Gene signature developed for predicting early relapse and survival in early-stage pancreatic cancer

Abstract

Abstract Background The aim of this study was to construct a predictive signature integrating tumour-mutation- and copy-number-variation-associated features using machine learning to precisely predict early relapse and survival in patients with resected stage I–II pancreatic ductal adenocarcinoma. Methods Patients with microscopically confirmed stage I–II pancreatic ductal adenocarcinoma undergoing R0 resection at the Chinese PLA General Hospital between March 2015 and December 2016 were enrolled. Whole exosome sequencing was performed, and genes with different mutation or copy number variation statuses between patients with and without relapse within 1 year were identified using bioinformatics analysis. A support vector machine was used to evaluate the importance of the differential gene features and to develop a signature. Signature validation was performed in an independent cohort. The associations of the support vector machine signature and single gene features with disease-free survival and overall survival were assessed. Biological functions of integrated genes were further analysed. Results Overall, 30 and 40 patients were included in the training and validation cohorts, respectively. Some 11 genes with differential patterns were first identified; using a support vector machine, four features (mutations of DNAH9, TP53, and TUBGCP6, and copy number variation of TMEM132E) were further selected and integrated to construct a predictive signature (the support vector machine classifier). In the training cohort, the 1-year disease-free survival rates were 88 per cent (95 per cent c.i. 73 to 100) and 7 per cent (95 per cent c.i. 1 to 47) in the low-support vector machine subgroup and the high-support vector machine subgroup respectively (P < 0.001). Multivariable analyses showed that high support vector machine was significantly and independently associated with both worse overall survival (HR 29.20 (95 per cent c.i. 4.48 to 190.21); P < 0.001) and disease-free survival (HR 72.04 (95 per cent c.i. 6.74 to 769.96); P < 0.001). The area under the curve of the support vector machine signature for 1-year disease-free survival (0.900) was significantly larger than the area under the curve values of the mutations of DNAH9 (0.733; P = 0.039), TP53 (0.767; P = 0.024), and TUBGCP6 (0.733; P = 0.023), the copy number variation of TMEM132E (0.700; P = 0.014), TNM stage (0.567; P = 0.002), and differentiation grade (0.633; P = 0.005), suggesting higher predictive accuracy for prognosis. The value of the signature was further validated in the validation cohort. The four genes included in the support vector machine signature (DNAH9, TUBGCP6, and TMEM132E were novel in pancreatic ductal adenocarcinoma) were significantly associated with the tumour immune microenvironment, G protein-coupled receptor binding and signalling, cell–cell adhesion, etc. Conclusion The newly constructed support vector machine signature precisely and powerfully predicted relapse and survival in patients with stage I–II pancreatic ductal adenocarcinoma after R0 resection.