Mechanistic Docking in Terpene Synthases using EnzyDock

Abstract

Terpene Synthases (TPS) catalyze the formation of multicyclic, complex terpenes and terpenoids from linear substrates. Molecular docking is an important research tool that can further our understanding of TPS multistep mechanisms and guide enzyme design. Standard docking programs are not well suited to tackle the unique challenges of TPS, like the many chemical steps which form multiple stereo-centers, the weak dispersion interactions between the isoprenoid chain and the hydrophobic region of the active site, description of carbocation intermediates, and finding mechanistically meaningful sets of docked poses. To address these and other unique challenges, we developed the multistate, multiscale docking program EnzyDock and used it to study many TPS and other enzymes. In this review we discuss the unique challenges of TPS, the special features of EnzyDock developed to address these challenges and demonstrate its successful use in ongoing research on the bacterial TPS CotB2.

1. INTRODUCTION

Enzymes play crucial roles in nature by enhancing the rate of chemical reactions and catalyze processes ranging from simple one-step reactions to multistep reaction cascades (Radzicka & Wolfenden, 1995; Arieh Warshel et al., 2006; Zhang & Houk, 2005). Natural product biosynthesis often entails complex synthetic processes resulting in intricate molecules that are rich in rings, stereocenters, and functional moieties (F. Zhou & Pichersky, 2020). Terpene synthases (TPS) and their downstream functionalizing enzymes produce the largest family of known natural products with over 80,000 terpenes and terpenoids identified to date (David W Christianson, 2017), and are found in mammals, plants, and microorganisms (Dickschat, 2016; Tholl, Rebholz, Morozov, & O’Maille, 2023; Weng, Philippe, & Noel, 2012). Terpenes and terpenoids possess rich aromas and flavors and have anti-microbial, anti-inflammatory, and many other medicinal properties (Cox-Georgian, Ramadoss, Dona, & Basu, 2019; Downer, 2020), and play important roles in defense mechanisms, chemical communication, and phytohormone regulation (Gershenzon & Dudareva, 2007; Rosenkranz, Chen, Zhu, & Vlot, 2021; Singh & Sharma, 2015). Due to their multifaceted roles, terpenes and terpenoids are widely used in a variety of industries including food, agriculture, pharmaceutical, cosmetic, textiles, detergent, and biofuels (Beller, Lee, & Katz, 2015; Cox-Georgian et al., 2019; Masyita et al., 2022; Ninkuu et al., 2021; Tetali, 2019; Weston-Green, Clunas, & Jimenez Naranjo, 2021).

The natural substrates of TPS are amphiphilic pyrophosphate esters, with a hydrophobic tail composed of linearly coupled isoprenoid units bound to a hydrophilic pyrophosphate (PP) head group. The TPS substrates are classified according to the number of isoprenoid units (n), e.g., geranyl diphosphate (GPP, n=2), farnesyl diphosphate (FPP, n=3), and geranylgeranyl diphosphate (GGPP, n=4), which yield mono-, sesqui-, and diterpene products, respectively (see Fig. 1A). TPS may be divided into two sub-families based on their mechanism. Class I terpene cyclases employ a trinuclear metal cluster to initiate the ionization of the isoprenoid diphosphate substrate to yield an allylic cation and inorganic pyrophosphate, while class II terpenoid cyclases rely on a general acid, like Asp, to protonate the terminal carbon−carbon double bond of the substrate alkyl chain (Wendt & Schulz, 1998). In class I TPS, the enzyme catalyzes the reaction with the help of two cofactors, inorganic pyrophosphate (PP), which originates from the substrate after diphosphate abstraction, and 3 divalent ions (e.g., Mg⁺²). Based on the known crystal structures of TPS, the enzyme domain architecture may be defined as $\alpha ,\alpha \beta$, or $\alpha \beta \gamma$ (David W Christianson, 2017). In recent years, the number of available X-Ray structures has increased significantly, and this data indicate that TPS undergo pronounced structural changes upon substrate binding while transitioning from the apo to the holo state. The active site in class I TPS is located in the middle of the $\alpha$-helical bundle domain (Fig. 1B). This catalytic pocket is biphasic, mirroring that of the amphiphilic nature of the substrate (Fig. 1C). One region contains highly polar and charged side chains which stabilize the charged PP-(Mg²⁺)₃ cluster and is well hydrated. This polar region is flanked by a binding pocket rich in hydrophobic and aromatic residues that host the isoprenoid moiety. Upon heterolytic cleavage of the bond between the pyrophosphate and isoprenoid groups, the pyrophosphate moiety remains tightly bound while the isoprenoid allyl cation is unleashed in the hydrophobic pocket. The carbocation then initiates a cascade of reactions involving cyclizations, methyl and methylene migrations, proton and hydride transfers, deprotonations and reprotonations, and reaction quenching by deprotonation or hydroxylation (D. E. Cane, 1999). Since these intermediates are highly reactive (i.e., intrinsic reactivity of carbocations (Dean J Tantillo, 2017)) and carbocation reactions are fast, it has been suggested that the role of the enzyme is to guide the reaction towards a specific product among many possible, rather than catalyzing it (Raz, Levi, Gupta, & Major, 2020).

Figure 1.

