3.2.1 Tethering

Wells and co-workers have used dynamic combinatorial chemistry (DCC) combined with fragment-based lead discovery to develop the ‘tethering’ approach. Reversible disulphide bonds between a cysteine residue in a protein and a thiol-containing fragment are used to capture otherwise weak-binding fragments and identify them by mass spectrometry (Erlanson et al. Reference Erlanson, Braisted, Raphael, Randal, Stroud, Gordon and Wells2000). Cysteines are often engineered into proteins local to the target site. This method was later developed into ‘tethering with dynamic extenders’ where a small-molecule ‘extender’ irreversibly alkylates the cysteine residue of a protein and also provides a thiol group to capture fragments (Erlanson et al. Reference Erlanson, Lam, Wiesmann, Luong, Simmons, Delano, Choong, Burdett, Flanagan, Lee, Gordon and O'Brien2003). However, the chemistry used in this approach is difficult and modification of the protein is a prerequisite for fragment screening. Furthermore, specific libraries composed of thiol-containing fragments are needed.

3.2.2 Screening fragments using biophysical methods

If tethering is not used, the low affinities often observed for fragment binding necessitate the use of relatively sensitive assays, in particular biophysical assays that can detect binding to one partner, where inhibition might not be unambiguously detectable in a bioassay. To attain sufficiently high-throughput, an initial coarse-grained screen is often used. Depending on the nature of the assay, this screen may provide some quantification of binding affinity or may serve simply to identify hits with a response above a given threshold. Hits can then be validated and characterised in greater detail but lower throughput, either using the same techniques with different experimental design or using additional techniques, to acquire a wider set of thermodynamic, kinetic and structural data to inform compound progression (see Fig. 2 for a general outline of the approach). We have found that this approach works well when targeting PPIs.

Fig. 2.

Strategy for a fragment-based screening campaign: methods for initial high-throughput fragment screening include thermal shift (DSF), SPR, ligand-based NMR and X-ray crystallography. Validation of potential fragment hits can be achieved by nanospray mass spectrometry, the kinetic parameters such as association and dissociation rates defined using SPR, and the major thermodynamic parameters derived using ITC. X-ray crystallography and NMR are used to define binding modes of the fragment and to guide fragment optimisation or purchase of related compounds (analogue search). Assessment of biological activity assists in the chemical optimisation of the fragments on their way to a drug lead.

The choice of technique to use at each stage of this screening cascade is determined by the physical properties of the target protein, the level of throughput required and the equipment available to the researcher. For example, in the initial stage of screening, if soakable crystals can be obtained, a lot of information on fragment binding can be obtained by X-ray crystallography in the context of a high-throughput screen. If one binding partner is relatively robust so that it can be regenerated on-chip, SPR can be used to screen many compounds with a very small amount of protein. Ligand-observed NMR is very sensitive and can be used to screen cocktails of compounds without the need for deconvolution. The fluorescence-based thermal shift assay is a relatively inexpensive and rapid technique that requires only a target protein with an experimentally accessible thermal denaturation transition.

In the later stages of screening, involved in validating and characterising the hits (Fig. 2), more precise measurements of affinity can be obtained using ITC and SPR. Both techniques can be used to perform competition experiments with a known ligand to confirm that compounds compete for binding at a specific site. In addition, ITC directly measures the change of enthalpy upon binding from which the change of entropy upon binding can be calculated, and SPR directly measures the rate constants for association and dissociation of the protein–fragment complex. At this point in a fragment-screening campaign, structural data obtained by X-ray crystallography or multi-dimensional heteronucleus-edited NMR spectroscopy can give invaluable insight into compound orientation and binding mode. Structure–activity relationships (SAR) can be rationalised at an atomic level and used to guide chemical synthesis of new compounds. These techniques are the most commonly used for detection and quantification of fragment binding and are described in detail later in this review. Many other biophysical techniques can be used at each stage of screening with appropriate experimental design. Examples include fluorescence resonance energy transfer (FRET) and the related amplified luminescent proximity homogeneous assay (ALPHA), fluorescence polarisation (FP) or fluorescence anisotropy (FA), and even single-molecule techniques such as fluorescence correlation spectroscopy.

Affinities (K _D) measured by any of the techniques mentioned above can be used to calculate the free energy of binding and so to calculate ligand efficiency (LE), which is a useful tool to monitor the effectiveness of compound optimisation. LE is defined as the free energy (ΔG) of binding of a ligand for a specific protein averaged for each heavy atom (non-hydrogen atom) (Hopkins et al. Reference Hopkins, Groom and Alex2004). Recently, group efficiencies (GE) have been used in fragment growing and optimisation campaigns (Saxty et al. Reference Saxty, Woodhead, Berdini, Davies, Verdonk, Wyatt, Boyle, Barford, Downham, Garrett and Carr2007). GE allow the estimation of an individual group's contribution towards the overall free energy of binding (ΔG), giving a quick and simple insight into how efficient one modification is over another.

Fragment hits are evolved into larger compounds by optimising and adding groups to pick up additional PLIs that result in improved affinity for the target. During the optimisation process, it is often useful to evaluate the compounds not only by their physicochemical parameters but also by their biochemical and biological activity. Enzyme assays and cell-based assays are now widely used to assess a compound's activity at an early stage of compound optimisation.

Most of the experience of fragment-based approaches has been gained from conventional targets, including kinases and proteases. Here, we review attempts to target PPIs based on these now well-established methods.

