scGAT: A Graph Attention Network Approach for Single-Cell Multi-Omics Data Analysis and Biomarker Discovery

Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of cellular heterogeneity; however, many analytical methods focus on differential expression, overlooking intercellular interactions and disease progression. Here, we present scGAT, a graph attention network (GAT)-based approach trained on scRNA-seq data to classify cell states and extract disease-associated biomarkers using a custom-designed interpretation module. To address the overrepresentation of regulatory RNAs (lncRNAs and rRNAs) in solid tumors, scGAT incorporates a dual-penalty mechanism to dynamically adjust their attention weights, enhancing robustness and target prioritization. Validated across 20 GEO datasets spanning diverse diseases, scGAT has been shown to identify novel biomarkers, including markers missed by traditional methods. The model is designed to enable direct multi-stage disease analysis and can also generalize to scATAC-seq data, offering a scalable, robust solution for multi-omics single-cell analysis. This approach has the potential for uncovering regulatory mechanisms underlying disease progression and accelerating therapeutic target discovery.