Machine learning methods for transcription data integration

Gene expression is modulated by transcription factors (TFs), proteins that generally bind adjacent to DNA coding regions. The relation between these regulatory proteins and their targets is generally many to many; each target can be regulated by more than one TF, and each TF can contribute to the regulation of more than one target. Among the consequences of combinatorial regulation is the emergence of control networks, which can regulate nuanced changes in phenotype in response to subtle environmental changes. A first step in developing a complete molecular understanding of transcriptional regulation is to associate each TF with the set(s) of genes that it regulates. In this paper we report on the ability of support vector machines to associate 104 TFs with their binding sites. Several types of data are used to train classifiers for TF binding prediction in the Saccharomyces cerevisiae genome: position-specific scoring matrix (PSSM) matches, conservation of PSSM matches in other genomes, gene expression, phylogenetic profiles, Gene Ontology functional annotation, TF-target expression correlation, and promoter sequence composition using various subsequence features. The classifiers are combined using a weighting scheme so that each type of genomic data can contribute to the classification of a binding site based on its performance. In each case the data are compared to known true positives taken from ChIP-chip data , Transfac, and the Yeast Proteome Database. The SVM works best when all genomic data are combined, and it also rank orders data types by performance. On average, we can classify binding sites with a sensitivity of 73% and a positive predictive value of almost 0.9. New ideas and preliminary work for improving SVM classification on biological data are also discussed.