Modelling three-dimensional protein structures for applications in drug design

Abstract

A structural perspective of drug target and anti-target proteins, and their molecular interactions with biologically active molecules, largely advances many areas of drug discovery, including target validation, hit and lead finding and lead optimisation. In the absence of experimental 3D structures, protein structure prediction often offers a suitable alternative to facilitate structure-based studies. This review outlines recent methodical advances in homology modelling, with a focus on those techniques that necessitate consideration of ligand binding. In this context, model quality estimation deserves special attention because the accuracy and reliability of different structure prediction techniques vary considerably, and the quality of a model ultimately determines its usefulness for structure-based drug discovery. Examples of G-protein-coupled receptors and ADMET-related proteins were selected to illustrate recent progress and current limitations of protein structure prediction. Basic guidelines for good modelling practice are also provided.

Introduction

The goal of drug discovery is to contrive bioactive molecules that efficaciously modify a disease in a way that is beneficial to the patient, while keeping adverse effects such as toxic responses controllable. On the molecular level these requirements translate into a picture where a drug molecule binds to one or more target proteins that are implicated in the pathophysiology of a disease and act as, for example, inhibitors, agonists or modulators. At the same time, binding to proteins that have a negative impact on efficacy, or cause unwanted side effects, has to be avoided. In this sense, drug design is an enterprise that aims to engineer molecules with a controlled interaction profile against a multitude of different target and off-target proteins in an organism. During the initial target validation and hit finding phases of a drug discovery programme the focus is usually on the main target and then, as a programme progresses into lead optimisation, the attention shifts to the interplay of the drug candidate with an increasing number of proteins. Obviously, a full characterisation of these interactions down to the 3D structural details would constitute a profound structural perspective of the mode of action (MOA) of a drug molecule, and hence greatly facilitate drug design. Nowadays, a vast amount of experimental structural data, mainly generated by X-ray crystallography, is available [1]. Yet, the number of known protein sequences vastly exceeds the number of corresponding 3D structures. This so-called sequence– structure gap implies that for many important proteins there are no structures available. Fortunately, 3D protein structure prediction often offers an appropriate remedy in such situations [2,3]. In this review, we discuss the current status, applicability and limitations of protein models derived from protein structure prediction methods. We briefly introduce the prevailing prediction methods, with a focus on their relevance in drug discovery. Using selected examples, we also demonstrate typical applications at various stages of the drug discovery process.

Although it might appear trivial, it is worth emphasising one of the most important achievements of protein 3D structure modelling: the transformation, integration and contextualisation of heterogeneous information, such as mutation and SAR data, in a 3D model. Numerous visualisation tools have been developed for inspecting, analysing and annotating such models [4]. Visualisation of valid models is not merely a decorative offshoot of modelling, rather a focal point where disparate facets of research efforts can amalgamate and converge into a detailed view of the underlying mechanistic basis, which in turn can become the driving force for further advances. It should be kept in mind that, even at relatively low resolution, ‘any level of physical characterisation of a protein, as opposed to its absence, is valuable’ [5].

Protein and binding site flexibility

Proteins are intrinsically dynamic systems that can exhibit significant flexibility and structural plasticity, also in their drug binding sites. A single structural model embodies only a static snapshot, regardless of whether it is an experimental or a predicted structure, and can therefore not always capture all the relevant characteristics of a protein. In essence, there is no entity such as ‘the’ structure of a protein, and this principle also applies to experimental structures for which, in addition to the issues associated with protein flexibility, experimental conditions, structural errors [6] and crystal packing effects [7] must be taken into account. Consideration of target and binding site flexibility is of paramount importance in computer-aided drug design (CADD), and disregarding them can dramatically hamper its success. Consequently, appropriate treatment of protein flexibility has become a major effort [8,9]. The ligand-steered modelling approaches outlined below have emerged as a result of these challenges.

Methods for protein modelling

Computational methods for predicting 3D protein models are widely used in the pharmaceutical industry, and much effort has been invested in improving model accuracy, and in expanding the scope of these methods (Table 1). Methods are generally categorised into template-based (i.e. homology) modelling and de novo modelling [10,11]. Traditional homology modelling (or comparative modelling) is considered to be the most accurate of these methods, and is thus most commonly applied in drug discovery research [12]. Homology modelling is based on the fundamental observation that all members of a protein family persistently exhibit the same fold, characterised by a core structure that is robust against sequence modifications [13]. It relies on experimentally determined structures of homologous proteins (templates), and enables the generation of models starting from given protein sequences (targets). The most accurate models can be obtained from close homologue structures; however, even with low sequence similarity (~20%) suitable models can be obtained [14,15].

