JMIR Publications

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. Multi-omics profiling—including genomics, transcriptomics, and epigenomics—offers a powerful avenue for characterizing this complexity and enabling more precise patient stratification.

Objective:

Address key challenges in multi-omics analysis, including high dimensionality, cross-modality heterogeneity, limited sample size, and the lack of explicit interaction modeling for an improved survival stratification of ovarian cancer patients, which were not fully covered by the literature. In response to this gap, an ensemble of deep learning techniques was employed.

Methods:

We integrate multi-omics data from The Cancer Genome Atlas (TCGA) and develop a five-stage deep learning pipeline that combines variational autoencoders (VAEs) for data augmentation and nonlinear dimensionality reduction with graph convolutional neural networks (GCNNs) for modeling interaction-aware representations. Latent embeddings derived from the VAE are clustered using k-means to obtain candidate survival subgroups, and their prognostic relevance is evaluated using Cox proportional hazards modeling and Kaplan–Meier survival analysis

Results:

A database of 292 patients, from The Cancer Genome Atlas (TCGA), with 99,322 molecular features, was screened. The final cohort used for survival modeling consisted of n = 49 patients. Two-dimensional PCA projection of the VAE with unsupervised clusters reports 3 clusters (k = 3) with a silhouette = 0.086. The coefficients of log–hazard ratios (HR) and 95% confidence intervals estimated using the Cox proportional hazards model lie close to zero. The Kaplan–Meier survival curves estimated for the three clusters exhibit a substantial overlap; nevertheless, Cluster 2 shows a somewhat steeper decline in survival probability. The cluster comparisons do not reach statistical significance based on pairwise log-rank tests (p = 0.7096 for Cluster 0 vs 1, p = 0.3341 for Cluster 0 vs 2, and p = 0.5186 for Cluster 1 vs 2).

Conclusions:

Although the observed cluster separation and survival differences are modest, the results demonstrate the feasibility of combining deep generative models with graph-based learning to uncover diffuse but potentially meaningful prognostic structure in ovarian cancer. These findings support the continued development of integrated deep learning frameworks for multi-omics-driven precision oncology.