A high-performance blood-based DNA methylation test for early detection of gastrointestinal cancers.

Abstract

3050 Background: Early detection of gastrointestinal cancers (GICs), including esophageal cancer (EC), gastric cancer (GC), and colorectal cancer (CRC), remains suboptimal in China due to low screening adherence and limited access to endoscopic procedures. Multi-cancer early detection (MCED) tests present a convenient alternative, yet the efficacy in detecting early-stage GICs has been inadequate. Accurate tumor localization, particularly differentiating between upper and lower GICs, is crucial for determining subsequent diagnostic procedures. In this study, we report the performance of an MCED assay utilizing targeted DNA methylation sequencing to detect GICs. Methods: This multicenter, case-control study prospectively enrolled a cohort of 667 GIC patients (203 EC, 263 GC, 201 CRC; stages: I 20.8%, II 26.7%, III 35.7%, IV 16.8%) and 667 non-cancer participants. Plasma cell-free DNA was sequenced using a panel targeting tumor-specific hyper- and hypo-methylation markers. A total of 5120 GIC-specific methylation features were captured. The GIC model was trained using a gradient-boosted tree model, and nested cross-validation was implemented to determine the optimal parameters and evaluate the model’s performance of cancer detection. To predict the tissue of origin (TOO), the top 256 features were first selected based on pairwise mutual information for each cancer type. An XGBoost classifier combined with Synthetic Minority Over-Sampling Technique (SMOTE) was trained to determine the TOO. Results: The GIC model exhibited robust performance, achieving an area under the curve (AUC) of 0.959 (95% CI: 0.949-0.970), with an overall sensitivity of 86.4% (83.5%-88.8%) at a specificity of 96.0% (94.2%-97.2%). Notably, for stage I-III GICs, which accounted for 83.2% of cases (a proportion consistent with that seen in prospective observational cohort studies of MCED, such as the SYMPLIFY study), the sensitivity reached 84.1% (80.9%-87.0%). The sensitivities for EC, GC, and CRC were 87.2%, 82.9%, and 90.0% respectively. For stage I CRC, the sensitivity of the GIC model reached 79.1% (64.8%-88.6%), comparable to that of multitarget stool DNA tests, and outperformed fecal immunochemical tests (FIT). Regarding tumor localization, the accuracy of TOO across all positive cases was 89.8% (87.0%-92.0%). For stage I-III GICs, the model maintained a high accuracy of 88.7% (85.5%-91.2%) in predicting TOO. Moreover, the model demonstrated exceptional accuracy in distinguishing between upper and lower GICs, with an accuracy of 95.5% (93.5%-96.9%). Conclusions: This study demonstrated the high performance of the MCED test for early detection of GICs in a large-scale, prospective cohort enriched with early-stage cancers. These findings highlight the potential of the GIC model to enhance early detection and precise localization of GICs, thus improving the efficiency of subsequent diagnostic procedures.