Table 1.

Frequently used servers and tools for protein structure homology modelling

Resource	Refs and URL
Protein Model Portal	[94]	http://www.proteinmodelportal.org
HHpred	[95]	http://toolkit.tuebingen.mpg.de/hhpred
ICM	[96]	http://www.molsoft.com/
IntFOLD	[97]	http://www.reading.ac.uk/bioinf/IntFOLD/
Modeller, ModWeb	[98]	http://salilab.org/modeller/
Phyre2	[99]	http://www.sbg.bio.ic.ac.uk/phyre2/
Robetta	[100]	http://robetta.bakerlab.org/
SWISS-MODEL	[101]	http://swissmodel.expasy.org/

A homology modelling pipeline generally comprises the following steps which can be repeated until a suitable model is obtained: (i) template selection for identifying the most suitable experimentally determined structures; (ii) target–template sequence alignment; (iii) 3D model structure building; (iv) model refinement; and (v) model quality estimation. Model refinement usually involves clash removal and geometrical regularisation of bond lengths and angles but can also involve additional more-sophisticated structural amendments. As a rule of thumb, most attention should be devoted to steps i, ii, iii and v, whereas global model refinement (iv) typically has a disappointing return on investment [16].

LSM: ligand-steered modelling

As mentioned above, appropriate modelling of the binding site and correct ligand placement are of the utmost importance in CADD. However, native protein ligands such as enzyme substrates or signalling molecules often exhibit only weak binding affinities and are therefore often lost during purification procedures. As a result, protein structures are often determined experimentally in the absence of ligands. Additionally, template selection procedures in traditional homology modelling are often based on sequence similarity as the only criterion, neglecting ligand information in the template structures. As a consequence of this, the resulting protein models often represent an unliganded state of the binding site.

Classically, docking approaches have been used to place the ligands into the binding sites of the final homology models as a post-processing step [17–19]. The shortcomings of this practice have been addressed by developing more ligand-aware approaches that treat ligands as an integral part of a model throughout the entire modelling process. Generally, two strategies can currently be distinguished. First, ligand-guided (or steered) receptor modelling (LSM) directly incorporates ligands in the modelling process for guiding the protein conformation sampling procedure. One pioneering approach is binding site remodelling, which uses restraints obtained from initially modelled complex structures to build a second refined model [20]. Such approaches often require expert knowledge and time-consuming manual intervention, and hence call for the development of fully automatic homology modelling pipelines. Dalton and Jackson [21] have developed and assessed two variants of LSM, both yielding significantly more accurate complex models than docking into static homology models, regardless of whether or not the ligand had been incorporated into the modelling process. The most successful variant utilises geometric hashing and shape-based superposition of the ligand to be built onto a known ligand in a template structure, prior to the modelling procedure. Generally, ligand-guided approaches can lead to highly accurate models but can be hindered by the fact that correct ligand placement is intrinsically linked to correct side-chain modelling, and even small inaccuracies can prevent the correct prediction of relevant interactions. The second approach, termed here: ligand-guided receptor selection, utilises a large number of homology models from which the model yielding the highest enrichment in docking calculations against known active and decoy compounds is determined [22]. Model generation usually encompasses extensive sampling of side chains in the binding cavity, but can also be extended to incorporate variations in the backbone conformation [23]. This method has recently been extended to a fully automated iterative sampling-selection procedure to generate an ensemble of optimised conformers [24]. This approach has the advantage that the models are optimised for a particular purpose; however, it is limited to cases where high-affinity ligands are known.

