An ultrasensitive approach for cancer screening and tissue of origin prediction based on targeted methylation sequencing of cell-free DNA.

10553 Background: Cancer-specific therapy requires the identification of the tumor origin. Detecting cancer and identifying the tumor origin in early stage can improve the survival and prognosis of cancer patients. However, early diagnosis of cancer is challenging as the early symptoms of cancer are mild or even absent. In order to improve the diagnostic efficacy and the patient compliance, we developed a two-stage model for cancer diagnosis and tissue of origin (TOO) classification, using DNA methylation profile evaluated through non-invasive blood-based testing as biomarker. Methods: cfDNA samples from 302 healthy and 598 cancer patients of 8 cancer types (colorectal, lung, liver, pancreatic, gastric, esophagus, breast and ovarian) were sequenced by our in-house-designed DNA methylation panel and randomly split into a training dataset (213:426) and a testing dataset (89:172). The testing dataset was reserved to evaluate the performance of models and would not be used during the feature selection and model development process. Methylation levels of cfDNA samples were measured by the methylated fragment ratios (MFRs) in methylated-correlated blocks (MCBs). We first selected pan-cancer markers by comparing the MFR values of MCBs between the healthy and the diseased individuals and developed a binary classifier (hyper-methylation score). Sample correctly predicted by the hyper-methylation score in the cancer group were then used for TOO marker selection and model fitting. TOO biomarkers were obtained by a series of pairwise comparisons of MFR values between any two cancer types. A deconvolution model was built to estimate the composition of ctDNA, which implied the tumor origin. Results: The hyper-methylation score model based on 135 MCB biomarkers achieved an area under the curve (AUC) of 0.89 in the training dataset and 0.85 in the testing dataset. Under a specificity of 94.8%, sensitivity was 66.2% in the training dataset. Using the 282 cases successfully being predicted, 583 TOO markers were selected to build the deconvolution model. In the testing dataset, the specificity was 95.5%, and the overall sensitivity was 66.3%. With increasing stage, sensitivity increased in all cancer types: 34.6%, 57.1%, 62.5% and 84.8% in stage I, II, III and IV respectively. In stage I-III, sensitivity was 52.3%. TOO classes were predicted in the 114 true positives from the hyper-methylation score model, producing a top 1 accuracy (true class matched the most probable class) of 75.4% and a top 2 accuracy (true class matched the first or the second most probable class) of 84.2%. Top 1 accuracy was 71.1% in stage I-III, vs. 84.0% in stage IV. Top 2 accuracy was 82.2% in stage I-III, vs. 90.0% in stage IV. Conclusions: Our cfDNA-based epigenetic method achieved outstanding performance either in pan-cancer detection or in TOO classification, and is a promising tool for early-stage cancer diagnosis.