Expanding the Conformational Selection Paradigm in Protein-Ligand Docking

Abstract

Conformational selection emerges as a theme in macromolecular interactions. Data validate it as a prevailing mechanism in protein-protein, protein-DNA, protein-RNA, and protein-small molecule drug recognition. This raises the question of whether this fundamental biomolecular binding mechanism can be used to improve drug docking and discovery. Actually, in practice this has already been taking place for some years in increasing numbers. Essentially, it argues for using not a single conformer, but an ensemble. The paradigm of conformational selection holds that because the ensemble is heterogeneous, within it there will be states whose conformation matches that of the ligand. Even if the population of this state is low, since it is favorable for binding the ligand, it will bind to it with a subsequent population shift toward this conformer. Here we suggest expanding it by first modeling all protein interactions in the cell by using Prism, an efficient motif-based protein-protein interaction modeling strategy, followed by ensemble generation. Such a strategy could be particularly useful for signaling proteins, which are major targets in drug discovery and bind multiple partners through a shared binding site, each with some—minor or major—conformational change.

1. Introduction

Proteins are involved in all molecular processes in living cells including metabolic, signaling, catalysis, viral entry, and regulation; cellular dysfunction due to inhibition, or to nonnative interactions of proteins can cause diseases (1, 2). Understanding the molecular and cellular activities in vivo and controlling their functions in disease requires analyzing the proteins, investigating their interactions, and elucidating their functions. Identifying protein interactions is important not only to understand how cells work, but also to elucidate disease mechanisms, discover effective drugs and figure out their effects on the entire cellular network (3, 4) to forecast side effects. Several experimental techniques (5), such as the yeast two-hybrid system (6), phage display (7), protein arrays (8), and affinity purification (9), generate massive amounts of protein interaction data. Yet, despite these, the complex nature of protein interactions is not entirely understood (10). As more data become available, computational methods which are able to analyze the large datasets are becoming increasingly important to make sense of experimental observations and use them to predict additional interactions, functional mechanisms, and protein and drug design.

Computational structural biology aims to introduce and apply effective methods that predict not only which proteins interact but also how they interact. Predictions of protein interaction can be carried out using docking or knowledge-based approaches. Although docking approaches are broadly used and are effective strategies, they cannot be applied on proteomic scales. The computation times are prohibitively long, and in particular, for reliable docking, additional biochemical data such as mutational information about protein interactions should be provided; in their absence, the number of false positive solutions can be astronomical and it is very difficult to distinguish between native and nonnative predictions (11). Knowledge-based approaches are faster compared to blind docking methods. Because they decrease the solution space by limiting possible orientations, the number of potential interactions is smaller which also leads to relatively shorter timescales. This enables knowledge-based methods to cope with large sets of data. In knowledge-based approaches, templates derived from known interacting proteins can be sequence-based (12–14), domain-based (15) or interface-based (16, 17). It has been widely accepted that the structure of the protein is evolutionarily more conserved than the sequence (18). Thus, in principle, prediction algorithms which are purely structure-based, where the methodology is completely independent from any sequence homology, can work; and this holds even in the absence of any sequence similarity. This is all the more so for protein interfaces, which are often more conserved than the overall structure (19). Analysis of the interfaces has shown that even if the global structures and functions differ, proteins can bind through similar interface architectures (20, 21). A structurally nonredundant dataset of protein-protein interfaces can be clustered into three types of groups according to the interface and global structures of the interacting protein pairs (see Fig. 1) (20, 22, 23): in Type I the interacting proteins have similar global structures and functions. This is the most common and expected type. In Type II cluster members have similar interfaces; however, the global structures and functions are different. This type contains examples that validate the paradigm that interface motifs can be conserved even in the absence of global structural similarity (24, 25). In Type III, only one side of the interface is similar and the surfaces of the complementary partners are somewhat different. Hub proteins are mostly clustered into this type; therefore, members of this cluster may help in the characterization of hub proteins and shared binding sites (23). From an energetic point of view, a subset of interface residues can act as “hot spots” (26). These residues contribute more to the binding free energy of complexes; that is, they play a more significant role in the affinity and stability of the interaction. There is a strong correlation between hot spots and conserved residues on structurally similar interfaces (27), which points to the importance of hot spots in determining binding sites. Since hot spots contribute most of the binding energy in the interaction, discovery of molecules that bind to hot spots (1, 28), which can be small molecule drugs (2, 29, 30) or inhibitory peptides (31–33), has gained importance in drug design.

Fig. 1.

