Abstract BACKGROUND: Neoadjuvant therapy is central to breast cancer treatment, yet response rates vary widely, and robust biomarkers to predict pathological complete response (pCR) are lacking. While expression-based signatures such as Oncotype DX and MammaPrint routinely guide adjuvant treatment decisions, no tools are approved for the neoadjuvant setting. Although some predictive markers have been explored for HR-positive and HER2-positive disease, triple-negative breast cancer (TNBC) lacks reliable predictors despite its aggressive nature and unmet clinical need. Recent single-cell transcriptomic studies reveal that multiple molecular subtypes can coexist within individual tumors, challenging the conventional one-subtype-per-patient paradigm. We hypothesized that modeling tumors as compositions of intra-tumoral subtypes would improve neoadjuvant therapy response prediction. METHODS: We developed BRIDGE (Breast Intratumoral Deconvolution of Gene Expression), the first computational framework that deconvolves bulk breast cancer transcriptomes to quantify the relative abundance of malignant subtype populations within each tumor and predict neo-adjuvant therapy response. BRIDGE involves three key steps. (I) Subtype Deconvolution : We constructed a reference matrix of ‘pure’ molecular subtype profiles from an integrated single-cell RNA-seq compendium and applied support vector regression to infer subtype abundances per tumor. (II) Response prediction from bulk expression: We trained logistic regression models to predict pCR taking as input subtype compositions using 33 datasets from 17 independent neoadjuvant cohorts. These included 11 HR-positive, 10 TNBC chemotherapy, and 12 anti-HER2 datasets. For each treatment, models were trained on the three largest datasets and validated on the remaining 24. (III) Response prediction from histopathology: we integrated BRIDGE with Path2Omics, a deep learning framework that infers gene expression from histopathology slides, enabling transcriptome-free prediction in six datasets with available image data. RESULTS: BRIDGE accurately estimated subtype composition, achieving a mean Spearman correlation of 0.80 on simulated pseudo-bulk data. For HR-positive chemotherapy response prediction, BRIDGE achieved a mean ROC-AUC of 0.78 and an odds ratio (OR) of 7.8 - outperforming existing biomarkers. For anti-HER2 therapy, BRIDGE reached a mean ROC-AUC of 0.77 and OR of 7.75, substantially outperforming the use of HER2-enriched subtype classification alone (OR = 3.1). For TNBC chemotherapy-treated cohorts, BRIDGE achieved a mean ROC-AUC of 0.71 and an OR of 4.7, addressing a major clinical gap due to the lack of established predictors. When applied to AI-inferred transcriptomes from H 0.85 for chemotherapy and 0.70 for anti-HER2), with response scores highly concordant with those derived from true transcriptomic data (correlation 0.7). CONCLUSION: BRIDGE is the first robust and interpretable framework for deconvolving malignant subtypes to predict neoadjuvant therapy response across diverse breast cancer subtypes. By modeling intra-tumoral heterogeneity and leveraging histopathology images, BRIDGE bridges the gap between expression-based profiling and routine pathology. Its low-cost, image-based approach supports personalized, subtype-informed treatment strategies and helps democratize precision oncology in the neoadjuvant setting. Citation Format: T. Cantore, D. Hoang, L. R. Pal, A. Stemmer, S. Dhruba, T. Chang1, S. Sammut, S. Lipkowitz, R. S. Padma1, C. Caldas, N. Ulhas Nair, E. Ruppin. Robust prediction of patients’ response to neoadjuvant therapy across breast cancer subtypes using transcriptomics and histopathology abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-12.
Cantore et al. (Tue,) studied this question.