3.1. Start and end points

The first step in construct design is to specify which domain(s) is (are) necessary for ligand binding or which function of the protein is being targeted by the ligand, as discussed above for BRC-Abl. The aim from a crystallo­graphic perspective is to identify the boundaries of the structured domain necessary to achieve ligand binding or maintain protein function, while excluding large unstructured regions which might hinder crystal formation. A good starting point is sequence alignment with closely related proteins of known structure, or the use of domain-recognition tools such as PFAM (Finn et al., 2014) or pDomTHREADER (Lewis et al., 2013).

In addition, secondary-structure and disorder prediction tools such as PSIPRED, RONN or DISOPRED (Buchan et al., 2013; Yang et al., 2005) can be used to avoid `cutting' into continuous secondary-structure elements and to minimize the incorporation of long disordered regions at the boundaries. The amino-acid composition at the termini of the protein fragment is also worth considering with respect to the occurrence of serine, proline or aromatic residues. Savitsky et al. (2010) analysed the amino-acid composition at the termini of different protein fragments compared with the termini of the corresponding full-length proteins. They found an over-representation of proline and serine residues, and an under-representation of hydrophobic residues, in the protein fragments which successfully crystallized. These residue types are also reported to have higher than average propensity for disorder (Linding et al., 2003). It can be speculated that the higher flexibility at the very termini might possibly contribute to construct solubility and fewer conformational restraints during crystal packing.

Once initial protein fragments have been designed, algorithms aimed at scoring these in terms of their probability to crystallize can be incorporated. For example, the data generated by structural genomics efforts on all of the protein constructs tested versus successful constructs were used to inform machine-learning approaches as implemented in XtalPred-RF (Jahandideh et al., 2014) for scoring different protein constructs.

One should keep in mind that the methods used for these predictions bear some degree of uncertainty, and the design and testing of multiple constructs around the predicted boundaries are recommended.

3.2. Conformational state

Many proteins depend on a chemical modification or a binding partner to switch between active and inactive states; the transition is often associated with extensive conformational change.

For protein kinases, the largest protein family in eukaryotes (Nolen et al., 2004), a number of different types of inhibitors have been developed. Type 1 (ATP-competitive) inhibitors bind to the kinase ATP-binding site and inhibit the protein in its active conformation. Non-ATP-competitive inhibitors inhibit the enzyme by stabilizing an inactive conformation through binding to a pocket that is not present in the active state. For some kinases, the switch between an inactive and the active conformation requires phosphorylation in the so-called activation loop, a flexible polypeptide region in the vicinity of the ATP and protein substrate-binding site, while others can also adopt the active state if nonphosphorylated (Cowan-Jacob et al., 2007). This means that for kinases, but also many other proteins, a further variable in the ligand co-crystallization experiment is the phosphorylation state, which will depend on the protein expression platform. It is not uncommon to find a mix of different phosphoryl­ation states in a protein prep­aration, and also variation in the phosphoryl­ation patterns between different protein preparations. As the resulting sample heterogeneity can hinder crystallization, it is recommended to separate the different states by ion-exchange chromatography. To shift the ratio towards the desired phosphorylation state, the protein can either be co-expressed with a phosphatase, phosphorylated in vitro or dephosphorylated during the purification process. Also, the presence of inhibitor during protein expression can have an effect on the phosphorylation profile. Cowan-Jacob et al. (2007) showed that Abl kinase is heterogeneously phosphorylated when expressed in Sf9 insect cells. Upon the addition of a kinase inhibitor during expression, they found that the amount of non­phosphorylated protein is increased, and in some cases also the overall protein yield. Alternatively, mutations can be introduced that either prevent phosphorylation by removing the site of phosphoryl­ation or mimic it by introducing a negative charge (Mace et al., 2013).

Conformational changes can also be induced through co-activators or repressors, and this might need to be considered in structural studies of ligand complexes. In the case of HCV NS3 protease, the activity of NS3 increases upon binding of the HCV NS4 peptide by anchoring the active-site residues, and crystal structures of NS3 bound to different protease inhibitors have been generated using the NS4 peptide, either by addition of the NS4 peptide to the NS3 protein prior to crystallization (Yan et al., 1998) or by constructing an NS3/NS4 fusion protein (Romano et al., 2012).

3.3. Regions of disorder

Within protein domains, there can be long stretches of disordered regions, which may hinder crystallization. In these cases, more sophisticated protein engineering might become necessary to displace these disordered residues. One approach would be to replace these regions of low complexity with equivalent residues of homologue proteins that are known or predicted to be less flexible.

