Mapping the binding sites of challenging drug targets

Abstract

An increasing number of medically important proteins are challenging drug targets because their binding sites are too shallow or too polar, are cryptic and thus not detectable without a bound ligand or located in a protein-protein interface. While such proteins may not bind druglike small molecules with sufficiently high affinity, they are frequently druggable using novel therapeutic modalities. The need for such modalities can be determined by experimental or computational fragment based methods. Computational mapping by mixed solvent molecular dynamics simulations or the FTMap server can be used to determine binding hot spots. The strength and location of the hot spots provide very useful information for selecting potentially successful approaches to drug discovery.

Introduction

Genome-scale CRISPR knockout screens can discover many novel and medically important drug targets [1], but it is predicted that traditional small-molecule drugs may not be used to modulate about half of these proteins [2], because they have binding sites that are either too large or too small, are highly lipophilic or highly polar, or are simply featureless. Given the properties of the binding site, one could frequently predict that the standard methods of drug discovery by experimental or computational high throughput screening of libraries of druglike small molecules are unlikely to work. Nevertheless, in the recent past, substantial efforts have been devoted to large-scale screenings for some targets with at most moderate success. Examples include targeting ZipA pockets in the interface with FtsZ [3] and the SI/II pocket between switch I and switch II of KRAS in the interface with SOS [4,5]. Although such targets are frequently considered undruggable [5,6], in some cases they can be successfully modulated by new chemical modalities including larger (beyond-the-rule-of-five, bRo5) compounds [7,8], macrocycles [9,10], cyclic or stapled peptides, or peptoid macrocycles [11,12]. Other possible approaches are finding allosteric sites [13], covalent inhibitors [14], or combinations of the two [14].

To determine whether a particular target needs a non-druglike chemical modality and if it does, which one, it is generally useful to determine the binding properties of the protein, particularly the geometry and chemistry of its binding sites. It is now well established that the binding sites of proteins include binding hot spots, defined as small regions where binding of ligands makes major contributions to the binding free energy [15,16]. As we argue in this review, the main value of understanding the hot spot structure of a target protein is that it yields information on the methods that are reasonable choices to target the site [17]. Hot spots can be determined by screening sets of small organic probe molecules for binding to the target protein by X-ray crystallography [15] or NMR [16]. The Multiple Solvent Crystal Structures (MSCS) method, involves determining the X-ray structure of the protein in aqueous solutions of various probe compounds [15]. The protein structures with bound organic molecules are then superimposed to derive a consensus X-ray structure. It was shown that the consensus clusters formed by overlapping probe clusters define consensus sites that are the binding hot spots. Similar results can be obtained by NMR based screening of small organic molecules against the ¹⁵N-labeled target protein is screened [16]. It was shown that the consensus clusters formed by multiple probe molecules indicate binding hot spots, and that the number of different probes in the consensus cluster predicts the importance of the site.

Computational mapping of protein binding sites

Since using experimental techniques for determining binding hot spots is generally costly and can even be limited by physical constraints such as limits on the solubility of probe molecules, several computational methods have also been developed. The FTMap algorithm [17] and the mixed solvent molecular dynamics (MSMD) approach [18-21] are both computational analogs of the MSCS or NMR based fragment screening experiments. FTMap exhaustively docks the molecular probes to the protein exploring billions of positions for each probe, selects favorable positions using empirical energy functions, and refines the selected poses by minimizing a more accurate energy function that includes molecular mechanics and structure-based terms. The energy landscape is efficiently sampled using a fast Fourier transform (FFT) based algorithm. The selected probe positions are refined by accounting for probe and limited protein flexibility. To determine the hot spots FTMap finds the consensus sites and ranks the strength of these sites in terms of the number of overlapping probe clusters [17]. The strength and arrangement of hot spots show whether the protein is suitable for binding small druglike ligands, or it is a challenging target and hence needs other type of modalities [17].

MSMD is an alternative hot spot mapping technique based on molecular dynamics (MD) simulations of proteins in binary solvent mixtures. Similar to FTMap, the technique can capture preferred binding sites of fragment-sized organic compounds. The best known methods are MixMD [19] and SILCS (Site-Identification by Ligand Competitive Saturation) [22]. Advantages are that MSMD allows for protein flexibility and accounts for the competition between the probe molecules and water. In contrast, apart from minor side chain motion, FTMap assumes a rigid protein and uses continuum solvation models, thereby missing specific protein-water and probe-water interactions. However, the advantage of FTMap is that the method is much faster than mixed MD, and therefore can be used with a much larger variety of molecular probes and can be applied to large sets of proteins. In particular, it is frequently useful to consider all X-ray structures available for a protein to explore the impact of large conformational changes that would be difficult to model using MD.

