Improved prediction of protein binding sites from sequences using genetic algorithm.

We undertook this project in response to the rapidly increasing number of protein structures with unknown functions in the Protein Data Bank. Here, we combined a genetic algorithm with a support vector machine to predict protein-protein binding sites. In an experiment on a testing dataset, we predicted the binding sites for 66% of our datasets, made up of 50 testing hetero-complexes. This classifier achieved greater sensitivity (60.17%), specificity (58.17%), accuracy (64.08%), and F-measure (54.79%), and a higher correlation coefficient (0.2502) than those of the support vector machine. This result can be used to guide biologists in designing specific experiments for protein analysis.