SCPSSMpred: A general sequence-based method for ligand-binding site prediction

In this paper, we propose a novel method, named SCPSSMpred (Smoothed and Condensed PSSM based prediction), which uses a simplified position-specific scoring matrix (PSSM) for predicting ligand-binding sites. Although the simplified PSSM has only ten dimensions, it combines abundant features, such as amino acid arrangement, information of neighboring residues, physicochemical properties, and evolutionary information. Our method employs no predicted results from other classifiers as input, i.e., all features used in this method are extracted from the sequences only. Three ligands (FAD, NAD and ATP) were used to verify the versatility of our method, and three alternative traditional methods were also analyzed for comparison. All the methods were tested at both the residue level and the protein sequence level. Experimental results showed that the SCPSSMpred method achieved the best performance besides reducing 50% of redundant features in PSSM. In addition, it showed a remarkable adaptability in dealing with unbalanced data compared to other methods when tested on the protein sequence level. This study not only demonstrates the importance of reducing redundant features in PSSM, but also identifies sequence-derived hallmarks of ligand-binding sites, such that both the arrangements and physicochemical properties of neighboring residues significantly impact ligand-binding behavior.