Large-Scale Validation of Mixed-Solvent Simulations to Assess

Article pubs.acs.org/jcim

Cite This: J. Chem. Inf. Model. 2018, 58, 784−793

Large-Scale Validation of Mixed-Solvent Simulations to Assess Hotspots at Protein−Protein Interaction Interfaces Phani Ghanakota,*,⊥ Herman van Vlijmen,± Woody Sherman,⊥,† and Thijs Beuming⊥ ⊥

Schrödinger, Inc., 120 West 45th Street, New York, New York 10036, United States Janssen Pharmaceuticals, Beerse, B-2340, Belgium

±

Downloaded via NEW MEXICO STATE UNIV on July 6, 2018 at 13:26:10 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

S Supporting Information *

ABSTRACT: The ability to target protein−protein interactions (PPIs) with small molecule inhibitors offers great promise in expanding the druggable target space and addressing a broad range of untreated diseases. However, due to their nature and function of interacting with protein partners, PPI interfaces tend to extend over large surfaces without the typical pockets of enzymes and receptors. These features present unique challenges for small molecule inhibitor design. As such, determining whether a particular PPI of interest could be pursued with a small molecule discovery strategy requires an understanding of the characteristics of the PPI interface and whether it has hotspots that can be leveraged by small molecules to achieve desired potency. Here, we assess the ability of mixed-solvent molecular dynamic (MSMD) simulations to detect hotspots at PPI interfaces. MSMD simulations using three cosolvents (acetonitrile, isopropanol, and pyrimidine) were performed on a large test set of 21 PPI targets that have been experimentally validated by small molecule inhibitors. We compare MSMD, which includes explicit solvent and full protein flexibility, to a simpler approach that does not include dynamics or explicit solvent (SiteMap) and find that MSMD simulations reveal additional information about the characteristics of these targets and the ability for small molecules to inhibit the PPI interface. In the few cases were MSMD simulations did not detect hotspots, we explore the shortcomings of this technique and propose future improvements. Finally, using Interleukin-2 as an example, we highlight the advantage of the MSMD approach for detecting transient cryptic druggable pockets that exists at PPI interfaces.

1. INTRODUCTION There are an estimated 350 000 protein−protein interactions in the human interactome alone,1 which offers a tremendous opportunity to modulate disease related pathways. However, protein−protein interaction (PPI) interfaces are typically shallow and extend over large surfaces, presenting unique challenges for small molecule design.2 Moreover, unlike other protein classes such as kinases and GPCRs, a diverse set of proteins with different folds and interaction motifs fall under the classification of PPI targets,3 thus presenting challenges for informatics-based approaches. It is possible that many PPIs may be undruggable with traditional small molecules, or at least very challenging, and as such it is valuable to understand the risks involved in pursuing a particular PPI target before embarking on a drug discovery program. Central to understanding this risk is to identify if hotspots exist at the PPI interface that can be leveraged in small molecule drug discovery programs such that sufficient potency can be gained without the need to design excessively large compounds that may present off-target liabilities and ADME/Tox risks. The challenges one faces when dealing with PPIs merit the development of computational approaches that can screen PPI interfaces for hotspots, thereby allowing one to assess the likelihood of success in disrupting a PPI interface. Computational © 2018 American Chemical Society

approaches to address druggability that rely on static conformations abound. These techniques rely on simple geometric and energetic metrics, such as the curvature of the protein surface and degree of hydrophobicity, to assess tractability of pursuing targets in small molecule discovery programs. Such techniques include those published by researchers at pharmaceutical companies such as Pfizer (Cheng et al.4) and AbbVie (Hajduk et al.5), as well as commercially available programs like Grid,6 FTMap,7 and SiteMap8,9 to name a few. In addition, techniques that involve computing the thermodynamic properties of water molecules at the surface of proteins have also been used to assess the druggability of PPIs.10,11 PPI interfaces are typically flat and undergo conformational changes when small molecule ligands bind.2 Therefore, it is particularly appealing to understand the existence of hotspots at PPI interfaces using molecular dynamics-based approaches, which allow for full protein flexibility. MSMD simulations are an attractive MD-based alternative to rigid receptor approaches to locate hotspots because they allow for the protein to reorganize in ways that might facilitate ligand binding, even if such sites are Received: January 28, 2018 Published: April 4, 2018 784

DOI: 10.1021/acs.jcim.7b00487 J. Chem. Inf. Model. 2018, 58, 784−793

Article

Journal of Chemical Information and Modeling

sites. The results can be visualized as clusters of “site points”, which denote patches on the protein surface that are druggable binding sites. SiteMap also generates scores (SiteScore, Dscore, volume, etc.) that can be used to assess druggability. In order to run SiteMap, the protein conformations were prepared using Maestro.25 All ligands, crystallographic additives, and water molecules were removed, and the Protein Preparation Wizard26 was run to add hydrogen atoms and predict the rotamer of ambiguous residues (Asn, Gln, and His). Subsequently, SiteMap was run using the default settings as well as with the “shallow binding site” settings. This process was repeated for all the PPI conformations for the proteins in this study, which are listed in Table 1.

