An efficient scRNA-seq dropout imputation method using graph attention network.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell sequencing reveals cellular heterogeneity but faces computational challenges.
• Dropout events (false zeros) in gene expression data complicate analysis.
• Current imputation methods struggle to accurately recover true expression from dropout noise.

Purpose of the Study:

• To develop an effective method for imputing dropout events in single-cell RNA sequencing (scRNA-seq) data.
• To address the limitations of existing dropout imputation tools.

Main Methods:

• Proposed GNNImpute, an autoencoder network utilizing graph attention convolution.
• Applied convolution operations on non-Euclidean space for scRNA-seq data.
• Aggregated multi-level similar cell information to reduce dropout noise.

Main Results:

• GNNImpute accurately and effectively imputes dropout events and reduces noise.
• Achieved superior performance metrics: 3.0130 MSE, 0.6781 MAE, 0.9073 PCC, 0.9134 CS.
• Demonstrated improved clustering effects with 0.8199 ARI and 0.8368 NMI.

Conclusions:

• GNNImpute effectively imputes dropout events in scRNA-seq data by leveraging shared information.
• Graph attention convolution and autoencoder structures show significant potential for single-cell dropout imputation.
• The method demonstrates robust performance across multiple real datasets.