Multi-omics transformer integration reveals an invasive-like molecular phenotype in ductal carcinoma in situ

Abstract Even though 20–30% of cases of ductal carcinoma in situ (DCIS) advance to invasive breast cancer (IBC), cancer histology is unable to accurately differentiate between progressing and indolent disease. An integrated model describing the transcriptomic, epigenetic, and immunological factors influencing this shift has not yet been developed. A hybrid multi-modal transformer comprising gene expression, an expression-derived methylation proxy, pathway scores, and immunological markers was developed using five GEO datasets (n = 415). In 102 IBC/DCIS samples, differential expression analysis revealed 2,996 DEGs, from which a 41-gene core consensus signature was confirmed across two independent validation cohorts. Model performance was tested using five-fold cross-validation, and epigenetic risk was validated using multivariate logistic regression in a separate clinical outcomes cohort (GSE281307, n = 185). The transformer achieved AUC-ROC = 0.9318 ± 0.0132 (overall held-out AUC = 0.8964), sensitivity = 0.8214, specificity = 0.7609, accuracy = 0.7941, and F1 = 0.7917. Conventional baselines trained on the same 75-gene expression matrix achieved numerically higher AUC values (SVM 0.952 ± 0.031, logistic regression 0.942 ± 0.028, random forest 0.944 ± 0.029). However, the transformer’s main benefit is that it produces 512-dimensional multi-modal fused embeddings that distinguish DCIS from IBC much more clearly than the expression features alone (silhouette coefficient 0.986 versus 0.125; permutation p = 0.0005). Additionally, single-modality classifiers are unable to detect a repeatable invasive-like DCIS subgroup. The 41-gene core signature was confirmed by > 80% directional concordance (33 ECM/stromal upregulated; 8 epithelial identity downregulated). Invasive-like samples were identified in 15 out of 46 DCIS samples (32.6%). DCIS development was independently predicted by participation in a high-risk methylation cluster (OR = 2.40, 95% CI 1.27–4.67, p = 0.008). The HER2-enriched subtype was independently predictive (OR = 3.71, 95% CI 1.19–12.51, p = 0.028). ITGA2, SNCA, EN1, COL4A6, COL4A3, and DEFB1 were ranked as the top model predictive features by SHAP analysis on the log-odds scale, with COL4A3 consistently showing up in all model iterations. A new multimodal transformer framework has been developed to identify a subset of invasive-like ductal carcinoma in situ (DCIS), comprising 32.6% of all DCIS samples. This framework uniquely avoids disease-stage label leakage and immunological group-mean leaking, marking the first independent validation of an invasive-like molecular subgroup within DCIS utilising per-sample immune marker expression without target label proxies.