INTERPRETABLE DEEP LEARNING ON INTEGRATED MULTI-OMICS RECOVERS SUBTYPE-SPECIFIC REGULATORY FEATURES AND ESTROGEN-RESPONSE SIGNALLING IN BREAST INVASIVE CARCINOMA

Abstract

Breast invasive carcinoma is a molecularly heterogeneous disease and the clinically relevant intrinsic subtypes are the result of a coordinated alteration of the genome, transcriptome and epigenome. With analysis limited to a single molecular layer, somewhat of this structure is obtained, but few analyses correlate the identity of subtypes with the regulatory features that differentiate them. We assembled matched gene expression, gene-level DNA methylation and microRNA expression profiles for 875 breast carcinoma samples annotated with PAM50 intrinsic subtype and asked whether an interpretable model trained on all three layers could both classify subtypes and nominate the molecular features responsible for each assignment. The embeddings of each omics layer were fed to their respective encoders and then a sample-wise attention mechanism was used to fuse the embeddings into a unified representation before the subtype prediction task. Feature attribution based on Shapley additive explanations was then used to rank genes, methylation features and microRNAs by their contribution to each subtype. On a held-out test set the integrated model reached a macro-averaged area under the receiver operating characteristic curve of 0.958, matching the strongest single-omics model (expression alone) while drawing on methylation and microRNA information that expression-only models cannot access. Attention weights varied across subtypes in a biologically sensible manner, with the luminal subtypes relying most heavily on expression and the normal-like group distributing weight toward microRNA. Genes prioritised by attribution were significantly enriched for established breast-cancer genes relative to the analysed background panel (3.2-fold, hypergeometric P = 4.1 × 10⁻⁷), and the enrichment held within four of the five subtypes. Estrogen-response signalling was the dominant enriched programme, most strongly in the HER2-enriched group. The prioritised features recovered canonical subtype markers including ESR1, FOXA1, GATA3, TFF1 and the estrogen regulated microRNA hsa-mir-375 without any prior knowledge supplied to the model. These results indicate that explainable multi-omics integration recovers biologically coherent subtype programmes and offers a transparent and reproducible route to candidate biomarker generation in breast cancer genomics.