Epigenetic profile drives accurate survival prediction in breast cancer via a multi-omics machine learning model

Technical Summary: Epigenetic Profile Drives Accurate Survival Prediction in Breast Cancer via a Multi-omics Machine Learning Model

Problem Statement
Accurate prediction of overall survival (OS) is fundamental to personalized treatment strategies in breast cancer. While tumor mutational burden (TMB) and somatic mutations have emerged as prognostic biomarkers in various malignancies, their utility in breast cancer is limited. Evidence suggests that high TMB does not reliably predict improved survival outcomes in breast cancer, unlike in other cancers where it correlates with better responses to immune checkpoint blockade. Existing prognostic models often rely on isolated data types (e.g., mutation profiles alone) or lack the integrative power to capture the full spectrum of tumor biology. Consequently, there is a need for robust models that integrate multi-dimensional tumor features—including genomic, transcriptomic, proteomic, and epigenetic data—to overcome the limitations of mutation-centric approaches and improve prognostic accuracy.

Methodology
The study leveraged data from The Cancer Genome Atlas (TCGA) Breast Invasive Carcinoma (BRCA) cohort, comprising 802 female patients with comprehensive multi-omic and clinical annotations. The workflow involved the following steps:

1. Data Integration and Feature Selection: The authors integrated genomic (somatic mutations, TMB), transcriptomic (RNA-seq), proteomic (Reverse-Phase Protein Array and mass spectrometry), epigenetic (DNA methylation), and clinical features. Initial correlation analyses (Pearson and Spearman) and Lasso-penalized Cox proportional hazards regression were used to identify features significantly associated with OS. This step reduced dimensionality and selected a sparse set of predictive variables.
2. Model Development (BANDOL): A Random Survival Forest (RSF) model, named BANDOL (Breast cancer Analysis with Neoplastic Data and Omics Learning), was trained on the selected features. RSF was chosen for its ability to handle high-dimensional, censored survival data without assuming proportional hazards.
3. Validation Strategy: The model's performance was evaluated using a nested cross-validation approach. The outer loop (5-fold) estimated generalization performance, while the inner loop (5-fold) optimized hyperparameters via grid search. Performance was measured using the Concordance Index (C-index) and time-dependent Receiver Operating Characteristic (ROC) curves with cumulative dynamic Area Under the Curve (AUC).
4. Interpretability: Shapley Additive Explanations (SHAP) values were computed to quantify the contribution of each feature to the model's predictions, identifying the most influential drivers of survival.
5. Transferability Testing: To assess generalizability, the breast cancer-trained model was applied to two independent TCGA cohorts representing distinct cancer types: Uterine Cancer (UEC, n=227) and Lower-Grade Glioma (LGG, n=189).

Key Results

• Limited Predictive Power of Mutations: Correlation and Cox regression analyses revealed that mutational features, including TMB, had negligible correlation with OS in the breast cancer cohort. Only a handful of clinical features (e.g., specific drug usage, age) and non-mutational molecular features showed significant associations with survival.
• Superiority of Multi-omics Integration: The BANDOL model trained on multi-omics data achieved a C-index of 0.7468, significantly outperforming a model trained solely on mutational data (C-index: 0.5292). Time-dependent AUC analysis further confirmed this, with the multi-omics model achieving AUCs between 0.9 and 1.0, compared to 0.4–0.9 for the mutation-only model.
• Epigenetic Dominance: SHAP value analysis identified epigenetic features as the strongest predictors of OS. Specifically, methylation levels of ME3, PPARG, OLIG3, and SLC25A22 were critical drivers of the model's predictions.
• Biological Insights:
  • Longer OS: Associated with immune activation signatures, including naive T cells, myeloid cells, IL-2 signaling, Type I IFN pathways, and specific mRNA/protein expression patterns (e.g., IRF2, ESRRG, MAP2K6).
  • Shorter OS: Linked to immunosuppressive signatures, including TREM2+ myeloid cells, B cells, leptin signaling, and elevated levels of PELO, MAFA, and phosphorylated NFKB1.
• Cross-Cancer Transferability: When applied to UEC and LGG cohorts, the model maintained moderate predictive performance (C-index 0.57 and 0.62, respectively), outperforming mutation-only models in these contexts, though performance was lower than in the original breast cancer cohort.

Significance and Claims
The paper claims that integrating multi-omics data via machine learning significantly enhances survival prediction in breast cancer compared to models relying solely on mutation profiles or clinical features. The study demonstrates that epigenetic profiles are the primary drivers of accurate survival prediction in this context, suggesting that DNA methylation patterns capture critical regulatory information missed by genomic sequencing alone.

Furthermore, the research identifies specific actionable biomarkers and potential therapeutic targets. The authors suggest that current therapies could be combined with strategies targeting the methylation of ME3, PPARG, OLIG3, and SLC25A22, as well as dephosphorylation of NFKB1 and targeting of MAFA and PELO. The study concludes that while the model shows promise for cross-cancer transferability, its primary strength lies in the breast cancer setting, where it provides a robust framework for estimating patient risk and survival probability. The authors acknowledge limitations, noting that feature selection via Lasso-Cox may favor linear relationships and that the external validation was performed on different cancer types rather than independent breast cancer datasets, framing the cross-cancer results as an assessment of transferability rather than formal external validation.