pmVAE

pmVAE learns interpretable pathway-level latent representations from single-cell RNA sequencing (scRNA-seq) data by integrating pathway gene sets into a pathway-module variational autoencoder architecture.

Key Features:

• Pathway-informed architecture: Leverages pathway gene sets as a biological prior and organizes the model into subnetworks corresponding to specific pathways.
• Pathway modules (mini-VAEs): Implements modules that act as mini Variational Autoencoders (VAEs) focusing exclusively on genes associated with their respective pathways.
• Latent space factorization: Factorizes the latent space according to pathway gene sets to enable direct interpretation of latent dimensions in a pathway context.
• Custom training procedure: Balances a module-specific local loss with a global reconstruction loss to encourage independence among modules while accounting for overlapping and hierarchical pathway structures.
• Multidimensional pathway representation: Models each pathway as a multidimensional vector to capture hierarchical organization and multiple downstream signals.
• Interpretable downstream analysis: Produces factorized representations that facilitate analysis of cell type differentiation and responses to biological stimuli at the pathway level.

Scientific Applications:

• Detection of perturbation-targeted pathways: Provides discriminative and consistent identification of pathways affected by experimental perturbations in scRNA-seq studies.
• Pathway-level interpretation of cell states: Enables analysis of cell type differentiation and stimulus responses within the context of individual pathways.
• Investigation of cellular mechanisms and targets: Supports elucidation of pathway-specific contributions to cellular mechanisms and the identification of candidate therapeutic targets.

Methodology:

Uses pathway gene sets as a biological prior and organizes the model into pathway-specific subnetworks where each module functions as a mini VAE encoding its pathway genes; the approach factorizes the latent space by pathway gene sets, models each pathway as a multidimensional vector, and trains by balancing a module-specific local loss with a global reconstruction loss to manage overlapping and hierarchical pathway correlations.