Common substrates of TPS and the structure of a class I TPS. (A) Substrates of mono-, sesqui-, and di-TPS, geranyl diphosphate (GPP, n=2), farnesyl diphosphate (FPP, n=3), and geranylgeranyl diphosphate (GGPP, n=4). (B) Structure of bornyl diphosphate synthase (BPPS) showing the α-helical bundle. (C) Zoom in on the active site shows the biphasic active site of class I TPS. Polar residues are shown in blue, hydrophobic residues are shown in yellow, carbocation carbons are colored green, and magnesium ions are colored purple.

Detailed studies of the chemistry of terpene biosynthesis coupled with structural biology (Allemann, 2008; D.E. Cane, 1990; D. E. Cane, 1999; D. W. Christianson, 2006; Croteau, 1987; Dickschat, 2016; Frick et al., 2013; Greenhagen, O’Maille, Noel, & Chappell, 2006; Wilderman & Peters, 2007; Xu, Wilderman, & Peters, 2007; K. Zhou & Peters, 2011), have allowed unprecedented insight into the workings of TPS. Additionally, quantum chemistry calculations of the gas-phase carbocation chemistry of terpenes has provided unparalleled insight into the plethora of possible reaction mechanisms (Das, Dixit, & Major, 2016; Hess, Smentek, Noel, & O’Maille, 2011; McCulley, Geier, Hudson, Gagne, & Tantillo, 2017; D. J. Tantillo, 2011; Dean J Tantillo, 2017; Weitman & Major, 2010). Full enzyme simulations have further provided extremely valuable insight into the catalytic strategies of TPS (Allemann, Young, Ma, Truhlar, & Gao, 2007; Rajamani & Gao, 2003). Our group has studied numerous TPS, with a focus on the catalytic control exerted by these enzymes (Das et al., 2016; Dixit, Weitman, Gao, & Major, 2017, 2018; Freud, Ansbacher, & Major, 2017; Dan T Major, 2017; Dan Thomas Major, Freud, & Weitman, 2014; D. T. Major & Weitman, 2012; Weitman & Major, 2010). The proposed roles of TPS in guiding and controlling chemistry are many (Raz, Levi, et al., 2020): Active site contour that correctly folds the substrate and sequester carbocation intermediates; (Deligeorgopoulou & Allemann, 2003; L. Sangeetha Vedula, Jiang, Zakharian, Cane, & Christianson, 2008) PP-(Mg²⁺)₃-enzyme cluster which activates the C-O bond; bifacial active site nature, which enables electrostatic guidance and control over the reactions (Das et al., 2016; Dixit et al., 2017, 2018; Freud et al., 2017; Dan T Major, 2017; Dan Thomas Major et al., 2014; D. T. Major & Weitman, 2012; Weitman & Major, 2010); specific position of active site water molecules or acids/bases that allow specific (de)protonation, hydroxylation, or epoxidation (David W Christianson, 2017). Many studies of TPS have shown that slight changes in the identity of the active site residues can divert the reaction trajectory towards alternative reaction channels (Greenhagen et al., 2006; Li et al., 2014; Ludwig et al., 2024; O’Maille et al., 2008; Whitehead, Leferink, Johannissen, Hay, & Scrutton, 2023) or may quench the reaction to give rise to different products (Raz, Driller, et al., 2020), which underscores the significant promiscuity in many TPS (Dixit et al., 2017; Jia, Potter, & Peters, 2016; D. T. Major & Weitman, 2012).

Although theoretical tools to study enzymes have existed for nearly half a century (A. Warshel & Levitt, 1976), methods that are capable of rapidly modeling complex reaction cascades, like those taking place in TPS, are still limited. Here we will describe a docking tool named EnzyDock, which was designed with enzymes in mind and especially enzymes catalyzing complex reactions.

2. CHALLENGES IN USING STANDARD DOCKING TOOLS FOR TERPENE SYNTHASES

Protein-ligand docking predicts in silico the preferred orientation of a ligand in a protein binding pocket. The docking process involves two main tasks, namely searching for favorable binding poses and scoring these poses. The search process can use sampling methods such as Molecular Dynamics (MD), Monte-Carlo (MC), Simulated Annealing (SA), or genetic algorithms. Scoring functions may employ energy based scoring functions like molecular mechanics (MM) (also called force fields, FF), e.g., CHARMM (J. Huang & MacKerell, 2013; Jing Huang et al., 2017; MacKerell et al., 1998; Vanommeslaeghe et al., 2010), AMBER (Cornell et al., 1995), and OPLS (William L. Jorgensen, Maxwell, & Tirado-Rives, 1996; W. L. Jorgensen & Tirado-Rives, 1988), or information based scoring functions, like DRUG-SCORE (Velec, Gohlke, & Klebe, 2005), IT-SCORE (Grinter, Yan, Huang, Jiang, & Zou, 2013), DSX (Neudert & Klebe, 2011), CHEMSCORE (Verdonk, Cole, Hartshorn, Murray, & Taylor, 2003), and SMoG (Ishchenko & Shakhnovich, 2002). Over the last several decades, more than 60 different docking programs have been developed, such as DOCK (Brooijmans & Kuntz, 2003; Jones & Willett, 1995; Kuntz, Blaney, Oatley, Langridge, & Ferrin, 1982; Shoichet & Kuntz, 1991), AUTODOCK (Goodsell & Olson, 1990; Morris, Goodsell, Huey, & Olson, 1996; Morris et al., 2009), GLIDE (Friesner et al., 2004; Friesner et al., 2006; Halgren et al., 2004), ROSETTALIGAND (Davis & Baker, 2009; Leaver-Fay et al., 2011; Meiler & Baker, 2006), GOLD (Verdonk et al., 2005; Willett, Glen, Leach, Taylor, & Jones, 1997), and CDOCKER (Ding, Hayes, Vilseck, Charles, & Brooks, 2018; Gagnon, Law, & Brooks, 2016; Pearce, Langley, Kang, Huang, & Kulkarni, 2009; Vieth, Hirst, Dominy, Daigler, & Brooks, 1998; Vieth, Hirst, Kolinski, & Brooks, 1998; Wu, Robertson, Brooks, & Vieth, 2003). In our group we developed EnzyDock (Das et al., 2019) using the CHARMM program (B. R. Brooks et al., 2009; Bernard R. Brooks et al., 1983). EnzyDock is a protein-ligand docking program with special emphasis on modeling enzyme reactions, ranging from simple one-step reactions to complex, multi-step reactions. Related docking tools targeting multistep reactions in TPS have been reported (O’Brien, Bertolani, Tantillo, & Siegel, 2016; O’Brien, Bertolani, Zhang, Siegel, & Tantillo, 2018; Tian et al., 2014).