Examples of Type I, II, and III interfaces. The interfaces are highlighted with boxes. (a) Members of Type I proteins use similar interfaces to bind each other. The two glutathione S-transferase complexes are homologous (PDB identifiers: 10gs and 1b48). (b) Members of Type I proteins are not related evolutionarily, but the interface structures are similar. The two complexes, cytochrome C and neuropeptide/membrane protein are examples of this type (PDB identifiers: 1bbh and 1rso). (c) In Type III, only one side of the interface has similar architectures, the complementary sides are different (dynein light chain 8, PDB identifier: 1f95AB; 4-oxalocrotonate tautomerase, PDB identifier: 1 otfAE).

Understanding the mechanism of binding is expected to help drug discovery, since it can lead to more effective methodologies. Over the years, Koshland’s “induced fit” scenario (34) has been widely accepted as the binding mechanism. According to the induced fit, binding of a protein to a ligand leads to a conformational change in the protein which is “induced” by the ligand and culminates in a favorable, tight fit. More recently, an alternative mechanism has been proposed, the so-called “conformational selection and population shift” (35–39). This proposition has been based on concepts derived from the free energy landscape (40). It argued that since proteins exist in solution in broad ensembles, among the conformational states present in the ensemble there should be some with binding sites matching the shape (and chemistry) of the ligand. While the energy of these states can be high, and thus they may be only sparsely populated, the binding will stabilize them, with a subsequent “population shift” toward these conformers, which maintains the chemical equilibrium. Recently, considerable experimental and computational data have accumulated (41–43) validating the conformational selection and population shift scenario for a broad range of binding events, and it has further been proposed to apply to drug discovery (44). Currently, conformational selection is believed to be the prevailing mechanism, with induced fit dominating in cases where the concentration of the ligand is extremely high (45). Of note, the timescales of induced fit are faster than those of conformational selection and population shift; this is because a shift in the population necessitates climbing barriers, and thus the times depend on the barrier heights. Following binding, there is an induced fit on a minor, local scale to optimize the interactions. The question arises in which way such a mechanistic scenario can help in drug discovery strategies. A reasonable way would be to generate an ensemble of states, and dock these separately to the small molecule drug. However this is an immensely complex task, since it critically depends on the sampling. Since high energy states also need to be considered, the sampling should not be confined to low energy conformations. Drug discovery is usually aimed at enzyme active sites; however, increasingly it also targets disruption or modulation of protein–protein binding sites. While enzyme active sites are known, this is not the case for the protein binding sites, where as we discussed above, data are available only for a (relatively small) fraction of the interactions. For these cases, combining prediction of protein-protein interactions and their binding sites as a first step, coupled with ensemble docking could be a strategy to consider.

Toward such a strategy, here we present a template-based protein-protein interaction algorithm, Prism (Protein Interactions by Structural Matching) (46, 47, 78) integrated with FiberDock (Flexible Induced-fit Backbone Refinement in Molecular Docking) (48). The Prism algorithm reveals possible interactions among a group of protein structures based on known protein–protein interfaces. Due to the existence of a limited number of distinct binding motifs in nature (49), similar interface architectures are shared among functionally and structurally different proteins (20). The method, which is independent of sequence data, utilizes structural and evolutionary similarity of a target protein with partners of an already known interaction to predict an interaction between two protein molecules. Although the structural similarity is detected via geometrical alignment of structures, evolutionary similarity is approximated by the conservation of hot spots. Besides the efficiency in prediction of protein interactions on the proteome scale, the prediction algorithm can be used to construct and analyze specific networks, such as the human cancer protein-protein interaction network (50), or to discover shared binding sites in hub proteins (51). Furthermore, increasing interest in targeting protein-protein interactions (52, 53), especially hot spots in interfaces (54), for drug discovery makes such a strategy particularly promising. Combining Prism with FiberDock is a powerful alternative to guide pharmacological research considering its ability to detect a potential interaction between a drug and its target protein or of a target protein with another protein in the network. Moreover, because the interacting residues can be sequentially discontinuous (see Fig. 2), an algorithm such as Prism which focuses on interfaces and is independent of the order of the residues on the chain is advantageous.

Fig. 2.

A two-chain interface (a) An example of a two-chain interface (PDB identifier: 1fq3; chains A and B). Black residues represent contacting residues which interact across the interface. Residues in their spatial vicinity (called nearby residues) are in whitish gray. The remaining residues in the chains A and B are shown in gray. (b) The interface consists of bits and pieces of each of the chains, and some isolated residues. The chain A side of the interface consists of five contacting and 24 nearby residues. There are nine contacting and 17 nearby residues in the chain B interface.

