Protein methylation is known as one of reversible post-translational modifications and plays an important role in regulation pathways. Experimental identification of protein methylation sites is in general time-consuming and costs much. Computational approaches to predicting methylation sites provide an efficient way for screening of candidate methylation sites. In this study, hydrophobicity and transfer energy retrieved from the Amino Acid Index database are utilized as the features, and Support Vector machines are employed to perform the classification task. Two training sets are obtained from the MASA dataset, where the numbers of samples, including both 50% positive and 50% negative samples, for modifications of lysine and arginine residues are 266 and 314, respectively. The experimental results are then evaluated by leave-one-out cross validation. The accuracy of predicting lysine residues as the modification sites is up to 79.32%, and 82.8% for the case of arginine residues. The results also show that using these two features mentioned above is simple yet effective in comparison with other features proposed in the literature.
