Insight

Abstract

Targeting the interfaces between proteins has huge therapeutic potential, but discovering small-molecule drugs that disrupt protein–protein interactions is an enormous challenge. Several recent success stories, however, indicate that protein–protein interfaces might be more tractable than has been thought. These studies discovered small molecules that bind with drug-like potencies to 'hotspots' on the contact surfaces involved in protein–protein interactions. Remarkably, these small molecules bind deeper within the contact surface of the target protein, and bind with much higher efficiencies, than do the contact atoms of the natural protein partner. Some of these small molecules are now making their way through clinical trials, so this high-hanging fruit might not be far out of reach.

There is probably no class of macromolecular interaction that rivals the complexity, diversity and regulatory impact of interactions between proteins^{1, 2, 3, 4, 5, 6}. There is much interest in targeting the interfaces between interacting proteins for therapeutic purposes — ideally with small 'drug-like' molecules, which generally are cheaper and can be administered orally instead of by injection — but the characteristics of these interfaces make this task a huge challenge. The contact surfaces involved in protein–protein interactions are large (1,500–3,000 Ų)^{7, 8} compared with those involved in protein–small-molecule interactions (300–1,000 Ų)^{9, 10}. In addition, the contact surfaces of proteins that interact with other proteins are generally flat and often lack the grooves and pockets present at the surfaces of proteins that bind to small molecules¹¹. Unlike the classic proteins for which small-molecule drugs have been designed (for example, enzymes and G-protein-coupled receptors), protein–protein interactions do not have natural small-molecule partners. Thus, efforts to discover drugs that bind to a protein–protein interface do not have the luxury of starting from a small natural substrate or ligand.

Most contact surfaces in protein–protein interfaces also involve amino-acid residues that are not contiguous in the polymer chain. For this reason, peptides derived from short contiguous sequences at the interface are generally poor chemical starting points. (There are notable exceptions, however, in which a protein displays a contiguous tripeptide or tetrapeptide sequence for which small-molecule peptidomimetics have been assembled; see refs 12 and 13 for examples.) Furthermore, high-throughput screening (HTS) does not routinely identify compounds that disrupt protein–protein interfaces^{14, 15}. And, although biopharmaceuticals such as monoclonal antibodies and polypeptide hormones almost always bind to protein–protein interaction surfaces, there are few approved small-molecule drugs that do so.

Despite these challenges, several lines of evidence provide hope for finding small molecules that target protein–protein interfaces. Although these interfaces are large, mutational studies show that a small subset of the residues involved contributes most of the free energy of binding^{16, 17, 18, 19, 20} (Fig. 1). Such 'hotspots' constitute less than half of the contact surface of a protein involved in the protein–protein interaction and are usually found at the centre of the contact interface. Proteins involved in protein–protein interactions can be 'promiscuous', binding to several targets using the same hotspot region²¹. Structural studies show that these promiscuous contact surfaces are adaptable, allowing one protein to engage a range of structurally diverse partners. Moreover, peptides selected for binding to one of the partners in a protein–protein pair (by using phage display) often compete with the natural protein partner for binding to the hotspot^{20, 21, 22, 23, 24}. Thus, there seem to be many chemical solutions for tight binding, and large contact surfaces can be engaged by more-compact structures.

Figure 1: Examples of protein–protein interface hotspots.

Alanine-scanning mutational analysis (replacing each amino acid, in turn, with alanine) was carried out on the contact surfaces of four pairs of interacting proteins. The resultant change in the binding free energy compared with interactions involving the wild-type protein (G) is shown by colour coding amino-acid residues, from red (the most-disruptive changes) to green (those having little or no effect). Therefore, in each case, most of the free energy is contributed by a small number of residues (red), and this region is known as the hotspot¹⁰⁰. VEGF, vascular endothelial growth factor; Z domain, a derivative of a domain from Staphylococcus aureus protein A. (Image courtesy of W. DeLano, DeLano Scientific, Palo Alto, California.)

High resolution image and legend (52K)

Research into finding small molecules that disrupt protein–protein interfaces has made considerable progress in the past five years (see refs ^{25, 26, 27, 28} for recent reviews). Here, we focus on six recently published examples of discontinuous protein–protein interfaces for which small molecules that directly compete with one of the protein partners have been discovered (Fig. 2; Table 1). These examples are particularly instructive because crystal structures that are publicly available in the Protein Data Bank (PDB) allow comparison of the protein–protein complexes and the protein–small-molecule complexes. This provides the opportunity to analyse structurally how a small molecule directly competes with a natural protein partner. We also compare the affinities of the protein–protein complexes with those of the protein–small-molecule complexes, using binding equilibrium dissociation constants (K_d) or, in the absence of direct binding data, dissociation constants from competitive binding experiments (K_i) or half-maximal inhibitory concentrations (IC₅₀) from functional assays. We then use these examples to address common myths about targeting protein–protein interfaces. Together, these case studies uncover patterns that should help to advance drug discovery in this important field.

Figure 2: Examples of small molecules that inhibit protein–protein interactions.

Six examples of protein–protein interactions and the small molecules that have been discovered to inhibit them are described in this article. These compounds sit in different areas of chemical space from each other and most other pharmacophores. SP4206 binds to IL-2. ABT-737 binds to Bcl-X_L. Nutlin-3 and the benzodiazepinedione shown above bind to HDM2. Compound 23 binds to HPV E2. Compound 1 binds to ZipA. And SP304 binds to TNF. Selected physical and biochemical characteristics of these molecules are listed in Table 1.

High resolution image and legend (31K)

Table 1: Comparison of protein and small-molecule binding partners

Full table

How small can we go?

