VAPOR: Variational autoencoder with transport operators decouples co-occurring biological processes in development

Area of Science:

• Computational Biology
• Developmental Biology
• Genomics

Background:

• Single-cell and spatial transcriptomics offer insights into developmental gene expression dynamics.
• Existing trajectory inference methods may aggregate distinct biological processes occurring on different timescales.
• A single pseudotime axis can obscure the nuances of simultaneous cellular processes like differentiation and maturation.

Purpose of the Study:

• To introduce VAPOR (variational autoencoder with transport operators), a novel computational method.
• To decouple and analyze distinct gene expression dynamics during development.
• To infer process-specific pseudotimes for multifaceted timescale analysis.

Main Methods:

• VAPOR utilizes a variational autoencoder to learn a latent space of gene expression dynamics.
• It decomposes this latent space into multiple subspaces, each governed by an ordinary differential equation.
• Process-specific pseudotimes are inferred to reveal distinct developmental timescales.

Main Results:

• VAPOR successfully recovered data topology and decoupled dynamic patterns in simulated datasets.
• Application to human brain scRNA-seq data identified key differentiation and maturation dynamics.
• Benchmarking demonstrated VAPOR's superior performance compared to existing methods.
• VAPOR analysis of spatial transcriptomics data revealed the 'inside-out' pattern of cortical layer formation.

Conclusions:

• VAPOR effectively disentangles complex gene expression dynamics during development.
• The method provides process-specific pseudotimes, offering a more nuanced view of cellular progression.
• VAPOR is open-source and applicable to various omics data, including epigenomic dynamics.