Validation of a Pancreatic Cancer Detection Test in New-Onset Diabetes Using Cell-Free DNA 5-Hydroxymethylation Signatures

Abstract

BACKGROUNDPancreatic cancer (PaC) has poor (10%) 5-year overall survival, largely due to predominant late-stage diagnosis. Patients with new-onset diabetes (NOD) are at a six-to eightfold increased risk for PaC. We developed a pancreatic cancer detection test for the use in a clinical setting that employs a logistic regression model based on 5-hydroxymethylcytosine (5hmC) profiling of cell-free DNA (cfDNA). METHODScfDNA was isolated from plasma from 89 subjects with PaC and 596 case-control non-cancer subjects, and 5hmC libraries were generated and sequenced. These data coupled with machine-learning, were used to generate a predictive model for PaC detection, which was independently validated on 79 subjects with PaC, 163 non-cancer subjects, and 506 patients with non-PaC cancers. RESULTSThe area under the receiver operating characteristic curve for PaC classification was 0.93 across the training data. Training sensitivity was 58.4% (95% confidence interval [CI]: 47.5- 68.6) after setting a classification probability threshold that resulted in 98% (95% CI: 96.5-99) specificity. The independent validation dataset sensitivity and specificity were 51.9% (95% CI: 40.4-63.3) and 100.0% (95% CI: 97.8-100.0), respectively. Early-stage (stage 1 and 2) PaC detection was 47.6% (95% CI: 23%-58%) and 39.4% (95% CI: 32%-64%) in the training and independent validation datasets, respectively. Sensitivity and specificity in NOD patients were 55.2% [95% CI: 35.7-73.6] and 98.4% [95% CI: 91.3-100.0], respectively. The PaC signal was identified in intraductal papillary mucinous neoplasm (64%), pancreatitis (56%), and non-PaC cancers (17%). CONCLUSIONSThe pancreatic cancer detection assay showed robust performance in the tested cohorts and carries the promise of becoming an essential clinical tool to enable early detection in high-risk NOD patients.