A Novel Distance-based Scoring Function Using Artificial Neural Network

Along with the development of computer hardware, high performance computing resources is more available, so we could easily perform calculation on HPC clusters. This urged the cross-disciplinary Computer Aided Drug Design(CADD) to develop rapidly. In addition, the development of structural biology was very breathtaking. The number of proteins which were deposited in PDB increased exponentially from 507 in 1990 to 13597 in 2000 and it is nearly 80000 now. The Structure-based Drug Design(SBDD) was benefited from the development of biological macromolecular structures and the database of small compounds which was result of the development of combination chemistry. SBDD became a noteworthy branch of CADD and molecular docking was the most commonly method in SBDD.In recent years, machine learning approaches were applied widely in predicting affinities of protein-ligand complexes which was an extremely challenging task in Virtual Screening (VS) large scale small compounds database. Although there were a number of excellent scoring functions beyond traditional fitting scheme, there was none of them was suitable to use in any biological targets for any purposes. There is plenty room to improve. Herein, we constructed a novel non-linear docking scoring function, YZ-Score, using Artificial Neural Network. With PDBbind-cn v2011 as large-scale training set including 2239 complexes, YZ-Score was proved to be a promising, competitive scoring function, especially getting high correlation coefficient(R = 0.804) in independent test set which was composed with 216 complexes with experimental affinity. In this paper we focused in the prediction affinity of scoring function. In addition, we evaluated the performance of YZ-Score on "scoring power", "docking power" and "ranking power". "Ranking power" was the most important factor in Virtual Screening, which mean that a docking scoring function could distinguish the potential lead molecule from the compounds database. YZ-Score could improve the hitting success after the Autodock. In totally, the result was dramatic, overall, the performance of YZ-Score is excellent.For the excellent performance of YZ-Score, we constructed the homology models of Neuropeptide FF receptors using YZ-Score. Finally we indicated the ligand binding composes with NPFF receptor and the key bind site.Furthermore, the more high-quality 3-D structures and the more accurate interaction features between protein-ligand complex, the better performance of the scoring function.