not present in the apo or protein-bound conformation. MSMD simulations are conceptually simple: simulations of proteins are performed in the presence of explicit water molecules and some fraction of one or more cosolvents.12 The motivation for including the cosolvents is that they can induce high-energy protein conformations that might be amenable to small molecule binding but are too rare to observe without some kind of small molecule probe. Information regarding hotspots can be extracted based on the probabilities (density distribution) of the various cosolvents. Importantly, MSMD simulations explicitly include protein flexibility and competition with water, thereby offering a realistic treatment of the PPI interface. Recent methodological advances in MSMD simulations have opened several potential applications for such simulations, which extend beyond hotspot mapping.12−14 The first MSMD technique proposed was MDmix,15 wherein MSMD simulations of proteins with isopropanol were performed to comment upon maximal affinity achievable at different patches on the protein surface. Subsequently several other approaches such as the Site Identification by Ligand Competitive Saturation,16 MixMD,17 Ligand Mapping MD,18,19 and others were developed.12 These techniques all have a similar approach in that systems are sampled with molecular dynamics, which includes protein flexibility, and cosolvents are intermixed with water at varying concentrations. All of these techniques have shown encouraging results and suggest that this general approach offers promise for tackling challenging problems in drug discovery. Here we have applied MSMD simulations on 21 PPI interfaces, choosing only those with known small molecule inhibitors bound. We reasoned that these PPI interfaces should contain hotspots as small molecules have been used to target them. Using this data set of 21 PPI interfaces extracted from the TIMBAL20 database, we performed MSMD simulations with acetonitrile, isopropanol, and pyrimidine cosolvents. These cosolvents were chosen as they are know to be water-miscible and mix evenly in MSMD simulations without phase separation.21 A recent study had shown that using MSMD simulations with these three probes could be used to successfully recapitulate allosteric sites on various proteins.22 Our choice of probes and workflow for performing MSMD simulations follows from previous work done in the development of the MixMD approach by Carlson and co-workers.17,21−24 To compare the MSMD approach with a simpler method that does not include protein flexibility or explicit solvent, we performed SiteMap calculations on the same PPI targets. The starting conformation for MSMD was then chosen randomly from those conformations where SiteMap did not detect the PPI interface. In this study, we find that MSMD simulations and SiteMap perform comparably in many targets. Those PPI interfaces that evade detection with SiteMap appear to do so as a result of the flat nature of the PPI interface. MSMD simulations, on the other hand, were not limited by such problems. Additionally, the benefits of including conformational flexibility were apparent for examples such as Interleukin-2 (IL-2). In the following sections we present our analysis and explore the advantages and limitations of MSMD simulations for hotspot detection at PPI interfaces.

Table 1. Dataset Used for This Studya MSMD

protein ZipA RAD51 TNFα S100B HRas MDMX Bcl-XL XIAP MDM2 Neuropilin MLL E2 Clathrin Integrin αIIbβ3 Plk1 PBD IL-2

number of proteinbound structures in PDB

number of structures where SiteMap detects a binding site at the PPI interface

starting conformation (chain)

does MSMD detect a hotspot (20σ)

1 1 17 17 35 17 97 9 36 1 6 1 4 14

0 0 0 9 8 13 76 7 36 1 6 1 4 12

1F47 (A) 4B3B (A) 1TNF (A, B) 1MQ1 (A, B) 4K81 (D) 3EQY (A) 1G5J (A) 1G3F (A) 1YCR (A) 4DEQ (B) 4GQ6 (B) 1R6N (B) 2XZG (A) 3ZDY (A, B)

yes yes yes yes yes yes yes yes yes yes yes yes yes no

28 2

28 0

3P35 (B) 2ERJ (D)

no no

a

Hotspot analysis on PPI proteins from the TIMBAL data set20 was performed using SiteMap and MSMD simulations. The number of PPI conformations in the PDB for each target, number of cases where SiteMap detects a binding site, and MSMD starting conformation are shown, in addition to whether MSMD was able to detect a hotspot at the PPI interface using a 20σ cutoff. Where available MSMD simulations were started from PPI conformations not mapped by SiteMap.

2.2. Mixed-Solvent Simulation Setup and Analysis. Our protocol for running MSMD simulations was inspired by the work done by Carlson and co-workers in their development of MixMD.22,24 We reimplemented this protocol to run with the Desmond MD engine.27 In brief, this approach requires the protein to be layered with a user-specified number of cosolvent molecules and placed in a box of TIP3P28 water molecules. Prior to the Desmond setup, the proteins were prepared using the Protein Preparation Wizard.26 Care was taken to correct tautomeric states of histidine residues. Asparagine and glutamine residue flips were also handled appropriately. The protein was then surrounded with an 8-Å layer of cosolvent molecules. Our choice of cosolvent molecules in this work is in line with the work of MixMD (acetonitrile, isopropanol, and pyrimidine). These probes allowed us to capture a range of interactions such as hydrophilic, hydrophobic, hydrogen bonding, and aromatic.

2. METHODS 2.1. SiteMap Setup and Analysis. SiteMap8,9 is a gridbased approach that relies on a rigid conformation of the protein and assesses druggability based on the curvature and energetic properties (hydrophobic and hydrophilic) of putative binding 785


Article


Figure 1. Example of MSMD grid maps with different isodensity cutoffs. ZipA MSMD simulation maps of acetonitrile (orange), isopropanol (blue), and pyrimidine (purple) overlaid and contoured at three different values 80σ, 20σ, and 5σ illustrate the nature of MSMD grid maps. ZipA MSMD maps when contoured too high (80σ) result in very little mapping of the protein surface. When mapped at (5σ), the entire surface is mapped. We have found it ideal to contour the maps at 20σ to detect hotspots. At this contour value the individual densities approach the volume of the probes used. This relationship is further explored in Figure 2.

bound crystal structure allows us to perform a direct comparison of the hotspot locations from MSMD and their relationship with the ligands from crystal structures. The TIMBAL20 database contained a curated data set of PPI interfaces that met this criteria. All 21 PPI targets present in TIMBAL (listed in Tables 1 and S1) were used for MSMD simulations. We chose to use protein-bound conformations for running MSMD simulations, as apo structures were not available for all 21 proteins and a small molecule-bound structure would be preinduced into a binding conformation that would not be a realistic scenario for the real-world application of interest (i.e., determining if a PPI can be inhibited with a small molecule). The lack of apo PPI conformations in this study may be considered a limitation for now and will be explored in future work. Nevertheless, where available the initial conformation we used for MSMD simulations was randomly picked from one of the PDB31 PPI conformations where SiteMap did not detect the PPI interface as druggable, thus indicating that these are all challenging interfaces for binding site and hotspot detection, albeit probably not as challenging as studying apo structures. Upon inspection of these conformations, we observed that the shallow nature of the binding site was the primary reason for not being detected by SiteMap. Curvature is an important component of several druggability prediction methods including SiteMap and therefore we expect that most methods that rely on the curvature of a rigid input structure would fail to detect binding sites for most of these PPIs. Our setup not only allowed use to interrogate if the lack of an explicit curvature term has an effect on pure physics-based method such as MSMD but also let us test the effectiveness of MSMD simulations to detect hotspots when all one has is a PPI conformation to start from. 3.2. Detecting Hotspots at PPI Interfaces. Druggability of a protein target is a concept that is deeply intertwined with the existence of hotspots. As our objective here is to approach the problem of druggability as a binary yes or no classification, we believe the best way to do this is by searching for the existence or absence of hotspots at PPI interfaces. Based on this simple definition, PPI interfaces that are devoid of hotspots would be considered undruggable and vice versa. Our definition of hotspots from MSMD simulations is similar to the one reported in experimental studies such as Multiple Solvent Crystal Structures (MSCS),32−35 NMR hit rates,5 and work done in MixMD development.22,24 Hotspots are defined as locations on the protein surface that are mapped by more than one type of probe. In order to realize this definition of a hotspot, it is important to understand the outcome of an MSMD simulation. MSMD simulations result in a grid map of where regions of high cosolvent probe densities are detected at the protein surface. The