The six examples of protein–protein interaction inhibitors that fit the criteria above are discussed in detail in this section.

IL-2 binders

The cytokine interleukin-2 (IL-2) has a key role in the activation of T cells and in the rejection of tissue grafts and is therefore of considerable medical interest. A series of small molecules that bind to IL-2 were produced at Sunesis Pharmaceuticals; the tightest binding of these molecules, SP4206 (Fig. 2; Table 1), has a dissociation constant in the mid-nanomolar range (K_i = 60 nM) and disrupts the interaction between IL-2 and the -chain of the IL-2 receptor (IL-2R)^{29, 30, 31}. These molecules were assembled in a fragment-based approach guided by X-ray structures and medicinal chemistry, and inspired by the previous drug-discovery efforts of Jefferson Tilley and co-workers at F. Hoffmann-La Roche³². Although the small molecules were assembled before the structure of the IL-2–IL-2R complex had been reported³³, they bind close to the centre of the receptor contact region on IL-2 (refs 34, 35) (Fig. 3a), and the interactions are specific because these molecules do not bind to IL-15, the closest homologue of IL-2 (ref. 31).

Figure 3: Four comparisons of how a protein interacts with its natural protein (or peptide) partner and with a synthetic small molecule.

The structures of protein–protein or protein–peptide complexes are shown on the left. The target protein is rendered as a filled surface (grey), and the binding protein or peptide is represented as a ribbon diagram (yellow), with selected side chains shown as sticks (with carbon in yellow, oxygen in red and nitrogen in blue). The contact surface on the target protein (within 4.5 Å of the binding partner) is shown in green. The structures of the protein–small-molecule complexes are shown on the right. The small molecule is shown in stick format, and the contact surface is shown in orange. In the centre, small molecules are shown superimposed on the protein in the conformation in which it binds to its natural protein or peptide partner, and the contact surface (on the target protein) of the natural interaction is shown in green. From these examples, it is clear that the protein–protein contact surface is much larger and flatter than the protein–small-molecule contact surface. a, IL-2 bound to its natural protein partner IL-2R (left), and IL-2 bound to the small molecule SP4206 (right). b, Bcl-XL bound to a peptide derived from one of its natural protein partners, BAD, and Bcl-XL bound to the small molecule ABT-737. c, HDM2 bound to a peptide derived from its natural protein partner p53, and HDM2 bound to the small molecule Nutlin-2 (upper) or a benzodiazepinedione (lower). d, HPV-18 E2 bound to HPV-18 E1, and HPV-11 E2 bound to the small molecule compound 18. The centre panel is not shown, because HPV-18 and HPV-11 are related but not identical. Images generated from files from the PDB, as indicated in Table 1.

High resolution image and legend (120K)

Several interesting features emerge when the structural and functional epitopes of IL-2 involved in binding to the small molecule SP4206 are compared with those involved in binding to the large protein IL-2R¹⁹ (Fig. 3a; Supplementary Movie 1). The contact epitope for the small molecule is about half the size of that for the receptor. But, because the small molecule and the receptor bind to IL-2 with nearly equivalent affinities (less than a tenfold difference), the ligand efficiency (that is, the free energy of binding per non-hydrogen (heavy) atom³⁶) for the small molecule is slightly less than twice that for the receptor (Table 1). The contact surface on IL-2 for binding to IL-2R is flat, as is typical of protein–protein complexes. By contrast, the small molecule traps a conformation of IL-2 in which a groove is present for small-molecule binding, and in which a loop of IL-2 has been repositioned to embrace the furanoic acid moiety at one end of the small molecule. Alanine-scanning mutational studies show that the small molecule and the receptor bind to the same hotspot residues on IL-2 (ref. 19). Although the structures of the small molecule and IL-2R differ markedly, the electrostatic potential of the surfaces presented is similar and probably reflects a need to establish electrostatic complementarity with IL-2. Electrostatic and surface-shape complementarity^{37, 38}, as well as specific hydrogen-bonding interactions, probably account for the high selectivity of these interactions.

These studies show that the binding surface on IL-2 is adaptive and can bind to a small molecule with high affinity using the same main hotspot residues. It is notable that the design of this series of IL-2-binding small molecules did not require knowledge of the structure of the bound receptor complex. Instead, the design was informed by fragment-binding data and by structures of compounds bound to IL-2, coupled with medicinal chemistry and structure–activity relationships (SAR). Thus, the small molecule SP4206 is not an accurate atomic mimic of the receptor, and it would not have been discovered if it had been assumed that the precise structure of the receptor-bound form of IL-2 needed to be captured.

Bcl-X_L binders

Members of the B-cell lymphoma 2 (Bcl-2) family are important regulators of apoptotic cell death^{39, 40}. These molecules can form homodimers and can form heterodimers with other family members (generating various combinations of pro-apoptotic and/or anti-apoptotic subunits). For example, Bcl-2 and Bcl-X_L inhibit apoptosis by binding a 16-residue -helical portion of the pro-apoptotic molecule BAK (Bcl-2-antagonist/killer)⁴⁰ or a 26-residue -helical portion of another pro-apoptotic molecule, BAD (Bcl-2 antagonist of cell death)⁴¹ (Fig. 3b). The importance of BAK and BAD as targets in the treatment of cancer has generated considerable interest in developing synthetic inhibitors of these protein–protein interactions. Several research groups have produced smaller helical molecules that mimic the key -helix involved in this interaction and have high affinities (K_i 5–100 nM in the best cases)^{42, 43, 44, 45}. Recently, a team at Abbott Laboratories generated high-affinity organic compounds that bind to the hydrophobic helical domain of Bcl-X_L, Bcl-2 and another anti-apoptotic molecule, Bcl-W. These small molecules do not bind well to other anti-apoptotic members of the Bcl-2 family such as MCL1 (myeloid cell leukaemia sequence 1) and Bcl-B⁴⁶. The most potent of these, ABT-737 (Fig. 2; Table 1), has a K_i of 0.6 nM and a molecular mass of 813 Da. Its affinity is therefore comparable to that of the -helix, and because it has a smaller contact region with the protein, its ligand efficiency is almost twofold higher. The group of compounds was discovered using a fragment-based nuclear magnetic resonance (NMR) method known as SAR by NMR, and their properties were improved by using NMR-structure-guided medicinal chemistry^{47, 48, 49}. The compounds were active in cell-based assays and in tumour xenograft models in animals, in which they showed synergy with several other chemotherapeutics and radiation. A derivative of ABT-737 — ABT-263 — is in phase I/II clinical trials for cancer (S. Fesik, personal communication).

