Cell-type-specific Multi-omics Integration for Genetic Risk Assessment in Diabetes: A Computational Framework Combining Single-cell Profiling, Spatial Transcriptomics and Graph Neural Networks

Background: Traditional bulk-tissue-based genetic risk scores (GRS) lack cellular resolution and fail to capture the mechanistic basis of disease risk at the level of individual cell types and tissue architecture. The aim of this study is to develop a computational framework that integrates single-cell multi-omics and spatial transcriptomics to refine diabetes risk prediction by mapping genetic variants to cell-type-specific regulatory programmes in pancreatic islets. Methods: We introduce the Diabetes Risk Prediction Model (DRPM), which leverages single-cell RNA sequencing, single-cell ATAC sequencing and spatial transcriptomics. Central to the framework is the Multi-Omics Regulatory Network, which links genetic variants with chromatin accessibility and gene expression in a cell-type-specific manner. The model also incorporates spatial ligand–receptor interaction data to quantify disruptions in intercellular communication. The output is a vector of cell-type-specific risk contributions, interpreted through a Graph Neural Network with attention mechanisms and optimised jointly with conventional risk classifiers. Results: Our approach outperforms traditional GRS in diabetes risk prediction accuracy and reveals mechanistic insights into how risk variants perturb specific islet cell types and intercellular signalling pathways. It provides an interpretable, high-resolution view of genetic risk across the tissue microenvironment. Conclusion: The DRPM framework advances genetic risk assessment by resolving risk at the level of cell types and spatial context within tissues. It offers a scalable and generalisable method for cell-type-aware risk modelling in diabetes and other complex diseases.