Considering the large number of docking programs available, one might wonder why there is a need for specialized docking approaches for TPS (and many other enzyme families)? To better understand why specialized docking tools are required when modeling TPS it is important to consider the challenges when docking ligands in TPS: (1) Although the binding pose of the PP moiety is well defined and highly conserved among all TPS, there is great variability in the folding of the isoprenoid moiety in the hydrophobic binding cavity. This is due to the difference in hydrophobic binding site contour, flexibility of the alkyl chain, and the weak dispersion interactions governing the binding. Predicting the correctly folded state of the isoprenoid moiety is critical to determine the correct reaction trajectory, yet this is difficult with standard docking tools. (2) Once the initial allyl carbocation is formed, subsequent cyclization reactions result in a reduction in the volume of the carbocation intermediates, reducing the steric match between protein and ligand and spurring tumbling of the intermediates. In fact, there is considerable experimental evidence from crystal structures indicating that the global free energy minimum structures of intermediates do not necessarily correspond to the state observed during reactions (L. Sangeetha Vedula et al., 2005; Whittington et al., 2002). Hence, standard docking tools might struggle to reliably predict such states. (3) The carbocation intermediates can be stabilized by cation interactions with e.g. $\pi$-systems, Met sulfur atoms, carbonyl groups, and anions. Since the position of the cation along the carbon frame of the carbocation changes throughout the reaction, these interactions do not necessarily provide direct clues regarding the correct binding mode. Additionally, these are not interactions that anchor the carbocation intermediate in well-defined binding modes, in contrast to hydrogen bonds, which are directional. (4) Although some ligand tumbling likely occurs during TPS reactions, due to the fast chemistry of carbocations and the “product-like” contour of binding sites, all bound poses along a reaction cascade are most likely quite similar (consensus binding). Standard docking tools cannot readily enforce such similar binding modes. (5) All bound states must match the correct final stereo configuration of the product. Considering the many possible stereochemical centers in terpenes and terpenoids, this puts additional restrictions on docking. Enforcing correct chirality, pro-chirality, and regiochemistry is not possible with most standard docking approaches. (6) If knowledge of the deprotonating agent or a possible hydroxylating water is known, docking of the ligands should allow this knowledge to be included. This can be included via restraints in some docking codes. (7) Chemical changes, like bonding to PP or varying protonation states that occur during the reactions are not easily accounted for with standard docking programs.

In our group, we have used EnzyDock to study numerous TPS and other enzymes (Gupta, Tarannam, Zev, & Major, 2023; Dan Thomas Major, Gupta, & Gao, 2023; Raz, Driller, et al., 2020; Zev et al., 2021). Below we will illustrate how one can use EnzyDock to address the challenges mentioned above.

3. MULTISTATE MULTISCALE DOCKING WITH ENZYDOCK

3.1. COTB2 – A TOY SYSTEM

To illustrate the unique capabilities of EnzyDock, we will explain the idea of mechanistic docking of the reaction states (substrate, intermediates, and product) in the bacterial Class I TPS CotB2 (Driller et al., 2018; Janke, Görner, Hirte, Brück, & Loll, 2014; Raz, Driller, et al., 2020). CotB2 catalyzes the formation of the tri-cyclic compound cyclooctat-9-en-7-ol from the acyclic substrate GGPP (see Fig. 2A). This reaction entails 12 distinct chemical steps, including cyclizations, hydride and methylene transfers, nucleophilic attack, and deprotonation. The product cyclooctat-9-en-7-ol includes 6 chiral centers and the reaction in CotB2 is highly specific (product > ~90%) (Driller et al., 2019). CotB2 has served as a useful toy system as the reaction chemistry is well characterized (Hong & Tantillo, 2015; Meguro et al., 2015), and a good-quality, biologically relevant crystal structure with a quenched intermediate bound in the active site (Driller et al., 2018). Although EnzyDock does not require a bound ligand in the active site, it can help guide the docking algorithm identify the correct poses. We note that EnzyDock can be used with homology models or deep learning models (e.g., AlphaFold), but some care should be exercised as these models do not reach the quality of experimental structures (Terwilliger et al., 2024). Additionally, a closed, holo-state of the enzyme should ideally be used. In these regards, the CotB2 structure (PDB ID 6GGI) is a near perfect structure.

