scGImpute: A hybrid BiLayer multi-head graph attention-based imputation framework for zero dropout in single-cell...

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Single-cell sequencing (SCS) is crucial for multi-omics profiling in various organisms.
• Advances in SCS aid disease diagnosis and treatment at the molecular level.
• Technical noise, particularly zero counts (dropouts), significantly hinders SCS data utility.

Purpose of the Study:

• To address high-frequency zero counts in SCS data.
• To develop a novel imputation method for omics SCS datasets.
• To preserve biological zeros while recovering technological zeros.

Main Methods:

• Proposed a hybrid two-layer multi-head graph attention-based neural network framework (scGImpute).
• Designed scGImpute to be generic for scRNA-seq, scATAC-seq, and CITE-seq data.
• Evaluated performance using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).

Main Results:

• scGImpute achieved the lowest RMSE (0.0039) and MAE (0.0032) for raw vs. imputed scRNA-seq data.
• Demonstrated superior performance on ground truth vs. imputed scRNA-seq data (RMSE: 0.3301, MAE: 0.2924).
• Outperformed existing imputation techniques and other proposed methods in evaluations.

Conclusions:

• scGImpute effectively reduces technical noise in SCS datasets.
• The method enhances downstream analysis for biological discovery.
• Provides a robust solution for zero imputation across multiple omics types.