Prediction of Eukaryotic Translation Initiation Sites Using Machine Learning

The computational identification of translation initiation sites (TIS) is a major component of every gene prediction system, and is thus of major importance in genome annotation projects. A large number of machine learning methods have been described to identify TIS in transcripts such as mRNA, EST and cDNA sequences. In this regard, most of the prediction methods have focused on recognizing TIS in transcripts. However, recognizing TIS in transcripts is different from recognizing TIS in genomic sequences, mainly because of the following reasons. To begin with, the transcription start sites, which are needed to predict TIS in transcripts using the well-known scanning models, are usually unknown in genomes. Next, transcripts typically contain at most a few TIS candidates, while eukaryotic genomes contain millions of TIS candidates. Moreover, eukaryotic genomes contain introns, which disrupt the coding structure downstream of the TIS. This article focuses on the identification of TIS at the genomic level. Recently, on the other hand, proteogenomics, the integration of proteomics and genomics, has been raised, since the high-throughput identification of proteins and their accurate partial sequencing by tandem mass spectrometry combined with capillary liquid chromatography are now feasible for any cellular model at a full genomic scale. Proteogenomics can provide correction of current gene annotation, together with the possibility of identifying novel genes. In other words, the TIS prediction at the genomic level becomes simpler by using mass spectrometry-based proteome data. In such a situation, all you have to do for the TIS identification is to search the restricted region upstream of the most upstream peptide used to identify an expressed protein. Distinct sequence information for TIS identification in the vicinity of the start methionine enables us to easily predict TIS. To ascertain whether there is distinct sequence information for TIS identification in the vicinity of the start methionine, we applied two machine learning methods, decision tree learning and support vector machine (SVM), to discriminate true or false start methionine codons.