An example of the impact of flexible regions is the human cGMP-specific phosphodiesterase PDE5A1, which is the target of the drug sildenafil (Viagra^TM). Several groups determined the structure of the catalytic domain of human PDE5A1 independently. Sung et al. (2003) chose a construct comprising residues 537–860. In the co-crystal structure with sildenafil, tadalafil and vardenafil, a loop which differs in length and sequence from other PDEs is largely disordered. Zhang et al. (2004) described a similar construct to determine the structure of PDE5A1 in the unbound state, and found that the loop was folded into and blocking access to the active site, further demonstrating the large conformational flexibility of the loop. To facilitate further structural studies, Zhang and coworkers replaced the loop region by the equivalent residues in PDE4B2B, which adopts an ordered helical conformation in the PDE4B structure (Xu et al., 2000). Unpublished data from our own group show that the flexible loop is prone to proteolytic digest, resulting in batch-to-batch variation of the protein preparation and low reproducibility of ligand co-crystallization. Mass-spectrometric analysis and N-terminal sequencing showed that the protein is proteolytically cleaved at several sites within a short stretch of the protein sequence (Fig. 3). We also engineered a chimera protein in which the residues of the disordered region were replaced with the equivalent PDE4B2B residues, and found that ligands bound to the chimeric protein with equal affinity compared with the wild-type catalytic domain. The chimera crystallization system was not only highly reproducible, but the new crystal form allowed ligand soaking.

Figure 3 Construct design of the PDE5/PDE4 chimera protein. (a) Domain boundaries of the original PDE5A1 catalytic domain construct with disordered residues highlighted in blue and the proteolytic sites marked. (b) Sequence alignment between the loop region of PDE5A1 (top row) with equivalent residues in PDE4B2B (bottom row) and the sequence of the PDE5A1/PDE4B2B chimera protein (middle row). (c) Cartoon representation of the crystal structure of PDE4B2B (PDB entry 4nw7) with the loop insertion region highlighted in orange, created using CCP4mg molecular-graphics software (McNicholas et al., 2011).

Alternatively, the residues predicted to be disordered can be cut out and the ends joined by short flexible linkers. To obtain the co-crystal structure of capsid protein (CA) N-terminal assembly domain (NTD) with PF-3450074, it was necessary to create a loop-truncation mutant in which a flexible loop is replaced by a single glycine residue (Blair et al., 2010). The capsid protein of HIV-1 is the primary structural protein of HIV, and is involved in both the assembly of viral particles and the infection of host cells. The N-terminal assembly domain is required for mature capsid assembly and it is possible to interfere with viral uncoating and formation of infective particles in infected cells with small-molecule drugs by targeting the NTD. Although the loop truncated for structural studies is important for viral infection, truncation does not affect compound binding, as validated by affinity measurements, and comparison of the crystal structures with other CA NTD structures with the loop residues present in the protein construct did not show any significant differences.

As these loop-deletion or replacement mutations introduce more extensive changes to the primary sequence of the protein, it is important to verify that they do not affect compound binding biochemically. One should further check for consistency in the structure–activity relationship of the compounds. The artificial nature of the construct should further be kept in mind when setting the structural information into a wider context, such as isoform specificity or selectivity.

3.4. Surface mutations

On their surface, proteins commonly harbour residues with flexible, charged side chains, which are often found in patches. These are suggested to have evolved as a method to prevent undesired protein–protein interactions in the cell (Doye et al., 2004). Crystal formation, however, is reliant on protein–protein interactions to mediate crystal packing. An early study by McElroy et al. (1992) describes the systematic mutation of nonconserved surface residues and their effect on crystallizability. For human thymidylate synthase, they created a panel of 12 single surface mutants in which charged residues were replaced with neutral amino acids or with amino acids of opposite charge. They found that many of these surface mutants were more crystallizable than the wild-type protein and that a single arginine-to-glutamate mutation produced seven times more crystal hits compared with the wild type. Later, Longenecker et al. (2001) reported an alternative approach in which, for patches of large hydrophilic side chains, the residues were mutated to alanine in order to overcome the resistance of a protein construct to crystallize. For human RhoGDI, they tested 13 mutants with one or several surface residues mutated to alanine, and found new crystal forms for two single and two triple mutants. Derewenda (2004) discusses a number of other examples where this approach of surface-entropy reduction (SER) have been successfully used. He and others subsequently developed the surface-entropy reduction prediction server SERp in order to help and identify possible sites of high surface entropy (Goldschmidt et al., 2007).

