Advancing the field of computational drug design using multicanonical molecular dynamics-based dynamic docking

Abstract

Multicanonical molecular dynamics (McMD)-based dynamic docking is a powerful tool to not only predict the native binding configuration between two flexible molecules, but it can also be used to accurately simulate the binding/unbinding pathway. Furthermore, it can also predict alternative binding sites, including allosteric ones, by employing an exhaustive sampling approach. Since McMD-based dynamic docking accurately samples binding/unbinding events, it can thus be used to determine the molecular mechanism of binding between two molecules. We developed the McMD-based dynamic docking methodology based on the powerful, but woefully underutilized McMD algorithm, combined with a toolset to perform the docking and to analyze the results. Here, we showcase three of our recent works, where we have applied McMD-based dynamic docking to advance the field of computational drug design. In the first case, we applied our method to perform an exhaustive search between Hsp90 and one of its inhibitors to successfully predict the native binding configuration in its binding site, as we refined our analysis methods. For our second case, we performed an exhaustive search of two medium-sized ligands and Bcl-xL, which has a cryptic binding site that differs greatly between the apo and holo structures. Finally, we performed a dynamic docking simulation between a membrane-embedded GPCR molecule and a high affinity ligand that binds deep within its receptor’s pocket. These advanced simulations showcase the power that the McMD-based dynamic docking method has, and provide a glimpse of the potential our methodology has to unravel and solve the medical and biophysical issues in the modern world.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12551-022-01010-z.

Evolution of molecular docking as a tool for drug design: from classical to dynamic approaches

Drug design is the process of developing new chemical compounds and using them to correct the defective molecular mechanism of a biological target. The first step to understanding the molecular mechanism of a compound is to determine how it binds to its bio-molecular target, more specifically, to its protein receptor. Molecular docking between a receptor and a compound, or a library of compounds, has been used for decades in the drug discovery process. Classical docking with programs such as DOCK (Kuntz et al. 1982), AutoDock (Goodsell and Olson 1990), FlexX (Rarey et al. 1996), GOLD (Jones et al. 1997), and myPresto/sievgene (Fukunishi et al. 2005) inherently treats the receptor as a rigid molecule, with only the ligand as fully flexible. However, the docking results strongly depend on the structure of the receptor. One early attempt to include receptor flexibility was to use ensemble docking, where either multiple experimental conformations of a protein were used (Craig et al. 2010; Wada et al. 2011), or by computationally generating the alternative conformations, e.g., via a rotamer library or by molecular dynamics (MD) simulations (Oda et al. 2018), after which the classical docking methodology would be applied to the ensemble to generate the docking poses. While such approaches take protein flexibility into account to some degree, the conformational space searched by such ensemble docking methods is still quite limited. In addition, the scoring functions used by the classical docking approaches lack accuracy due to many approximations and the usage of empirical parameters. Furthermore, to properly understand the molecular mechanism of the drug, it is important to not only understand the final bound state, but also the binding process, including intermediary states. Therefore, a method that can accurately predict protein–ligand complex formation is required.

One solution is to exhaustively sample the conformational space of the receptor and the configurational space of protein–ligand complexes simultaneously using conventional (i.e., to brute force it) MD simulations. Although this would correctly model the binding/unbinding between the receptor and the ligand, the time required to gather enough statistics from the binding/unbinding events to identify the native binding configuration and binding mechanism makes this approach effectively useless, even on expensive, hardware accelerated MD supercomputers such as ANTON (Shaw et al. 2008). In the past decade, an alternative to these expensive computers has become the usage of general-purpose graphics processing units (GPGPU) accelerated PC clusters, enabling larger and longer calculations at much lower costs than before. This has also given rise to a new class of docking methodologies, which have been coined dynamic docking (Gioia et al. 2017). One of the key advantages of dynamic docking techniques is the high accuracy of the binding configuration prediction because it treats both the receptor and the ligand as flexible. However, this still takes an unsurmountable amount of time if performed with conventional MD simulations, due to the existence of high energy barriers that are commonly found in bio-molecular systems. Therefore, using enhanced sampling MD methods such as multicanonical MD (McMD) (Nakajima et al. 1997), replica exchange MD (REMD) (Sugita and Okamoto 1999) , Filling Potential (Fukunishi et al. 2003), MetaDynamics (Laio and Parrinello 2002), Accelerated MD (Hamelberg et al. 2004), mD-VcMD (Hayami et al. 2019), and hypersound accelerated MD (Araki et al. 2021) should enable effective dynamic docking. Each of these methods adds a bias to the system, either by increasing the temperature of the whole or part of the system (e.g., McMD, REMD, AMD), by applying a bias to the center-of-mass (or landmark atom) of a molecule based on a pre-determined reaction coordinate (e.g., FP, MetaD, mD-VcMD), or via non-equilibrium approaches (e.g., hypersound accelerated MD).

