Protein sumoylation sites prediction based on two-stage feature selection

Protein sumoylation is one of the most important post-translational modifications. Accurate prediction of sumoylation sites is very useful for the analysis of proteome. Though the putative motif Psi K XE can be used, optimization of prediction models still remains a challenge. In this study, we developed a prediction system based on feature selection strategy. A total of 1,272 peptides with 14 residues from SUMOsp (Xue et al. [8] Nucleic Acids Res 34:W254-W257, 2006) were investigated in this study, including 212 substrates and 1,060 non-substrates. Among the substrates, only 162 substrates comply to the motif Psi K XE. First, 1,272 substrates were divided into training set and test set. All the substrates were encoded into feature vectors by hundreds of amino acid properties collected by Amino Acid Index Database (AAIndex, http://www.genome.jp/aaindex ). Then, mRMR (minimum redundancy-maximum relevance) method was applied to extract the most informative features. Finally, Nearest Neighbor Algorithm (NNA) was used to produce the prediction models. Tested by Leave-one-out (LOO) cross-validation, the optimal prediction model reaches the accuracy of 84.4% for the training set and 76.4% for the test set. Especially, 180 substrates were correctly predicted, which was 18 more than using the motif Psi K XE. The final selected features indicate that amino acid residues with two-residue downstream and one-residue upstream of the sumoylation sites play the most important role in determining the occurrence of sumoylation. Based on the feature selection strategy, our prediction system can not only be used for high throughput prediction of sumoylation sites but also as a tool to investigate the mechanism of sumoylation.