Integrating Multiomics Data Using a Correlation Based Graph Attention Network for Subtype Classification in Lower Grade Glioma

Accurate classification of cancer subtypes is crucial to personalized therapies and targeted intervention. In this study, we proposed BioGAT-LGG, a deep learning framework that integrates the multi-omics data, namely mRNA, miRNA, and DNA methylation, by a correlation-based Graph Attention Network (GATv2) for biomarker discovery and subtype classification of Lower-Grade Glioma (LGG). Unlike existing methodologies that depend on external biological priors such as protein-protein interaction networks or reference biological graphs, BioGAT-LGG creates sample-driven correlation graphs so that the model can learn biologically meaningful molecular interactions. For the sake of improving feature interpretability and dimensionality reduction, LASSO regression is performed initially before training. The model attains an accuracy of 98.03% among the variables in precision (98.12%), recall (97.74%), and F1-score (97.87%), all supported by stratified 10-fold cross-validation. hsa-mir-3936, MTCO1P40, and CCND2 represent key markers, whereas KEGG enrichment throws insights into other pathways like PI3K-Akt signaling, Small Cell Lung Cancer, and Transcriptional Misregulation in Cancer. The findings thus suggest that BioGAT-LGG can potentially serve to support clinically relevant subtype classification and biomarker-driven decision-making. This framework thus lays a scalable foundation for the multi-omics integration in oncology, which can further be adopted in other tumor types.