One powerful, but woefully underutilized enhanced sampling methodology is McMD, which has a long history at Prof. Nakamura’s lab (Nakajima et al. 1997; Kamiya et al. 2002, 2007, 2008; Ikebe et al. 2011; Higo et al. 2012; Bekker et al. 2017), with a theoretical description of the methodology described in Section S1. Here, a bias, determined via a pre-run, is applied to the system to enable uniform sampling over a specific potential energy range, facilitating a random walk within this energy range, where each energetic state is equally probable to be sampled. This enables exploring a wide conformational space without conformational trapping at local minima, as the method can randomly transition between the higher and lower temperature regions. One of the merits of the McMD method is that the physically accepted ensemble, i.e., the canonical ensemble, can be obtained by a reweighting procedure (Nakajima et al. 1997), which determines the probability that a structure could be sampled at a given temperature given its potential energy. The McMD method first forms the multicanonical ensemble, covering a wide temperature range. Then, the conformational ensemble at 300 K is obtained from the multicanonical ensemble by the reweighting method, enabling structural analysis of the reweighted multicanonical ensemble, to extract representative structures and rank them. REMD is another popular generalized ensemble method (Sugita and Okamoto 1999) and has also been applied to find native binding configurations (Kokubo et al. 2011). For both the McMD method and the REMD method (as well as its derivates REST (Liu et al. 2005) and REST2 (Wang et al. 2011)), the canonical ensemble at 300 K can be reobtained, although with McMD data from the entire energy range can be integrated. With REST and REST2, only part of the system is accelerated, while the other part remains in the canonical ensemble. Nakajima et al. (Nakajima 1998) also developed a two-component McMD version to only accelerate the solute. Unlike McMD, REMD and REST do not require a pre-run, but require manual setting and tuning of parameters such as the temperature of each of the replicas, which increase as the system size increases, as overlap between potential energy distributions of the neighboring replicas is required for the method to work. In addition, REMD requires its replica simulations to be executed simultaneously and synchronously, increasing the computational requirements. With McMD, multiple parallel trajectories are only used to increase statistics and improve sampling efficiency, with each parallel trajectory covering the entire potential energy range. Furthermore, parallel trajectories are executed independently and can thus be executed serially, so McMD could even be used on small clusters or even a single machine, thus making McMD more practical for smaller research groups.

We have taken the powerful McMD-enhanced sampling method and developed our own dynamic docking methodology and toolset (Bekker and Kamiya 2021a), with a schematic overview of our dynamic docking protocol shown in Fig. 1. Using our methodology, we successfully predicted the native binding configuration between lysozyme to an oligosaccharide (Kamiya et al. 2008), cyclin-dependent kinase 2 to a small compound inhibitor (Bekker et al. 2017), β-secretase (BACE) to a medium-sized compound inhibitor (Bekker et al. 2019a), an antibody to its peptide antigen (Bekker et al. 2020b), heat shock protein 90 (Hsp90) to a small compound inhibitor (Bekker et al. 2020a), the binding of medium-sized compounds to the cryptic binding site of B-cell lymphoma-extra large (Bcl-xL) (Bekker et al. 2021b), the binding of a small compound inhibitor to the β₂-adrenergic G-protein Coupled Receptor (β₂AR GPCR) (Bekker et al. 2021a), and the binding of an antigen peptide to Human Leukocyte Antigen-A*2402 (Bekker and Kamiya 2021b). Here, we will describe our latest efforts in advancing the field of computational drug design by describing the first application of our methodology to exhaustively sample the configurational space between an inhibitor and Hsp90, the dynamic docking of two ligands to the cryptic binding site of Bcl-xL and the docking of a ligand to a membrane-embedded GPCR molecule.

Fig. 1.

