Multi-omic cancer subtype clustering with random-walk-refined cross-view contrastive learning

Cancer heterogeneity complicates precision oncology, and robust multi-omic subtyping remains challenging. Existing contrastive integration methods often force heterogeneous omics into a single shared latent space and treat all non-matching samples as negatives, which can suppress modality-specific structure and separate biologically similar patients. We present Subtype-DIVIDE, a self-supervised multi-view contrastive framework for integrative cancer subtyping. The model encodes each omic modality separately and performs target-conditioned cross-view alignment, using the target modality’s affinity structure to guide inter-view consistency. To reduce false negatives, Subtype-DIVIDE further refines contrastive supervision with per-omic random-walk soft-positive masks constructed from modality-specific affinity graphs. Across nine TCGA cohorts spanning copy number, DNA methylation, mRNA, and miRNA data, Subtype-DIVIDE achieved statistically significant survival separation in all datasets and the highest mean number of clinically enriched variables among the compared methods. Cox models, leave-one-cluster-out analyses, and technical covariate tests supported the clinical relevance and robustness of the inferred subtypes. Ablation studies showed that cross-view alignment is essential, while diffusion-based soft positives improved robustness. A KIRC case study identified coherent subtype programs involving metabolism, vascular and inflammatory signaling, extracellular-matrix remodeling, and epithelial transport. External validation on two independent cohorts, performed without hyperparameter retuning, supported the generalizability of the inferred subtypes. In 222 CPTAC-3 clear-cell renal carcinoma cases, the subtypes were significantly associated with pathologic stage and T category across all training repeats and recapitulated the KIRC subtype programs. In 1,980 METABRIC breast tumors, they consistently showed significant survival separation and enrichment for age, nodal status, and stage across all repeats.