scCMA: A Contrastive Masked Autoencoder Framework for Robust Representation Learning of scRNA-seq Data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) data analysis faces challenges like high dimensionality, sparsity, noise, and batch effects.
• These issues hinder accurate cell clustering and downstream analyses, impacting biological discoveries.

Purpose of the Study:

• To develop a novel computational framework, scCMA, for robust scRNA-seq data analysis.
• To generate stable and biologically meaningful cell embeddings that overcome common data challenges.

Main Methods:

• scCMA integrates discriminative representation learning with a masked autoencoder architecture.
• A contrastive module enhances cell-type distinctions and mitigates batch effects implicitly.
• A masked autoencoder captures transcriptional dependencies and reduces noise/sparsity impact.

Main Results:

• scCMA demonstrated superior performance in clustering precision across diverse datasets.
• The framework effectively corrected batch differences while preserving biological variance.
• scCMA showed proficiency in identifying rare cell subsets and modeling cell development trajectories.

Conclusions:

• scCMA provides a powerful tool for accurate and robust scRNA-seq data analysis.
• The generated cell embeddings facilitate deeper understanding of cellular heterogeneity and developmental processes.