{k + 1}$$ $$f^{k + 1} = f^{k} + f - Au^{k + 1}$$ $$b_{v}^{k + 1} = b_{v}^{k} + u^{k + 1} - v^{k + 1}$$ Eq leads to two sub-problems: the first one, which contains only L₂ norm terms, is solved iteratively using a Krylov space solver; the second one, with the L₁ terms, is solved using analytical formulas. Eqs, and are the Bregman iterations that impose the data constraint and surface constraint, respectively. # Evaluation with real systems We have tested the proposed method using two different C-arm devices: a commercial C-arm based on an image intensifier (SIREMOBIL, SIEMENS) and an in- house C-arm prototype based on a flat panel detector. The calibration was performed for each system using a calibration phantom with two circular patterns of diameter 49.1 mm separated 35 mm, each one formed by eight ball bearings (with 0.8 mm of diameter) symmetrically located around a cylinder with internal and external diameters of 45.5 mm and 49.5 mm respectively. For testing we used one hand and one foot of a PBU-60 anthropomorphic phantom (Kyoto Kagaku Co., Kyoto, Japan). Both phantoms were acquired beforehand in a Toshiba Aquilion/LB helical scanner and reconstructed as a CT volume of 512×512×1645 voxels, with 0.931×0.931×0.5 mm voxel size. A simulated mask was obtained from the CT volume by thresholding. Quantitative evaluation was done using two metrics. The reduction of the artifacts due to the low number of projections was evaluated using the Streak Artifact Indicator (SAI), which measures the total variation of the difference between reconstructed images and the CT volume. The compensation of the distortion due to the limited angular span was measured with the Limited View Artifact metric (LiVA). LiVA was calculated as the RMSE between the reconstructed images and the CT volume within an ROI delimited by the external contour of the sample, as shown in for the two phantoms used. Data were reconstructed on a computer with an Intel(R) Core(TM) i7- 7700 processor at 3.6 GHz and one NVidia GeForce GTX 1060 6GB GPU with the FDK-based method proposed in and SCoLD with *μ* = 35, *λ* = 5, *γ* = 0.02 and 35 iterations. ## C-arm Model SIREMOBIL The C-arm Model SIREMOBIL Compact L of Siemens consists of an X-ray source and a 23 cm-diameter image intensifier attached to a C-arm gantry. The detector of the SIREMOBIL system is an X-ray image intensifier (XRII) connected to a TFT monitor for image visualization. We exported the acquired images to a PC through a USB port using an external video-capture device (model NPG USB RealStudio II), which copied static images and video data onto the connected computer. To evaluate the lines distortion and vignetting of the intensifier, we used a phantom consisting of an old electronic board with radio-opaque cupper straight lines at right angles (left) attached to the intensifier outer casing. Profiles taken along the line patterns on the projection images showed a straight pattern indicating no significant line distortion or vignetting in the image intensifier. Nevertheless, the projection showed a slightly oval shape (yellow lines in, right) that does not match the circular shape of the image intensifier case. A difference of 8% found between small and large diameter was corrected by the geometrical transformation: $$\left( \begin{array}{l} {x´} \\ {y´} \\ 1 \\ \end{array} \right) = \begin{pmatrix} 1 & 0 & 0 \\ 0 & 1.08 & 0 \\ 0 & 0 & 1 \\ \end{pmatrix}\left( \begin{array}{l} x \\ y \\ 1 \\ \end{array} \right)$$ The rotation around *y* axis (vertical rotation) is severely non-isocentric due to the slipping of the arm over the base. To maintain a reasonable Field of View (FOV) size seen from all projection angles, we acquired 60 projections rotating around the horizontal supporting arm (horizontal rotation), following an orbit close to be isocentric. As with other conventional C-arms, not designed for tomography, the movement of our device is manual and the only angular positioning information provided is a rudimentary scale drawn on the arm (left). To increase the accuracy of angular positioning, we implemented a position-recording device based on an ADIS16209 digital inclinometer, with an accuracy of 0.1 degrees (left). This system was connected to the computer by a single-board microcontroller, the model TM4C123G of LaunchPad board (Texas Instruments). To avoid inaccuracies due to possible vibration of the C-arm, we averaged ten consecutive readings from the inclinometer. Position values, measured as relative increments in the range of ±90 degrees, were transformed into absolute values ranging from 0 to 360 degrees to be used in the reconstruction. The direction of rotation was estimated from previous recorded positions. shows axial, sagittal and coronal views of the image reconstructed with FDK (before and after *adaptive refinement of geometrical calibration*) and reconstructed with SCoLD. ## Evaluation with in-house built C-arm prototype The in-house built C-arm prototype is based on a wireless, flat panel detector, the XRpad 4336 (PerkinElmer Inc., US), with an imaging area of 35 cm × 43 cm. The X-ray generator is a light weight integrated system Transportix (Radiologia, Algete, Spain)) with 4 KW, 125 kVp, 100 mA, 0.001–10 s, 0.1–250 mAs, and dual focal spot 0.6–1.5 mm. The distance from source to detector is 125 cm, with a useful FOV for reconstruction of 17 cm. We obtained 49 projections, with a matrix size of 444×540 and pixel size of 0.8 mm, within an angular span of 120 degrees using the rotation along C-arm plane of both hand and foot, as shown in. shows values of horizontal offset and skew of the detector for two calibrations obtained consecutively. The difference between the curves clearly demonstrates the non-repeatability of the source and detector positions and the need for a calibration refinement procedure. Reconstructions using FDK based on any single system calibration show severe artifacts (white arrows). The application of our calibration refinement algorithm to refine geometrical parameters *H*_*corr*, *V*_*corr*, and *S*_*corr* leads to a much-improved reconstruction (panels B and D). shows the reconstructed volume of the hand using the refined geometrical calibration with FDK and SCoLD with the mask, for limited data with different number of projections. Reconstruction of a volume of 250×250×500 voxels (0.6 mm isotropic) took 5 and 6 minutes for 25 and 49 projections respectively. We evaluated a possible implementation of the surface extraction using a *3D Artec Eva* scanner (Artec3D, Luxembourg), with a maximum spatial resolution of 0.5 mm and a 3D point accuracy of 0.1 mm, to scan the surface of the foot. The software *Artec Studio* was used to create a polygonal mesh from which a 3D binary mask of the surface was obtained. We could not replicate this experiment with the hand phantom since it is wrapped in a plastic cover that prevents the acquisition of its real surface with a surface scanner, as it can be seen in (top-right). shows the result of the reconstructed image using the refined geometrical calibration with FDK and SCoLD using the mask obtained with the surface scanner. Reconstruction of a volume of 512×512×364 voxels (0.6 mm isotropic) took 15 minutes. SCoLD presented significantly lower SAI and LiVA than FDK for the three cases, with an average reduction of 81% and 67.44% respectively. # Discussion and conclusion We propose a novel method to achieve tomographic capabilities using basic non- motorized C-arms, intended for planar imaging. There are three main challenges to address: (1) the trajectory of the source-detector pair may differ from a circular path and the system may suffer mechanical strains that modify the relative positions of the source and detector for different projection angles; (2) the exact position of the source and detector elements may not be repeatable for consecutive rotations due to low mechanical precision, thus preventing an accurate geometrical calibration of the system, and (3) the limitation of the angular span and the relatively low number of projections pose a great challenge to the image reconstruction, that leads to severe artifacts when using conventional reconstruction methods. The method proposed here is based on exploiting the surface information of the sample, which can be obtained with a 3D surface scanner. The results using data from two real C-arm systems, based on IR and flat panel detectors respectively, showed the feasibility of the proposal. The use of the surface of the sample enables the fine-tuning of geometrical calibration parameters and the removal of the artifacts derived from a low number of projections and limited angular span. To address the issue of system calibration, Cho et al. proposed a method specifically designed to obtain calibration parameters of a cone-beam system individually for each projection. However, this approach does not provide a complete estimation of the inclination angles (pitch and roll) because it is based on solving an optimization problem under the assumption that one of them is always different to zero. In the case of planar systems, not designed for obtaining tomographic data, this assumption may lead to artifacts, and for this reason we have extended the method by Cho et al. to include the estimation of the two inclination angles of the detector with respect to the horizontal and vertical directions independently. Regarding the mechanical inaccuracy, which prevents a precise repetition of source-detector position between acquisitions, our results show that a periodic system calibration, standard in CT systems, is not enough to avoid misalignment artifacts in the reconstructed image. The *adaptive geometrical calibration* proposed here enables the refinement of three geometrical parameters, horizontal offset, vertical offset and skew, for the specific acquisition. Although possible errors in the rest of the geometrical parameters are neglected, results showed that this refinement increased accuracy so as to obtain images free of misalignment artifacts. The first step of the refinement algorithm used mutual information to co-register the surface with the preliminary FDK reconstruction of the limited data. This might be difficult if the preliminary reconstructed volume is severely distorted due to a very low angular span or to truncation artifacts if the sample falls out of the field of view. This problem could be tackled by using markers on the sample, visible both in the acquired projection data and the surface scanner, which would allow a point-based registration instead. Given the high spatial resolution of the ARTEC EVA scanner used, radiopaque markers commonly used in the clinical practice (around 2 mm) could be used for this step. Regarding the reduced angular span and the difficulty of obtaining a high number of projections, our *surface-constrained image reconstruction* reduces significantly the streak artifacts derived from the low number of projections, similarly to previous proposals in the literature. However, its main advantage over previous works is that the restriction on the search space by exploiting the surface-based support results in a complete recovery of the external contour of the sample and adjacent areas even for extremely reduced angular spans. The reconstructed images showed an average SAI reduction of 81%, and limited angular span, with an average LiVA reduction of 67.44% when using SCoLD algorithm. Some horizontal patterns visible in the coronal views can be explained by the fact that the present implementation of SCoLD calculates the derivatives only in 2D, as it can be seen in Eqs–. We are currently implementing the 3D version, which we expect to avoid this effect. The nominal accuracy of the ARTEC EVA scanner (0.1 mm) proved to be enough for the generation of the mask, since it is under the voxel size. However, the structured-light scanner showed two limitations: its elevated cost and possible errors due to the body hair or to the use of baggy clothes that may provide a wrong surface of the patient. The first problem could be solved using less expensive instrumentation, such as the Kinect camera system, and the second problem using infrared cameras. The reconstruction time of several minutes may limit the use of the proposed method in some applications, like image-guided surgery where quasi real time is required. Optimization of this reconstruction time was out of the scope of this work, since our goal was to develop a proof of concept. Anyhow, we expect to reduce reconstruction times down to 50% by using high-performance workstations with multi-GPU architecture, as we have previously shown for similar algorithms. Besides, a dedicated implementation for specific acquisition protocols would further reduce computational burden. Evaluation was done with high contrast phantoms (soft tissue and bone). Further evaluation on more complex studies with more subtle soft-tissue contrast (abdomen) or tissue heterogeneity (chest) are advisable. The use of the proposed method with other X-ray systems originally designed for planar imaging is straightforward provided that they allow moving the X-ray source and/or detector, opening the possibility of obtaining 3D information in other conventional radiology scenarios, such as tomosynthesis. This work was supported by Spanish Ministry of Economy and Competitiveness (projects TEC2013-47270-R and RTC-2014-3028-1), Spanish Ministry of Economy, Industry and Competitiveness (projects DPI2016-79075-R AEI/FEDER, UE—Agencia Estatal de Investigación and DTS17/00122 Instituto de Salud Carlos III—FIS), and co-financed by ERDF (FEDER) Funds from the European Commission, “A way of making Europe”. The CNIC is supported by the Spanish Ministry of Economy, Industry and Competitiveness and the Pro CNIC Foundation, and is a Severo Ochoa Center of Excellence (SEV-2015-0505). [^1]: The authors have declared that no competing interests exist.

# Introduction Forced oscillation techniques (FOT), in particular impulse oscillometry (IOS), have been introduced as complementary approach to conventional pulmonary function testing. FOT determines mechanical characteristics of the respiratory system via the assessment of impedance, providing the real and the imaginary part, i.e. resistance and reactance, respectively, of impedance over a wide range of frequencies. It is an effort-independent and patient-friendly technique, requiring only minimal cooperation by the patient without special breathing maneuvers. Therefore, FOT is well accepted by pediatricians for pulmonary function testing in young children who cannot perform spirometry properly, and its feasibility and clinical usefulness have been demonstrated. In adults, its diagnostic value for obstructive lung disease in comparison to spirometry or body plethysmography is still controversial. Despite this it could be suspected that FOT is especially suited for lung function testing in elderly patients who cannot perform demanding maneuvers due to mental or physical impairments. Pezzoli and colleagues showed that almost 20% of patients aged between 65 and 94 years were unable to perform spirometry satisfying ATS/ERS requirements. Moreover, the time needed for spirometry was as long as 20 to 30 minutes which means a handicap for both clinical use and epidemiological studies. FOT may serve as an alternative in these patients in view of the fact that the population is getting older, lung diseases are becoming more frequent and feasible diagnostic tools are needed. Although total respiratory impedance is known to depend on age, only few reference studies exist in adults,, with only one covering the advanced age for IOS indices in Japanese adults. Reduced lung elasticity and increased stiffness of the chest wall contribute to an impairment of lung function with age. Further characteristics are airspace dilatation and increased collapsibility of small airways. Correspondingly, forced vital capacity (FVC) and forced expiratory volume in 1 second (FEV₁) decline while residual volume increases. In contrast, airways resistance as measured by body plethysmography is much less age-dependent than spirometric indices. Based on this, the aim of the present study was to determine the age-dependence of IOS indices in healthy subjects of middle and advanced age and to establish up-to-date reference equations. For this we used two representative population samples from the South of Germany in which IOS measurements had been carried out in 1990 subjects covering an age range for reference equations from 45 to 85 years. # Methods ## Ethics Statement The KORA studies (Cooperative Health Research in the Augsburg Region, Germany) were approved by the Ethics Committee of the Bavarian Medical Association and written informed consent was obtained from the study participants. ## Subjects and Lung Function Spirometry and IOS were determined in two cohorts of age 45 to 65 years (KORA-F4L, n = 1,050) and 65 to 91 years (KORA-AGE 1, n = 960) within a representative population cohort of the Augsburg region. KORA F4L is the 3-year follow-up of the KORA F4 cohort, where spirometry has been carried out in 1,321 subjects between 2006 and 2008. Due to decease, relocation, or other reasons 1,293 subjects have been invited for the F4L follow-up in 2009. From those 1,050 subjects (81.2%) performed spirometry and IOS measurements in KORA-F4L. The entire KORA AGE 1 cohort is comprised of 5,991 subjects from which a randomly selected sample of 1079 participants has been invited for physical examination to the study centre. Due to physical conditions, contraindications, or exhaustion 119 subjects had to be excluded from lung function so that 960 subjects (89%) were examined by spirometry and IOS. Both studies were approved by the Ethics Committee of the Bavarian Medical Association and informed consent was obtained from the study participants. Conditions for lung function measurements including the main examiners were the same in both cohorts with IOS measurements preceding spirometry. Standing height and weight were measured on the day of examination while subjects were wearing light clothes without shoes. Standard spirometry was performed in line with ATS/ERS recommendations in a sitting position using the Masterscope PC spirometer (Erich Jaeger/CareFusion, Germany) while subjects were wearing nose clips. The Lilly-type pneumotachograph was calibrated daily using a calibration syringe supplied and certified by the manufacturer. Additionally, daily self-testing of the examiners (biological control) was performed. Under the guidance of the experienced examiners at least 3 and at most 8 trials were recorded to obtain at least 2 acceptable and reproducible flow-volume curves. After completion of each test, the curves were visually inspected, maneuvers with artefacts excluded and results evaluated according to ATS/ERS recommendations. Spirometric indices measured were maximal forced expiratory volume in 1 second (FEV₁), maximal forced vital capacity (FVC), and the Tiffeneau index FEV₁/FVC. IOS measurements and their quality control were conducted in line with ERS Task Force recommendations using the Masterscreen IOS (Erich Jaeger/Care Fusion, Germany). For quality control the Lilly-type pneumotachograph flow transducer (screen resistance of 36 Pa^.s^.L⁻¹, flow accuracy ±1 mL; volume computer integrated, accuracy 1 mL) was calibrated daily using a calibration syringe supplied and certified by the manufacturer. Additionally, a reference impedance (0.2 kPa s L⁻¹) provided by the manufacturer was used every two weeks to check the pressure transducer calibration and the overall spectral accuracy of the IOS set-up. Throughout the measurement period values for Rrs (0.2 kPa s L⁻¹) and Xrs (0.0 kPa s L⁻¹) spectra were frequency independent and variations smaller than the recommended 10% or 0.01 kPa s L⁻¹. IOS measurements were performed in a sitting position. A nose clip was used and subjects were asked to support their cheeks with their palms and keep their lips sealed tightly around the mouth piece while breathing quietly. After adaptation to the setup, the subjects performed a minimum of 3 consecutive measurements of 30 s each with an impulse interval of 0.2 seconds, i.e. 5 impulses per second. During data acquisition time trends of flow, volume, and respiratory system impedance at 5 Hz (Zrs5) were monitored online by the examiner. After each test, respiratory system resistance (Rrs), reactance (Xrs) and the coherence (Co) between 5 Hz and 35 Hz as well as the volume dependence of Zrs5 were visually inspected and checked for artefacts, such as irregular breathing, hyperventilation, leakages or swallowing. Measurements with artefacts were discarded and, if possible, within the time schedule repeated. As outcome measures we used respiratory system impedance at 5 Hz (Zrs), resistance (Rrs) and reactance (Xrs) between 5 Hz and 35 Hz in 5 Hz increments (R5–R35 and X5–X35, respectively), resonant frequency (Fres), integrated area of low-frequency reactance (AX) from zero line between X5 and resonant frequency, the absolute difference of R5 and R20 (R5–R20), and the relative difference ((R5–R20)/R20). IOS was performed in 1990 subjects. From those, IOS measurements with coherence values below 0.6 between 5 Hz and 15 Hz or below 0.7 for frequencies \>20 Hz were rejected, as well as measurements with atypical resting breathing patterns, i.e. tidal volume above 2.5 L for men and 2.0 L for women or a breathing frequency above 30 min⁻¹. Although care was taken to avoid artefacts during measurements normal spirometry associated with extreme resistance and reactance values and abnormal flow were suspicious for artefacts during the Zrs estimates. In case a measurement is considered artefactual, the guidelines recommend rejection of both, Rrs and Xrs. Accordingly and to be on the safe side, those measurements were discarded from analysis (139 subjects). Further, within subject variability as assessed by the coefficient of variation of Zrs5 had to be smaller than 15%, with typical values below 10%. As a result of all quality criteria, 1811 subjects with valid IOS measurements were left. From those 397 lung healthy subjects, aged 45 to 89 years, were identified after exclusion of subjects with spirometric values below the lower limit of normal, acute respiratory tract infections, a history of allergic rhinitis, asthma, or COPD (due to symptoms, medication (inhaled beta-agonists, steroids, and/or anticholinergic drugs) and/or diagnosed by a physician), and ever-smokers. These conditions were assessed by questionnaire or from the Pharmazentralnummer (PZN) of the medication taken. Questionnaires and information on medication (beta- blockers, diuretics, and antiarrhythmics) were used to assess the prevalence of cardiac diseases covering myocardial infarction, heart failure, coronary artery disease, and cardiac arrhythmias including atrial fibrillation. ## Statistical Analyses Quantile regression models were used to calculate sex specific reference equations for IOS indices. In contrast to linear regression, quantile regression requires no distributional assumption and therefore is more robust to outlying observations. Thus, it allows an adequate estimation of percentiles of the outcome. In the present study, prediction equations were derived separately for the median as well as the 5th and 95th percentile, i.e. the lower and upper limit of normal (LLN and ULN, respectively) because most of the IOS indices exhibited no normal distribution (Shapiro-Wilk test). From the methodological point of view the 2.5^th and 97.5^th percentiles would be more appropriate for normalcy but conventionally the 5^th and 95^th percentiles have been used in respiratory medicine. Age and height were included as independent variables in the model and the additional influence of weight was tested using the analysis of deviance. Beside linear models we also tested more complex regression models with quadratic terms of age, height and weight. It turned out, however, that the linear models fitted best according to the analysis of deviance. The derived reference equations were compared to prediction equations for IOS indices proposed by Newbury et al. and Vogel and Smidt. Differences between predicted and observed values are presented as mean difference in percent of mean observed values and mean squared difference. Furthermore, the proportion of observed values above the predicted upper limit of normal (ULN) and below the lower limit of normal (LLN), respectively, were calculated. Bland Altman plots illustrate the differences between predicted values based on the equations from the current study and those from other studies. Additionally, predicted values for selected resistance and reactance between 5 and 35 Hz for males and females at the age of 50 and 80 years were calculated using reference equations from the present study and other studies based on mean height and weight. Differences between males and females for resistance and reactance values at different frequencies were tested using Wilcoxon rank sum test and p-values below 0.05 are used to indicate statistical significance. All analyses were performed using the statistical software R, version 2.15.1 (<http://www.r-project.org>). # Results ## Study Population From the 1990 subjects studied 179 did not fulfil the quality criteria due to suspicion of mouth piece leakage, atypical resting breathing patterns, low coherence values, or large within subject variability leaving 1811 subjects. Due to the strict inclusion criteria 397 of those subjects were considered as lung healthy, with a sex ratio of 1.5 to 1 in favour of females. FEV₁, FVC, and the FEV₁/FVC ratio showed an age-dependent decline between KORA F4L and AGE. Percent predicted values of both cohorts were within the normal range. Tidal volume (Vt) and respiratory rate (BF) reflected resting conditions but upper and lower limits indicated extremes that occur during lung function testing of large samples. Both sexes showed an equal age distribution throughout the age range of 45 to 85 years, with rarefication at older ages. The aged subjects addressed in the study represented almost 60% of the population. The exclusion rates were relatively evenly distributed throughout ages but lower in females than in males mainly due to higher rates of never smokers in females. ## Frequency Dependency of Resistance and Reactance At all frequencies women showed higher resistance values than men (p\<0.05), while reactance at low frequencies, up to 20 Hz, was lower (p\<0.05). Most of the IOS indices, particularly resistance indices, were not normally distributed. ## Age Dependency of IOS Indices Individual values from all subjects as a function of age are provided in. Moreover, for a ficticious subject showing median height and weight of the study population (phantom subject), lower and upper limits of normal (5th and 95th percentile) and the median were evaluated by our newly derived reference equations. Among the IOS indices, a significant age dependency was observed for the difference between R5 and R20, AX, Fres in both sexes and X5 in females. ## Reference Equations For both sexes, predictive equations for the median, the 5th and 95th percentile of commonly applied IOS indices are provided in (men) and (women); the equations for other indices are given in the supplement. Body weight more often contributed to the model statistically significantly in women than in men. Exclusion of subjects with cardiac diseases did not affect the reference equations significantly. This may be related to the fact that participants visiting the study center typically have early and less severe disease stages which may not affect lung function. Comparisons of the measured resistance and reactance data with the equations by Newbury et al. (n = 125), and Vogel & Smidt (n = 584) are illustrated in. Regarding the Newbury equations, the Bland-Altman plots for R5, R20 and X5 showed obvious systematic differences for both sexes, except for R20 in men and X5 in women. Correspondingly, frequency-dependent resistance and reactance for 50 and 80 years old phantom subjects showed clear differences for R5, whereas those for X5 appeared to be smaller. In our lung healthy study population mean differences between observed values and predictions were marked for R5 and X5 in men when comparing Newbury with our equations, 29.2% versus 3.9% and 21.3% vs. −0.3%, respectively. In women distinct differences were observed for R20 (−17.8% vs. 1.4%). Obvious differences to Vogel & Smidt equations were detected for R20 and X5 in both, men (−25.5% vs. 5.2% and 76.3% vs. −0.3%, respectively) and in women (−33.8% vs. 1.4% and 22.6% vs. 2.0%, respectively). The percentages of values above the ULN or below the LLN (5% expected by definition) also varied. In men much higher values were achieved for R5 (16.2% vs. 5.8%) and X5 (13.0% vs. 3.9%) when compared to Newbury et al. With respect to the Vogel & Smidt equations, almost no in women and only a low percentage in men (\<1.3%) were beyond the limits of normal. To further characterize differences in the discriminative power between the equations, comparisons between observed and predicted values and the number of subjects beyond the limits of normal were performed in the entire F4L and AGE population with acceptable IOS measurements (n = 1811). Mean differences of percentages in men ranged between −17.7% and 81.4% and −24.7% and 38.6% in women. With respect to Newbury et al. the percentage of values above the ULN was higher for R5 in men (21.6% vs. 18.5%) and half as much for R20 (5.9% vs. 12.8%) compared to our equations. Substantial differences also occurred for X5 in men (24.9% vs. 15.1%). In women, lower values were seen for R5 (5.3% vs. 16.0%), for R20 (1.4% vs. 8.1%) as well as for X5 (8.3% vs. 18.6%). With regard to the Vogel and Smidt equations, values for R5, R20 as well as X5 were much lower in both sexes compared to our equations. Thus, the number of subjects with IOS values beyond the limits of normal depended substantially on the reference equations used. # Discussion Based on a random population sample from Germany of Caucasian origin, reference equations for all IOS indices were determined for adults particularly covering the advanced age range. Care was taken to include only lung healthy subjects with no history of smoking or lung disease and spirometric values above the LLN. Subjects were fairly evenly distributed across the age range studied, with some emphasis on advanced age but thinning-out beyond 85 years. Based on the characteristics of our population we suggest that the reference equations derived for the median, ULN and LLN of IOS indices should be applied only in subjects aged between 45 to 85 years, with typical body weights (55 to 100 kg) and heights (160–190 cm in men and 145–170 cm in women). As previously reported, Rrs was higher in females than males, while the opposite was true for Xrs at low frequencies. Some of the IOS indices were age-dependent, but not as pronounced as spirometric indices. In line with physiological and morphological changes occurring in old age, age-dependent changes were most prominent for X5, AX, Fres, and R5–R20. All of these are suggested to be associated with peripheral lung function, although other mechanisms such as an airway shunt proximal to an elevated resistance or inhomogeneous constriction of peripheral airways may also hold true. In addition to the linear models we also tested more complex regression models, e.g. additive polynomial ones, with various functions of age, height, weight, or BMI. It turned out, however, that the linear models, with age, height and body weight as predictors of IOS indices, fitted best. This is in accordance with previous findings although other models for IOS indices have been reported in Japanese population. While body weight was a significant predictor for most IOS parameters in females, its relevance was less obvious in men, particularly for the 95% percentile. Respiratory mechanics is known to be affected by increased body weight, resulting in decreased FRC, expiratory reserve volume and compliance, as well as abnormalities of ventilation even with a modest increase in weight. This raises the risk of expiratory flow limitation and airway closure. Although our epidemiological setting does not provide evidence, particularly the latter may be involved in the functional link between body weight and various IOS indices in our study. Predictive equations for IOS indices of adults are scarce. Only two studies, by Vogel & Smidt and Newbury et al., have been published for Caucasians and one study for Japanese. There were marked differences between our and the published equations for Caucasians, which were more pronounced in men than women when quantified by the differences between observed and predicted values. The comparison of estimates in ficticious, average height and weight subjects underlined the differences in Rrs in both sexes, while values of Xrs appeared to be more comparable. Several reasons may account for these differences. Firstly, they could be related to the inclusion criteria of study populations. Our study was based on a randomly selected population sample, while Newbury et al. used purposive sampling of healthy individuals which could have introduced a sampling bias. On the other hand, it provided equal numbers for both sexes, which we did not achieve as females outnumbered males by 50%. Smokers were included in the population by Vogel & Smidt, and Newbury et al. included some past smokers, whereas we strictly excluded subjects with a smoking history. Moreover, the number of subjects varied between studies, the lowest being just above 100. Both published studies did not cover ages beyond 75 years, while we put emphasis on the middle and advanced age range. In all studies subjects were Caucasians but they originated from different countries, Australia vs. Germany (Vogel & Smidt and our study). As for spirometry, this could affect reference equations. It is also important that different statistical approaches have been employed. The published equations were based on standard linear regression models that rely on data properties which may be violated by IOS data, especially regarding normal distribution and homoscedasticity. Indeed, the Shapiro-Wilk test showed that most of resistance and reactance indices were not normally distributed. Therefore, we used quantile regression which is free of assumptions on the distribution of outcome measures and more robust against outliers or skewed distributions. The comparison of prediction equations thus requires some caution, since different statistical parameters are involved, e.g. mean vs. median. For instance, in case of the right skewed distribution of resistance indices the mean value is somewhat higher than the median. The generalized additive models for location, scale and shape, which have been recently applied to spirometric indices could have been an alternative to quantile regression. However, quantile regression appeared to be better suited with respect to the large but limited number of subjects included in our monocentric study. It also better allows for the comparison and interpretation of results. As a result of the possible explanations for differences in IOS predictive equations we would like to suggest that the present equations are the best to use in Caucasians, particularly for ages between 45 to 85 years. We also provide reference values for recently recognized IOS indices such as AX. Subjects of advanced age represent a population of great interest for IOS as the cooperation in spirometry is often limited. However, the clinical and predictive relevance of IOS in that age range remains to be investigated in clinical studies. The potential clinical implications of the use of different reference equations were studied by applying them to the entire KORA study population. The mean differences between observed values and predictions were considerably higher in both sexes for R20 and X5 when using the Vogel & Smidt equations. The percentages of subjects beyond the limits of normal were comparable between our and these predictions for R5 but much lower for R20 in both sexes, while small differences were found for X5 in woman. This demonstrates that with the reference equations currently applied by the IOS providing company evaluation of data results in systematic differences. ## Conclusions Our study provides up-to-date reference equations for all common indices of IOS in Caucasians of age 45 to 85 years. To facilitate the detection of functional alterations by IOS in this age range, we suggest to the use of the novel prediction equations. Still, further studies from Caucasians of advanced age are required to confirm our results and provide a more sound basis of IOS reference equations for clinical and epidemiological use. # Supporting Information We thank the KORA research platform (KORA, Cooperative Research in the Region of Augsburg) and all technicians involved in the study for conducting the study. Particularly we thank Kathrin Ernst for lung function measurements. We also thank Hans-Jürgen Smith, Care Fusion, for his valuable comments and critical reading of the manuscript. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: HS JB MH RH RMH RAJ DN AP HEW JH SK. Performed the experiments: HS CF RH AP HEW JH SK. Analyzed the data: HS CF RH AP HEW JH SK. Contributed reagents/materials/analysis tools: HS CF SK. Wrote the paper: HS CF SK. Critical revision of the manuscript: JB MH RH RMH RAJ DN AP HEW JH.