Model validation and quality estimation

Homology models are computationally derived approximations of a protein structure and can contain significant errors and inaccuracies. It should be noted that the quality required for a model depends largely on its intended use. For example, low-accuracy models can be completely sufficient for designing mutagenesis experiments, whereas structure-based virtual screening (SBVS) applications require greater accuracy [15], and for mechanistic studies the highest level of accuracy possible is essential [2,11]. Although the accuracy of a protein modelling method can be evaluated a posteriori based on experimental structures [14], the quality of an individual model can vary significantly and the a priori estimation of model quality it therefore of great importance. Common methods for estimating model quality use a combination of stereochemical plausibility checks, knowledge-based statistical potentials, physics-based energy functions or model consensus approaches [25–28]. Different scores have been developed for tasks ranging from ranking of an ensemble of models on a relative scale to the prediction of the absolute accuracy on a per residue basis.

Hit finding and virtual screening

Virtual screening (VS) has matured into an invaluable approach for identifying active compounds against drug targets by means of ‘smart’ computational approaches [29]. Basically, SBVS is the automated positioning (docking) of different 3D conformational models of compounds (poses) into a suitable binding site of a 3D protein structure. Subsequent post-processing of these poses aims to identify the compounds that are most likely to be active. See, for example, the reviews by Klebe [30], Waszkowycz [31] and Cheng et al. [32] for overviews. In the absence of appropriate experimental 3D structures, homology models can be used as an alternative. The usefulness of homology models in SBVS against many different targets has been demonstrated in various retrospective analyses [33–36]. A comprehensive survey of the scientific literature on prospective VS campaigns has also been published, analysing a total of 322 SBVS campaigns [37]. Out of these, homology models were successfully utilised in a total of 73 studies. Surprisingly, the potency of the hits identified using homology models was on average higher than for hits identified by docking into X-ray structures.

The selection of the most suitable model for docking from a pool of generated models remains a problem. Notably, there is only a weak correlation between global model quality parameters and docking success [15,34,38], indicating that the discrepancy between global and local structural accuracy cannot be adequately expressed by these measures. One fundamental reason is that protein flexibility, and in particular the adaptation of the ligand binding cavity to the bound ligand, can have a more significant impact on docking success than relatively small global modelling errors. Moreover, binding site residues generally exhibit a higher degree of conservation than the rest of a structure [39]. As a result, fairly accurate binding site models can be obtained from models with relatively poor overall quality [14,40].

There is a notion that models built on templates with sequence identity >50% are accurate enough for SBVS; however, this rule should be treated with caution. In a large-scale benchmarking study on 38 targets, Fan et al. [15] showed that the most enriching single models generally outperformed the apo X-ray structures, and even performed comparably to the holo X-ray structures. Docking into multiple models combined with consensus scoring further increased the enrichment rates, and was comparable to using the holo structure. This shows that the success of SBVS campaigns is intimately linked to adequate treatment of binding site flexibility. Approaches for handling protein flexibility in docking have been reviewed recently [41,42], and include docking to an ensemble of structures (ensemble docking) and ligand-guided homology modelling described above. Recently, basic guidelines for selecting the best binding site for docking from an ensemble of X-ray structures have been published [43], and it will be interesting to examine whether such rules could be conferred to homology models.

Applications

Homology modelling has been applied to various areas of drug discovery including structure-related aspects of target validation such as site-directed mutagenesis and druggability assessment [44,45]. This review is confined to exemplifying recent progress in ligand-associated protein modelling, namely SBVS, and mapping of protein–ligand interactions and mutations. Many compelling studies involving diverse intricate proteins have been published recently, and cannot be appreciated here in detail but deserve attention. These include, for example, antibodies [46], gamma-aminobutyric acid (GABA)-A receptor [47], ion channels [48], cystic fibrosis transmembrane conductance regulator (CFTR) [49] and epigenetic targets such as DNA methyltransferases (DNMTs) [50], histone deacetylases (HDACs), sirtuins and HMTs [51].

G-protein-coupled receptors

G-protein-coupled receptors (GPCRs) represent the most prominent target class against which almost a third of all FDA-approved drugs are targeted [52]. Owing to the enormous difficulties of crystallising GCPRs, the era of GPCR structural biology began relatively late [53], with the first human GPCR structure, β2-adrenergic receptor (β2-AR), solved in 2007 [54]. At present, a total of 16 unique class A GPCR structures are available, and large efforts are being made to characterise more representative GPCR structures, including class B and C GPCRs [55]. Given that there are more than 800 GPCR genes in the human proteome, it is evident that there will be great demand for reliable structural models of GPCRs in forthcoming years.

