Breast cancer patients exhibit variable responses to neoadjuvant therapy (NAT), necessitating robust predictive biomarkers. We developed an artificial intelligence (AI)-driven integrated predictive model (IPM) combining histopathological, clinical, and immune features to address this challenge. Using whole-slide images from 1035 patients across four centers, we compared tumor epithelium (TE-score), stroma (TS-score), and whole-tumor (TR-score) deep learning biomarkers, identifying TR-score as optimal (AUC = 0.729 vs. 0.686/0.719 for TE/TS-scores). The IPM, incorporating TR-score and clinical variables, demonstrated superior NAT response prediction versus clinicopathological models (validation AUC = 0.780 vs. 0.706, p < 0.001), with 10% higher accuracy. Immune profiling further enhanced performance (AUC = 0.831 vs. 0.822, p = 0.183). These results establish the biological and clinical validity of TR-score for characterizing tumor-stroma interactions, with IPM providing a generalizable framework for precision oncology. The model's stability across multicenter cohorts (AUCs 0.781-0.816) and incremental value of immune data suggest its utility in guiding NAT decision-making and trial stratification.