# Introduction Parkinson's disease (PD) is a neurodegenerative disorder mainly characterized by motor deficits generated by the decrease of dopaminergic neurons within the substantia nigra pars compacta (SNpc). PD patients endure severe sleep disturbances, such as excessive daytime sleepiness, rapid eye movement sleep (REM) behavior disorders, “sleep attacks” and cessation of dream recall while L-DOPA therapy has been found to enhance vivid dream recall. Recent studies show increasing evidence that the disturbance of sleep and wakefulness in PD is most intimately related to the disease itself and does not exclusively represent a secondary phenomenon. This primary effect of PD on sleep may be linked to the previously recognized, although only recently fully appreciated, role of dopamine (DA) in modulating sleep-wake states. PD is a disease characterized by impairment of the nigrostriatal pathway. This pathway is known to control and initiate motor plans, from the dorsal striatum to the motor cortex and back to the cortex via the thalamus. Patients with PD who have extensive loss of dopaminergic cells within the SNpc, and less so within the ventral tegmental area, often have increased sleepiness, which is made worse by the presence of dopaminergic D₂ receptor agonists. Such involvement of DA has been described subsequent to sleep deprivation as being directly involved in the generation of dopaminergic D₂ receptor supersensitivity. Moreover, recent findings demonstrate that partial DA depletion causes disturbances of REM sleep without affecting motor functions. Additionally, a robust increase in the electrophysiological activity of dopaminergic neurons of the ventral tegmental area has been identified during REM sleep. Along with the influence of dopaminergic neurotransmission on generating sleep disturbances, the evidence further supports the hypothesis that the dopaminergic nigrostriatal pathway, specifically, is fundamental in the regulation of sleep patterns. To test this hypothesis we examined the electrophysiological activity along the sleep-wake cycle of rats submitted to a surgically induced lesion of the SNpc induced by 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP). Bilateral guide canullas were implanted in the rats, allowing for the microinjection of the neurotoxin directly into the SNpc. Once microinjected, MPTP is converted, by monoamine oxidase-B in glial cells, to 1-methyl-4-phenylpyridinium (MPP⁺), which is taken up via the neuronal dopamine (DA) transporter and accumulates in dopaminergic neurons. MPP⁺ then becomes concentrated in mitochondria where it inhibits complex Ι reducing ATP generation and causing increased free- radical production and subsequent neuronal death. Intranigral MPTP administration produces a burst of dopaminergic neuronal death, and also a specific modulation in the tyrosine hydroxylase (TH) protein expression within the substantia nigra (SN). In view of these considerations we sought to determine the fluctuations of TH protein expression within the SN, concomitantly to the sleep-wake pattern, in order to examine the participation of this protein in mediating events that regulate the sleep-wake cycle. # Materials and Methods ## Subjects All experiments were conducted in accordance with National Institutes of Health (USA) guidelines for the care and use of animals and abided by an approved animal protocol from our university's ethical committee for animal experimentation (#0737/06). Male wistar rats weighing 280–320 g at the beginning of the experiments were used. Rats were housed individually in transparent acrylic cages and maintained in standard laboratory conditions (22±2°C, 12 h light/dark cycle, lights on 7:00 A.M.) with food and water provided *ad libitum*. Maximal efforts were employed to reduce the number of animals used in the experiments yet enough to ensure unambiguous and reliable statistical analysis and data interpretation. ## Stereotaxic surgery Rats were distributed at random into the two groups named sham and MPTP. Animals were anesthetized with diazepam (10 mg/kg i.p.) and ketamine (90 mg/kg i.p.) and they were mounted in a classical stereotaxic frame (Insight Instruments). Body temperature was maintained at 37°C with a regulated electric heating pad (Harvard Apparatus). Two bipolar electrodes with four stainless-steel screws (Ø 0.9 mm) were placed into the skull through small holes bored into the right lateral fronto-parietal (one pair) and in the left medial fronto-parietal (another pair) in order to monitor bipolar electroencephalogram (EEG). The free ends of the electrodes were soldered to a socket that was attached to the skull with acrylic dental cement. Two nickel-chromium flexible wires were inserted into the neck muscles in order to record the electromyogram (EMG). Additionally, the animals were implanted guide cannulae bilaterally (20 mm×0.6 mm) 2.0 mm above the SNpc according to the following coordinates: anteroposterior (AP): −5.0 mm from the bregma; mediolateral (ML): ±2.1 mm from the midline; dorsoventral (DV): −7.8 mm from the skull. All the rats received penicillin (20,000 U in 0.1 ml, i.m.) and sodium diclofenac (25 mg/ml, i.p.) after surgery. One week after surgery, the sockets were connected via flexible recording cables and a commutator to a polygraph and computer. After three days with the cables and the cannulae implants, the rats progressively habituated to the apparatus allowing the performance of the 48h basal sleep-wake states recording. ## Intranigral microinjections of MPTP After the basal sleep-wake recording of states the animals were manipulated and gently immobilized in order to perform the neurotoxin microinjection. Microinjections were performed during the early part of the light period (7:00 A.M. to 9:00 A.M). Each infusion was performed with a 30-gauge stainless injection needle bilaterally introduced in the guide cannulae through which 2 µL of MPTP (100 µg/µL, Sigma, prepared in sterile saline 0.9%,) were administered. The control of the flow of the microinjection was made by using an electronic pump (Insight Instruments) at a rate of 0.40 µL/min for 2.5 min, followed by 2 min with the needle in the injection site to avoid reflux. Sham microinjections followed the same procedure but using 2 µL of sterile saline 0.9%. Immediately after the microinjection procedure, the rats were placed in their home cages and the data acquisition initiated for a period up to 5 days (for details of the experimental protocol see). ## Sleep-wake state identification Electrophysiological signals were recorded on a digital polygraph (Neurofax QP 223A ©Nihon Kohden) at a sampling rate of 200 Hz. Filter settings of EEG were at 35 Hz and the time constant was 0.1 s. For the EMG the filter setting was at 70 Hz and the time constant was 0.03 s. EEG and EMG signals were calibrated at 50 µV pulses. In homeothermic animals, sleep was divided into two main distinct stages: slow wave sleep (SWS) and REM sleep. The former is characterized by high-voltage slow oscillations in the EEG associated with weak EMG activity. During REM sleep, activity in the forebrain is comparable to waking levels with pronounced and sustained theta rhythm in the hippocampus and a complete loss of muscle tone. Recordings of epochs were displayed at 15 s intervals on a high resolution PC monitor and visually classified as wakefulness (W), SWS and REM sleep. ## TH immunohistochemistry of the SNpc neurons For the immunohistochemical study of neuronal dopaminergic population within the SNpc, rats were deeply anesthetized with ketamine and intracardially perfused with saline and 4% of the fixative solution of formaldehyde in 0.1 M phosphate buffer (pH 7.4) during the period the sleep wake states were recorded, which amounted to five days. Brains were removed from the skulls and were immersed for 1 week in that fixative solution at 4°C. Subsequently, the brains were placed in 30% sucrose solution for 48h before sectioning. Series of 30 µm thick sections were cut on a cryostat in the frontal plane and collected at −0.48 mm to −0.56 mm from the bregma. Tissue sections were incubated with primary antibody anti- TH, raised in rabbits, diluted in PBS containing 0.3% Triton X-100 (1∶500; cat \#AB152 Chemicon) and reacted overnight at 4°C. Biotin conjugated secondary antibody incubation (1∶200 cat \#S-1000 Vector Laboratories), was performed for 2 h at room temperature. After several washes in PBS, antibody complex was localized using the ABC system (Vectastain ABC Elite kit cat \#PK6101, Vector Laboratories) followed by 3,3′-diaminobenzidine reaction with nickel enhancement. The sections were then mounted onto gelatin-coated slides and coverslipped after dehydration in ascending concentrations of ethanol-xylene solutions. ## Quantification of dopaminergic neurons Unbiased quantification of TH-labeled neurons from the SNpc relied on the software Freeware NIH Image, 1.63. Counts were done on 8–10 tissue sections (one in four series), and an average count per section was determined for each animal. The selected areas were digitized with a digital camera DP71 (Olympus Optical) using an Olympus microscope BX50. ## Determination of TH protein expression To determine TH expression a different set of animals underwent the same protocol of stereotaxic surgery, cannulae implantation, and lesioning with MPTP microinjection. To investigate the TH expression fluctuations, generated by MPTP among the sleep-wake cycle, 10 rats (5 of sham group and 5 of MPTP group) were killed daily, by decapitation, at the same time of the MPTP microinjection. After the decapitation the brains were quickly dissected with the isolation of the SN which was immediately frozen in dry ice and stored at −80°C until the lysis procedure. The lysis were performed in 1.5 mL eppendorf tubes by sonication in the presence of an ice-cold buffer containing 50 mM Tris (pH 8.0), 250 mM NaCl, 1% NP-40, 0.1% sodium dodecyl sulfate (SDS), 0.25% sodium deoxycholate, 2 mM EDTA, 1mM dithiothreitol (DTT), 20 µM phenylmethylsulfonyl fluoride (PMSF), and protease inhibitors (Complete tablet; Roche). After incubation on ice for 30 min extracts were centrifuged at 12,000× g for 40 min at 4°C, and the supernatants for protein extracts were collected and stored at −80°C for further western blotting analysis. The aliquot of supernatant was collected for total protein analysis. Samples containing equal amounts of total protein (5 µg per lane) were boiled with SDS sample buffer and electrophresed in 10% SDS-polyacrylamide gels in a Mini Protean II Dual Slab Cell (Bio-Rad). Proteins were electrophoretically transferred to nitrocellulose membranes by means of a Mini transblot electrophoretic transfer cell (Bio-Rad). Each membrane was blocked for 1 h in 10% nonfat dry milk/0.5% Tween-20 in Tris-buffered saline. Subsequently each membrane was probed overnight at 4°C with mouse monoclonal antibodies against TH (1∶5,000; cat \#T2928 Sigma) or β-tubulin ΙΙΙ (1∶500; cat \#MAB1637 Chemicon) followed by several washes in TBST and incubation with an adequate secondary horseradish peroxidase-conjugated antibody (1∶5,000; cat \#30021019 GE) for 60 min, and visualized by chemiluminescence (cat \#sc 2048 Santa Cruz Biotechnology). The bands were quantified by using the software ImageJ 1.32j. ## Statistical methods Differences in number of cell counts and TH protein expression data underwent analysis of variance (ANOVA) followed by the Newman-Keuls test. Sleep parameters were analyzed by ANOVA and the Tukey test was used as *post hoc* when indicated. Pearson correlation coefficients were calculated for comparison of loss of dopaminergic neurons or TH protein expression with alterations in sleep parameters. Differences were considered significant if p\<0.05. The values were expressed as mean±S.E.M. # Results ## Characterization of dopaminergic neuronal loss Representative photomicrographs of TH immunohistochemistry in the SNpc are shown for both sham and MPTP groups. The TH-ir neurons were prominently detectable in the SNpc of sham rats. The bodies and fibers of dopaminergic neurons showed intense staining with evident immunopositive processes in the sham group. The mean number of TH-ir neurons within the SNpc (sham, 9,750±345 vs MPTP, 4,873±176) indicated that MPTP inflicted an expressive dopaminergic neuronal loss of 50% \[F_(10.47) = 17.25, p\<0.01\] restricted to the SNpc, with no detectable damages in the ventral tegmental area. ## Effects of dopaminergic cell loss in the SNpc upon sleep-wake cycles The data indicated that a 50% dopaminergic neuronal loss restricted to the SNpc, inflicted by MPTP, was able to produce a robust and constant decrease in the latency to the onset of SWS during the 5 days of recording in both light \[F_(22.16) = 72.46, p\<0.0001\] and dark \[F_(22.16) = 75.0, p\<0.0001\] periods. In addition, this nigral dopaminergic depletion also increased the latency to REM sleep during the light \[F_(22.16) = 14.73, p\<0.001\] and dark \[F_(22.16) = 14.70, p\<0.0001\] periods 1 day after neurotoxin microinjection. MPTP-induced dopaminergic neuronal loss also generated pronounced increase in the percentage of sleep efficiency during the first 4 of the 5 days of recording \[F_(21.15) = 21.48, p\<0.0001\] in comparison to the sham group. The percentage of SWS was increased in the MPTP group on days 2 and 3 only in the dark period \[F_(21.15) = 4.54, p\<0.0001\] compared to the sham group. In contrast, the reduction in the dopaminergic neuronal population, resident in the SNpc provoked an ablation in the percentage of REM sleep during days 1 (for light and dark periods), 2 and 3 (for the light period only). Moreover, there was a strong correlation between the number of TH- ir neurons lost and the percentage decrease of REM sleep on the first day of recording after MPTP lesion (r = 0.91; p\<0.0001). On day 4, REM sleep presented an increase in both periods \[F_(22.16) = 2.46, p\<0.0007\]. All the sleep parameters return to basal values at day 5 after the MPTP exposure, with the exception of latency to SWS. This observation indicated the end point of the examination paradigm. ## Fluctuations in the nigral TH protein expression after MPTP microinjection A single band was observed at the expected molecular mass of 60 kDa for the groups tested at the distinct time-points examined. Denser TH-positive bands were consistently observed in the sham group. There was a significant decrease of 50.1% of TH expression in the SN of the MPTP group in comparison to the sham group \[F_(9.99) = 2.95, p\<0.004\] on the first day after the neurotoxin microinjection. This dramatic decrease in TH expression at day 1 was not seen in later time-points. Two days after the microinjection TH expression in the MPTP group was 98% in comparison to the sham group of the respective day. A similar situation was observed for the following time-points: 3 (p = 0.07, 86%), 4 (p = 0.06, 82%) and 5 (p = 0.11, 100%) days after the microinjections. Weak correlations were observed between TH protein expression and sleep parameters (sleep efficiency r = −0.64, p = 0.042; SWS r = −0.30, p = 0.39; REM r = 0.20, p = 0.57) along the time-points. # Discussion The main findings herein are that the dopaminergic neurons located within the SNpc played an important role in regulating sleep patterns in rats, and that disturbances in this particular neuronal population produced severe complications in all the sleep parameters examined, especially in REM sleep. To the extent of our knowledge this is the first paper that provides direct evidence of SNpc function in sleep. The impairment in REM sleep presented a transient characteristic which culminated a few days after the nigral disruption with a strong increase in the percentage of REM sleep, during the light and dark periods. The present findings dovetail with growing evidence showing the influence of DA in the regulation of sleep-wake pattern and the significant repercussions that eventual reductions of this neurotransmitter may generate, particularly in PD patients. A recent experiment has shown a coordinated bursting mode of activity of dopaminergic neurons, located in the ventral tegmental area, initiated 10–20 s before the onset of REM sleep and lasting until the end of this sleep stage. It has been proposed that during the sleep-wake cycle the temporal pattern of firing rates alters in midbrain dopaminergic cells. The available evidence tends to indicate that during wakefulness there occurs an increase of burst firing activity of dopaminergic neurons, and enhanced release of DA in the nucleus accumbens, and a number of forebrain structures. Hence on the basis of our findings, different features of the dopaminergic neurons are revealed, considering that they are not exclusively related to wakefulness as initially postulated, according to previous lesion studies. Nevertheless, a large range of different cell populations are damaged with the systemic administration of a neurotoxin or even by electrolytic protocols. In this sense, the present study promoted a dopaminergic lesion compatible with the extension of the area that was examined, i.e. the SNpc. We adopted the intranigral microinjection of MPTP as a more reliable and accurate choice to produce only specific dopaminergic neuronal death. This protocol resorts to precise delivery of MPTP to the SNpc, relying on its specificity to the dopaminergic transporter. The TH immunohistochemistry examination revealed that MPTP produced a reduction of 50% in the dopaminergic neurons in the SNpc. In addition, neuronal morphology was apparently preserved in the remaining neurons of the MPTP group, despite the remarkable apoptosis inflicted by MPTP. At this point, a note of caution should be added: MPTP microinjection, in our experience, can achieve a plateau of dopaminergic neuronal loss of around 50%, without affecting adjacent areas. In this sense, our hypothesis is limited exclusively to the influence of SNpc neurons in the patterns of the sleep-wake cycle. To ensure an unbiased lesion protocol, we preferred not to produce a more extensive lesion, guaranteeing that the physiological effects occurred only as consequence of SNpc lesion. The electrophysiological data indicated that the absence of half of the dopaminergic neurons within the SNpc provoked a major impairment in the sleep- wake parameters. Some manifestations, as indicated by the decrease in the latency to SWS, initiated almost immediately after the neurotoxin microinjection, corroborating a previous study with systemic administration of MPTP in cats, and those alterations were continuously identified until the end of the experimentation. In relation to these findings, it is noteworthy to consider that the impairments of these parameters could be characterized as an overwhelming episode of sleep, similar to “sleep attacks” which is a manifestation experienced by 1–21% of PD patients. We also demonstrated that latency to REM sleep presented a punctual augment on the very first day after the nigral lesion, with a return to baseline values on the subsequent days of recording. This finding is strengthened by the dramatic reduction in the percentage of REM sleep observed in the first day after MPTP, which is highly correlated to the number of dopaminergic TH-ir neurons lost within the SNpc. Otherwise, TH protein expression in the SN did not present significant correlation to any recorded sleep parameter. Such evidence indicates that sleep- wake patterns are more closely related to function of nigral dopaminergic neurons than exclusively to variations of TH expression. An analogous pattern of deficit in REM sleep was observed the following days, with some level of fluctuation in the percentage of REM sleep although on the 4^th day of recording, REM sleep robustly increased in both light and dark periods. Such unexpected result suggests the activation of a compensatory mechanism which led to an irrefutable augmentation in the percentage of REM sleep. It is not unreasonable to suggest that this phenomenon can be considering a rebound of REM sleep. The dopaminergic nigrostriatal lesion promoted a significant reduction in the TH protein expression, and suggestively in DA biosynthesis. Nevertheless, 50% of the dopaminergic nigral neurons survived generating a potential plastic response to this neurotoxic assault. This response is possibly mediated by the TH protein up-regulation and also by an important post-synaptic event defined as D₂ dopaminergic supersensitivity, which is unleashed mainly by reduction of DA in the synaptic cleft. Considering the percentage of SWS we observed an increase during the 2^nd and 3^rd days of recording followed by a return to baseline values after those time-points. Notably, the reduction of the dopaminergic nigral population generated a pronounced potentiation of SWS to the detriment of REM sleep. In addition, sleep efficiency was found to be augmented indicating increased sleepiness during the four first days in the lesioned rats. We proposed that dopaminergic neurons present in the SNpc possess a fundamental role in the regulation of sleep processes, particularly in promoting REM sleep. It also should be mentioned that flanking areas of the SNpc, like the ventral tegmental area, have recently been reported to be involved in the regulation, and perhaps genesis, of REM sleep as we demonstrated for the SNpc. Besides, clinical evidence demonstrate a transient restoration of motor control in PD patients during REM sleep, representing, in the light of our study, an intersection between these two physiological functions, with the nigrostriatal pathway playing a key role. In conclusion, we reported herein that dopaminergic neurons resident in the SNpc possess a central role in the regulation of sleep processes, particularly of REM sleep. This evidence directly demonstrates that the SNpc is an integrative area of biological function, well known to be responsible for voluntary motor control, and herein described as an important center of sleep regulation. The authors would like to thank Adriano Zager, Waldermaks Aires Leite and Marilde Aires Costa for the capable technical assistance. [^1]: Conceived and designed the experiments: ML MA AR MV ST. Performed the experiments: ML AR. Analyzed the data: ML AR. Contributed reagents/materials/analysis tools: ML MA ST. Wrote the paper: ML MA. [^2]: The authors have declared that no competing interests exist.

