Predicting neoadjuvant therapy response in breast cancer from preoperative biopsy via spatial-semantic-differential learning and interpretable clinicopathological-guided fusion.

Predicting pathological complete response (pCR) to neoadjuvant therapy (NAT) in breast cancer remains challenging due to high tumor heterogeneity and disparities across data modalities. This study introduces a multimodal learning framework that integrates whole-slide image (WSI) from preoperative biopsy with clinicopathological (CP) variables to predict pCR. The framework is built on two novel components: (1) a spatial-semantic-differential (SSD) learning layer that jointly models the multi-view heterogeneity of the tumor microenvironment in WSIs, and (2) an interpretable, CP-guided (ICG) fusion strategy that leverages CP variables to steer the fine-grained integration of WSI representations, further enriched by transcriptomic profiling. This design ensures dual-layer biological interpretability-semantic (linking CP variables to tissue types) and molecular (connecting decisions to pathways). Evaluated on a retrospective multi-center cohort of 950 breast cancer patients, our method achieved ROC-AUCs of 0.845 (95% CI: 0.801-0.886) on the internal set and 0.815 (95% CI: 0.755-0.873) on the external set, outperforming state-of-the-art benchmarks. Subgroup analysis confirmed robust performance across molecular subtypes (Luminal, HER2+, TNBC), and disease-free survival stratification affirmed its prognostic relevance, highlighting its potential to guide personalized treatment planning.