As for construct design in general, it is also recommended to test several combinations of surface mutants in parallel. Sorrell et al. (2016) describe their efforts to generate the first crystal structure of the BIKE kinase domain. BIKE (BMP-2-inducible kinase) is a Ser/Thr kinase of the family of Numb-associated kinases (NAKs), and in an initial attempt the authors prepared over 50 BIKE kinase domain constructs, of which less than 20% showed soluble expression but none could be crystallized. Six different SER mutants were prepared for the construct which showed the highest expression, mutating up to three lysine residues at a time to alanines. Well diffracting crystals were obtained for a double mutant in which the alanine residues are not directly involved in crystal packing but a longer, charged side chain instead of the alanine side chain would have caused repulsion and/or steric clash for one of the mutations.

An example of increasing protein solubility via surface mutations is the glucocorticoid receptor (GR). GR receptors belong to the superfamily of nuclear receptors (NRs). They are regulated through hormone binding to the ligand-binding domain (LBD), and the LBD is targeted by drugs such as prednisolone. Bledsoe et al. (2002) reported their approach to increase the soluble expression and purification levels of the GR LBD. Through sequence alignment with other NRs, they identified a single phenylalanine that, when mutated to serine, facilitated soluble expression of the NR in Escherichia coli in the presence of ligand. When Schoch et al. (2010) attempted to use this mutant for co-expression with a different ligand, they did obtain soluble protein, but were not successful in generating co-crystals for structure determination. They therefore screened the existing GR LBD structures for clusters of lysine and glutamic acid residues on the surface and distant from the ligand-binding site, and identified a double alanine mutant that showed increased thermal stability in the presence of ligand compared with the wild-type sequence. The new construct could be used to determine the structure of their ligand of interest.

3.5. Troubleshooting

The most obvious problem in crystal structures aimed at determining protein–ligand interactions is when parts of the protein construct block or bind to the active site, as discussed for PDE5 above. Another example is the apo structure of the catalytic domain of PDE2a, where the terminus of a neighbouring chain binds into the active site (Iffland et al., 2005).

In the case of HDAC4, inspection of the first ligand-bound structures of the catalytic domain in the public domain suggested crystal packing to have a distorting effect on the binding mode of the ligand. To overcome this, we designed an alternative protein construct using the boundaries of the successfully crystallized catalytic domain of HDAC7 as a guide (Schuetz et al., 2008), in particular the first and last residues modelled in the HDAC7 structure (PDB entry 3c0y; Fig. 4a). Co-crystallization trials with this construct yielded ligand structures that were not biased by the same crystal-packing effects observed for the original construct, but further attempts to co-crystallize other ligands or to repeat the initial trials failed. Inspection of the crystal packing revealed that a leucine residue of a neighbouring chain packs closely against the ligand (Fig. 4b). Within the protein chain, the leucine was distant from the active site and at the terminus of a disordered loop. Hypothesizing that mutation of the leucine to a less bulky residue would allow a slightly looser packing interaction without affecting protein function, we introduced a leucine-to-alanine mutation, which retained enzyme activity compared with the wild-type protein, as expected. Crystallization conditions were readily identified for the single point mutant in the presence of a variety of different ligands, yielding a novel crystal form devoid of any packing interactions around the active site, and with the formerly disordered loop visible in the electron density (Bürli et al., 2013).

Figure 4 Rounds of construct optimization for the HDAC4 histone deacetylase domain. (a) The construct boundaries of the original HDAC4 catalytic domain structure (PDB entry 2vqj, top) and the HDAC7 catalytic domain in PDB entry 3c0y (bottom) are shown. The HDAC7 numbering is aligned with the HDAC4 numbering. The first residue modelled in the original HDAC4 structure corresponds to the first residue modelled in the HDAC7 structure and was used as the starting point for the optimized HDAC4 construct (middle). The C-terminus of the optimized HDAC4 model was chosen based on the last residue visible in the HDAC7 structure. This shortened the new HDAC4 construct by about 20 residues compared with the original HDAC4 structure, and the new C-terminal boundary is highlighted in the cartoon representation of HDAC4 (PDB entry 2vqj). It was speculated that this truncation would help to generate an alternative crystal form with more favourable packing contacts. (b) Crystal structure of the new HDAC4 construct with bound ligand (PDB entry 4cbt). The new HDAC4 construct did result in an alternative crystal form. Here, close crystal contacts are observed between ligand (yellow) and Leu728 (magenta) at the terminus of a disordered loop in a neighbouring chain (blue). These contacts were thought to hinder co-crystallization with larger ligands. (c) In a second round of construct optimization, Leu728 was mutated to alanine (green) and the mutant readily crystallized in the presence of ligand in yet another crystal form devoid of packing interactions at the ligand-binding site (PDB entry 4cby). In addition, a loop that was disordered in the corresponding wild-type structure could be modelled into the electron density (orange). Figures were created with the CCP4mg molecular-graphics software (McNicholas et al., 2011).

