SCMBench: benchmarking domain-specific and foundation models for single-cell multi-omics data integration.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell sequencing generates vast multi-omics data, necessitating advanced integration tools.
• Existing multi-omics integration tools lack comprehensive evaluations and clear guidelines for researchers.

Purpose of the Study:

• To benchmark Domain-specific Models (DMs) and Foundation Models (FMs) for multi-omics data integration.
• To evaluate model performance on integration accuracy, biomarker detection, trajectory inference, and batch effect correction.
• To identify limitations of FMs and propose strategies for improvement in single-cell multi-omics analysis.

Main Methods:

• Benchmarking 23 multi-omics integration methods, including DMs and FMs, with optimized hyperparameters.
• Assessing performance based on integration accuracy, biomarker preservation, trajectory inference, and batch effect correction.
• Developing and testing a lightweight adaptation strategy to enhance FM performance.

Main Results:

• Foundation Models (FMs) generally underperform state-of-the-art Domain-specific Models (DMs) in multi-omics integration tasks.
• The proposed adaptation strategy significantly improves FM effectiveness in integrating single-cell multi-omics data.
• Comprehensive analysis highlights the importance of preserving cellular characteristics and developmental trajectories.

Conclusions:

• Researchers should carefully select integration methods based on specific single-cell analysis goals.
• Foundation Models show promise for multi-omics integration but require adaptation strategies for optimal performance.
• This study provides essential guidance for selecting and developing effective multi-omics integration tools.