Detecting the need for beyond rule of five (bRo5) compounds

Lipinski’s rule of five (Ro5) was developed to define the chemical space of orally bioavailable compounds. However, the concept is too restrictive [23], as over 30% of approved kinase inhibitors and around 50% of protein-protein inhibitors discussed in the scientific literature are beyond the rule of five (bRo5) compounds [23]. The need for using a bRo5 compound to target a protein can be effectively determined by mapping of the binding hot spots [8]. Targets can benefit from bRo5 drugs if they have complex hot spot structures with four or more binding hots spots, including some strong ones. Although such targets are conventionally druggable using molecules that are bRo5 compliant, reaching additional hot spots improves binding affinity, which creates options for improving pharmaceutical properties by adding or replacing some functional groups that otherwise would be detrimental to binding. For example, the only FDA approved nonpeptidic direct thrombin inhibitor Argatroban extends to all five hot spots in the binding site and has a molecular weight of just over 500 Da [8]. Although some lower molecular weight thrombin inhibitors also have high affinity, they turned out to have problems, including but not limited to poor selectivity, weak oral bioavailability, poor metabolic stability, innate liver toxicity, rapid elimination from the blood, high-plasma protein binding, and low anticoagulant activity. Therefore, it may be reasonable to consider as many hot spots as possible in drug design, despite the increase in molecular weight. Many protein kinases also have multiple strong hot spots, but bRo5 inhibitors were generally designed to improve selectivity rather than affinity [8]. Interestingly, targets that have simple hot spot structures with less than four hot spots that are too weak to provide conventional druggability also must use larger compounds that can form interactions with surfaces outside the hot spot region to reach acceptable affinity [8]. More recent studies focus on the pharmaceutical properties of novel bRo5 modalities such as peptidomimetics, with particular emphasis on membrane permeability [24].

Identification of protein-protein inhibitor binding sites

Intercellular protein-protein interaction (PPI) interfaces are challenging targets because the cavities available for binding druglike molecules are generally less defined than the pockets of traditional drug target proteins [25]. In addition, ligand binding may depend on the flexibility of the pocket, therefore potential conformational changes must be considered. Fragment based methods have been important for developing PPI inhibitors [26]. Computational fragment screening by FTMap [27] and SILCS [22] was shown to identify binding hot spots amenable to inhibitor binding based on mapping the structures of the interacting proteins. A more complex method involving molecular dynamics simulations and protein docking gave similar results [28]. Frequently, the hot spot residues in protein-protein interfaces, identified by alanine scanning, extend into binding hot spots of the partner protein, thus the two hot spot concepts are related [29]. Many recent studies search for hot spots residues to find targetable sites [30], primarily with application to cancer [31]. Based on the outcome of drug discovery campaigns, it appears that high affinity inhibitors bind to pockets that are at least partially formed in the protein-protein complex [27], and the tractability of such sites can be reliably determined by mapping either unbound or protein-bound structures [27,28]. In many cases the protein interacting with the target can be reduced to a peptide, most frequently an alpha-helix [32], but beta-turn structures also occur [33]. These secondary structures can be then stabilized to form cyclic peptides or peptidomimetic inhibitors [32].

Searching for allosteric sites

In some medically important proteins targeting the main orthosteric site is challenging, because the site may not provide sufficient selectivity and the inhibitor must compete with the endogenous ligands. For example, high affinity active site inhibitors of tyrosine phosphatases would need to emulate the charged nature of the phosphorylated substrate and achieving selectivity may require fairly large compounds due to the similarity of residues directly surrounding the site [34]. For such targets, allosteric drugs may provide a critical advantage due to non-competitive and highly specific regulation [13]. Identification of allosteric sites involves two aspects, first finding an appropriate site, and second showing allosteric communication to the orthosteric site. We restrict consideration to the first aspect as it appears to be critical, and don’t discuss specialized algorithms such as Allosite [35] and AlloFinder [36].

