LAUSR

e22518 Background: Early detection of cancer has the potential to provide a clinical benefit to patients, especially for those with hard-to-treat diseases like ovarian and pancreatic cancer. The THUNDER (the unintrusive detection of early-stage cancers) program was designed to develop and validate a multicancer detection blood test (ELSA-seq) through a series of prospective, observational case-control studies. The second sub-study of THUNDER program (THUNDER-II) includes a discovery, training/validation, and single-blind test (SBT) phase. This new analysis, comprising data from the SBT set of THUNDER-II, evaluated performance of ELSA-seq in an independent group of participants, including individuals at high risk of certain cancer types for the first time. Methods: In the SBT group reported here, all participants were recruited from sites that were not represented in training or validation. The cancer group focused on six deadly cancer types (lung, colorectal, liver, pancreatic, esophageal, and ovarian cancer), and the non-cancer controls were stratified to ‘healthy’ and ‘high-risk’ subgroups based on examination results of low-dose spiral chest CT scan (LDCT), abdominal ultrasound, and Hepatitis B serologic test (HBsAg). 10ml of blood from each participant was collected and methylation-sequencing libraries were prepared by ELSA-seq as described previously. The detection results were automatically reported by a locked computational pipeline to mimic a real-world test scenario (Burning Rock Biotech, China). All samples were delinked from the donors’ cancer status and experiments were performed by blinded personnel. Results: In general, the cancer and non-cancer groups were comparable with respect to age, sex, and tobacco/alcohol-consumption history. The assessment is still in progress, and we have studied 360 eligible participants including 202 cancer cases and 158 non-cancer controls (76 ‘healthy’; 82 ‘high-risk’). Overall, ELSA-seq achieved a specificity of 98.1% (95%CI: 94.1-99.5%) and a sensitivity of 74.8% (95%CI: 68.1-80.5%), consistent with the performance in training and validation1. Among six pre-specified cancer types, sensitivity was 53.0% (95%CI: 40.4-76.3%) in stage I, 73.3% (95%CI: 57.8-84.9%) in stage II, 90.4% (95%CI: 78.2-96.4%) in stage III, and 92.3% (95%CI: 78.0-98.0%) in stage IV. In terms of identifying the diseased organ(s), 80.8% (95%CI: 73.4-86.6%) of the predications matched the true TOO status. Conclusions: ELSA-seq demonstrated consistent results in sites that did not contribute to training, supporting that performance was not subject to sample-/site-dependent bias. It also showed a very low false positive rate in asymptomatic individuals with or without certain cancer-prone conditions, suggesting a potential applicability to broad populations.