CHARMM‐GUIhigh‐throughput simulator for efficient evaluation of protein–ligand interactions with different force fields

Abstract

Molecular docking is one of the most popular computational tools for the hit discovery step in drug design. However, there is ample room for improvement of docking's ability to identify correct binding modes and discriminate active from decoy compounds. Molecular dynamics (MD) simulations of protein–ligand docking structures have been shown to be effective in improving docking results. Here, we present CHARMM‐GUI high‐throughput simulator (HTS) that prepares MD simulation systems and inputs for multiple protein–ligand complex structures in a high‐throughput manner. HTS supports commonly used MD programs (NAMD, GROMACS, AMBER, OpenMM, GENESIS, Desmond, LAMMPS, and Tinker) along with various force field combinations for protein and ligand, including CHARMM36m, Amber (ff19SB/ff14SB), OPLS‐AA/M, CGenFF, GAFF2, and OpenFF. Validation tests using Miller and the directory of useful decoys‐enhanced (DUD‐E) datasets demonstrate that short MD simulations using HTS‐generated systems and simple ligand RMSD calculations consistently outperform docking results. Specifically, MD simulations can better identify correct ligand‐binding modes among top 10 binding poses as compared to docking scores. In addition, MD simulations can better discriminate active from decoy compounds in the DUD‐E dataset than docking scores for both soluble and membrane proteins. We expect that HTS can be a useful tool to facilitate the hit discovery process in drug design by improving docking results.

CHARMM-GUI Implicit Solvent Modeler for Various Generalized Born Models in Different Simulation Programs

“…In the explicit solvent MD simulations from Guterres et al, among 88 protein–ligand structures, ligand RMSD values in 49 systems are less than 5 Å, while our implicit solvent MD simulations show 46 ligands with RMSDs less than 5 Å. It is worth noting that our implicit solvent MD simulation find the same 42 stable protein–ligand systems out of 49 stable systems determined by explicit solvent MD simulations.…”

“…Studies have shown that the good initial ligand-binding modes could improve toward crystal structure poses during short MD simulations. , The docking structures from Guterres et al were divided into two classes (good modes: ligand RMSD < 5 Å and bad modes: ligand RMSD > 5 Å) based on the ligand RMSD relative to the corresponding crystal structures (Table S1). Therefore, according to their docking results, 20 out of 88 belong to the good binding modes.…”

“…Initial ligand-binding modes (pink) from top 8 AutoDock vina cross-docking of ligand 8DX (PDB ID 3FVK) onto Glur5 kainate receptor (PDB ID 3FUZ) and the binding poses (green) from the last snapshot of 50 ns implicit solvent MD simulations. The binding poses were ranked based on the vina docking score from pose 1 (P1) to P8 from Guterres et al The ligands in P1 and P5 are stable in the binding pocket throughout the MD simulations. The ligand in P2–4 and P6–8 escaped from the binding pocket during the implicit solvent MD simulations, so the binding poses from last snapshots are not shown.…”

“…To further test ISM, we performed MD simulations for protein–ligand systems generated by ISM. Recently, Guterres et al used 208 protein–ligand structures to demonstrate that explicit MD simulations can efficiently distinguish the good from bad ligand-binding modes . To examine the capability of implicit simulations, we randomly picked 88 out of 208 protein–ligand docking structures (Table S1).…”

“…Recently, Guterres et al used 208 protein−ligand structures to demonstrate that explicit MD simulations can efficiently distinguish the good from bad ligand-binding modes. 85 To examine the capability of implicit simulations, we randomly picked 88 out of 208 protein−ligand docking structures (Table S1). These 88 protein−ligand systems were simulated with GB-OBC2 using AMBER for 50 ns with three replicas.…”