The mapping methods SILCS [21], MixMD [18,37], and FTMap [38,39] were all used to identify allosteric binding sites. SILCS was shown to detect more potential sites than FTMap, but several such sites appear to be false positives [21]. Most applications focused on kinases and GPCRs, two important target families whose allosteric sites have been extensively studied. Although the majority of currently approved kinase inhibitors target the ATP binding site, there is substantial interest in allosteric sites and allosteric drugs [40]. While type II and type III allosteric inhibitors bind at or near the ATP binding site, the literature identifies ten regions that have been reported as regulatory hot spots and are therefore potential target sites for type IV inhibitors. Kinase Atlas, a collection of binding hot spots located at each of the ten allosteric sites was constructed using the FTMap results for all kinase structures in the PDB. Kinase Atlas https://kinase-atlas.bu.edu) displays summarized results including the presence of binding sites and their druggability for all structures of a particular kinase. Additionally, users may view hot spot information for individual kinase structures [38].

Allosteric modulators represent a very important strategy against GPCR targets. Despite the growing number of GPCR structures, only 39 have been co-crystallized with allosteric inhibitors, and thus identification of allosteric sites is important. FTMap has been applied to several GPCRs by the McCammon group [41,42]. More recently the method was shown to successfully identify allosteric sites within the seven-membrane region of GPCRs [39]. However, FTMap is parameterized for analysis of soluble proteins and may fail to identify allosteric sites in receptor-lipid interfaces. A recent probe confined dynamic mapping protocol developed for GPCRs predicts the location of allosteric sites at both intracellular and extracellular regions and within the receptor-lipid interface [43]. The method enhances sampling of probe molecules within a defined region of a GPCR and prevents membrane distortion during molecular dynamics simulations by applying a harmonic wall potential. In addition, the method uses a set of probes derived from structures of GPCR allosteric ligands [43]. Another recent study used exhaustive docking of small molecular probes, considering the different electrostatics of the transmembrane and solvent-exposed parts of the receptors, resulting in the “pocketome” of G protein-coupled receptors [44].

How useful are cryptic sites for drug discovery?

Some proteins have binding sites that are difficult to detect in ligand-free structures and only become apparent after ligand binding [45]. This is frequently the case for allosteric sites. Attempts to find alternative ways to drug challenging targets lead to the development of computational methods for the identification and analysis of such cryptic sites [45,46]. Cimermancic et al. [46] created a benchmark set of 93 ligand-free and ligand-bound pairs of proteins from the PDB with cryptic binding sites which they also used to build a machine learning model (CryptoSite) to predict cryptic sites in the apo structures. The original CryptoSite data set was expanded by Beglov et al. [47] by adding all ligand-free structures in the PDB for each of the 93 proteins. Mapping of apo structures by FTMap revealed that cryptic binding sites are generally located near a strong binding hot spot and that the sites exhibit above-average flexibility [47]. While the FTMap results were in good agreement with those of CryptoSite, both methods account only for the limited flexibility of the proteins. There is no question that more realistic simulations that reveal the multiplicity of potential conformational states help to identify cryptic sites. MixMD was able to identify cryptic and allosteric sites in ligand-free structures of some proteins that were not found by FTMap [19,37].

Markov state models, MSMD simulations, and collective variable enhanced sampling methods were shown to open transitional pockets [48,49]. The problem is that the number of pockets that are open in more than 10% of the simulation time can be very high [50]. However, based on FTMap results, proteins generally have only three different sites with substantial ligand binding capability [47], and hence most of the newly created pockets are too weak for drug discovery. Results supporting this observation were reported by Bowman et al. [51], who identified multiple hidden allosteric sites in TEM-β lactamase using Markov state models. Although small compounds covalently bound at some of these sites were shown to have an allosteric effect [51], non-covalent modulators designed for the site had only moderate impact [52]. In particular, we have observed that pockets that open solely by the movement of some side chains can bind ligands with at most high micromolar affinity [47], most likely because the side chains protruding into the site compete with the ligands for binding. Since the side chains are always at the site, their local concentration is very high, resulting in substantial competition.