# Introduction Avian influenza viruses (AIV) infect animals and humans and thus are of major relevance under the one health context. They are enveloped viruses belonging to Alphainfluenzavirus from *Orthomyxoviridae* family. These negative sense and single-stranded RNA viruses present a genome that is fragmented into eight segments which encode for eleven proteins. These proteins possess individual function or form complexes during viral replication. The mechanisms underlying AIVs evolution, comprising antigenic shift and drift, point mutations, reassortments and others, are induced mainly by their segmented genome and the receptor specificity, including a wide host range into the animal kingdom. Hemagglutinin (HA) and neuraminidase (NA) are structural proteins responsible for attachment and viral release, respectively. These proteins play a fundamental role in AIV replication. In addition, HA and NA have also been used for viral classification into subtypes. Thus, eighteen HAs and eleven NAs have been described to date. From these, sixteen types of HA (H1-16) and nine of NA (N1-9) were identified in avian species, including wild and migratory birds from *Anseriformes* and *Charadriiformes* orders. Other two subtypes, H17N10 and H18N11, have been identified in bats from Guatemala and Peru. Most AIV subtypes have been classified as low pathogenicity avian influenza viruses (LPAIV) and generally induce minimal impact in the health of domestic and wild birds, although there have been reports of sporadic cases in humans related to H5, H6, H7, H9 and H10 avian subtypes. Animal reservoir carrying these viruses spread AIV on the ecosystem, facilitating an effective transmission between vulnerable hosts. Furthermore, genetic changes in relevant proteins, such as HA and NA, can turn them into highly pathogenic avian influenza viruses (HPAIV) with high consequences for other species. Infections by HPAIV have a major social and economic impact in poultry because it can result in massive deaths in the farms, and has a potential for causing disease in humans. In addition, global population expansion has provided new opportunities for spillover events. Altogether, active epidemiological surveillance under a one health approach is essential to develop novel strategies that will allow us to implement measures in a concerted manner before epidemics begins. Hence, it is essential to get a better understanding of viral circulation patterns in our ecosystem, providing fundamental information for AIV risk assessment and detecting probable hotspots of novel AIV emergence. Waterfowl has been considered the main natural reservoir for all AIV subtypes representing an important source for transmission to other species. Moreover, Peru has a central location within the migration route of wild birds in Americas, and therefore is of relevance for viral spreading. Hence, our major objective was the isolation and genetic subtyping and characterization of the AIV circulating in wild birds in Peru, using the HA and NA genetic information, and establishing relationships to other isolates previously detected in America. # Materials and methods ## Experimental design and samples A descriptive, qualitative, cross-sectional study of active epidemiological surveillance for avian influenza in animal reservoir in the Coast of Lima (Peru) was conducted in the period of March 2019 and March 2020. Sampling was carried out targeting different population of waterfowl species belonging to *Charadriiformes* and *Anseriformes* orders. This included several residents and migratory waterfowl in Peruvian territory. Fecal samples were collected from the ground using swabs and Universal Transport Media (UTM) (Copan Diagnostics Inc., USA) in cool chain boxes. Samples were taken in 5 natural resting areas for waterfowl in Lima (Puerto Viejo 76.7026578°W 12.5714006°S, Pantanos de Villa 76.9949416°W 12.2065568°S, Albufera de Medio Mundo 77.6626868°W 10.9353162°S, Laguna La Encantada 77.5522327°W 11.1364147°S and Poza de La Arenilla 77.1590185°W 12.0721548°S). Avian species were determined under close visual observation and photographic record, and classified according to phenotypic characteristics. Thus, fecal samples were taken only when the avian species could be determined. Samples were transported to the Avian Pathology Laboratory of the School of Veterinary Medicine, Universidad Nacional Mayor de San Marcos (Lima). Fecal samples were processed and clarified by centrifugation (8000 x g for 10min) and supernatants were filtered through 0.22-micron syringe filters. Filtered samples were then pooled according to species, collection date and geographical location into 4 to 5 samples per pool, supplemented with 1000 U/mL Penicillin G, 1 mg/mL Streptomycin and 10 μg/mL amphotericin B, and incubated at 4°C for 1 hour prior to inoculation into embryonated eggs. ## Virus isolation in embryonated chicken eggs Viral isolation was performed by inoculation of 0.2 mL of pooled samples in ten- day-old specific-pathogen-free (SPF) embryonated eggs through allantoic sac route in five replicates. This process was followed by incubation at 37°C for 120 hours. Embryo survival was checked daily. Allantoic fluid from each sample was tested by hemagglutination assay as initial screening. Positive samples were titrated using the same technique by serially diluting with two-fold volume of PBS (pH 7.4) and then incubating with equal volume of 1% chicken red blood cells (RBC). After positivity evidence, the positive pool had their samples evaluated individually. ## Viral detection Rapid immunochromatographic assay (QuickVue Influenza A+B Test, US) was used for rapid identification of Influenza A virus in allantoic fluid samples that tested positive for hemagglutination test. For confirmation, we used a PCR-based method to amplify the NP gene from the viral genomic material. Thus, the viral RNA was extracted from 140 μL of allantoic fluid using QIAamp Viral RNA Mini Kit (Qiagen, US), according to manufacturer instructions. This was followed by a step of reverse transcription using GoScript™ Reverse Transcription System kit (Promega, USA). In this step, we used the primer described by Hoffman et al. (Uni-12 `5’-AGCAAAAGCAGG-3’`) that allows amplification of all influenza A segments. NP gene was detected by PCR using GoTaq® G2 Green Master Mix (Promega, USA) and primers described by Wright et al. (1995) Anp-1 (`5’- ATCACTCACTGAGTGACATC-3’`) and Anp-2 (`5’-CCTCCAGTTTTCTTAGGATC-3’`) for 306-bp. The thermal cycling conditions were 94°C for 3 minutes; 35 cycles of 94°C for 30s, 58°C for 30s and 72°C for 60s; and 72°C for 5 minutes. PCR products were electrophoresed on 1.5% agarose gels and visualized by ethidium bromide staining. ## Next generation sequencing (NGS) First strand cDNA was prepared using SuperScript™ III First-Strand Synthesis kit (Invitrogen TM), following manufacturer recommendations and using the primer FR20RV (`5’-GCCGGAGCTCTGCAGATATC-3’`) prior to library preparation. The libraries were prepared using Nextera XT method and sequenced on a MiSeq platform using paired 150 base read chemistry (Cambridge Technologies, Worthington, MN, USA). ## Assembly and annotation The raw data was submitted for end trimming using *Trimmomatic*, a flexible tool for Illumina NGS data and quality control evaluation based on Phred quality scores. Reads that presented a high Phred score were submitted for *de novo* genome assembly (DISCOVAR). Segments that were not properly assembled were submitted to a new round using PRICE. The annotation was performed using “Influenza Virus Resource” tool, from the “National Center for Biotechnology Information”–NCBI, available at <https://www.ncbi.nlm.nih.gov/genomes/FLU/annotation/>. After annotation, the segments were edited manually. ## Sanger’s sequencing For some isolates, complete nucleotide sequences of the 5’-end and 3’-end non- coding and coding regions of few gene segments were not obtained by NGS. Therefore, primer pairs were designed to amplify the non-coding and coding sequences of these gene segments by RT-PCR, as described earlier. The resulting amplicons were purified and DNA was directly sequenced or cloned into a pGEM-T vector, and the plasmid DNA was sequenced using Sanger’s dideoxy chain termination method. (Institute of Marine and Environmental Technology, Baltimore, MD, USA). Thus, complete nucleotide sequences for the whole genome of Peruvians AIVs were attained, namely PE-01 to PE-5, and the data was deposited into the GenBank. ## NA and HA subtyping and phylogenetic analysis Following sequencing, we focused on the subtyping of AIV isolated based on the HA and NA genetic analysis. Each HA and NA segment assembled was submitted to genetic characterization using MEGA. Criteria for databank construction considered the evolutionary pattern of AIV in the ecosystem (host, space and time). The coding region of HA and NA from all AIV subtypes from America available in the webpage of National Center for Biotechnology Information (NCBI), Influenza Virus Genome Set (<https://www.ncbi.nlm.nih.gov/genomes/FLU/Database/nph- select.cgi?go=genomeset>) were included and sequences that presented degenerated nucleotide were excluded. The alignment was conducted using MAFFT online service (<https://mafft.cbrc.jp/alignment/server/>) according to default parameters and the aligned sequences were posteriorly manually edited using BioEdit 7.2 software (<https://bioedit.software.informer.com/7.2/>). Phylogenetic analysis was performed using MEGA X (<https://academic.oup.com/mbe/article/35/6/1547/4990887>) determining the best evolution model by using the jModelTest2 software; clustering was obtained using the maximum-likelihood (ML) method under General Time Reversible gamma distribution and proportion of invariable sites parameters and 1,000 replicates of bootstrap. # Results To survey the presence of AIVs in wild birds from Peru, 421 fecal samples were inoculated in SPF chicken embryos and tested by hemagglutination assay (HA), which was performed by mixing equal volume of allantoic fluid and chicken red blood cells (1:1 dilution). Out of these, eight samples tested positive by HA. Further screening of all eight samples by immuno-chromatography and by RT-PCR amplification of the NP gene evidenced only five positive samples of AIVs. From five AIV samples collected from the wetlands; two were from Poza de la Arenilla, two from Pantanos de Villa and one from La Encantada, all located in Lima province. Two positive samples were isolated from two franklin gulls in Poza de la Arenilla; one from a kelp gull and one from a whimbrel in Pantanos de Villa; and one from a mallard in La Encantada. A summary of these data, including general information about species and location is presented in. Following detection of these AIV isolates, we performed the whole genome sequencing (WGS), which allowed us to characterize and identify close genetic relatives for each viral segment. Here, we focused on the annotation of HA and NA which reveal that our PE isolates correspond to the following subtypes: H6N8 (PE-01), H2N6 (PE-02), H6N2 (PE-03), H13N6 (PE-04) and H6N2 (PE-05) commonly found in wild birds. The information related to accession numbers and the closest relatives for each gene segments is presented in. An amino acid analysis was conducted and the list of amino acid substitutions for each HA and NA segments, including additional information is presented in and, respectively. Using HA sequences, we deployed multiple phylogenetic trees, including those HA genetic information from virus subtypes that have been circulating in America. Using Mega 11, we created subtree versions of the complete phylogenetic analysis, which are shown in Figs (H2), (H6) and (H13). Thus, H2 (PE-02) has emerged within a north American clade including isolates from 2011–2015. All H6 genes (PE-01, PE-03 and PE-05) share a common ancestor and formed a sister group with another monophyletic clade formed by three isolates from US identified in 2015 (ALX30476, H6N1; ALX30488, H6N1 and ALX30464, H6N2). The H13 (PE-04) sequence was more similar with a mixed strain from USA/2015 and a strain H13N8 identified in the US in 2017 (US/Minnesota/2017) with a 99.47% of amino acid identity. The phylogenetic tree showed a monophyletic group between our isolate with other two isolates from US identified in 2015, a mixed strain APH09815 and a strain H13N8 APY16670. Interestingly, we found multiple amino acid changes compared to their closest relatives. Thus, H2 (PE-02) has multiple changes which had the larger number of changes among our isolates and were also evidenced in the phylogenetic tree since that branch was more dissimilar than others in that group. Other HAs have also shown some amino acid changes but in lower proportions. In contrast, one of our H6N2 isolates (PE-03) did not evidence changes compared to its closest relative. Based on NA sequence analysis, we also deployed phylogenetic trees including that NA genetic information from virus subtypes that have been circulating in America. Similar to HA analysis, Figs (N2), (N6) and (N8) show the phylogenetic relationship among NA genes reported to date, presented as subtree versions of complete phylogenetic trees. Hence, we identified that both N2 from our two H6N2 isolates arose from two divergent ancestors. N2 (PE-05) presented the highest similarity with a strain H5N2 (US/Arkansas/2015), with 99.56% of amino acid identity, while N2 (PE-03), clustered with three different strains from Chile in 2016, presenting 98.69% amino acid identity with the closest relative. Furthermore, the phylogenetic tree showed that the both N2, isolated in the same site from Peru (77.1595302°W, 12.0717358°S), from the same wild bird species (Franklin’s gull), sampled one week apart (13^th and 19^th of December, 2019), grouped in two different clades. For N2 (PE-05), it grouped with all US isolates. The other strain (PE-03) exhibited the closest relationship with those identified in Chile and Argentina. In terms of N6, the one from PE-02 had the highest similarity with a strain from Chile in 2017 (Chile/2017, H7N6), whereas, the N6 sequence from our H13N6 isolate was closely related to the US isolate (US/Minnesota/2017/H13N6) with 99.57% amino acid identity. The phylogenetic tree presented the topology with the same characteristics, showing that Peruvian H13 strain formed a monophyletic group with Chilean strains identified in 2013 (H3N6, ANH21898 and H7N6, ANH21875) and the other, with strains from US also identified in 2013 (four H13N6 isolates, AIN76265, APY16561, ALZ46774, ALZ46750 and a mixed N6, APH09819). Moreover, N8 from PE-04 had more similarity with the US isolate (US/New Jersey/2017/H5N8) having 99.57% amino acid identity. # Discussion To conduct surveillance of avian influenza A virus (AIV) circulating among the wild birds in Peru, we collected and tested fecal samples from several species of birds along the coast of Peru. We isolated five AIVs and their whole genome was sequenced. Genetic analysis of HA and NA genes revealed that these viruses are of subtypes H6N8, H13N6, H6N2 and H2N6. Our results also demonstrated a high HA and NA variability in our isolates, as compared to others previously identified, indicating the active viral circulation within the populations assessed. Despite the multiple orders evaluated, *Anseriformes* and *Charadriiformes* orders are the main natural reservoir of AIV in Peru. According to Rejmanek et al. (2015), H6N2 and H6N8 have been frequently isolated from mallard and goose (*Anseriformes*), while H13 and H16 have been exclusively isolated from gulls (*Charadriiformes*). Our findings show that this pattern varies, which might explain in part, the higher variability of HA and NA observed. Further studies are required to assess the implications of host variation on viral evolution. Nevertheless, all isolates found in the study belonged to subtypes commonly known as low pathogenic avian influenza viruses (LPAIV). The HA protein present 10⁻⁴ amino acid substitutions / site / year under action of purifying selection, mainly driven by the host immune responses. This suggests that viral genetic structure tends to be determined by “space”, “territoriality” under a sympatric speciation model, indicating that evolutionary events within the same migration route are more likely to occur. In our study, sites where we detected those AIV isolates are common geographical sites for migratory birds feeding and resting within their migratory route, considering that these places create potential sites for emergence of novel strains. On the other hand, according to HA specificity in the host cell receptor, each AIV subtype has a group of susceptible animals to infections, and in multiple cases, include humans in the center of this circulation pattern. In this context, public policy actions must consider a one-health dynamics, in which all factors must be considered, from factors such as environmental preservation, animal and human health, since only this perspective will have the possibility of success in regional and global epidemiological surveillance for emergency containment of highly pathogenic strains. Peru is a biodiverse country for multiple avian species and a special location to identify new strains of AIV since it is located right at the crossroads of Pacific and Mississippi migratory routes. Our study identified five different isolates, including four subtypes obtained from four different migratory bird species. Our isolates presented hemagglutinin whose subtypes originated from US (H2, H6 and H13 presented closest relationship with isolates identified in US and Canada); while the neuraminidase of our isolates originated from Chile (N2), Chile and US (N6). Since our isolates were obtained from migratory birds, these findings suggest that Peru, with its geographic position is an important site for emergence of novel reassortants of AIV. A schematic representation of the genetic flow across Americas involved in the evolution of AIV in Peru is shown in. The H2 identified in a mallard presented a special attention since it could be transmitted to humans directly or through domestic poultry infections and possibly associated to cause the next flu pandemic. However, little is known in terms of viral interaction with its natural host. AIV H2 produced a severe pandemic in 1957 and was reported circulating in North America in the end of 90’s. On the other hand, the H6 was identified in two different species, in a whimbrel and in a Franklins’ gull. Interestingly, domestic poultry has been reported to be extremely susceptible to this subtype, representing an economic and social threat. H6 segment has been intensely identified in mallard from Canada and presents global distribution. Moreover, H6N2 and H6N8 subtypes have been reported in several parts of America for more than forty years. For instance, H6N2 was identified in USA (1986–2015), in Mexico (2007) and Argentina (2010) and Chile (2016–2017). On the other hand, H6N8 has been circulating in America between 1977–2016 in USA, detected in Argentina and Peru in 2021; and in Chile, in 2016 and 2017. Thus, H6N8 is described as a high-priority subtype. Similar to H2 subtype, H6 has been used experimentally in ferrets to produce vaccine candidates. Finally, the H13 identified in a kelp gull can be spread to domestic poultry and other mammals. Interestingly, H2 and H13 have shown global distribution, while H2 subtype has a closest phylogenetic relationship with H5, indicating a more recent differentiation between these subtypes. This close phylogenetic relationship between H2 and H5 might induce events for viral reassortments. In regard to NA, N2 subtype isolated from two Franklin’s gulls has shown to infect swine, marine mammals, dogs, poultry, and humans. For instance, H9N2 subtype has presented high prevalence in domestic and wild birds and has been isolated from swine and humans, which has strong ability for inter-species transmission. N6 has also been described as a high-priority subtype and was identified in two different strains in our study, one obtained from an *Anseriformes* (mallard) and other from a *Charadriiformes* (kelp gull). This N6 subtype needs special attention because in addition to infecting aquatic mammals, it can infect swine and thus suitable for genetic changes. Finally, N8 subtype has been widely identified in America, sometimes in association with the subtype H6, as we found in our study, and is considered as a high-priority subtype. Some evidence highlights that strains identified in the same flyway are more similar between them than those found in others flyway during a period. The America continent presents three different flyways: the Pacific, the Mississippi and the Atlantic. In our study, samples were obtained from animals from the Pacific flyway. Further studies are required to assess the viral circulation in flyways crossing other areas in South America. Hence, considering the spatial evolution, AIV strains seem to be spatially structured. Nevertheless, in multiples places, the overlapping of the flyways facilitates the genic flow and the possibility to generate novel strains. Additionally, in the northern region such as Canada and Alaska, there is a confluence of multiple flyways from several parts around the world and thus AIV strains do not show spatial genetic structure. Finally, our results indicate that multiple AIV subtypes are circulating in wild birds in Peru. Even though those LPAIV subtypes have been previously detected, this work has revealed that these isolates possess unique features at the nucleotide and protein level. Although we have not defined what the implications of these amino acid changes represent on HA and NA function, our results provide novel insights into AIV evolution for genetic surveillance. # Conclusion Our study shows the active circulation of multiple AIV subtypes in wild birds in Peru. Despite their close genetic relationship with other common subtypes previously found in America, our findings have identified that these isolates have gained multiple unique changes during the viral evolution in America. Altogether, this work has shed the light on the importance of Peru, which confers ecological advantages as relevant geographic location for the emergence of novel AIV in America with pandemic risk. # Supporting information We would like to acknowledge Dr. Hermelinda Rivera Geronimo, Dr. Mercy G. Ramirez-Velasquez, Dr. Sonia Y. Calle-Espinoza and Cusi Flores, from Laboratory of Microbiology and Parasitology, Faculty of Veterinary Medicine, Universidad Nacional Mayor de San Marcos for professional assistance and recommendations, as well as Dr. Claudia Carranza from Laboratory of Avian Pathology, Faculty of Veterinary Medicine, Universidad Nacional Mayor de San Marcos and Dr. Manolo Fernández from Farmacologicos Veterinarios S.A.C. (FARVET SAC) for help with logistic management. [^1]: The authors have declared that no competing interests exist.

