GE-Impute

GE-Impute employs a graph embedding-based neural network to impute dropout zeros in single-cell RNA sequencing (scRNA-seq) data by reconstructing cell-cell similarity networks to improve expression recovery.

Key Features:

• Graph Embedding Neural Network Model: Learns a graph representation for each cell to reconstruct a cell-cell similarity network from scRNA-seq data.
• Improved Imputation Accuracy: Uses the reconstructed similarity network to allocate neighbors and impute dropout zeros, improving recovery of missing expression values for both droplet-based and plate-based scRNA-seq.
• Enhanced Biological Interpretation: Improves identification of differentially expressed genes and unsupervised clustering by providing more accurate gene expression profiles.
• Marker Gene Identification and Cell Type Assignment: Facilitates detection of marker genes and supports assignment of cell types to clusters based on imputed expression.
• Trajectory Analysis Improvement: Enhances reconstruction of differentiation trajectories in time-course scRNA-seq data by refining expression inputs.

Scientific Applications:

• Complex cellular heterogeneity: Enables more accurate characterization of heterogeneous cell populations by reducing dropout-related noise in expression matrices.
• Developmental biology and lineage reconstruction: Improves inference of developmental trajectories and lineage relationships from time-course scRNA-seq data.
• Disease modeling and biomarker discovery: Enhances detection of disease-associated expression patterns and marker genes by recovering dropout-masked signals.

Methodology:

GE-Impute learns a neural graph representation for each cell, reconstructs a cell-cell similarity network from those representations, and uses the network to impute dropout zeros by borrowing information from allocated neighboring cells.