Prediction of DNA-binding residues from sequence features.

Protein-DNA interaction plays a pivotal role in transcriptional regulation, DNA metabolism and chromatin formation. Although structural data are available for a few hundreds of protein-DNA complexes, the molecular recognition pattern is still poorly understood. With the rapid accumulation of sequence data from many genomes, it is important to develop predictive methods for identification of potential DNA-binding residues in proteins. In this study, neural networks have been trained using five sequence-derived features for prediction of DNA-binding residues. These features include the molecular mass, hydrophobicity index, side chain pKa value, solvent accessible surface area and conservation score of an amino acid. Interestingly, the side chain pKa value appears to be the best feature for prediction, suggesting that the ionization state of amino acid side chains is important for DNA-binding. The predictive performance is enhanced by using multiple features for classifier construction. The classifier that has been constructed using all the five features predicts at 72.71% sensitivity and 67.73% specificity. This is by far the most accurate classifier reported for prediction of DNA-binding residues from sequence data. The classifier has also been evaluated by using the Receiver Operating Characteristic curve and by examining the predictions made for different classes of DNA-binding proteins. Supplementary materials including the datasets are available at http://bioinformatics.ksu.edu/pdi/feature.html.