# Introduction Following injury the adult mammalian spinal cord shows little capacity for functional repair involving regeneration of axotomised neurites. In spite of many years of intensive research to promote such regenerative growth at the site of injury, no new therapies for patients have resulted and attempts to replicate apparently promising studies have so far failed. In recent years it has become apparent that a more effective approach may be to build on much older studies of Sherrington and others indicating that even when isolated from the brain, the adult spinal cord retains the ability to produce rhythmic activity of the limbs when they are stimulated by sensory input. However, such rhythmic activity is generally not accompanied by an ability to support body weight, except in some animal models after intensive training (e.g., spinalised cats, ; see for review of different species). Extensive studies have given rise to the notion of a central pattern generator (CPG) in the lumbar spinal cord with the intrinsic ability to generate rhythmic movements of the hindlimbs. Further studies have suggested that in spinal animals, these CPGs may be able to be activated when there is peripheral sensory input via the hindlimbs. Systematic attempts to use this information in patients consisted of clinical trials in spinal cord injured individuals whose body weight was supported in a harness over a treadmill. The results, however, did not show any consistent improvement in unsupported body weight locomotion compared to overground walking therapy. In both animal and patient studies body weight supported treadmill training has been extended by use of a feedback robot, but in patients a recent systematic review concluded that the evidence for efficacy was lacking because of inadequate experimental design. A few robotic systems have also been developed for animal studies. However, it seems that the mechanisms involved in rhythmic limb movements originating from central pattern generators and the mechanisms involved in body weight support are distinct. Thus loss of supraspinal input in spinal cord injury results in loss of body weight support that, except in particular specific circumstances, does not recover following treadmill training with or without robotic assistance. The addition of electrical stimulation of the spinal cord together with administration of drugs (e.g., agonists for serotonin, dopamine and noradrenalin receptors) appears to be required to raise the state of excitability of the propriospinal neural circuits that are responsible for weight bearing in a normal individual. These approaches are quite complex and time consuming for the patients and it has yet to be shown how applicable they will be in the wide range of injuries sustained following trauma to the spinal cord. As an alternative approach we have followed up on our recent report on the ability of the Grey short-tailed opossum (*Monodelphis domestica*) to exhibit weight bearing locomotion when injured in the postnatal period, an ability that is possible even when there is no evidence of axonal growth across the site of a complete spinal transection made at 4 weeks of age and without any interventions like treadmill training, drugs or electrical stimulation. Lesions made earlier (one week of age) were followed by developmental and regenerative axon growth across the lesion with subsequent development of essentially normal locomotor function. This has also been demonstrated *in vitro* in preparations of cultured *Monodelphis* spinal cord, where axon growth across a spinal lesion occurred following injury in the neonatal but not in two-week-old spinal cords. Similar findings of axon growth across a lesion, providing it had been made early in development, have been reported in North American opossums, *Didelphis virginiana*. Marsupial species such as *Monodelphis* and *Didelphis* are particularly favourable for studies of the response of the developing spinal cord to injury because the animals are born at a very early stage of development. A response involving axon growth across a spinal cord injury in eutherian mammals such as rodents has only been described in fetuses. Rodents with spinal cord injuries made in the first week of life do show some limited locomotion with only variable body weight support, but the recovery is also not associated with axon growth across the lesion. In the present study we have focused on the development of weight bearing locomotion in *Monodelphis* transected at two different stages of spinal cord maturity: at an age (4 weeks) when no axon growth occurs across the lesion (more akin to the clinical situation in humans with severe spinal cord injuries) and compared these spinal animals with those transected at an even earlier stage of development (one week) when supraspinal innervation is restored. For reasons of experimental consistency we have used a transection injury rather than a contusion, which would have been more like the clinical situation The ability to walk with weight bearing steps in spite of a lack of axonal connections across a spinal cord lesion is an interesting fundamental property of the spinal cord at a particular stage of its development. Understanding the biological basis for this ability might eventually lead to development of less complex interventions in spinal cord injured patients than are currently being contemplated. # Methods ## Ethics Statement The University of Melbourne Animal Ethics Committee approved all animal experiments described in this study, under ethics no. 111198. All experiments were conducted following National Health & Medical Research Council (NH&MRC) guidelines. ## Animal Care South American Grey short-tailed opossums, *Monodelphis domestica,* were supplied from a breeding colony maintained by the Biological Research Facility (BRF) at the University of Melbourne. A comprehensive description of these animals’ husbandry has been published previously, , and is only outlined briefly here. Opossums are housed in polycarbonate boxes in temperature- and humidity- controlled (27°C; 60% humidity) rooms in the Animal House facility with a 14∶10 hour inverted light/dark cycle. Food and water are given *ad libitum.* Following a gestational period of 13 days mothers give birth to litters of 5–8 pups, which remain attached to the maternal teats until approximately post-natal day (P) 14, after which they start to detach for increasing time intervals. Pups are weaned at P60–65 and begin to reach sexual maturity at about 5 months of age. ## Numbers of Animals used in this Study In total 27 animals were used in this study. Control (n = 11), P7-injured (n = 7) and P28-injured opossums (n = 9) were all allowed to grow to adulthood (P90–110) before being assessed using a series of behavioural tests (see below). Following these tests, animals were either used for retrograde labelling or gene expression analysis as described below. summarises numbers of animals used for the different experiments. ## Anaesthesia and Post-operative Care Surgical procedures were similar to those described previously. Anaesthesia was induced using inhaled isofluorane. Animals were anaesthetised to a surgical level and maintained with 3–4% isofluorane in O₂-enriched air, and all procedures were performed under sterile conditions. For the duration of the operation animals were placed on a heated pad (25–28°C). During recovery from anaesthesia opossums were placed under a heat-lamp. Post-operative pain was managed using intraperitoneal injections of buprenorphine (0.06 mg/kg). No post- operative infections were observed. No manual voiding of the bladder was required for any animal in this study. ## Spinal Cord Injuries Spinal cord injuries were made at two different ages, post-natal day (P) 7 or P28. Although detailed procedural methodology has been described previously, a brief description follows here. At P7 pups are still attached permanently to the mother’s teats making it necessary to anaesthetise the mother for the duration of the surgery. After anaesthetic induction the mother was placed supine to expose the pouchless abdominal area to which the pups attach. In addition, P7 pups were individually anaesthetised by placing over their snout a 1.5 mL tube filled with isofluorane- soaked cotton wool. The skin overlaying the mid-thoracic vertebrae was cut and using a fine ophthalmic blade (Sharpoint, 15° stab blade) a laminectomy was performed to expose the spinal cord. Spinal transection was made using fine scissors (5 mm blade, Fine Science Tools). Completeness of the transection was confirmed by running the point of a fine ophthalmic blade through the lesion site in contact with the spinal bones to ensure that no tissue bridges remained. Following tissue closure, the wound was sealed with surgical glue (Vetbond tissue adhesive, 3 M) and washed thoroughly with sterile saline. After recovery from anaesthesia the mother was returned to the animal facility. Pups injured at P7 heal very quickly and no tissue scar forms making identification of injured pups difficult, therefore all pups in a litter were subjected to spinal injury, while age-matched un-operated animals from different litters served as controls. By P28 pups detach from the mother’s teats for increasing periods of time and can be removed from the cage for surgical interventions. Pups were individually anaesthetised using inhaled isofluorane delivered via a facemask and were maintained on a heated pad under continuous anaesthesia. A longitudinal incision was made in the skin overlying the lower thoracic vertebrae, the musculature of the spinal column was incised to reveal the spinal column and a laminectomy was performed at T10. The spinal column was stabilized throughout using a stereotaxic frame. The spinal cord was completely transected using a fine ophthalmic blade, which was then passed repeatedly through the injury site to ensure no tissue bridges remained intact. Once the transection was done, the skin incision was sealed using surgical glue. Control animals (from the same litter) were anaesthetised but no operation was performed. All pups were returned to their mother after they had recovered under a heat lamp for 1 hour. We have confirmed previously the effectiveness of this lesioning technique (always performed by the same operator; BJW) by randomly selecting P7 and P28 pups immediately after surgery and fixing the spinal cords for histological analysis which consistently showed that spinal transections were complete. ## Behavioural Studies All animals grew to adulthood in their normal environment and were not interfered with further until they had reached 90–100 days of age when their locomotor abilities were assessed using an array of behavioural tests that we have used and described previously. ### Open field locomotion The Basso, Beattie, Bresnahan (BBB) Locomotor Rating scale was used to assess animals as they moved about freely in a standard polycarbonate animal box (30 cm×40 cm) with a smooth non-slip floor. The Locomotor Rating scale was developed for use in rat models of spinal cord contusion and transection but has been adapted for use in the two species of opossum, and published previously. *Monodelphis domestica* are a furtive species; consequently, it was found that attempts to assess them in large open spaces (as recommended in the original BBB publications) led to the animals running too rapidly from point to point for fine assessments to be made of limb placement and coordination. Therefore the size of the assessment area was modified (30 cm×40 cm) and this resulted in more accurate assessment due to more controlled, slower movements of the animals. Two blinded assessors observed *Monodelphis* for 4 minutes as they moved about the box and a score on the 21-point BBB scale was assigned. The test was performed under low ambient light and the animals were encouraged into constant exploratory locomotion by gently tapping the sides of the box. ### Gait Analysis Digital video recordings were made as *Monodelphis* walked on a treadmill moving at 6 metres per minute. A single camera was placed at 90° to the direction of locomotion (i.e., perpendicular to the sagittal plane of the animal). No modifications in treadmill speed were necessary for any injured animal. Footage (taken at 30 frames/second) was manually analysed frame-by-frame and each paw placement and lift-off was plotted from a continuous period of 8–10 seconds and a representative analysis period (typically encompassing 6–10 step cycles) was selected that was free of pauses in gait and other anomalies. Gait parameters were computed from these gait traces. Regularity Index (RI) was calculated by expressing the number of normal step sequence patterns (NSSPs; the sequential use of all four limbs in any order) as a percentage of total number of steps taken. Total step cycle, stance and swing durations were also calculated together with phase lags. Phase lag is defined as the point in time during the total step duration of one limb that another limb is placed. Phase lags can be calculated for any pair of limbs, but here, since opossums mostly use an alternating gait pattern (–LF–RH–RF–LH–;), only “in sequence” and girdle phase lags were calculated. ### Swimming test Swimming tests an animal’s ability to move its limbs in an environment of reduced sensory feedback. *Monodelphis* were recorded as they swam lengths of a narrow swimming pool (100 cm long×20 cm wide×60 cm deep) and climbed onto a visible platform placed at one end. The video footage was later examined for any indication of hindlimb movement during swimming, which would indicate the involvement of supraspinal control. ## Retrograde Axon Tracing Studies Following behavioural testing, quantitative anatomical tracing was employed to examine to what extent neuronal pathways had been remodelled following spinal injury in the neonatal period. ### Tracer injection A unilateral injection of fluororuby (tetramethylrhodamine-labelled dextran amine, 10,000 MW; Molecular Probes) was made into the lumbar spinal cord (L1/2) below the injury site, as described previously. Briefly, in deeply anaesthetised animals, after blunt dissection of the musculature overlying the spinal column, a small hole was drilled through the dorsal plate on the right side of the L1/2 vertebra through which a single injection of fluororuby (0.55 µl per injection; 25% w/v dissolved in 0.1 M Tris buffer with 2.5% (v/v) Triton X-100) was made using a pulled glass micropipette (70 µm outer diameter) under gentle pressure. The area was washed with sterile saline and packed with Gelfoam and the wound was sealed using number 4.0 silk sutures (Ethicon) and tissue glue. Animals were killed 6 days later and perfuse-fixed with ice-cold paraformaldehyde (4% in PBS). Brains and spinal cords were dissected and post-fixed in the same fixative overnight at 4°C. ### Neuronal counting and mapping Following post-fixation the brains and spinal cords were embedded separately in 4% agar blocks. The spinal cord was divided into smaller segments as appropriate for embedding and analysis purposes (see Results below). Brain and spinal cord tissue was sectioned at 100 µm using a vibrating microtome (Leica VT1000S) and mounted on glass slides under Fluorescent Mounting Media (DAKO). Brain was sectioned in the coronal plane and spinal cord was sectioned in the transverse plane. Sections were stored at 4°C. All labelled cells were counted manually in sections viewed under an appropriate filter specific to the wavelength of the fluorophore (Texas Red filter, Olympus BX50 microscope with a DP70 digital camera attached). For brainstems, all sections were analysed and every fluorescently labelled neuron was morphologically identified and manually counted under 20×magnification. For spinal cords, 10 evenly spaced sections through each area of interest were manually counted (as above for brainstems) and these counts were pooled. Additionally, fluorescent neurons in all analysed spinal cord sections were mapped using the layer function in Adobe Photoshop, where each dot represented one labelled neuron. These neuron maps were then overlaid to produce 2-dimensional reconstructions of fluororuby-labelled neurons throughout the areas of interest. These overlays were not used for counting purposes, but were instead used to give a general map of the distribution of labelled neurons. ### Morphometric analyses After fluorescent neurons had been counted, three sections per spinal segment were used for morphometric measurements and analysed under light microscope where grey and white matter areas can be clearly distinguished. Total section area and separate white and grey matter areas of each cross-section were determined. These values were then plotted to give a geometric profile of cross sectional areas along the entire spinal cord. ## Gene Expression Studies: qRT-PCR Following behavioural assessment, a subset of animals was killed and the spinal cord was dissected out, divided into segments (using the same divisions used for propriospinal neuron counts) and snap frozen in liquid nitrogen. Gene expression studies were performed on two of these samples: the L1–2 and L3–5 segments, both caudal to the injury site (which was performed at the T10 level; see Methods). Total RNA was extracted using the Trizol method and quantified using a NanoDrop ND-1000 UV-VIS spectrophotometer (Thermo Scientific). Quantitative real-time PCR (qRT-PCR) was used to determine the expression of selected neurotransmitter receptor genes. Reverse transcription from RNA to cDNA was performed where a 9 µl aliquot was used for samples with the lowest concentration (up to 2 µg total RNA) and other higher concentration samples were diluted to match to the lowest sample. qRT-PCR was performed on all samples using RT² SYBR Green ROX qPCR Mastermix (Qiagen) using gene specific primers based on published sequences for both *Monodelphis domestica* and human genome. Three genes from each class of receptors: excitatory, inhibitory and modulatory were chosen. shows forward and reverse primers used for the genes analysed in this study. The design of each primer was checked for efficiency by running standard curves and dissociation curves on all plates. The reaction was performed using an Applied Biosystems 7900 HT Real Time System. All primer pairs were tested for specificity (by dissociation curves). Cyclophilin was used as a general internal control reference gene for whole tissue. Average threshold cycle (Ct) values were obtained from triplicates for each sample. Analysis was completed for each sample by normalizing this average Ct value for each gene to its internal cyclophilin Ct value (to give the corresponding ΔCt value) followed by a comparison between spinal cord injured and control animals (to give ΔΔCt values). Fold changes were calculated relative to control using the 2^−ΔΔCt calculation and expression ratios (Log₂ transformations) used to express these changes graphically. ## Statistical Analyses Data are presented as mean ± sem. All data were plotted and analysed using GraphPad Prism software. For the majority of studies using multiple comparisons One-way Analysis of Variance (ANOVA) was used for analysis, with a Tukey post- test where appropriate. # Results In order to establish to what extent functional recovery had occurred following a complete spinal transection at either P7 or P28 the behavioural characteristics of spinally injured *Monodelphis* were determined and compared with age-matched control animals. Following the completion of behavioural testing, propriospinal connections in the spinal segment caudal to the original injury were investigated using fluorescent axonal tracing probe, fluororuby. Finally gene expression studies were performed on segments caudal to the injury site in an attempt to identify the pattern of changes in neuronal innervation that could explain locomotor abilities of spinal *Monodelphis*. ## Behavioural Testing Open field locomotion was assessed using the semi-quantitative BBB scale , and the results are illustrated in. All opossums injured at either P7 or P28 were able to take body-weight supporting steps with their hindlimbs. All P7-injured animals achieved consistent FL–HL coordination, although trunk instability and rotational errors in placing the hind paws led to a BBB score of 15.7±0.8. This was significantly different from the uninjured control group, which showed normal locomotion (BBB = 21). In contrast to P7 injured animals, no *Monodelphis* injured at P28 achieved consistent FL–HL coordination in spite of the full weight supported plantar stepping. These animals scored 12.3±0.2 on the BBB scale, significantly different from both control and P7-injured opossums. To further assess the opossums’ locomotion, video footage of these animals walking on a treadmill was analysed frame-by-frame to produce gait traces from which coordination and other gait characteristics could be computed. Regularity index measurements showed that both control (96.8±1.7%) and P7-injured (93.0±2.7%) opossums walked with highly regular FL–HL paw placement coordination, but P28-injured opossums did not (37.5±3.4%). In the P28-injured opossums this decrease in coordination reflected the fact that animals took different numbers of FL and HL steps leading to a FL:HL step ratio of 1.7∶1; this was in contrast to control (1.01∶1) and P7-injured (1.02∶1) opossums, where each forelimb and hindlimb was used equally. This inequality of limb usage following injury at P28 resulted from a shortening of step duration in the forelimbs relative to controls, rather than any change in the durations of the hindlimb steps, which did not differ from controls. In P28 injured animals stance and swing phases of forelimb stepping were both altered but no change was observed in either stance or swing duration in the hindlimbs. P7-injured animals displayed step durations that were slightly decreased for all limbs compared with controls. This resulted from a decrease in swing duration for the forelimbs, but for the hindlimbs the decrease was in stance duration. The Regularity Index assesses the order in which paws are placed but does not take into account the timing of their placements. For this purpose phase lags were determined. Results are shown in and summarized briefly here. No differences were found for either fore or hindlimb girdle phase lags for either injury group when compared to controls. However, the forelimb girdle phase lag was increased slightly in P28-injured animals compared with P7-injured animals. Since P28-injured animals recorded very few normal step sequence patterns (NSSPs) and, when they did, seldom employed any one repeated gait pattern, no other phase lags (diagonal or ipsilateral) were calculated for this group of animals. The control and P7-injured animals on the other hand, were highly regular in their use of an alternating gait (AltB;), so “in-sequence” phase lags were calculated for these groups of animals. The diagonal lags LF–RH and RF–LH both appeared to be increased following injury at P7, though only LF–RH was statistically significant. No differences were found for ipsilateral lags between control and P7-injured opossums. Examination of footage of swimming opossums allows their assessment in an environment of low sensory feedback, reducing the influence of peripheral reflexes on hindlimb movements. This is a critical measure of an animal’s capacity to make supraspinally mediated movements of the hindlimbs. Opossums are strong swimmers and typically use all four limbs during swimming. P7-injured opossums also used all four limbs and were likewise able to swim well. P28-injured opossums, in contrast, did not use their hindlimbs when swimming and relied only on their forelimbs to propel them forwards. Once these P28-injured animals touched the submerged portion of the exit platform with their hindlimbs, they were able to climb out of the pool and use their hindlimbs for weight- bearing locomotion. Results from these locomotor and swimming studies confirmed similar observations from a different previously published study, where representative video footage can be seen. ## Brainstem Neuron Labelling In order to determine if any axons had grown across the lesion site or if any spinal cord circuits had remodelled following injury, quantitative retrograde neuronal labelling was conducted in a subgroup of each of the three groups of animals studied (control, P7-injured and P28-injured; see in Methods) at completion of behavioural testing (above). Fluororuby was injected into the right hand side of the spinal cord at L1–2 and the spinal cord and brainstem were examined for fluorescently labelled cell bodies (see Methods). Brainstems were examined to establish whether any supraspinal fibres had crossed the injury site. The brainstems of control and P7-injured animals showed many labelled cell bodies (total numbers not significantly different between these two groups) suggesting pronounced regrowth of brainstem fibres across the injury site after spinal transection at P7. In contrast, no fluorescent labelling was detected in the brainstems of P28-injured animals suggesting no regrowth across the lesion was present following injury at this age. These data closely match previous, more extensive observations of brainstem labelling using bilateral fluororuby injections. ## Morphometric Analysis Morphological measurements were made along the entire length of the spinal cords dissected out from animals at the completion of behavioural and axon tracing studies. Whole mounts of control, P7- and P28-injured spinal cords are shown in. It was apparent that P7-injured spinal cords had regrown dense tissue across the injury site, but in P28-injured animals loose, transparent tissue had bridged the site of transection. Serial sections that were cut for labelled propriospinal neuron counting were examined and morphometric measurements were computed for all spinal segments. Under light microscope, a clear distinction between grey and white matter is visible in these sections without further staining; results are shown in. Whole cord cross-sectional area was decreased at the injury site of both P7- and P28-injured opossums compared to controls. In both injured groups this decrease in cross-sectional area appeared to persist throughout the majority of the thoracic spinal cord but in the lumbar and cervical regions the cross-sectional area between the groups was no longer different. There was a sharp, significant decrease in grey matter area but this only occurred at the site of injury in both injured groups and quickly returned to normal levels where it remained closely matched to the control cords for the much of the remainder of the spinal tissue. The pattern of white matter area changes showed a persistent decrease throughout the thoracic cord that was even more pronounced than that of the whole cord cross-sectional area. The values approached those of controls in both cervical and lumbar segments. ## Propriospinal Labelling In order to identify regrowth of, or modifications to, neuronal pathways that could be involved in the locomotor recovery of spinally injured animals, fluororuby was injected unilaterally into L1–2 spinal cord (see Methods) and labelled neurons were counted in sections throughout the spinal cord. Results of these counts are shown in and representative images are shown in. In control animals fluorescently labelled neurons were found in every spinal segment of the cord. These were located both ipsilateral and contralateral to the injection site. Rostral to the injury site, labelled propriospinal neurons were at all spinal levels in P7-injured animals, though fewer in number than in controls. No fluorescently labelled neurons were found in any spinal segment rostral to the injury site in the P28-injured animals. However, caudal to the injury site in the contralateral grey matter of L1–2 segments an increase in the number of labelled neurons in P28-injured but not P7-injured animals, compared to controls was found. In the L3–5 segments, on the other hand, a decrease in the number of labelled neurons in the contralateral grey matter for both P7- and P28-injured animals, compared to controls was observed. No obvious differences in the distribution of the labelled neurons were found in either the L1–2 or L3–5 segments. ## Gene Expression: qRT-PCR In order to identify the characteristics of propriospinal neuron cell populations in the lumbar spinal cords of control and injured opossums qRT-PCR was employed to examine gene expression of selected neurotransmitter receptors in the lumbar cord, caudal to the injury site. Gene targets were chosen to incorporate 3 receptors each involved in excitatory, inhibitory and neuromodulatory signalling, and included receptors of glutamate, GABA, glycine and serotonin. Results are shown in. ### Excitatory neurotransmitters Quantitation of the expression of three excitatory neurotransmitter receptor genes *Grin1* (NMDA receptor 1), *Gria2* (AMPA receptor 2) and *Grm5* (metabotropic glutamate receptor 5) is shown in (red bars). In the L1–2 segments, all three excitatory neurotransmitter receptor genes in P7-injured opossums were moderately upregulated compared to uninjured controls, though none reached statistical significance. For P28-injured animals the expression of *Grin1* was decreased significantly, while *Gria2* and *Grm5* did not change. A different pattern of expression was observed in the L3–5 cord segment. *Gria2* and *Grm5* expression was significantly increased in P28-injured animals; but neither changed in P7-injured opossums. *Grin1* expression was not different from controls in either P7 or P28 injury groups. ### Inhibitory neurotransmitters Quantitation of the expression of the three inhibitory neurotransmitter receptor genes, *Gabrb2* (GABA-A receptor beta-2 subunit), *Glra1* (Glycine receptor alpha-1 subunit) and *Glrb* (Glycine receptor beta subunit) is shown in (blue bars). In the L1–2 segments, the expression of all three inhibitory receptors in P7-injured animals appeared higher than controls although only expression of *Glra1* was statistically significant. For P28-injured opossum in this segment all appeared to be downregulated, though none reached statistical significance. In the L3-5 segments, expression was generally consistent with downregulation with *Glra1* showing moderately decreased expression in P7-injured animals and *Glrb* appeared to moderately downregulated in both P7- and P28-injured animals, although no change was statistically different from controls for any gene at either age. ### Neuromodulatory neurotransmitters The expression of serotonergic receptor genes *Htr1a* (5-HT_1a), *Htr2a* (5-HT_2a) and *Htr7* (5-HT₇) was quantitated and the results are shown in (black bars). In the L1–2 segments, serotonin receptor expression in P7-injured animals was generally decreased relative to controls where *Htr2a* (non-significantly) and *Htr7* (significantly) were both downregulated, but *Htr1a* showed a modest increase (not significant). In P28-injured animals, *Htr7* expression was significantly upregulated. In the L3–5 segment the expression pattern was similar to the L1–2 segment. The expression of two serotonin receptors was moderately decreased (*Htr2a* and *Htr7*) and one was unchanged (*Htr1a*) in P7-injured animals, while in P28-injured animals only *Htr7* was moderately, though non-significantly, increased. # Discussion In this study we have used behavioural testing, retrograde axonal tracing and gene expression studies to begin to unravel the biological basis of our observation that at a particular stage of *Monodelphis* development (P28) spinal injured animals develop locomotion in the absence of supraspinal innervation. This is in contrast to animals injured at P7 who grow axons across the lesion and develop locomotor abilities that approach those of uninjured animals. We have shown previously that opossums with complete spinal injury performed at four weeks of age (P28) are able to walk with weight-supporting hindlimb steps when adult. This was despite the complete absence of axon regrowth across the injury site and lack of restitution of supraspinal control, as evinced by the animals’ inability to swim using their hindlimbs. Part of the present study was an independent confirmation of previous observations, which was essential to repeat so that we could compare their locomotor abilities with the axonal tracer and gene expression studies. The high degree of over-ground locomotor function described in *Monodelphi*s with spinal transections in a neonatal period is in contrast to the behaviour observed following a similar transection in the adult, suggesting that P28-injured animals are somehow better able to access and utilize the spinal cord circuitry responsible for locomotion. This led to the speculation that in the absence of axonal regrowth across the site of injury local spinal circuitry may be modified or developed *de novo* as these animals grew into adulthood. A notable finding of the present study is the extent of morphological remodelling of propriospinal neurons in the lumbar spinal cord, caudal to the injury site following a complete transection at T10, as shown by fluorescent labelling. In P28-injured animals there were marked differences in the number of labelled neurons in the lumbar spinal cord, which have been quantified for the first time. In the L1–2 segments, an increase in the number of labelled propriospinal neurons was found, but in the L3–5 segments the number was decreased, suggesting a rearrangement of neuronal signalling architecture. In P7-injured animals, this remodelling was less pronounced: we found no change in labelling in the L1–2 segments, but a decrease, similar to P28 injured animals, in L3–5 segments. The observed remodelling, especially in P28-injured opossums, where regrowth of fibres across the injury site does not occur, may be particularly important in mediating their behavioural responses following injury. The morphological changes of propriospinal circuitry detected by labelling experiments were strengthened by the expression studies of genes coding for different neurotransmitter receptors (see below). ## Locomotion and Propriospinal Labelling In control adult *Monodelphis* the step durations for all limbs are approximately equal. This was not the case for P28-injured animals where unequal numbers of forelimb and hindlimb steps were observed. This suggests that in these animals the forelimbs and hindlimbs are operating at different cycle speeds and we speculate that their low but measurable regularity index represents the fact that four individual footsteps involving each limb may, by chance, fall in a coordinated manner, giving the impression of occasional coordination. This unequal usage of the limbs was a result not of a change in rate of hindlimb movements, but was due to shortening of forelimb step duration, suggesting that the forelimbs were required to compensate for the poor control and balance of the hindquarters. These timing irregularities also help to explain the variable level of observed coordination that led to a BBB score of 12.3±0.2 for P28-injured opossums. In P28 injured animals injection of fluorescent tracer into the lumbar spinal cord did not result in any labelling of neurons in the brainstem or in the spinal cord rostral to the injury site, confirming the observations of our previous study that regrowth of supraspinal or local fibres across the injury site does not occur in these animals and therefore can not account for the behavioural responses seen here. This lack of supraspinal drive is the major contributing factor to these animals’ inability to swim using their hindlimbs. We have compared P28-injured animals with opossums injured earlier in development (P7). In spite of their substantially normal locomotion P7-injured opossums retained several deficits in their locomotor organisation. In the open- field testing all animals displayed abnormalities in foot placements leading to a lower than normal BBB score. On the treadmill, although full coordination was observed, differences in the timings of foot placements compared to controls led to differences in diagonal phase lags suggesting that a reorganisation of step timing and execution had occurred. Neuronal labelling was used to show that supraspinal and propriospinal neurons had grown across the injury site and into the L1–2 segments in the P7-injured animals. The mere presence of labelled supraspinal neurons does not itself necessarily indicate that they are functionally involved in any behavioural responses but the fact that animals injured at P7 walked with normal FL–HL coordination and were able to utilize their hindlimbs during swimming suggests that supraspinal control was likely to have been reinstated and was playing an important role in these animals. To demonstrate further that the lumbar spinal cord, even when not connected to the brainstem, is capable of producing weight-bearing locomotion of the animal, we have performed a preliminary re-transection experiment on one P28-injured opossum. Results are illustrated in and show that re-transection had little impact on the animal’s gait. Although a reduction in regularity index occurred, the animal was still able to perform weight-bearing stepping. Spinal labelling after re-transection revealed that no fibres crossed the injury site. Taken together these observations suggest that the initial recovery was not due to spared axons, since surgically severing these would be expected to result in substantial loss of motor function. ## The Biological Significance of Gene Expression Studies Gene expression studies following spinal cord injury are typically aimed at analysing changes in early response inflammatory or regeneration-associated genes at the centre of the injury site (e.g.,). Changes in gene expression of neurotransmitter receptors are less commonly performed, and as a consequence only limited comparative information is available. However, two relevant papers are those