TRUST: An MRI-based AI framework for non-invasive prediction of treatment response and prognosis in breast cancer.

e12556 Background: The tumor microenvironment (TME) plays a pivotal role in therapeutic response and prognosis in breast cancer. Although stromal tumor-infiltrating lymphocytes (sTILs) and tumor-stroma ratio (TSR) are established biomarkers, they capture immune infiltration and stromal architecture separately and are limited by biopsy sampling bias and invasiveness. We aimed to establish an integrated immune-stromal classification and develop a non-invasive, MRI-based artificial intelligence framework, the TILs-TSR Unification System (TRUST), to predict immune-stromal phenotypes and guide treatment stratification. Methods: sTILs and TSR were independently assessed to define four immune-stromal phenotypes: sTILs-high/TSR-high (HH), sTILs-high/TSR-low (HL), sTILs-low/TSR-high (LH), and sTILs-low/TSR-low (LL). These joint labels served as pathological ground truth for training TRUST using multiparametric MRI from a multicenter cohort of 687 patients across eight institutions. Model performance was evaluated in independent cohorts for neoadjuvant therapy (NAT; n = 351) response and disease-free survival (DFS; n = 190). Spatial proteomic profiling using PhenoCycler-Fusion was performed on tissue microarrays (n = 48) to characterize biological features across subtypes. Results: The integrated sTILs-TSR classification demonstrated stronger associations with NAT response and DFS than either metric alone. The HH subtype exhibited the highest pathological complete response (pCR) rate (78%), whereas the LL subtype showed the poorest prognosis. TRUST achieved high performance for tumor detection (AUC, 0.921; accuracy, 0.843) and immune-stromal phenotype prediction in training (AUC, 0.861; accuracy, 0.748) and validation cohorts (AUC, 0.821; accuracy, 0.721). Notably, in external cohorts (TNBC and Luminal subtypes), TRUST alone stratified pCR rates in neoadjuvant chemotherapy (73.0% vs. 10.1%) and chemoimmunotherapy with anti-PD-1 (67.8% vs. 14.3%) between predicted NAT-sensitive (HH) and NAT-resistant (LL) groups. TRUST also significantly stratified DFS, with no 3-year DFS events observed in the predicted HH group. Spatial proteomics revealed distinct immune-stromal architectures, characterized by enriched Granzyme B⁺ CD8⁺ T cells, abundant antigen-presenting cells, and close tumor-immune interactions in HH tumors, while LL tumors exhibited dominant myCAF-rich stroma, increased M2-like macrophages, and sparse CD8⁺ T-cell infiltration. Conclusions: The integrated sTILs-TSR classification provides superior prognostic and predictive value by capturing coordinated immune and stromal architecture. Leveraging multiparametric MRI and deep learning, TRUST establishes a biologically interpretable, non-invasive imaging biomarker that enables clinically actionable risk stratification and supports therapeutic escalation or de-escalation.