2. Materials and Methods

Prism attempts to predict protein-protein interactions based on structural similarity of the proteins to the complementary sides of a known interface. If it is known that there is an interaction between proteins A and B, and protein A′ is structurally similar to protein A and protein B′ is structurally similar to protein B, it is claimed that A′ and B′ may interact with each other (46). Prism considers a potential binary interaction by querying whether target interfaces structurally and evolutionarily complement each other in a way similar to template interfaces. Then, by using FiberDock, flexible refinement of docking solution candidates is performed by optimizing the side chain orientations. Binding energy is also calculated for the refined structures. To carry out such a protocol, the first step involves the availability of target structures and generation of template datasets. A flowchart summarizing the prediction algorithm is given in Fig. 3.

Fig. 3.

Flowchart summary of the prediction algorithm of Prism together with FiberDock.

2.1. Template Dataset

All interfaces of two chain protein complexes available in the Protein Data Bank (55) were extracted. Interfaces consist of interacting residues between two chains and neighboring residues. Neighboring residues are in the spatial vicinity of interacting residues and constitute the scaffold of the interface. Two residues from two different chains are considered as interacting if they are at a distance smaller than the sum of van der Waals radii plus a threshold of 0.5 Á. In addition, a noninteracting residue whose Cα is closer than 6.0 Á to the Cα of any interacting residue is marked as a neighboring residue. In order to obtain a nonredundant dataset, 49,512 two-chain interfaces (as of February 2006) extracted in the first step were clustered structurally following an iterative all-against-all structural comparison in a sequence order-independent way (20, 51). 8,205 clusters were obtained. Interface members of each cluster are structurally similar to the representative interface. A cluster should contain at least five nonhomologous sequences.

The template interface can be constructed in several ways: one can use (1) all representatives (8,205) of the interfaces, or (2) a subset of the representatives, for example, the heterodimeric protein interfaces (1,036), or the nonobligate protein interfaces (158 interfaces) (see Note 1). The type of reduction of the template set is determined with respect to characteristics of the query molecules. Computational hot spots are found by using the HotPoint web server (56) (see Note 2). Prism then searches for a potential interaction by comparing the surfaces of target proteins to the partners of known template interfaces while accounting for evolutionary conservation.

2.2. Target Dataset

Proteins in a target dataset are searched for a potential interaction (see Note 3). The data of query proteins are extracted from the PDB. Multimeric proteins are split into their monomers, and homologous chains are counted only once (see Note 4). The surfaces of the molecules are extracted by using the NACCESS program (described in Subheading 2.3).

2.3. Prediction of Protein-Protein Interaction

Prism suggests a possible interaction between two target proteins A′ and B′, if protein A′ shares structural similarity with one side of template interface I, which is extracted from a known interaction between protein A and protein B, and protein B′ is structurally similar to the other side of the interface I.

The surfaces of target proteins are extracted using the NACCESS program (57) (see Note 5). NACCESS calculates the relative surface accessibilities (RSAs) of residues, which are the percent accessibility with respect to the accessibility of the residue type X in an extended ALA-X-ALA tripeptide (58). Residues whose RSA values are greater than 15% are considered as surface residues. “Nearby” residues are then added to the surface shell as described above, but the threshold value is chosen as 5.0 Á. Structural similarity between target and template interfaces is assessed using MultiProt (59, 60). MultiProt aligns the target surface with each complementary partner of the representative template interfaces and determines the common geometrical cores between structures. MultiProt’s output is the ten best alignments for substructural matching of a target protein surface with a template interface. Target surfaces should geometrically match with 50% of the residues of the template chains if the template chains contain at most 50 residues. This matching threshold is 30% for the larger template chains. In addition, at least one conserved hot spot should be correctly matched between the template interface and the target surface (see Note 6). Moreover, at least five pairs of matched residues from each side of the template interface should be against each other in order to guarantee the correct matching for the left and right partners (see Note 7).

Target proteins which pass the alignment process and match with the partners of the same template interface are next checked if it is physically possible for them to constitute a complex. If the Cα atom of a residue from one partner is at a distance shorter than 3 Á to the Cα atom of a residue from the complementary partner, those two residues are considered as clashing. A threshold of five clashes makes the interaction physically impossible.

Finally, FiberDock (48) is used for flexible refinement of the predicted complexes and for calculation of the energy of the interaction. Steric clashes of side chains due to their orientations are solved via conformational adjustment of the side chains and the binding energy of the final transformed structures is calculated (see Note 8). FiberDock ranks the docked solutions by the calculated energies. Hence, FiberDock checks if a potential interaction estimated by Prism is favorable in terms of global energy.

