TriPDCL: A Tri-Pathway Prototype-Driven Contrastive Learning Framework for Cross-Modality Single-Cell Integration.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell sequencing technologies enable multiomics data integration.
• Challenges include technical batch effects, biological variability, data sparsity, and intercellular heterogeneity.
• Existing methods struggle with joint analysis of sparse, heterogeneous single-cell multiomics data.

Purpose of the Study:

• To develop a framework for effective multiomics data integration, focusing on cellular heterogeneity.
• To address challenges in aligning cross-modal heterogeneity and robust learning in joint analysis.
• To improve downstream analyses by creating a unified latent representation of single-cell multiomics data.

Main Methods:

• Proposed TriPDCL, a prototype-based contrastive learning framework.
• Employed an iterative prototype-learning update mechanism for heterogeneity information transfer.
• Utilized learnable prototype centers for constructing reliable positive-negative sample pairs.

Main Results:

• TriPDCL effectively transfers cellular heterogeneity information.
• The framework enables precise construction of reliable sample pairs for contrastive learning.
• Comparative assessments on five datasets demonstrated the superiority of TriPDCL over seven representative methods.

Conclusions:

• TriPDCL offers a robust solution for multiomics integration, particularly for single-cell data.
• The framework successfully aligns cross-modal heterogeneity and enhances learning robustness.
• This approach facilitates more reliable downstream analyses in multiomics research.