关键词: 胰腺癌, 外周血单个核细胞, DNA甲基化, 早期诊断, 分子标志物

Abstract: Objective To provide a new approach for the early diagnosis of pancreatic cancer by establishing and validating a combined diagnostic model for pancreatic cancer based on the methylation of the ataxia telangiectasia mutated(ATM)gene and clinical test indicators using the random forest algorithm. Methods Retrospectively, 118 specimens of peripheral blood mononuclear cells(PBMCs)from pancreatic cancer patients(pancreatic cancer group)and 73 specimens from healthy controls(healthy control group)were collected and their clinical test results were recorded. They were divided into discovery set, training set and validation set. In the discovery set, 935K methylation chip was used to analyze the differentially methylated points(DMPs)of PBMCs samples from both groups. In the training set and validation set samples, pyrosequencing was used to validate the candidate DMPs. In the training set, random forest algorithm was used for variable selection and to construct a combined diagnostic model. The diagnostic efficacy of the model in pancreatic cancer at different stages and in patients with negative carbohydrate antigen 19-9(CA19-9)was evaluated in the validation set using the receiver operating characteristic(ROC)curve. Results The results of the DNA methylation beadchip indicated that, with the screening condition of |Δβ|≥0.1 and P<0.01, 132 differentially methylated points(DMPs)were obtained. Among them, the methylation level of the ATM gene DMPs in PBMCs of the pancreatic cancer group was significantly higher than that of the healthy control group, and it could distinguish the pancreatic cancer group from the healthy control group(AUC=0.871, P<0.001). The pyrosequencing results further indicated that the ATM gene DMPs in PBMCs of the pancreatic cancer group was in a high-methylation level. Based on the variable selection of the random forest algorithm, three variables, ATM gene DMP, CA19-9, and albumin(ALB), were obtained. In the training set and validation set, the expression level of CA19-9 was significantly increased in the pancreatic cancer group(P<0.001), and the expression level of ALB was significantly decreased(P<0.001). Using the expression levels of the above three variables as features, the combined diagnostic model("AmCA")constructed by random forest in the training set had an AUC of 0.992(95%CI: 0.952-1.000)for pancreatic cancer diagnosis. The AUC of the model in the validation set was 0.982(95%CI: 0.895-1.000), which was superior to the AUC value of CA19-9 alone(0.840, 95%CI: 0.705-0.930); the model had good diagnostic efficacy for early-stage pancreatic cancer, with AUC values of 1.000(95%CI: 0.863-1.000)and 0.979(95%CI: 0.840-1.000)for stage I and stage II pancreatic cancer, respectively. The AUC of the model for diagnosing CA19-9-negative pancreatic cancer patients was 0.751(95%CI: 0.639-0.843), with a sensitivity of 52.2% and a specificity of 98.1%. Conclusion The ATM gene is hypermethylated in PBMCs of pancreatic cancer patients. Characterized by ATM hypermethylation,high CA19-9 and low ALB levels,the combined diagnostic model built with the random forest algorithm holds great clinical value for early pancreatic cancer diagnosis,and can make up for the diagnostic deficiency of the conventional biomarker CA19-9.

Key words: Pancreatic cancer, Peripheral blood mononuclear cells, DNA methylation, Early diagnosis, Molecular marker