3.4.1 Fluorescence-based thermal shift assays

The fluorescence-based thermal shift assay is a rapid and sensitive tool for monitoring thermostability of proteins. Temperature-induced protein denaturation is monitored using an environmentally sensitive dye such as Sypro Orange (Invitrogen) that fluoresces in the environment of the hydrophobic core as the protein unfolds (Pantoliano et al. Reference Pantolino, Petrella, Kwasnoski, Lobanov, Myslik, Graf, Carver, Asel, Springer, Lane and Salemme2001; Holdgate & Ward, Reference Holdgate and Ward2005; Leung et al. Reference Leung, Senisterra, Ritchie, Sadis, Lepock and Hightower1996). The interaction between ligands and proteins usually induces changes in protein thermal stability with modifications in the midpoint denaturation temperature, enthalpy of unfolding and heat capacity. These modifications are due to the coupling of unfolding with binding equilibrium and can be attained by changes in protein structure and conformational flexibility induced by ligand interaction. Hence, an increased thermal stability of the protein–ligand complex corresponds to a reduced conformational flexibility of the protein or a stabilisation. These effects were studied in detail by Celej et al. using bovine serum albumin (BSA) interacting with three different anilinonaphthalene sulphonate (ANS) derivatives and employing different biophysical techniques (Celej et al. Reference Celej, Montich and Fidelio2003). The increase and decrease of protein thermal stability upon ligand binding and influence of ligand concentration and affinity were recently described in a quantitative model by Cimmperman et al. (Reference Cimmperman, Baranauskiene, Jachimoviciute, Jachno, Torresan, Michailoviene, Matuliene, Sereikaite, Bumelis and Matulis2008).

Using β-site amyloid precursor protein-cleaving enzyme 1 (BACE1) as a test system, Lo et al. tested 13 compounds and determined their unfolding temperature (ΔT _m) (Lo et al. Reference Lo, Aulabaugh, Jin, Cowling, Bard, Malamas and Ellestad2004). Using the shift of the unfolding temperature obtained in the presence of ligands relative to that obtained in the absence of ligands, they assessed the ligand-binding affinity at T _m and compared it with values obtained from ITC, which proved to correlate well. However, one should be cautious in extracting further information such as the potency ranking order and the magnitude of binding constants from the ΔT _m values because the same binding affinity can give rise to different values of ΔT _m (and vice versa) depending on the enthalpy, entropy and change in heat capacity of binding and protein folding (Brandts & Lin, Reference Brandts and Lin1990; Holdgate & Ward, Reference Holdgate and Ward2005). Furthermore, this study demonstrates that positive hits can be identified only if thermal shift assays are conducted at appropriate compound concentrations relative to their dissociation constants, i.e. at least 2–3 times K _D of the compound. Vedadi and co-workers determined the thermal stability of 61 recombinant proteins in the presence and absence of a range of physiologically relevant compounds with the aim of finding a stabilising compound that would aid purification and crystallography. No T _m could be determined for 17 proteins. They found that T _m shifts of more than 4K translate into values for IC₅₀ of better than 1μM (Holdgate & Ward, Reference Holdgate and Ward2005; Vedadi et al. Reference Vedadi, Niesen, Allali-Hassani, Fedorov, Finerty, Wasney, Yeung, Arrowsmith, Ball, Berglund, Hui and Marsden2006; Waldron & Murphy, Reference Waldron and Murphy2003).

The fluorescence-based thermal shift assay is unsuitable for proteins that do not show a well-defined sigmoidal unfolding curve in an unliganded state. Proteins with hydrophobic binding pockets or cavities accessible to the dye tend to display an atypically high initial fluorescence. Other difficulties might arise from poorly folded proteins, proteins with high thermal stability or from multiple, independently unfolding domains (Vedadi et al. Reference Vedadi, Niesen, Allali-Hassani, Fedorov, Finerty, Wasney, Yeung, Arrowsmith, Ball, Berglund, Hui and Marsden2006). Additionally, since aggregation of a denatured protein buries hydrophobic surface area, it causes the fluorescent signal to decrease; where this occurs data must be excluded from the analysis. In some cases, compounds themselves can have an influence on the outcome of this assay. Highly fluorescent compounds or compounds that bind both the native and unfolded states of the protein are not suitable for this assay.

Figure 3 shows unfolding curves for two proteins that interact with each other, an example of targeting protein–protein interactions: the Met receptor shows a well-defined apo-unfolding curve and was used for fragment screening whereas NK1 (comprising the first two domains of HGF) does not show a suitable unfolding curve, probably because it has two independent domains, one of which has many disulphides, as well as a tendency to dimerise through domain swapping (Sigurdardottir, Reference Sigurdardottir2012).

Fig. 3.

Thermal shift curves of Met and NK1. (A) Fragments interacting with Met cause a shift in melting temperatures. Compounds 1 (– – –) and 2 (– ·· –) cause a slight destabilisation of the protein as compared with control (——), whereas compound 3 (·····) stabilises the protein. (B) Melting curves of NK1 obtained using different buffers at varying pH (30mM phosphate, pH 7, 150mM NaCl (——), 50mM Bis Tris, pH 6.7 (·····), CAPS, pH 10.5 (– – –) indicate complex melting processes, possibly related to the presence of two domains and disulphide bridges, as well as a tendency to dimerise through domain swapping. One of the curves (– ·· –) shows a fragment that interacts with the fluorescent dye.

The fluorescence-based thermal shift assay is the cheapest of those described in this review – both in initial equipment cost and in consumables used per experiment. However, relatively high protein and fragment consumption can be limiting factors in some projects. Furthermore, this technique is the least sensitive of the techniques presented here confirming only 8–23% of NMR hits against three kinase targets (for examples, see Hubbard & Murray, Reference Hubbard and Murray2011) indicating that binders in millimolar range may be difficult to detect.

However, the simplicity and general applicability of the thermal shift assay still make it attractive for identification of fragment hits. Thus, for targeting the PPI in the Notch ankyrin domain Abdel-Rahman et al. used a library of 1201 rule-of-three-compliant fragments, mainly from the Maybridge library that was not target-tailored but was reasonably diverse (Abdel-Rahman et al. Reference Abdel-Rahman, Martinez-Arias and Blundell2011). The fluorescent-based thermal shift assay identified 36 positive hits with a shift in thermal unfolding temperature (T _m) of 0.5K or more, an impressive outcome when the protein interface is relatively featureless and flat.