GPCR structure prediction faces a number of challenges, inter alia the adoption of multiple conformations depending on their activation state [56] and interaction with different adaptor proteins that actuate multiple signalling pathways [57]. Whereas, for example, the differences between the orthosteric β2-AR binding pockets in its active and inactive state are small, the binding site volumes differ [58]. Related to this, subtle changes in chemical structure can turn agonists into antagonists, and vice versa [59]. By contrast, orthosteric GPCR binding sites represent deep cavities that are well buried between the transmembrane (TM) helices, and should therefore be particularly suitable for SBVS.

The usefulness of GPCR models in SBVS has been assessed in a number of publications: see, for example, a brief overview [58], and a description of an automated SBVS workflow [60]. Guidelines for building GPCR models to be used for docking have been worked out recently [61]. Although loop modelling, in spite of recent progress [62,63], is still a challenge, it has also been demonstrated that loop-less models can be successfully applied in SBVS [64]. In fact, carefully built models can even outperform X-ray crystal structures in docking performance, subject to selecting templates representing the correct activation state [65]. For a comprehensive discussion of retrospective and prospective SBVS campaigns with experimental and modelled GPCRs, including model building and VS approaches, incorporation of experimental data and results, we refer the reader to the extensive recent review by Kooistra et al. [66].

In a community-wide assessment it has been concluded that GPCR structures in complex with small molecules can be reliably predicted, and can ‘approach the level of accuracy observed in the experiment’, if templates exhibiting >35% sequence homology (in the TMD) are used [67]. Using this criterion, it is estimated that roughly 20% of nonolfactory class A GPCRs can be reliably modelled based on currently available structural data [55]. It is also likely that the recent discovery of a conserved network of noncovalent contacts between the TM helices [68] will aid the further development of GPCR homology models.

Whereas reasonable enrichments can be achieved with imprecise models in SBVS, the correct mapping of protein–ligand interactions in 3D models is a more ambitious task, because optimal accuracy is imperative. A recent example on the mapping of protein–ligand binding in H4R has conclusively established that this is feasible [69]. Homology models from various templates have been built, and a combination of docking and molecular dynamics (MD) has been employed. The determination of ligand-binding modes that could explain experimental data was complicated by a quasi-symmetric distribution of certain residues in the binding site, leading to different plausible configurations. Nevertheless, careful analysis of the ligand poses and pocket volumes in connection with the experimental data, consideration of ligand protonation states and, for one ligand class, quantum-mechanical evaluation of low-energy conformations [70] enabled the elucidation of binding models that could expound ligand-specific mutation effects and subtle experimental SAR data.

ADMET, cytochrome P450s

Managing the interplay of drug molecules with ADMET-relevant off-targets such as, for example, metabolic cytochrome P450 enzymes (CYPs), drug transporters or the hERG channel is paramount in lead-optimisation campaigns. The CYP gene family comprises 57 functional genes in man, and plays a major part in metabolising xenobiotics into more polar compounds with improved elimination [71]. In 2004, merely three CYP structures were available, which were intensively used for the homology modelling of various CYP subtypes [19]. Nowadays most CYP structures and many other ADMET-relevant ones have been solved, allowing structure-based methods in ADMET to be increasingly applied [72,73]. The impact of CYP homology models in ADMET prediction is therefore likely to attenuate; however, one should not conclude that an X-ray structure is necessarily the best possible representation of a particular structural state. A recent study for predicting the substrate site of metabolism (SOM) in CYP2D6 alludes to the drawbacks related to protein flexibility raised above. A model of CYP2D6 was generated based on published X-ray crystal structures of the substrate-bound CYP2C5 [74]. During the study, the structure of apo CYP2D6 also became available. Docking calculations were performed with the new structure and the homology model. Although the homology model was, overall, in good agreement with the CYP2D6 crystal structure, the model consistently outperformed the experimental structure; this has been attributed to structural differences in the substrate recognition sites (Figure 1).

Figure 1.