An example: Mapping of KRAS

We show that mapping of KRAS structures provides valuable information on the relative tractability of the binding sites. Figure 1a shows the results of mapping a KRAS structure (PDB ID 6MBT [53]), which has no bound ligand apart from ADP and Mg²⁺ that are removed before the mapping. Three ligands superimposed from bound structures are added in Figure 1b. The strongest consensus site that includes 36 probe clusters binds the GDP molecule. The second strongest consensus site is located close to residue 12, in a pocket that accommodates the covalent G12C inhibitor AMG 510 in the structure 6OIM [54]. The site binds 17 probe clusters suggesting limited druggability and the need for a covalent drug [55]. Finally, the third consensus site is in the extensively targeted shallow polar pocket between switch I and switch II (SI/II pocket) in the KRAS-SOS interface. This consensus site includes only 9 probe clusters, which suggests that the site is too weak to bind druglike molecules with high affinity [55]. In fact, the site binds the small inhibitor, developed by the Fesik group in 2012 with only K_d = 420 μM [4]. This inhibitor was co-crystallized with KRAS, and the resulting structure 4EPW was also mapped by FTMap after removing the ligands. While the strongest consensus site is still at the GDP binding pocket (Figure 1c), the binding of the inhibitor slightly expands the SI/II pocket, which now binds 16 probe clusters. In addition, ligand binding induces a second hot spot with 10 probe clusters in the SI/II pocket (Figure 1c), and the inhibitor binds to both hot spots (Figure 1d). In collaboration with Boehringer Ingelheim, the Fesik group recently developed a larger inhibitor (MW = 512 g/mol) that binds to the SI/II pocket with Kd = 750 nM [5]. Given the limited druggability of the SI/II site [55] this is an extraordinary achievement, and the compound, shown in Figure 1b can be used as a chemical probe. However, based on the mapping results it is unlikely that any small druglike compound binding at the SI/II site can be developed into a drug, emphasizing the importance of the covalent allosteric inhibitors binding at the G12C site [54,56]. Interestingly, the binding of an inhibitor at the SI/II pocket weakens the hot spot at the allosteric site that binds the covalent inhibitor, suggesting bidirectional allosteric communication between the SI/II pocket and the G12C site.

Figure 1.

Mapping of KRAS. (a) Hot spots on the wild-type KRAS bound to GDP and Mg²⁺ (PDB ID 6MBT). All ligands are removed prior to mapping. Only probes at cluster centers are shown, represented as lines. The consensus sites define the binding hot spots. The most important consensus site (cyan) includes 36 probe clusters. The second site (blue) and the third site (magenta) bind 17 and 9 probe clusters, respectively. (b) Same as (a) with ligand superimposed from bound structures. The ligands are shown as sticks. The GDP molecule (orange) binds at the strongest consensus site. The covalent inhibitor AMG 510 bound to G12C from PDB ID 6OIM (blue) binds at the second strongest site [54]. The third site (with 9 probe clusters) is in the SI/II pocket and binds both the inhibitor developed by the Fesik group in 2012 (K_d = 420 μM from PDB ID 4EPW, yellow) [4] and the more recent direct inhibitor in the same pocket (K_d = 750 nM from PDB ID 6GJ7, green) [5]. (c) Mapping the KRAS structure co-crystallized with the low affinity inhibitor (PDB ID 4EPW) following the removal of the inhibitor. The most important consensus site (cyan) now includes only 22 probe clusters. The second consensus site (orange) and the third one (magenta) bind 16 and 10 probe clusters, respectively. The fourth site (blue) also has 10 probe clusters. (d) The top site (cyan) still binds the GDP as in (b). However, the SI/II pocket now includes the second and third consensus sites, both interacting with the inhibitor from 4EPW, shown as yellow sticks. The fourth consensus site is located at the binding site of the covalent inhibitor at KRAS G12C.

We also mapped the 282 KRAS structures in the PDB with < 90% sequence identity to 4EPW and determined the consensus clusters formed by all mapping results. Interestingly the large-scale mapping provided the same top sites obtained by mapping only the unbound structure 6MBT and the inhibitor bound structure 4EPW. The strongest consensus site, located at the GDP binding site, on average binds 17.4 ± 6.1 probe clusters. The second consensus site is in the SI/II pocket, with 8.8 ± 6.8 probe clusters, and the third consensus site is at the pocket that binds the G12C inhibitors and includes 8.6 ± 6.1 probe clusters. These results reveal that the SI/II pockets and the G12C site have almost the same strength, but both have limited druggability [55]. Therefore. the ability of AMG 510 to covalently bind to the cystine residue at position 12 in the G12C mutant is crucial. Unfortunately, the weakness of the site implies that extending the approach to other G12 oncogenic mutants will be very challenging.