3.4.2 Surface plasmon resonance (SPR)

SPR is a label-free optical biosensing technique that measures adsorption of material onto planar metal (typically gold) surfaces. The SPR method is based on optical measurement of refractive index changes associated with the binding of analyte molecules to molecules immobilised on the SPR sensor surface (Jonsson et al. Reference Jonsson, Fagerstam, Ivarsson, Johnsson, Karlsson, Lundh, Lofas, Persson, Roos and Ronnberg1991). Its low protein and compound (analyte) consumption, and high sensitivity make it invaluable for fragment-based drug design. Case studies demonstrate the successful identification of selective low molecular weight inhibitors for pharmacologically relevant drug targets through the SPR screening of fragment libraries (Giannetti, Reference Giannetti2011). Fragment hits as weak as 5mM can be detected in short time frames, with screening of a 5000-membered library and the first stage of hit validation possible in one month. Fragment concentrations used in a screening campaign usually range from 100μM to 1mM (Giannetti, Reference Giannetti2011). In addition, most SPR instruments are multiplexed allowing multiple targets to be screened simultaneously, thereby increasing turnaround speed. Proteins and other biomolecules can be immobilised on a variety of chip surfaces leaving room for individual requirements in assay design. However, great care must be taken in data correction, scaling and normalisation (Giannetti, Reference Giannetti2011). Additionally, solvent effects need to be taken into account, especially if dimethyl sulphoxide (DMSO) is used to solubilise the fragments in the fragment library (Navratilova & Myszka, Reference Navratilova and Myszka2006).

It is helpful to use a known binder as means of identifying over-stoichiometric binders often caused by fragment stacking or other self-interactions. High compound concentrations can result in signals arising from a number of secondary effects, for example non-specific binding (i.e., poorly defined interactions to a single site) and methodological limitations (e.g. carryover, insufficient correction for buffer-related signals and drifting baselines) (Geitmann et al. Reference Geitmann, Elinder, Seeger, Brandt, De Esch and Danielson2011). Screening against different proteins (e.g. mutants with disabled active sites) enables identification of promiscuous binders and ensures specificity against the target protein (Hamalainen et al. Reference Hamalainen, Zhukov, Ivarsson, Fex, Gottfries, Karlsson and Bjorsne2008). A stepwise hit selection and validation protocol is useful for compensating for these secondary effects. For example, a primary screen with relatively generous response cut-off criteria could be used in the first instance to detect binders. This screen should, if possible, highlight promiscuous binders by using one or more suitable reference proteins. These preliminary hits could then be applied to a second screen where competition against a known binder is assessed or the K _D is determined. Using a series of screens, promiscuous binders, fragment stacking or false positives can be easily identified.

The literature provides several examples where SPR has successfully been used as the screening method (Elinder et al. Reference Elinder, Geitmann, Gossas, Kallblad, Winquist, Nordstrom, Hamalainen and Danielson2011; Giannetti, Reference Giannetti2011; Perspicace et al. Reference Perspicace, Banner, Benz, Muller, Schlatter and Huber2009; Xiang et al. Reference Xiang, Hirth, Asmussen, Biemann, Bishop, Good, Fitzgerald, Gladysheva, Jain, Jancsics, Liu, Metz, Papoulis, Skerlj, Stepp and Wei2011). In our hands, SPR is more sensitive than the thermal shift assay for identifying hits in millimolar range. The affinity range that can be reliably determined is limited primarily by the aqueous solubility of the fragment. Kobayashi and co-workers compared SPR with solution ligand-observed NMR in fragment screening. They found that SPR detection is most suited to ligands that bind with K _D better than 1mM. However, they concluded that the good correlation between SPR and potency in a bioassay makes this a good method for hit validation and characterisation studies (Kobayashi et al. Reference Kobayashi, Retra, Figaroa, Hollander, Ab, Heetebrij, Irth and Siegal2010).

Summarising different studies at Vernalis, Hubbard and Murray reported that 84–100% of NMR hits of five different targets were confirmed in SPR with affinities ranging from 0.1 to 1000μM (Hubbard & Murray, Reference Hubbard and Murray2011). Given the screening speed, sensitivity and ability to drive SAR down to biochemical potency, SPR methods represent a robust approach to initiating and supporting a fragment-based lead identification, but additional experiments such as protein-observed NMR or X-ray crystallography need to be performed to determine structural information about the protein–fragment complex.

We have used the approach to target successfully protein–protein interactions with the aim of interrupting the binding between hepatocyte growth factor (HGF)/scatter factor (SF) and Met receptor (Jubb et al. Reference Jubb, Higueruelo, Winter and Blundell2012; Sigurdardottir, Reference Sigurdardottir2012). Protein consumption could be reduced 5000-fold using SPR as compared with using thermal shift assay for fragment screening. A fragment library of ∼1300 members was screened against NK1, SPH-domain and Met receptor at concentrations of 0.5 or 1mM. Promiscuous binders were easily identified by comparing binding levels of each fragment against each target (see Fig. 4). Promiscuous binders show binding to all three proteins and over-stoichiometric binders or aggregating fragments can be identified by extremely high responses. In Figure 4, fragments 1 and 8 are specific for NK1, fragments 2, 3, 4, 6, 12 and 15 are specific for Met and fragment 11 is specific for SPH. Fragment 9 is probably aggregating on the chip surface and might be omitted from the library in future. Fragments 5, 7, 10, 13 and 14 bind to all proteins equally well or do not bind at all. These fragments are probably promiscuous binders, which need to be excluded from further experiments. Hit rates for each protein were around 5–10%, which is consistent with results from other screening campaigns against more conventional targets (Giannetti, Reference Giannetti2011; Hubbard & Murray, Reference Hubbard and Murray2011; Perspicace et al. Reference Perspicace, Banner, Benz, Muller, Schlatter and Huber2009). Only fragment hits that showed clear selectivity for one target were taken forward for hit validation. After employing this selectivity filter, hit rates for each target protein were 0.1% (SPH), 5% (NK1) and 8% (Met). These data are encouraging and indicate that SPR could find wide applicability for screening PPIs.

Fig. 4.

SPR screening. Binding levels of 15 fragments from fragment screening showing hits, over-stoichiometric binders and promiscuous binders to NK1 (green), Met (black) and SPH domain (red).