Selection of an appropriate structural template for model building is crucial for its successful application in structure-based virtual screening (SBVS), and often the most appropriate template is not necessarily the one with the highest sequence identity to the target protein. In a study by Unwalla et al. [74] docking calculations performed on the apo X-ray structure of cytochrome P450 (CYP)2D6 were consistently outperformed by a model of CYP2D6, generated based on the substrate-bound structure of CYP2C5. A superposition of the CYP2D6 apo structure (red) (PDB: 2F9Q) and the CYP2C5 holo structure (blue) (PDB: 1N6B) is shown here. Panels a and b display the binding site surfaces of apo CYP2D6 and holo CYP2C5, respectively. Panel a highlights the fact that the binding pocket observed in the apo structure is too small to accommodate the ligand. Panel c shows backbone structural variations close to the binding site. Graphics were produced with OpenStructure [105,106].

Several CYPs, including CYP2D6, appear in numerous polymorphic forms. These inter-individual variations result in large differences for drug clearance and clinical response of patients [10], delineating ADMET as a highly individual-specific phenomenon. Currently, more than 90 2D6 alleles are known, corresponding to roughly 30 non-synonymous protein-coding single nucleotide polymorphisms (nsSNPs). Amino acid residues corresponding to these SNPs are distributed across the whole CYP structure; however, distinct clusters in proximity to the haem cofactor or the substrate access channel could be found [10]. It is evident that structural modelling of these variants can facilitate the interpretation of SNP-specific genotype–phenotype relations. This approach can be seamlessly extended to examine variations across different species. Such ortholog models have, for example, been derived for Macaca CYPs [75] and various estrogen receptors [76], and can assist the selection of appropriate models for pharmacological studies. It can be hoped that, ultimately, and with the advance of NGS, such models will enable the prediction of individual and species-specific drug responses, thus truly promoting personalised medicine and reducing animal sacrifice.

Multidrug resistance protein 1

Multidrug resistance protein 1 (MDR1), also known as P-glycoprotein (P-gp) or ABCB1, is a membrane-bound ATP-driven efflux pump that belongs to the family of ABC transporters. MDR1 is a promiscuous transporter that extrudes numerous hydrophobic compounds and drugs as its substrates from cells. MDR1-mediated drug efflux thus impairs drug delivery, and hence plays a key part in ADMET [77]. In many cancer cell types, upregulation of MDR1 activity causes fast removal of drugs, triggering multidrug resistance [78]. Hence, the interplay of MDR1 with its substrates, modulators and inhibitors is crucial in different areas of drug discovery. Homology modelling of MDR1 had been hampered by the low sequence identity and even structural errors in the available templates [79,80]. Even with the murine structure [81] as a template with 87% homology to the human protein there are many obstacles, which can probably be attributed to the huge polyspecific ligand-binding site composed of several subsites, the low resolution of the template structures and the large dynamic rearrangements that occur during the transport cycle [82], see also the review by Ravna and Sylte [83] on homology modelling of transporters. Despite these complications, an intriguing study establishing detailed binding hypotheses for known MDR1 inhibitors (profanone derivatives) has been published recently [84]. To capture the different catalytic states of the transport cycle, 100 models were generated for each of the apo (open-inward) and the nucleotide-bound (open-outward) states; the templates used reflected the various states. The best models were selected using the MOE (Chemical Computing Group, Montreal, QC) geometry check functions and validated with different quality estimation methods. Exhaustive docking of five selected propafenone ligands was performed using the entire TM area as the binding site. Post-processing and analysis was accomplished using protein–ligand interaction fingerprints, common scaffold clustering and incorporation of SAR knowledge. This SAR-guided docking protocol led to the selection of binding hypotheses that convincingly concur with experimental data and previous pharmacophore studies. Other recent stimulating studies using MDR1 homology models include an explanation of enhanced substrate bioavailability [85] and modelling of catalytic transitions based on targeted MD [86].

hERG

The hERG K⁺ channel represents one of the primary ADMET-related anti-targets, because undesired binding of drugs can lead to life-threatening arrhythmias and sudden death [87]. In spite of major efforts hERG homology models have been notorious for being descriptive at best, and not predictive for applications in drug discovery [72]. Clearly, segments in the hERG sequence with very low sequence identity to all possible templates are a major impediment of hERG modelling. An analysis of seven published hERG models with inconsistent alignments in the S5 sectors emphasises the crucial importance of sound alignments [88]. The analysis enabled the selection of a model that exhibited good quality criteria, and conformed best to experimental data. It also stressed the potentially detrimental impact of alignment errors in the drug binding site, even in segments remote from drug interactions. Further progress in hERG channel modelling concerns the elucidation of channel blocker trapping. Open- and closed-state homology models in combination with docking of propafenone derivatives have been reported [89]. In another study, 12 blockers with known activity along with mutagenesis data were used for validating open- and closed-state models, also using docking [90]. It was concluded that the models were reliable enough for explaining inhibitor binding and for identifying strong binders; however, 3D pharmacophores and QSAR methods still remain more predictive than the 3D models.