Schematic overview of our McMD-based dynamic docking protocol. The McMD-based dynamic docking protocol can be thought of to consist of eight steps, which are described in Section S2. This image uses data from our WEHI – Bcl-xL docking simulations (Bekker et al. 2021b). The figures were generated using Molmil (Bekker et al. 2016), a molecular viewer developed by Protein Data Bank Japan (Kinjo et al. 2017, 2018; Bekker et al. 2022)

Methodology and application of exhaustive McMD-based dynamic docking to Hsp90 and a high-affinity ligand

In our early works, we focused only on predicting the binding configuration with the assumption that the binding site was known. In those studies (Bekker et al. 2017, 2019a, 2020b), we used a cylinder to restrain the ligand to a binding site and the nearby bulk region, which leads to improved prediction accuracy and simultaneously suppresses the need for extremely long simulations, while still attaining sufficient binding/unbinding event statistics due to the reduced search space. In our Hsp90 docking work, we docked a modified ligand (VER49008) (Schmidtke et al. 2011) with a known high affinity to Hsp90 (Bekker et al. 2020a). However, unlike our previous works, we did not use a cylinder to restrain the ligand, but instead performed an exhaustive search of the configurational space of the entire N-terminal domain of human Hsp90, where its natural ligand, ATP, normally binds. Thus, we intended to not only predict the native binding configuration, but simultaneously the native binding site with this exhaustive simulation. To prevent unfolding at high temperatures during the McMD simulations, the Cα atoms corresponding to the residues belonging to either an α-helix or a β-sheet as calculated by DSSP (Kabsch and Sander 1983) were position restrained during the McMD simulations. For the McMD simulation, we first executed a pre-run to estimate the density of states for a total duration of 10.81 µs and an additional 10.24 µs to equilibrate the system. Finally, a production run lasting 32 µs was executed to sample the binding configurations.

As part of this work, we also introduced a more refined and robust analysis method of the multicanonical ensemble (Fig. 1). In our preceding works, we used simple K-means clustering with k = 10 clusters on the first three principal components (PCs) obtained by principal component analysis (PCA) to obtain a set of representative structures. Here, we made several changes to this methodology. First, instead of just using three PCs, we dynamically calculated the number of PCs so that the cumulative explained variance of the selected components is at least 90%, so that most of the dynamics are described in the coordinates used for the clustering, which amounted to 5 PCs for our Hsp90 simulation data. Subsequently, we performed K-means clustering on this five-dimensional space (PC1–PC5), using a total of k = 1000 clusters and selected a single representative structure for each cluster. Finally, we compared the intermolecular contact matrices of all representative structures with each other using our R-value analysis (Bekker et al. 2019a) to merge similar clusters (R-value > 0.75) together and then ranked the clusters and their corresponding structure based on the cluster free energy (CFE), which is the sum of the McMD-reweighted probabilities of all the structures assigned to each cluster. Here, the R-value is a measure of the similarity rate between the short-range contact matrices of two arbitrary configurations and/or conformations, where a value of 1.0 corresponds to matching contact matrices. The new methodology has several advantages compared to the one we used in our preceding works (Bekker et al. 2019a, 2020b). Specifically, first, the automatic PC selection makes for a robust and scalable method, while it coincidentally also leads to an increase in the number of selected PCs, increasing the descriptiveness of the coordinate space used during the clustering step. Secondly, the significant increase in the number of clusters during the K-means clustering step means that outlier structures are less likely to be assigned to an inappropriate cluster, making the clusters more homogenous and the picked representative structures similarly more representative for their respective cluster. Finally, we obtain a set of the uniquely sampled representative structures along with their relative probabilities (i.e., their CFE) after merging clusters with similar structures (as the initial number of clusters is generally larger than required for the configurational space within the chosen structural similarity cutoff), which we can then use to rank and filter the structures.

