In this study, we propose a supercomputer-assisted drug design approach involving all-atom molecular dynamics (MD)-based binding free energy prediction after the traditional design/selection step. Because this prediction is more accurate than the empirical binding affinity scoring of the traditional approach, the compounds selected by the MD-based prediction should be better drug candidates. In this study, we discuss the applicability of the new approach using two examples. Although the MD-based binding free energy prediction has a huge computational cost, it is feasible with the latest 10 petaflop-scale computer. The supercomputer-assisted drug design approach also involves two important feedback procedures: The first feedback is generated from the MD-based binding free energy prediction step to the drug design step. While the experimental feedback usually provides binding affinities of tens of compounds at one time, the supercomputer allows us to simultaneously obtain the binding free energies of hundreds of compounds. Because the number of calculated binding free energies is sufficiently large, the compounds can be classified into different categories whose properties will aid in the design of the next generation of drug candidates. The second feedback, which occurs from the experiments to the MD simulations, is important to validate the simulation parameters. To demonstrate this, we compare the binding free energies calculated with various force fields to the experimental ones. The results indicate that the prediction will not be very successful, if we use an inaccurate force field. By improving/validating such simulation parameters, the next prediction can be made more accurate.Key words computational drug design; molecular dynamics; binding free energy; high-performance computing; force field parameter As predicted by the Moore's law, 1) computational power progresses exponentially each year. The K computer (RIKEN, Japan) was the first one to reach the computational speed of 10 petaflops (PF).2) Subsequently, the United States and China released 10 PF-scale supercomputers. Thus, such huge computational power is expected to be utilized effectively for progress in a wide variety of science and technology fields. The ultimate purpose of this study is to develop and demonstrate an efficient drug design method with the help of such a stateof-the-art supercomputer.To overcome the difficulty of drug development, many computational structure-based drug design (SBDD) methods have been proposed in the last two decades. One important advantage of these computational methods is that they are free of experimental difficulty; thus, the computational methods are expected to reduce the experimental effort involved in the SBDD. The in silico SBDD methods can be categorized into two groups: virtual screening [3][4][5][6][7][8] and de novo drug design. 9-11)In virtual screening method, drug candidates are selected from libraries of chemical compounds by predicting their binding free energies approximately. In de novo drug de...

The computational structure-based drug design (SBDD) mainly aims at generating or discovering new chemical compounds with sufficiently large binding free energy. In any de novo drug design methods and virtual screening methods, drug candidates are selected by approximately evaluating the binding free energy (or the binding affinity). This approximate binding free energy, usually called "empirical score," is critical to the success of the SBDD. The purpose of this work is to yield physical insight into the approximate evaluation method in comparison with an exact molecular dynamics (MD) simulation-based method (named MP-CAFEE), which can predict binding free energies accurately. We calculate the binding free energies for 58 selected drug candidates with MP-CAFEE. Here, the compounds are generated by OPMF, a novel fragmentbased de novo drug design method, and the ligand-protein interaction energy is used as an empirical score. The results show that the correlation between the binding free energy and the interaction energy is not strong enough to clearly distinguish compounds with nM-affinity from those with µM-affinity. This implies that it is necessary to take into account the natural protein motion with explicitly surrounded by water molecules to improve the efficiency of the drug candidate selection procedure. Key words computational drug design; molecular dynamics; binding free energy; high-performance computingBecause the drug discovery becomes more costly and more time-consuming despite the continuous progress of technologies, more efficient and logical drug development methods are required.1) The computational structure-based drug design (SBDD) approach 2-4) is one promising prescription, and thus many methodologies and programs have been developed for the last two decades. The computational SBDD methods are categorized into two groups: the virtual screening [5][6][7][8] and de novo design.9,10) The virtual screening method selects bioactive compounds from chemical compound libraries, elements of which are usually available or already known. On the other hand, the de novo drug design generates chemical compounds from scratch.In any computational SBDD methods, drug candidates, having high binding affinity with the target protein, are selected from the chemical compound libraries or the computationally designed compound groups. For the selection, it is necessary to evaluate the binding affinity (the binding free energy) for every compound. Thus, the accuracy of the binding free energy evaluation is critical to the success of the computational SBDD. Even if the compound group includes several highaffinity compounds but if the selection method does not work well, it is impossible to discover strong candidates.Because the standard chemical compound library includes ca. 1000000 compounds, it takes huge computational time to calculate the binding free energies for many compounds. To reduce the required computational resource, the computational SBDD usually introduce approximations to the binding free energy. (The ...

No abstract