Scoring Ensembles of Docked Protein:Ligand Interactions for Virtual Lead Optimization

Abstract

Ensembles of protein structures to simulate protein flexibility are widely used throughout several applications including virtual lead optimization where they have been shown to improve ligand ranking. Yet, there is no established convention for weighting individual scores generated from ensemble members. To investigate the best method for weighting ensemble scores for proper ligand ranking, a series of dihydrofolate reductase inhibitors was docked to ensembles of Candida albicans dihydrofolate reductase (CaDHFR) structures created from a molecular dynamics (MD) simulation. From a single MD simulation, two ensemble collections were generated, one of which was subjected to a minimization procedure to create a group of structures of equal probability. As expected, ligand ranking accuracy was significantly improved when Boltzmann weighting was applied to the energies of this ensemble (60%), relative to that achieved with averaging (36%). However, accuracy was further improved (72%) by averaging docking scores across a minimized ensemble. To examine whether this accuracy results from structural variation in the single trajectory versus the possibility that error is minimized by averaging, a third collection of receptor structures was created in which each member was taken from an independent molecular dynamics simulation after minimization. Comparison of the docking accuracy results from the single trajectory (72%) to this third collection (61%) showed decreased accuracy, suggesting that ligands are more accurately oriented and assessed when docked to the minimized ensemble from a single MD trajectory, an effect that is more than simply error minimization. Averaging docking scores over a minimized ensemble of another target, influenza A neuraminidase, yielded a ligand ranking accuracy of 83%, representing a 24% improvement over other methods tested.

PL-PatchSurfer2: Improved Local Surface Matching-Based Virtual Screening Method That Is Tolerant to Target and Ligand Structure Variation

“…This improvement is consistent with a previous study, which used the Boltzmann weighting scheme for scoring 40 . The performance of BS with different parameter values, β and N conf , is provided in Supporting Information (Table S6).…”

“…First, the lowest Conformer Score (LCS) was taken as the final score to the ligand. Second, we computed a weighted average of Conformer Score which is inspired by the Boltzmann distribution (BS) 40,41 : $$\text{Boltzmann}-\text{Weighted}\text{Ligand}\text{Score}false(\mathrm{normalP},\mathrm{normalL}false)=\frac{true\sum _{\mathrm{normalC}}^{{\mathrm{N}}_{\text{conf}}}\text{PS}(\mathrm{P},\mathrm{C})\times expfalse[-\mathrm{normal\beta }\times \text{PS}false(\mathrm{normalP},\mathrm{normalC}false)false]}{true\sum _{\mathrm{normalC}}^{{\mathrm{N}}_{\text{conf}}}exp[-\mathrm{\beta }\times \text{PS}(\mathrm{P},\mathrm{C})]}$$…”

Challenges and advances in computational docking: 2009 in review

“…Anderson and co-workers have looked at this problem from two points of view. In the absence of general agreement about how to weight scores generated by individual ensemble members, they have carried out an extensive evaluation of different weighting schemes using structures of Candida albicans dihydrofolate reductase (CaDHFR) and influenza A neuranimidase from MD simulations (42). The schemes they have tested and their respective ligand ranking accuracies included simple averaging (36%), Boltzmann weighting without (60%) and with (72%) structural minimization, and taking initial structures from independent MD runs (61%).…”

Computer-aided drug design platform using PyMOL

“…Ensemble docking [18][19][20] has been put forward as an approach to include protein flexibility beyond the standard side-chain level during docking. The AutoDock Vina plugin allows the user to import an NMR style PDB file containing multiple protein structures or a single Amber trajectory file with multiple frames.…”

“…for AutoDock [6], each individual program usually requires a specific input structure and produces a specific output format not recognized by other computational programs. When this situation occurs, novice users are not capable of combining different programs such as using MD simulations to refine docking poses [14][15][16][17] or to perform ensemble docking based on a MD trajectory [18][19][20].…”