SysBio-Traj

SysBio-Traj is the dataset released with RegimeFlow, a regime-aware flow matching framework for probabilistic forecasting of biological trajectories across dynamical systems.

This dataset accompanies the paper "A Regime-Aware Trajectory Prediction Framework for 1000+ Systems Biology Models", accepted to ICML 2026. It contains 1,050 biological dynamical systems, each organized as a self-contained model folder with a simulated trajectory, the source SBML model, curated initial conditions, and per-species regime metadata.

SysBio-Traj is designed for regime-aware time-series modeling, systems biology analysis, and reproducible simulation.

Quick Facts

Item	Description
Number of systems	1,050
Organization unit	One folder per model
Trajectory length	512 uniformly sampled time points
Trajectory format	CSV with time + one column per species
Model source	SBML (.xml)
Extra metadata	Initial conditions + regime annotations
Global index	SysBio-Traj_index.csv
Companion code	RegimeFlow
Simulation backend	Tellurium
Model ID format	BIOMD... or MODEL...

Directory Layout

SysBio-Traj/
├── README.md
├── SysBio-Traj_index.csv
├── scripts/
│   └── simulate_sbml.py
└── Data/
    ├── BIOMD0000000013/
    │   ├── Poolman2004.csv
    │   ├── Poolman2004.xml
    │   ├── initial_conditions.json
    │   └── Poolman2004_conditions.json
    ├── BIOMD0000000144/
    │   ├── Calzone2007.csv
    │   ├── Calzone2007.xml
    │   ├── initial_conditions.json
    │   └── Calzone2007_conditions.json
    └── ...

Each model folder contains exactly four files:

File	Purpose
model_name.csv	Simulated multivariate trajectory
model_name.xml	Source SBML model
initial_conditions.json	Calibrated initial state used for reproduction
model_name_conditions.json	Per-species regime labels and summary metadata

How to Use the Dataset

1. Start with SysBio-Traj_index.csv to find the model_id, model_name, and simulation time range.
2. Open Data/<model_id>/ to access the trajectory, SBML file, and metadata for that model.
3. Use scripts/simulate_sbml.py if you want to reproduce a trajectory from the released SBML + initial conditions.
4. Use the companion RegimeFlow repository for the model training and evaluation workflow built around this dataset.

SysBio-Traj_index.csv

Column	Description
model_id	Unique folder name
model_name	Base filename for the model files
time_start	Simulation start time
time_end	Simulation end time
time_span	Total simulated time span

Example:

model_id,model_name,time_start,time_end,time_span
BIOMD0000000013,Poolman2004,0,0.4,0.4
BIOMD0000000144,Calzone2007,0,300,300
...

Regime Metadata

Each *_conditions.json file stores regime annotations for the released trajectory.

Field	Meaning
trajectory_type	Numeric regime label used by RegimeFlow
trajectory_type_name	Released dataset label
bounds	Minimum and maximum values observed in the trajectory
period	Period measured in sample-index units

The benchmark follows the six-class regime taxonomy used in RegimeFlow. In the released JSON files, trajectory_type_name retains the compact dataset labels for consistency with the provided annotations, while the terms in parentheses denote the corresponding names used in the paper: directly_stable (complex), inc_stable (increasing-stable), dec_stable (decreasing-stable), oscillation, increasing (monotonic increasing), and decreasing (monotonic decreasing).

Reproducing the Released Trajectories

This section describes how to reproduce the released dataset trajectories. Trajectories were generated by numerically simulating each SBML model with Tellurium, using the released time range and sampling exactly 512 time points. The file initial_conditions.json provides the curated initial state needed to reproduce the released trajectory.

Install the required Python packages:

pip install -r requirements.txt

The pinned environment is intended for Python 3.10. Tellurium and libRoadRunner may not provide compatible wheels for newer Python releases.

Example:

python scripts/simulate_sbml.py \
  --model-id BIOMD0000000013 \
  --model-name Poolman2004 \
  --start-time 0 \
  --end-time 0.4 \
  --num-timepoints 512 \
  --use-ic-json

Data Source and License

The source SBML models are derived from BioModels, a repository of mathematical models of biological systems. BioModels states that its encoded models and annotations are distributed under the Creative Commons CC0 Public Domain Dedication.

SysBio-Traj adds simulated trajectories, curated initial conditions, regime annotations, and benchmark-level metadata for RegimeFlow. This released dataset is provided under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use SysBio-Traj or RegimeFlow, please cite:

@inproceedings{rao2026regime,
  title     = {A Regime-Aware Trajectory Prediction Framework for 1000+ Systems Biology Models},
  author    = {Rao, Heng and Zhang, Jason Zipeng and Gu, Yu and Liu, Zhenghao and Yu, Ge and Su, Jeffrey and Cao, Yang and Yang, Fan and Chen, Minghan},
  booktitle = {Forty-third International Conference on Machine Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=sI3UUkXJxs}
}

References

• [R1] Hucka, Michael, et al. "The systems biology markup language (SBML): a medium for representation and exchange of biochemical network models." Bioinformatics 19.4 (2003): 524-531.
• [R2] Medley, J. Kyle, et al. "Tellurium notebooks—an environment for reproducible dynamical modeling in systems biology." PLoS computational biology 14.6 (2018): e1006220.
• [R3] Malik-Sheriff, Rahuman S., et al. "BioModels—15 years of sharing computational models in life science." Nucleic acids research 48.D1 (2020): D407-D415.