MSMD simulations for each protein were run with the three different probes separately. The setup involved placing a layer of cosolvent molecules around the surface of the protein followed by the placement of this system within a box of water molecules. The number of water molecules was adjusted such that a ratio of 5% cosolvent:water was achieved. Simulations of this setup were performed in Desmond using the OPLS3 force field,29 which includes improvements in the treatment of small molecules30 and proteins as compared with previous versions of the OPLS force field. The system was equilibrated using the standard Desmond NPT relaxation protocol where in the protein is subjected to a series of heating and minimization stages. The initial system was minimized with a restraint of 50.0 kcal/mol on the solute heavy atoms using the NVT ensemble first with Brownian dynamics for 100 ps followed by Langevin dynamics for 12 ps at a temperature of 10 K. This system was then further equilibrated using the NPT ensemble with similar restraints on solute heavy atoms for 24 ps at 300 K. It was then subjected to a short simulation of 24 ps using the NPT ensemble without restraints. This was followed by a 20 ns production simulation using the NPT ensemble. Ten such simulations were performed for each protein and cosolvent probe, randomizing the starting location of the cosolvents during the setup. This resulted in a cumulative total of 200 ns for each protein for each unique cosolvent. For the analysis, only the last 5 ns of the production run across all 10 runs for each cosolvent were taken into consideration. The cosolvent heavy atoms were spatially binned using a 0.5 Å grid. The raw grid data was then normalized by using the following equation (x − μ)/σ where x is the raw grid count, μ is the mean of the grid, and σ is the standard deviation of the grid. This normalization process allowed us to compare results across different cosolvent simulations. The processed grids were then contoured at a value of 20σ to locate highly occupied regions in our MSMD simulations. All our MSMD maps are color coded in a similar manner used in MixMD: acetonitrile maps are colored orange, isopropanol maps are colored blue, and pyrimidine maps are colored purple.22,24

3. RESULTS AND DISCUSSIONS 3.1. System and Starting Conformation Selection TIMBAL Data Set of PPI Targets. PPI targets selected for MSMD simulations were chosen such that crystal structures of small molecules that disrupt the PPI interface exist. This was important for two reasons. First, the presence of small molecules binding at the interface confirms that it is possible for small molecules to bind with at least some amount of detectable affinity, thereby providing us with a data set to assess if MSMD simulations can detect binding hotspots. Second, the ligand 786


Article


Figure 2. Volume of the largest isolated density surface across all the MSMD simulations at each contour value. We find that, at 20σ, the maximum volume of each separate probe density does not exceed the average volume of the probes used (78.17 Å3, green line).

Figure 3. Comparison of hotspots detected in MSMD simulations and SiteMap for ZipA and MDM2. The small molecules that disrupt the PPI interface are rendered in green sticks. All MSMD maps are contoured at 20σ and are color coded such that acetonitrile maps are orange, isopropanol maps are blue, and pyrimidine maps are purple. (A) Results from ZipA (PDB ID 1Y2F) show strong hotspot activity detected by MSMD simulations whereas SiteMap fails to detect the PPI interface. (B) MSMD and SiteMap detect the PPI interface in MDM2 (PDB ID 4JV7), although MSMD shows a more complete coverage of the sub pockets occupied by the ligand.

20σ is a suitable contour value to detect hotspots. We have systematically explored this relationship to identify in an objective manner the ideal contour value to use for detecting hotspots. By contouring all the MSMD maps from all cosolvents and proteins we observe that at 20σ the volume of each isolated density surface from every MSMD grid map never exceeds the average volume of the cosolvent molecules used (see Figure 2). In other words, when the MSMD maps are contoured at 20σ we obtain probe density volumes that resemble the volumes of probes used in the MSMD simulations. Thus, in all our MSMD simulations, we have chosen to contour the MSMD maps at 20σ

maps generated from these probe densities are contoured at a given isodensity value to provide a visual indication of where probes bind on the protein surface. The contour value for MSMD simulations can have a profound impact on how one detects cosolvent binding on the protein surface, which ultimately influences what is considered a hotspot. This is illustrated in Figure 1, when MSMD maps for the protein ZipA are contoured at 80σ, very little of the protein surface is mapped. On the other hand when contoured at 5σ the entire protein surface is mapped. In our implementation of the MSMD grid map generation, we have noticed through visual inspection that 787


Article


Figure 4. Comparison of MSMD, SiteMap default settings, and SiteMap shallow binding site settings for (A) Bcl-XL and (B) XIAP. The ligands that displace the PPI partner are shown as sticks and colored green to delineate the PPI interface. The ligands were obtained from PDB IDs 2YXJ and 3YEL, for Bcl-XL and XIAP respectively. Each SiteMap site is color coded differently and the extent of the site is depicted by site points that are shown as spheres. SiteMap default settings underpredicted the PPI interface; whereas, SiteMap shallow binding site settings overpredict the extent of the PPI interface. MSMD simulations appear to find the right balance in detecting hotspots at the PPI interface.

