Predicting Citrullination Sites in Protein Sequences Using mRMR Method and Random Forest Algorithm

As one of essential post-translational modifications (PTMs), the citrullination or deimination on an arginine residue would change the molecular weight and electrostatic charge of its side-chain. And it has been found that the citrullination in protein sequences was catalyzed by a type of Ca2+-dependent enzyme family called peptidylarginine deiminase (PAD), which include five isotypes: PAD1, 2, 3, 4/5, and 6. Citrullinated proteins participate in many biological processes, e.g. the citrullination of myelin basic protein (MBP) assists the early development of central nervous system. However, abnormal modifications on citrullinated proteins would also lead to some severe human diseases including multiple sclerosis and rheumatoid arthritis.

Therefore, it is necessary and important to identify the citrullination sites in protein sequences. The information about the location of citrulliantion sites in protein sequences will be useful to investigate the molecular functions and disease mechanisms related to citrullinated proteins.

In this study, we investigated the peptide segments that contain the citrullination sites in the centers, which were encoded into numeric digits from four aspects. Thus, we yielded a training set with 116 positive samples and 232 negative samples. Then, a reliable feature selection technique, called maximum-relevance-minimum-redundancy (mRMR), was applied to analyze these features, and four algorithms, including random forest (RF), Dagging, nearest neighbor algorithm (NNA), and support vector machine (SVM), together with the incremental feature selection (IFS) method were adopted to extract important features.

Finally an optimal classifier derived from RF algorithm was constructed to predict citrullination sites. 44 most prominent features were comprehensively analyzed and their biological characteristics in citrullination catalysis were also revealed.

We believed that the biological features obtained in this pioneering work would provide some useful insights into the formation and function of citrullination and the optimal classifier could be a useful tool to identify citrullination sites in protein sequences.