Identification Of Genes Related To Tumor Recurrence Of Patients And Development Recurrence Predictive Models With Validation For Early Gastric Cancer

Objective:Compared with advanced gastric cancer(AGC),the prognosis of early gastric cancer(EGC)is much better.One of the important outcomes for patients with early gastric cancer is the recurrence.Previous studies have discovered several clinicopathological and genetic factors related to the tumor recurrence but there is still a lack of accurate recurrence predictive models.This study aims to construct predictive models of EGC recurrence by combining traditional statistic methods and machine learning algorithms including clustering analysis and decision tree,and prospectively collect new patients as test data to validate the model.Method:Based on two sets of transcriptome data(data set GSE130823 and GSE55696),monotonically changing differentially expressed genes(mcDEGs),whose expression level changes monotonically from control tissue,to low-grade intraepithelial neoplasia(LGIN),to high-grade intraepithelial neoplasia(HGIN)and to EGC,have been screened out.Then based on another transcriptome data set GSE62254 with full prognosis data,T test and univariable COX regression analysis have been performed to determine target genes that potentially related to tumor recurrence from those mcDEGs.Three predictive models:risk score model based on multivariable COX regression analysis,unsupervised clustering analysis model and decision tree model,have been constructed and trained by patients in stage I or stage II in data set GSE62254.After that,16 patients diagnosed with HGIN or EGC have been prospectively collected.Among them,4 patients have been detected tumor recurrence during the follow-up period.The expression levels of selected mcDEGs of each patient tumor tissue samples have been detected using quantitative real-time polymerase chain reaction(qRT-PCR).Three trained prognostic models have been validated using the new test data.Finally,sensitivity and specificity of each predictive model were calculated.Results:1.25 mcDEGs potentially related to EGC recurrence have been determined based on the transcriptome data set of GSE130823 and GSE55696.2.The risk-score predictive model based on multivariable COX regression analysis contains 13 mcDEGs after further selection.This model has a great predictive performance in training data set but fails to show good predictive performance in the test data set.3.Predicative models based on machining learning algorithm show excellent predictive performance both in training data set and in test data set.Predictive model based on clustering analysis contains 10 mcDEGs after further selection.It has a sensitivity of 100%and a specificity of 58.3%in the test data set.Predictive model based on the decision tree algorithm contains 8 mcDEGs after further selection.This model has a sensitivity of 100%,a specificity of 58.3%,and an area under the curve(AUC)value of 0.792 in the test data set.Conclusions:1.Several genes whose expression levels change monotonously during the different development stages of gastric cancer can be used to predict the risk of tumor recurrence.2.The predictive models based on machine learning algorithms are able to discover the complex relationship between gene expressions and tumor recurrence.Those predictive models show excellent predictive performance in both training data set and test data set and could help to guide doctors to individually make follow-up plans of each patients with EGC.