Abstract 7402: Foundation AI models predict molecular measurements of tumor purity

Abstract

Abstract Background: Molecular assays, such as comprehensive genomic profiling, play a critical role in characterizing patient disease and guiding treatment selection in oncology but require sufficient tumor nucleic acids in tissue samples for reliable results. Automated, accurate and reproducible quantification of tissue area and tumor nuclei from H&E slides to determine suitability for molecular testing could reduce the amount of time spent assessing tissue quality. To address this need, we developed pan-indication foundation AI models to quantify tissue and cell features within H&E whole-slide images (WSI) and compared tumor purity estimates to orthogonal molecular and manual methods. Methods: Pan-indication artifact, tissue, and cell models were each trained using at least 2.8 million pathologist annotations across over 100,000 WSI from 14 stains, 25 organs, and 9 scanners and validated against pathologist ground-truth labels. This suite of foundation models was deployed on H&E WSIs of colon adenocarcinoma (COAD; N=442), melanoma (N=363), prostate cancer (PRAD; N=490), and non-small cell lung cancer (NSCLC; N=996) from TCGA1. Model outputs allowed for the segmentation of tumor regions and enumeration of cancer cells for the purpose of determining tumor purity scores, calculated as the fraction of cancer cells within tumor tissue on a WSI. For comparison, tumor purity values determined based on DNA copy number (ABSOLUTE), gene expression (ESTIMATE), DNA methylation (LUMP) and manual assessment2 were obtained. Concordance between AI model scores and other tumor purity values was assessed using the intraclass correlation coefficient, ICC(2,1). Results: AI model scores were correlated with orthogonal molecular measures of tumor purity in all four indications. AI model purity scores corresponded the most with ABSOLUTE (ICC range: 0.22-0.46), followed by LUMP (ICC range: 0.02-0.50) and ESTIMATE (ICC range: 0.03-0.34). Across indications, the highest agreement observed between AI model scores and molecular methods was in melanoma (ICC range: 0.34-0.50), while the lowest agreement was in PRAD (ICC range: 0.02-0.22). Compared against ABSOLUTE in NSCLC and PRAD, AI model scores showed the highest concordance, while the other two molecular methods and manual assessment all overestimated ABSOLUTE tumor purity. Conclusions: Here, we describe a pan-indication AI-based approach for the quantification of tumor features in H&E-stained WSI. Foundation AI models successfully segment tumor tissue regions and enumerate cancer cells for determining tumor purity, and these predictions correlate with molecular metrics in four cancer indications. This work provides further evidence of how AI could improve the efficiency of molecular testing and enhance precision medicine approaches in oncology.