of Wienecke *et al.,* who studied gene expression changes in motor neurons following spinal transection and Aimone *et al.,* who studied gene expression changes following spinal cord contusion; both studies were in adult rats. These are particularly important for comparison with our study because we were unable to carry out extensive spinal transection experiments in adult *Monodelphis* because of autophagy of the animals’ hindlimbs, which led to early termination of the experiments for ethical reasons. Expression of several excitatory glutamate receptors subtypes changed following spinal injury, mostly in P28-injured animals where NMDA-1 was downregulated in the L1–2 segment and AMPA-2 and Glu-R metabotropic receptor-5 were upregulated in the L3–5 segment. NMDA-1 was found upregulated in the adult rat. In P7-injured animals, two out of the three excitatory receptors appeared to increase their expression (NMDA-1 and AMPA-2) in the L1–2 segment but were not changed in the L3–5 segment. In the adult rat NMDA-1 was shown to increase its expression. Glutamatergic transmission has been implicated as a key player in locomotor circuit activation in the cat and rat and it is thought that the primary effects are mediated through the NMDA receptors. These ionotropic glutamatergic receptors are also involved in fast network interaction between interneurons and motoneurons. That NMDA agonism is commonly used as a circuit activator in studies into defining the essential components of the central patter generator – points to its importance as a positive modulator of locomotor output. Following injury, when other descending systems are lost, glutamatergic systems have been shown to be involved in the recovery of treadmill stepping and preliminary studies by the same group demonstrated that receptor density was maintained in an upregulated state following injury, which was supported by Wienecke *et al.,* who found upregulation of kainite and NMDA receptor subunits and associated proteins in lumbar motoneurons following adult spinal cord injury. This is similar to the finding in the present study where an upregulation at L1–2 of excitatory receptor genes in P7-injured spinal cord but contrasted with downregulation in P28-injured cord. The difference between P28-injured animals and those injured when adult may indicate important differences in circuit reorganisation at the two ages following spinal cord injury. For reasons mentioned above, it was not possible to do spinal injury experiments in adult *Monodelphis* and keep the animals for long enough for possible modifications to spinal circuitry to develop. A further contrast suggesting an increase in *Gria2* and *Grm5* in the L3–5 segments of P28-injured animals suggests that significant reorganization occurs not only between ages but also along the rostrocaudal axis of the cord. *Gria2* has been shown to be expressed more highly in interneurons than motoneurons in the lumbar cord, pointing to a potential difference in the balance of these neuronal subtypes in the lower lumbar segments of P28-injured animals. The two major inhibitory neurotransmitters in the spinal cord are GABA and glycine. There was a general trend towards increased expression of one GABA- and two glycine-receptor subunits in P7-injured animals in the L1–2 segments which contrasted with the general decrease of these genes in P28-injured opossums. Upregulation of *Glra1* has been shown following injury to the adult rat. These inhibitory receptors are thought to be expressed on interneurons in the lumbar cord and consequently differences in their expression may suggest a reorganization of the regulation of central pattern generation. The level of inhibitory drive may be modified after spinal injury. An increase in GABA synthesis following spinal injury has been shown in cats and an increase in GABA-A receptor subunits in certain motoneurons following injury in the rat has been observed. Administration of GABA receptor blockers or glycine receptor blockers resulted in improved stance and locomotion suggesting that the inhibitory GABA- or glycinergic systems in the spinal cord interfere with locomotor generation. Similar increases in the amount of glycine receptor have been shown in the lumbar spinal cords of rats following spinal injury, similar to the P7-injured opossums. Locomotor training, which has positive benefits on the stepping ability of completely transected animals, normalizes the level of these inhibitory neurotransmitter receptors and synthesis molecules, pointing to a use-dependency of these systems. The P28-injured animals in the present study were able to walk extremely well with no training of any kind. Taken alongside the evidence of decreased inhibitory transmitter receptor expression in their upper lumbar spinal cord this suggests that this downregulation may have helped to reduce the inhibitory drive, so that locomotion was possible. P7-injured animals generally displayed increased levels of inhibitory neurotransmitter receptor gene expression. It is possible that the presence of their re-established supraspinal innervation enabled them to overcome any increased inhibitory drive resulting from this overexpression. Overall the results suggest that there was a change in the balance between excitatory and inhibitory inputs from local spinal and supraspinal innervation in P7-injured animals compared to P28-injured animals. Serotonergic receptor expression also changed following spinal cord injury. In the L1–2 segments of P7-injured animals, there was a general decrease in expression of *Htr2a* and *Htr7* but not *Htr1a*, suggesting a decrease in sensitivity to serotonin in these animals. This contrasted with the increased *Htr7* expression in P28-injured animals at L1–2 suggesting an increase in sensitivity. Serotonergic systems in particular have been shown to make important contributions to descending pathways that mediate signals from the brainstem to all levels of the spinal cord. The regulation of the neuromodulatory receptors in the injured spinal cords here may be a reflection of the differing supraspinal drive that is present in these animals. It is likely that at least a portion of the brainstem neurons whose axons span the lesion site in the P7-injured animals would be serotonergic since retrograde axonal labelling consistently labels neurons in the areas where these systems are known to originate, the raphe nuclei and sub coeruleus respectively. However, since only a proportion of these regrow following injury at P7, the decrease in the neuromodulatory receptor gene expression could be an adaptation reflecting this limited level of supraspinal input. In the P28-injured opossums, in which no supraspinal drive was present, there was a decrease in expression of this serotonin receptor, and no change in any of the others that were examined. This change may reflect downregulation resulting from the lack of use of this system. Previous studies have also revealed that following spinal injury adaptive changes occur for many receptors in the lumbar spinal cord. Giroux *et al.,* found that in the first weeks after injury there was an upregulation in receptors of the adrenergic and serotonergic systems, as measured by autoradiographic receptor binding studies, but these later declined to control levels. The authors speculated that this change in receptor regulation demonstrated the continued importance of these systems after complete spinal transection. Increased expression of a wide array of neuromodulatory receptors were found in the late phase after SCI in adult rat, analysed by PCR. ## The Role of Propriospinal Input More research is becoming focused on the role that the isolated spinal cord itself may play in recovery. Recent studies have indicated that propriospinal remodelling plays an important role in recovery of function in adult animals with incomplete spinal lesions. Propriospinal relay connections have been demonstrated to reconnect pyramidal cells in the cortex with their target lumbar motoneurons and mediate the reinstatement of supraspinal motor control despite the absence of any direct supraspinal innervation. However, studies such as these rely on animals with partial lesions where spared tissue provides a bridge through which reconnections are made. Adult animals with complete spinal transections do not recover to the same degree unless external stimulation is given in the form of cutaneous feedback, serotonergic agonist administration, electrical stimulation, or a combination of each. In contrast to other reports in which weight supported stepping has been reinstated following injury, the recovery of spinal injured opossums in the present study is independent of any interventions and consistently occurs in overground locomotion as well as on the treadmill. ## Limitations of the Study ### Spinal injury in adult animals One limitation to the approach taken in the present study is the inability to perform complete spinal transections in control, adult opossums and keep the animals long enough to study their locomotion post-injury. This experimental approach was attempted but unfortunately the animals that were injured for the very first time while adult were found to chew at their insensate hindlimbs and tail so aggressively that studies had to be halted a week post injury (also described). The autophagia continued in spite of attempts to dissuade the animals by painting their hindlimbs with various foul-tasting liquids normally used to stop people biting their nails. This group of animals would have added significant value to the comparison with P7-, P28-injured opossums and those that were re-transected as adults, particularly in respect to the labelling and gene expression studies. However we were able to use this difficulty for comparative purposes in observations of re-transected animals. Thus the P28-injured animal, when re-transected as adult, did not chew on its hind limbs indicating differences in its spinal sensory circuits. Instead results of similar experiments in *Didelphis virginiana* and/or rats have been used in the present study for comparison where appropriate. ### Labelling protocol In a previous study where the sole intention of the spinal labelling experiments was to establish whether or not re-growing supraspinal axons (i.e., originating in the brainstem neurons) could be found, bilateral injections of fluororuby were used and the entire brainstem was sectioned and counted to establish the extent of the regrowth. In the current study, however, unilateral injections were employed so that propriospinal neurons, both ipsilateral and contralateral that projected into the injection area could be counted and mapped to assess remodelling rather than regrowth. As all animals were injected in a similar manner, a direct comparison of results obtained for labelled neurons between different groups should reflect differences in their connectivity. ### qRT-PCR gene expression The *Monodelphis* genome is not well annotated and no commercially available molecular tools are available. An initial study was performed using a commercially available Profiler Human Neurotransmitter Receptor PCR array plate (Qiagen) to screen for differentially expressed genes. However there was a lack of consistent and repeatable hybridization with *Monodelphis* cDNA, leading to large variability in control values so that no statistical differences between operated and control animals could be detected. Other studies using more conventional laboratory species have successfully employed genetic screens (usually in the form of microarrays –). We limited our studies to three representative genes from each category of transmitters: excitatory, inhibitory and neuromodulatory. It is possible that other neurotransmitter receptors might have shown other differences. An immunocytochemical demonstration of some of these gene products would have added important information. Unfortunately none of the commercially available antibodies to these antigens from other animal species cross-react with *Monodelphis* tissue. ## Conclusions The experiments described in the present study provide the beginnings of a basis for understanding the mechanisms that allow developing quadrupedal animals to recover from a complete spinal transection and be able to exhibit weight-bearing locomotion in the absence of any axon growth across the lesion site and indicate the important role that propriospinal remodelling may play. We are primarily interested in the biological mechanisms underlying the substantially normal development that occurs following spinal cord injury when the lesion is made very early in development. We do not make any specific claims about how this approach to spinal cord injury might contribute to new therapies in spinal patients, although this may be possible once the mechanisms are better understood. From the present study it seems likely that the observed weight bearing and alternating hindlimb movements in spinally transected animals are primarily a function of the sensory input from the hind limbs during over ground walking, as these movements are lost when the animal is placed in water. However, the forward locomotion may be generated by the stretching of trunk and limb muscles and joints when the animal initiates movement with the forelimbs. This raises concerns about the suitability of rodents and other quadrupeds for studies of spinal cord injuries, where an attempt may be made to apply the results to human patients. The present study suggests that significant weight bearing and forelimb–hindlimb movements can occur even in the absence of neural connections across a lesion and emphasizes the need for retrograde labelling to show whether or not apparent improvements in the BBB or BMS scores do show evidence for supraspinal re- innervation following spinal cord injury. If weight-bearing locomotion in spinally injured animals is indeed initiated by the forelimbs then this will be of limited value for patients. A bipedal model of spinal cord injury would be more appropriate; this is approached by the use of robotic harnesses in rats, which oblige the rats to walk on their hindlimbs, otherwise species that are normally bipedal need to be considered. # Supporting Information [^1]: The authors declare that no competing interests exist. [^2]: Conceived and designed the experiments: BJW NMN PJC KMD NRS. Performed the experiments: BJW NMN SW MZ JST. Analyzed the data: BJW NMN SW JST KMD NRS. Contributed reagents/materials/analysis tools: BJW NMN SW JST WDD HZ PJC KMD NRS. Wrote the paper: BJW NMN KMD NRS.

# Introduction The introduction of combined active antiretroviral therapy (cART) for HIV infection was followed by a significant decline in the number of deaths attributable to opportunistic infections. As the survival of chronically HIV- infected individuals increased, so did the prevalence of major chronic degenerative disorders, including metabolic, cardiovascular, renal, and bone- related conditions. For example, the widespread use of cART caused a marked drop in the incidence of rapidly progressive HIV-associated nephropathy (HIVAN), but HIV infection is now associated with indolently developing chronic kidney disease (CKD). Currently, renal disease risk in patients with HIV infection appears to be compounded by ethnicity, chronic comorbidities, concurrent viral infections, and the use of antiretroviral drugs with nephrotoxic potential. Regardless of the underlying cause, CKD is considered a significant independent risk factor for mortality among HIV-infected patients in both developed and resource-limited settings. Given these adverse associations, all HIV-infected persons are advised to undergo kidney function evaluation. In 1991, the Brazilian Ministry of Health launched a pioneer program to provide free access to antiretroviral treatment for all HIV-infected individuals as needed. A marked drop in morbidity and mortality rewarded Brazil’s unique position of a middle-income economy with more than 200.000 patients receiving cART. While successful, increasing patient life expectancy and HIV infection prevalence is anticipated to place a growing burden on limited local specialized human resources and health care logistics. Because the first step toward preparation and planning involves collecting evidence, the objectives of this study were to screen for decreased glomerular filtration rate (GFR) and associated risk factors in an outpatient cohort of HIV-infected patients followed at Instituto de Pesquisa Clinica Evandro Chagas (IPEC)/Fiocruz, a referral center for HIV care and research in the city of Rio de Janeiro, Brazil. # Methods ## Ethics Statement The project was submitted to the Institutional Review Board of Instituto de Pesquisa Clínica Evandro Chagas - Fundação Oswaldo Cruz and approved on Sept 13, 2010 as \# 044/2010. A translation of the conclusion of the board is read as follows: “In substitution to written informed consent, following declaration \#196 of 1996 from the Brazilian National Health Council, the investigators in charge (PS and BG) subscribed to a term of agreement pledging confidentiality and anonymity of the participants of the study”. The project was not specifically funded. ## Cohort Identification and Study Design The IPEC clinical HIV service has been providing patient care since 1986. The HIV cohort database was started in 1998. Longitudinal data are updated regularly using clinic and inpatient clinical documentation, laboratory results, and pharmacy records. Medical providers and support staff record clinical information, including prescriptions for antiretroviral drugs. Trained data abstractors collect relevant information onto standardized forms for further processing. We performed a cross-sectional study including all patients aged 18 years or older who attended the IPEC outpatient clinic and had at least one serum creatinine evaluation during the calendar year of 2008. HIV infection was diagnosed according to the Brazilian Ministry of Health criteria : either two ELISA tests <u>or</u> two rapid tests followed by a confirmatory test, either a Western blot or immunofluorescence assay. Creatinine was measured with the Roche enzymatic assay at the IPEC Safety Laboratory, which is certified by the American College of Pathology. We excluded patients with missing ethnicity data and those whose creatinine results were obtained during or up to six weeks after admission to a hospital. The outcome of interest was the estimated glomerular filtration rate (eGFR), calculated from a single fasting creatinine measurement using the CKD-EPI equation. The GFR results were stratified according to the NKF-K/DOQI staging system. All of the patients classified in stages 3 through 5 (i.e., with a GFR \<60 mL/min/1.73 m²) were considered to have a decreased GFR. The data that were obtained concurrently with the assessment of renal function and evaluated as explanatory variables for reduced GFR included gender, age, ethnicity, HIV transmission group, AIDS defining illness (according to the 1993 AIDS CDC definition**)**, time since the HIV infection diagnosis, co-infection with the hepatitis B (HBV) virus (defined as a positive HBsAg test), co- infection with the hepatitis C (HCV) virus (defined as positive anti-HCV serology), diabetes (defined as the use of anti-diabetic drugs or a record in the patient’s chart), hypertension (defined as a record in the chart or the use of antihypertensive therapy), hyperlipidemia (defined by the prescription of lipid-lowering drugs or documentation in the chart), current HIV1-RNA plasma viral load, current CD4+ cell count, nadir CD4+ cell count, past and current ARV exposure, and exposure to selected nephrotoxins in the three years prior to the GFR assessment (aciclovir, valaciclovir, ganciclovir, gentamicin, amikacin, amphotericin B, and meglumine antimoniate). ## Statistical Analysis The prevalence ratio (PR) of decreased GFR was estimated with generalizing linear models created assuming a Poisson distribution with logarithmic link function and robust variance. The covariates that were significantly associated with renal dysfunction in the univariate analysis (p\<0.10) were entered in the multivariate models. Given the low prevalence of the study outcome, the variables were retained if they met the condition p≤0.10 in a backward elimination strategy <u>and</u> if they were considered potential confounders (e.g., when removed, a change equal to or greater than 20% in the prevalence ratio of any other variable of the model was observed). All of the analyses were performed using STATA/SE version 10.1. # Results ## Study Population Between January 1 and December 31, 2008, 2,345 patients were seen at the IPEC outpatient clinic. A total of 375 patients (16%) were excluded from further analyses for the following reasons: lack of serum creatinine testing (310), creatinine testing performed exclusively during hospitalization or within 6 weeks after hospital discharge (to prevent the inclusion of patients with acute kidney injury) (63), and lack of ethnicity data (thus interfering with the estimation of GFR by the CKD-EPI equation) (2). We analyzed data from 1,970 patients. The median age was 41.6 (interquartile range \[IQR\] 34.0–48.2) years; 63.6% were male and 57.1% were white. The patients were followed at IPEC for 10,456 person-years. Hyperlipidemia, hypertension, and diabetes mellitus were present in 46.9%, 26.6%, and 9.3% of the patients, respectively. The median time since the diagnosis of HIV infection was 78.5 (IQR 29.1–136.2) months, and 42.8% met the clinical definition for AIDS. The median nadir and current CD4+ cell counts were 189 (IQR 78–308) and 460 (IQR 307–650) cells/mm³, respectively. The plasma viral load was undetectable in 70.5% of the subjects. Overall, 1,634 patients (82.9%) had been exposed to ART, with a mean cumulative exposure (MCE) of 67.7 months (SE: 1.3) and a total of 9206 person-years of follow up. Of the patients, 807 (41.0%) had been exposed to tenofovir (MCE 25.0±19.3 months), 305 (15.5%) had been exposed to indinavir (MCE 31.8±26.5 months), 490 (24.9%) had been exposed to atazanavir (MCE: 25.7±18.8 months), and 516 (26.2%) had been exposed to lopinavir (MCE: 27.4±21.3 months). At the time of evaluation, 744 patients were on tenofovir (TDF), and 67.5% (n = 502) had been exposed for at least one year. Among the patients exposed to TDF in the past (n = 63), 52.4% had been exposed for more than one year. A total of 860 patients were exposed to atazanavir and/or lopinavir (ATV⋁LPV); 546 were currently using one of these drugs, and of these patients, 77.7% (n = 424) had at least one year of exposure. Among the patients exposed to ATV⋁LPV in the past (n = 314), 72.9% had been exposed for at least one year. Prior exposure to potentially nephrotoxic non-ARV drugs was recorded in 345 subjects (17.5%) as follows: aciclovir, 13.2%; valaciclovir, 4.9%; ganciclovir, 0.6%; amikacin, 0.2%; gentamicin, 0.1%; amphotericin B preparations, 0.1%; and meglumine antimoniate, 0.1%. The median eGFR for the entire cohort was 111.4 mL/min/1.73 m². presents the categories of eGFR prevalence according to the NKF-K/DOQI stages. Of the 1,970 patients, 1,647 (83.6%) had an eGFR above 90 mL/min/1.73 m², 249 patients (12.6%; 95% CI: 11.2–14.2) had an eGFR between 60 and 89 mL/min/1.73 m², 61 patients (3.1%; 95% CI: 2.4–4.0) had an eGFR between 30 and 59 mL/min/1.73 m², six patients (0.3%; 95% CI: 0.1–0.7) had an eGFR between 15 and 29 mL/min/1.73 m², and seven patients (0.4%; 95% CI: 0.1–0.7) had an eGFR below 15 mL/min/1.73 m². Notably, 74 patients (3.8%; 95% CI: 3.0–4.7) fulfilled the primary outcome criterion of an eGFR below 60 mL/min per 1.73 m². also shows that in comparison with the patients with GFR ≥60 mL/min per 1.73 m², the patients with decreased GFR were more frequently aged 50 years or older (51.4% vs. 18.7%) and had higher prevalence rates of hyperlipidemia (58.9% vs. 46.6%), hypertension (54.1% vs. 25.5%), diabetes (27.0% vs. 8.6%), and positive anti-HCV serology (11.0% vs. 5.9%). These patients also had been diagnosed with HIV for a longer time, with a median time of 117.7 (IQR 50.7–158.7) vs. 77.1 (IQR 28.9–134.2) months, and they were more likely to have an AIDS diagnosis (52.7% vs. 42.4%) and a current CD4+ count below 350 cells/mm³ (50.7% vs. 31.2%). Exposure to TDF (60.3% vs. 40.2%), current (42.5% vs. 37.6%) and previous use of TDF (17.8% vs. 2.6%), and exposure for at least one year (43.0% vs. 26.5%) were more common among the patients with decreased GFR. The same trend was observed for exposure to ATV⋁LPV (68.5% vs. 42.6%), current use of ATV⋁LPV (49.3% vs. 26.9%), and exposure to these drugs for at least one year (58.9% vs. 32.2%). Past exposure to indinavir was also more common among the patients with reduced GFR (37% vs. 14.7%). Exposure to potentially nephrotoxic non-ARV drugs was observed in 19.7% of the patients with GFR \<60 mL/min/1.73 m² vs. 17.4% of the patients with GFR ≥60 mL/min/1.73 m². In the univariate analysis, age, diabetes, hypertension, hyperlipidemia, hepatitis C, time since the HIV diagnosis, opportunistic diseases, current CD4+ count, indinavir exposure, and past and current exposure to TDF and ATV⋁LPV were significantly and positively associated with the primary outcome. In the multivariate adjusted Poisson regression model, the following factors were significantly associated (p\<0.05) with a GFR \<60 mL/min per 1.73 m²: age ≥50 years (PR = 3.4; 95% CI: 1.7–6.8), diabetes (PR = 2.0; 95% CI: 1.2–3.4), hypertension (PR = 2.0; 95% CI: 1.3–3.2), current CD4+ cell count \<350 cells/mm³ (PR = 2.1; 95% CI: 1.3–3.3), and past exposure to TDF (PR = 4.7; 95% CI: 2.3–9.4) and IDV (PR = 1.7; 95% CI: 1.0–2.8). Among the patients using ATV or LPV at the time of creatinine evaluation, the prevalence of decreased GFR was higher than the prevalence of those never exposed to such drugs, although this association was not statistically significant (PR = 1.7; 95% CI: 0.9–3.1; p = 0.09). # Discussion The main findings of our study were as follows: (1) as far as we know, our study investigated the prevalence of decreased GFR in the largest cohort of HIV patients highly exposed to antiretroviral drugs from a middle-income country; (2) we found that only 3.8% of the participants had a GFR below 60 mL/min per 1.73 m² as estimated by the CKD-EPI equation; and (3) the results indicate that a decreased GFR was positively and independently associated with markers of the severity of HIV infection as well as with other chronic degenerative diseases and exposure to nephrotoxic antiretroviral drugs. Large cohort studies (i.e., with more than one thousand members) investigating the prevalence of CKD in HIV-infected individuals have been conducted with either cART-treated patients from high-income countries – or predominantly treatment-naïve subjects from low-income settings. Healthcare in middle-income countries fits between these two extremes. While low-income countries still struggle to meet the needs of their people, middle-income economies have already developed the ability to deliver basic health services but face hurdles regarding coverage, priority setting, financing, and efficiency. With regard to HIV treatment, despite undeniable progress, middle-income countries exhibit marked differences in access to cART and treatment consistency. The epidemiological features of our cohort, including the prevalence of CKD risk factors, are akin to those in higher-income countries (e.g., the predominance of urban white males in early middle age (median age 41.6 years), access to cART, and a moderate prevalence of chronic comorbidities). Decreased GFR was uncommon (3.8%) among the HIV-infected patients in our cohort. Differences in the baseline sociodemographic data, the cohort inclusion criteria and the accrual period as well as exposure to different antiretroviral medications preclude exact comparisons among published reports. Nevertheless, data from the high-income country cohorts mentioned above show that the median ages and prevalence rates of CKD (defined by GFR \<60 mL/min/1.73 m²) for the three main creatinine-based GFR estimation methods were as follows: an age of 42.8 years and a CKD prevalence of 4.2% for CG, an age of 39.3 years and a CKD prevalence of 4.5% for MDRD, and an age of 41.3 years and a CKD prevalence of 4.5% for CKD-EPI GFR. Notably, 16.4% of the members of the cohort had estimated GFRs below 90 mL/min/1.73 m². Previously, similar double figure rates have been highlighted as an indication of a high burden of CKD in HIV-infected cohorts. However, a sizable proportion of healthy normal adults have eGFR values within the range of 60 to 89 mL/min/1.73 m². Consequently, it is feared that such a high cut-off inappropriately inflates the prevalence of CKD. In any case, even this threshold has been shown as an independent risk factor for cardiovascular mortality in apparently healthy adults. Concerns about the over-diagnosis of CKD are extensive for our chosen eGFR cut- off of 60 mL/min/1.73 m². Nevertheless, this boundary is widely used for screening purposes. More importantly, a single eGFR value under 60 mL/min/1.73 m² is independently associated with an increased risk of cardiovascular, – and all-cause mortality. Furthermore, most cases of false- positive CKD labeling involve older individuals, mostly females with low muscle mass. We studied a targeted group of young to middle-aged men with known risk factors for chronic kidney damage, therefore enhancing the positive predictive value and minimizing the negative predictive value of the eGFR test. Cross-sectional studies in healthy individuals show that directly measured glomerular filtration rated are lower in older adults than younger adults and that this gap is magnified when GFR is estimated with creatinine-based equations. We, like others, did find that a lower eGFR was independently associated with increasing age. This finding was particularly relevant given the rising incidence of HIV infection in older adults, and the remarkable success of cART, which is allowing younger infected adults to age. In addition, HIV- associated inflammation, exposure to potentially toxic antiretroviral drugs, accelerated aging, and chronic comorbidities contribute to the premature occurrence of “non-AIDS” degenerative diseases, including CKD. With regard to this matter and in agreement with previous reports, we noted that both diabetes and hypertension were significantly and independently associated with a low eGFR. Previous studies have provided evidence that cART controls kidney cell damage and preserves renal function and that its withdrawal is associated with HIVAN and kidney function deterioration. In our cohort, 18.1% of the patients receiving cART had detectable viral loads, and 32% had a current CD4+ cell count below 350, indicating some degree of uncontrolled infection, due to either viral resistance or non-adherence. Overall, our patients had fairly advanced HIV disease, as shown by the median nadir of 189 CD4+ cells/mm³ and by the fact that 42.8% had a previous AIDS defining illness. Against this background, we did not find an association between reduced GFR and either of these variables, but we noticed that a reduced GFR was significantly associated with a current CD4+ count below 350 cells/mm³, a marker of incomplete immune reconstitution. The rationale behind cART is to administer multiple drugs targeting different stages of HIV cell entry and the cell life cycle to achieve long-lasting viral suppression. This strategy exposes patients to multiple potential drug interactions and entails a lifelong risk of drug toxicity, including nephrotoxicity. We observed that the past use of tenofovir and indinavir was associated with a decreased glomerular filtration rate. An association between current atazanavir and/or lopinavir use and decreased glomerular filtration rate was also observed but it did not reach statistical significance (p = 0.09). Tenofovir was used by 41% of our patients, and 84.8% of those patients were still using the drug. Even after accounting for demographics, HIV-related factors, comorbidities, and other antiretroviral drugs, past TDF use remained independently associated with low GFR compared with those with no previous exposure. This result suggests that kidney damage and dysfunction do not quickly reverse after the exposure ends. In fact, among the patients who discontinue TDF because of renal impairment, only a minority returns to their pre-TDF GFR. Furthermore, data from a large cohort in the US showed that past users of TDF remained at an increased risk of proteinuria and rapid renal function decline compared with those never exposed to TDF. Because TDF is a first-line treatment for HIV infection that is increasingly used in all antiretroviral regimens and in pre-exposure prophylaxis, attention to the early signs of renal toxicity, perhaps even above the 60 mL/min/1.73 m² cutoff, is necessary. Contrary to previous use, the current use of TDF was not associated with a decreased GFR. This finding differs from other reports that did find such an association. Several possible underlying explanations can be given for this result, including differences in study definitions, in the background characteristics of the study populations, and in the length of exposure to tenofovir, among others. Nevertheless, we believe that the broadly reported association between TDF and renal dysfunction , led to careful and repeated measurements of serum creatinine, resulting in earlier withdrawal and prevention of unchecked toxic exposure to TDF. In fact, recent data from the D:A:D study showed that decreased GFR was associated with a significant rate of TDF discontinuation, which prevented further deterioration of renal function. Interestingly, this behavior was not replicated when the GFR declined in patients exposed to other antiretroviral agents. Both indinavir and atazanavir have been associated with crystalluria, crystal nephropathy, nephrolithiasis, and CKD in previous studies. In our cohort, past users of indinavir had a higher risk of GFR decline that reached borderline significance (PR = 1.7; 95% CI: 1.0–2.8). In recent years, the newer protease inhibitors, particularly atazanavir and lopinavir, have replaced indinavir. Contrary to the above mentioned studies, past or current use of atazanavir or lopinavir was not significantly associated with a GFR below 60 mL/min/1.73 m². However, in the multivariate regression model, there was a trend toward an association of current use of ATV⋁LPV and low GFR (PR = 1.7; 95% CI: 0.9–3.1). It is possible that the low prevalence of the decreased GFR end-point in the cohort contributed to this negative finding. Further investigation is necessary to delineate the impact of these drugs on kidney function. This study does have several limitations. First, as in all cohorts, the sample includes adherent individuals with extensive and complicated medical histories that may obscure the associations of interest. Second, because of the cross- sectional design, we can draw associations between events but cannot establish temporal sequences. Third, the diagnosis of low GFR was based on an isolated creatinine measurement. In spite of this, as discussed above, this strategy has been used widely in screening investigations related to CKD and associated risk factors, , and a single eGFR below 60 mL/min/1.73 m² is associated with cardiovascular, – and non-vascular mortality. We were also particularly careful to exclude patients who developed acute kidney injury by excluding any creatinine measurement obtained during a period of hospitalization or up to six weeks thereafter. Fourth, given the low prevalence of the low GFR outcome in our cohort, we may have missed factors associated with renal disease. Finally, kidney disease in subjects with eGFR values above 60 mL/min/1.73 m² may have been underestimated because we lacked data on proteinuria and renal tubular involvement, both early and significant indicators of renal disease in HIV patients. This study also had multiple strengths. It was conducted at a referral center for HIV/AIDS care and research in one of the epicenters of the Brazilian AIDS epidemic on a very well characterized cohort with mixed ethnic background and universal access to first-line NNRTI-based regimens, as well as free access to PI-based regimens for subsequent treatment steps after first- and second-line virological failure. Additionally, we employed the recently developed CKD-EPI equation, which provides a more accurate estimate of GFR than other creatinine- based equations, an observation that was also confirmed in patients with HIV infection. In summary, in this well-characterized cohort of Brazilian patients with HIV/AIDS with long-term exposure to cART, an age over 50 years, a current CD4 count less than 350 cells/mm³, diabetes, hypertension, previous use of tenofovir, and previous use of indinavir were associated with a decreased glomerular filtration rate (p\<0.05). The current use of atazanavir or lopinavir showed a weaker association with the outcome (p = 0.09), and this result may be related to the low prevalence of the outcome and the study population size. We were able to confirm that the current risk factors for CKD in HIV-infected individuals are broad and include virus-related factors as well as degenerative and nephrotoxic factors. With long-term exposure to cART, within the context of premature aging and cumulative chronic comorbidities, cumulative exposure to even low-grade nephrotoxins may eventually prove to have deleterious effects. Nevertheless, it is reassuring that despite the potential for nephrotoxicity from some pivotal antiretroviral drugs, in the short-term, advanced renal disease remains very rare. [^1]: The authors have declared that no competing interests exist. [^2]: Conceived and designed the experiments: BG RKF AVO RIM SWC VGV JHRS. Performed the experiments: PS CBC LEC. Analyzed the data: PS BG RKF CBC RIM SWC VGV JHRS. Contributed reagents/materials/analysis tools: AVO CBC. Wrote the paper: PS BG RKF SWC PML JHRS.

