Background: Cancers of the gastrointestinal (GI) system, including esophagus, stomach, pancreas, gallbladder, liver, bile duct, colon, and rectum are estimated to account for 38% of all cancer incidences and nearly 46% of cancer-related deaths in China. We conducted a multi-center study to evaluate the feasibility of using genetic and epigenetic abnormalities in plasma cfDNA to diagnose and locate GI cancers.
Methods: We performed parallel genetic and epigenetic profiling of plasma cfDNA from hepatocellular carcinoma (HCC), colorectal cancer (CRC) and pancreatic cancer (PC) patients as well as age-matched healthy individuals by ultra-deep sequencing targeting cancer driver genes, and by targeted bisulfite sequencing covering genome-wide CpG islands, shelves, and shores.
Results: Using a pre-specified mutation scoring system, we found that cfDNA mutation profiling achieved a sensitivity of 59.6%, 67.2%, and 46.8% for detecting HCC (n = 322), CRC (n = 244) and PC (n = 141) respectively, with a specificity of 95% in healthy controls (n = 207). For 901 plasma cfDNA samples that underwent methylome profiling, we first applied a machine learning approach to classify each cancer type versus healthy controls in the training cohort (HCC: n = 125; CRC: n = 105; PC: n = 97; healthy individuals: n = 84). Random Forest models with 10-fold cross validation achieved an AUC of 0.96±0.04,0.89±0.06, 0.91±0.07 for HCC, CRC, and PC, respectively. Further analyses were performed on the validation cohort, including 172 HCC patients, 162 CRC patients, 60 PC patients, and an independent cohort of healthy individuals (HCC validation: n = 63; HCC independent validation: n = 109; CRC validation: n = 104; CRC external validation: n = 58; PC validation: n = 60; healthy controls: n = 96). The trained model achieved a sensitivity of 83.1% (specificity = 95.8%), 89.5% (specificity = 95.8%), and 76.7% (specificity = 91.7%) for HCC, CRC, and PC, respectively. Using regional methylation markers from diagnostic models for individual cancer types, we built a tissue-of-origin classification model, which achieved a cross-validation accuracy of 83.3% in the training cohort and an accuracy of 80.1% in the validation cohort in assigning correct cancer types.
Conclusions: Plasma cfDNA methylome profiling identified effective biomarkers for the detection and tissue-of-origin determination of GI cancers, and outperformed mutation-based detection approach. Therefore, a liquid biopsy test capable of detecting and locating GI cancers is feasible and may serve as a valuable tool for early detection and intervention.