Constructing a Foundational Platform Driven by Japan’s K Supercomputer for Next‐Generation Drug Design

Abstract

The cost of pharmaceutical R&D has risen enormously, both worldwide and in Japan. However, Japan faces a particularly difficult situation in that its population is aging rapidly, and the cost of pharmaceutical R&D affects not only the industry but the entire medical system as well. To attempt to reduce costs, the newly launched K supercomputer is available for big data drug discovery and structural simulation-based drug discovery. We have implemented both primary (direct) and secondary (infrastructure, data processing) methods for the two types of drug discovery, custom tailored to maximally use the 88 128 compute nodes/CPUs of K, and evaluated the implementations. We present two types of results. In the first, we executed the virtual screening of nearly 19 billion compound-protein interactions, and calculated the accuracy of predictions against publicly available experimental data. In the second investigation, we implemented a very computationally intensive binding free energy algorithm, and found that comparison of our binding free energies was considerably accurate when validated against another type of publicly available experimental data. The common feature of both result types is the scale at which computations were executed. The frameworks presented in this article provide prospectives and applications that, while tuned to the computing resources available in Japan, are equally applicable to any equivalent large-scale infrastructure provided elsewhere.

Molecular dynamics simulation-guided drug sensitivity prediction for lung cancer with rare EGFR mutations

“…The above findings regarding EGFR mutation diversity and the efficacy of EGFR-TKIs for a subset of mutants prompted us to apply our in silico drug sensitivity prediction model (31, 32). We have previously indicated that a differential sensitivity of two representative rare EGFR mutants, exon 20 insertion mutants A763_Y764insFQEA and A767_V769dupASV, to two EGFR-TKIs, afatinib and osimertinib, revealed that A763_Y764insFQEA was more sensitive than A767_V769dupASV (36).…”

“…First, to evaluate whether our supercomputer-based binding free energy (Δ G bind ) calculation model could predict the difference in sensitivity conferred by these mutations, we calculated Δ G bind values for the binding of afatinib to these EGFR mutants. In this model, the structures of EGFR molecules harboring rare mutations were built by homology modeling, and their binding affinities for EGFR-TKIs were evaluated using massively parallel computation of absolute binding free energy with a well-equilibrated system (the MPCAFEE method) (31, 32, 37). The modeled structure with bound afatinib is shown in SI Appendix , Fig.…”

“…We have previously developed the supercomputer-based binding free energy calculation model utilizing MD simulation (31, 32) and applied our model to secondary ALK and RET mutants, which appeared during therapy using TKIs (33, 34). Based on our previous work, we hypothesized that our supercomputer-based model would allow us to predict the sensitivity of rare EGFR mutants to EGFR-TKIs at a clinically relevant level.…”

Improving the Accuracy of Protein‐Ligand Binding Mode Prediction Using a Molecular Dynamics‐Based Pocket Generation Approach

“…These results demonstrate that although MD/MM‐PBSA may not always predict the correct binding mode because of the large calculation error, it can restrict the search space to approximately five candidates. These candidates can then be clearly discriminated by standard binding free‐energy computation methods which have higher computational costs per candidate , with simulation times of several tens to hundreds of nanoseconds. Discrimination may also be enabled by incorporating optimization methods such as “Best Arm Identification algorithm” into MD/MM‐PBSA scoring, which decreases the computational cost for significantly unstable binding poses and intensively evaluates the comparatively more stable ones …”

A secondary RET mutation in the activation loop conferring resistance to vandetanib

“…The MP-CAFEE method was used to calculate the protein-compound binding free energy 17 , 36 , 37 . For each protein-compound complex and solvated compound, a 32 λ parameter set for the Coulomb and van der Waals interactions was used 17 , and six independent simulations were performed with different initial velocities for each λ parameter.…”