Another common problem is the need to switch to an alternative crystal form when the original system does not allow crystal soaking. This can be achieved by the identification of alternative crystallization conditions, subtle changes to the construct boundaries as described above or by more drastic changes to the construct. Clifton et al. (2015) describe how an MBP-fusion protein was used to successfully generate a soakable crystal system for Mcl-1. Mcl-1 is a member of the Bcl-2 family of proteins that regulate apoptosis. The sequence N-terminal to the Bcl-2 domain is predicted to be of low structural complexity and to be highly disordered, and structures of Mcl-1 co-crystallized with ligand comprised the Bcl-2 domain only. Involvement of the ligands in packing inter­actions in all but one of these structures suggested crystallization success to be highly ligand-dependent. Clifton and coworkers therefore engineered a MBP-Mcl-1 fusion protein, undergoing multiple rounds of construct optimization. Firstly, they fused a range of different partners including MBP, lysozyme, thioredoxin (Trx) and SUMO N-terminal to the Bcl-2 domain of Mcl-1, but none of the fusion proteins tested yielded crystals in the absence of ligand. To aid crystallization, they further introduced SER mutations on the Mcl-1 protein (K194A, K197A and R201A) in a stretch of ten highly disordered residues. Further, they added linker residues between Mcl-1 and MBP to reduce steric clashes between the two domains: a short GS and a longer GSGGGG linker. The short linker in combination with the SER mutations and addition of the MBP ligand maltose allowed crystallization of the protein in the absence of Mcl-1 ligands. Analysis of the structure showed that the overall fold of Mcl-1 is similar to other Mcl-1 structures in the PDB derived from both NMR as well as X-ray diffraction data. The SER mutations introduced are involved in key packing interactions. Also, the ligand-binding groove is accessible from the solvent channels, and it is possible to generate ligand co-crystal structures via soaking with this new Mcl-1 crystal form.

5.1. Soaking

A prerequisite for soaking is the existence of a soakable crystal form. Inspection of a crystal structure can indicate whether soaking should be possible: close crystal-packing contacts around the ligand-binding site (PDE2A, as mentioned above) or around regions expected to undergo conformational change upon ligand binding might be problematic. When checking for potential crystal-packing issues, it is recommended to not only look at the atomic model but also at the electron-density maps. Flexible regions or affinity tags might not have been modelled, and features in the electron-density maps might reveal whether these are close to and possibly interfering with the ligand-binding site.

The protein does not necessarily need to have been crystallized in the ligand-free form to be suitable for soaking. Co-crystallized ligands can often be displaced, even by ligands with equal affinity. In the case of apoptosis signal regulating kinase 1 (Ask1), the co-crystallized inhibitor was replaced by an equally potent compound (IC₅₀ for both in the sub-micromolar range) by soaking at 10 mM for a couple of days (PDB entries 4bib and 4bic; Singh et al., 2013). Indeed, for ligands which induce large conformational changes that are not tolerated by crystal packing in the apo form, co-crystals of a similar ligand would seem to be a better starting point for soaking trials. For allosteric modulators, it might be necessary to have a ligand present in the active site, either to block the site to avoid false-positive hits or to induce a particular conformation required for binding. AGC kinases have a small phosphate-binding site in the N-lobe, which is believed to play a key regulatory role. This binding pocket is disrupted in the crystal structures of AGC kinases in the inactive conformation but is ordered in the active conformation. For the AGC kinase PDK1, small molecules that bind to this phosphate-binding pocket were successfully introduced by soaking into crystals of ATP-bound PDK1, i.e. the protein in its active conformation (PDB entry 3hrf; Hindie et al., 2009).

