One generative model that forecasts the future dynamics of 1,050 biological systems — and generalizes to systems it has never seen. Instead of hand-crafting a new ODE for every system, RegimeFlow learns the shared “grammar” of biological dynamics from simulated trajectories, and predicts future states.
MSE comparison between Ours and five competitive baselines across ten most frequent biological system taxonomies.
Overall MSE across all models (lower is better).
1,050 curated ODE mechanism models, 1–1,000 species each.
Numerically solved trajectories, 512 time points per system.
Regime-aware flow matching, trained once across all systems.
Past 96 points → future 256 points, with uncertainty.
Stable / oscillatory / monotonic behavior, injected as conditioning signal C.
Bayesian linear regression with per-regime basis functions replaces standard Gaussian noise.
Learns a vector field that pushes prior samples to target trajectories.
Linear-complexity state-space backbone with adaptive layer norm.
Biological systems are wildly diverse, yet their large-scale behavior collapses to a handful of regimes. By conditioning on regime and seeding generation with a regime-aware prior instead of Gaussian noise, RegimeFlow shortens the transport distance the model must learn — yielding 31% lower MAE and 17% better-calibrated uncertainty (CRPS) than baselines, while generalizing to unseen systems with no new equations and no retraining.
1,050 Biological Models