Integrin αIIbβ3, leaving IL-2 as the only protein where both methods fail. Overall in 8 of the 16 PPI targets, a conformation of the PPI interface not mapped by SiteMap was available where MSMD could be shown to detected hotspots at the PPI interface (ZipA, RAD51, TNFα, S100B, HRas, MDMX, BCLXL, XIAP). In ZipA, we observed several hotspots at the PPI interface, as shown in Figure 3A. These hotspots appear to be in strong agreement with the location of all ligand molecules from the crystal structures. One of these ligands is shown overlaid with the MSMD maps (Figure 3A, ligand in green stick representation). SiteMap did not detect the PPI interface in ZipA, which was understandable given the flat nature of the PPI interface (see Figure 3A). Unlike techniques such as SiteMap, pure physics based methods such as MSMD do not require any knowledge of the curvature of the protein surface or other descriptors to assess the existence of hotspots. On this note, it was pleasing to observe that PPI interfaces such as ZipA were correctly predicted to have hotspots using MSMD simulations. We also noticed several systems where both SiteMap and MSMD simulations performed comparably. One such example from this set is MDM2. MDM2 is known to be a highly druggable system where several groups have reported diverse high affinity inhibitors.37 As such, it might be expected that both SiteMap and MSMD agree that this system comprises hotspots at the PPI interface. In Figure 3B, the PPI interface (delineated with a small molecule rendered as green sticks) is shown mapped by both MSMD and SiteMap. Interestingly, MSMD simulations appear to provide more detailed information on specific sub pockets into which the ligand functional groups protrude. This is illustrated for MDM2 in Figure 3B, MSMD simulations reveal three sub pockets targeted by the small molecule from PDB ID 4JV7. The subpockets are clearly defined by the MSMD simulations where

and identify hotspots as overlapping densities from different cosolvent MSMD simulations. 3.3. MSMD Complements SiteMap in Several PPI Targets. MSMD simulations revealed hotspots at PPI interfaces in 18/21 cases studies here using a contour of 20σ (see Table 1). Our data set of PPI targets had five proteins that were part of the bromodomain family, which in some ways resemble a small molecule-binding site due to the deep binding pocket present for lysine recognition and therefore are not necessarily reflective of the properties one would associate with a typical PPI interface. Indeed, a recent study on bromodomains using SiteMap demonstrated these PPI interfaces were druggable.36 We observed the same in our SiteMap calculations and MSMD simulations. Thus, we chose to analyze these systems separately, and the results are presented in the Supporting Information. The MSMD and SiteMap results for the remaining 16 PPI targets are summarized in Table 1. Interestingly, in three systems (ZipA, RAD51, TNFα) MSMD identified hotspots at the PPI interface where SiteMap did not detect the PPI interface in any of the available PPI conformations. Inspection of these conformations suggested that for the most part, the flat nature of the PPI interface was responsible for these edge cases to be missed by SiteMap. Similar but less frequent missed mapping of the PPI interface was seen in five protein systems (HRas, MDMX, BclXL, XIAP). In these systems SiteMap mapped the PPI interface in a subset of the crystal structures. MSMD mapped the PPI interface on a conformation randomly chosen from the ones where SiteMap did not detect the PPI interface. In five protein targets, MSMD and SiteMap compared similarly where all the PPI conformations were mapped by both techniques (MDM2, Neuropilin, MLL, E2, Clathrin), whereas in the remaining three targets MSMD did not identify hotspots at the PPI interface. Of these three targets, SiteMap detects sites in PLK1 PBD and 788


Article


Figure 5. PLK1 PBD MSMD simulations do not detect hotspots at the PPI interface. These results are in agreement with literature data reflecting the weak affinity of ligands after several rounds of optimization. All MSMD maps are contoured at 20σ and are color coded such that acetonitrile maps are orange, isopropanol maps are blue, and pyrimidine maps are purple. (A) PLK1 PBD displays no hotspot activity in MSMD simulations. (B) SiteMap detects the PPI interface. The reference ligand shown was obtained from PDB ID 4H71.

four distinct hotspots occupy this region. On the contrary, the SiteMap cluster in this region (see Figure 3B) has a partial overlap with the ligand position and misses one of the hotspots where a phenyl group from the ligand is positioned. This increased level of detail from MSMD present a clear benefit over SiteMap, which can aid in the formulation of new design ideas for small molecule development. In addition to the default settings, SiteMap provides a shallow binding site mode, which lowers the thresholds for pocket detection, allowing one to detect more shallow binding sites. Applying these settings to the current set of PPI interfaces, we observed a deterioration of sensitivity, identifying large areas of the protein surface as a potential binding site. As small molecules are not expected to occupy such large regions of the protein surface, these predictions have little to offer in understanding druggability from the perspective of small molecule development. The impact of the shallow binding site mode in SiteMap is demonstrated using Bcl-XL and XIAP. As seen in Figure 4, SiteMap shallow binding site settings suggest large portions of the protein surface as a binding site in both proteins. MSMD simulations provided a more accurate depiction of the hotspots at the protein surface. For this reason, we limit the comparison of MSMD with the default settings of SiteMap in the remainder of the manuscript. Similar observations were made for the other proteins in this data set and can be found in the Supporting Information (Figures S17−S36). Overall, our analysis reveals that when SiteMap maps a PPI interface, it is more than likely that MSMD simulations identify hotspots at the same location. The advantage of speed with a SiteMap analysis that finishes in a matter of seconds needs to be balanced with the greater sensitivity of MSMD simulations that come at a cost of several hours of compute time needed to complete and analyze the simulations. If the goal is to assess pocket druggability, the application of the more computationally intensive MSMD calculations are most valuable in those cases where SiteMap does not detect the PPI interface. However, when additional insights are of interest, such as a more detailed understanding of the sub pockets, or the presence of additional sub pockets that might be induced by the cosolvent probes during the MD simulations, MSMD simulation offer significant added value over SiteMap. Figures depicting SiteMap and MSMD results for other systems are presented in the Supporting Information (Figure S1−S16). The scope of our current work was limited to assess if MSMD simulations could detect hotspots at the PPI interface. Examination of the MSMD maps across all the PPIs suggest that most hotspots appear as clusters within the PPI interface with the remaining few being scattered across the protein surface. Previous crystallographic studies using the MSCS have