3.4.3 NMR (ligand-observed)

NMR spectroscopy has been very popular as an initial screening method and was the first experimental method used for screening fragments (Shuker et al. Reference Shuker, Hajduk, Meadows and Fesik1996). NMR methods for fragment screening are classified as either ligand- or receptor-observed methods, depending on whether ligand binding is monitored through ligand or receptor resonances. Generally ligand-based methods are used for primary screening and receptor-based methods for mapping the binding site and further development. Therefore, receptor-based methods are discussed in detail in Section 3.5.4.

Ligand-based methods detect changes in the characteristics of the ligand (such as relaxation and diffusion) that occur when it transiently binds to the receptor (Lepre, Reference Lepre2011). The typical ligand-observed method compares the NMR spectrum of a mixture of ligands in the presence and absence of a protein. This approach renders the molecular weight of the protein irrelevant. Additionally, there is no need to produce isotope-labelled protein. Currently, the most widely used ligand-based methods are saturation transfer difference (STD; Mayer & Meyer, Reference Mayer and Meyer1999) and water-ligand observed via gradient spectroscopy (WaterLOGSY; Dalvit et al. Reference Dalvit, Pevarello, Tato, Veronesi, Vulpetti and Sundstrom2000). STD relies on intermolecular magnetisation transfer directly via the protein, whereas WaterLOGSY experiments utilise the transfer via the bulk water for the detection of molecules interacting with the receptor. These methods consume very little protein and are capable of detecting ligands with affinities ranging from 10⁻⁸ to 10⁻³M (Lepre, Reference Lepre2011). Another useful screening technique is 1D ¹H relaxation experiments (Hajduk et al. Reference Hajduk, Olejniczak and Fesik1997), which detect small molecule binding by comparing line shapes in the presence or absence of receptor. Binding-induced ligand transverse autorelaxation rate (R ₂ = 1/T ₂) enhancements may be visible as simple broadening of proton resonance lines upon the addition of receptor (Lepre et al. Reference Lepre, Moore and Peng2004). Initial ligand-based NMR screening is usually carried out with cocktails containing a mixture of fragments thereby decreasing the amount of protein needed and data collection time. If changes in the ligand spectrum are detected, single fragments from that cocktail are then analysed separately (so-called deconvolution). Cocktails usually contain 8–10 fragments and are designed to avoid chemical reactions between the fragments and overlapping resonance spectra. A competition-binding step during NMR screening has been used to test the specificity of an identified ligand and for extracting the dissociation-binding constant using titration experiments.

The clear advantage of ligand-based NMR methods is that they require only soluble protein; there is no need to crystallise, immobilise or tag the target protein. It is possible to detect fragment binding over a very broad affinity range making this technique suitable for weak binders up to a K _D of 1mM (Kobayashi et al. Reference Kobayashi, Retra, Figaroa, Hollander, Ab, Heetebrij, Irth and Siegal2010). However, this technique is susceptible to false positives arising from fragments that aggregate in solution, but this can be detected by running control experiments. NMR also consumes relatively large amounts of reagents compared with other methods. To screen a library of 1200 fragments, about 30–40mg of protein is required which is considerably more than needed for SPR. Additionally, NMR has lower throughput than biochemical HTS. However, when fragment cocktails are used for screening, the throughput can be increased substantially.

Many examples have been reported for classical targets. For example, Stockman et al. (Reference Stockman, Kothe, Kohls, Weibley, Connolly, Sheils, Cao, Cheng, Yang, Kamath, Ding and Charlton2009) conducted a study to identify small, ligand-efficient fragments that might inhibit the 3-phosphoinositide-dependent kinase-1 (PDK1) in the phosphoinositide 3-kinase (PI3K) signalling cascade. Using STD NMR experiments, 372 fragment hits were identified from a fragment library containing 10237 compounds. In a second round of experiments, fragments were investigated for binding either the ATP (adenosine-5′-triphosphate) site or the allosteric PDK1 interacting fragment (PIF) pocket of PDK-1. In a second example, WaterLOGSY experiments were performed to screen a 2000-compound general fragment library against the aspartic protease β-secretase (BACE1), which is a well-established target for the treatment of Alzheimer's disease. Hits were further characterised using SPR, and a series of six-substituted isocytosines could be identified as a novel scaffold for BACE-1 inhibitors (Geschwindner et al. Reference Geschwindner, Olsson, Albert, Deinum, Edwards, De Beer and Folmer2007).

Ligand-based NMR has also been used successfully for targeting protein–protein interactions, for example with B-cell leukaemia/lymphoma X_L (Bcl-X_L), which inhibits apoptosis by binding a 16-residue α-helical portion of the pro-apoptotic molecule Bcl-2 homologous antagonist/killer (BAK) or a 26-residue α-helical portion of another pro-apoptotic molecule Bcl-2-associated death (BAD). Two different fragments were shown to bind simultaneously to Bcl-X_L and were linked to a single molecule (Oltersdorf et al. Reference Oltersdorf, Elmore, Shoemaker, Armstrong, Augeri, Belli, Bruncko, Deckwerth, Dinges, Hajduk, Joseph, Kitada, Korsmeyer, Kunzer, Letai, Li, Mitten, Nettesheim, Ng, Nimmer, O'CONNOR, Oleksijew, Petros, Reed, Shen, Tahir, Thompson, Tomaselli, Wang, Wendt, Zhang, Fesik and Rosenberg2005). We describe this successful campaign at targeting PPIs in Section 5.

3.4.4 X-ray crystallography (Screening)

X-ray crystallography is an information-rich screening technique, which is widely used in fragment-based drug discovery (Blundell et al. Reference Blundell, Jhoti and Abell2002; Hartshorn et al. Reference Hartshorn, Murray, Cleasby, Frederickson, Tickle and Jhoti2005; Hubbard et al. Reference Hubbard, Davis, Chen and Drysdale2007; Nienaber et al. Reference Nienaber, Richardson, Klighofer, Bouska, Giranda and Greer2000; Verlinde et al. Reference Verlinde, Kim, Mande and Wgj1997). In addition to allowing the detection of binding hits, it provides information about both the site and mode of binding in a single experiment. It is now an established part of several drug development platforms for example the Pyramid at Astex Therapeutics (Davies et al. Reference Blundell, Sibanda, Montalvao, Brewerton, Chelliah, Worth, Harmer, Davies and Burke2006), SeeDs approach at Vernalis (Hubbard et al. Reference Hubbard, Davis, Chen and Drysdale2007) and at Johnson and Johnson (Hubbard et al. Reference Hubbard, Davis, Chen and Drysdale2007; Spurlino, Reference Spurlino2011).

