Early detection and localization of multiple cancers using a blood-based methylation assay (ELSA-seq).

Abstract

459 Background: Early detection of cancer can potentially offer clinical benefits, particularly for those without effective screening methods. The PREDICT study (pan-cancer early detection project, NCT 04383353) is a prospective, multi-center, longitudinal study that aims to identify multiple cancers non-invasively before symptoms become apparent. As a pilot project, the THUNDER (the unintrusive detection of early-stage cancer) study is designed for development and validation of ELSA-seq, a sensitive targeted methylation sequencing assay that interrogates epigenetic alterations from circulating cell-free DNA (cfDNA). Herein we report results from the second THUNDER sub study (THUNDER-II), which focused on malignancies developed in liver, colon/rectum, esophagus, pancreas, lung and ovary, of which four are gastrointestinal (GI) cancer types. Methods: THUNDER-II comprises four independent steps: marker discovery, model training, validation, and single-blind test. The marker discovery work was conducted by profiling 5.5 million CpG sites in 247 tissue samples (116 cancer, 131 normal/benign, 6 cancer types) using a SeqCap Epi CpGiant system (Roche). By combining data generated in-house and from public sources, a custom hybridization capture panel was designed to target 161,984 CpG sites. For cfDNA applications, 625 patients and 483 non-cancer controls were enrolled and divided into a training set (274 cancer and 195 non-cancer) and an independent validation set (351 cancer and 288 non-cancer). Results: The cancer patients and non-cancer controls were generally comparable with respect to age, gender, and smoking status. Various stages were represented in the cancer group, and 79.5% patients were diagnosed at early stages (I-III). At 99.5% training specificity (95%CI: 96.7-100%), the cross-validated sensitivity was 79.9% (95%CI: 74.6-84.4%). The results were consistent in the validation set, with 98.3% specificity (95%CI: 95.8-99.4%) and 80.6% (76.0-84.6%) sensitivity across stages and cancer types. In terms of tracking diseased organ(s), the classifier returned a tissue-of-origin (TOO) result in 98.6% cases, and 81.0% (95%CI: 77.2-84.3%) of these predictions were correct. Conclusions: Results from the THUNDER-II study demonstrated that early cancer signals could be identified by ELSA-seq with high specificity. This method also enabled accurate prediction of TOO, offering guidance for subsequent diagnostic work-up. Together these findings highlight the potential implementation of this sensitive and robust assay as a multi-cancer detection test.