scGACL: a generative adversarial network with multi-scale contrastive learning for accurate single-cell RNA...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for studying cell heterogeneity.
• Dropout events in scRNA-seq data necessitate accurate imputation for reliable downstream analysis.
• Existing imputation methods often cause over-smoothing, losing critical cell-to-cell variation.

Purpose of the Study:

• To develop an advanced imputation method that overcomes the over-smoothing problem in scRNA-seq data.
• To preserve both fine-grained cell-to-cell heterogeneity and macroscopic cell-type variations.
• To improve the accuracy of downstream analyses in scRNA-seq studies.

Main Methods:

• Proposed scGACL, a novel method combining a generative adversarial network (GAN) with multi-scale contrastive learning.
• Employed GAN architecture to ensure imputed data distribution mirrors real data distribution.
• Implemented cell-level contrastive learning to retain cell heterogeneity and cell-type-level contrastive learning for biological variation.

Main Results:

• scGACL effectively addresses the over-smoothing issue inherent in other imputation techniques.
• The method accurately recovers gene expression profiles from scRNA-seq data.
• Demonstrated superior performance of scGACL across simulated and real-world datasets compared to existing methods.

Conclusions:

• scGACL provides accurate imputation for scRNA-seq data while preserving crucial biological heterogeneity.
• The method significantly enhances the performance of downstream analyses, including cell clustering and differential gene expression.
• scGACL represents a significant advancement for scRNA-seq data analysis, enabling more robust biological insights.