Improve the prediction of RNA-binding residues using structural neighbours

The interactions between rna-binding proteins (rbps) with rna play key roles in managing some of the cell's basic functions. the identification and prediction of rna binding sites is important for understanding the rna- binding mechanism. computational approaches are being developed to predict rna-binding residues based on the sequence- or structure-derived features. to achieve higher prediction accuracy,improvements on current prediction methods are necessary. we identified that the structural neighbors of rna-binding and non-rna-binding residues have different amino acid compositions. combining this structure-derived feature with evolutionary (pssm) and other structural information (secondary structure and solvent accessibility) significantly improves the predictions over existing methods. using a multiple linear regression approach and 6-fold cross validation,our best model can achieve an overall correct rate of 87.8% and mcc of 0.47,with a specificity of 93.4%,correctly predict 52.4% of the rna-binding residues for a dataset containing 107 non-homologous rna-binding proteins. compared with existing methods,including the amino acid compositions of structure neighbors lead to clearly improvement. a web server was developed for predicting rna binding residues in a protein sequence (or structure),which is available at http://jeele.go.3322.org/rna/. © 2010 bentham science publishers ltd.