NMR structures of small fragments that bind weakly to Bcl-X_L (K_d 0.3–4 mM) show ligand conformations similar to those of analogous groups in the elaborated high-affinity compounds these fragments gave rise to, such as ABT-737 (Supplementary Movie 2). Compared with the -helix, however, there are marked differences (Fig. 3b). Alanine-scanning mutational analysis of the BAK-derived peptide identified several residues that are crucial for binding to Bcl-X_L: Val 74, Leu 78, Ile 81, Asp 83 and Ile 85 (refs 40, 45). The small molecule ABT-737 binds to the same region of Bcl-X_L as these residues of the BAK-derived peptide; however, it does not closely mimic the atomic details of the peptide. Instead, the small molecule traps a slightly different conformation of Bcl-X_L, binding in deeper cavities with more puckered grooves.

HDM2 binders

The human protein double minute 2 (HDM2) has emerged as an excellent drug target for cancer treatment. Initially, it was found that the mouse homologue of HDM2 (known as MDM2) binds to the tumour-suppressor protein p53 and increases its degradation, thus blocking the transcriptional activity of p53 that results in tumour suppression (see ref. 50 for a review). MDM2 can bind to a 15-residue -helical region of p53 (K_d 600 nM), and the structure of the complex shows an interface that is largely hydrophobic⁵¹. Alanine-scanning mutational analysis of the 15-residue peptide identified three dominant amino acids in the centre of the interface: Phe 19, Trp 23 and Leu 26 (ref. 52). In search of inhibitors, a subsequent HTS and medicinal-chemistry effort at F. Hoffmann-La Roche (in Nutley, New Jersey) identified a series of tetra-substituted imidazoles, which the researchers named Nutlins. After considerable chemical optimization, the most potent of these small molecules, Nutlin-3 (Fig. 2; Table 1), was found to disrupt HDM2–p53 complexes with an IC₅₀ of 90 nM, and it showed potent p53-blocking activity in vitro and activity against tumour xenografts in vivo⁵³. At Johnson & Johnson, 338,000 compounds were screened in parallel for binding to HDM2, by monitoring changes in thermostability (with a ThermoFluor instrument; Johnson & Johnson), and this resulted in the identification of a series of benzodiazepinediones⁵⁴. After chemical optimization, one of these molecules (Fig. 2; Table 1) was found to have a high affinity for HDM2 (K_d = 67 nM, IC₅₀ = 420 nM)⁵⁵. Although these compounds were initially selected for binding to HDM2 and not for functional disruption of the complex, they promoted rapid dissociation of p53 from HDM2 in cells overproducing HDM2 (ref. 56). Furthermore, a benzodiazepinedione that had been modified with a solubilizing moiety inhibited the proliferation of tumour cells in vitro with an IC₅₀ of 10 M and showed synergistic activity with the chemotherapeutic drug doxorubicin against tumours in mice⁵⁶.

The structures of the benzodiazepinedione bound to HDM2, and two Nutlins bound to the same protein, have been solved: HDM2–Nutlin-2 by X-ray crystallography⁵³, HDM2–Nutlin-3(analogue) by NMR spectroscopy⁵⁷, and HDM2 in complex with the benzodiazepinedione by X-ray crystallography⁵⁴ (Fig. 3c; Supplementary Movie 3). These three compounds bind to HDM2 within the same region as the -helical portion of p53, and they insert aromatic or aliphatic moieties into hotspot pockets on HDM2 that bind to on HDM2 that bind to key residues on p53: Phe 19, Trp 23 and Leu 26. The contact epitopes for the small molecules are again about half the size of the minimal peptide-binding epitope. The conformation of HDM2 is more open at the ends when it binds to the peptide, whereas it closes more tightly over the small molecules, resulting in a more concave contact surface, as is the case for IL-2 and Bcl-X_L. It is remarkable that these dissimilar small-molecule scaffolds, discovered from markedly different starting points, have a similar mode of binding.

HPV E2 binders

Human papilloma virus (HPV) is of considerable interest, because it is the causative agent of warts and some cervical cancers. At present, there is no small-molecule drug that can treat these conditions. The interaction between the viral transcription factor E2 and the viral helicase E1 is crucial for the viral life cycle and thus is an important protein–protein-interface target. By using HTS, a research group at Boehringer Ingelheim identified a class of indandiones that moderately disrupts this interaction (K_d 20 M)⁵⁸. Medicinal-chemistry efforts allowed further optimization of the affinity, with IC₅₀ values as low as 6 nM^{59, 60, 61} — for compound 23, for example (Fig. 2; Table 1). Direct binding of ³H-labelled indandiones and isothermal titration calorimetry both showed that these small molecules bind to the transactivation domain of E2 with one-to-one stoichiometry. Interestingly, an X-ray structure of one of these small molecules, compound 18, bound to the E2 transactivation domain showed two copies of the small molecule; one penetrates into a cavity in the three-helix domain of E2, and the other sits on top⁶⁰.

