Flexible CDOCKER: Hybrid Searching Algorithm and Scoring Function with Side Chain Conformational Entropy

Abstract

The binding of small-molecule ligands to protein or nucleic acid targets is important to numerous biological processes. Accurate prediction of the binding modes between a ligand and a macromolecule is of fundamental importance in structure-based structure-function exploration. When multiple ligands with different sizes are docked to a target receptor, it is reasonable to assume that the residues in the binding pocket may adopt alternative conformations upon interacting with the different ligands. In addition, it has been suggested that the entropic contribution to binding can be important. However, only a few attempts to include the side chain conformational entropy upon binding within the application of flexible receptor docking methodology exist. Here, we propose a new physics-based scoring function that includes both enthalpic and entropic contributions upon binding by considering the conformational variability of the flexible side chains within the ensemble of docked poses. We also describe a novel hybrid searching algorithm that combines both molecular dynamics (MD)-based simulated annealing and genetic algorithm crossovers to address the enhanced sampling of the increased search space. We demonstrate improved accuracy in flexible cross-docking experiments compared with rigid cross-docking. We test our developments by considering five protein targets, thrombin, dihydrofolate reductase(DHFR), T4 L99A, T4 L99A/M102Q, and PDE10A, which belong to different enzyme classes with different binding pocket environments, as a representative set of diverse ligands and receptors. Each target contains dozens of different ligands bound to the same binding pocket. We also demonstrate that this flexible docking algorithm may be applicable to RNA docking with a representative riboswitch example. Our findings show significant improvements in top ranking accuracy across this set, with the largest improvement relative to rigid, 23.64%, occurring for ligands binding to DHFR. We then evaluate the ability to identify lead compounds among a large chemical space for the proposed flexible receptor docking algorithm using a subset of the DUD-E containing receptor targets MCR, GCR, and ANDR. We demonstrate that our new algorithms show improved performance in modeling flexible binding site residues compared to DOCK. Finally, we select the T4 L99A and T4 L99A/M102Q decoy sets, containing dozens of binders and experimentally validated nonbinders, to test our approach in distinguishing binders from nonbinders. We illustrate that our new algorithms for searching and scoring have superior performance to rigid receptor CDOCKER as well as AutoDock Vina. Finally, we suggest that flexible CDOCKER is sufficiently fast to be utilized in high-throughput docking screens in the context of hierarchical approaches.

Figure 1:

Receptor binding pocket complexity. (A) Distribution of number of receptor flexible side chains. (B) Distribution of number of receptor rotatable bonds.

Figure 2:

2OUN flexible self-docking results using the flexible receptor docking algorithm. These two docking poses (pink) belong to the same cluster and are native-like docking poses. The corresponding flexible side chain GLU726 adopts two different conformations. The backbone atoms of the two GLU726 conformations are shown in orange, while the side chain atoms are shown in yellow and blue, respectively. The side chain amide groups adopt two different orientations.

Figure 3:

Flexible docking searching algorithm.

Figure 4:

Average RMSD distribution of ligand docking poses in the initial generation. The RMSD values are binned with a 0.5 increment.

Figure 5:

Average population of native-like poses vs generation. Population for each docking measurement is calculated by dividing number of native-like poses by 500 (the number of trials in a generation). Average population of native-like poses for all 6 datasets is plotted with their corresponding error bars constructed by computing the standard deviation.

Figure 6:

Cumulative docking accuracy for the (A) T4 L99A dataset, the (B) T4 L99A/M102Q dataset and the (C) Riboswitch dataset. Distribution of ligand properties: (D) rotatable bonds, (E) logP and (F) molecular weight. A rank of N means the correct docking pose is within the top N solutions. AutoDock Vina trials use an exhaustiveness of 20. Rigid CDOCKER uses 500 docking trials for each cross-docking experiment.

Figure 7:

Cumulative docking accuracy for the (A) PDE10A dataset, the (B) DHFR dataset and the (C) Thrombin dataset. Distribution of ligand property: (D) rotatable bonds, (E) logP and (F) molecular weight. A Rank of N means the correct docking pose is within the top N solutions. AutoDock Vina trials use an exhaustiveness of 20. Rigid CDOCKER uses 500 docking trials for each cross-docking experiment.

Figure 8:

Cumulative docking accuracy for the (A) DHFR dataset and the (B) Thrombin dataset with some ligands starting with their native pose internal conformations. Top rank pose prediction accuracy of flexible docking is 74.54% and 74.18% using the proposed flexible receptor docking algorithm for each dataset, respectively.

Figure 9:

Properties of compounds in T4 L99A decoy set and T4 L99A/M102Q decoy set.