Identification of protein-interacting nucleotides in a RNA sequence using composition profile of tri-nucleotides

- Present study is an attempt to predict protein-interacting nucleotides (PINs) from RNA sequences
- Tri-nucleotide composition of sliding window patterns is most appropriate input features to develop predicting PINs.
- It was found that SVMlight based machine learning performed better than other classifiers.
- It was also found that certain di- and tri-nucleotides are preferred in the interaction with proteins.
- A user-friendly web-server 'RNApin' has been developed for the help of global scientific community.

The RNA-protein interactions play a diverse role in the cells, thus identification of RNA-protein interface is essential for the biologist to understand their function. In the past, several methods have been developed for predicting RNA interacting residues in proteins, but limited efforts have been made for the identification of protein-interacting nucleotides in RNAs. In order to discriminate protein-interacting and non-interacting nucleotides, we used various classifiers (NaiveBayes, NaiveBayesMultinomial, BayesNet, ComplementNaiveBayes, MultilayerPerceptron, J48, SMO, RandomForest, SMO and SVMlight) for prediction model development using various features and achieved highest 83.92% sensitivity, 84.82 specificity, 84.62% accuracy and 0.62 Matthew's correlation coefficient by SVMlight based models. We observed that certain tri-nucleotides like ACA, ACC, AGA, CAC, CCA, GAG, UGA, and UUU preferred in protein-interaction. All the models have been developed using a non-redundant dataset and are evaluated using five-fold cross validation technique. A web-server called RNApin has been developed for the scientific community (http://crdd.osdd.net/raghava/rnapin/).