Discovering governing equations of biological systems through representation learning and sparse model discovery

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Biological systems are complex, exhibiting nonlinear and high-dimensional dynamics.
• Analyzing single-cell RNA sequencing (scRNA-seq) data to understand these dynamics is challenging.
• Existing methods often struggle to capture the intricate regulatory mechanisms driving cellular processes.

Purpose of the Study:

• To present CLERA, a novel computational framework for uncovering dynamical models and active gene programs from scRNA-seq data.
• To integrate prior biological knowledge for simultaneous dimensionality reduction and dynamical system identification.
• To provide a tool for robustly reconstructing gene expression dynamics and identifying key regulatory elements.

Main Methods:

• CLERA employs a supervised autoencoder architecture integrated with Sparse Identification of Nonlinear Dynamics (SINDy).
• The framework leverages prior knowledge to extract low-dimensional representations and uncover underlying dynamical systems.
• Network analysis, including Personalized PageRank, is used to identify central genes and active gene programs.

Main Results:

• CLERA demonstrates robust performance in reconstructing gene expression dynamics across various cell types.
• The framework successfully identifies key regulatory genes driving cellular processes.
• Dynamic interaction networks are generated, highlighting temporal patterns and regulatory mechanisms.

Conclusions:

• CLERA provides a powerful approach for modeling complex biological systems using scRNA-seq data.
• The framework offers new insights into the regulatory mechanisms underlying cellular processes by identifying active gene programs.
• CLERA enhances our ability to understand the dynamics of biological systems at the single-cell level.