# Introduction Leishmaniasis is a neglected tropical disease (NTD) caused by multiple species of *Leishmania* parasites and transmitted by the bite of female sand flies. It is endemic in 98 countries, and is mostly concentrated in low-and middle-income countries in South Asia, East Africa, Latin America and in the Middle East. The disease presents in different forms depending on the species and geographical location. Visceral leishmaniasis (VL; also known as kala-azar) presents with fever, weight loss, hepatosplenomegaly and may have neurological manifestations. If untreated, it has a fatality rate of over 95%. Post kala-azar dermal leishmaniasis (PKDL), occurring as a consequence of VL can cause erythematous or hypopigmented macules, papules, nodules and patches. Cutaneous leishmaniasis (CL) patients present with plaques, nodules and / or ulcers and, in the case of mucocutaneous leishmaniasis (MCL), symptoms manifest on the mucous membranes of the nasal and oral cavities and surrounding tissues. These forms of leishmaniasis invariably leave visible disfiguring lesions and lifelong scars on the skin. Although not fatal, CL lesions have been described in the literature as a source of distress and discomfort. Such visible manifestations have been linked to social stigma that could potentially lead to isolation and self-stigma and psychosocial morbidity (PM). For example, in Afghanistan, the incorrect belief that the disease can be transmitted by human contact results in social exclusion that can range from not sharing utensils to severe physical and emotional isolation. In some cultures, women are considered unfit for marriage or are separated from their children when they have the disease. A study involving high school students in Morocco showed awareness of the stigma attached to CL with self-stigmatization including feelings of shame, embarrassment, depression, and self-contempt. Although inflammation of the liver and spleen are the most well-documented clinical manifestations of VL, inflammation of the nervous system has also been reported. Neurological manifestations in human VL include peripheral and cranial nerve dysfunction, neurological tremors, meningitis, seizures, paresis, and haemorrhagic stroke. In addition, naturally infected dogs with canine VL as well as rodent models of VL show neuroinflammation. Despite these findings and research showing close links in general between neuroinflammation and mental health problems, there is limited evidence of a direct link between the neuroinflammation present in leishmaniasis and mental health problems. There is a growing body of evidence suggesting a causal link between systemic and localised inflammation and depression in other mental health disorders including two meta-analyses that show statistically significant raised levels of inflammatory markers in depressed subjects. Patients with depression often show cardinal features of inflammation, including increased expression of pro- inflammatory cytokines and their receptors as well as an upregulation of acute- phase reactants, chemokines and soluble adhesion molecules in peripheral blood and cerebrospinal fluid (CSF). In VL, a link between mental illness (MI) and systemic or neuroinflammation has been postulated. At the time this review was conducted, there had been no high quality comprehensive systematic review published on possible aetiological associations and the impact of leishmaniasis on mental health and psychosocial wellbeing. A search of the Cochrane Database of Systematic Reviews and the Database of Abstracts of Reviews of Effects (DARE) identified no systematic reviews of the effects of healthcare interventions for the potential mental health consequences of leishmaniasis. A detailed search of PROSPERO indicated that no other systematic reviews were in progress. Subsequently, one scoping review has recently been published making a ‘preliminary assessment of the extent of the literature’ focusing solely on localised cutaneous leishmaniasis and with no quality appraisal for included studies. In this systematic review, by mental illness (MI), we are referring to diagnosed mental or psychiatric disorders that are listed in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5). The term psychosocial morbidity (PM) does not have an official definition but is used widely in the literature\[–\], and will be defined in this study by any morbidity caused by psychosocial factors. The psychosocial approach looks at individuals in the context of themselves and the combined influence of their surroundings and looks at its influence on its mental and emotional wellbeing and the environment. PM is highly influenced by culture and society. It can manifest in the form of mental illness, emotional distress, where feelings such as fear, sadness, despair or shame manifest, with stigma playing a chief role here. Quality of life (QoL) is defined by the WHO as an individual’s perception of their position in life in the context of the culture and value systems in which they lie and in relation to their goals, expectations, standards and concerns. It is affected by the person´s physical health, psychological state, personal beliefs, social relationships and their relationship to salient features of their environment. These three terms (MI, PM and QoL) overlap substantially (although not completely) and together encompass all the outcomes we were interested in assessing by using a broad and sensitive approach. The objective of this study was, therefore, to conduct a comprehensive systematic review (SR) of MI, PM and quality of life (QoL) related to all forms of leishmaniasis. Secondary research questions included: i) is there evidence that inflammation resulting from VL is directly associated with MI?; ii) is there evidence for social stigma against people with VL? If so, is this associated with MI?; iii) do cognitive and physical impairments resulting from VL result in MI or PM?; iv) are stigma and discrimination of patients with post kala-azar dermal leishmaniasis and cutaneous leishmaniasis associated with MI?; and v) do co-morbidities in people with leishmaniasis have an association with decreased QoL, increased MI or PM? From this research, we set out to develop a conceptual model for the association between leishmaniasis, MI, PM and QoL. This model may serve to prompt further research and debate and to inform health workers, government bodies and the scientific community about the nature of and the mental health implications of leishmaniasis and the unanswered questions surrounding these associations. # Methods The review methodology including the search strategy for this review was published in PROSPERO following CRD and PRISMA guidelines. The recommendations in the Cochrane Handbook were adhered to for reporting the review. ## Databases and search strategy The following primary electronic databases were searched after expert advice from experienced information specialists working in this field: MEDLINE, EMBASE, PsycINFO, LILACS, POPLINE, Global Health, IndMED, ArabPsyNet and AfricanIndexMedicus (AIM). There were no restrictions on the date and language of publication. MEDLINE and EMBASE were chosen as comprehensive databases for life sciences and biomedical research. PsycINFO is a robust database which contains psychology-related articles. POPLINE and Global Health are population and public health-related databases. LILACS, IndMED, ArabPsyNet and AIM were searched as they contain literature from geographical locations where leishmaniasis is endemic. The search strategy was designed to be more sensitive than specific (search strategy included all possible terms to answer the broad PICOS as opposed to one with less descriptors) as potentially unknown exposures could exist, and because of the lack of a previous robust synthesis of this topic. Backward and forward citation tracking was performed for included studies to find any relevant studies not included in the databases. After retrieving the results for each database, citations and abstracts were exported to EndNote to remove duplicates and for screening. ## Rationale for the chosen outcomes The outcomes were a broad scope of the psychosocial consequences that can result from leishmaniasis. They were based on three concepts: 1) the biological link between the neuroinflammation caused in VL and MI; 2) the social impact of stigma, isolation and discrimination around CL and PKDL on QoL; 3) any further burden on the patient in terms of QoL or mental health (including impact on their relatives) such as, for example consequences secondary to outcomes related to mental health (e.g. financial or physical co-morbidities). ## Rationale for chosen study designs Cohort studies, case-controls and cross-sectional studies that quantitatively measured psychosocial outcomes such as mental illness or QoL using a validated tool in this population were included. Systematic reviews (including and in addition to the review by Bennis et al 2018) were also included as a mechanism for identifying relevant publication for screening. ## Study eligibility criteria The inclusion and exclusion criteria used to select studies is shown in the protocol, which used an adapted PICOS framework for observational studies. As this study was examining observational studies and not randomised control studies, there was no “intervention” being studied, and instead, the term “exposure” was used. There was also no comparator in observational studies, so this domain was not used. ## Rationale for chosen population Due to the broad exploratory nature of this systematic review, all patients with any form of leishmaniasis regardless of age and gender were included. We also included assessments of how the disease has wider impacts on family and community members due, for example, to social stigma and financial strain. ## Rationale for chosen exposures The exposures included the social determinants of health associated with leishmaniasis including: poverty, low social economic status, co-morbid infections, social and cultural norms, as well as the physical and psychosocial factors that could lead to a decrease in QoL, PM and MI. ## Data extraction Titles and abstracts were downloaded onto EndNote and de-duplicated. These were then transferred onto an excel spreadsheet and screened independently by two reviewers (MP and VC). Any disagreement between these two authors were resolved by a third reviewer (RC). Screening codes for titles and abstracts were also created. After selecting the included studies, a standard form as used to extract the relevant data. Extracted information included: study setting; study design; year; time period; study population; sample size; inclusion rate and attrition (where relevant e.g. for cohort studies); age range of participants; mean age; sex ratio (M/F); outcome (e.g. anxiety or depression); statistical test; measurement tool and main findings. ## Data synthesis and quality assessment A narrative synthesis of the findings from the included studies was performed. This was organised according to characteristics of the studied population; outcome and how these were measured as well as measures of effect. It was expected that the studies would not be similar enough to conduct a meta-analysis due to the heterogeneity of the primary studies included in the review e.g. cutaneous, visceral, PKDL; different countries, different studied populations, different outcome measures and different study designs). ## Quality assessment and risk of bias The tool to assess quality and risk of bias for the cohort and case-control studies was the Newcastle-Ottawa Scale (NOS) as suggested in the Cochrane Handbook. For cross-sectional studies, an adapted version of the NOS was used. For any SRs identified during the selection process, the ROBIS tool for assessing the risk of bias was used. The tool is comprised of three phases: 1) assess relevance, 2) identify concerns with the review process, and 3) judge risk of bias. ## Methodological quality appraisal of studies The methodological quality of the studies was assessed for the 13 primary studies using the NOS risk of bias tool, as described above; the systematic review was assessed using ROBIS. An adapted version of NOS was used to appraise the quality of the cross-sectional studies. The case-control studies and cohort study were appraised using the original version of NOS. ROBIS is a very comprehensive tool used to assess the risk of bias in systematic reviews, and not in primary studies. It was used to assess the risk of bias in the systematic review that met our inclusion criteria. # Results ## Study characteristics A total of 14 publications (consisting of 13 full articles and 1 conference proceeding) met all of the inclusion criteria for this systematic review. All were independent studies, except two where one was the baseline study, and the other the follow-up study, and published between 2004 and 2018. Three studied VL; 10 studied CL and one PKDL. All studies were performed in LMICs. ## Population and demographics Of the VL studies, two were about HIV patients co-infected with VL, which was part of the inclusion criteria for the population. The populations measured in the studies vary greatly in age. Ages of the participants in the studies ranged from the ages of 7 to 80 years, according to the values reported in the studies. Two of the studies were in children only. Most of the studies had both female and male participants although two had an only female population. The combined number of study participants in the 14 studies was 2565. Sample sizes of the primary studies ranged from 20 to 590. Female to Male ratio in percentage of female was 57% showing that more women were studied over-all (excluding the review). All studies, including the systematic review were performed in low-and middle-income countries (LMICs). ## Study design Nine studies were cross-sectional, three were case-control (, one was a prospective cohort study and one a systematic review. The characteristics of the 14 included studied that met the eligibility criteria were assessed for quantitative synthesis. ## Diagnostic criteria The different instruments to measure the outcomes for each of the 14 publications are shown in. ## Outcomes The outcomes mentioned in this section are summarised below in. ### Visceral leishmaniasis Two studies by the same authors conducted in Northwest Ethiopia reported the results of a prospective longitudinal study at baseline and followed up six months later, measuring QoL in HIV patients with VL (HIV-VL) and HIV patients. At baseline, HIV-VL patients had lower mean scores in all domains of the QoL questionnaire showing poor quality of life. Importantly depression was strongly and consistently associated with all the QoL domains in HIV-VL patients (as it was in the HIV group). The mean (SD) depressive-symptoms scale scores were higher 2.67 (±0.7) for HIV-VL patients compared to HIV patients 1.61 (±0.5) (p = 0.001). Mean scores for social relationship among co-infected patients were significantly lower compared to the HIV group (p = 0.001). After 6 months of treatment with both antiretroviral treatment (ART) and anti-leishmanial drugs, the follow-up study, showed there was improvement in all the QoL domains analysed at baseline in both groups. The social relationship score (one domain of the WHO-HIV-BREF) was calculated according to how the participant felt in terms of being accepted by the people they know, about their sex life, if they had support from friends, and if they were satisfied with their personal relationships. Another study looked at knowledge attitudes and practices (KAP) about VL among adults in a community in India. It was found that 4.7% of the participants agreed with the statement that the incidence of VL in the family should not be disclosed. Forty-three percent reported that the illness affects mental health, causes stress (27%), irritation (3.7%), depression/fear of death (5.8%) and other (16,9%). Almost 74% thought that VL in the family has financial consequences, and reported that it caused impoverishment, led to the need for loans, and that they could be forced to sell property. ### Cutaneous leishmaniasis The cross-sectional study by Chahed et al, 2016, measured the QoL in girls and women with CL scars and explored the psychological and psychosocial consequences of CL using the Revised Illness Perception Questionnaire (IPQ-R), World Health Organization Quality of Life-26 (WHOQOL-26) and the Psoriasis Life Stress Inventory (PLSI) in 41 girls and women with CL scars in the Sidi Bouzid region, Tunisia. The correlation analyses performed on inter and intra-subscales showed that the emotional representations associated with CL were correlated with a loss of self-esteem and feelings of inferiority (r = 0.77, p\<0.05). The more patients knew about CL, the more pessimistic they got about the prospects of recovery. Patients who had a more coherent perception of CL had stronger emotional reactions (p\<0.001). Moderate correlations were found between the total number of body scars and experiences of rejection (r = 0.31, p\<0.05). The number of body scars had a strong link to experience with stigma. Experiences of rejection and avoidance of stress negatively correlated with age (r = -0.33, p\<0.05, and r = -0.31, p\<0.05), suggesting that younger women experience social stigma more than older women. The WHOQOL-26 and the PSLI questionnaire results showed that the domains of Social QoL and anticipation avoidance of stress, and social QoL and total stress correlated significantly (r = -0.36, p\<0.05 and r = -0.32, p\<0.05). The Mental QoL domains, however, did not correlate significantly with any of the PSLI domains. QoL was measured using the Dermatology Life Quality Index (DLQI) in a cross- sectional study of 20 patients with CL. In 70% (n = 14), CL resulted in a moderate to large effect on QoL in the “work and school” domain, and the “symptoms and feelings” domain followed. The domain with the least impact was “personal relationships”. Handjani and Kalafi, 2013 measured the impact of dermatological diseases on the quality of life of healthy families of patients with skin diseases which included five patients with CL using the 10-item validated Persian version of the Family Dermatology Life Quality Index (FDLQI) questionnaire. The FDLQI scores for each of the groups showed that there was no statistically significant difference in QoL of families found between the groups of different skin diseases studied (vitiligo, psoriasis, pemphigus and leishmaniasis). However, due to there only being 5 CL families in this study, statistical power is too low to draw meaningful conclusions. The WHOQOL-BREF was used to study QoL, PM and MI of CL patients. The psychological and environment domains had the lowest median scores. Forty (90.9%) interviewees presented negative feelings (blue mood, anxiety, despair, depression). Of these, eight (18.18%) reported experiencing such feelings always, 19 (43.18%) very often, nine (20.45%) quite often, and four (9.09%) rarely. 50% were dissatisfied with the support received from family and friends, and in their intimate lives. Layegh et al., 2017, measured the QoL, depression, and anxiety (MI) in children with CL using the Children's Dermatology Life Quality Index (CDLQI), Children's Depression Inventory (CDI), and State-Trait Anxiety Inventory for Children (STAIC) questionnaires. This study enrolled 42 children by convenience sampling. The prevalence of low quality of life, state anxiety, and trait anxiety was 57.1, 76.9, and 15.8%, respectively; 32% of patients had depression. Cases of low quality of life (54.1%), state anxiety (56.6%), and trait anxiety (53.8%) were more common with the acute form of leishmaniasis. Low quality of life (70.83%), state anxiety (76.66%), trait anxiety (83.3%), and depression (84.6%) were more prevalent in females. The face was the most common location of involvement in patients with low quality of life (63.3%), state anxiety (70.4%), trait anxiety (83.3%), and depression (54.5%). However, the authors report that no significant difference was found between psychological factors in patients and sex (p \> 0.05), acute or chronic type of disease (p \> 0.05), presence of any other skin or systemic diseases (p \> 0.05), location of lesions (p \> 0.05), number of lesions (p \> 0.05), and duration of involvement (p \> 0.05). No information is available about the statistical tests used and with only 42 subjects in total, the power to detect differences or associations is very low. Simsek et al., 2008 conducted a cross-sectional study to assess mental health disorders in women in the region using the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I).0.5). Fifteen out of 270 women had cutaneous leishmaniasis (6.1%) of whom 8 (53.3%) had a mental disorder. Women with CL have a 2.13 higher OR (95% CI: 1.25–7.31) compared to women without CL of suffering from a mental disorder mainly depression and anxiety. Turan et al., 2015 assessed the psychiatric morbidity and QoL in children and adolescents with CL and their parents using the Child Depression Inventory (CDI), the State-Trait Anxiety Inventories for Children (STAIC) and the Pediatric Quality of Life Inventory Parent and Child Versions (PedQL-P and C, respectively). Fifty-four subjects, of mean age 12.0 (±3.2) years of whom 46.3% were female, were matched with 40 healthy controls of mean age 11.5 (±2.3) years, 50% of whom were female. In addition, the mother or other caregiver was included for each child. From 2011 to 2013, subjects were recruited at the paediatrics department of a hospital in Sanliurfa, Turkey and the controls were children receiving vaccinations or undergoing routine health checks, matched for age, gender, and parents’ level of education. Scores for depression were higher in patients compared to the controls and QoL was lower in patients and their mothers. All results were statistically significant (p\<0.001 for CDI and p\<0.05 for PedQL-P and PedQL-C). However no statistically significant difference in scores for anxiety (STAIC) was found. Yanik et al., 2004 looked at the psychological impact of CL using the Hospital Anxiety Depression Scale (HAD), the Body Image Satisfaction Scale (BIS) and the Dermatology Quality of Life Scale (DQL) to measure anxiety, psychosocial morbidity and quality of life, respectively. Ninety-nine subjects in 3 equal groups, those with active lesions (n = 33), patients with healed lesions (n = 33) and a healthy group (n = 33) were studied. Results showed that lesions on the face and hands, disease active for over a year, permanent scars and social stigmatisation led to anxiety and depression. Body image satisfaction and quality of life were also decreased. Higher scores were obtained in patients with active CL and scores were also statistically significant in patients with healed scars compared to controls in HAD and BIS. Patients with active lesions had lower QoL scores compared to those with healed scars. The correlations between the subscale of HAD and DQL showed a moderate correlation (anxiety r = 0.490, p\< 0.001; depression r = 0.291, p = 0.040). The comparison between the HAD scale and the BIS scale had a significant correlation (anxiety r = 0.201, p = 0.047; depression r = 0.256, p = 0.011). The quality of life in patients with CL using the Dermatology Quality of Life (DLQI) Index was carried out. QoL was significantly affected. Highest scores were seen in the symptoms and feelings domains; the lowest effect was observed in the treatment domain of the DLQI. The appearance of the lesion and type of the lesion significantly affected the QoL (p\<0.05) as patients with ulcerated lesions had lower quality of life compared to those with nodular (P = 0.003) and plaque lesions (P = 0.005). The activity of the disease, location of the lesions and gender did not affect the scores significantly. A scoping review on the impact of localised cutaneous leishmaniasis on psychosocial wellbeing has recently been conducted. Eight quantitative studies five qualitative studies and two mixed-methods studies were included in their review. It combines the results of the quantitative studies through narrative synthesis looking at anxiety and depression, low QoL, stigma and fear of scars. The three last cited quantitative studies were picked up by us during the title screen but did not fulfil the abstract or full-article screen criteria for this review, and were thus excluded (refer to protocol for inclusion and exclusion criteria). Our set of inclusion criteria was different from that of the scoping review carried out by Bennis et al (2018), hence only six studies included in their study overlap with this systematic review. The quantitative study by Abazid et al., 2012 showed no mental health or psychosocial outcomes; An RCT was found that falls under our exclusion criteria. Qualitative studies were not included in our study. Bennis et al did not include three studies on CL that were included in this review. ### Post kala-azar dermal leishmaniasis Pal et al., 2017 assessed the perceptions, stigma and quality of life in patients with PKDL patients compared to healthy controls using the Dermatology Life Quality Index (DLQI) and the 36-Item Short Form Health Survey (SF-36). The type and severity of the lesions were also noted. A range of independent variables were compared. Quality of life was significantly lower in patients under the age of 20 years (p = 0.03) and in those with more severe lesions (p = 0.001). Initiation of treatment for PKDL improved the scores (p = 0.04), while gender, duration and location of the lesions had no impact. The SF-36 showed that mental health, social functioning, body pain and general health were more severely affected in the patients compared to the control group (p\<0.05). ### Risk of bias assessment Based on NOS, 6 of the 14 studies were of good quality, 3 were of fair quality and the remaining 4 were of poor quality. Alemayehu et al 2017 and 2018 both showed the highest scores among all 14 articles. Most studies with poor or fair scores were found to have inadequate sample sizes (too small) with no justification. In contrast, recruitment methods for subjects seemed to be justified for all studies. Out of the nine cross-sectional studies, only five controlled for at least one confounder. Thus, the remaining four could be severely biased due to confounders such as the age or gender of the participants. Moreover, what contributed further to risk of bias was that five of the studies did not report confidence intervals, which is considered inadequate statistical reporting. However, some important inconsistencies between the NOS and the adapted-NOS, as well as the different study designs, could potentially have had an impact on the overall score and quality rating. The maximum rating that can be given in the adapted NOS-version is 10 stars whereas for the original NOS versions (cohort and case-control studies) Quality rating does not take into account the overall star rating, but rather the rating per section: “selection”, “comparability” and “exposure”. The comparability section was always assessed in terms of whether the study controlled for specific confounding variables. Gender and age were considered the most important variables to control for. The ROBIS (– Tables) assessment showed that in the review by Bennis et al 2018 the risk of bias is unclear overall. As shown in the assessment, the lack of a previously published protocol made it very difficult to assess the reliability of the review. # Discussion The primary research question of this review was: *What is the impact of leishmaniasis on mental health and psychosocial wellbeing*? This systematic review shows preliminary evidence of associations between leishmaniasis and mental health, but also shows several lacunae in the evidence. There are different manifestations of the different forms of leishmaniasis and it would be important not to conflate these. We have brought all forms together in one review in the same way as previous authors have done in relation to descriptions of clinical treatments, epidemiological findings and burden of disease, and clinical practice guidelines. This affords us an opportunity to assess the current state of research and see where the gaps are. The systematic review only yielded studies about VL, PKDL and CL. The first evident large gap in the existing literature is the lack of quantitative observational studies on the impact of mucocutaneous leishmaniasis on mental health. All 14 studies found examined at least one of the mental health outcomes (QoL, PM or MI). Of these studies: nine found significant associations between leishmaniasis and QoL; seven found significant associations with MI; and six with PM. Several of the included papers appear to show an association between all forms of leishmaniasis and depression. Mechanisms for this are likely to be complex and interwoven as with other NTDs but evidence to date suggests that scarring in CL has an increased association with social and family rejection and anxiety/depression and severe nodular lesions also carry an association with depression. The effect may be more severe in