Soon after the release of this X-ray structure, the structure of the transactivation domain of HPV type 18 (HPV-18) E2 in complex with E1 was reported⁶². The contact surface between E1 and E2 in HPV-18 completely spans the E2–compound-18 contact interface in HPV-11 (Fig. 3d). Of the 20 residues of E2 that are in contact with E1, compound 18 is in contact with only 7. Importantly, compound 18 accesses a cavity that is not observed in the protein–protein interface (Supplementary Movie 4). Compound 23 achieves higher ligand efficiency than E1 (Table 1), presumably by deeply burying its hydrophobic surface area rather than spreading it across the interface.

ZipA binders

The separation of bacterial cells during cell division, and therefore their replication, depends on the formation of a septal ring. In certain Gram-negative bacteria, this ring consists of two or more proteins: FtsZ, a homologue of eukaryotic tubulin; and ZipA, a membrane-anchored protein. These molecules form a complex by interacting through their carboxy-terminal domains. A high-resolution (1.5 Å) X-ray structure revealed that a 17-residue peptide derived from the C terminus of Escherichia coli FtsZ binds to a cavity in ZipA as an extended -strand followed by an -helix⁶³ (Supplementary Movie 5). The FtsZ-derived peptide (K_d 20 M) binds about 100 times weaker than full-length FtsZ but is a useful surrogate. Although ten of the peptidyl side chains directly interact with ZipA, alanine-scanning mutational analysis shows that only four of these side chains (three hydrophobic and one acidic) dominate the binding affinity and constitute a hotspot. The structure of the unbound form of ZipA was also reported⁶³ and was found to be roughly similar to ZipA in complex with the FtsZ-derived peptide. A comparison of the pre-bound (apo) and the peptide-bound ZipA structures shows that loop adjustments and side-chain flips occur in ZipA to facilitate binding of the peptide. ZipA thus presents an adaptive surface for binding.

Using NMR spectroscopy to screen a diverse set of 825 compounds, Wyeth yielded 7 molecules that bound to ZipA at the same site as FtsZ⁶⁴. Even though this is a high 'hit' rate (0.8%), indicating that ZipA might be 'druggable'⁶⁵, extensive medicinal-chemistry and SAR efforts starting from selected hits did not generate any high-affinity small molecules⁶⁶. In a search for other possibilities, HTS of 250,000 compounds identified a pyridylpyrimidine, compound 1 (Fig. 2; Table 1), with a K_i of 12 M⁶⁷. The X-ray structure of compound 1 in complex with ZipA shows that the small molecule binds to ZipA entirely within the region bound by the 17-residue peptide and is in contact with only 740 Ų of ZipA compared with 1,350 Ų for the peptide (Supplementary Movie 5). Although the surface of this small molecule is more complementary to the surface of ZipA than the peptide surface is, these molecules could not penetrate deep into the surface of ZipA, unlike the other small molecules described here to bind IL-2, Bcl-X_L, HDM2 and HPV E2.

TNF disruptors

The cytokine tumour-necrosis factor (TNF) is a key factor in inflammatory responses and is therefore an important drug target. Biological therapeutics that target TNF have been approved for treating arthritis. Not surprisingly, there is considerable interest in developing small molecules or peptides that can disrupt the interaction between TNF and its receptors, TNFR1 and TNFR2. For example, small (13-residue) TNFR1-derived peptides that bind to TNF with moderate affinity (K_d 5 M) have been found⁶⁸, and photoactive small molecules that inhibit the TNF–TNFR1 interaction by labelling a site near where the receptor binds have also been discovered⁶⁹.

More recently, another class of small molecule that targets TNF was discovered, by using fragment screening⁷⁰. These molecules disrupt TNF by binding (up to K_d 13 M) and displacing one of the three monomers that constitute TNF. More specifically, they bind to an adaptive cluster of tyrosine residues at the core of the trimer interface, which is illustrated in Fig. 4a for the small molecule SP304 (Fig. 2; Table 1). Two aromatic side chains from SP304 occupy the same position as the tyrosine residues from the displaced monomer (Supplementary Movie 6). Small molecules of this class are not seriously considered as drug candidates because of their moderate affinities; however, this finding shows that even constitutive interfaces in oligomeric proteins can bind to small molecules. Another recent example of this is the discovery of small molecules that can inhibit the activity of the anti-apoptotic protein survivin (also known as BIRC5) by binding at the interface between survivin homodimers⁷¹.

Figure 4: Disruption of TNF by a small molecule.

a, The structure of TNF, which is composed of three monomers, is shown on the left. Two of the TNF monomers are rendered as a filled surface (grey), and the other monomer is represented as a ribbon diagram (yellow). The contact surface on the TNF dimer (within 4.5 Å of the third monomer) is shown in green. The structure of the TNF dimer in complex with the small molecule SP304 is shown on the right. The small molecule is shown in stick format (with carbon in yellow, oxygen in red, nitrogen in blue and fluorine in white), and the contact surface on the TNF dimer is shown in orange. Images generated from files from the PDB, as indicated in Table 1. b, There are two models for how small molecules could block the formation of TNF trimers. In model 1, one of the monomers of TNF must completely dissociate before the small molecule can bind. In model 2, the small molecule can intercalate into the TNF complex and associate, which facilitates dissociation of a monomer. SP304 accelerates the rate of monomer dissociation (by more than 600-fold), which supports model 2.

