Title (ID: 3927860): RS-DLDock: a ligand docking method using Deep Learning (DL) scoring function coupled with realspace, X-ray/Cryo-EM refinement

Session:

DIVISION: Division of Computers in Chemistry
SESSION: Drug Design
SESSION TIME: 8:00 AM – 12:00 PM

DAY & TIME OF PRESENTATION: Thursday, August 17, 2023 from 9:45 AM – 10:00 AM
ROOM & LOCATION: Room 201, South Bldg. – Moscone Center

Abstract:

X-ray crystallography and cryogenic electron microscopy (cryo-EM) are the primary experimental techniques used to determine the three-dimensional (3D) structure of protein:ligand and protein:protein complexes, and these methods play a central role in Structure Based Drug Design (SBDD). To overcome weaknesses in conventional X-ray refinement using stereochemical-restraints, we incorporated QM and QM/MM functionals into crystal structure refinement [1, 2] and demonstrated the critical role of this refinement in tautomer/protomer determination and binding affinity prediction in SBDD [3]. During the model building stage of X-ray and Cryo-EM refinement, a significant challenge in ligand model placement is determination of the correct ligand orientation especially when the experimental density is weak or incomplete. We have addressed this ligand placement and refinement problem with the use of our MovableType fast free energy based docking algorithm (MTDock) [4] coupled with a built-in realspace (RS) refinement engine in which geometry energy and gradients – derived using a QM, MM or mixed QM/MM Hamiltonian – are combined with electron density gradients derived from X-ray data / Cryo-EM maps [1, 2]. In the present work, in order to significantly improve the predictability and accuracy of the scoring function, we implemented a Deep Learning (DL) algorithm and integrated it with our RS-dock procedure. With this RS-DLDock approach, each placed and refined ligand pose is scored with the random forest algorithm coupled with protein-ligand pairwise interactions derived from MovableType. This new RS-DLDock protocol has been added to our DivCon Discovery Suite and validated using 480 protein:ligand structures taken from the Iridium and PDBBind sets, and we will discuss our pose prediction and binding free energy prediction results.