Biophysical modeling with variational autoencoders for bimodal, single-cell RNA sequencing data

Area of Science:

• Computational Biology
• Genomics
• Molecular Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-throughput gene expression profiling.
• Understanding the biophysical processes governing RNA dynamics is crucial for interpreting scRNA-seq data.
• Existing methods often lack the capacity to directly model these underlying molecular mechanisms.

Purpose of the Study:

• To introduce biVI, a novel computational framework for scRNA-seq data analysis.
• To integrate variational autoencoder (VAE) principles with biophysical models of RNA transcription and splicing.
• To enable genome-wide inference of RNA biophysical parameters from scRNA-seq data.

Main Methods:

• Developed biVI by combining the scVI VAE framework with kinetic models for RNA transcription and splicing.
• Applied biVI to both simulated and experimental scRNA-seq datasets.
• Evaluated biVI's performance in capturing cell type structure and inferring biophysical parameters.

Main Results:

• biVI successfully retains the cell type structure identification capabilities of scVI.
• biVI enables genome-wide inference of RNA biophysical mechanisms, including system burst sizes and degradation rates.
• The model provides insights into the molecular processes driving observed gene expression patterns.

Conclusions:

• biVI offers a powerful new approach for analyzing scRNA-seq data by integrating VAEs with biophysical modeling.
• This framework facilitates a deeper understanding of the molecular underpinnings of cell-to-cell variability in gene expression.
• biVI opens avenues for exploring RNA kinetics and regulatory mechanisms across diverse biological systems.