4. A Drug Target: Insulin Receptor

Mutations in protein kinases contribute to diseases or pathophysiological states, including cancer, autoimmune disorders, cardiac diseases, and inflammatory conditions (61). Therefore, recent effort increasingly focuses on inhibitor and small molecule drug design to modulate these enzymes. The insulin receptor (IR) is a member of the tyrosine kinase receptors. In addition to diabetes, it appears to be related to Alzheimer’s disease and cancer (62–66).

IR exists on the surfaces of cells and interacts with insulin, the hormone having a significant role in regulating the energy and glucose metabolism in the body. Insulin receptor substrate 2 (IRS2) is one of the substrates of IR. The conformational change of the insulin receptor tyrosine kinase domain (IR_TKD) through binding to IRS2 is shown in Fig. 4a. In the figure, the molecular structure of free IR_TKD is transparent black (PDB identifier: 1 irk); black and the whitish gray molecules represent the IR_TKD complex with IRS2 (PDB identifier: 3bu3; black: IR_TKD, chain A; whitish gray: IRS2, chain B). IR is inhibited by growth factor receptor-bound protein 14 (Grbl4) and the molecular structure is given in Fig. 4b (PDB identifier: 2auh; black: IR_TKD, chain A; whitish gray: Grbl4, chain B). The beads shown in Fig. 4a, b represent the computational hot spots of IR_TKD extracted by using the HotPoint web server (56). Dark gray beads (Leu¹¹⁷¹, Val¹¹⁷³ and Gln¹²⁰⁸) are the common hot spots of IR_TKD in receptor/substrate (3bu3) and receptor/inhibitor complexes (2auh). Although the interacting partners are different molecules and a different conformational change is observed following the binding, IR_TKD interacts through the same hot spot residues. Interaction of both the inhibitor and the substrate through the same hot spots indicates the importance of targeting hot spots in drug discovery (53, 67). Several studies have focused on the discovery of small molecules that bind with drug-like potencies to hot spots at the interface (68–70).

Fig. 4.

Insulin receptor tyrosine kinase domain complex with its substrate and its inhibitor (a) Molecular structure of free insulin receptor tyrosine kinase domain (IR_TKD, black in transparent, PDB identifier: 1 irk) and its complex with insulin receptor substrate 2 (IR_TKD/IRS2, black/whitish gray, PDB identifier: 3bu3). Computational hotspots on IR_TKD are shown with ball representation (light gray and dark gray). Dark gray balls are common hotspots of IR_TKD in IR_TKD/IRS2 and IR_TKD/Grb14 (b). (b) Molecular structure of insulin receptor tyrosine kinase domain complex with growth factor receptor-bound protein 14 (IR_TKD/Grb14, black/whitish gray, PDB identifier: 2auh). Computational hotspots on IR_TKD are shown with ball representation (light gray and dark gray). Dark gray balls are common hotspots of IR_TKD in IR_TKD/Grb14 and IR_TKD/IRS2 (a).

5. Discussion and Conclusions

Conformational selection and population shift is currently the accepted paradigm for molecular recognition. The question arises how to use it to improve experimental and computational strategies. Here our focus is on docking. A knowledge-based docking approach such as Prism, which follows a rationale that if a binding site motif is similar between two proteins it is likely to interact with a common motif of a partner protein, implicitly follows the conformational selection concept. As such, it can also be used toward small molecule ligand and peptide docking. As targets, above we focused on protein-protein interfaces. Our approach considers two steps: in the first the pathways are modeled to obtain their protein-protein interfaces. This is because the PDB contains only a small fraction of the interactions. In the second, ensembles would be generated, and candidate drugs would be docked to representatives of the ensemble clusters. Signaling proteins are particularly good targets: they are at the crossroads of pathways and their binding sites can be shared by a large number of partners (54). Drug binding will elicit allosteric effects which not only will change the conformations of their protein-protein binding sites elsewhere, but will also propagate in the pathway.

Ensemble docking has been a strategy long in use, even if for different consideration—to overcome the technical difficulties in flexible docking. A quick literature search produces hundreds of papers devoted to the subject; among these is the work by Lorber and Shoichet (71) which to our knowledge is the first. Conformational selection has also been used directly in docking (72, 73). However, it is difficult to apply this concept on a comprehensive scale. Docking of a large ensemble is currently prohibitive, because of the timescales. Nonetheless, rapid sampling methods (74) perhaps coupled with semiatomistic approaches (75) or effective filters (76) or other useful strategies (77), hopefully will eventually help in this endeavor which mimics real life mechanisms.