High resolution image and legend (51K)

It is known that TNF monomers in the trimer can be exchanged for free monomers, albeit slowly. Remarkably, the small molecules discussed above can increase the kinetics of monomer dissociation by more than 600-fold. Thus, instead of waiting for a monomer to dissociate completely (Fig. 4b, model 1), the small molecule can intercalate into the dynamic trimer complex and displace the monomer (Fig. 4b, model 2). Presumably, the dynamic motions of proteins (see page 964) enable the small molecule to intercalate into the interface and prevent the displaced monomer from re-forming a high-affinity complex with the remaining dimer.

Myths about disrupting protein–protein interfaces

Until recently, lack of success in identifying small molecules that disrupt protein–protein interactions has led to several misconceptions about the prospects for new discoveries.

Protein–protein contact surfaces

One myth is that the large and flat contact surfaces seen in structures of protein complexes are rigid and do not present cavities for small molecules to bind. However, all of the contact surfaces described here show some adaptability, and cavities that are not seen in structures of either the free protein or the protein–protein complex are available for binding. Most of this flexibility involves motions of side chains and small perturbations of loops. In each case in Fig. 3, the small molecule accesses small pockets or grooves, which the larger, more constrained protein or peptide does not. Thus, it should not be assumed that the best binding site for a small molecule can be observed from static structures of either the free protein target or the protein–protein complex. For example, Bcl-X_L seems to have a rather flat surface in the static apo state, but during computer simulations of molecular dynamics in the absence of small molecules, transient pockets open in less than 1 nanosecond^{72, 73}. Similar transient openings of binding pockets were found in simulations with IL-2 and HDM2 (ref. 73).

Screening for protein–protein interface inhibitors

Another myth is that screening does not work for protein–protein interfaces. All of the examples presented here, however, involved empirical screening, either fragment screening or traditional HTS involving small-molecule libraries. In several examples, the starting compounds were identified by HTS, by using large numbers of compounds (more than 250,000) to identify moderate hits (K_i in the mid-micromolar range). Extensive biophysical techniques were applied to check that these hit compounds were 'real' and stoichiometric before any investment in medicinal-chemistry approaches. In four of the cases described here, medicinal-chemistry approaches improved the properties of these hit compounds to generate molecules with a K_d in the mid-nanomolar to low-nanomolar range; in two cases, ZipA and TNF, they did not. The ability to improve a hit compound was not well predicted by the behaviour of the initial compound or by inspection of the binding site, but it might be indicated by hit rates from fragment libraries⁶⁵ or by druggability indices⁹ applied to possible protein–ligand conformations identified from computer simulations⁷².

One explanation for why HTS is not more successful is that the libraries of compounds used for screening are derived mostly from historical medicinal-chemistry efforts in pharmaceutical companies. These 'chemical phenotypes' (chemotypes) are dominated by past drug-discovery research into G-protein-coupled receptors, enzymes and other traditional druggable targets. New classes of target often require new chemotypes. Thus, it is probable that each protein–protein interface will require a new chemotype. As a small-scale analysis, we took high-affinity inhibitors of protein–protein interactions by IL-2, Bcl-X_L, HDM2 and HPV E2, and we compared these small molecules with sets of compounds directed against targets in the chemical databases MDL Drug Data Report (MDDR; Symyx Technologies) and World of Molecular Bioactivity (WOMBAT; Sunset Molecular Discovery), by using a compound similarity ensemble approach⁷⁴. The protein–protein interaction inhibitors did not show high similarity to any set of compounds against other known targets. Thus, if traditional libraries are used for screening, large collections of compounds might be required to find bona fide hit compounds with a K_d in the 10–100 M range⁷⁵. Moreover, it should not be assumed that there are a few 'privileged' scaffolds that will unlock this entire target class, as has been the case for protein kinases and G-protein-coupled receptors. Except for close homologues, each protein–protein interface is different, so the chemotypes of their inhibitors are likely to be more isolated in chemical space.

It is possible that fragment screening will be more successful than HTS when applied to protein–protein interfaces. Several successes have been achieved by using fragment screening, even though there have probably been far fewer fragment-screening efforts than HTS efforts. In theory, fragments (150–250 Da) have higher ligand efficiencies than typical compounds discovered by HTS (400–500 Da) and allow a greater search of chemical space^{36, 76}.

Affinity of protein–protein interactions

A further myth is that native protein complexes have a higher affinity than protein–small-molecule complexes and cannot be competed away. In most of the cases described here, the optimized small molecule bound with an affinity comparable to that of the native partner protein or peptide. In several examples (IL-2, HDM2 and HPV E2), the K_i or IC₅₀ was in the mid-nanomolar to low-nanomolar range and comparable to the binding affinity (K_d). This indicates that in equilibrium conditions, the small molecule is not at a disadvantage when it comes to displacing the protein partner.

From a kinetic perspective, small molecules might have an advantage over a large protein competitor, such as an antibody. For example, in the case of TNF, the inhibitory compounds accelerated the dissociation of a monomer from the TNF trimer by more than 600-fold. Thus, inhibition was not rate limited by the off-rate of a TNF monomer. It would be interesting to determine whether the other small molecules described here can accelerate dissociation of the respective protein–protein complexes. Recent paramagnetic NMR studies of protein complexes show that protein–protein interfaces might be inherently 'wobbly'^{77, 78}. If this is generally the case, then a small molecule could penetrate these dynamic 'encounter' complexes and have a kinetic advantage over a large antibody therapeutic, the association of which depends on complete (not partial) dissociation of the competing protein partner.