For our Hsp90 docking, 12 representative structures were selected within a CFE cutoff of 2.5 kcal/mol (Fig. 2a), where the most top-ranking structure (r₁) had an R(native)-value of 0.89 and a root-mean-square-deviation (RMSD) of 0.76 Å, indicating that both the intermolecular interactions between the ligand and Hsp90, as well as the absolute positioning of the ligand, were well reproduced by our simulations. Subsequent canonical MD simulations at 300 K and 400 K, which can be used to validate the results if no experimental data is available (Bekker et al. 2019a, b), validated our results, showing that the most top-ranking structure (r₁) is also the most stable one. The second-ranking structure, which was also bound to the same pocket but in a different orientation, also showed a high score during the subsequent validation step, suggesting that this binding configuration would most likely also contribute to the inhibiting effect of the ligand. Finally, our results showed that a loop structure near the binding site of Hsp90, which we referred to by “FHL” (Fig. 2b), had attained a helical conformation in the most top-ranking representative structure r₁ (Fig. 2c, e), while a loop-in structure was observed for the experimental structure of our ligand’s analog (Fig. 2b, d), which was postulated by Amaral et al. to be caused by the size of the ligand, where only large ligands would result in a helical state (Amaral et al. 2017). Analysis of our multicanonical ensemble showed however that more than 50% of the near-native binding configurations were in a helical state, compared to only 6% of the unbound structures, which suggested to us that in explicit solvent, some of the interactions between the ligand and Hsp90 in the residues in the FHL as observed in the experimental structure become unstable, causing the FHL to predominantly form a helical structure. Thus, with our first attempt at using McMD-based dynamic docking to perform an exhaustive search of the ligand’s binding configurations did not only accurately reproduce the native binding configuration, it also reproduced the native binding site. In addition, we also predicted an alternative binding configuration in the same site and showed that the FHL region is likely to attain a helical configuration upon binding a relatively small ligand. Our work with Hsp90 has not only provided insight into the molecular mechanism of ligand binding to the protein, but has also enabled us to benchmark exhaustive McMD-based dynamic docking and pioneer a new, more refined, and robust analysis method that we have since used in all our subsequent research.

Fig. 2.

Docking results obtained from our Hsp90-VER49008 ligand dynamic docking simulations, data available from BSM-00010. (a) Overview of the top-ranking 12 structures, colored by their cluster free energy from blue to red, with the experimental structure in white and the helices and sheets labeled. (b) Cartoon representation of the experimental structure from the front, with the “FHL” loop in pink. This flexible lid region has a sequence of INNLGTIA, spanning from Ile104 to Ala111. (c) Cartoon representation of the top-ranking structure r₁ from the front, with the “FHL” loop in pink. (d) Stick representation of the binding site of the experimental structure from the side, with nearby crystal waters shown (note: hydrogens were generated). (e) Stick representation of the binding site of the top-ranking structure r₁ from the side, with nearby water molecules shown

Dynamic docking of two medium-sized compounds to the cryptic binding site of Bcl-xL

In our subsequent work, we studied two medium-sized compounds that are known to bind to the cryptic binding site of Bcl-xL (Bekker et al. 2021b), an anti-apoptotic protein that in various cancers becomes overexpressed, with cells losing their ability to inhibit apoptosis, thus making Bcl-xL an important drug target. For this protein, the conformation of the binding site differs greatly between the apo and the holo states, thus requiring large conformational changes to the protein to occur to enable ligand binding. We selected two medium-sized ligands, ABT-737, which was the first BH3-mimic inhibitor developed and mimics the Bcl-2 homology 3 only (BH3-only) motif found in BCL-xL’s natural ligands such as Bim, Puma, and Bad (Petros et al. 2000; Liu et al. 2003), and the second ligand WEHI-539, which was the first Bcl-xL-selective inhibitor developed (hereafter referred to by as ABT and WEHI, respectively).

To give ourselves the most difficult challenge for this project, we had decided to perform an exhaustive binding simulation starting from the apo state of Bcl-xL, thus requiring our McMD-based dynamic docking simulations to not only sample all the binding configurations, but to also make it find the correct binding site first. Unlike our above-described Hsp90 simulation where the apo and holo structures do not differ significantly, here, we have considerable differences between the structures. Furthermore, we were doing two independent simulations, one for each ligand. One of the biggest hurdles was that large conformational changes of Bcl-xL were required, but on the other hand we also had to restrain our protein structure to prevent unfolding during the McMD simulations. To allow for sufficiently large conformational changes to occur, but to prevent complete unfolding, we had decided to go with a weak distance restraint protocol, as the position restraints used during our Hsp90 docking would have been far too stringent. We had previously used distance restraints with our lysozyme and BACE dynamic docking works (Kamiya et al. 2008; Bekker et al. 2019a), but in those cases a very large number of restraints was used, leading to both a more rigid structure and a higher computational load. Instead, we used distance restraints between the backbone oxygen and nitrogen atoms for the residues that form hydrogen bonds to stabilize the secondary structure determined from the apo state. However, as these restraints would only prevent the secondary structure from breaking at high temperature, we also added restraints between a set of manually picked Cα atom pairs to restrain the tertiary structure. For both systems, the pre-run lasted for 24 µs, which was then followed by a production run, each lasting 32 µs. For the analysis of the data, we used the analysis protocol as described above in the Hsp90 section (see also Fig. 1 and Section S2).