The main technical challenge of the method is to obtain crystal structures with different compounds using either crystal soaking or co-crystallisation. At the screening stage, soaking is preferable as it provides higher throughput and enables a higher concentration of compound to be used. Indeed, a high compound concentration allows the characterisation of even weak-binding fragments (Carr & Jhoti, Reference Carr and Jhoti2002). In practice, the maximal fragment concentration is limited only by solubility and the tolerance of the crystals for fragment and co-solvent. The most common co-solvent used to maintain fragment solubility in the soaking solution is DMSO, so the crystal form used must be stable in the presence of DMSO.

In the simplest case, successful soaking requires a crystal form where the binding site is unoccupied by ligand and unobstructed by crystal contacts. Targeting PPIs offers new challenges to the use of X-ray crystallographic screening. The fact that PPIs evolved to bind to proteins suggests that they are likely to form crystal contacts if their natural partner is absent. This along with the additional difficulties in obtaining soluble unliganded protein for PPI targets, as described in Section 3.3 of this review, makes obtaining a suitable crystal form for screening more difficult. A technique that we have found particularly useful for targeting the RAD51–BRCA2 interaction is microseed matrix screening combined with cross seeding between different mutants (D'Arcy et al. Reference D'arcy, Villard and Marsh2007). This has enabled us to both crystallise previously un-crystallisable mutants and to engineer crystal forms with unobstructed binding sites to be optimal for soaking. The crystals must be readily reproducible in order to obtain optimal results because screening a library of fragments will require 100–1000s of soaks depending on whether fragment cocktails are used and the number of fragments in one cocktail.

At Astex Therapeutics, an average of four fragments is used per soak with the fragments selected to be as chemically diverse as possible within a particular cocktail (Davies et al. Reference Blundell, Sibanda, Montalvao, Brewerton, Chelliah, Worth, Harmer, Davies and Burke2006; Hartshorn et al. Reference Hartshorn, Murray, Cleasby, Frederickson, Tickle and Jhoti2005). This both minimises the chance of more than one fragment binding per soak and eases de-convolution if more than one fragment binds. A different approach, adopted by Johnson and Johnson (Spurlino, Reference Spurlino2011), is to assemble fragments into groups of five similarly shaped compounds. They maintain that this approach increases the chances of multiple fragment hits in one cocktail and reinforces areas of density where multiple fragments are bound. In this case, the hits are de-convoluted by carrying out multiple rounds of refinement with each different fragment from the cocktail as a starting point. However, in our experience, it is desirable to carry out single-molecule soaks for all cocktails where binding is observed.

Screening a large number of crystals requires the acquisition and processing of a concomitantly large number of datasets. In the past, this would have presented a significant obstacle. However, technical improvements have expedited all stages from crystallisation to structure determination. Advances in laboratory X-ray sources now allow efficient in-house data collection. Our in-house X-ray diffraction system (X8 PROTEUM from Bruker (www.bruker.com)) routinely enables datasets to be collected in less than 2h and up to a maximum resolution of 1.5Å. Improvements in automated crystal mounting in synchrotron sources not only allow data to be collected remotely (Gabadinho et al. Reference Gabadinho, Hall, Leonard, Gordon, Monaco and Thibault2008), but can improve the efficiency of data collection (Muchmore et al. Reference Muchmore, Olson, Jones, Pan, Blum, Greer, Merrick, Magdalinos and Nienaber2000), especially by using simple scripting methods to increase automation.

Typical hit rates observed in fragment screening against conventional targets range from 0.5 to 10%. Thus, although the fragment libraries contain only 500 − 1000 compounds, they are able to sample significant chemical space (Hartshorn et al. Reference Hartshorn, Murray, Cleasby, Frederickson, Tickle and Jhoti2005). However, X-ray crystallography has the disadvantage of screening fragments in a protein environment that may be partially constrained by crystal contacts, whereas NMR retains flexibility of protein side chains and conformational changes upon binding of the fragment.

In practice, to increase efficiency of the initial screening process, a pre-screening step is often implemented before screening potential hits with X-ray crystallography (Murray & Blundell, Reference Murray and Blundell2010). However, screening all fragments of the library with X-ray crystallography offers the advantage of evaluating fragment hits based solely on their binding mode and not only on potency. For example, a weaker potency fragment whose orientation offers more opportunity for fragment growing could be preferred over a tighter binding fragment that would be more difficult to develop.

Fragment-based approaches are proving successful in several protein–protein interaction campaigns. An example has been the interface between the human recombinase RAD51 and the hub protein BRCA2 (see Section 2.3 above). Figure 1 shows that the interface on the RAD51 is fairly flat, but there are distinct pockets for the phenylalanine and alanine side chains that bind the FxxA motif. Eighty-one of 1338 fragments screened using thermal shift showed induced shifts greater than 0.8K and were selected for further analysis (Scott et al. Reference Scott, Ehebauer, Pukala, Marsh, Blundell, Venkitaraman, Abell and Hyvönen2012). Of these, 54 bound almost entirely in the pocket that binds Phe. Figure 5 shows that the chemistry and interactions of the fragments are rather more varied than one might assume from the selectivity of binding of the Phe in the BRCA2 sequence (see sequences of binding motifs in Fig. 1). The fragments exploit both the lipophilic nature of the pocket as well as opportunities to form well-defined H-bonds, particularly with the main chain carbonyl function of Leu 214 and the nearby side chain of Gln 217. The affinities are in the range of 400–1000μM. This shows that small side chain pockets in PPIs may act as hotspots that undergo little conformational rearrangement on ligand binding.

Fig. 5.