younger patients and women/girls although this needs to be investigated further. Lesions on the face have an increased association with anxiety/ depression. Evidence for a societal link was found between VL and mental health outcomes (MI, QoL, PM). Survey work assessing the perceptions and attitudes of the community in Bihar, India, found that there is social stigma attached to VL, as well as PM MI, and decrease in QoL due to financial loss. While this supports the hypothesis that financial burden may also lead to decreased mental well-being, our study did not find more evidence associating VL with financial problems and social drift. A more complete study would involve economic evaluations to be part of our inclusion criteria and require different search terms. When looking at subtypes of leishmaniasis more specifically, there are several gaps in the literature concerning the effect of VL on mental health outcomes, despite it having been shown that VL did decrease the quality of life of patients directly. We found no evidence that neuroinflammation or inflammatory responses play a role in the development of mental health problems associated with leishmaniasis as this had not been specifically explored in this population. This is a surprising gap in the literature given that inflammation associated with other diseases has been shown to be associated with mental health (in a bidirectional manner), although these reported diseases and syndromes are all chronic. Neurological manifestations can occur in VL, which can both decrease quality of life and increase the likelihood of mental illness in Leishmaniasis. Only one review (non-systematic) looks at the neurological and psychological consequences of visceral leishmaniasis in humans and animals but the only article that documents mental health outcomes in human VL patients was an article published by Carswell in 1953 which did not meet our inclusion criteria and included no validated assessments of mental health outcomes. The subject of stigma in VL was only examined indirectly in a single study from the 14 included studies. This showed a reluctance in the community to report the occurrence of the disease, indicating fear of the negative consequences this would have upon the patient and their families. The association of VL and MI was, however, indirectly examined, where 43% of the community perceived that VL could cause changes in mental health, although no patient outcomes were reported. The prospective longitudinal study from Ethiopia addressed the question partly because it was related to HIV-VL co-infected patients and not just VL infection. This study reported a higher mean score for depressive-symptoms compared to HIV-patients alone. However, it would be speculative to conclude that this was due to VL alone and it is possible that any other co-morbidity could exacerbate the symptoms of MI. Carswell reported "mental depression” and “apathy” in all of 96 (100%) patients of VL, but no clinical or diagnostic measures were used and these findings have therefore not been tested in more methodologically robust research using validated tools to assess MI, QoL or PM. Bennis et al., 2018 focuses on the stigmatising impact that localised cutaneous leishmaniasis can have, by dividing “stigma” into the categories of social stigma (when society rejects or excludes against people even if the stigmatized disagree with the way they are being treated) and self-stigma (the internalised mechanism within the person who is being stigmatized who faces rejection in an anticipatory manner). Even though Pal et al 2017 do find a significant decrease in quality of life in the personal relationship domain, one cannot jump to conclusions about whether this was related to stigma, and whether PKDL patients suffered from MI. CL patients are victim to social stigma, self-stigma and suffer from PM, MI and decreased QoL. In Chahed et al., 2016, the number of body scars was weakly correlated to an experience of stigma (p = 0.06, r = 0.29). They also observed anticipation avoidance of stress, indicating not only social stigma but also self-stigma. These results should be interpreted with caution, because the sample size was small (n = 41). The study showed stronger correlations between CL and loss of self-esteem and feelings of inferiority (r = 0.77, p\<0.05), and it was shown that at younger ages, women experienced higher levels of rejection and avoidance of stress. Bennis’ review included and reported on the results of qualitative studies (not included in the study design for this systematic review) showing: social isolation; social contempt; social exclusion; marriage difficulties; embarrassment; shame; sadness; disgust; shyness; and decreased marriage prospects. The scoping review concludes that stigma is closely linked to psychosocial morbidity. Although PKDL and CL are caused by different *Leishmania* species, both have visible and disfiguring clinical manifestations. The parasitic aetiology itself does not create any common links but these are mediated by factors such as stigma, which whilst operating in different ways depending on the symptomatology and the local culture, beliefs and attitudes, never-the-less allows us to consider an important cause of mental health problems with all its various pathways. Stigma was found to play an important role in the mental health outcomes associated with CL and PKDL but the measurement tools used in the quantitative observational studies (part of inclusion criteria) could not quantify how much stigma a person or their family faces. Several generic validated tools to measure stigma among NTD patients are readily available in the NTD Morbidity & Disability (NMD) Toolkit. Bennis et al., 2018 call for the development of a standardized tool to measure stigma in CL. Overall, the clinical manifestations for PKDL and CL both led to decreased body satisfaction as well as misconceptions within society about potential disease spread. These findings complement those of Bailey et al, in their recent systematic review of the psychosocial impact of CL. They used the data from qualitative and quantitative studies to demonstrate a high burden of co-morbid depression in both active and inactive forms of the disease. The only studies including a co-morbidity were part of the same study assessed at baseline (cross-sectional) (prospective cohort) where the impact of VL as a comorbidity of HIV was assessed pre- and post-treatment with anti-parasitic treatment for leishmaniasis. The baseline study shows that HIV-VL patients showed significantly worse quality of life scores in all domains compared to the HIV-alone patients (p = 0.001). Regarding psychosocial morbidity and mental illness, the Kessler Psychological Distress Scale correlated significantly with psychological health, social relation, and environmental domains of the WHOQoL had correlation coefficients of -0.335, -0.295, and -0.350 with the Kessler scale (p = 0.001), further showing that a decrease in quality of life correlates with an increased mean psychological distress score. The mean (SD) depressive symptoms scale score was higher in HIV-VL patients compared to HIV patients, 2.67 (±0.7) vs 1.161 (±0.5) respectively. After therapy, at the 6-month follow- up stage, HIV-only patients showed no significantly different QoL scores compared to HIV-VL patients, showing that QoL of life scores improve after the disease is treated. Usually patients return to their normal physical appearance, as opposed to the other forms of leishmaniasis. This is starkly different in comparison to people who have CL and PKDL lesions. When the cutaneous lesions are healed, these become visible, disfiguring scars on the patient. Studies have shown that the suffering is long-term with the consequences of the disease for life. We used the results of the review to develop a conceptual framework outlining the relationships visually between the forms of leishmaniasis, mental illness, psychosocial morbidity and quality of life. This shows the strength of evidence or the absence of evidence where there are hypothesised links. This takes the form of a model that may help to direct future research and public health policy. ## Limitations The variability in outcome measures and study designs contributed to large inter-study heterogeneity. Therefore, a narrative description of the results was the only option. Even if the outcome measures had been more consistent or if more papers had been identified, the lack of consistent good quality studies, would have hindered the conduct and interpretation of any meta-analysis. This highlights the need for better literature in this area. The searches and inclusion criteria for this review did not include qualitative studies which could also have revealed useful information regarding the studied outcomes. However, the large body of evidence in existing qualitative studies for CL compliment the results observed in the included quantitative studies. Feelings of shame, social and self-stigma, isolation, depression, anxiety and even suicidality among patients and their family have been well documented in qualitative studies. ## Future research Further research is recommended with larger sample sizes. Research should include both mixed-methods and qualitative methods, as it is only through combining quantitative and qualitative measures of mental health outcomes that we will begin to understand the dimensions of the problem at hand. There has been no research to follow-up on Carswell’s observation in 1953, that most patients with VL showed marked signs of depression. Thus, a more sophisticated, and modern large cross-sectional study with VL patients to study if there is associated mental illness in these patients would be an important aspect to investigate. This research needs to be designed in culturally appropriate ways given the variety of low- and middle-income country settings that the disease is to be found. ## Clinical and policy implications Governments of endemic countries could invest more in research to find out exactly what is the best way to intervene with the CL and PKDL mental health crisis at hand. There is evidence that stigma is a serious problem associated with leishmaniasis, and for this reason policies addressing informational gaps and misconceptions may be a good way forward. Interventions ranging from information-based interventions, to contact between persons with the condition and the community, health professionals, or others, using agents of change or popular opinion leaders (e.g heads of village), peer counselling, and skills building and empowerment are all interventions that have proven effective in the control stigma for NTDs (leprosy) and other highly stigmatising conditions such as HIV and TB. Women and children were found to be impaired significantly in these studies. Women’s health is of paramount importance in public health. There is a need for effective promotion of good mental health for women and children. Hence good quality evidence-based research and practice would add considerably to this field. In the LMIC settings, where VL is endemic, there is very little investment in mental health. If VL has a causative link with depression it may be going undiagnosed. If so, this could (after treatment) have a detrimental impact on the patients’ quality of life, and wider societal and economic impacts. # Conclusion This wide exploratory systematic review has shown that there are substantial gaps in our knowledge and in the research literature and that there is a lack of methodological quality in many of the existing studies. This systematic review shows preliminary evidence that leishmaniasis impacts upon individuals and families affecting their social status, causing stigma, with effects on quality of life and raising the risk of mental health problems. This work allows us to build a preliminary model that we present here to scaffold future attempts to better understand the effects of leishmaniasis on MI, PM and QoL. # Supporting information [^1]: The authors have declared that no competing interests exist.

# Introduction Visual field defects are often caused by damage to early visual areas in the brain’s occipital cortex and a common consequence of stroke or trauma. Damage in these post chiasmatic areas of the brain usually lead to partial loss of vision—affecting approximately 12% of patients suffering from traumatic brain injury and 35% of patients suffering from stroke. The extent of the visual field defect coheres with location and dimension of the cortical lesion. Most common forms are blindness in half (hemianopia) and in a quarter (quadrantanopia) of the visual field. “Visual Restitution Therapies” (VRT) postulate that damaged areas in the visual cortex can be reactivated by repeated visual stimulation. By presenting bright light-impulses on a computer-screen in the individual transition zone of the intact and defect visual fields, neurons are stimulated and form new connections among residual cortical structures. Animal studies indicate that damaged neurons in the visual cortex can indeed recover and reconnect by bright light stimulation. However, existing evidence for “training-induced neuroplasticity” in the human primary visual cortex is inconsistent. While some studies found a diminution of the blind field after intervention, other studies found no significant amelioration and attribute improvement after training to compensatory eye movements during perimetric pre-, post-, and follow-up assessments. To achieve robust and credible evidence for the effects of interventions based on the principles of VRT and for the issue of training-related neuroplasticity within the human visual cortex, we developed an eye-tracking based perimetric methodology (“Eye Tracking Based Visual Field Analysis” \[EFA\]). The EFA is based on static automated perimetry and additionally compensates individual eye movements during diagnosis in real time while providing a standard error of measurement of 0.44° of visual angle. This eye-tracking based feature precludes that the diagnosis is unduly influenced by (involuntary) compensation strategies such as (micro) saccades during pre-, post-, and follow-up diagnosis. Study results from patients suffering from cortical lesions (and glaucoma) further indicate that the EFA is applicable for clinical use. Thus, the unprecedented accuracy of the EFA and its insusceptibility to compensation strategies makes the methodology an ideal choice for upcoming investigations regarding results from studies, which found significant amelioration after intervention with VRT based approaches, ascribing post-intervention perimetric changes to the successful stimulation of neuroplasticity in the visual cortex. In order to realize an evaluation study of VRT based approaches, we sought to develop an ambulant training instrument. The existing concepts and available instruments are coded for Personal Computers and in this way additionally require the purchase of a chin and head rest. This makes therapy inflexible as it is place-bound to a stationary personal computer. Therefore, our newly developed VRT based instrument should be—besides being reliable—portable and easy to use in order to instill high compliance in the trainee. We envisioned that these requirements could be met by implementing the concept of VRT as a Virtual Reality (VR) application and named the device “Salzburg Visual Field Trainer” (SVFT). The advantages over existing VRT systems is that SVFT is compact, portable and comfortable to wear, making the training device ideal for both randomized controlled trials (RCT) and clinical usage during rehabilitation. In combination with the EFA, which provides highly accurate and reliable visual field assessment, another crucial aspect for experiments with the SVFT regarding potentially ameliorating effects of VRT is enabled. That is the unprecedented exact positioning of training stimuli on the patients’ individually shaped border areas. In the following, we describe the device and present a proof-of-concept study of its reliability, which was assessed in a sample of 40 healthy participants. # Materials and methods Patients or participants were not directly involved in the design of this study. Written informed consent was obtained from all participants who received ECTS credits or 10€ for participation in the study. The study design was approved by the ethics commission of the University of Salzburg (Reference No. 39/2018). ## Hardware The SVFT was developed using Unity® with a Google® Cardboard solution, consisting of ZEISS® VR ONE Plus goggles (Carl Zeiss AG, Oberkochen, Germany) in combination with Celly® Bluetooth® remote controls (Celly S.p.A., Vimercate, Italy) and Samsung Galaxy S9® smartphones (Samsung, Seoul, South Korea), running on the Android® Pie operating system. ZEISS® VR ONE Plus houses two aspherical biconvex lenses with +32.5 diopter, enabling the patient to accomodate on the smartphone display, which is inserted in a front tray of the head mounted system in a distance of 44 mm. From our clinical experience the device is wide enough to be worn with most correctional glasses. ## Functionality Similar to conventional VRT software on Personal Computers, the patient fixates her/his gaze on a fixation cross in the middle of the screen. No prior ocular calibration or manual diopter adjustment is necessary before using the ZEISS® VR ONE Plus goggles (Carl Zeiss AG, Oberkochen, Germany). Bright light stimuli with approx. 1,000 cd/m² and a size of 3° appear for 750 ms in randomized order on a dark background. The presentation of each stimulus is interrupted for 2000 ms (plus a randomized duration of up to 1500 ms) to give the patient time to react to the displayed stimulus via remote control (see below). Training time duration and all other settings mentioned above can be adapted by qualified personnel, allowing a flexible optimization of the therapy setting. Patients’ behavioral feedback is necessary as the theoretical foundation of VRT states that stimulus presentation should lead to reconnection of lesioned cortical areas, resulting in gradual amelioration of damaged visual field areas. Therefore, the SVFT runs a stimulus-location adaptation process during rehabilitation, based on user feedback by clicking (or not clicking) the Bluetooth® remote control during the break after every displayed stimulus. If the user detects the stimulus peripherally, its location is moved towards a custom reference point within the central area of the individual visual field defect. This allows the VRT system to automatically adapt to an improving visual field or small inaccuracies from minor gaze deviations from the fixation cross. On the contrary, if the stimulus was not detected, it is moved away from the reference point towards the center of the patient’s intact visual field, balancing out the therapeutic process in case of inaccuracies, misclicks or loss of already gained training progress. Thus, this behavioral procedure produces a steady approximation of the “real” border across the intact and defect visual field and compensates for minor gaze inaccuracies of the patients during the training procedure. A potential additional benefit of this behavioral procedure is that by pressing a button when concentrating on the peripherally displayed stimuli, attention during training may be further improved as the rehabilitation process becomes an interactive task. In the course of rehabilitation of patients with visual field defects based on restitutive methods (the currently implemented rehabilitation procedure in the SVFT), patients are not allowed to wear glasses during training. This is because glasses potentially distort the principle of restitution therapy due to their specific refractive indices. In other words, due to the wearing of glasses, the light rays from presented stimuli falling into the retina are potentially deflected in such a way that the intended areas in the visual cortex are no longer "hit", even if the patient reliably looks exactly at the fixation cross. However, other approaches—such as so-called "compensatory approaches"—can be performed with glasses. The ZEISS model, which we used in the current study, is specially designed for this purpose—as indicated by the manufacturer and which we can confirm from our clinical work with patients. The entire input required during the training process can be controlled from the Bluetooth® remote control (consisting of one pressable button), allowing the smartphone to be permanently fixed to the ZEISS® VR ONE Plus socket. Through a side opening, the smartphone can also be charged while it is inserted in the virtual reality goggles. With a double click on the remote control, the SVFT application can be started directly from the smartphone’s standby mode. Then, one single click is required to start training. Subsequently, single clicks during training act as patients’ behavioral feedback, which is required for the stimulus adaptation procedure (see below for details) during rehabilitation sessions. When training is complete, SVFT closes automatically, allowing the smartphone to fall back autonomously into standby mode. Individual stimuli adaptation progress is stored by the SVFT and retrieved when the next training session starts. Optionally, progress is documented but changed stimuli coordinates are deleted after training and the initial coordinates positions are used for the next training session. These training settings—like stimuli display location and duration, reference point location, break duration between stimuli and total training duration—are fully customizable. Additionally, based on individual diagnosis, parameters like stimulus or fixation cross size and brightness are adjustable to the needs of the patient. To allow researchers and healthcare specialists tracking the training progress of patients, SVFT periodically saves the current training progress into a protocol file. This is done multiple times—which is customizable—during (in case the training is cancelled) and at the end of each training session. For each training session, a new file is created including time and date, overall session runtime and the number of feedback clicks on the remote control. These parameters allow researchers and healthcare professionals to track technical problems, as well as user compliance and general behavior during the training sessions. Additionally, the current locations of the training stimuli are saved, facilitating comparisons with previous sets of stimuli locations and therefore investigations into the patients’ therapy progress are enabled. ## Technical evaluation The crucial point of a VRT based training device is a reliable and precise presentation of the training stimuli across the individually shaped border area in the patients’ visual field. This central aspect depends on two factors. First, the patient is required to fixate continuously on the central fixation cross in order to perceive the presented stimuli correctly from a parafoveal position. Because patients expect a therapeutic effect from training with SVFT, they are—based on our clinical experience—usually highly motivated to conduct the training procedure as accurately as possible. Second, the individual characteristics of the patients’ visual field defects must be assessed as precisely as possible and—most importantly—presented stimuli must be placed exactly across the patients’ individual border area during the training procedure. This crucial factor is achieved as follows: In a first step the patients’ visual fields are diagnosed with the help of the EFA. Due to its eye- tracking-based adaptation process, highly accurate data regarding the patients’ perimetric status can be obtained. Thus, this enables—in a second step—individually custom-tailored positioning of test stimuli in the SVFT, based on every patients’ respective visual field properties and the resulting individual characteristics of the border area between intact and defect visual field. In principle, visual field diagnosis can also be performed with other—comparably exact (portable or stationary) diagnostic instruments. ## Experimental design To investigate whether training stimuli are displayed correctly in their intended locations in the SVFT, we conducted a proof-of-concept study with a group of 40 healthy participants. Accurate displaying of training stimuli in the individually shaped border area of patients is crucial for the potential therapeutic benefit of VRT based approaches. Therefore, we defined the blind spot—a naturally occuring scotoma in the human visual field—as an objective marker for the degree of precision that SVFT displays stimuli in the patients’ visual fields. This experimental approach is based on the blind spot fixation control mechanism in clinically established automated perimeters (such as the Humphrey Field Analyzer from ZEISS). In these devices the retinal natural blind spot serves as an indicator for patients’ reliability to maintain central fixation during visual field diagnosis. This control mechanism is based on the assumption that if a patient reacts too often to stimuli presented in his/her blind spot, the perimetric assessment needs to be repeated in order to ensure validity. Because humans are unable to detect stimuli presented in their blind spots (due to missing retinal receptor cells), patients’ reactions to such displayed stimuli suggest low compliance rate, stemming from unreliable central fixation and thus leading to biased visual field test results. The blind spot is classified as an absolute scotoma and is located between approximately 12° medial, 17° lateral and 1° below the horizon in a temporal location from the fovea centralis. Using the blind spot as a marker of visual field measurement has a tradition in visual field research and the German Ophthalmological Society (DOG) states that the blind spot should serve as a referential scotoma in every perimetric examination. We already used this experimental paradigm successfully to investigate and show the high accuracy of the EFA. We evaluated stimulus presentation monocularly in the SVFT by displaying 15 stimuli in the participants’ individual blind spot locations (“blind spot stimuli”) as diagnosed by means of the EFA and placed another 85 stimuli in the rest of the normal-sighted visual field (“detectable stimuli”). The 40 healthy participants had a mean age of 23 years (SD = 3) and suffered from no reported eye diseases or other clinically ophthalmologic or neurologic issues. Participants were excluded from the study if dioptre was greater than +3 dpt or if other clinical conditions were reported. Our hypothesis was that if the SVFT displays stimuli as intended participants should be able to see (and react accordingly) only to detectable stimuli, but not to blind spot stimuli, which they naturally cannot see due to their presentation in their absolute scotomas. All stimuli had a size of 0.33° and were displayed for 200 ms, following a time period of 2000 ms to react to the stimulus. A break, randomly ranging from 0 to 1000 ms, followed subsequently. Then the next stimulus was presented. The color of the fixation cross changed randomly after a time period of 3000 ms to 7000 ms. We implemented this color change because we learned from previous experiences that color shifts facilitate fixation of the cross during visual field diagnosis. A similar effect was also reported by Steinman. Similarly, as in automated static perimetry assessment, we instructed the participants to react to a set of peripherally presented stimuli via Bluetooth® remote control, while looking steadily at the central fixation cross of the SVFT. Based on the visual field assessment with the EFA, we configured the location of the 15 “blind spot stimuli” in the SVFT individually for every participant, so the stimuli would fit exactly with the previously assessed participants’ blind spot central location. This means that an “optimal” SVFT test run should result in 85 true positives of “detectable stimuli”, 0 false positives of “blind spot stimuli”, 15 true negatives of undetectable “blind spot stimuli” and 0 false negatives of “detectable stimuli”. After trial completion, individual participants detection results were automatically stored in a data file on the smartphone, allowing calculation of results regarding sensitivity, specificity, positive predictive value, negative predictive value, hit rate, random hit rate and RATZ-Index for the SFVT. A total number of 150 stimuli were randomly presented to the participants. The first 25 stimuli were used to accustom the participants to the experimental procedure. Therefore, no “blindspot stimuli” were presented in this block and data was excluded from analysis. The last 25 stimuli were used as a buffer for saving results in the SVFT, hence no “blindspot stimuli” were displayed in this phase of the experiment either and data was also excluded from the analysis. In consequence, testing with the SVFT resulted in a mean total duration of 6 minutes and 45 seconds. Because the EFA and the SVFT are based on two different 2-dimensional coordinate systems, visual field data results from the EFA—which is the basis for the correct