Although no cylinder was used, the top four ranking structures were all located in the native binding site for both WEHI (Fig. 3a) and ABT (Fig. 3b). For WEHI, the docking resulted in a top-ranking structure (r₁, Fig. 3c) that only had an R(native)-value of 0.447, as the ligand had only bound at the experimental binding site partially, while the other part was in a more easily attainable state pointing outwards, instead of binding inside the cryptic pocket. Notably, this is similar to the configuration of W1191542, an analog of WEHI, as found in the PDB structure 3INQ (Lee et al. 2009). On the other hand, the third-ranking structure (r₃, Fig. 3d) had an R(native)-value of 0.752 and an RMSD of 3.18 Å, showing that the experimental configuration can be reproduced, but is difficult to attain due to the required conformational changes of Bcl-xL. The fourth-ranking structure only had an R(native)-value of 0.491, but had an RMSD of only 1.51 Å, showing that the ligand had positioned itself similar to the experimental structure, but that Bcl-xL was in a different conformational state. For the ABT docking, better results were observed, with the top-ranking structure having an R(native)-value of 0.674, while the third-ranking structure had an R(native)-value of 0.844. Thus, for WEHI, we obtained a number of stable binding configurations, presumably due to WEHI’s difficulty of binding deeply inside the pocket and the large conformational changes required of Bcl-xL for this to occur. We subsequently performed our usual canonical MD simulations for validation, showing the best results for the third-ranking structure of WEHI (which is also the most buried configuration), as well as for the top-ranking structure of ABT.

Fig. 3.

Docking results obtained from our Bcl-xL dynamic docking simulations, data available from BSM-00021. (a) Tube representation of the receptor and stick model for the ligand for the top four ranking structures from our WEHI-ligand docking simulation in green, cyan, magenta and black, respectively. (b) Tube representation of the receptor and stick model for the ligand for the top four ranking structures from our ABT-ligand docking simulation in green, cyan, magenta, and black, respectively. (c) Cartoon representation of the receptor and stick model for the ligand for the r₁ structure of our WEHI-ligand docking. Also shown are the nearby residues as stick models. (d) Cartoon representation of the receptor and stick model for the ligand for the r₃ structure of our WEHI-ligand docking. Also shown are the nearby residues as stick models

Analyzing the binding/unbinding pathway (step 7 in Fig. 1) indicated that the configurational ensemble had low probability regions along the binding pathway, which suggested that the binding event is sudden and thus follows a population shift between the receptor’s open and closed structures, where Bcl-xL would first need to attain a favorable structure, before the ligand can bind. It also showed that the structure r₁ of WEHI (Fig. 3c) was actually an intermediary structure to the final bound state (r₃, Fig. 3d). Although both structures started from the same apo state, the different ligands had a profound effect on the conformational ensemble of Bcl-xL, with large conformational changes required to enable the ligand to bind, especially for WEHI, where the ligand binds following a two-step model: first to our top-ranking structure (unbound/intermediary state to r₁, Fig. 3c), then after sufficient conformational changes of Bcl-xL so that the cryptic binding site appears, to the near-native third-ranking (Fig. 3d) state (i.e., r₁ to r₃). This also gives a clue to the low ranking of r₃; r₁ is a much more easily attainable state, while r₃ requires considerable conformational changes of Bcl-xL, followed by a partial flip of WEHI to enable this binding configuration to form.