demonstrated that binding sites are comprised of a cluster of hotspots. It is appropriate in this regard that PPI interface studied here which are know to bind small molecules are predominantly the locations were a cluster of MSMD hotspots are observed. In a prospective setting, one could use such information to detect and prioritize binding sites for drug discovery efforts. In the next sections, we explore in detail three PPI interfaces where MSMD had difficulty detecting hotspots. These systems allowed us to understand the scope and limitations of MSMD simulations in their current form. 3.4. PLK1 Polo Box DomainA Difficult PPI Interface to Target. To our surprise, PLK1 Polo Box Domain (PLK1 PBD) was one of the protein targets where SiteMap had mapped the PPI interface; whereas MSMD simulations did not. The MSMD results show that the PPI interface is only partially mapped with isolated pyrimidine and acetonitrile densities when contoured at 20σ (Figure 5A). As no overlapping densities from different probes could be found, none of these isolated densities from cosolvents at the PPI interface were considered to be hotspots. On the contrary, SiteMap maps the entire region where the ligand binds at the PPI interface where the SiteMap cluster of interest is colored cyan (Figure 5B). We only simulated the PBD domain of PLK1, which contains the PPI interface. However, the full length PLK1 protein also has a kinase domain. While kinases are generally considered to be druggable, targeting the kinase has been associated with unintended off target effects with other similar kinase domains.38 Thus, great interest has been shown in pursuing the PPI interface on the PBD domain.39 Interestingly, high affinity inhibitors for PBD domain are yet to be discovered.39 Molecules reported in the literature that bind at the PPI interface are predominantly cyclic peptides.39 Additionally the most potent small molecule for this system is Poloxin-2 with an IC50 of 1 μM.40 This protein target drew parallels to what was observed in a recent study with the tyrosine phosphatase PTP1B in MixMD.25 Efforts to identify high affinity and orally bioavailable small molecules for PTP1B that required to be bound to a charged binding site were met with limited success. Similarly, the PPI interface in the PBD domain is known to bind a phosphorylated serine.41 Encouraged by our MSMD results, we speculate that this is a difficult target to pursue. However, given the small number of literature reports, it remains to be seen if PLK1 PBD domain can be targeted by high affinity small molecules. 3.5. Integrin αIIbβ3Metal Binding Proteins Expose a Limitation of MSMD Simulations. Integrins are a class of proteins that are known to play a central role in coagulation.42 Disrupting the PPI interaction of Integrin αIIbβ3 with fibrinogen has been used as an effective strategy to generate anticoagulant drugs.43 Several therapeutics that target Integrin αIIbβ3 exist, 789


Article


Figure 6. Integrin αIIbβ3 MSMD simulations do not detect hotspots at the PPI interface. All ligands disrupting the Integrin αIIbβ3 PPI interface overwhelmingly rely on metal interactions for achieving potency, suggesting a lack of hotspots. All MSMD maps are contoured at 20σ and are color coded such that acetonitrile maps are orange, isopropanol maps are blue, and pyrimidine maps are purple. The αIIb domain is colored green, and the β3 domain is colored cyan. (A) MSMD results for Integrin contoured at 20σ show no hotspots. (B) SiteMap does not detect the Integrin PPI interface. (C) Close-up view of Tirofiban (rendered in stick model and colored yellow), an FDA approved drug (PDB ID 2VDM), binding at the Integrin PPI interface overlaid with the MSMD results contoured at 20σ. The carboxylic acid of Tirofiban and the metal ion it chelates with (colored as a black sphere) are highlighted using a red outline.

which range from macromolecules, including Abiciximab44,45 (monoclonal antibodies) to Eptifibatide46 (cyclic peptides) and small molecules such as Tirofiban.47 Integrin αIIbβ3 is a heterodimer of the αIIb and β3 domains. These domains can be seen in Figure 6A overlaid with the MSMD maps, where αIIb is colored green and β3 is colored cyan. MSMD simulation did not detect hotspots at the Integrin αIIbβ3 PPI interface (see Figure 6A) while SiteMap succeeded in a majority of the conformations (Table 1). Visual inspection of the PPI conformations suggested that the angle at which the two heterodimers (αIIb and β3 domains) contact each other influences the concavity and ultimately the end results from SiteMap. However, SiteMap maps the PPI interface in 12 of the 14 PPI conformations (Table 1) even though it does not do so in the conformation we used to start MSMD simulations as seen in Figure 6B Based on the aggregate analysis of all Integrin αIIbβ3 PPI conformations, this PPI interface may be considered to be druggable according to SiteMap. Tirofiban is an FDA approved small molecule that is administered intravenously.47 Crystal structures of Tirofiban bound to the PPI interface allowed us to examine its location in conjunction with the MSMD simulation results. A close up view of the PPI interface and Tirofiban can be seen in Figure 6C. As with PLK1 PBD (Figure 5), there are isolated mapping from different cosolvents at the PPI interface. However, no overlapping cosolvent densities meant that we did not identify hotspots at the PPI interface. We were interested in analyzing these discrepancies in greater detail. Our analysis on the structure activity relationship data for Tirofiban revealed a key observation. Removing the metal binding warhead of Tirofiban resulted in a complete loss of activity.48 This may explain to a certain extent the lack of hotspots at the PPI interface. However, several other orally bioavailable analogues such as RUC-4 are also being

pursued in the clinic, suggesting this protein to ultimately be druggable.49 Against this backdrop, it is fair to consider that MSMD simulations with our current set of probes did not perform optimally for Integrin αIIbβ3. Detecting metal binding sites using MSMD simulations may require further screening for a suitable set of probes. 3.6. Interleukin-2Cryptic Pockets at Lower Populations in MSMD Simulations. IL-2 was the only protein in this study where both MSMD and SiteMap did not detect the PPI interface. The MSMD maps for IL-2 showed in Figure 7A display no hotspot activity at the PPI interface. We believe these difficulties for both techniques are associated with conformational fluctuations on the side chain level. The IL-2 PPI interface harbors a well-known cryptic pocket,50 where a single phenylalanine residue (Phe 42) flip is critical for opening this pocket. This conformational change is depicted in Figure 7B, where holo (Green) and PPI (white) conformations of IL-2 are overlaid along with a ligand that takes advantage of this cryptic pocket. The opening and closing of the cryptic pocket can be easily tracked using the CHI1 angle of Phe 42. In the holo conformations, the CHI1 angle is close to −176.3° whereas in the PPI conformation it is closer to −59.5° (Figure 7C). This straightforward way of tracking the opening and closing of the pocket allowed us to ask the following question. Is the holo conformation present in our MSMD simulations? To answer this question we first surveyed all our IL-2 MSMD simulations to understand the distribution of CHI1 angle of Phe 42. This distribution is show in Figure 7D, where the CHI1 angle histogram is plotted for acetonitrile, isopropanol, and pyrimidine MSMD simulations separately. We observed two different populations of Phe 42 rotameric state representing the holo and PPI conformations. The holo conformations are present in our MSMD simulations but occurred infrequently. In order to 790


