Abstract 1316: A methylation-based molecular tumor typing classifier for tissue samples from cancers of unknown primary.

Abstract Introduction: Determining tissue of origin is critical for guiding treatment decisions in cancers of unknown primary (CUP), metastatic tumors with ambiguous diagnostic findings, and cases where molecular confirmation may inform therapy selection. We developed a methylation-based molecular tumor type (MTT) prediction algorithm for malignant tissue samples, designed to operate robustly across a wide range of tumor fractions, to aid in determining cell of origin as an adjunct to conventional histopathology and immunohistochemistry. Methods: The classifier was trained on 24 of the most prevalent solid cancers using the methylation profiles of more than 2,000 differentially methylated regions derived from Guardant360 Tissue (Guardant Health, Palo Alto, CA) from 6,047 FFPE tissue samples. Each MTT prediction was assigned a confidence score between 0 and 1, and the highest-confidence predictions were reported for evaluable samples. Evaluability criteria required a confidence level of the top prediction being greater than 0.5. The MTT classifier reports up to two top predictions based on the confidence scores. Performance was evaluated in 6,056 non-CUP samples and 191 CUP samples (including 35 CUP samples with a suspected diagnosis). Results: The MTT classifier was first evaluated on 6,056 tumor tissue samples against the clinical diagnoses provided on the test requisition forms. 97.85% (5,926/6,056) were evaluable. The overall accuracy was 91.56% (5,426/5,926) for primary predictions and 93.76% (5,556/5,926) for the top 2 predictions. Among evaluable samples, 91.29% (5,410/5,926) have high-confidence primary predictions, with an accuracy of 94.05% for primary prediction in this subset.A notable subset of discordant primary predictions was among tumor types that could include mixed or closely related subtypes, such as nonsmall cell lung carcinoma (NSCLC) subtypes (22.20% (111/500)), or tumor types with related histogenesis, such as pancreaticobiliary tumors (7.60% (38/500)). The overall accuracy increases to 94.08% (5,575/5,926) and 95.46% (5,657/5,926) for top-1 and top-2 predictions, respectively, when NSCLC subtypes are merged into a single category and the pancreatic and biliary categories are merged to form a pancreaticobiliary category.The MTT classifier was also tested on 191 CUP samples; 88.48% (169/191) were evaluable. In 35 CUP samples with suspected diagnosis, evaluability was 97.14% (34/35) with accuracy of 85.29% (29/34) for primary predictions and 91.18% (31/34) for top-2 predictions. Conclusions: This methylation-based MTT classifier demonstrates high evaluability and accuracy in determining tissue of origin for both non-CUP and CUP samples, providing valuable information to guide clinical decision-making when used as a supplement to conventional diagnostic methods. Citation Format: Elmira Forouzmand, Lauren Lawrence, Jack Tung, Sheila Solomon, William W. Young Greenwald, Yupeng He. A methylation-based molecular tumor typing classifier for tissue samples from cancers of unknown primary abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1316.