OncoGAT- a Novel Multimodal Graph Attention Network for Precision Oncology

Overview

Precision oncology faces a fundamental challenge: how can we combine diverse molecular data types to uncover patterns hidden from traditional single-omics analyses? Current computational frameworks remain fragmented, addressing isolated tasks such as subtype classification or survival prediction without offering the comprehensive decision support clinicians urgently need. We hypothesized that graph-based attention mechanisms could bridge this divide by modeling molecular relationships more effectively than existing approaches. OncoGAT introduces a novel multimodal Graph Attention Network that represents each patient as a node in a molecular similarity graph constructed from genomics, transcriptomics, methylation, and copy-number variation data. Our methodology began with LASSO regression using five-fold cross-validation to reduce roughly 27,000 molecular features to 500 informative markers, ensuring computational efficiency while preserving biological relevance. Using cosine similarity metrics, we constructed patient similarity graphs and implemented a three-layer GATv2Conv architecture with multi-head attention mechanisms. This framework jointly performs PAM50 subtype classification, survival estimation through Cox proportional hazards modeling, and treatment plan recommendation —unifying three critical clinical tasks within a single interpretable system. The attention mechanism automatically identifies the most relevant molecular neighbors for each patient, providing unprecedented insight into personalized cancer signatures. Our validation using the METABRIC dataset (n=1,411) demonstrates that OncoGAT significantly outperforms existing approaches across all clinical tasks. Subtype classification accuracy reached 92.1%, outperforming prior models by 15–20%. Challenging subtypes including Basal-like and Claudin-low achieved 97% recall, addressing a critical diagnostic need for rare subtype identification. Prognostic modeling yielded a concordance index of 0.713 (p<0.001) while treatment recommendations aligned with clinical guidelines. Crucially, attention weight analysis revealed biologically validated biomarkers including ERBB2, ESR1, and TP53, confirming the model's interpretability and clinical relevance. To translate this research into practice, we developed a user-friendly Streamlit-based clinical decision support application that delivers real-time predictions with interpretable attention visualizations. Clinicians can explore patient-specific molecular signatures through natural language queries, reducing diagnostic time from hours to minutes while maintaining complete transparency in decision-making. The system is currently piloted in tier-2 cities across developing regions, where specialist access remains limited. Future work will validate the framework across additional cancer types, potentially establishing this approach as a cornerstone technology for precision medicine implementation.