A Novel Prediction Method for Metal-Ion Binding Sites in Protein Sequence Based on Ensemble Learning

The identification of metal ion-binding sites is important for detecting the protein structures and understanding its biological functions. However, in Protein Data Bank (PDB) which collects the known crystal structures of proteins, only less than one percent are membrane proteins even though they play a significant role in material exchange for cells and have a close relationship in drug target design. In this work, we develop an efficient prediction method for six different types of metal ion-binding sites in membrane proteins. In order to solve the imbalance problem in the dataset, multiple random down-sampling technique is used to obtain multiple training subsets with equal number of binding residues and non-binding residues. The support vector machines (SVM) and random forest (RF) classification models are built based on these subsets and their results are combined by ensemble learning algorithm which efficiently reduce the number of false positive samples in the final prediction. On an independent testing set, our proposed method achieves the average accuracy of 0.991 and average MCC of 0.681 which outperform a recently proposed prediction method, . The superiority in performance has demonstrated that our proposed method is expected to be an accurate tool for prediction of metal ion-binding sites in membrane proteins and it should provide assistant in design of new drug targets.