DMFF: a deep learning-based multi-omics fusion framework for survival prediction and subtype classification in breast cancer

Abstract

Intra-tumor molecular heterogeneity contributes substantially to survival differences among patients with breast cancer. Although deep learning-based multi-omics integration has shown promise for prognostic modeling, existing approaches may be affected by information leakage during model evaluation and often focus on either survival prediction or molecular subtype discovery. In this study, we developed the Deep Multi-modal Fusion Framework (DMFF), an integrated computational framework that supports continuous patient-level survival risk estimation and downstream de novo molecular subtype discovery within a unified analytical workflow. DMFF uses modality-specific denoising autoencoders to learn compact representations from mRNA and miRNA data, while a protein–protein interaction network-guided graph convolutional network is used to encode proteomic features. The resulting modality-specific representations are adaptively integrated through an attention-based fusion module. A Cox partial likelihood loss and a pairwise ranking loss are jointly optimized to generate continuous survival risk scores. To minimize information leakage, data imputation, standardization, feature preselection, hyperparameter optimization, and model checkpoint selection were performed within a nested five-fold cross-validation framework. Under nested cross-validation, the pure-omics DMFF model achieved a C-index of 0.656. After integrating the learned multi-omics representations with clinical characteristics using a Random Survival Forest, the joint model achieved a C-index of 0.706 and outperformed the evaluated baseline models in prognostic discrimination. Multivariable Cox regression further indicated that the DMFF-derived pure-omics risk score provided prognostic information independent of conventional clinical variables. Unsupervised clustering of the fused multi-omics representations identified three molecular subtypes with distinct survival outcomes, and repeated subsampling analysis supported the internal stability of the three-subtype solution. Tumor microenvironment analysis revealed distinct immune-cell enrichment patterns across the three subtypes. Comparison with PAM50 intrinsic subtypes suggested that the DMFF-derived and PAM50 classifications capture partially complementary aspects of breast cancer heterogeneity. Because the present study did not include immunotherapy exposure or treatment-response data, the potential relationship between the identified immune profiles and immunotherapy response should be regarded as exploratory and hypothesis-generating. Overall, DMFF provides a leakage-controlled framework for multi-omics survival prediction and downstream molecular subtype discovery, although its generalizability requires further validation in independent multi-center cohorts.