A multi-use graph neural network framework for single-cell multi-omics data

The advances of single-cell multi-omics profiling technologies in biomedical research offer an unprecedent opportunity for understanding cell heterogeneity and subpopulations. There are many statistical and computational challenges in the integrative analyses of these rich data, including sequencing sparsity, complex differential patterns in gene expression, and different platforms and panels used to generate multiple single-cell multi-omics. In this presentation, we introduce a multi-use graph neural network framework that can effectively impute and predict missing sequencing panels, integrate multi-omics single-cell datasets, and formulate and aggregate cell–cell relationships with graph neural networks. Comprehensive simulations and applications on multiple single multi-omics datasets demonstrate that our proposed method is a powerful tool for general single-cell data multi-omics analyses that outperforms the existing methods for protein prediction, gene imputation and cell clustering.