2662 - Pathway-Informed Graph Neural Networks for Cancer Phenotype Classification in Large-Scale Cancer Datasets

Purpose/Objective(s): Cancer is not a one-size-fits-all disease. Each patient has a distinct molecular profile, and even within a single individual, tumor sites may demonstrate genomic variability. In addition, tumors evolve under therapeutic pressure, further complicating biomarker development and treatment selection. Although technological advances now enable comprehensive multi-omics profiling, translating these high-dimensional datasets into clinically actionable insights remains challenging, in part because many analytical approaches fail to reflect the biological organization underlying these data. Cancer biology operates through coordinated networks rather than isolated gene effects. However, many existing models emphasize molecular expressions but underutilize their underlying pathways. Our hypothesis is that by developing a pathway-informed graph learning framework that integrates dynamic molecular expression data with curated pathway architecture, we can more faithfully model tumor biology and improve prediction of clinically relevant cancer phenotypes.

Materials/Methods: We propose a graph-based learning framework that integrates established biological pathway architecture with sample-specific molecular activity. Curated pathway networks were obtained from a public repository of manually curated and biologically validated molecular interaction maps, where each pathway represents a defined biological process and encodes coordinated gene-gene interactions. For each patient, the most transcriptionally active genes were identified, and pathways containing these genes were merged into a single patient-specific molecular interaction graph. A graph neural network classifier was then trained directly on these pathway-informed graphs to predict clinically relevant disease phenotypes. The framework was evaluated across The Cancer Genome Atlas (TCGA), a prostate cancer cohort, the Tabula Sapiens atlas, and the Zheng 68K dataset from 10x Genomics.

Results: The proposed model achieved classification accuracy of 94.3% on the TCGA pan-cancer cohort, outperforming structure-agnostic and feature-based graph learning approaches by up to 50.5%. These findings demonstrate that incorporating biological structure at the graph level provides meaningful additional information for cancer classification tasks.

Conclusion: The proposed framework demonstrates that explicitly encoding pathway-level biological structure improves cancer prediction. By moving beyond treating genes as independent features and instead representing tumors as interconnected molecular systems, this framework provides a scalable and interpretable foundation for precision oncology applications, with potential to enhance patient stratification and biomarker development.