Figure 2.

Mechanistic docking using EnzyDock. (A) Proposed mechanism of cyclooctat-9-en-7-ol from the linear substrate GGPP. (B) Illustration of multistate multiscale docking in CotB2.

3.2. THE MULTISTATE CONCEPT

The overarching goal of multistate docking is to generate accurate binding poses of all important reaction states (e.g., substrate, intermediates, and product) in a consistent and biochemically meaningful manner. In the case of TPS, this requires that the substrate isoprenoid chain is folded in a product-like manner; at each step of the reaction the pro-chiral approach of each electrophile-nucleophile pair must match the chirality of the following intermediate (i.e., re vs. si face attack); hydroxylation sites must be positioned near plausible water binding pockets with correct pro-chirality; deprotonation/protonation sites must be located in the vicinity of a base/acid; all poses along the reaction pathway are geometrically similar to a certain degree (consensus poses); the isoprenoid carbon of the initial C-O bond cleavage is likely somewhat close to its oxygen bond-partner throughout the reaction. Ideally, the carbocations generated during the reaction are also stabilized throughout the reaction by active site amino acids or the PP cofactor. Hence, for example in the case of CotB2 (Fig. 2A), GGPP must be folded so that the initial cation at C1 approaches C11 in a si-face manner, while C10 approaches C14 in a re-face manner to generate carbocation A. Further, the double bonds in B must face one another with regiochemistry that will generate the correct stereochemistry for two new chiral centers in intermediate C. As the reaction proceeds towards the final product the intermediates become increasingly better folded and fewer stereochemical challenges remain for the docking program. However, as the intermediates become increasingly more product-like and dense with a smaller effective volume, they tend to tumble more in the active site and present a challenge for docking.

In EnzyDock, we chose to tackle this challenge by adopting a multistep docking approach (Das et al., 2019; O’Brien et al., 2016; O’Brien et al., 2018; Tian et al., 2014). This means that multiple steps are docked sequentially, in a predetermined order, while transmitting information from one ligand docking step to the next. For instance, in the case of CotB2, we docked all states, starting with GGPP and concluding with intermediate I (i.e., the range GGPP:I). This allows validating that the sets of poses of docked states in the range GGPP:I agree with the above mentioned requirements. Inspection of the sets of docked states (e.g., GGPP:I) allows researchers to judge the possible involvement of active site elements in substrate folding, stabilization of carbocations, deprotonation/protonation. Such insights may be helpful for generating mechanistic understanding and provide guidance for site-directed mutational experiments and enzyme design. Additionally, once designs have been proposed, these can be screened prior to experiments by applying EnzyDock to variants and check whether the new enzyme designs are likely to improve the reaction of interest or not.

3.2.1. ENZYDOCK WITHOUT RESTRAINTS

When docking multiple states with EnzyDock, for each of the N states the docking algorithm produces many plausible poses (typically 10²-10³). Therefore, after an EnzyDock run, one has a very large number of poses to analyze (N·10²-10³) before one can generate biochemical insights. For instance, in the case of CotB2, one would like to analyze sets of consensus binding modes that may constitute a likely in-enzyme reaction path for GGPP:I. However, to identify the sets of all matching poses GGPP:I is a challenge (Fig. 2). To aid in this search for consensus poses, EnzyDock has a path-finder tool which searches through all bound poses and identifies geometrically matching poses (Raz, Driller, et al., 2020). Once all such sets of consensus poses have been obtained, these may be visually analyzed using tools like PyMol. However, in many cases it is challenging for EnzyDock to identify sets of consensus paths including all states without any additional help. The reason for this is not necessarily a limitation of the docking methods, but rather an inherent difficulty in enzyme catalysis and in particular in TPS related to the fact that the global free energy minimum structures of intermediates do not necessarily correspond to the state observed during reactions. This will be discussed further below. In cases where it is difficult to identify all states using pathfinder, restraints may be applied. Such restraints, which are often based on the chemical intuition of the modeler, are discussed in the following section.

3.2.2. ENZYDOCK WITH RESTRAINTS

In practice, the user must determine a docking seed, which is the first state docked, and this state may be used as a template for all subsequent states via restraints. For instance, in the case of CotB2, the crystal structure was resolved with a bound ligand which is similar to intermediate B. Intermediate B already has 3 out of 6 chiral centers formed, but its effective volume is still intermediate. Hence, B could be a suitable docking seed for a multistate docking simulation in CotB2 because it is most likely to dock correctly. Hence, carbocation B would be the first in a series of states docked in a single EnzyDock run. To ensure consensus docking (i.e., similar poses), EnzyDock allows application of Nuclear Overhauser Effect (NOE) restraints of all subsequent states relative to the seed state. Presumably, the seed state will be docked in a multitude of poses. For each pose generated for the seed, the poses of subsequent states will be biased towards the pose of the seed via NOE restraints. In such a manner, a series of consensus poses conforming with possible hypothetical reaction pathways are generated and may be inspected using the docked energy score and biochemical intuition to guide the user.