placement of the stimuli in the patient’s visual field—were converted into SVFT specific dimensional properties. Based on the users’ viewing distance to the display screens in the EFA (400 mm) and the SFVT (44 mm) (1 retinal angle degree being 1.54 angle degree in the SVFT, based on specifications of the ZEISS^® VR ONE Plus goggles) every stimulus (s) location (x and y coordinate from the EFA diagnosis) can be converted from the EFA to the SFVT as follows: $2*arctan\left( \frac{s}{2 \times 400} \right) \times 1.54$. # Results In total, 4000 of 6000 displayed stimuli were used for calculations (see above for details); in total 588 stimuli (mean of the subject = 14.7; SD = 0.6) were recorded as true negative (TN), 28 stimuli (mean of the subject = 0.7; SD = 1.4) as false positive (FP), 12 stimuli (mean of the subject = 0.3; SD = 0.6) as false negative (FN) and 3372 stimuli (mean of the subject = 84.3; SD = 1.4) as true positive (TP). This translates to 14.7% correct non-reactions (optimum: 15%), 0.7% false reactions (optimum: 0%), 0.3% false non-reactions (optimum: 0%) and 84.3% correct reactions (optimum: 85%). The calculation of accuracy (ACC) results in a value of 99.0% $\left( ACC = \frac{(TP + TN)}{(TP + FP + TN + FN)} \right)$. shows individual results, that is sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), hit rate (HR), random hit rate (RHR) and RATZ-Index (RATZ) of every participant (*SUBJ*). SEN denotes the ratio of stimuli correctly identified as “blind spot stimuli” (true negatives). SPE is the ratio of stimuli correctly identified as “detectable stimuli” (true positives). PPW is the probability that a “blind spot stimulus” is correctly identified and NPW is the probability that a “detectable stimulus” is correctly identified. HR denotes the ratio of correct classifications in “blind spot stimuli” and “detectable stimuli”. RHR describes the probability to respond to “blind spot stimuli” by chance. RATZ denotes the increase of HR compared to RHR. Following, the sample’s mean values and standard deviation for every category—based on results from - are illustrated in. Additionally, Spearman-Brown formula was used to assess the (split-half) reliability of the SVFT. Results of true negative and false positive values indicate a good reliability with a correlation of.547 and a Spearman-Brown coefficient of.707. # Discussion In the present study we validated the visual presentation method of our newly developed virtual reality-based rehabilitation methodology “Salzburg Visual Field Trainer” (SVFT), which is based on the principles of “Visual Restitution Therapies” (VRT). Results of this proof-of-concept study indicate that the SVFT and our developed methodology of translating perimetric test results from the “Eye Tracking based Visual Field Analysis” (EFA) to the SVFT is highly reliable and accurate, resulting in precise display of visual stimuli in predefined areas of the users’ visual field. This process is crucial for a potential therapeutic benefit of VRT based instruments, as such methodologies and rehabilitation procedures must be designed to display light impulses as accurately as possible across desired retinal areas (and hence stimulate the appropriate “residual” cortical areas) in the patients’ visual fields. Thus, our data indicates that we created a feasible and exact virtual reality training device and procedure that can help both to facilitate therapy for patients and contribute to the ongoing scientific discussion regarding the “true” effectiveness of VRT. Furthermore, the results indicate that the participants adhered to the experimental instructions and looked principally at the central fixation cross during the testing. Because other neuropsychological approaches (such as compensatory based rehabilitation) also depend on exact and precise presentation of training objects and stimuli in the patients’ visual fields, our results indicate that the technical setup of the SVFT can be used for these methodologies as well. It must be noted that the clinical use of the SVFT is, of course, strongly dependent on the individual clinical disorders of affected brain-damaged patients. Since the SVFT is based on a behavioral methodology, the ability to understand and execute the therapeutic instructions in principle is a basic prerequisite for training with the neuropsychological instrument. These aspects must be considered individually before any training with the SVFT. In the present proof-of-concept study, however, we have focused on an "ideal situation" (with healthy patients) in order to evaluate the accuracy of the instrument as well as the algorithmic procedures behind it. First clinical studies further indicate that the principle of the SVFT works adequately for patients suffering from visual field issues. Compared to traditional personal computer and monitor based VRT approaches, our virtual reality based SVFT device offers a number of advantages—in part already discussed by other research groups—which all potentially increase patients’ motivation and compliance to engage in regularly repeated neuropsychological rehabilitation. (1) The training situation becomes significantly more comfortable and convenient as no chin and head rest is necessary. The patient can sit and move freely and perform training e.g., sitting on the couch while listening to music. (2) The general technical and neuropsychological conditions for the training process itself are further improved. Because the distance between the eyes and the presented stimuli is always exactly the same, head movements or imprecise configurations and adjustments conducted by the patient regarding chin and head rest or PC monitor do not affect the crucial aspect of exact stimulus presentation (across the visual border area). (3) The immersive design of the virtual reality goggles makes it possible to rule out external factors like light conditions or other visual distractions in the surrounding environment, potentially confounding the therapeutic effect. After the experiment, the participants of the study confirmed to us that the virtual reality goggles are indeed very comfortable to wear and intuitive to use. We also received similar feedback from the first patients—suffering from visual field loss—who already trained with the instrument. These patients also exhibited very good compliance (for which the ease of application and the comfortability of the instrument is certainly conducive) for the entire duration of the training (up to 6 month). Obviously, a central aspect in VRT-based methods and other neuropsychological approaches is user compliance. Patients need to steadily fixate the centrally presented cross in order to enable visual stimulation of respective cortical areas from peripherally displayed stimuli. Because in the current version of the SVFT there is no technical possibility to control for central fixation during the experiment, we emphasized the significant importance of maintaining a steady fixation to the participants. We were aware that there are some behavioral approaches to “control” for central fixation, such as naming out loud changes of the color of the fixation cross. However, we decided against this concept in our proof-of-concept study, because we did not want to distract participants unnecessarily during the experiment. Furthermore, based on our clinical experience, we can state that these control methodologies do not prevent patients reliably (if they really plan to do so) from performing saccades in order to detect more stimuli. Instead, we concentrated strongly on participants’ compliance and so we chose only undergraduate psychology students, who we educated extensively about the experimental design and the importance of complying with the experimental instructions. Results from the proof-of-concept study indeed indicate that the participants’ compliance was high and they acted during testing as instructed. Furthermore, our data show that no participant was “trigger happy” or the contrary, indicating both high compliance during the experiment and high validity of the SVFT. 95% of all participants (N = 40) committed fewer than 3 false positive reactions. One participant committed 3 such reactions and only one participant showed—with 8 false positive reactions—conspicuous behavior in this regard. Regarding future patients—who will train with the SVFT to potentially improve their defect visual field—the issue of compliance will tend to be a smaller problem as in the experimental condition with healthy participants. The reason is simply because patients have a strong natural interest in achieving significant positive therapeutic effects. Compensatory actions, such as the performance of saccades, would run counter their motivation to improve their visual field functionality. We also learned from the present proof-of-concept study that we will focus—especially in the beginning of the upcoming clinical trials with the SVFT—not only on patients’ compliance but also on patients’ motivation. This is because patients—in order to achieve a potentially measurable positive therapeutic effect—need to train self-reliantly with the SVFT for 6 months, 6 days a week, twice a day for 30 minutes, which requires a high level of motivation. Therefore, it is crucial to help these patients to keep their training motivation as high as possible. We believe that this can be achieved with regular feedback and contact. Although we know from clinical practice that neurological patients are usually highly motivated to improve their deficits, it is essential to pay attention to the patient’s motivation to carry out the therapy as well as possible. From participants’ feedback we additionally learned that the VR goggles are indeed comfortable to wear and that the functionality of SVFT is easy to grasp. This further indicates that—with the help of SVFT—patients will be enabled to conduct independent rehabilitation in a comfortable and uncomplicated way over a long period of time, which will ultimately be essential in order to state robust evidence for any potential effects of VRT or other neuropsychological interventions. # Conclusion Our findings indicate that the SVFT can help—in combination with our recently developed perimetric methodology EFA (or any other highly precise and reliable visual field assessment methodology or instrument)—to investigate the true effects of VRT based approaches (or other approaches) and the underlying mechanisms of training-related neuroplasticity in early regions of the visual cortex. # Supporting information The authors would like to thank Christian Kebsak for ophthalmic advice and counseling and Luise Zellner & Leoni Bernstorf for participants management, data acquisition, data analysis and support during the experiment. [^1]: The authors have declared that no competing interests exist.

# Introduction Benford’s Law states that, in naturally occurring systems, the frequency of numbers’ first digits is not evenly distributed. Numbers beginning with a “1” are far more common than numbers beginning with “9”—more than six times as frequent. The exact frequency *P* predicted for a digit *d* is given by this formula: $$\begin{array}{r} {P(d) = \log_{10}\left( 1 + \frac{1}{d} \right)} \\ \end{array}$$ Benford’s Law is frequently used in forensic accounting, where a distribution of first digits that is outside the expected distribution may indicate fraud. Research has also shown that it applies to genome data, scientific regression coefficients, election data, the stock market, and even to JPEG compression. We conducted an analysis over five of the most popular social networking websites and found that Benford’s Law applies to the social network structure in all of them. Specifically, the first significant digit (FSD) of users’ friend and follower counts on Facebook, Twitter, Google Plus, Pinterest, and LiveJournal all follow Benford’s Law. Users’ numbers of posts also conform to Benford. To our knowledge, this is the first time Benford’s Law has been applied to social networks. We show that exceptions to this rule can uncover configurations within social media systems that lead to unexpected results. We also show that, for any individual, the distribution of friend counts within his or her egocentric network also follows Benford’s Law. When the expected distribution is violated, it indicates unusual behavior. A preliminary analysis of over 20,000 Twitter accounts showed that the 100 users whose egocentric networks deviated most strongly from the Benford’s Law distribution were all engaged in suspicious activity. We discuss how these results lead to the possibility of Benford’s Law being used to detect malicious or irregular behavior on social media. We also show that it could be used to validate the sampling in social media datasets. # Benford’s Law: Background and Related Work It was astronomer Simon Newcomb who first formulated what came to be known as Benford’s Law in the 1880s. He noticed that books with logarithm tables showed a lot more wear toward the front, where the numbers beginning with 1 were, than in the back toward the 9s. Concluding that numbers beginning with 1 must be more common, he calculated the probability formula mentioned above. Physicist Frank Benford noticed the same phenomenon. He validated the observation by collecting naturally occurring numbers from many sources: the surface area of rivers, atomic weights, and numbers appearing in Reader’s Digest. All values followed the pattern. Although they were not a perfect match, the principle was established. The formula for the law, $P(d) = \text{log}_{10}(1 + \frac{1}{d})$, provides a theoretical distribution of expected first digits, shown in. On the surface, Benford’s Law is quite counterintuitive. Why would numbers beginning with 1 be any more common than those beginning with 9? Nevertheless, the law holds across many variations in measurement. Temperatures that follow Benford’s Law do so regardless of whether they are measured in Fahrenheit, Celsius, or Kelvin. Distances follow whether measured in miles, kilometers, or smoots. Most persuasively, Hill provided a proof for the Benford’s Law in 1995. As a simple demonstration for skeptics, he suggests that they jot down all the numbers that appear on the front pages of several newspapers, or randomly select data from the Farmer’s Almanac. Benford’s Law describes all these naturally occurring sets of numbers, and more. Specifically, some applications of Benford’s Law are more relevant to our work. Benford often applies to systems that follow a power law distribution. Power laws are commonly found in social network structures and social media. Although no one has yet investigated how well Benford’s Law describes social networks (online or offline) or social media, it has been shown to describe online human behavior through price distributions in eBay auctions. # Data, Data Sets, and Collection We analyzed data from five major social networking websites: Facebook, Twitter, Google Plus, Pinterest, and LiveJournal. We collected the number of friends in each network and followers when appropriate. On Google Plus, Twitter, and LiveJournal, we also had access to egocentric network data. For each user, we obtained a list of friends and the count of outgoing edges for each of those friends. With this data, we could analyze the distribution of FSDs for an individual person’s social network On Twitter and Pinterest, we also had access to the number of posts each person made. This provides another interesting insight into the general patterns of behavior on social media and whether Benford’s Law applies. We collected some of these datasets ourselves and used other datasets that had been created by others. The following sections detail our process with each network. ## Facebook We accessed user profiles using the Facebook Graph API with requests for friend counts of a numeric Facebook user ID. Once we accessed a user’s data, we incremented the user ID by 10,000 and make the next request. If there was no data available for a given user, we incremented the user ID by 1 and tried the next person until we found a match. We collected friend counts for 18,298 users. ## Twitter We collected the numbers of followers and “friends” (i.e. people the user is following) the user had, and the number of people each of those friends were following. This allowed us to analyze the distribution of FSDs within egocentric social networks. In addition to this network data, we collected the number of status updates for each user. Although there are existing Twitter social network datasets online, we collected our own data in this project in order to work with non-anonymized users so we could later analyze their account activity. We accessed data via the Twitter API, using users’ numeric Twitter user ID. Our process was to access a user’s data and then increase the user ID by 50,000. This gave us a fairly uniform distribution of users, and we were able to collect the data in a reasonable amount of time (over a few weeks) given Twitter’s API limits. If an a user ID was protected or not linked to an account, we incremented user the ID by 1 and tried the next person until we found a match. Because we considered counts for friends, followers, and status updates, we only included users in our sample who had at least one of each. This allowed us to consider that same set of users for all three attributes. For egocentric network analysis, we only included users who had at least 100 friends so the distribution of FSDs would be measured over a reasonably large sample. Our final dataset had 78,225 users. For 21,135 of these, we also had egocentric networks with friend, follower, and status counts for all the people they were following. ## Google Plus The Google Plus network is part of the Stanford Network Analysis Project (SNAP) datasets, The social network is provided as an adjacency list. We made one pass through the “combined” dataset, counting the number of friends a person had. The network is directed, so we counted outgoing edges. In addition to using the friend counts for each person, we were able to get the friend count for each of their friends, thus allowing us to construct a FSD distribution for each egocentric social network. After processing, we had data for 72,271 users. ## Pinterest The Pinterest data was provided as part of the Social Curation Dataset. We used the “Pinterest User Information” data, which contained follower, following, and pin (i.e. post) counts. After filtering out users with no followers or pins, we had data for 39,586,033 users. ## LiveJournal LiveJournal Dataset is one of the SNAP datasets. We followed the same processing procedure as we used for the Google Plus dataset. After processing, we had data for 4,307,491 users. # Results We found that the distribution of FSD among friends in all five datasets closely followed the values expected from Benford’s Law, with one interesting exception: the Pinterest following relationship. We discuss this later, but we will set this aside for now. shows the distribution of FSDs for each of the six datasets. For Facebook, Google Plus, and LiveJournal friends, Twitter friends and followers, and Pinterest followers, all the distributions of FSDs followed Benford’s Law. Note that with datasets of this size, it is not appropriate to conduct a statistical hypothesis test for goodness of fit; over tens of thousands or millions of people, even a very tiny deviation would cause us to reject the null hypothesis. Furthermore, conformance with Benford’s Law has never been about a perfect statistical match to the predicted values—not even in Benford’s original work on the subject. Rather, the relative frequencies of FSDs are the guiding principle. Pearson correlations are a common way to measure how closely a distribution adheres to Benford’s Law. The correlation between the FSDs of the friend–follower counts and what Benford’s Law predicts are extremely strong. As shown in, all the *r* values are \> 0.990. We also ran Kolmogorov–Smirnov tests to check the fit of the data with the Benford’s Law distribution. The p-values, which indicate the probability that the social network’s FSD distribution is the same as Benford’s, are also shown in. These values are all \> 0.97. Other user behavior also fit Benford’s Law as well. We had data for the number of posts users made on Pinterest (number of pins) and Twitter (number of tweets). In both cases, correlation with Benford’s predictions was extremely high: 0.9998 and 0.9960, respectively. However, we mentioned above that one dataset did not follow Benford’s predictions. On Pinterest, users have both a follower count, which represents incoming social connections, and a following count for outgoing edges. The follower count is what we presented above, and it follows the expected distribution. The *following* count did not adhere to Benford’s Law. The percentages are very far off the law’s prediction, and the dominance of FSDs of 5 is especially striking. Is this simply an exception to the rule, or is something else going on? When Benford’s Law is applied in forensic accounting, auditors know to look for explanations of data that appears unusual. For example, a company may have a high percentage of FSDs of 3, not because anything fraudulent is happening, but because they happen to frequently purchase an item that costs \$39.99. We investigated this issue on Pinterest more deeply and found the explanation for the frequent 5s. When new users sign up for Pinterest, they are prompted to choose “interests” to follow. Users *must* select at least five before continuing with the registration process. This creates at least five initial following relationships for users. Though users can go in and later delete those follows, few do, and this initiation process affects the entire distribution of FSDs. When we looked at the edges in the opposite direction (considering incoming follower edges rather than outgoing), the FSDs adhered to Benford’s Law, as shown above. This exposes an important point about applying Benford’s Law: it can be violated when there is external influence over people’s natural behavior. In the Pinterest case, we discovered the influence was an artifact of the system configuration. Benford’s Law extends to second digits, as well. The distribution is much flatter, ranging from 0.1197 for the frequency of 1s to 0.0850 for the frequency of 9s. Our networks agreed with the second digit distribution quite well. The Pearson correlations for all five networks, including Twitter friend counts and follower counts, are all \> 0.97. The Kolmogorov–Smirnov p-values are \> 0.98 for Twitter friends, Facebook, Pinterest, and Google Plus. This second digit analysis eliminated the frequent 5s artifact seen in Pinterest. The Kolmogorov–Smirnov p-values for LiveJournal and Twitter follower counts are a bit less impressive (0.79 and 0.76, respectively), though some existing literature suggests this might be the expected with the flatter distribution of the second digit. ## Extension to Egocentric Networks The adherence to Benford’s Law carries through into FSD distributions within individual egocentric networks. Using data from Twitter, Google Plus, and LiveJournal, we selected individuals with at least 100 friends, and then obtained the number of social connections that each of those friends had. We then determined FSD distributions in the friend-of-friend counts of each egocentric network. Overall, the vast majority of egocentric networks conformed to Benford’s Law. On Google Plus, 91.5% of users’ egocentric networks’ FSD distributions had a correlation of over 0.9 with Benford’s Law predictions. This was true for 85.1% of LiveJournal egocentric networks. In the Twitter data, 89.7% of users had a correlation of over 0.9. Of our 21,135 users, only 170 (\< 1%) had a correlation under 0.5. Since we had non-anonymized data for Twitter, we were able to investigate these accounts with low correlations. Nearly every last one of the 170 accounts mentioned above appeared to be engaged in suspicious activity. Some accounts were spam, but most were part of a network of Russian bots that posted random snippets of literary works or quotations, often pulled arbitrarily from the middle of a sentence. All the Russian accounts behaved the same way: following other accounts of their type, posting exactly one stock photo image, and using a different stock photo image as the profile picture. While we are currently investigating the purpose of these bot accounts’ existence, their deviation from Benford’s Law made it quite easy to identify their highly unusual behavior. Of the 170 accounts, only 2 seemed to belong to legitimate users. Figs and show examples of a spam account and Russian bot account detected by this method. # Discussion and Conclusions We have shown that Benford’s Law applies to relationships in online social networks. This is true for social networking sites as a whole, and for individual users’ egocentric networks. Data from Twitter and Pinterest also suggest that it applies to the number of posts users make on social media sites, as well. In the one network where Benford’s Law did not hold, closer inspection revealed that it due to a feature of the system that altered users’ behavior. Next are some applications for these results, followed by some closing thoughts on the work. ## Applications First, Benford’s Law can be used to detect users who are behaving in unexpected ways. As we found in our Twitter dataset, the vast majority of accounts that strongly deviated from the expected FSD distributions were engaged in unusual behavior. As is the case with forensic accounting investigations using Benford’s Law, a deviation does not necessarily mean there is fraud happening. Given the large number of users on social media, it would be statistically unusual to have *no* accounts that naturally deviate from expected patterns; rather, deviation from a Benford distribution can flag accounts for additional review. These insights can also be used to validate experimental datasets. It is often the case that data can be hard to collect from social media sites, especially when researchers are looking for detailed personal information. Truly random or representative sampling is difficult to do—and essentially impossible when connected components of a social network are important to the analysis. This raises the question as to whether the sample of accounts collected by a research team seriously deviates from normal patterns. While Benford’s Law only addresses one aspect of expected behavior, major differences between a sample’s FSD distribution and Benford’s Law could indicate serious sampling problems. We tested this by analyzing the FSD distributions on a number of datasets collected for various projects and experiments. We randomly selected 50 Twitter-based networks from the NodeXL Graph Gallery (<http://nodexlgraphgallery.org>). These were all generated by collecting the networks of users who had tweeted a given search term. For each graph, we analyzed friend, follower, and tweet counts for the users in each dataset. On all graphs and each of the three measures, the Pearson correlation with Benford’s expected values was \> 0.990. This shows that, structurally, the networks look like we would expect. We found similarly strong correlations and agreement with Twitter data collected for a research project that posted a survey on a popular psychology website. Subjects included their Twitter IDs in their survey responses. The distributions of FSDs for friends, followers, and tweets all correlated with Benford’s Law distributions with *r* \> 0.990 and very close values. However, not all datasets were a good match. One Facebook dataset of 151 users had a Pearson correlation of 0.761, and there were large differences between the predicted and actual frequencies. All but two FSDs saw deviations over 25% from expected values, and some saw deviations over 80%. This was true on another Facebook dataset with 220 users (supplied by a colleague). In this example, friend counts were self-reported, and 94.5% of those began with a 1—more than triple the expected 30.1%. Such deviations do not necessarily imply a problem with the data; indeed, this distribution may be irrelevant to the analysis being performed. However, it hints that the subjects are not reflecting an expected distribution, and thus may vary from the larger population in other ways. Further research is needed to understand the implications of deviation in experimental samples. There is a growing understanding of the subtle patterns of natural behavior which humans have difficulty replicating in unnatural circumstances. The applicability of Benford’s Law to social media is a new tool for analyzing user behavior, understanding when and why natural deviations may occur, and ultimately, detecting when abnormal forces are at work. Thanks to Tanya Lokot for her help in translating and analyzing the Russian bot accounts we detected. [^1]: The author has declared that no competing interests exist. [^2]: Conceived and designed the experiments: JG. Performed the experiments: JG. Analyzed the data: JG. Contributed reagents/materials/analysis tools: JG. Wrote the paper: JG.