Fragments binding to the Phe pocket of the RAD51 interface with the BRC4 peptide of BRCA2. The fragments bind with the following K _Ds (A) 540μM (B) 1000μM (C) 600μM (D) 730μM (E) 430μM (F) 460μM (from Scott et al. Reference Scott, Ehebauer, Pukala, Marsh, Blundell, Venkitaraman, Abell and Hyvönen2012).

3.5.1 Thermodynamics of fragment binding: isothermal titration calorimetry

ITC measures the heat evolved or absorbed by all chemical processes that occur when one component of a reaction mixture is titrated into a solution of another component at constant temperature (Wiseman et al. Reference Wiseman, Williston, Brandts and Lin1989). It can be used to monitor any process that has an associated change in enthalpy (ΔH). In the case of macromolecular association, the enthalpy of binding, which can be exothermic or endothermic, is measured as one partner is titrated against another. When the heat of dilution has been accounted for, the heat of binding per mole of injectant decreases in magnitude as the fractional saturation of the complex increases. For the simplest case of a 1:1 association, the given curve can be fitted to a modified Langmuir isotherm (Hill equation) to obtain a value for the equilibrium dissociation constant K _D and the stoichiometric ratio (n). The entropy of binding (ΔS) can be calculated from the Gibbs free energy of binding (ΔG, calculated from the K _D) and the measured enthalpy of binding (Leavitt & Freire, Reference Leavitt and Freire2001; Pierce et al. Reference Pierce, Raman and Nall1999). Accurate fitting of ITC data is dependent on careful experimental design. Both the concentration of the binding partner in the ITC cell and the molar excess of the titrant in the syringe can be manipulated to give data that can be reliably interpreted for a wide range of dissociation constants (nM–mM) and binding enthalpies (Turnbull & Daranas, Reference Turnbull and Daranas2003; Wiseman et al. Reference Wiseman, Williston, Brandts and Lin1989). Although traditionally regarded as low throughput, data acquisition has been expedited by the introduction of commercial small volume ITC instruments that are amenable to automation. In the future, the development of arrays of nano-calorimeters could allow the collection of ITC data in a high throughput screen (Torres et al. Reference Torres, Recht, Coyle, Bruce and Williams2010).

The concentration of protein used typically ranges from 5 to 100μM in a cell volume of 200–1800μl, depending on the enthalpy and stoichiometry of binding. The titrant is typically at a 10-fold molar excess for tight binding (nM to low μM), with a greater excess being required for association stoichiometries greater than 1:1 or for weaker binding titrants.

In line with other binding techniques, ITC may be used to monitor competition experiments where one ligand displaces another from a binding site on a protein. A general solution for the fitting of competitive binding of two ligands (Wang, Reference Wang1995) has been modified to fit ITC data (Sigurskjold, Reference Sigurskjold2000). These experiments ensure that a putative competitive inhibitor is binding to the expected site on a target protein and also can be of particular use measuring extremely tight association (low K _D) for which simple binding data could not be reliably fitted (Velazquez-Campoy et al. Reference Velazquez-Campoy, Todd and Freire2000, Reference Velazquez-Campoy, Kiso and Freire2001).

One of the earliest reported applications of ITC to fragment-based drug discovery was to probe hot spots at the co-factor binding site of the E. coli ketopantoate reductase using fragments derived from the natural ligand, nicotinamide adenine dinucleotide phosphate hydrogen (NADPH) (Ciulli et al. Reference Ciulli, Williams, Smith, Blundell and Abell2006). The binding affinities and thermodynamics of the fragments were measured by ITC and the fragment binding sites were located by a combination of site-directed mutagenesis and ITC. In a second example, ITC in competition with a natural phosphopeptide ligand was used to characterise the binding affinity and thermodynamics of fragments found by virtual screening and ligand-directed NMR to bind to the phosphotyrosine-binding pocket of the viral oncoprotein viral-sarcoma (v-Src) (Taylor et al. Reference Taylor, Gilbert, Williams, Pitt and Ladbury2007). Recently, Edink et al. (Reference Edink, Rucktooa, Retra, Akdemir, Nahar, Zuiderveld, Van Elk, Janssen, Van Nierop, Van Muijlwijk-Koezen, Smit, Sixma, Leurs and De Esch2011) optimised fragment hits toward high-affinity lead compounds that bind to the acetylcholine-binding protein (AChBP). Analysis of fragment binding by ITC contributed to structure-based optimisation of the fragments. Comparison of ligand efficiencies and thermodynamic data from different compounds and protein mutants revealed interactions with a ligand-induced AChBP subpocket. Induced changes in the pocket were shown to be important for selectivity and may be of use in the design of subtype-selective ligands for human nicotinic receptors. In a similar study, Hung et al. (Reference Hung, Silvestre, Wen, Ciulli, Blundell and Abell2009) used a fragment-growing approach of an initial fragment hit, 5-methoxyindole, to enhance binding to pantothenate synthetase (PS) for development of new tuberculosis therapeutics. 5-methoxyindole was found to bind with low affinity of 1.1mM and binding could be enhanced 1000-fold by elaborating the original fragment scaffold.

ITC data can also be used in a more general fashion to inform and direct compound progression. Since the value of ΔH is directly measured in an ITC experiment, and the value of ΔS is calculated using equilibrium constant that is fitted independent of the enthalpy, those parameters are less prone to correlated errors than estimates from van't Hoff analysis of binding monitored by other techniques. Therefore, ITC-derived thermodynamic data are better suited to statistical analysis and correlation with the physico-chemical properties of ligands and binding sites. The Structure/Calorimetry of Reported Protein Interactions Online (SCORPIO) database collated by the Ladbury group holds structural and calorimetric data on nearly 100 unique protein ligand complexes (Olsson et al. Reference Olsson, Williams, Pitt and Ladbury2008). Analysis of these data reveals striking differences in the range of affinities and dominant thermodynamic contributions to binding energy for biological ligands compared with compounds from medicinal chemistry. Further analysis allows assessment of the relative importance of burial of polar and apolar surface area in determining ligand selectivity and affinity respectively, and undermines confidence in commonly assumed physicochemical explanations for increased affinity upon increasing compound lipophilicity.