To validate the pathway and to better analyze it, we also performed path sampling simulations starting from these structures (step 8 in Fig. 1). The calculated affinities showed slightly higher values compared to the experimental ones, but more peculiar was the difference between the two ligands when comparing the experimental to the computational results. Experimentally, ABT was shown to bind a bit stronger to Bcl-xL than WEHI, and WEHI furthermore has faster association and dissociation rates than ABT (Lessene et al. 2013). However, our dynamic docking simulations showed that binding of WEHI (in the r₃ configuration, Fig. 3d) to be much more difficult, following a two-step binding mechanism, indicating that binding of WEHI would be slower than ABT; thus, WEHI would have slower association and dissociation rates than ABT to the final native binding configuration. However, since the experimental binding kinetics consider all configurations, it is quite likely that our top-ranking configuration of WEHI (r₁, Fig. 3c), which is only the partially bound structure, is the major contributor to the binding kinetics, suggesting that in explicit solvent, the binding populations might be more similar to our dynamic docking results, rather than just the experimental crystal structure, which might have been stabilized due to crystallization effects. This is also corroborated by another PDB structure, 3INQ (Lee et al. 2009), where an analogous ligand of WEHI is bound to Bcl-xL in a configuration similar to WEHI’s top-ranking structure r₁. Thus, given this evidence, we can assume that r₁ is likely the native binding configuration of WEHI, with r₃ a rare, deeper bound state that might sometimes be attained, and that we had thus accurately predicted the native binding configuration for both ligands and their binding free energy.

Dynamic docking of a high-affinity antagonist to a membrane-embedded GPCR

One large difficulty has long been the enhanced sampling of membrane-embedded proteins such as GPCRs, due to their complicated design, which also includes a lipid bilayer, and due to the difficulty of sampling the configurational space of a binding site that is deeply buried inside the receptor. However, we have successfully applied our McMD-based dynamic docking method between a membrane-embedded β₂AR GPCR molecule and its high-affinity antagonist, alprenolol (Bekker et al. 2021a). To overcome the difficulty of enhanced sampling with a membrane, we employed a clever application of simple position restraints to only restrain the movement of a subset of membrane atoms along the axis perpendicular to the membrane, thus allowing the lipids to freely move within the membrane without rupturing it during the enhanced MD. The ligand was placed on the extracellular side of the GPCR in the bulk region and a cylinder (perpendicular to the membrane) was used to prevent the ligand from “teleporting” to the intracellular side through the bulk (due to the presence of periodic boundary conditions), and to prevent the ligand from getting trapped near the charged membrane surface. The GPCR was restrained using distance restraints to protect the secondary (backbone N–O restraints) and tertiary (Cα-Cα restraints) structure. First, a 68.16 µs pre-run was performed, followed by a 48 µs production run, which was then analyzed using our analysis protocol as described above (Fig. 1).

Analysis of the ensemble gave us 11 representative structures within a CFE of 1.5 kcal/mol (Fig. 4a), with the top-ranking one (r₁) matching that of the experimental structure with an R(native)-value of 0.999 and an RMSD of 0.72 Å (Fig. 4b). The subsequent canonical MD simulations corroborated these results, with r₁ also being the most stable one. Analysis of the binding pathway from the multicanonical ensemble enabled us to determine the molecular mechanism of ligand binding. We found that a gatekeeper residue Phe193, along with its neighboring residues Asp192 and Phe194 and their interacting residues, is central in guiding the ligand from the outer pocket (r₂₆₁, Fig. 4c) at the extracellular side of the receptor, into the inner pocket (r₂₆₄, Fig. 4d), where the ligand binds in its native configuration to the orthosteric site. As the ligand approaches the pocket entrance from the bulk region, it hits the gateway residue Phe193. The ligand then pushes Phe194 inside, which in turn pushes the gateway residue to the back of the outer pocket (Fig. 4c). Simultaneously, the positively charged nitrogen atom of the ligand starts interacting with Asp192, which in turn weakens Asp192’s charge-charge interaction with Lys305, leading their salt bridge to weaken and eventually break (Fig. 4c). This causes the pocket to widen and the loop that is formed by the gateway forming residues Asp192, Phe193, and Phe194, to destabilize, opening the path to the inner pocket, allowing the ligand to finally pass through, after which the gateway can close up again after the ligand has migrated to the inner pocket to bind to the orthosteric site (Fig. 4d).

Fig. 4.

