Abstract 2432: Automated detection of genome-doubled tumors in H&E whole-slide images: towards a novel visual biomarker for predicting overall survival

Abstract Background: Several studies have outlined whole genome-doubled (WGD) tumor tissue characterized by elevated genome ploidy as an important biomarker positively associated with immunotherapy response and overall survival of cutaneous melanoma patients (Liu NatMed2019, Tarantino BioRXiv2022), as well as conferring therapeutic vulnerabilities (Quinton Nat2021). A robust and cost-effective method for detecting this biomarker would significantly improve patient risk assessment and treatment optimization. Unfortunately, current approaches depend on whole-exome sequencing of both tumor and normal tissue - a costly process rarely performed in standard clinical practice. Methods: We focus on H&E-stained whole-slide images (WSIs) routinely collected to diagnose cancer patients. We use publicly available melanoma WSIs (TCGA SKCM) and ploidy calls (Conway NatGen2020) to train a machine learning pipeline for classifying WSIs as diploid or WGD. After splitting the WSI into tiles we apply a wide array of recently released deep-learning foundation models that have been pre-trained on vast amounts of WSIs and were shown to be effective at extracting biologically meaningful features. To demonstrate the utility of these features for predicting WGD, we use tile-level features to train several types of slide-level classifiers. To evaluate the prognostication potential of H&E-inferred WGD, we use it to fit a univariate Cox survival model using publicly available survival data (PanCanAtlas). We perform a cluster analysis on tile-level features and model-assigned importance scores leveraging the help of a board-certified pathologist to identify visually interpretable morphological features significantly correlated with WGD. Results: We performed both an internal cross-validation using our training cohort and an external validation cohort collected at Dana Farber Cancer Institute. We evaluate 100s of deep learning pipelines, each characterized by a unique combination of foundation model, slide classifier, and various other hyperparameters. Best-performing pipelines predict WGD with AUC >0.7 on both internal and external validation cohorts, demonstrating a robust association between tissue morphology and ploidy. Furthermore, in univariate Cox survival analyses, best-performing pipelines achieve a hazard ratio <0.7 with a p-value <0.05, underlining the ability of WGD to serve as a visually identifiable biomarker for predicting melanoma outcomes. Conclusion: Our study demonstrates the potential of H&E-based deep learning pipelines to robustly predict WGD in melanoma tissue with reliable performance across internal and external cohorts. Furthermore, WGD inferred from tissue morphology shows significant prognostic value for overall survival, offering a cost-effective and clinically feasible biomarker for personalized treatment strategies. Citation Format: Bojan Karlaš, Marc Glettig, Giuseppe Tarantino, Jiajia Chen, Guihong Wan, Tyler J. Aprati, Haitham A. Elmarakeby, Christine G. Lian, Yevgeniy R. Semenov, David Liu. Automated detection of genome-doubled tumors in H&E whole-slide images: towards a novel visual biomarker for predicting overall survival [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 2432.