DStruBTarget: Integrating Binding Affinity with Structure Similarity for Ligand-Binding Protein Prediction

Motivation: Identification of ligand-binding proteins is an important issue for drug development. Most of the current computational approach is developed only utilizing ligand structure similarity. However, the ligand structure similarity has failed to reflect the binding quality between the ligand and the target protein, which limited the performance of current methods. Results: The present study integrated two-dimensional (2D) and three-dimensional (3D) ligand structure similarity between query ligand and template with known ligandprotein binding affinity (BA) to identify proteins binding with the query ligand. This method is named as DStruBTarget. The performance of DStruBTarget was evaluated by 10-fold cross-validation in a dataset containing 9197 ligands and 1111 ligand-binding proteins (DBD dataset). This dataset was constructed by excluding the ligands with similar structures and the proteins with high sequence identity. The DStruBTarget achieved a hit rate of 77% in top 1 prediction, which is 4.80 and 3.00% better than the methods only using 2D structure similarity, and the method integrating 2D and 3D structure similarity (2D + 3D), respectively. An independent test of DStruBTarget was performed in a publicly available dataset constructed by SwissTargetPrediction. In this dataset, the top 1 hit rate of DStruBTarget reached 44.02%, which was better than the SwissTargetPrediction, and also outstands other methods, such as 2D, 3D, 2D + 3D, 2D integrating binding affinity (2D + BA), and 3D integrating binding affinity (3D + BA). DStruBTarget was compared to another newly published method HitPickV2 and achieved 52.17% hit rate of the top 1 prediction, which was significantly better than the result of HitpickV2 (30.43%). Finally, DStruBTarget was integrated with protein BLAST to predict the ligand-binding proteins not limited in a certain database. DStruBTarget with BLAST was tested in the DBD dataset. Its top 1 hit rate was 51.15%, which is lower than DStruBTarget without BLAST. Further comparison was on the ligands that bind to multiple numbers of proteins, which illustrated that DStruBTarget with BLAST performed better than without BLAST when the number of binding proteins of the query ligands is larger than six. Meanwhile, the prediction power of the DStruBTarget with BLAST in top 1 prediction was found to be positively correlated with the number of proteins binding with the query ligands, while the top 1 prediction power of DStruBTarget without BLAST was negatively correlated with the number of binding proteins for query ligands. Thus, DStruBTarget with BLAST is a potentially useful approach for predicting novel proteins for ligands that bind to multiple proteins.