Docking results obtained from our GPCR dynamic docking simulations, data available from BSM-00024. (a) Overview of our top 11 structures, with the ligand, alprenolol, structures colored based on their cluster free energy from blue to red, with the experimental structure in white. (b) Close-up of the orthosteric binding site, with the top-ranking structure (blue), the experimental structure (green), and the sidechains of nearby residues shown. (c) Structures r₁ and r₂₆₁ shown in black and magenta, respectively, where ligand of r₂₆₁ is positioned in the outer pocket and the receptor is in a closed state, similar to the native structure, except for the gateway loop, which is in a non-native state. (d) Structures r₁ and r₂₆₄ shown in black and magenta, respectively, where the ligand of r₂₆₄ has started to enter the inner binding site (as it passes through where Phe194 is positioned in the natively bound configuration), but the receptor’s gateway loop still in an open state

We also found that the second-ranking binding configuration might also contribute to the inhibitory effects of the ligand, as, although it does not enter the inner pocket, it binds to the GPCR in the outer pocket, before the gateway residue Phe193. It is likely that the ligand in this binding configuration can actually prevent the GPCR’s natural ligand epinephrine from passing through, thus providing an inhibitory effect. Since alprenolol would have to perform some acrobatics to get through the GPCR’s gateway, while to epinephrine, which is a molecule that is much smaller than alprenolol, and therefore would have much less trouble passing the gateway, the inhibitory effect of alprenolol in the second-ranking binding site might actually be considerable, as it would be able to inhibit the GPCR without having to enter the inner pocket. Thus, we found that alprenolol might not only function as an orthosteric inhibitor, but also as an allosteric one, where improvements to the chemical structure of the molecule could potentially be made to make it a better allosteric inhibitor, with improved affinity and specificity to the outer binding site.

Conclusion

Over the past decade, we have pushed the boundaries of the computational drug design field with our McMD-based dynamic docking approach. As part of Prof. Nakamura’s lab, we have developed the McMD-based dynamic docking approach, and since then, we have further refined it as shown in Fig. 1, and looked for new boundaries to surpass. We have shown that our methodology can not only be used to predict the correct binding configuration, but also to predict it in conjunction with it having to predict the correct binding site with our exhaustive simulations. Even for the difficult Bcl-xL docking, where its cryptic pocket differs greatly between the apo and holo states, we were able to accurately predict the binding configuration of the two large ligands. We also showed that our methodology can be used for membrane-embedded systems by clever usage of position restraints that prevent the membrane from rupturing, without significantly inhibiting membrane dynamics. One of the next stages in pushing the boundaries would be to apply our dynamic docking approach to protein–protein docking, with both molecules of similar size, unlike the protein-peptide dockings we have done thus far (Bekker et al. 2020b; Bekker and Kamiya 2021b).

One of the great benefits of McMD-based dynamic docking is that it not only accurately predicts the native binding configuration, but it can also be used to determine the binding/unbinding pathway, and thus provides insight into the molecular mechanism of binding. By understanding the molecular mechanism of binding, one can start to tune the compound by modifying its functional groups. Thus, our McMD-based dynamic docking approach would be a useful tool for lead optimization, where dynamic docking would first be performed using an initial version of the lead. Then, analysis of the molecular mechanism of binding can provide insight into how the compound binds to the native site, intermediary sites, and non-native sites, to improve the binding profile, e.g., by improving specificity to the native site or early intermediary binding sites to improve the association constant, with subsequent McMD-based dynamic docking simulations for validation and additional stages of optimization. With further increases of computational performance, as well as by tuning the computational parameters of the McMD-based dynamic docking simulation (simulation length, timestep, restraints, etc.), the McMD-based dynamic docking approach will become an attractive tool to model the binding mechanism not only in an academic environment, but also an industrial one, where McMD-based dynamic docking could be used to optimize a lead compound into a high-affinity and high-specificity drug.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

This work was supported by Japan Agency for Medical Research and Development (AMED) to N.K., and by the Grand-in-Aid for Scientific Research from the Japan Society for the Promotion of Science (JP20H03229). It was performed in part under the Cooperative Research Program of the Institute for Protein Research, Osaka University, CR-21–05 and CR-22–05. Computational resources of the TSUBAME3.0 system, Tokyo Institute of Technology, were provided by the HPCI Research Project (hp190018, hp190021, hp190027, hp200011, hp200025, hp200063, hp210002, hp210005, hp210048, hp220002, hp220015, and hp220026).

Data availability

The representative structures of our McMD-based dynamic docking simulations have been submitted to the Biological Structure Model Archive (https://bsma.pdbj.org) under BSMIDs BSM-00002, BSM-00007, BSM-00008, BSM-00010, BSM-00021, BSM-00024, and BSM-00029 (Bekker et al. 2020c).

The McMD-enhanced version of GROMACS is available from https://gitlab.com/gjbekker/gromacs, along with analysis tools.

Declarations

Ethics approval

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Conflict of interest

Not applicable.