Skip to Main Content
Protein methylation modification has been discovered for half a century but still far less been studied than other modifications. Computational analysis is recently introduced to discover other unknown methylation sites based on few known ones. To effectively predict possible methylation, sophisticated classification strategy should be well devised. In this paper, we first extracted informative features from methylated fragments in many protein sequences, including the physicochemical properties, secondary structure information, evolutionary profiles, and solvent accessibility of surrounding residues. Then, an efficient feature selection method (mRMR) is applied to eliminate redundant features but keep important ones. Since methylated residues are far less than non-methylated, the collected data is relatively imbalanced. Thus, we propose to use the granular support vector machine (GSVM) which is specially designed for imbalanced classification problems. A 7-fold cross validation shows that our strategy generates comparable predication accuracy with many current methods or even better. Meanwhile, our method provides insights to identify the underlying mechanisms of protein methylation.