Technical performance of cancer detection and TOO identification of PanSeer7, a targeted bisulfite sequencing assay for non-invasive multi-cancer detection.

Abstract

e16338 Background: Many cancers are symptoms free in early clinical stages, resulting in nearly half of cancer patients diagnosed in advanced-stages when therapeutic options are limited. Early cancer detection is key to improve clinical outcomes. We developed PanSeer7, a multi-cancer detection assay based on targeted bisulfite sequencing of circulating cell-free DNA (cfDNA) and evaluated its technical performances of reproducibility and sensitivity. Methods: The panel of PanSeer7 consists of 2447 markers which were either differentially methylated between healthy and cancer samples, or distinctively methylated in a specific cancer. We assessed PanSeer7’s reproducibility by using it to sequence technical replicates of cfDNA samples. For its limit of detection (LOD), we prepared samples mimicking cancer plasma DNA by diluting fragmented cancer cell line DNA, which represent 7 common cancer types, into GM12878 control at ratios of 1/10,000 to 1/100. We further tested PanSeer7’s ability to identify tissue of origin (TOO) by analyzing DNA samples from formalin-fixed paraffin-embedded (FFPE) tissues and healthy plasma. Results: We analyzed PanSeer7’s reproducibility by sequencing 40 replicates of synthetic healthy cfDNA samples on four independent batches and with different inputs (from 2ng to 20ng). Results show that at a minimum of 10ng input, PanSeer7 produced highly consistent methylation levels among replicates. As to cancer signal detection, we found that PanSeer7’s technical LOD was 1/10,000 for lung cancer cell line H1650, liver cancer line HepG2, gastric cancer line HGC27, esophageal cancer line KYSE150 and colorectal cancer line SW480; it was slightly lower as 5/10,000 for pancreatic cancer line PANC1 and breast cancer line MDA-MB-231. To evaluate PanSeer7’s accuracy of TOO identification, we sequenced 38 healthy plasma and 121 FFPE tissues (17 of liver cancer, 13 of pancreatic cancer, 21 of gastric cancer, 18 of esophageal cancer, 16 of colorectal cancer, 14 of lung cancer, and 22 of breast cancer). Clustering analysis showed that they were segregated according to their TOO based on methylation levels. We also generated 3500 sets of simulated data by mixing the reads of cancer tissue into those of healthy plasma at ratios of 1/10,000 to 1/100, and trained TOO-predicting models with train data sets. At a ratio of 5/10,000, the model predicted TOO of test data sets with an accuracy of over 95%. Conclusions: PanSeer7 required as low as 10ng input DNA for high reproducibility. Its technical LOD in detecting cancer signal was no lower than 5/10,000 for all 7 cancer cell lines tested, and has an in silico TOO detection LOD of 5/10,000. Thus, PanSeer7 had excellent performances in both cancer signal detection and TOO identification, showing promise to be clinically applied for non-invasive multi-cancer detection after future optimization and validation.