CMIGAT: Joint learning via Cyclic Modality-Interaction Graph attention for multi-omics integration.

Abstract

With the rapid development of high-throughput sequencing technology, integrating multi-omics data has become a necessary means to elucidate complex disease mechanisms and achieve precision diagnosis. However, existing methods still face two major challenges: (1) the difficulty of effectively and accurately extracting cross-omics shared representations; and (2) the lack of effective strategies to combine specific and shared representations. To address these challenges, we propose the Cyclic Modality-Interaction Graph Attention Network (CMIGAT), which unifies specificity extraction, shared alignment, and topological fusion in an end-to-end framework. CMIGAT comprises three coupled modules. Omics-specific features are first extracted via graph convolutional encoders with reconstruction regularization and confidence learning. We then extract shared features directly from the raw omics inputs through a lightweight linear alignment that bypasses the deep modality-specific encoders, and apply dual-alignment constraints (Maximum Mean Discrepancy and semantic consistency) to ensure cross-modal distributional and semantic agreement. For multi-omics integration, we propose the Cyclic Modality-Interaction Graph Integration Module (CMIGM). In this module, a Cyclic Modality-Interaction Graph (CMIG) is designed to integrate the shared and specific features of each omics, and a Graph Attention Network (GAT) is used to execute cross-modal information propagation, whereby effective information interaction and robust feature aggregation are achieved. Extensive experiments on six public benchmarks (ROSMAP, BRCA, LGG, KIPAN, GBM, and OV) show that CMIGAT achieves the best or competitive performance, ranking first on the large majority of metrics across the benchmarks. The two four-omics datasets (GBM and OV) further show that the framework scales naturally to more modalities. Ablation studies confirm the necessity and complementarity of each module. Shapley-based biomarker analysis on BRCA, together with KEGG and GO enrichment analyses, identifies biologically meaningful features closely associated with cancer-related pathways. CMIGAT effectively addresses the challenges of cross-omics shared representation extraction and specific-shared feature combination, achieving superior classification and interpretable biomarker identification. It provides a useful computational tool for multi-omics tasks such as cancer subtype classification and biomarker screening.