Early detection and localization of major gastrointestinal cancers by combining DNA methylation and fragmentomic signatures generated from a single cfDNA targeted sequencing assay.
Xinrong Yang,Dongli He,Zhiguo Xiong,Bin Yan,Quanlin Li,Dezhen Guo,Ao Huang,Zhen Feng,Pinxiang Lu,Qi Guo,Mengjiang He,Chengcheng Ma,Minjie Xu,Yiying Liu,Mingyang Su,Qiye He,Rui Liu,Yunshi Zhong,Jia Fan,Jian Zhou
DOI: https://doi.org/10.1200/JCO.2023.41.4_suppl.330
IF: 45.3
2023-01-26
Journal of Clinical Oncology
Abstract:330 Background: Gastrointestinal (GI) cancers totally account for more than one third of the cancerous deaths, yet there is no cost-effective blood-based assay for the early detection of GI cancers. We sought to develop GutSeer, a noninvasive test based on cell-free DNA (cfDNA) methylation and fragmentation signatures derived from one single targeted DNA methylation sequencing panel, for early detection and localization of five major GI cancers, including colorectal (CC), gastric (GC), liver (LC), esophageal (EC), and pancreatic cancer (PC). Methods: A DNA methylation targeted sequencing panel with 1656 target regions was designed. It was then verified in a large cohort of retrospective cancer and control plasma samples for feature selection and modeling. The participants were randomly divided into a training cohort and a validation cohort in a 1:1 ratio. DNA methylation and fragmentomic features were calculated based on GutSeer sequencing data. An ensemble stacked machine learning approach was built to classify cancer and healthy samples in training cohort and tested in validation cohort. We also constructed a TOO model to predict the tissue of origin of detected cancer samples. Results: To develop GutSeer assay, we have enrolled and tested a total of 1844 retrospective plasma samples (787 healthy, 342 LC, 239 GC, 209 EC, 180 CC, and 87 PC), over half of the cancer samples were diagnosed with early-stage disease (TNM stage I 35.6%; stage II 23.3%; stage III 21.7%; stage IV 12.5%). Cancer- vs-healthy model was built on training cohort and tested in validation cohort, achieving an AUC of 0.94 (sensitivity=77.7%, specificity=96.4%) with methylation features, and 0.95 (sensitivity=77.1%, specificity=95.9%) with fragmentomic features. Combining these features could achieve AUC of 0.963 (sensitivity = 86.2%, specificity = 96.7%). For individual cancer types, the sensitivity was 93.3% (CC), 81.1% (EC), 70.3% (GC), 96.5% (LC) and 86.4% (PC), respectively. For predicted cancer samples, we achieved an 82% top-one (66.7% CC, 87.0% GC/EC, 89.0% LC, 63.2% PC) and 95.2% top-two (86,9% CC, 98.2% GC/EC, 97.6% LC, 89.5% PC) TOO accuracy (ACC, accuracy of predicting the most likely, and the top 2 most likely tissue or organ types where the identified cancer was located, respectively) in validation cohort with TOO model combined all features. Conclusions: Based on a single targeted DNA methylation sequencing assay, GutSeer, which combined cfDNA methylation and fragmentomic signatures, could detect and localize the major five GI cancers with high accuracy but low cost. Although this is a pilot study with limited sample size, GutSeer demonstrated the potential to be further optimized into non-invasive diagnostics for blood-based early screening and diagnosis for GI cancers.
oncology