Ensemble docking for intrinsically disordered proteins

Abstract

Intrinsically disordered proteins (IDPs) are implicated in many human diseases and are increasingly being pursued as drug targets. Conventional structure-based drug design methods that rely on well-defined binding sites are however, largely unsuitable for IDPs. Here, we present computationally efficient ensemble docking approaches to predict the relative affinities of small molecules to IDPs and characterize their dynamic, heterogenous binding mechanisms at atomic resolution. We demonstrate that these ensemble docking protocols accurately predict the relative binding affinities of small molecule α-synuclein ligands measured by NMR spectroscopy and generate conformational ensembles of ligand binding modes in remarkable agreement with experimentally validated long-timescale molecular dynamics simulations. Our results display the potential of ensemble docking approaches for predicting small molecule binding to IDPs and suggest that these methods may be valuable tools for IDP drug discovery campaigns.

Keywords: Drug Discovery; Ensemble docking; Intrinsically Disordered Proteins; Molecular Docking; Molecular Dynamics.

Figure 1.. Ensemble docking for intrinsically disordered proteins.

(A) A series of ligands that bind α-synuclein with experimental affinities previously determined by NMR spectroscopy (11). (B) A schematic illustration of the AutoDock Vina ensemble docking protocol for IDPs proposed here. For each conformation in an IDP ensemble, docking calculations are performed restricting the docking search space to a cubic volume surrounding the center-of-mass of each residue. One docked pose is returned for each residue, and the docked pose with the best docking score is selected as the docked pose for that conformation. This returns a docked ensemble containing one docked pose per conformation.

Figure 2.. IDP ligand affinity predictions from ensemble docking are consistent with experimental affinities from NMR spectroscopy and simulated affinities from MD simulations.

(A) Comparison of the relative ligand affinities predicted by the proposed ensemble docking protocols with experimental ligand affinities determined from NMR chemical shift perturbations (CSPs) and affinity predictions from long timescale MD simulations (11). Relative experimental affinities from NMR and affinity predictions from MD and each docking approach are scaled such that tightest binding ligand has a relative affinity value of 1.0. (B) Distributions of normalized docking scores from holo and apo Autodock Vina and DiffDock ensemble docking calculations. Docking scores obtained from each docking method have been normalized with min-max normalization using the highest and lowest docking score observed in calculations of all three ligands.

Figure 3.. Comparison of ligand binding poses obtained from long timescale MD simulations and IDP ensemble docking calculations.

Comparisons of Ligand 47 binding modes obtained from a long timescale MD simulation (top) and AutoDock Vina holo ensemble docking (bottom) for a representative conformational substate of α-syn-C-term identified by t-SNE clustering. The MD ligand-bound ensemble contains all frames from an unbiased MD simulation where at least one heavy atom of α-syn-C-term is within 6Å of at least one heavy atom of the ligand. In holo docking calculations, the ligand is removed from all conformations in the ligand-bound MD ensemble, and a docked pose is predicted for each conformation. (A) Overlay of representative snapshots of ligand-bound conformations from MD and from ensemble docking in the selected t-SNE cluster. α-syn-C-term residues are colored by a gradient corresponding to the contact probability of Ligand 47 with each residue in the selected cluster. (B) Populations of intermolecular interactions between Ligand 47 and each residue of α-syn-C-term in ligand-bound ensembles in the selected t-SNE cluster. (C) The probability that Ligand 47 simultaneously forms contacts with each pair of residues of α-syn-C-term residues in the selected t-SNE cluster.

Figure 4.. Comparison of ensemble averaged intermolecular protein-ligand interactions and dual-residue contact probabilities between Ligand 47 and α-syn-C-term obtained from a 200μs MD simulation with Ligand 47, AutoDock Vina holo docking calculations, and AutoDock Vina apo docking calculations.

(A) Average populations of intermolecular interactions between Ligand 47 and each residue of α-syn-C-term in ligand-bound ensembles, weighted by the populations of each cluster identified by t-SNE for the original MD simulation. (B) The average probabilities that Ligand 47 simultaneously forms contacts with each pair of residues in α-syn-C-term, weighted by the populations of each cluster identified by t-SNE for the original MD simulation.

Figure 5.. Comparison of the ligand RMSD between docked poses and MD bound poses pf Ligand 47.

Distribution of frame-matched RMSD values (A) and best-matched RMSD values (B) obtained from AutoDock Vina holo docking calculations (red) and DiffDock holo docking calculations (blue) of Ligand 47. The fraction of frames with ligand RMSD values less than 3Å are displayed for each docking method. (C) A representative frame showing the α-syn-C-term protein coordinates used for docking in a blue-to-red gradient. The docked ligand poses predicted by AutoDock Vina ensemble docking (red) and by DiffDock docking (blue) are compared to the ligand pose in the original MD bound pose (gray), and the ligand RMSD values between the docked pose and the MD pose are displayed (top). The best-matched MD poses are shown for the AutoDock Vina docked pose (middle) and DiffDock docked pose (top). The α-syn-C-term coordinates and ligand coordinates of the best-matched MD poses are colored tan. The ligand RMSD values between the best-matched MD pose and each docked pose are displayed.