Size of small molecules that disrupt protein–protein interactions

Another myth is that small molecules that target protein–protein interfaces are too large to be drugs. For good oral absorption (or bioavailability), most orally active drugs are less than 500 Da^{79, 80}, and drugs to treat neurological conditions usually need to have even lower molecular masses to cross the blood–brain barrier. Such criteria, derived from the limited set of known drugs, have notable exceptions; for example, cyclosporin A is 1,000 Da. In addition, ABT-737 is 813 Da (Table 1) and has a reasonable (70%) bioavailability in rodents⁴⁶, and a derivative of comparable size, ABT-263, has entered clinical trials. Moreover, numerous drugs, including many antibiotics and anticancer drugs, are administered by injection, so in these cases, considerations of molecular mass are not driven by oral bioavailability.

There is always a trade-off between compound binding affinity and properties such as pharmacokinetics, solubility, toxicity and ease of synthesis, which together determine the probability that a compound will succeed as a drug. For optimum values of the latter properties, it is clearly better if the molecular mass is lower. All of the protein–protein interaction inhibitors described here with K_i values of less than 1 M have molecular masses of 500–900 Da. We therefore wondered whether there is a limiting relationship between compound potency and size for small molecules that inhibit protein–protein interactions.

To analyse this, we selected only compounds for which there were extensive medicinal-chemistry data, as well as solved structures showing the compounds or close analogues bound to their targets. For the highest-affinity fragments and optimized compounds that bind to these target proteins, we plotted the binding free energy against the number of non-hydrogen atoms in the ligand (Fig. 5). These data have a reasonably linear distribution with a correlation coefficient of 0.77. It is remarkable that the small molecules that bind to these markedly different targets have similar ligand efficiencies, even though they belong to different chemotypes. The slope of the line gives a ligand efficiency of 0.24 kcal per mol per non-hydrogen atom. This value is considerably less than that for the tightest-binding small molecules such as biotin binding to avidin (1.15 kcal per mol per non-hydrogen atom)³⁶, but it is not dissimilar to that of many protein-kinase inhibitors (0.3–0.4 kcal per mol per non-hydrogen atom) and is comparable to that of many protease inhibitors (0.25–0.35 kcal per mol per non-hydrogen atom)^{36, 81}.

Figure 5: Relationship between compound potency and size for small molecules that inhibit protein–protein interactions.

For the highest-affinity fragments and small molecules that target protein–protein interfaces, the binding free energy (-G) is plotted against the number of non-hydrogen atoms. K_d values were converted to free energy (kcal per mol) using standard-state conditions of 1 M concentration at a temperature of 300 K. Where a direct binding dissociation constant was not available or was not the lowest measured, the K_i or IC₅₀ was used instead. The slope can be described by y = 0.24x, and the correlation coefficient is 0.77. The linear relationship implies that there is a uniform ligand efficiency for these targets. Note that the first occurrence of Compound 1 (black circle) denotes the compound discussed in this article and depicted in Fig. 2, and the second occurrence (orange circle) denotes a molecule with a different chemical structure.

High resolution image and legend (41K)

A survey of several less-optimized small molecules that inhibit protein–protein interfaces (Table 2) shows, with some exceptions, that these have ligand efficiencies similar to the small molecules listed in Table 1. Assuming a value of 0.24 kcal per mol per non-hydrogen atom, a compound with a K_d of 10 nM (typical of many drugs) would need to have 46 non-hydrogen atoms (and therefore have a molecular mass of 645 Da). We suggest that medicinal-chemistry efforts that generate molecules above this curve are doing exceptionally well. But for molecules that are considerably below this curve, much optimization is required if nanomolar affinity and oral bioavailability are desired.

Table 2: Ligand efficiencies of other small molecules that inhibit protein–protein interactions

Full table

Prospects and challenges for drug discovery

In the past five years, there has been remarkable progress in identifying, characterizing and developing small molecules that bind to protein–protein contact surfaces. In addition to binding to the contact surface itself, it is also possible to inhibit protein–protein interactions through allosteric sites^{82, 83} and by promoting aberrant interactions (for example, by cyclosporin A, which inhibits calcineurin by promoting an inhibitory interaction between cyclosporin A and cyclophilin A)⁸⁴. However, there is still a long way to go before protein–protein interface inhibitors can be discovered routinely. It is not clear that the optimal region of compound space is being screened or that the compounds that are found can be easily optimized for these diverse interfaces. Fragment-screening methods offer excellent opportunities to cover a wider swathe of synthetically feasible chemical space per atom. The existence of hotspots means that ligand-efficient 'footholds' can be established by the initial fragments. However, except for the HDM2 binder, compounds that bound to the hotspot alone were not high-affinity inhibitors. To target IL-2, Bcl-X_L, HPV E2 and ZipA, additional sources of affinity were needed and were subsequently found (except for ZipA). The highest-affinity small molecules often engaged residues that the natural protein partner did not, exploiting 'cryptic' pockets within the protein contact interface.

For fragment screening to become more widely adopted, it will need to become cheaper, more sensitive and higher throughput. The biggest challenge in applying such fragment-screening technologies to drug discovery could be 'growing' fragments into higher-affinity small molecules. Improved computational methods to design ligand-efficient 'elaborated' compounds that bind to flexible protein sites would help to focus medicinal-chemistry efforts on these adaptive targets. In a recent study, high-affinity inhibitors of IL-2, Bcl-X_L and HDM2 were computationally docked to protein-conformation snapshots obtained from 10-nanosecond molecular-dynamics simulations⁸⁵. In each case, one or more protein-conformation snapshots had a docked ligand conformation similar to the one observed experimentally. These results suggest that most, but not all, of the conformational differences seen when comparing unbound structures with inhibitor-bound structures result from conformational selection by the ligand. It is notable that the structural changes seen at these protein–protein interfaces are generally smaller than those that appear in classic examples of biologically evolved induced fit, such as hexokinase^{86, 87, 88}.

