Sequence-based protein-Ca2+ binding site prediction using SVM classifier ensemble with random under-sampling

Abstract

Calcium ions (Ca2+) are crucial for protein function. They participate in enzyme catalysis, play regulatory roles, and help maintain protein structure. Accurately recognizing Ca2+-binding sites is of significant importance for protein function analysis. Although much progress has been made, challenges remain, especially in the post-genome era where large volume of proteins without being functional annotated are quickly accumulated. In this study, we design a new ab initio predictor, CaSite, to identify Ca2+-binding residues from protein sequence. CaSite first uses evolutionary information, predicted secondary structure, predicted solvent accessibility, and Jensen-Shannon divergence information to represent each residue sample feature. A mean ensemble classifier constructed based on support vector machines (SVM) from multiple random under-samplings is used as the final prediction model, which is effective for relieving the negative influence of the imbalance phenomenon between positive and negative training samples. Experimental results demonstrate that the proposed CaSite achieves a better prediction performance and outperforms the existing sequence-based predictor, Targets.