GeneVector

GeneVector constructs dense gene vector representations using mutual information to identify and analyze transcriptional programs and model gene co-expression in single-cell RNA-seq data.

Key Features:

• Mutual information-based vectors: Defines dense gene vector representations using mutual information between gene expression profiles.
• Vector space model: Implements a vector space model that captures gene-specific co-expression relationships rather than aggregating sparse signals.
• Low-dimensional embeddings: Produces low-dimensional gene embeddings that reflect co-expression patterns.
• Dimensionality reduction on gene co-expression: Performs dimensionality reduction focused on gene co-expression patterns to address sparsity in scRNA-seq data.
• Latent space arithmetic: Enables latent space arithmetic within the gene embedding framework to interrogate transcriptional programs.
• Cell-type classification: Supports classification and annotation of cell types from scRNA-seq data using gene embeddings.
• Batch effect correction: Performs batch effect correction across experimental conditions.
• Longitudinal pathway variation detection: Identifies pathway variations associated with treatment over time.
• Demonstrated performance: Validated on four distinct single-cell RNA-seq datasets.

Scientific Applications:

• Transcriptional program discovery: Detection and characterization of transcriptional programs from single-cell data.
• Cell-type classification and annotation: Classification and annotation of cell types based on gene co-expression embeddings.
• Phenotype-specific pathway analysis: Extraction of phenotype-specific pathways and pathway-level signals.
• Batch correction and integration: Correction of batch effects to ensure consistency across experiments.
• Longitudinal treatment-response analysis: Tracking pathway changes and transcriptional dynamics with treatment over time.

Methodology:

Computes mutual information between gene expression profiles to define dense gene vectors and constructs a vector space model, performs dimensionality reduction on gene co-expression to obtain low-dimensional embeddings, and applies latent space arithmetic within the gene embedding framework.