ITC has been used extensively as part of the RAD51–BRC4 screening effort to monitor the thermodynamics of fragment and compound binding. Figure 6 shows examples of binding isotherms for a fragment, a peptide and a peptide fragment chimaera to humanised RadA.

Fig. 6.

Example binding isotherms of a fragment, a peptide and a peptide fragment chimaera to humanised RadA. (A) Example fragment titration with K _D = 1.5mM. (B) Example peptide titration with K _D = 280μM. (C) Example peptide fragment chimaera K _D = 3μM (Scott et al. Reference Scott, Ehebauer, Pukala, Marsh, Blundell, Venkitaraman, Abell and Hyvönen2012).

3.5.2 Kinetics of fragment binding (SPR)

As we have seen in Section 3.4.2, SPR has become a valuable tool for screening fragment binding to drug targets. Here, we focus on the use of SPR to carry out kinetic analysis by determining the on and off rates and enabling calculation of the binding affinity constant (K _D) of the compound. Additionally, the method is also useful in competition studies to map the binding locus.

There exists a wealth of studies described in the literature where SPR had been used to determine binding affinities for compounds and to monitor compound optimisation (Cole et al. Reference Cole, Olland, Jacob, Brooks, Bursavich, Czerwinski, Declercq, Johnson, Joseph-Mccarthy, Ellingboe, Lin, Nowak, Presman, Strand, Tam, Williams, Yao, Tsao and Fitz2010; Geitmann et al. Reference Geitmann, Elinder, Seeger, Brandt, De Esch and Danielson2011; Nordin et al. Reference Nordin, Jungnelius, Karlsson and Karlsson2005). Most fragments show a squared response curve due to fast on and off rates with little or no curvature, and a 1:1 Langmuir-binding model cannot be fitted. In these cases, kinetic experiments are carried out using different concentrations of the compound and determining their respective equilibrium-binding levels. Response levels versus concentration are therefore fitted using a global fit that is able to determine steady-state affinity constants (Giannetti, Reference Giannetti2011). In our experience, K _D values of a protein–protein-binding event obtained using both fitting methods vary slightly but are generally in good agreement. However, K _D values obtained by other methods such as ITC can differ because the protein in SPR is immobilised on a surface possibly introducing restrictions in flexibility or induced conformational changes upon ligand binding.

Kinetic studies have been used in our laboratory to determine the binding affinity of fragments targeting a PPI as a means of hit validation (Sigurdardottir, Reference Sigurdardottir2012; Jubb et al. Reference Jubb, Higueruelo, Winter and Blundell2012). After data correction, 13% of our initial fragment hits selective for Met were found to bind with an affinity of 100μM or lower, and 34% were found to bind with affinities lower than 500μM (see Fig. 7a). This amounts to 3% of the whole fragment library used. However, nearly 20% of the initial fragment hits did not show the expected binding curves but instead either a linear concentration dependency or no binding; they were therefore rejected from further analysis. This behaviour might be due to low solubility, aggregation or unwanted interaction with the chip surface, and these fragments are not considered hits. Figure 7b shows four examples of binding curves of fragments with different binding affinities.

Fig. 7.

Evaluation of fragment binding using SPR. (A) Distribution of steady-state binding constants (K _D) among fragment hits for Met. (B) Concentration-dependent binding of selected fragments to immobilised Met.

SPR can also be used to characterise drug–drug interactions by evaluating binding of one fragment, while keeping the concentration of the other fragment constant. This type of competition assay can give valuable information about whether or not fragments bind to the same locus on a protein; this is especially useful in the absence of structural information. Results can be used to focus on medicinal chemistry efforts to either combine fragments that target neighbouring sites or to evolve two different fragment scaffolds independently.

3.5.3 Structure (NMR: protein based)

NMR can be used to identify the binding site of a fragment hit to its protein target. When a ligand binds to a protein, the chemical shifts of both ligand and protein proton resonances are affected, but mostly nuclei located within the binding site (Meyer & Peters, Reference Meyer and Peters2003). Once the protein has been enriched with ¹⁵N (or ¹³C) and the chemical shift assignments are known, the protein is titrated with a ligand and chemical shift changes are monitored. 2D ¹H-¹⁵N heteronuclear single quantum coherence (HSQC) spectroscopy is the most common method to monitor backbone chemical shifts. Only the protein resonances are visible in such spectra and, therefore, the ligand can be included at a high concentration (Homans, Reference Homans2004). By comparing the HSQC spectra in the absence of a ligand with a similar spectrum acquired in the presence of a ligand, information on where the ligand is binding can be obtained. Additional information on binding constants and stoichiometry can be obtained by titrating the ligand into the protein sample and observing chemical shifts of selected residues.

HSQC measurements have been used to detect binding of some common buffer molecules to NK1 (Sigurdardottir, Reference Sigurdardottir2012). NK1 is a natural splice variant of HGF/SF and contains the high affinity binding site for Met. Buffers and some other common molecules are often found in crystal structures. The buffer molecules occupy the binding pocket itself, making it possible to use this information, combined with fragment information, to aid the drug discovery process. Figure 8a shows overlaid HSQC spectra of NK1 in the absence and presence of 4-(2-hydroxyethyl)-1-piperazineethanesulphonic acid (HEPES) at one concentration. The dissociation constant K _D was calculated by plotting the weighted average ¹H and ¹⁵N chemical shifts of the residues experiencing significant variations as a function of ligand concentration added during the titration and considering the one site binding model (Garrett et al. Reference Garrett, Seok, Peterkofsky, Clore and Gronenborn1997). A K _D of 3.3mM was calculated for HEPES.

Fig. 8.

Evaluation of HEPES binding to the lysine-like pocket of NK1. (A) Overlaid 2D ¹H-¹⁵N HSQC spectra of 15N enriched NK1 in the absence (blue) and presence (red) of a ligand. (B) X-ray structure of the same compound obtained by soaking.

