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

Area of Science:

• Computational biology
• Single-cell genomics
• Biophysics

Background:

• Gene expression data often exhibits bimodal distributions, reflecting distinct biological states.
• Existing variational autoencoder methods may not fully capture the causal relationships in multimodal single-cell data.

Approach:

• We present biVI, integrating the scVI framework with bivariate, biophysically motivated models.
• biVI explicitly models the biophysical processes governing nascent and mature RNA distributions.

Key Points:

• Simulated benchmarking shows biVI accurately captures cell type structure and parameter values.
• biVI effectively recapitulates copy number distributions in gene expression data.
• The method provides a scalable approach for analyzing multimodal single-cell genomic datasets.

Conclusions:

• biVI offers a novel strategy for analyzing complex, multimodal single-cell data by modeling underlying biophysical mechanisms.
• This framework is generalizable to various high-throughput single-cell genomic assays, advancing the understanding of gene expression regulation.