scGUMI: a single cell Graph-based Unified Multi-modal Integration approach

Abstract

In the rapidly evolving landscape of single-cell genomics, integrated analysis of multi-modal single-cell data has emerged as a pivotal challenge to extract comprehensive biological insights. Motivated by the need to harness the complementary information across modalities, we developed scGUMI, a novel Graph-based Unified Multi-modal Integration approach. scGUMI leverages the power of graph convolutional networks to bridge the modalities. In addition, scGUMI adopts an attention mechanism to facilitate self-adapted integration, and allocate explainable weights for each modality. We applied scGUMI to cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) and 10x Chromium single cell multiome ATAC and gene expression data, demonstrating its effectiveness in revealing higher resolution of cellular heterogeneity in multiple tissue types. Notably, scGUMI excels in cross-modality imputation, addressing the inherent data sparsity and dropout issues. Compared to other single-cell imputation methods, scGUMI's cross-omics mutual imputation enhances the imputation accuracy by integrating information from the other modality. Even under significant masking, scGUMI's imputation was able to reconstruct the masked omics measurement accurately. Through this advancement, scGUMI explores the way for a deeper understanding of cellular states, highlighting the synergistic power of integrated multi-modal analysis.