Prediction Of Protein-ligand Complex Structures And Calculation Of Binding Free Energies

For the searching of biologically active small molecules,accurate prediction of protein-small molecule complex structure and binding free energy can largely reduce the blindness during the process of screening,thereby saving time and cost.This article focused on this topic and conducted the following researches:1.Using the Alanine Scanning-based molecular mechanics Generalized Born surface area-Interaction Entropy(ASGBIE)method developed by our group,the binding free energies of 34 small molecules from the 4^thGrand Challenge competition bound to the BACE protein were predicted,with the help of the modification on the complex crystal structure and molecular dynamics simulations.The Pearson linear correlation between our calculated results and the experimental values ranked the 2^ndplace among all the 53 groups of results submitted by contestants around the world.2.A complex trajectory data set was constructed through system screening,molecular dynamics(MD)simulation and trajectory stability analysis.The data set contains 670stable MD trajectories of totally 134 complexes,and the simulation time of each trajectory was 25 ns.Based on this data set,the performance of the ASGBIE method on binding free energy prediction was systematically assessed.The results showed that the overall performance of ASGBIE for different complexes is slightly lower compared to traditional methods such as the Molecular Mechanics Generalized Born Surface Area(MMGBSA)approach,due to some inherent limitations in this method.3.A scoring function named APBScore(Atom Pair Based Scoring function)was developed to predict the binding free energy of protein-ligand complex optimized with the Amber force fields using linear fitting on a total of 4197 complexes in the training set.Element type based pair-wised van der Waals energies,total electrostatic interaction energies,and hydrogen bond energies between the receptor and ligand were used as descriptors.The APBScore scoring function showed good performance on the overall free energy prediction for the test set containing 1964 complexes,as compared to other commonly used non-machine learning scoring functions,and it was also well-suited for discriminating the native binding pose from the decoy complex structures.4.A machine learning scoring function was trained with the 3D-CNN network architecture based on the machine learning scoring function K_DEEP,on a data set with about 12,000 complex crystal structures with known activities,by using element type channeled receptor ligand voxel as descriptors.This scoring function showed excellent scoring and ranking powers on the test set,but failed to differentiate the native binding pose from the decoy complex structures.5.Through molecular docking with Autodock Vina on the above-mentioned 12,000complexes,a docking data set was constructed containing more than 1.17 million docking poses.Based on this data set,a machine learning model named Deep BSP(Deep Binding Structure rmsd Prediction model)was developed to predict the root mean square deviation(RMSD)of a ligand docking pose with reference to its native binding pose by using the RMSD value of the ligand as the fitting target.For the test data set,by combining the Deep BSP model with Autodock Vina,the successful rate of selecting docked ligand poses with the RMSD value below 2.0?could increase from68%(when only Vina score was used)to about 80%.The above researches provided new theoretical tools in improving the accuracy of the structure and binding free energy prediction of protein-ligand complexes,and also provided physical foundation for follow-up researches.