5.1.1. Experimental setup

The simplest way to soak a compound into a crystal is by adding either a concentrated stock solution or pure compound directly to the crystals in the crystallization drop. This reduces crystal manipulation, especially for fragile crystals, and allows automation of the process (Krojer et al., 2017). The main disadvantage is that only one ligand can be explored per drop, potentially wasting many crystals if multiple crystals are present. When the ligand is added, especially in the solid form, it is recommended to first remove any skin that might have formed on the crystallization drop, both to allow diffusion of the compound and to not hinder crystal harvesting later. In the case where cryoprotection is applied after crystal soaking, ligand should also be present in the cryo-buffer to avoid soaking the ligand back out of the crystals. Depending on the binding kinetics, this can happen on a very short timescale.

Alternatively, the crystals can be harvested into a stabilizing solution before soaking. This solution can either be the crystallization well solution, possibly supplemented with cryoprotectant, or have a more complicated composition, as binding of ligand to the protein can change protein solubility. This is sometimes revealed by disappearing protein crystals upon ligand addition. Increasing the precipitant concentration can counteract this increase in protein solubility and help to maintain diffraction of the crystal after soaking, and in some cases, effectively through crystal dehydration, improve resolution (Russo Krauss et al., 2012). When salt is used as the major precipitant in the crystallization, high concentrations can decrease solubility, especially for hydrophobic ligands, and also cause issues for cryoprotection. It has been shown that addition of polyethylene glycol (PEG) to an aqueous solution as a co-solvent can help to increase compound solubility (Rytting et al., 2005). Replacing salt with PEG in order to increase ligand solubility is often possible without damaging the crystals. Also, the pH of the stabilizing solution can be shifted into the range required for ligand binding. If a protein was co-crystallized in the presence of another ligand, the stabilizing solution can be used to soak this ligand out, possibly in several steps, by transferring the crystal into fresh drops over hours or days before soaking with an alternative ligand.

The ligand concentration required for soaking depends on the affinity of the ligand and, together with the soaking time, needs to be optimized. The aim is to achieve that in excess of 90% of protein (P) in the sample is bound to ligand (L). This will depend on the ligand concentration [L] and the dissociation constant of the PL complex K_d,

The ratio of bound to total protein can therefore also be expressed as

For >90% of protein to be bound in the complex, the required free ligand concentration is ten times the K_d or greater. In practice, ten times the total ligand concentration is a good guide for soaking experiments. Yet, often, the binding affinity is not known. In these cases, we would aim for 20–50 mM for fragment-like compounds and 0.1–1 mM for higher molecular-weight ligands, their solubility permitting. Rapid visual deterioration of the crystals would suggest that the concentration chosen was too high, as would heavy precipitation of the ligand in the stabilization buffer. Crystals harvested at different time points (<1 h, several hours, overnight) can be checked both for diffraction as well as ligand occupancy. Also, a slow, stepwise increase in ligand concentration can help to maintain diffraction. If crystals show signs of deformation (especially plate-like or needle-like crystals) or appear cracked, it is still worth checking the diffraction before discarding the experiment (Fig. 5). If soaking optimization did not help to reduce crystal decay, data collection at modern synchrotrons and using a single-photon-counting pixel detector can further help to obtain processable data and, more importantly, interpretable electron-density maps. If the diffraction is poor despite optimizing the ligand concentration and stabilizing solution, cross-linking with glutaraldehyde prior to soaking might help to stabilize the crystals (Lusty, 1999). In essence, protein crystals are equilibrated through the gas phase against a solution containing glutaraldehyde. The glutaraldehyde reacts with and forms cross-links between lysine side chains on the protein surface, ideally tightening the crystal contacts and making the crystal more resilient to conformational changes during soaking. In terms of experimental setup, we generally add 5 µl of 10–25% glutaraldehyde to a microbridge or sitting-drop well and add 50 µl of crystallization or stabilization solution to the reservoir. Crystals are placed on a cover slip, either in the crystallization or a stabilizing solution, suspended over the well and equilibrated from between 5 and 10 min up to several hours before harvesting for the soaking experiment.

Figure 5 Diffraction image for a crystal apparently damaged during soaking. For the example shown, a data set could be collected and the structure solved without significant loss in resolution or map quality compared with nonsoaked apo crystals.

5.2. Co-crystallization

For co-crystallization, the ligand can be added to the protein during crystallization setup, and Gelin et al. (2015) even suggest `dry' co-crystallization by pre-coating crystallization wells with ligand. More commonly, the protein is pre-incubated with the ligand. Initial crystallization screens are performed in the same way as any protein-crystallization experiment (Ng et al., 2016), avoiding conditions that are known to disrupt ligand binding (Müller et al., 2011). If available, the drops can be cross-seeded using crystals of either the apo-protein or another ligand complex.