To check that EnzyDock generates biochemically meaningful poses, it is prudent to first perform a docking experiment with the seed alone to make sure that it identifies a reasonable binding pose. In a case where the seed (e.g., B) does not dock correctly, one may apply a variety of restraints that guide the docking towards an enzymatically meaningful pose. For instance, if the correct pro-chiral binding pose is difficult to obtain, one may apply, for example, dihedral restraints during docking to increase the likelihood of obtaining correct prochirality. In the case of intermediate B, the double bonds must face one another with correct regiochemistry to form the correct stereochemistry for two new chiral centers in intermediate C, and this can be achieved via dihedral restraints. However, if one is unable to obtain the correct orientation of the seed in the active site pocket, i.e., the C1 position is located deep inside the active site pocket far from the PP cofactor which it was initially bound to, one might need additional restraints. EnzyDock allows addition of NOE restraints between any ligand atom with any other atom in the system (e.g., between C1 and a PP oxygen) or a point in space. This can help the docking search algorithm identify biologically meaningful poses. We note that such restraints can be applied to any docked state (e.g., the states GGPP:I in CotB2).

An important question is of course why such restraints are needed. Shouldn’t the docking algorithm be able to identify the correct pose? First, docking algorithms might not be able to identify certain poses due to a need to cross high energy barriers to reach a given state. Second, the docking scoring function is only an approximate, empirical energy function and hence might not always accurately rank the correct pose as the lowest energy pose. However, in the case of TPS (and likely other enzymes as well), intermediate states are often bound in modes that are clearly not the global minimum. For instance, in the case of bornyl diphosphate synthase (BPPS), inspection of several crystal structures with bound substrate and intermediate mimics, as well as the final product, revealed that a key intermediate was clearly bound in an off-path pose (Whittington et al., 2002). Similar conclusions were drawn based on crystal structures of trichodiene synthase (L Sangeetha Vedula, Cane, & Christianson, 2005; L. Sangeetha Vedula et al., 2005). Hence, even a perfect docking method would not predict the biologically meaningful and correct binding mode for all intermediates.

An important point which arises in multistate docking with EnzyDock is the differing covalent bonding appearing in different states along a reaction path. All TPS have a covalent bond between the cofactor and the isoprenoid moiety in the substrate state, but mostly not in other bound states. In EnzyDock such changing bonding situations are dealt with via so-called residue patching, or bond-stitching, on the fly during docking. The patching is applied by generating a harmonic C-O bond and all associated valence terms using a standard force field. Hence, in the case of CotB2, during docking of GGPP a bond is formed between the PP cofactor and the allyl cation, whereas all other states are docked without such a bond. In the case of BPPS, where both the initial substrate GPP and the final product BPP, are bonded to the PP cofactor, such patches are applied to these two states during the docking. The use of patches is preferred over docking the whole ligand with the covalently bound PP, as patching avoids the need to dock the PP moiety and instead allows the PP moiety to reside in the experimentally resolved position.

We note that template-based docking, like docking using a crystal structure with a bound ligand as the seed for subsequent docked states, is also possible in recent versions of EnzyDock.

3.3. THE MULTISCALE CONCEPT

A key aspect of EnzyDock is that it can dock multiple states, including intermediate and transition states. This raises problems not faced in most traditional docking approaches regarding the scoring of the docked ligands. Intermediates, like carbocations, are fleeting structures which often have unusual geometries, like non-classical carbocations, and these are not well treated using simple scoring functions. For instance, most force fields were not parametrized for intermediate structures. Furthermore, transition states are even more difficult to treat, yet most enzymes have evolved to stabilize the charge distribution developed at the transition states. To be able to treat such charge distributions it is necessary to employ quantum chemistry methods. In large systems, like enzymes, it is usually not possible to treat the entire system using quantum chemistry due to the computational cost and instead so-called multi-scale methods are employed. QM/MM methods (quantum mechanics-molecular mechanics)(Dixit, Das, Mhashal, Eitan, & Major, 2016; van der Kamp & Mulholland, 2013; A. Warshel & Levitt, 1976), treat the ligands and possible parts of the active site residues and cofactors as QM while the remaining protein, solvent, and ions are treated using MM (Fig. 3). However, even QM/MM calculations are too expensive for the kind of screening inherent in docking, and therefore in EnzyDock a 3-level funneled approach is adopted, wherein the computational cost increases as one moves down the funnel (Fig. 3). In this approach the first level employs docking scoring on a 3D grid, the second level an all-atom force field (FF) scoring, while the third level is carried out using QM/MM.

Figure 3.

EnzyDock’s 3-level funneled approach.

Level 1: The first and main part of the docking process of all the ligands is performed using a force field (FF) approach wherein all atoms in the system are treated as classical particles. The enzyme is described by a 3D grid, as is commonly done in docking, while flexible residues, cofactors, and the ligands are treated explicitly within the grid. It is during this stage that the main conformational search is performed (described further below), and the consensus docking takes place and restraints may be applied. In TPS, cofactors like PP, Mg²⁺ ions, and water molecules are treated explicitly and are restrained to their original position to avoid perturbation of their structure during docking. Potential catalytic waters can be placed to assess plausible mechanisms involving water. At the end of the main docking stage performed at level 1, EnzyDock performs geometric clustering of the docked ligand poses and then passes the lowest energy pose from each cluster on to the remaining levels.

Level 2: In this stage, EnzyDock moves from a grid description to an all-atom, off-grid, description. During this stage, only local minimization is performed to attain a minimum, all-atom structure; implicit solvation (e.g., Generalized Born solvation) can be included at this stage. This stage is essential in preparation for the next level.

