Prediction Research Of Pancreatic Cancer Markers Based On Support Vector Machine

Pancreatic cancer is a digestive tract malignant tumor.According to reports,it has been reported a very high malignant degree and the five-year overall survival rate was less than 8% between 2006 and 2012.Among them,ductal adenocarcinoma accounts for about 80% to 90%.The incidence of the disease is very high and it is rising.In addition,its overall malignancy is quite high,and there are major difficulties in early screening and treatment.According to the statistics of the World Health Organization,the results of the 2008 World Cancer Survey show that in all cancer rankings,the incidence of pancreatic cancer is ranked 13 th,and the mortality rate is 7th.Since the disease usually does not have obvious symptoms at the beginning,it is difficult to confirm the diagnosis at an early stage.It is difficult to find an accurate serum or urine marker that can be used for early cancer screening through traditional biochemical experiments.In the early 21 st century,pancreatic cancer is still an unsolved major health problem.Considering the transcriptome data has provided a relatively complete background,and has obvious advantages in the description of the biological characteristics for cancer,more and more researchers have been using feature selection method based on machine learning to find biomarkers for cancers.In this research,we proposed a recursive feature elimination(for short RFE)method,which is based on support vector machine(for short SVM)and large margin distribution machine(LDM)to identify potential biomarkers of pancreatic cancer.The dataset we used in the experiment came from the GEO database.In the experiment,we have strengthened the RFE process to achieve better performance.Through the experiment,we identified 12 genes as biomarkers for pancreatic cancer with 91.28 percent accuracy.On the dataset of another platform,we verified the universality of the candidate genes we obtained,which reached above 80% classification accuracy.In addition,by using the SVM,LDM and BP Neural Networks classifiers,we compared the ordered feature sets generated by our proposed method with T-test,SVM-RFE and LDM-RFE,and the results showed that the proposed method achieved higher average classification accuracy.Meanwhile,functional analysis of 12 genes and pathway analysis on 200 genes showed some significant correlations with pancreatic cancer.Finally,we found that among the 12 genes,six feature genes(FN1,GABBR1,MGP,SRGN,MMP7 and DCN)may encode urinary excretory proteins and are closely related to the survival rates of pancreatic patients.The findings may be used to help medical experts to make better diagnose for pancreatic cancer in the future.