TSC_ATP: A two-stage classifier for predicting protein-ATP binding sites from protein sequence

It is well known that adenine triphosphate (ATP) binds with proteins to play important roles in metabolism, cell signaling, and as cofactor. Therefore it is crucial to identify ATP-binding sites on proteins to understand these mechanisms. In this paper, we present a computational method that can accurately predict ATP-binding sites on proteins using sequence-derived information. The algorithm is organized in a two-layered structure. The method first makes prediction by a K-Nearest Neighbors (K-NN) method using evolutionary profile information of each residue. The output of the first-layer classifier serves as an input feature together with other 11 features to a second-layer classifier. The final method achieved an AUC (area under the ROC curve) of 0.829 and 0.860 respectively on two benchmark datasets.