Level 3: In the final scoring stage, EnzyDock uses QM/MM to minimize the structure of the ligand and possibly also the nearby environment. The QM region includes the ligand atoms only, while everything else is treated using MM (i.e., a FF). A range of QM methods can be used as is detailed below. In the case of TPS, only the carbocations are treated as QM, and hence the substrate is described as PP-allyl cation ion-pair. The effects of the surrounding MM region (protein, cofactors, water, ions) on the QM region (ligand) are accounted for using so-called electrostatic embedding, i.e., the classical MM charges are allowed to polarize the QM region. For the CotB2 system, several works used hybrid QM/MM potentials, and the lowest energy poses match the available crystal structure very well (Driller et al., 2018; Raz, Driller, et al., 2020).

3.4. ENZYDOCK WORKFLOW

EnzyDock’s workflow commences with a pre-docking conformational search of all the ligands in vacuum (Fig. 4). For each ligand, a torsional clustering algorithm (Schulz-Gasch, Schärfer, Guba, & Rarey, 2012) is applied to group the many possible ligand conformers into clusters and the lowest energy conformer from each cluster is used for docking, thereby reducing unnecessary redundancy in the docking process. Following this process, the docking begins by placing the ligand seed randomly into the active site and Monte Carlo (MC) simulations are performed. The first MC stage entails rigid translations and rotations of the ligand in the active site, which allows initial rough placement of the ligand in favorable poses. Following this stage, a series of MC and optional molecular dynamics (MD) simulated annealing (SA) is performed (SAMC and SAMD, respectively). During SAMC, ligand torsional MC moves are included in addition to rigid translations and rotations of the ligand. During this stage, flexible active site sidechains can rotate and explicit water molecules undergo rigid translations and rotations. After the SA stage, a final minimization is performed. Following complete docking of the seed, the remaining ligands are docked. At this stage consensus docking restraints relative to the seed pose can be applied. During the entire docking process of the seed and subsequent ligands, restraints can be applied as discussed in 3.2.2 above to implement chemical knowledge and intuition.

Figure 4.

Workflow of the EnzyDock program.

During the above sampling of the ligand states (MC, SAMC, SAMD), a soft van der Waals (vdW) potential is employed, which reduces the repulsive interactions at close distances, allowing for more rapid crossing of energy barriers. During final on-grid minimization, the standard vdW potential is switched on.

Following the grid-based docking, rescoring is performed in an all-atom environment, commencing with standard MM, followed by MM in an implicit solvent environment (optionally), and finally using QM/MM.

Users can define the grid size and center, grid resolution, the temperature range for SA, number of sampling steps, the type of the solvation (GBSW, PBEQ, or none), and whether and how to perform QM/MM minimization. MM scoring employs the standard CHARMM c36 FF (Jing Huang et al., 2017), TIP3P for water (William L Jorgensen, Chandrasekhar, Madura, Impey, & Klein, 1983), and all ligands use the CGenFF FF (Vanommeslaeghe et al., 2010). The QM part can use a range of methods, ranging from semi-empirical methods (MNDO, AM1, PM3)(Ojeda-May & Nam, 2017) to density functional theory (e.g., M062X) or specialized fast quantum chemistry methods (HF3c) (Shao et al., 2015).

3.5. ENZYDOCK INPUT

The input files needed for EnzyDock runs are PDB files for the protein and cofactors (e.g., PP, Mg²⁺ ions, and waters for TPS), and FF parameters for the cofactors (if not part of the regular CHARMM topology and parameter files). If one prefers, CHARMM CRD and PSF files for the protein (e.g., from the web-server CHARMM-GUI, see more details below) may also be used (B. R. Brooks et al., 2009; Jo, Kim, Iyer, & Im, 2008; Lee et al., 2016). We note that the protein must be complete, and preferably in a closed (holo) state without breaks in the chains, although missing side chains are allowed as these will be completed by EnzyDock. The ligands can be provided as a CSV file listing the SMILES of the ligands (in this case, ligand parameters will be generated on the fly during EnzyDock’s run using CGenFF) or PDB and matching CHARMM topology and parameter files. However, when docking cations, the CHARMM topology and parameter files must be provided by the user as these cannot be generated by CGenFF.

Restraints may be supplied as a CSV file with columns listing the pairs of atoms to be restrained at a certain distance, e.g., H-bonded atoms or ionic interactions. Similarly, consensus restraints may be included as a CSV file. Restraining atoms to distances within specified points in space can also be included. Dihedral restraints must be included in an appropriate restraints file.

Currently, a module named EnzyDocker is being implemented in the CHARMM-GUI web-platform (Jo et al., 2008) to enable an easy workflow starting from a PDB file (e.g., directly from the RCSB web site) all the way to the docking results.

3.6. PRACTICAL ASPECTS OF ENZYDOCK

EnzyDock relies on the CHARMM program, and therefore before you start, you must have an executable of a recent version of CHARMM (https://academiccharmm.org/). CHARMM typically runs on Linux machines. Additionally, CGenFF (https://silcsbio.com/) is required to generate force field parameters for the ligands and in some cases cofactor. CHARMM and CGenFF are freely available for academic use. EnzyDock also relies on Python and therefore the location of a Python interpreter must be set in the script scripts/python_wrapper.sh via the PYTHON4ENZYDOCK shell variable, and the following Python libraries must be installed in the Python environment defined in python_wrapper.sh:

1. NumPy

2. Pandas

3. OpenBabel

4. RDKit

These libraries can conveniently be installed using anaconda or miniconda.

