Abstract PD11-03: Predicting treatment outcomes in breast cancer from H&E slides using pathology foundation models with multiple instance learning

Abstract Background. Pathologic complete response (pCR) is the absence of residual invasive cancer in the breast and axillary lymph nodes after neoadjuvant therapy. In breast cancer treatment, pCR is a proven surrogate for long-term outcomes. However, accurately predicting pCR at diagnosis remains a clinical challenge. Current tools primarily rely on clinical, genomic, or transcriptomic data. Advances in computational pathology and deep learning enable the extraction of meaningful features from H&E-stained whole slide images (WSIs) to identify phenotypic biomarkers of response. Methods. We apply attention-based multiple instance learning (MIL) to predict pCR from pre-treatment H&E-stained frozen tumor biopsy WSIs in the I-SPY2 trial. A total of 3,306 WSIs from 911 patients across 13 treatment arms were tiled and filtered. Vectors of 1024 features were extracted per tile using the UNI pathology foundation model. The MIL model was independently trained and evaluated with 3-fold cross-validation on these vectors for each arm and tumor subtype. We compared model performance by AUROC between MIL and two elastic net regression models: one trained on pathologist-assessed features (tumor grade, DCIS, invasive histology, and lymphovascular invasion), and another adding clinical features: pre-treatment MRI functional tumor volume (FTV) and transcriptome-derived response predictive subtypes (RPS). Results. 298 of 911 patients achieved pCR (142 HR+/HER2+, 347 HR+/HER2-, 85 HR-/HER2+, and 337 HR-/HER2-). MIL performance varied by arm (AUROC 0.501-0.893) with 6 arms achieving statistically significant performance (95% CI > 0.5). Highest model performance was in HER2+ cohorts: (1) Paclitaxel + Trastuzumab and (2) Paclitaxel + Pertuzumab + Trastuzumab (AUROC = 0.893, 0.785) (Table 1). Of the 6 arms, MIL outperformed the elastic net trained on pathologist-assessed histology features in 5 arms. After including FTV and RPS in the elastic net, MIL still outperformed in 3 arms. Across subtypes, the model predicted better in HR+ subgroups (HR+/HER2- AUROC = 0.706, HR+/HER2+ AUROC = 0.677) than in HR- subgroups (HR-/HER2+ AUROC = 0.533, HR-/HER2- AUROC = 0.548). Conclusion. These findings demonstrate the feasibility of applying MIL vision models to predict treatment-specific response in breast cancer, even with frozen section WSIs and limited data. MIL detects important histology patterns not captured by conventional pathology. Even with added MRI and transcriptomic data, the model provides complementary predictive value. This approach enables early, accurate predictions from routine histology and supports personalized, less toxic treatment—particularly in under-resourced settings. Citation Format: A. Sun, S. Venters, C. Yau, D. Wolf, G. Hirst, M. Campbell, A. Asare, W. Symmans, L. Brown-Swigart, N. Hylton, J. Perlmutter, A. DeMichele, D. Yee, H. Rugo, A. Borowsky, F. Howard, L. Esserman, L. van't Veer, A. Basu. Predicting treatment outcomes in breast cancer from H&E slides using pathology foundation models with multiple instance learning [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 PD11-03.