Cancer Detection Using Co-Training of SNP/Gene/MiRNA Expressions Classifiers

摘要

Recent studies have explored using SNPs, miRNA and gene expression profiles for detection of different types of cancer.Many attempts have been proposed in literature to build machine learning classifiers for these genomic data types. However, these studies did not totally exploit the relations between the three data types. These studies also suffer from the scarcity of microArray labeled data. In this paper, we propose a new system for detecting cancer and its subtypes using co-training of SNPs, gene and miRNA classifiers. We leverage the relations between SNPs and genes, and the relations between miRNAs and genes, to predict one type of expression from the other. We also introduce a new method to predict the SNP microarray data from the gene expression data and the opposite. We evaluated our mapping method on a paired dataset with SNPs and gene expression data. The results gives 6% average error for the predicted expression. Evaluation of the overall system on two types of cancer shows that our approach enhances the accuracy by up to 7.6% over the baseline individual classifiers.