Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OmicsCL: Unsupervised Contrastive Learning for Cancer Subtype Discovery and Survival Stratification

Atahan Karagoz

TL;DR

OmicsCL presents an unsupervised, survival-aware contrastive framework for integrating three omics modalities (gene expression, DNA methylation, miRNA) to discover cancer subtypes and stratify survival without subtype labels. It introduces a survival-aware contrastive loss layered on top of cross-omics NT-Xent objectives, producing joint embeddings that reflect both molecular similarity and prognostic heterogeneity. Evaluated on TCGA-BRCA, OmicsCL achieves a test-set C-index of 0.7512 and significant Kaplan–Meier separation, demonstrating prognostic relevance without supervision and revealing configurable trade-offs between survival stratification and subtype coherence. The method uses lightweight per-modality encoders, a simple fusion strategy, and demonstrates the potential of contrastive objectives for unsupervised, survival-informed discovery in high-dimensional, heterogeneous omics data.

Abstract

Unsupervised learning of disease subtypes from multi-omics data presents a significant opportunity for advancing personalized medicine. We introduce OmicsCL, a modular contrastive learning framework that jointly embeds heterogeneous omics modalities-such as gene expression, DNA methylation, and miRNA expression-into a unified latent space. Our method incorporates a survival-aware contrastive loss that encourages the model to learn representations aligned with survival-related patterns, without relying on labeled outcomes. Evaluated on the TCGA BRCA dataset, OmicsCL uncovers clinically meaningful clusters and achieves strong unsupervised concordance with patient survival. The framework demonstrates robustness across hyperparameter configurations and can be tuned to prioritize either subtype coherence or survival stratification. Ablation studies confirm that integrating survival-aware loss significantly enhances the predictive power of learned embeddings. These results highlight the promise of contrastive objectives for biological insight discovery in high-dimensional, heterogeneous omics data.

OmicsCL: Unsupervised Contrastive Learning for Cancer Subtype Discovery and Survival Stratification

TL;DR

OmicsCL presents an unsupervised, survival-aware contrastive framework for integrating three omics modalities (gene expression, DNA methylation, miRNA) to discover cancer subtypes and stratify survival without subtype labels. It introduces a survival-aware contrastive loss layered on top of cross-omics NT-Xent objectives, producing joint embeddings that reflect both molecular similarity and prognostic heterogeneity. Evaluated on TCGA-BRCA, OmicsCL achieves a test-set C-index of 0.7512 and significant Kaplan–Meier separation, demonstrating prognostic relevance without supervision and revealing configurable trade-offs between survival stratification and subtype coherence. The method uses lightweight per-modality encoders, a simple fusion strategy, and demonstrates the potential of contrastive objectives for unsupervised, survival-informed discovery in high-dimensional, heterogeneous omics data.

Abstract

Unsupervised learning of disease subtypes from multi-omics data presents a significant opportunity for advancing personalized medicine. We introduce OmicsCL, a modular contrastive learning framework that jointly embeds heterogeneous omics modalities-such as gene expression, DNA methylation, and miRNA expression-into a unified latent space. Our method incorporates a survival-aware contrastive loss that encourages the model to learn representations aligned with survival-related patterns, without relying on labeled outcomes. Evaluated on the TCGA BRCA dataset, OmicsCL uncovers clinically meaningful clusters and achieves strong unsupervised concordance with patient survival. The framework demonstrates robustness across hyperparameter configurations and can be tuned to prioritize either subtype coherence or survival stratification. Ablation studies confirm that integrating survival-aware loss significantly enhances the predictive power of learned embeddings. These results highlight the promise of contrastive objectives for biological insight discovery in high-dimensional, heterogeneous omics data.
Paper Structure (27 sections, 3 equations, 4 figures, 1 table)

This paper contains 27 sections, 3 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Definition
  5. Model Architecture
  6. Contrastive Objective Across Omics
  7. Survival-Aware Contrastive Loss
  8. Joint Training
  9. Clustering and Evaluation
  10. Experiments
  11. Dataset
  12. Data Preprocessing
  13. Implementation Details
  14. Evaluation Metrics
  15. Clustering Metrics
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Combined survival histograms over training epochs. This visualization captures the evolution of censored and deceased event distributions throughout training, highlighting the temporal dynamics our model encodes into the embedding space.
  • Figure 2: Kaplan–Meier curves stratified by predicted clusters. Clear survival separation is observed, especially between clusters 0 and 2.
  • Figure 3: 2D t-SNE visualization of embeddings colored by predicted clusters. Distinct subpopulations emerge despite unsupervised training.
  • Figure 4: 2D UMAP visualization of embeddings colored by PAM50 subtypes. The model partially recovers subtype structure without supervision.