Inferring stochastic dynamics by biophysical Neural ODE using single-cell transcriptomics.

Single-cell RNA sequencing has revolutionized our ability to dissect cellular heterogeneity and study cell fate mechanisms, yet inferring stochastic dynamics from static snapshots remains a fundamental challenge. Current approaches face a critical trade-off: mechanistic models impose rigid assumptions limiting biological realism, while data-driven methods sacrifice interpretability for deeper mechanistic explorations. Here, we present DynNet, a deep learning method that integrates Neural ODEs with biophysical models and prior knowledge of gene expression dynamics. DynNet learns the stochastic dynamics of gene regulatory systems for cell fate decisions. Benchmarking on synthetic data shows DynNet's ability to infer stable cell states, reconstruct dynamical trajectories, and characterize multi-stable cell fate transitions. Using hepatocyte differentiation data, DynNet demonstrates its capability to infer developmental trajectory and the underlying cell fate landscape, revealing the stability and transition probabilities among distinct cell states. Applied to Epithelial-mesenchymal transition (EMT) data, DynNet further captures critical gene regulations and transition paths during EMT.