Abstract 2464: Inferring genomic properties from H&E whole-slide images of >60,000 solid tumors

Abstract Background: Clinical tumor-normal panel sequencing has enabled personalized approaches to targeted therapies. However, the impact of somatic mutations on tissue morphology remains incompletely understood, and financial hurdles limit universal adoption of personalized approaches. Methods: We developed machine learning tools to infer genomic properties from hematoxylin and eosin (H&E) whole-slide images (WSIs). We assembled a pan-cancer cohort of 62, 266 sequenced tumor samples matched to 315, 383 H&E WSIs from 54, 446 patients, the largest to date. Using these data, we developed transformer models to aggregate embeddings of tissue by part to infer genomic features. We then tested performance on 172 OncoKB-annotated predictive or prognostic biomarkers across 127 histologies and analyzed how the models could be used to identify indicated targeted therapies from H&E WSIs alone. Results: We highlight four putative use cases of our models, two each for predictive and prognostic biomarkers. First, BRAF mutational status, inferred with AUROC 0.85 (precision 0.95, recall 0.63) in papillary thyroid cancer has level two evidence for dabrafenib and trametinib in metastatic differentiated thyroid cancer. Second, FGFR3 mutational status, identified with AUROC 0.92 (precision 0.95, recall 0.42) in bladder cancer indicates erdafitinib, surpassing the prior state of the art of AUROC 0.73. Prognostically, the models accurately identify unfavorable STK11 mutations in lung adenocarcinoma (AUROC 0.92) and favorable POLE exonuclease domain mutations in uterine endometrioid carcinoma (UEC) (AUROC 0.90), which could guide shared decision making around adjuvant therapy. Furthermore, disaggregating disease subtypes reveals that phenome-genome relations cited by previous works vary significantly by subtype. For example In UEC, WNT pathway aberrations are detected with AUROC 0.87 (95% C.I. 0.82-0.91) in microsatellite-stable UEC, whereas in MSI UEC, there is no detectable phenome-genome relation (AUROC 0.61, 95% C.I. 0.44-0.76). MSI status itself can be inferred accurately from both colorectal adenocarcinoma (AUROC 0.99, 95% C.I. 0.98-0.99) and UEC (AUROC 0.91, 95% C.I. 0.87-0.95). Conclusions: Taken together, our work uncovers phenome-genome relations for granular subtypes lumped together in prior studies. Such approaches advance our knowledge of how somatic mutations affect tissue morphology and have the potential to improve access to precision oncology. Citation Format: Kevin M. Boehm, Arfath Pasha, Armaan Kohli, Raymond Lim, Tom Pollard, Darin Moore, JianJiong Gao, Nikolaos M. Dimitriou, Madison Darmofal, Meera Hameed, Ahmet Dogan, Natasha Rekhtman, Klaus J. Busam, Lora H. Ellenson, Jie-Fu Chen, Fresia Pareja, Daniel Gomez, Hikmat Al-Ahmadie, Nancy Lee, Eric Sherman, Jason Chang, Himanshu Nagar, Britta Weigelt, James Fagin, Susan Fu, Jonathan Alarcon, Neelraj Patil, Areej Alsaafin, Georgios Asimomitis, Anika Begum, Andrew Aukerman, Justin Jee, Christopher Fong, Michele Waters, Lior Z. Braunstein, Michael F. Berger, Luke R. Pike, Pedram Razavi, Paul Romesser, Walid K. Chatila, Aasiya Islam, Elana Sverdlik, Ino de Brujin, Karl Pichotta, Nikolaus Schultz, Sohrab P. Shah, Francisco Sanchez-Vega. Inferring genomic properties from H&E whole-slide images of >60, 000 solid tumors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2464.