Explainable Graph Learning for Multimodal Single-Cell Data Integration

Abstract

AbstractIntegrating multi-omic single-cell data is essential for uncovering cellular het- erogeneity and identifying specialized subpopulations. However, achieving both explainable and expressive integration remains challenging due to the complex relationships between modalities. Here, we introduce Single-Cell PROteomics Vertical Integration (SCPRO-VI), a novel algorithm that integrates paired multi- omic data through similarity graph fusion, enhanced with a multi-view variational graph auto-encoder. SCPRO-VI incorporates a biologically guided distance met- ric and a multi-view graph-based embedding approach to effectively capture cross-modality relations. Extensive benchmark on multi-omic CITE-seq datasets shows that SCPRO-VI significantly enhances inter-cell type heterogeneity and identifies biologically meaningful sub-clusters that remain indistinguishable by existing methods. These results demonstrate robustness of SCPRO-VI and its potential to address key challenges in single-cell multi-omic data integration.