Learning stochastic cellular dynamics from snapshots through bidirectional causal neural stochastic differential...

Area of Science:

• Computational Biology
• Systems Biology
• Machine Learning

Background:

• Single-cell sequencing data provides snapshots but obscures cell lineages and gene dynamics due to cell destruction.
• Existing methods often use univariate expression, neglecting cell types and rotational effects in non-steady states, leading to limited accuracy and interpretability.
• Cellular fates are complex, involving divergence and potential reversibility, which current models struggle to capture.

Purpose of the Study:

• To develop a robust computational framework for analyzing single-cell time-series data that overcomes limitations of existing methods.
• To accurately model cell differentiation pathways, including gene dynamics and lineage reconstruction.
• To enhance the interpretability of cell fate decisions and construct potential landscapes.

Main Methods:

• A bidirectional causal neural network based on stochastic differential equations (SDEs) was proposed.
• The model jointly trains forward and reverse processes of a diffusion model for reversible consistency and learning stability.
• The forward process models cell differentiation using causal relationships within principal components, while the reverse process incorporates rotational effects and learns the drift term via reverse modeling.

Main Results:

• The proposed multivariate dynamic modeling approach demonstrates improved expressiveness and interpretability compared to univariate methods.
• The model ensures consistency in cell-type distribution between predictions and real data.
• An interpretable Waddington potential landscape was constructed by learning the drift term through reverse modeling.

Conclusions:

• The bidirectional causal neural network effectively models complex cell differentiation dynamics and gene expression from time-series single-cell data.
• The framework enhances the stability, interpretability, and accuracy of analyzing cellular processes.
• This approach provides a powerful tool for understanding cell fate decisions and constructing potential landscapes in systems biology.