PHYPERNET: P-LAPLACIAN REGULARIZATION HYPERGRAPH LEARNING FOR MULTI-OMICS CANCER SUBTYPING

Abstract

Precise cancer subtype identification plays a crucial role in precision oncology, driving targeted therapies, personalized prognoses, and a deeper understanding of tumor complexity. Nevertheless, there is a challenge in integrating multi-omic data to identify subtypes, as these datasets are heterogeneous, high-dimensional, and noisy. To address these problems, we propose PHyperNET. This integrative model is characterized by resilient correlation-based multi-view fusion, Hypergraph-based modeling of high-order structure, and p-Laplacian regularization of complex, nonlinear relationships between patient samples. It was tested using five TCGA cancer datasets: LUAD, COAD, KIRC, GBM, and BRCA, encompassing gene-expression, DNA methylation, and copy-number variation data. The model demonstrated better performance in terms of clustering accuracy and clinical significance compared to SNF, NEMO, MOFA, and DeepMO. In particular, PHyperNET achieved the maximum average silhouette score of 0.52, as well as very small log-rank p-values (e.g., 0.0006 in GBM and 0.0014 in KIRC), implying good stratification of survival by subtypes. Moreover, the enriched pathway analysis revealed that the uncovered subtypes are associated with known cancer pathways, such as WNT signalling in COAD and VEGF signalling in KIRC. These findings demonstrate that PHyperNET provides both improvements in biomarker subtype separation and clinical interpretation, as well as biological consistency. PHyperNET delivers a robust, scalable, but interpretable architecture that can capture interactions in multi-dimensional data to help achieve cancer subtyping. PHyperNET facilitates a more integrated approach to multi-omic data, informing the direction of cancer subtyping studies and potentially offering translational relevance and clinical applications.