Abstract Neoadjuvant dual HER2 blockade with trastuzumab and pertuzumab plus chemotherapy represents the current standard-of-care for HER2-positive breast cancer. However, treatment responses remain heterogeneous, underscoring the lack of clinically practical tools for predicting treatment efficacy and informing personalized therapy. Here, we developed HER2-LADDER (Layered AI-based Dual-targeteD anti-HER2 Recommendation), a spatially interpretable and clinically accessible artificial intelligence framework that integrates clinicopathological and spatial topological features from routine hematoxylin and eosin (H&E) and HER2 immunohistochemistry (IHC) slides. Using these spatially derived features, HER2-LADDER accurately predicted response to neoadjuvant TCbHP/PCbHP, achieving AUCs of 0.944 in the model construction cohort ( N = 276), 0.917 in the temporal validation cohort ( N = 82), and 0.869 in the trial-based validation cohort ( N = 85). On the basis of HER2-LADDER scores, patients were stratified into Low (highly responsive), Medium (responsive), and High (resistant) groups, identifying candidates for treatment de-escalation (THP or TCbH/PCbH), standard-of-care (TCbHP/PCbHP), or alternative regimens (e.g., next-generation anti-HER2 antibody‐drug conjugates), respectively. Importantly, Xenium in situ profiling further revealed biological correlates underlying model predictions, including HER2-enriched tumor cell aggregation and neutrophil-helper T-cell interactions, thereby highlighting the mechanistic interpretability of the model. Collectively, HER2-LADDER unites digital pathology and high-resolution spatial profiling into a clinically accessible AI framework, offering a robust, transparent, and biologically grounded tool to tailor individualized HER2-targeted therapy optimization.
Wu et al. (Thu,) studied this question.