Article


Figure 7. Trajectory clustering approach which reveals cryptic pockets in IL-2. A ligand that binds into this cryptic pocket is shown in all figures to orient the viewer to the PPI interface (PDB ID 1M48). (A) MSMD results of IL-2 contoured at 20σ. (B) Movement of Phe 42 reveals a new pocket where ligands bind. (C) CHI1 angle of Phe 42 adopts two different positions in the PPI and Holo structures of IL-2. (D) Histogram of the CHI1 angle of Phe 42 in MSMD simulations reveals the holo conformation in lower population. (E) Generating MSMD density maps for the holo and PPI like conformations of Phe 42, reveals high hotspot activity at the cryptic pocket. All MSMD maps are contoured at 20σ and are color coded such that acetonitrile maps are orange, isopropanol maps are blue, and pyrimidine maps are purple.

clustering approach in its current form relies on prior knowledge. This is possible in our instances as we start from a PPI conformation, so clustering is targeted to specific residues at the PPI interface. However, a more robust protocol without including prior knowledge would require advanced clustering techniques. We are currently investigating several such methods and intend to report our observations in a more focused study across a wide range of protein targets not necessarily limited to PPIs. Taken together, our analysis on IL-2 opens up exciting prospects for coupling conformational analysis with MSMD simulations to detect hotspots.

detect any changes in hotspot activity based on the conformation of the PPI interface, we generated MSMD maps for the holo and PPI-like conformations. To our surprise we observed hotspots at the PPI interface when we examined the holo-like conformations (Figure 7E). These hotspots were absent in the MSMD maps generated for the PPI conformations. To the best of our knowledge, this is the first instance where subpopulations of MSMD simulations have been analyzed for hotspot activity, although other works have explored clustering of conformations from MD trajectories followed by SiteMap calculations to detect subpopulations.51,52 While our default MSMD analysis reveals hotspots by taking into consideration all the simulation data, the results here suggest there is much to be gained from analyzing specific subsets of conformations from MSMD simulation. Interesting as it may be, we realize that this

4. CONCLUSIONS Mixed-solvent molecular dynamics simulations were applied across 21 PPI targets to detect and characterize the interfaces 791


Journal of Chemical Information and Modeling where small molecules bind. We determine that a visual inspection of the MSMD maps for the presence or absence of hotspots provides valuable insights into the druggability of a PPI interface. Our comparisons between a grid-based rigid receptor approach (SiteMap) and MSMD highlighted cases where MSMD detected a binding site at the PPI interface when SiteMap failed to do so. Interestingly, no hotspots were found for three PPI interfaces using MSMD simulations. We examined these PPI interfaces in greater detail, which provided valuable insights into the advantages and limitations of MSMD. The first PPI interface was PLK1 PBD, where literature evidence points to difficulties in generating high affinity small molecules. The second PPI interface was Integrin αIIbβ3 where we noticed metal binding interaction was paramount for high affinity. MSMD simulations did not pick up these metal interaction sites as hotspots and indicates a weakness in the current approach, which could be improved by including more diverse cosolvent probes (e.g., highly polar or charged probes). The third PPI interface was IL-2 where significant conformational fluctuations at the side chain level were necessary for mapping hotspots. Our conformational analysis of IL-2 MSMD simulations identified a lower population of holo like conformations where hotspots were present at the PPI interface, which could be detected once the conformations from the MD trajectory were clustered. To the best of our knowledge, this is the first study to analyze subpopulations of MSMD simulation to assess hotspots. We believe that further development of this protocol can assist in investigating challenging PPI interfaces that involve significant conformational changes, and will be our focus in future works. Additional advantages can arise from using grid-based tools that rely on a static protein structure such as SiteMap in combination with MSMD simulations for understanding the druggability of PPI interfaces. Based on our observations, we recommend first running a fast druggability analysis tool such as SiteMap before proceeding to use more advanced approaches such as MSMD to interrogate the existence of hotspots at PPI interfaces.

■

■

ABBREVIATIONS

■

REFERENCES

MSMD, mixed-solvent molecular dynamics; MD, molecular dynamics; PDB, Protein Data Bank; PPI, protein−protein interaction; PBD, Polo Box Domain; IL-2, Interleukin-2; GPCR, G-Protein-Coupled Receptor; ADME, absorption, distribution, metabolism, and excretion.