The file hierarchy of EnzyDock is as follows:

Input files describing the enzyme system:

EnzyDock requires the three-dimensional structure of the protein, one or more ligands to dock, and possible cofactors and water molecules.

• Protein. PDB file or CRD coordinate and PSF topology files.

• Ligands. Can be provided as SMILES strings or PDB and CHARMM topology/parameter files.

• Cofactors. PDB and CHARMM topology/parameter files.

• Waters. PDB files.

These files can be prepared manually or for example using the web-interface CHARMM-GUI (https://www.charmm-gui.org/). All PDB files must be CHARMM-compatible. The protein should be complete and preferably in a closed state without any missing loop regions, and protonation states must be determined in advance (e.g., His tautomers).

In the case of class I TPS, the protein could be the α-helical bundle domain that includes the active site (i.e., in cases where other domains are missing, docking results should not be significantly influenced). The cofactors PP and (Mg²⁺)₃ should be provided as separate PDB files, while all water molecules are grouped together in a single PDB file. Although EnzyDock can technically run without water molecules, in the case of TPS it is recommend to include the water molecules stabilizing the charged cofactors and any structural water molecules. The ligand SMILES are provided as a CSV file. The naming conventions for all the files should follow that described in the documentation (https://github.com/majordt/EnzyDock).

EnzyDock environment control file (param.str):

• Set the ‘proj’ variable to the name of the project directory.

• Set the ‘DIR0’ variable to the absolute path to the directory that contains the ‘proj’ directory.

• Set the ‘scr1DIR’ to the path of a scratch directory (if not known, write ‘@scrDIR’).

• Set the ‘cgenff’ variable to the path to the CGenFF executable.

Excerpt from param.str:

set proj cotb2 ! Project name
set DIR0 “~/charmm/workspace/dock” ! If mixed case path or file names, must be in quotation marks
…
set scrDIR @DIR/scr
set scr1DIR path_to_scratch_disk/enzydock/scr ! ‘path_to_scratch_disk’ is the absolute path
set cgenff path_to_cgenff_executable ! ‘path_to_cgenff_executable’ is the absolute path

EnzyDock run control file (userparam.str):

The following variables must be set:

• ‘proteinname’

• ‘numligands’ (e.g., 3 if you have substrate, product and one intermediate)

• ‘numcofactor’ (e.g., 2 if you have PP and (Mg²⁺)₃). In this case, you must set the names of the cofactors via the variables ‘cofact1’ and ‘cofact2’.

Excerpt from userparam.str:

set proteinname 6ggi
set numligands 1
set numcofactor 2
set cofact1 pop
set cofact2 mg

The center of the binding site must also be set. In the case of TPS this would be the center of the hydrophobic pocket, which can be defined manually or using tools like DoGSiteScorer. Errors of few Å in the definitions of the center are negligible.

‘bsitex’, ‘bsitey’, ‘bsitez’

The size of the grid:

‘gridsize’

Excerpt from userparam.str:

! Center of mass (bound ligand): [−23.472, −17.123, 12.969]
set bsitex −23.472
set bsitey −17.123
set bsitez 12.969
set gridsize 30.0

Important run parameters that control the docking (for full details and advanced options, see https://github.com/majordt/EnzyDock):

• ‘ligmcsteps’ (The number of MC steps used to sample ligand conformations)

• ‘numheatsteps’ (The number of SAMD heating steps)

• ‘heattemp’ (The high temperature of the SAMD)

• ‘numcoolsteps’ (The cooling temperature of the SAMD)

• ‘cooltemp’ (The low temperature of the SAMD)

• ‘maxmit’ (The number of micro-iterations per ligand)

Excerpt from userparam.str:

set ligmcsteps 10000
set numheatsteps 5000
set heattemp 600.0
set numcoolsteps 10000
set cooltemp 100.0
set maxmit 25

EnzyDock PDB input file names:

EnzyDock expects PDB files names that match ‘proteinname’ and ‘cofact1’, ‘cofact2’, …

Example file names for proteinname=‘6ggi’, cofact1=‘pop’, cofact2=‘mg’ are:

pdb/6ggi_1.pdb (for protein chain 1)
pdb/6ggi_pop.pdb (for diphosphate)
pdb/6ggi_mg.pdb (for all mg ions)
pdb/6ggi_wat.pdb (for all water molecules)
pdb/ligand_1.pdb (for ligand 1, for additional ligands number sequentially)

FF topology and parameter file should be named:

 local_top/6ggi_pop.str (for diphosphate)

The ligand PDB and topology/parameter files are only required if the ligand is not supplied as a SMILES string. In these files the ligand should have the residue name ‘LIG’. Ligands that are covalently attached to cofactors (e.g., TPS substrates) require additional patch files, and for GPP, FPP, and GGPP these are included in EnzyDock.

EnzyDock restraint file (stream/consdef/user_consensus.csv):

User-defined restraints can be supplied via a CSV file named ‘user_consensus.csv’, which is read by EnzyDock. The most recent versions of EnzyDock automatically identifies matching atoms for consensus restraints (unpublished results), while additional restraints, like distance restraints between the ligands and the active site, can be added manually via the ‘user_consensus.csv’ file. The strength of the restraint is defined by setting the parameter ‘harmforce’ in the userparam.str file mentioned above. To turn on the consensus restraints the user must set a few variables as described in the documentation.