If we assume that protein–protein interactions have a lower 'ceiling' for ligand efficiency than more traditional targets, then the drug-discovery community will need to improve its management of the absorption, distribution, metabolism and excretion (ADME) properties of larger compounds. Although compounds that inhibit protein–protein interfaces are larger than typical drugs, these compounds are specific for their targets, as shown here for IL-2, Bcl-X_L, HDM2 and HPV E2. The compelling biology of protein–protein interfaces and the fact that several small molecules that inhibit protein–protein interactions are making their way through clinical trials provide hope that more of these drugs might be on the shelf in the future. Clearly, recent efforts have lifted us a rung higher in the quest to reach this class of high-hanging fruit.

References

1. Krogan, N. J. et al. Global landscape of protein complexes in the yeast Saccharomyces cerevisiae. Nature 440, 637–643 (2006). | Article | PubMed | ISI | ChemPort |
2. LaCount, D. J. et al. A protein interaction network of the malaria parasite Plasmodium falciparum. Nature 438, 103–107 (2005). | Article | PubMed | ISI | ChemPort |
3. Komurov, K. & White, M. Revealing static and dynamic modular architecture of the eukaryotic protein interaction network. Mol. Syst. Biol. 3, 110 (2007). | Article | PubMed | ChemPort |
4. Strong, M. & Eisenberg, D. The protein network as a tool for finding novel drug targets. Prog. Drug Res. 64, 191–215 (2007). | PubMed | ChemPort |
5. Pu, S., Vlasblom, J., Emili, A., Greenblatt, J. & Wodak, S. J. Identifying functional modules in the physical interactome of Saccharomyces cerevisiae. Proteomics 7, 944–960 (2007). | Article | PubMed | ChemPort |
6. Collins, S. R. et al. Functional dissection of protein complexes involved in yeast chromosome biology using a genetic interaction map. Nature 446, 806–810 (2007). | Article | PubMed | ISI | ChemPort |
7. Jones, S. & Thornton, J. M. Principles of protein–protein interactions. Proc. Natl Acad. Sci. USA 93, 13–20 (1996). | Article | PubMed | ChemPort |
8. Conte, L. L., Chothia, C. & Janin, J. The atomic structure of protein–protein recognition sites. J. Mol. Biol. 285, 2177–2198 (1999). | Article | PubMed |
9. Cheng, A. C. et al. Structure-based maximal affinity model predicts small-molecule druggability. Nature Biotechnol. 25, 71–75 (2007). | Article |
10. Smith, R. D. et al. Exploring protein–ligand recognition with Binding MOAD. J. Mol. Graph. Model. 24, 414–425 (2006). | Article | PubMed | ChemPort |
11. Hopkins, A. L. & Groom, C. R. The druggable genome. Nature Rev. Drug Discov. 1, 727–730 (2002). | Article |
12. Marsters, J. C. Jr et al. Benzodiazepine peptidomimetic inhibitors of farnesyltransferase. Bioorg. Med. Chem. 2, 949–957 (1994). | Article | PubMed | ChemPort |
13. Zobel, K. et al. Design, synthesis, and biological activity of a potent Smac mimetic that sensitizes cancer cells to apoptosis by antagonizing IAPs. ACS Chem. Biol. 1, 525–533 (2006). | Article | PubMed | ChemPort |
14. Robin, W. S. High-throughput screening of historic collections: observations on file size, biological targets, and file diversity. Biotechnol. Bioeng. 61, 61–67 (1998). | Article | PubMed |
15. Cochran, A. G. Antagonists of protein–protein interactions. Chem. Biol. 7, R85–R94 (2000). | Article | PubMed | ISI | ChemPort |
16. Clackson, T. & Wells, J. A. A hot spot of binding energy in a hormone-receptor interface. Science 267, 383–386 (1995). | Article | PubMed | ISI | ChemPort |
17. Clackson, T., Ultsch, M. H., Wells, J. A. & de Vos, A. M. Structural and functional analysis of the 1:1 growth hormone:receptor complex reveals the molecular basis for receptor affinity. J. Mol. Biol. 277, 1111–1128 (1998). | Article | PubMed | ISI | ChemPort |
18. Muller, Y. A. et al. Vascular endothelial growth factor: crystal structure and functional mapping of the kinase domain receptor binding site. Proc. Natl Acad. Sci. USA 94, 7192–7197 (1997). | Article | PubMed | ChemPort |
19. Thanos, C. D., DeLano, W. L. & Wells, J. A. Hot-spot mimicry of a cytokine receptor by a small molecule. Proc. Natl Acad. Sci. USA 103, 15422–15427 (2006). | Article | PubMed | ChemPort |
20. Moreira, I. S., Fernandes, P. A. & Ramos, M. J. Hot spots — a review of the protein–protein interface determinant amino-acid residues. Proteins 68, 803–812 (2007). | Article | PubMed | ChemPort |
21. DeLano, W. L., Ultsch, M. H., de Vos, A. M. & Wells, J. A. Convergent solutions to binding at a protein–protein interface. Science 287, 1279–1283 (2000). | Article | PubMed | ISI | ChemPort |
22. Sidhu, S. S., Lowman, H. B., Cunningham, B. C. & Wells, J. A. Phage display for selection of novel binding peptides. Methods Enzymol. 328, 333–363 (2000). | PubMed | ISI | ChemPort |
