Prototype contrastive alignment with clinical attention enables multi-modal breast cancer molecular subtyping

Accurate molecular subtyping of breast cancer is essential for treatment planning, yet computational methods face challenges from class imbalance and variable-quality histopathology features. We propose a multi-modal framework that integrates histopathology image features with genomic expression data through prototype-based contrastive alignment and clinical guideline-aware attention. The framework computes class-conditional prototypes in each modality’s latent space and aligns them via a bidirectional contrastive loss, decoupling representation learning from sample-level pairing requirements. A clinical attention mechanism encodes established breast cancer biomarker profiles as structured prior knowledge within the fusion architecture. Evaluated on 916 patients from TCGA-BRCA with PAM50 molecular labels using CLAM histopathology features and RNA-Seq expression data, the framework achieves macro-averaged F1 of 0.843 ± 0.012 in 5-fold cross-validation. Clinical attention is the most impactful component (contributing + 5.8% macro F1 over uniform fusion), followed by genomic-supervised feature learning via knowledge distillation (+ 2.1%). An exploratory Histogenomic Discordance Score measuring cross-modal prediction disagreement shows directionally consistent but non-significant association with disease-free survival (HR = 1.91, p = 0.25). These results demonstrate that structured clinical knowledge injection through guideline-aware attention provides measurable improvements in multi-modal breast cancer subtyping under realistic feature quality constraints.