Constructing feature models for citrullination sites prediction

Protein citrullination is catalyzed by peptidylarginine deiminase (PAD), during which the positive charge of arginine is changed to the neutral charge of citrulline. Some human diseases such as rheumatoid arthritis, autoimmune diseases, and Alzheimer’s disease are known to be associated with PAD enzymes and citrullinated proteins. However, none of the existing prediction tools for citrullination have resulted in a good outcome. This study was conducted to evaluate the performance of the catalyzing rules of PADs, which have been described in previous studies. Machine-learning approaches were used to construct a prediction model for citrullination sites with eight features, i.e., catalyzing rules, sequence similarity, evolutionary information, physicochemical and biochemical properties of amino acids, secondary structure, and disorder and surface accessibility, which were derived from previous studies. We then designed small data modeling and proposed a feature model selection to construct the evaluation of feature model selection (FMS) model that could predict unknown citrullination candidates. Finally, our prediction model was able to achieve an accuracy of up to 0.90 and a Matthews correlation coefficient (MCC) of 0.80, while the selected features were almost similar to those in previous biological analyses.