23. Wrighton, N. C. et al. Small peptides as potent mimetics of the protein hormone erythropoietin. Science 273, 458–464 (1996). | Article | PubMed | ISI | ChemPort |
24. Livnah, O. et al. Functional mimicry of a protein hormone by a peptide agonist: the EPO receptor complex at 2.8 Å. Science 273, 464–471 (1996). | Article | PubMed | ISI | ChemPort |
25. Arkin, M. R. & Wells, J. A. Small-molecule inhibitors of protein–protein interactions: progressing towards the dream. Nature Rev. Drug Discov. 3, 301–317 (2004). | Article |
26. Yin, H. & Hamilton, A. D. Strategies for targeting protein–protein interactions with synthetic agents. Angew. Chem. Int. Ed. Engl. 44, 4130–4163 (2005). | Article | PubMed | ChemPort |
27. Fry, D. C. Protein–protein interactions as targets for small molecule drug discovery. Biopolymers 84, 535–552 (2006). | Article | PubMed | ChemPort |
28. Arkin, M. Protein–protein interactions and cancer: small molecules going in for the kill. Curr. Opin. Chem. Biol. 9, 317–324 (2005). | Article | PubMed | ChemPort |
29. Arkin, M. R. et al. Binding of small molecules to an adaptive protein–protein interface. Proc. Natl Acad. Sci. USA 100, 1603–1608 (2003). | Article | PubMed | ChemPort |
30. Braisted, A. C. et al. Discovery of a potent small molecule IL-2 inhibitor through fragment assembly. J. Am. Chem. Soc. 125, 3714–3715 (2003). | Article | PubMed | ISI | ChemPort |
31. Raimundo, B. C. et al. Integrating fragment assembly and biophysical methods in the chemical advancement of small-molecule antagonists of IL-2: an approach for inhibiting protein–protein interactions. J. Med. Chem. 47, 3111–3130 (2004). | Article | PubMed | ChemPort |
32. Tilley, J. W. et al. Identification of a small molecule inhibitor of the IL-2/IL-2R receptor interaction which binds to IL-2. J. Am. Chem. Soc. 119, 7589–7590 (1997). | Article | ISI | ChemPort |
33. Rickert, M., Wang, X., Boulanger, M. J., Goriatcheva, N. & Garcia, K. C. The structure of interleukin-2 complexed with its receptor. Science 308, 1477–1480 (2005). | Article | PubMed | ISI | ChemPort |
34. Thanos, C. D., Randal, M. & Wells, J. A. Potent small-molecule binding to a dynamic hot spot on IL-2. J. Am. Chem. Soc. 125, 15280–15281 (2003). | Article | PubMed | ISI | ChemPort |
35. Emerson, S. D. et al. NMR characterization of interleukin-2 in complexes with the IL-2R receptor component, and with low molecular weight compounds that inhibit the IL-2/IL-R interaction. Protein Sci. 12, 811–822 (2003). | Article | PubMed | ISI | ChemPort |
36. Kuntz, I. D., Chen, K., Sharp, K. A. & Kollman, P. A. The maximal affinity of ligands. Proc. Natl Acad. Sci. USA 96, 9997–10002 (1999). | Article | PubMed | ChemPort |
37. Lee, L. P. & Tidor, B. Optimization of binding electrostatics: charge complementarity in the barnase–barstar protein complex. Protein Sci. 10, 362–377 (2001). | Article | PubMed | ChemPort |
38. Midelfort, K. S. et al. Substantial energetic improvement with minimal structural perturbation in a high affinity mutant antibody. J. Mol. Biol. 343, 685–701 (2004). | Article | PubMed | ISI | ChemPort |
39. Adams, J. M. & Cory, S. The Bcl-2 protein family: arbiters of cell survival. Science 281, 1322–1326 (1998). | Article | PubMed | ISI | ChemPort |
40. Sattler, M. et al. Structure of Bcl-x_L–Bak peptide complex: recognition between regulators of apoptosis. Science 275, 983–986 (1997). | Article | PubMed | ISI | ChemPort |
41. Petros, A. M. et al. Rationale for Bcl-x_L/Bad peptide complex formation from structure, mutagenesis, and biophysical studies. Protein Sci. 9, 2528–2534 (2000). | Article | PubMed | ISI | ChemPort |
42. Sadowsky, J. D., Murray, J. K., Tomita, Y. & Gellman, S. H. Exploration of backbone space in foldamers containing - and -amino acid residues: developing protease-resistant oligomers that bind tightly to the BH3-recognition cleft of Bcl-x_L. Chembiochem 8, 903–916 (2007). | Article | PubMed | ChemPort |
43. Yin, H. et al. Terphenyl-based Bak BH3 -helical proteomimetics as low-molecular-weight antagonists of Bcl-x_L. J. Am. Chem. Soc. 127, 10191–10196 (2005). | Article | PubMed | ChemPort |
44. Walensky, L. D. et al. Activation of apoptosis in vivo by a hydrocarbon-stapled BH3 helix. Science 305, 1466–1470 (2004). | Article | PubMed | ISI | ChemPort |
45. Sadowsky, J. D. et al. (/+)-peptide antagonists of BH3 domain/Bcl-x_L recognition: toward general strategies for foldamer-based inhibition of protein–protein interactions. J. Am. Chem. Soc. 129, 139–154 (2007). | Article | PubMed | ChemPort |
46. Oltersdorf, T. et al. An inhibitor of Bcl-2 family proteins induces regression of solid tumours. Nature 435, 677–681 (2005). | Article | PubMed | ISI | ChemPort |
47. Bruncko, M. et al. Studies leading to potent, dual inhibitors of Bcl-2 and Bcl-xL. J. Med. Chem. 50, 641–662 (2007). | Article | PubMed |