Protein-ATP interactions are ubiquitous in a wide variety of biological processes. Correctly locating ATP binding sites from protein information is an important but challenging task for protein function annotation and drug discovery. However, there is no method that can optimally identify ATP binding sites for different proteins. In this study, we report a new composite predictor, ATPbind, for ATP binding sites by integrating the outputs of two template-based predictors (i.e., S-SITE and TM-SITE) and three discriminative sequence-driven features of proteins: position specific scoring matrix, predicted secondary structure, and predicted solvent accessibility. In ATPbind, we assembled multiple support vector machines (SVMs) based on a random undersampling technique to cope with the serious imbalance phenomenon between the numbers of ATP binding sites and of non-ATP binding sites. We also constructed a new gold-standard benchmark data set consisting of 429 ATP binding proteins from the PDB database to evaluate and compare the proposed ATPbind with other existing predictors. Starting from a query sequence and predicted I-TASSER models, ATPbind can achieve an average accuracy of 72%, covering 62% of all ATP binding sites while achieving a Matthews correlation coefficient value that is significantly higher than that of other state-of-the-art predictors.