M3FusionNet: Cross-cohort multimodal prediction of breast cancer biomarkers.

BACKGROUND: Accurate prediction of breast cancer biomarkers - including oestrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), molecular subtypes (PAM50), and the proliferation marker Ki-67 (MKI67) - is essential for treatment stratification and prognosis. Existing computational approaches predominantly rely on single-cohort data, limiting their clinical generalisability. METHODS: We propose M3FusionNet (Multi-modal, Multi-scale, Multi-source Fusion Network), a comprehensive multimodal deep learning framework integrating haematoxylin-and-eosin (H&E) whole slide images (WSIs), RNA-seq, miRNA, proteomics, and clinical variables for simultaneous prediction of five breast cancer biomarkers. Feature extraction from WSIs was performed using ResNet50 (2,048-dimensional embeddings), with subsequent feature-level fusion across modalities. Gradient-boosted models (XGBoost, CatBoost) were employed for both classification (ER, PR, HER2, PAM50) and continuous regression (MKI67). Crucially, M3FusionNet is evaluated through three experimental protocols spanning TCGA-BRCA n ≈ 1036) and CPTAC-BRCA (n ≈ 134), with comprehensive domain adaptation via Macenko stain normalisation and ComBat batch correction. RESULTS: In-distribution performance on TCGA reached AUC values of 0.99 (ER), 0.96 (PR), 0.98 (HER2), and 0.96 (PAM50 macro-AUC), with an overall AUC of 0.97. Cross-dataset generalisation exhibited a controlled 13-22% AUC reduction (E1: Overall AUC 0.82), substantially mitigated by harmonised preprocessing and multimodal fusion. The combined TCGA + CPTAC model (M3FusionNet-E3) demonstrated superior robustness with Overall AUC of 0.90, particularly for HER2-enriched and triple-negative subtypes. Proteomics integration (CPTAC-exclusive) improved HER2 AUC by 2-3 %age points. CONCLUSIONS: M3FusionNet establishes one of the first comprehensive cross-cohort multimodal evaluation framework for breast cancer biomarker prediction, demonstrating that proteogenomic integration and domain-adapted fusion substantially improve generalisation. The protein-level Ki-67 regression from CPTAC mass spectrometry data constitutes a novel, clinically meaningful regression target. M3FusionNet offers a scalable pathway towards robust, multimodal computational pathology applicable across diverse clinical settings.