PIRPath: A region-guided weakly supervised pathomics framework for neoadjuvant immunotherapy response prediction across multiple cancer types
Background Neoadjuvant immunotherapy is now standard for multiple solid tumors; however, substantial heterogeneity in treatment response remains. Current biomarkers such as programmed death-ligand 1 (PD-L1) capture only limited molecular dimensions of tumor biology. We aimed to predict pathological response from pretreatment hematoxylin and eosin (H&E) whole-slide images (WSIs), benchmark it against PD-L1, and test cross-cancer transferability. Methods Across 320 patients, 377 pretreatment H&E WSIs were processed, yielding 20,247,535 tissue patches. We developed PIRPath (Pathomics for Immunotherapy Response Prediction), a region-guided weakly supervised framework, internally validated in a breast cancer cohort using five-fold cross-validation and further evaluated in two independent external cohorts. Exploratory cross-cancer transfer was assessed in lung cancer and esophageal squamous cell carcinoma cohorts. Results In breast cancer, PIRPath-BC achieved an AUC of 0.9318 (95% CI, 0.8836–0.9713) in internal validation and external AUCs of 0.7991 (95% CI, 0.6000–0.9509) and 0.7292 (95% CI, 0.5333–0.8947). High-score patients showed significantly higher pathological complete response (pCR) rates than low-score patients in both cohorts (78.6%vs 31.2%, P = 0.010; 70.6% vs 35.3%, P = 0.039). In cross-cancer transfer, PIRPath reached 0.7107 for major pathological response (MPR) and 0.7037 for pCR, versus 0.6125 and 0.4590 for PD-L1. Incremental analyses demonstrated additional predictive value beyond PD-L1. Exploratory transcriptomic analyses linked predictions to reduced stromal features and immune activation-related states. Conclusions H&E WSIs contain response-associated histological information not fully captured by conventional biomarkers. PIRPath provides a computational pathology framework for extracting this and complementing existing biomarkers in neoadjuvant immunotherapy stratification, although the modest external cohorts require prospective validation.
Authors
- Peng Yuan (ORCID: https://orcid.org/0000-0002-5968-0368)
- Jian Yue
- Anjie Zhu
- Jinyang Liu
- Zhao Bi
- Jialiang Yang
- Shufang Shi
- Anqing Liu
- Wenyu Yang
Institutions
- Changsha Medical University (CN)
- Chinese Academy of Medical Sciences & Peking Union Medical College (CN)
- National Cancer Center (US)
- Shandong Tumor Hospital (CN)
- Cancer Hospital of Chinese Academy of Medical Sciences (CN)
- Ministry of Education (ET)
- Shandong First Medical University (CN)
Publication Details
- Journal
- Translational Oncology
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1016/j.tranon.2026.103127
- Primary Topic
- AI in cancer detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00