Abstract 6186: Histopathology-derived molecular portrait of serous ovarian cancer

Abstract

Abstract Background: The progression of ovarian cancer (OV) is often evaluated through clinical staging and molecular profiling of genomic alterations and subtypes, with the molecular profiling of tumors playing a crucial role in cancer management by informing therapeutic decisions. However, collecting molecular data is expensive, time-consuming, and can be sensitive to sample handling protocols. These factors prevent routine implementation at many care centers, thereby limiting benefit to patients. This work examines the viability of an image-centric method for inferring molecular endophenotypes and predicting overall survival in OV, leveraging commonly gathered histopathology scans. Methods: Embedding representations of digitized Hematoxylin and Eosin slides were produced using a self-supervised histology model. First, we assessed the ability of the embeddings to predict overall survival using Cox proportional hazards models by calculating the concordance index. Next, these embeddings served as input features to machine learning models trained to predict genomic patterns within tumors, including molecular subtypes, driver gene alterations, and mutational processes, leading to a more comprehensive analysis of the prognostic potential of our approach. Results: Histopathological embeddings were found to be more predictive of patient survival than traditional factors such as age, histological grade, and clinical stage. Embedding risk combined with clinicopathological risk had a higher predictive value than clinicopathological risk alone (0.578 vs 0.569, p = 0.025). Moreover, embedding risk alone was more predictive of survival (0.577 vs 0.569, p = 0.038). Risk scores derived from embeddings and all clinicopathological features were uncorrelated (R = 0.03), indicating that embeddings provide unique prognostic information. Motivated by these findings, we investigated the reasons behind the enhanced accuracy in survival prediction. We found that models trained on embeddings predicted previously defined molecular subtypes (TCGA, 2011), achieving area under the receiver operating characteristic curve (AUROC) scores as follows: Differentiated (0.86), Immunoreactive (0.85), Mesenchymal (0.9), and Proliferative (0.89). Similarly, our models identified crucial oncogenic events, including the loss of TP53 and CDKN2A, and instances of whole genome duplication (AUROCs of 0.85, 0.99 and 0.8). Conclusions: Taken together, this work showcases the potential utility of standard of care digitized histopathology scans for imputing genomic patterns and predicting overall survival in OV. Our results suggest that imputation could provide a fast and scalable solution for deploying patient screening biomarkers, thereby playing a pivotal role in shaping the future of precision medicine. Citation Format: Yajas Shah, Zachary R. McCaw, Anna Shcherbina, Christopher Probert. Histopathology-derived molecular portrait of serous ovarian cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6186.