(1) Stumpf, M. P. H.; Thorne, T.; de Silva, E.; Stewart, R.; An, H. J.; Lappe, M.; Wiuf, C. Estimating the Size of the Human Interactome. Proc. Natl. Acad. Sci. U. S. A. 2008, 105, 6959−6964. (2) Wells, J. A.; McClendon, C. L. Reaching for High-Hanging Fruit in Drug Discovery at Protein−Protein Interfaces. Nature 2007, 450, 1001−1009. (3) Arkin, M. R.; Tang, Y.; Wells, J. A. Small-Molecule Inhibitors of Protein-Protein Interactions: Progressing toward the Reality. Chem. Biol. 2014, 21, 1102−1114. (4) Cheng, A. C.; Coleman, R. G.; Smyth, K. T.; Cao, Q.; Soulard, P.; Caffrey, D. R.; Salzberg, A. C.; Huang, E. S. Structure-Based Maximal Affinity Model Predicts Small-Molecule Druggability. Nat. Biotechnol. 2007, 25, 71−75. (5) Hajduk, P. J.; Huth, J. R.; Fesik, S. W. Druggability Indices for Protein Targets Derived from NMR-Based Screening Data. J. Med. Chem. 2005, 48, 2518−2525. (6) Goodford, P. J. A Computational Procedure for Determining Energetically Favorable Binding Sites on Biologically Important Macromolecules. J. Med. Chem. 1985, 28, 849−857. (7) Brenke, R.; Kozakov, D.; Chuang, G.-Y.; Beglov, D.; Hall, D.; Landon, M. R.; Mattos, C.; Vajda, S. Fragment-Based Identification of Druggable ‘Hot Spots’ of Proteins Using Fourier Domain Correlation Techniques. Bioinformatics 2009, 25, 621−627. (8) Halgren, T. New Method for Fast and Accurate Binding-Site Identification and Analysis. Chem. Biol. Drug Des. 2007, 69, 146−148. (9) Halgren, T. A. Identifying and Characterizing Binding Sites and Assessing Druggability. J. Chem. Inf. Model. 2009, 49, 377−389. (10) Beuming, T.; Che, Y.; Abel, R.; Kim, B.; Shanmugasundaram, V.; Sherman, W. Thermodynamic Analysis of Water Molecules at the Surface of Proteins and Applications to Binding Site Prediction and Characterization. Proteins: Struct., Funct., Genet. 2012, 80, 871−883. (11) Vukovic, S.; Brennan, P. E.; Huggins, D. J. Exploring the Role of Water in Molecular Recognition: Predicting Protein Ligandability Using a Combinatorial Search of Surface Hydration Sites. J. Phys.: Condens. Matter 2016, 28, 344007. (12) Ghanakota, P.; Carlson, H. A. Driving Structure-Based Drug Discovery through Cosolvent Molecular Dynamics. J. Med. Chem. 2016, 59, 10383−10399. (13) Kimura, S. R.; Hu, H. P.; Ruvinsky, A. M.; Sherman, W.; Favia, A. D. Deciphering Cryptic Binding Sites on Proteins by Mixed-Solvent Molecular Dynamics. J. Chem. Inf. Model. 2017, 57, 1388−1401. (14) Oleinikovas, V.; Saladino, G.; Cossins, B. P.; Gervasio, F. L. Understanding Cryptic Pocket Formation in Protein Targets by Enhanced Sampling Simulations. J. Am. Chem. Soc. 2016, 138, 14257−14263. (15) Seco, J.; Luque, F. J.; Barril, X. Binding Site Detection and Druggability Index from First Principles. J. Med. Chem. 2009, 52, 2363− 2371. (16) Guvench, O.; MacKerell, A. D., Jr Computational FragmentBased Binding Site Identification by Ligand Competitive Saturation. PLoS Comput. Biol. 2009, 5, e1000435. (17) Lexa, K. W.; Carlson, H. A. Full Protein Flexibility Is Essential for Proper Hot-Spot Mapping. J. Am. Chem. Soc. 2011, 133, 200−202. ́ (18) Tan, Y. S.; Sledź , P.; Lang, S.; Stubbs, C. J.; Spring, D. R.; Abell, C.; Best, R. B. Using Ligand-Mapping Simulations to Design a Ligand Selectively Targeting a Cryptic Surface Pocket of Polo-Like Kinase 1. Angew. Chem. 2012, 124, 10225−10228. (19) Tan, Y. S.; Spring, D. R.; Abell, C.; Verma, C. S. The Application of Ligand-Mapping Molecular Dynamics Simulations to the Rational

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.jcim.7b00487. MSMD and SiteMap results for the remaining targets; PPI conformations used for all the SiteMap calculations, along with average volume of the MSMD cosolvents (PDF)

■

Article

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Phone: +1 212 295 5800. Fax: +1 212 295 5801. ORCID

Phani Ghanakota: 0000-0002-2249-3681 Present Address †

Silicon Therapeutics, 300 A Street, Boston, MA 02210.

Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS We would like to thank Thomas Holder for valuable discussions on the use of PyMOL. 792


Article

Journal of Chemical Information and Modeling Design of Peptidic Modulators of Protein−Protein Interactions. J. Chem. Theory Comput. 2015, 11, 3199−3210. (20) Higueruelo, A. P.; Jubb, H.; Blundell, T. L. TIMBAL v2: Update of a Database Holding Small Molecules Modulating Protein−Protein Interactions. Database 2013, 2013, bat039. (21) Lexa, K. W.; Goh, G. B.; Carlson, H. A. Parameter Choice Matters: Validating Probe Parameters for Use in Mixed-Solvent Simulations. J. Chem. Inf. Model. 2014, 54, 2190−2199. (22) Ghanakota, P.; Carlson, H. A. Moving Beyond Active-Site Detection: MixMD Applied to Allosteric Systems. J. Phys. Chem. B 2016, 120, 8685−8695. (23) Lexa, K. W.; Carlson, H. A. Improving Protocols for Protein Mapping through Proper Comparison to Crystallography Data. J. Chem. Inf. Model. 2013, 53, 391−402. (24) Ung, P. M. U.; Ghanakota, P.; Graham, S. E.; Lexa, K. W.; Carlson, H. A. Identifying Binding Hot Spots on Protein Surfaces by MixedSolvent Molecular Dynamics: HIV-1 Protease as a Test Case. Biopolymers 2016, 105, 21−34. (25) Schrödinger Release 2015-2: Maestro, Version 10.2; Schrödinger, LLC: New York, NY, 2015. (26) Schrödinger Release 2015-2: Schrödinger Suite 2016-2 Protein Preparation Wizard; Epik Version 3.2, Impact Version 6.7, Prime Version 4.0; Schrödinger, LLC: New York, NY, 2015. (27) Bowers, K. J.; Chow, E.; Xu, H.; Dror, R. O.; Eastwood, M. P.; Gregersen, B. A.; Klepeis, J. L.; Kolossvary, I.; Moraes, M. A.; Sacerdoti, F. D.; Salmon, J. K.; Shan, Y.; Shaw, D. E. Scalable Algorithms for Molecular Dynamics Simulations on Commodity Clusters. Proc. CM/ IEEE Conf. Supercomputi. (SC06) 2006, 84. (28) Jorgensen, W. L.; Chandrasekhar, J.; Madura, J. D.; Impey, R. W.; Klein, M. L. Comparison of Simple Potential Functions for Simulating Liquid Water. J. Chem. Phys. 1983, 79, 926−935. (29) Harder, E.; Damm, W.; Maple, J.; Wu, C.; Reboul, M.; Xiang, J. Y.; Wang, L.; Lupyan, D.; Dahlgren, M. K.; Knight, J. L.; Kaus, J. W.; Cerutti, D. S.; Krilov, G.; Jorgensen, W. L.; Abel, R.; Friesner, R. A. OPLS3: A Force Field Providing Broad Coverage of Drug-like Small Molecules and Proteins. J. Chem. Theory Comput. 2016, 12, 281−296. (30) Shivakumar, D.; Harder, E.; Damm, W.; Friesner, R. A.; Sherman, W. Improving the Prediction of Absolute Solvation Free Energies Using the Next Generation OPLS Force Field. J. Chem. Theory Comput. 2012, 8, 2553−2558. (31) Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235−242. (32) Allen, K. N.; Bellamacina, C. R.; Ding, X.; Jeffery, C. J.; Mattos, C.; Petsko, G. A.; Ringe, D. An Experimental Approach to Mapping the Binding Surfaces of Crystalline Proteins. J. Phys. Chem. 1996, 100, 2605−2611. (33) Mattos, C.; Ringe, D. Locating and Characterizing Binding Sites on Proteins. Nat. Biotechnol. 1996, 14, 595−599. (34) Ringe, D.; Mattos, C. Analysis of the Binding Surfaces of Proteins. Med. Res. Rev. 1999, 19, 321−331. (35) Mattos, C.; Bellamacina, C. R.; Peisach, E.; Pereira, A.; Vitkup, D.; Petsko, G. A.; Ringe, D. Multiple Solvent Crystal Structures: Probing Binding Sites, Plasticity and Hydration. J. Mol. Biol. 2006, 357, 1471− 1482. (36) Vidler, L. R.; Brown, N.; Knapp, S.; Hoelder, S. Druggability Analysis and Structural Classification of Bromodomain Acetyl-Lysine Binding Sites. J. Med. Chem. 2012, 55, 7346−7359. (37) Vassilev, L. T.; Vu, B. T.; Graves, B.; Carvajal, D.; Podlaski, F.; Filipovic, Z.; Kong, N.; Kammlott, U.; Lukacs, C.; Klein, C.; Fotouhi, N.; Liu, E. A. In Vivo Activation of the P53 Pathway by Small-Molecule Antagonists of MDM2. Science 2004, 303, 844−848. (38) Park, J.-E.; Kim, T.-S.; Meng, L.; Bang, J. K.; Kim, B. Y.; Lee, K. S. Putting a Bit into the Polo-Box Domain of Polo-like Kinase 1. J. Anal. Sci. Technol. 2015, DOI: 10.1186/s40543-015-0069-y. (39) Berg, A.; Berg, T. Inhibitors of the Polo-Box Domain of Polo-Like Kinase 1. ChemBioChem 2016, 17, 650−656. (40) Scharow, A.; Raab, M.; Saxena, K.; Sreeramulu, S.; Kudlinzki, D.; Gande, S.; Dötsch, C.; Kurunci-Csacsko, E.; Klaeger, S.; Kuster, B.;

