Explainable graph learning for multimodal single-cell data integration

Understanding cellular heterogeneity and identifying functionally distinct subpopulations are central goals in single-cell analysis. Integrating paired multi-omic data, such as transcriptomic and proteomic profiles from the same cells, offers a more comprehensive view of cell states. However, current integration methods often struggle to balance expressiveness with interpretability, largely due to the complex and non-linear relationships across modalities.

We present Single-Cell PROteomics Vertical Integration (SCPRO-VI), a new algorithm designed to integrate paired single-cell multi-omic data. SCPRO-VI introduces a biologically informed distance metric to construct modality-specific similarity graphs. These graphs are then used to learn omic-wise cell embeddings through variational graph auto-encoders. The resulting embeddings are fused using an auto-encoder to produce a unified representation of each cell. This architecture allows for both effective cross-modality integration and interpretability by enabling backtracking of cell relationships via the initial similarity graphs. We evaluated SCPRO-VI using multiple CITE-seq datasets and observed a significant improvement in distinguishing cell types compared to existing approaches. The method also uncovered biologically relevant subpopulations that remained indistinct in other integrated representations.

SCPRO-VI offers a robust and interpretable framework for integrating paired single-cell multi-omic data. Its ability to enhance cell type separation and uncover meaningful sub-clusters suggests its utility in advancing our understanding of cellular diversity and regulatory mechanisms in complex tissues.