A hybrid method for splice site prediction based on Markov model and codon information

Predicting splice sites is very important for gene identification. In this paper, we propose a hybrid splice site prediction method, SVM with Markov model and Codon usage (MC-SVM). The sequence features used for MC-SVM contain the codon bias information and the Markov probabilistic dependence information between adjacent nucleotides. Feature selection is performed using an F-score-based method, and then MC-SVM employs SVM to predict splice sites for both the acceptor and the donor sites. The test on the HS3D data set shows MC-SVM performs well for human gene sequences. The prediction accuracy of MC-SVM is 94.0% for donor splice sites, and 91.5% for acceptor splice sites on the data set with an equal amount of true and false splice site sequences. Compared with many other methods, MC-SVM achieved an improved prediction performance.