Running EnzyDock:

When everything is set, run EnzyDock from the dock directory as follows:

charmmexecutable < enzydock_main.inp > enzydock.out

where ‘charmmexecutable’ is the name of the CHARMM executable.

See results in the ‘results’ directory and follow EnzyDock’s progress in the ‘enzydock.log’ file. Analysis of results can be performed using visualization tools like PyMol. For more details, see next section.

3.7. ENZYDOCK OUTPUT AND ANALYSIS

The results of an EnzyDock run are multiple structures of the docked ligand(s) and their associated energies (PDB and log files), and geometric root mean square deviation (RMSD) lists are also provided. The RMSD is calculated with respect to the initial structure provided by the user. Useful analyses may include comparison to a co-crystalized ligand, e.g., 2-fluoro-3,7,18-dolabellatriene in the CotB2 crystal structure (PDB ID 6GGI)(Driller et al., 2018). To find a consensus set of docked poses, one can use the “pathfinder” analysis (Fig. 4) (Raz, Driller, et al., 2020). Typically, visual inspection is helpful in examining the role of the enzyme environment in folding the substrate or stabilizing the many intermediates occurring during a TPS reaction. For instance, one might look for particular π-cation, sulfur-cation, dipole-cation, or anion-cation interactions. System-specific analyses, like specific intermolecular distances between key atoms and pro-chirality of early steps in the reaction can readily be assessed from the docked structures. Since one often gets many possible reaction paths using EnzyDock, our group employs automated Python scripts that generate PyMol inputs with preset visual effects.

3.8. ENZYDOCK EXAMPLE

In CotB2, we performed simultaneous docking simulations of key reaction coordinate states (GGPP, A, B, C, E, G, H, and I) for WT and several variants: N103A, F107A, F107L, F107Y, F149L, F185 A, W186H, W186L, W186F, W288F, and W288G (Raz, Driller, et al., 2020). In these docking simulations, no mechanistic restraints were applied, and to generate consensus poses pathfinder was applied. High-level docking scoring was performed using QM(HF3c)/MM. In the case of the WT enzyme, the initial EnzyDock consensus poses were used as starting points for multiscale free energy modeling. The WT modeling commenced with the substrate GGPP and terminated with the carbocation precursor to the final product cyclooctat-9-en-7-ol. All carbocation intermediates were found to be stabilized via strong π−cation, dipole−cation, and ion-pair interactions with the enzyme and the PP cofactor (Fig. 5). The EnzyDock and free energy simulations also proposed the position of the catalytic water, which was not observed in the crystal structure. Subsequent crystallography studies of CotB2 variants have shown that this proposed catalytic water pocket indeed hosts a water molecule (unpublished results). The 11-step carbocation cascade in CotB2 was also compared with the same reaction in the gas phase. Remarkably, the free energy profiles in gas phase and in CotB2 were found to be surprisingly similar, despite the multitude of strong interactions between all intermediates in the reaction cascade and the enzyme. This suggests a remarkable balance of interactions in CotB2, which was ascribed to the similar magnitude of the interactions between the carbocations along the reaction coordinate and the enzyme environment. In the case of CotB2, the above-mentioned 11 CotB2 mutations were studied using EnzyDock, machine learning, and X-ray crystallography, and the effect of the mutations on product distribution was rationalized.

Figure 5.

Key cations along the reaction trajectory are stabilized via strong non-covalent interactions.

When docking the substrate (e.g., GPP, FPP, or GGPP) one must pay attention to which of the two PP oxygens the isoprenoid moiety links to. A recent study has shown that this depends on whether the TPS originates from plants or not (Schwartz, Zev, & Major, 2024).

Here a distinction between EnzyDock simulations and free energy simulations is in place. The advantage of performing free energy simulations for an enzyme mechanism is the deep insight, detail, and accurate free energy profiles generated. Yet, this comes at a price as these simulations are very costly. EnzyDock presents a low-cost alternative to free energy simulations, as it can generate all important states along a reaction coordinate, including energetic information, at an affordable computational cost. The optimal poses generated using EnzyDock can serve as starting points for subsequent free energy simulations.

4. CONCLUDING WORDS

To conclude, in this review we highlighted unique features of the advanced multistate, multiscale docking program EnzyDock and its applications to the natural product enzyme family TPS, which catalyzes complex reaction cascades. Docking ligands in TPS brings about significant challenges, including: (1) Flexibility of the substrate isoprenoid chain and the weak dispersion interactions governing its binding; (2) tumbling of the carbocation intermediates; (3) lack of directional interactions that anchor the carbocation intermediates in well-defined binding modes; (4) all bound poses along a reaction cascade are most likely very similar (consensus binding); (5) all bound states must match the correct final stereo configuration of the product; (6) the docked intermediates must be in correct poses for bases/acids to deprotonate/protonate the intermediate; (7) chemical changes often occur in the protein-cofactor during the docking of multiple states. EnzyDock was designed to be able to tackle these challenges, which are difficult with standard docking tools. The uses of EnzyDock’s capabilities were exemplified on the bacterial TPS class I enzyme, CotB2, where a set of consistent docked poses of intermediates was found, matching experimental knowledge and several design experiments were rationalized. The EnzyDock code is available at https://github.com/majordt/EnzyDock.