Structure based prediction of binding residues on DNA-binding proteins

Annotation of the functional sites on the surface of a protein has been the subject of many studies. In this regard, the search for attributes and features characterizing these sites is of prime consequence. Here, we present an implementation of a kernel-based machine learning protocol for identifying residues on a DNA-binding protein form the interface with the DNA. Sequence and structural features including solvent accessibility, local composition, net charge and electrostatic potentials are examined. These features are then fed into Support Vector Machines (SVM) to predict the DNA-binding residues on the surface of the protein. In order to compare with published work, we predict binding residues by training on other binding and non-binding residues in the same protein for which we achieved an accuracy of 79%. The sensitivity and specificity are 59% and 89%. We also consider a more realistic approach, predicting the binding residues of proteins entirely withheld from the training set achieving values of 66%, 43% and 81%, respectively. Performances reported here are better than other published results. Moreover, since our protocol does not lean on sequence or structural homology, it can be used to annotate unclassified proteins and more generally to identify novel binding sites with no similarity to the known cases.