Biomarker rediscovery using Random Forests-based gene selection from microarray data of colon cancer

摘要

Background: After DNA microarray technique first described in 1995 and quickly adopted by the research community, it brings great convenience to our studies on identifying biologic markers, classifying types of tumours, predicting the outcome and response to chemotherapy.Although researchers have proposed a lot of approaches to deal with the data come from microarray, it is still a challenge to select feature genes containing most information about the diseases, as well as better understand underlying biological phenomena. Random Forests (referred as RF following) is an important method of classification and regression and can return feature gene set meanwhile.Additionally, to gain insight into biological phenomena, ways taking account of not merely single-gene level but also multi-gene level can be used, e.g., pathway analysis and gene ontology (referred to as GO following). These methods may lead to a better biological interpretation and supply us a method exploring feature genes independent of data.Results: In this paper, we selected 60 genes finally, which produced the smallest OOB error rate with RF. Furthermore, by using certain pathway and gene ontology databases, most of the genes selected by our approach were confirmed to have notable association with the formation and progression of various kinds of human cancer.Conclusions: We adopt a RF-based method of feature gene selection incorporating with backward eliminating thought to process the data of colon cancer and finally identified 60 genes which can better classify tumour and normal samples. For the purpose of biological interpretation and verifying these genes,pathway analysis and gene functional annotation were also referred to. The results showed that our approach was available for identifying feature genes and precise for molecular classification of colon tumour.