GNNImpute

GNNImpute performs imputation of dropout events in single-cell RNA sequencing (scRNA-seq) data using an autoencoder integrated with graph attention networks to reconstruct true gene expression profiles for downstream analyses.

Key Features:

• Graph Attention Network (GAT): Uses graph attention convolution to aggregate multi-level similar-cell information and handle non-Euclidean relationships in scRNA-seq data.
• Autoencoder Structure: Employs an autoencoder framework integrated with GAT to reconstruct gene expression from dropout-affected measurements.
• Performance Metrics: Evaluated using mean square error (MSE = 3.0130), mean absolute error (MAE = 0.6781), Pearson correlation coefficient (PCC = 0.9073), and cosine similarity (CS = 0.9134) across four real datasets.
• Clustering Effectiveness: Improves clustering quality with an Adjusted Rand Index (ARI) of 0.8199 and Normalized Mutual Information (NMI) of 0.8368.

Scientific Applications:

• Cellular heterogeneity analysis: Reduces dropout noise in scRNA-seq data to enable more accurate characterization of cellular heterogeneity.
• Cell differentiation and developmental biology: Provides denoised gene expression profiles to support studies of cell differentiation and developmental processes.
• Disease mechanism investigation: Facilitates investigation of disease mechanisms by supplying more reliable single-cell expression measurements.
• Downstream analyses (clustering and differential expression): Enhances the reliability of clustering and differential expression analyses by imputing dropout events.

Methodology:

Constructs a cell graph with nodes representing individual cells and edges reflecting gene-expression similarity, applies a graph attention convolution within an autoencoder framework to prioritize relevant cellular interactions for context-aware imputation; the approach is distinguished from DCA, MAGIC, scVI, scImpute, and SAVER.