Deep Generative and Graph-Based Representation Learning for Multiomics Survival Stratification in Ovarian Cancer: Secondary Analysis.

Abstract

Background: Ovarian cancer remains one of the most lethal gynecologic malignancies, largely due to pronounced molecular heterogeneity, nonspecific clinical presentation, and frequent diagnosis at advanced stages. Multiomics profiling-including genomics, transcriptomics, and epigenomics-offers a powerful avenue for characterizing this complexity and enabling more precise patient stratification.

Objective: This study aimed to address key challenges in multiomics analysis, including high dimensionality, cross-modality heterogeneity, limited sample size, and the lack of effective approaches for survival stratification of patients with ovarian cancer through deep representation learning.

Methods: We analyzed multiomics data from The Cancer Genome Atlas and developed a 5-stage deep learning pipeline centered on variational autoencoders (VAEs) for nonlinear dimensionality reduction and latent representation learning. A graph convolutional neural network component is described as a proposed extension for modeling interaction-aware representations but was not empirically evaluated in this study. Latent embeddings derived from the VAE were clustered using k-means, and their prognostic relevance was assessed using Cox proportional hazards modeling and Kaplan-Meier survival analysis.

Results: Following correction of a clinical-molecular harmonization issue, the final matched cohort comprised 291 patients. Silhouette analysis identified k=2 as the optimal clustering solution (silhouette=0.272). Kaplan-Meier analysis demonstrated significantly different overall survival between the two clusters (log-rank χ²₁=10.0; P=.002). Cox proportional hazards modeling estimated a hazard ratio of 0.519 (95% CI 0.343-0.785; P=.002), indicating that patients assigned to cluster 1 exhibited an approximately 48% lower hazard of death than those in cluster 0. These results demonstrate that the learned latent representations capture prognostically relevant structure within the integrated multiomics data.

Conclusions: The proposed VAE-based framework identified 2 prognostically distinct patient subgroups with significantly different overall survival. These findings demonstrate the potential of deep generative representation learning for multiomics-based survival stratification in ovarian cancer and provide a foundation for future validation in independent cohorts and the evaluation of graph-based extensions.