Schwalbe, H.; Strebhardt, K.; Berg, T. Optimized Plk1 PBD Inhibitors Based on Poloxin Induce Mitotic Arrest and Apoptosis in Tumor Cells. ACS Chem. Biol. 2015, 10, 2570−2579. (41) Kim, J. H.; Shim, J.; Ji, M.-J.; Jung, Y.; Bong, S. M.; Jang, Y.-J.; Yoon, E.-K.; Lee, S.-J.; Kim, K. G.; Kim, Y. H.; Lee, C.; Lee, B. I.; Kim, K. T. The Condensin Component NCAPG2 Regulates Microtubule− kinetochore Attachment through Recruitment of Polo-like Kinase 1 to Kinetochores. Nat. Commun. 2014, 5, 4588. (42) Cox, D.; Brennan, M.; Moran, N. Integrins as Therapeutic Targets: Lessons and Opportunities. Nat. Rev. Drug Discovery 2010, 9, 804−820. (43) Ley, K.; Rivera-Nieves, J.; Sandborn, W. J.; Shattil, S. IntegrinBased Therapeutics: Biological Basis, Clinical Use and New Drugs. Nat. Rev. Drug Discovery 2016, 15, 173−183. (44) Coller, B. S. A New Murine Monoclonal Antibody Reports an Activation-Dependent Change in the Conformation and/or Microenvironment of the Platelet Glycoprotein IIb/IIIa Complex. J. Clin. Invest. 1985, 76, 101−108. (45) Use of a Monoclonal Antibody Directed against the Platelet Glycoprotein IIb/IIIa Receptor in High-Risk Coronary Angioplasty. N. Engl. J. Med. 1994, 330, 956−961; DOI: 10.1056/ NEJM199404073301402. (46) Phillips, D. R.; Scarborough, R. M. Clinical Pharmacology of Eptifibatide. Am. J. Cardiol. 1997, 80, 11B−20B. (47) Straub, A.; Azevedo, R.; Beierlein, W.; Wendel, H.; Dietz, K.; Ziemer, G. Tirofiban (Aggrastat®) Protects Platelets and Decreases Platelet-Granulocyte Binding in an Extracorporeal Circulation Model. Thorac. Cardiovasc. Surg. 2006, 54, 162−167. (48) Hartman, G. D.; Egbertson, M. S.; Halczenko, W.; Laswell, W. L.; Duggan, M. E.; Smith, R. L.; Naylor, A. M.; Manno, P. D.; Lynch, R. J. Non-Peptide Fibrinogen Receptor Antagonists. 1. Discovery and Design of Exosite Inhibitors. J. Med. Chem. 1992, 35, 4640−4642. (49) Li, J.; Vootukuri, S.; Shang, Y.; Negri, A.; Jiang, J.; Nedelman, M.; Diacovo, T. G.; Filizola, M.; Thomas, C. J.; Coller, B. S. RUC-4 A Novel ΑIIbβ3 Antagonist for Prehospital Therapy of Myocardial Infarction. Arterioscler., Thromb., Vasc. Biol. 2014, 34, 2321−2329. (50) Thanos, C. D.; Randal, M.; Wells, J. A. Potent Small-Molecule Binding to a Dynamic Hot Spot on IL-2. J. Am. Chem. Soc. 2003, 125, 15280−15281. (51) Osguthorpe, D. J.; Sherman, W.; Hagler, A. T. Exploring Protein Flexibility: Incorporating Structural Ensembles From Crystal Structures and Simulation into Virtual Screening Protocols. J. Phys. Chem. B 2012, 116, 6952−6959. (52) Osguthorpe, D. J.; Sherman, W.; Hagler, A. T. Generation of Receptor Structural Ensembles for Virtual Screening Using Binding Site Shape Analysis and Clustering. Chem. Biol. Drug Des. 2012, 80, 182− 193.

793


Large-Scale Validation of Mixed-Solvent Simulations to Assess

Recommend Documents