Abstract 2308: Classifier performance of a cfDNA-based multi-cancer detection test on uncommon cancer types

Abstract

Abstract Introduction: We previously reported multi-cancer detection capabilities across 20 cancer types in the Circulating Cell-free Genome Atlas study (CCGA; NCT02889978). Applied to a screening population, a multi-cancer early detection test will confront the test with malignancies not represented during test development (ie, untrained cancer types [UCT]). The ability of a test to classify UCT and uncommon clinical scenarios (eg, multiple primaries [MP] or metastatic cancers with uncertain primary [UP]), can indicate whether a multi-cancer early detection test is generalizable for detection of cancer types not included in the training set. Methods: CCGA plasma cfDNA samples were subjected to a targeted methylation (TM) assay; methylation states per DNA fragment across targeted genomic regions were input to create a classifier that detects cancer signal and predicts tissue of origin (TOO) for 20 prespecified prediction classes (breast, ovarian, uterine, cervical, anal, prostate, bladder & urothelial, renal, colorectal, liver & bile duct, pancreas & gallbladder, upper GI, lung, head & neck [H&N], thyroid, sarcoma, melanoma, plasma cell-, myeloid-, lymphoid neoplasms). Test results were independently obtained for 44 participants (2 tubes each) with UCT (cancer other than the 20 prespecified cancer classes), for 16 participants with MPs, and 26 with UP; results were compared to previously published detection results for the 20 prespecified classes. Results: UCT were detected for 2 / 1 / 0 tubes in 19 / 2 / 23 cases, respectively. Of 15 vulva, vagina, and penis cancers, 10 were detected in at least 1 tube with TOO prediction of HPV-driven cancer types. Of 7 mesothelioma of pleura or peritoneum, 4 were detected in both tubes (4 / 0 / 3) with TOO prediction of lung, indeterminate localization, and H&N. Of 22 other cancer types, 7 were detected with TOO prediction that was consistent with biologic features of trained classes. MP were detected in 12 out of 16 cases (11 / 1 / 4). TOO prediction was consistent with the cancer class with the stronger signal in 11 out of 12 cases. UP was detected in 23 out of 26 cases (23 / 0 / 3) and TOO prediction was narrowed down to a median of 4 cancer types (out of the 20 prespecified classes). Sensitivity for detection in cfDNA at >99% specificity by stage (I / II / III / IV) was 36% / 39% / 66% / 64% for UCT, 50% / 67% / 90% / 100% for MPs (using higher stage of 2 primaries), and 88% for UPs. For comparison, sensitivity by stage for cases with the 20 prespecified classes was 18% / 44% / 78% / 90%. Conclusion: This TM multi-cancer early detection test detected cancers beyond the trained types from plasma samples with a sensitivity similar to those for trained cancer types. This suggests that the test classifier exploits general features indicative of malignancy and handles superposition of signals from MPs, supporting its generalizability to population-scale multi-cancer early detection and potential application to UP. Citation Format: Joerg Bredno, Samuel Gross, Alexander P. Fields, Kathryn N. Kurtzman, Rita Shaknovich, Jessica Yecies, Xiaoji Chen, Jan Schellenberger, Eric Scott, Zhao Dong, Eric T. Fung, Anne-Renee Hartman, Earl Hubbell, Arash Jamshidi, Alexander M. Aravanis, Oliver Venn. Classifier performance of a cfDNA-based multi-cancer detection test on uncommon cancer types [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2308.