Another example of the applicability of the HSQC spectrum to identify and validate binding modes of fragments for protein–protein targets is the perturbation of the ZipA–FtsZ complex (Tsao et al. Reference Tsao, Sutherland, Jennings, Li, Rush, Alvarez, Ding, Dushin, Dushin, Haney, Kenny, Malakian, Nilakantan and Mosyak2006). Interaction of membrane-anchored zinc-regulated and ion-regulated transporter proteins (ZipA) with the cytosolic filamenting temperature-sensitive mutant Z (FtsZ) protein is required for bacterial cell division. Inhibitors of this complex might be a new antibiotic class that could help to overcome bacterial drug resistance. Less than 1000 fragments (with molecular weights of 200–250Da with c log P < 2.5) were screened in cocktails of six or fewer compounds by ¹H-¹⁵N HSQC spectrum of ZipA. In total, 16 fragments were found to bind to ZipA, of which seven bound to the FtsZ-binding pocket.

Although protein-based NMR is not high-throughput it is an appealing method to validate binding modes of ligands because it is simple, easy to use and affordable. It is very sensitive and particularly valuable for screening weakly binding ligands that otherwise might be missed. It also gives information about where the ligands bind. In contrast, X-ray crystallography has the advantage of providing information on not only where but also how the ligand binds.

3.5.4 X-ray crystallography

Used as a validation tool, X-ray crystallography affords the opportunity to visualise compound binding in atomic detail. This information can then be used to monitor the progression of the compound at every iterative round of the optimisation process. The ability to use structural information dramatically increases the success of the fragment-based drug design process (Blundell et al. Reference Blundell, Jhoti and Abell2002; Hajduk & Greer, Reference Hajduk and Greer2007; Hartshorn et al. Reference Hartshorn, Murray, Cleasby, Frederickson, Tickle and Jhoti2005). The lower throughput requirements of a validation technique mean that the disadvantages of co-crystallisation, i.e. screening for new crystallisation conditions with each different compound are less problematic and therefore sustainable in a drug discovery campaign.

Fragment-based approaches are proving successful in targeting protein–protein interactions. An example has been the interface between the human recombinase RAD51 and the hub protein BRCA2. In this case, the analysis has progressed by two routes: the exploration of short peptides based on the natural ligand and fragment screening. Around 20 peptide–protein structures have been studied by X-ray crystallography and these demonstrated the subtle changes of conformation that allow peptide H-bonds to be optimised as the sequence changes. These structures of the peptide–protein complexes along with structures of protein–fragment complexes were used to design fragment-peptide chimaeric molecules, and these in turn served as templates for the design of other molecules (see Figure 9).

Fig. 9.

Structures of humanised RadA bound to a fragment and peptides. X-ray crystal structures of humanised monomeric RadA bound to a fragment (red), peptide (blue) and a fragment/peptide chimaera (green). Protein (grey) and compounds are all shown as van-der-Waals spheres (Scott et al. Reference Scott, Ehebauer, Pukala, Marsh, Blundell, Venkitaraman, Abell and Hyvönen2012).

An example of how a potential difficulty can be utilised to give useful binding information is seen from our own work on the HGF/SF splice variant NK1. This protein was crystallised by chance with the buffering component HEPES present in a pocket thought to be important in receptor binding. This gave valuable information on how ligands bind in this pocket and how it could be exploited for the development of inhibitors of the interaction with the Met receptor (see Fig. 8b). In an effort to investigate this further, seven buffer compounds have now been successfully soaked into NK1 crystals (Sigurdardottir, Reference Sigurdardottir2012). This information, in combination with results from fragment screening, will be used to evolve fragments into a lead-like molecule.

The propensity of PPIs to form crystal contacts can be exploited in the discovery of new binding sites as exemplified by recent work (Sledz et al. Reference Sledz, Zheng, Murzyn, Chruszcz, Zimmerman, Chordia, Joachimiak and Minor2010). Through crystal packing analysis of 19 different crystal forms of the polo-box domain (PBD) of polo-like kinase 1 (Plk1), Sledz et al. observed a previously unreported binding site formed by rearrangement of surface residues involved in crystal packing. This binding site was shown to participate in binding to a peptide derived from a biologically relevant ligand by using biophysical techniques, mutagenesis and X-ray crystallography.

3.5.5 Bioassays

Biochemical assays exploiting the enzymatic activity of the target protein have the advantage of being high throughput and low cost, so there has been interest in utilising such assays in fragment-based drug discovery. However, bioassays require inhibition of interactions, whereas most biophysical methods detect binding to a protein interface of one component rather than disrupting the interaction. Nevertheless, Geitmann et al. (Reference Geitmann, Elinder, Seeger, Brandt, De Esch and Danielson2011) employed a primary SPR-based fragment screen followed by an interaction and an enzyme inhibition assay for stepwise hit selection and validation of a novel human immunodeficiency virus type 1 reverse transcriptase (HIV-1RT) inhibitor. For non-enzyme targets, a variety of biochemical, biophysical and cell-based biological assays may be useful, such as ALPHA-screen (PerkinElmer), assays assessing enzymatic activity of down-stream effectors such as the phosphorylation assay (Ferraris et al. Reference Ferraris, Gherardi, Di, Heinz and Niemann2009) or migration assays (Boyden, Reference Boyden1962).

In some fragment-based screening campaigns, biological assays have proved remarkably helpful. For example, apo-crystals of the Met receptor were not available for compound soaking, so instead a variety of biological and biophysical assays were explored to assess the biological activity of fragments. Phosphorylation assays proved to be a fast and easy way to probe for inhibition of downstream signalling [extracellular-signal-regulated kinase (Erk) and Akt phosphorylation] caused by the fragments and early compounds, in the presence of stimulating HGF/SF. In this assay, Vero cells were grown to confluency, starved and treated with 0.1nM HGF/SF in the presence of compound. Several fragments and compounds found from initial fragment hits using substructure and similarity searches were found to prevent Akt and Erk phosphorylation indicating biological activity (A. Winter and E. Gherardi, unpublished results).

However, most fragments are too small to exhibit biological activity and several rounds of fragment growth or evolution are usually required before cell-based assays can be utilised.