key: cord-0025005-rmjobgdw
authors: Li, Zhe; Li, Hui; Yu, Kunqian; Luo, Hai-Bin
title: Perspective of drug design with high-performance computing
date: 2021-06-18
journal: Natl Sci Rev
DOI: 10.1093/nsr/nwab105
sha: 3262009852eafe9342dc9e628ec13744faf15773
doc_id: 25005
cord_uid: rmjobgdw

The representative applications, recent advances and possible future directions of computational drug design were summarized, aiming to accelerate the drug discovery with the assistance of the fast-developing high-performance computing.

compounds in ZINC database has already reached 230 million, while that of 'make-on-demand synthesis' compounds can be 10 billion. For the scoring functions usually implemented in docking methods, ultra-high-throughput virtual screening is now possible by HPC. Although the accuracy of the docking-based virtual screening can be limited, the results are still promising because of the large number of compounds. In 2019, virtual screening of 138 million compounds was performed against D4 receptor, and 30 hits showed sub-micromolar activity including a 180 pM subtype-selective agonist [2] . Recently, an open-source virtual screening platform VirtualFlow was successfully conducted to screen 1.4 billion compounds and identified a potent KEAP1 inhibitor with nanomolar affinity [3] .

Accurate binding affinity prediction is another important issue in drug dis- [5, 6] such as scaffold hopping, ligand selectivity study, etc. Recently, a highly potent PDE10 inhibitor with subnanomolar affinity was discovered by FEP ABFE-based hit-to-lead optimizations [5] . However, these accurate binding prediction methods have not been widely applied in real drug discovery works because of high computational costs. With the continuous development of enhanced sampling methods and computational ability, these methods will have more impact on drug discovery in the near future.

Computational drug discovery by HPC has showed greater potential than at any time before in the development of promising agents against COVID-19.

Since the outbreak of COVID-19, billionlevel ultra-high-throughput virtual screenings were performed to find potential therapeutic agents. The work from LeGrand et al. using AutoDock-GPU in the summit supercomputer was shortlisted for the 2020 Gordon Bell Prize. Recently, we also finished a billion-level virtual screening by AutoDock-GPU against RdRp of SARS-CoV-2 in less than 24 hours on a domestic supercomputer. In another work, FEP was used to evaluate the docking-based virtual screening results towards the FDAapproved drug database [6] to give a more accurate ranking in binding affinity. As a result, 16 hits were identified from 25 selected drugs after bioassay. Among them, dipyridamole showed promising outcomes in the subsequent clinical trials [7] . Additionally, the roles of glycans in modulating the conformational dynamics of the SARS-COV-2 spike protein were also revealed by extensive MD simulations, which provided insight for further drug and vaccine design [8] .

Millions or even billions of binding affinity predictions via accurate methods could be one of the ultimate goals for computational drug discovery. However, this would require 10 3 to 10 6 times the computational resources of the fastest supercomputer in the world. Currently, combining calculation methods with different speed and accuracy, for example docking-based high-throughput prescreening and FEP-based binding affinity prediction, is an applicable strategy (Fig. 2) . Deep learning technologies can also accelerate calculations in drug discovery such as by sampling the Boltzmann distribution, and thus overcome the rare events problems [9] . The convolutional neural networks can be trained to predict the binding affinity from 3D structures [10] . In the foreseeable future, with the development of HPC and technologies such as enhanced sampling and deep learning, more innovative algorithms can be developed, which will further speed up the drug design and discovery.

Advance access publication 18

This work was supported by the National Key R&D Program of China (2017YFB0202600), the National Natural Science Foundation of China (21877134, 81903542 and 22077143) and the Science Foundation of Guangzhou City (201904020023).