scVAE

scVAE models single-cell RNA sequencing (scRNA-seq) count data using variational auto-encoders (VAEs) to learn latent representations and enable likelihood-based model comparison.

Key Features:

• Direct Use of Raw Count Data: Models raw scRNA-seq count data directly without requiring intermediate normalization or transformation steps.
• Latent Representation Learning: Learns a latent representation for each cell to capture underlying biological variability and cellular heterogeneity.
• Deep Generative Models: Employs variational auto-encoders as deep generative models to represent complex gene expression distributions across cells.
• Likelihood-Based Model Comparison: Uses likelihood-based criteria to compare and evaluate alternative models on scRNA-seq data.
• Incorporation of Count Likelihood Functions: Tests and employs various count likelihood functions to model gene expression counts accurately.
• A Priori Clustering in Latent Space: Integrates a priori clustering within the latent space to inform and improve cell clustering outcomes.

Scientific Applications:

• Cell Type Identification and Clustering: Produces latent-space clustering of single cells that reflects distinct cell types for cellular taxonomy and population analyses.
• Gene Expression Estimation: Estimates expected gene expression levels from modeled count distributions to support downstream analysis of gene regulation and differential expression.

Methodology:

Applies variational auto-encoders to model raw count scRNA-seq data, evaluates multiple count likelihood functions, performs likelihood-based model comparison, incorporates a priori clustering in latent space, and is implemented in Python using TensorFlow.