Abstract 2471: Interpretable HPV detection in head and neck cancer using foundation models and synthetic digital pathology

Abstract Introduction: Detecting human papillomavirus (HPV) status is crucial for treating Head and Neck Squamous Cell Carcinomas (HNSCCs). While p16 immunohistochemistry is the current standard for HPV detection, its moderate sensitivity and complex implementation limit its global utility. The ability to diagnose HPV status in HNSCC has become increasingly critical worldwide, as rising HPV-positive HNSCC rates observed in high-income countries may signal a global trend, and HPV status remains essential for treatment selection. Although hematoxylin and eosin (H&E) stained slides are clinically ubiquitous, artificial intelligence (AI) methods applied to these images have not matched molecular assays’ performance nor provided needed clinical interpretability. Recent advances in vision transformer-based foundation models for computational pathology offer a promising approach to address this unmet need. Methods: We analyzed H&E images from 981 HNSCC patients across four datasets: TCGA (n=HPV+ 33/total 401), CPTAC (n=1/109), UCH (n=159/364), and PENN (n=106/106). Fifty percent of the patients were used for validation. Using UNI, a foundation self-supervised learning (SSL) model, we extracted feature vectors from 10x effective magnification tile images across each whole slide. We identified an HPV feature axis using recursive support vector machine and principal component analysis to isolate SSL features that differentiated HPV tumors, then interpreted the relevant histologic features using HistoXGAN to generate synthetic histology images. This approach enabled the isolation of histologic features specific to HPV+ tumors, whereas real histology images contain multiple sources of variation. An expert pathologist validated the biological relevance of the identified HPV-axis features. We developed a predictive model for HPV status by varying the percentages of tiles within each slide required to exceed a binary threshold on the HPV-axis. This model was robust across threshold choices. Results: Our method identified an HPV-axis that robust HPV detection performance (sensitivity 0.83, specificity 0.88) across all datasets (balanced accuracy - TCGA 0.77, CPTAC 0.98, UCH 0.78, PENN 1.00). Using synthetic histology images and pathologist validation, we identified key, morphological features aligning with established HPV-associated histology, including nuclei size, color, and cell borders. Together with our Grundium slide scanning pipeline the time-to-prediction for a slide is less than 3 minutes. Conclusions: Foundation models and synthetic digital pathology enabled HPV detection from histology with accuracy comparable to current diagnostic standards, while providing pathologist-interpretable predictions. This accessible, rapid, and explainable method holds promise for expanding testing of HPV and potentially other molecular features in resource-limited settings. Citation Format: Hanna M. Hieromnimon, Anna Trzcinska, Frank Wen, Frederick M. Howard, James M. Dolezal, Emma Dyer, Sara Kochanny, Jefree Schulte, Cindy Wang, Heather Chen, Jeffrey Chin, Elizabeth Blair, Nishant Agrawal, Ari Rosenberg, Everett Vokes, 1 Rohan Katipally, Aditya Juloori, Evgeny Izumchenko, Mark W. Lingen, Nicole Cipriani, Jalal B. Jalaly, Devraj Basu, Samantha J. Riesenfeld, Alexander T. Pearson. Interpretable HPV detection in head and neck cancer using foundation models and synthetic digital pathology [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 2471.