Benchmarking and analysis of DNA-binding site prediction using machine learning

We benchmarked the performance of machine learning based publicly available Web servers, predicting DNA-binding residues. A blind test on data sets derived from protein-DNA complexes submitted to PDB after the publication of these Web servers was performed. It was discerned that models trained on unusually large number of parameters show exaggerated performance during training, which could not be sustained on new proteins submitted after these publications. Also discussed are the optimum definition of binding site and a correspondence between residue propensity and its bias for predictability in positive and negative class.