Robust semi-supervised scRNA-seq integration from virtual adversarial learning.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) integration methods often fail to distinguish closely related cell subtypes.
• Over-mixing of cell populations reduces biological resolution and interpretability.
• Existing marker gene integration approaches are limited by data variability and complexity.

Purpose of the Study:

• Introduce scCRAFT+, a semi-supervised integration model.
• Enhance scRNA-seq data integration and cell type auto-annotation accuracy.
• Improve robustness to noisy or incomplete marker gene information.

Main Methods:

• Developed scCRAFT+, a semi-supervised integration model.
• Incorporated marker gene information using Virtual Adversarial Training (VAT).
• Jointly optimized marker-derived supervision and transcriptome-wide representations.

Main Results:

• VAT enforced local prediction smoothness, improving robustness to noisy marker annotations.
• Achieved enhanced integration quality and cell type auto-annotation.
• Demonstrated superior performance compared to existing unsupervised and supervised integration methods.
• Significantly improved annotation accuracy and robustness, especially with incomplete marker sets.

Conclusions:

• scCRAFT+ offers a robust and accurate approach for scRNA-seq data integration.
• The model effectively leverages marker gene information for improved biological resolution.
• scCRAFT+ enables more biologically meaningful sub-cell type auto-annotations.