Detection and localization of gastrointestinal cancers based on multi-dimentional signatures from a single cfDNA targeted sequencing assay.

Abstract

4169 Background: Five major gastrointestinal (GI) cancers - colorectal (CRC), gastric (GC), liver (LC), esophageal (EC), and pancreatic cancer (PC) - are responsible for hundreds of thousands of mortalities annually worldwide. Unfortunately, there is a lack of cost-effective, blood-based screening method for their early detection. To address this issue, we aimed to develop GutSeer, a noninvasive, targeted methylation sequencing-based test by leveraging methylation and fragmentomic signatures carried by cell-free DNA (cfDNA). Methods: The panel of GutSeer consists of 1656 target regions which were either differentially methylated between healthy and cancer samples, or distinctively methylated in a specific GI cancer. Cancer and healthy participants were recruited and randomly divided into a training and a validation cohort. Their plasma DNA samples were analyzed to generate DNA methylation and fragmentomic features. These multi-dimensional features were integrated to build ensemble stacked machine learning models to differentiate cancer against healthy, and to determine the tissue-of-origin (TOO) of the cancer. Results: A total of 1844 cases (787 healthy, 342 LC, 239 GC, 209 EC, 180 CRC, and 87 PC cases) were recruited for this study. A cancer- vs-healthy model achieved an AUC of 0.94 and 0.95 (sensitivity of 77.7% and 77.1% under the specificity around 96%) using either methylation or fragmentomic features only, respectively. Combining both methylation and fragmentomic features further improved performances, achieving an AUC of 0.96 (sensitivity = 86.2% at a specificity of 96.7%). For individual type of cancer, GutSeer has a sensitivity of 93.3% for CRC, 81.1% for EC, 70.3% for GC, 96.5% for LC, and 86.4% for PC. An independent test using 629 benign cases as controls achieved a specificity of 87.1%. A separate TOO model was built using all features and achieved an overall accuracy of 82% for all cancer cases (66.7% for CRC, 87.0% for GC and EC combined, 89.0% for LC, and 63.2% for PC). Same as the cancer detection model, using multi-dimensional features in TOO prediction yielded higher accuracy than when models using only methylation or fragmentomics features (accuracy = 75.6% or 75.4%, respectively). When compared with whole-genome sequencing (WGS) based approaches, GutSeer showed a comparable performance in cancer detection but a higher accuracy in TOO identification, further confirming its effectiveness for detection of GI cancers. Conclusions: GutSeer, a non-invasive test integrating multi-dimensional features, was demonstrated to detect and localize the 5 main types of GI cancer with high accuracy. Our results further showed that a reasonably sized panel can perform comparably or even better than WGS-based methods in cancer detection and TOO localization, indicating GutSeer may be a low-cost solution for blood-based early screening for GI cancers.