Concluding remarks and future outlook

With the tremendous growth of available structural data one could naively assume that the need for protein structure prediction would decline. On the contrary, we anticipate the opposite. As discussed above, the availability of an experimental structure does not necessarily imply that this structure represents the best starting point for CADD studies. Although, in general, an experimental structure will undoubtedly be better than a predicted model, this should not be presupposed axiomatically. Because structures can vary depending on their functional state and form (apo vs holo form; agonist (active) vs antagonist (inactive) state), their applicability for CADD should always be appraised in their functional context.

Another aspect that should not be underestimated is the rapid availability of models, particularly in the light of advanced fully automated modelling pipelines. In lead optimisation, X-ray crystal structures of relevant lead compound complexes are undoubtedly always desirable; however, a good structural model can be instrumental and guide research a long time before the experimental structures become available.

Most importantly, the availability of new structures opens avenues for reliably modelling many more structures than previously possible, as shown for GPCRs and MDRs. Because it cannot be expected that experimental structures will ever be available for all structures of interest, it is probable that the impact of protein structure prediction will even increase for such target classes in drug discovery. As new experimental structures take over the arenas that have been dominated by predicted models, protein modelling will utilise those as seeds for pioneering new areas of structural relevance. The example CYP polymorphism in context with personalised medicine described above supports this optimism. Increasing reliability of fully automated modelling pipelines allows models to be generated on a proteomic scale, as for the structural characterisation of the entire human kinome [91]. Another enthralling area is the prediction of protein–protein complexes on an omics scale [5,92,93]. It remains to be seen whether or not this enterprise will become conducive to drug discovery and when, for example in the design of protein–protein interaction inhibitors.

To summarize, the recent literature endorses the supposition that accurate homology modelling suitable for many drug discovery applications can be achieved (Box 1). We trust that the co-action of newly solved X-ray crystal structures, constantly growing number of sequences, improved modelling techniques and quality assessment methods will furnish powerful homology modelling techniques to advance the complex structural challenges further in drug discovery.

Box 1. Tips to remember and questions to address when applying homology modelling and hints for best practice.

Template selection	Carefully select the appropriate template(s). The most appropriate one is not necessarily the one with the highest sequence identity to the protein to be modelled. Depending on the functional context (apo/holo, agonist/antagonist), a template with lower sequence identity might be more suitable. Assess as to whether or not available template structures convey information about protein flexibility, and in particular plasticity of the binding site of interest. This can, for example, be achieved by superimposing known 3D structures with different ligands, orthologs or paralogs. Sequence identity of the templates should not be handled as an absolute criterion; nevertheless, can serve as a rough guideline for the expected ‘resolution’ of a model [11,102]. Consider the resolution and crystallographic refinement information of template structures. A high resolution structure will yield better models than a low resolution structure. If in doubt about structural details, inspect the electron density [103], for example using DeepView [102] or COOT [104]. Check whether the areas of particular interest for modelling are involved in crystal contacts. Use state-of-the-art structure model validation methods and metrics.
Sequence alignment	Carefully inspect target–template sequence alignments and assess as to whether alternative alignments yield models more suited to the particular question at hand.
Modelling	Automated modelling programs will normally produce better results than a nonexpert user can achieve with manual manipulation.
Refinement	Be wary of refining your model with molecular dynamics (MD), it can be boon or bane. Going beyond just refining cavity side-chains can be precarious.
Validation	Take maximal advantage of available experimental data, such as SAR and mutation data, for validating your model(s). The same holds true for independent computational methods.
Application	Do not think of structures (models and experimental) as true or false. Ponder on their suitability in the context of the questions you would like to answer. Based on the considerations above, surmise which answers your model can most probably answer and which not. Use your model(s) as a hypothesis (i.e. as a working tool) and always be prepared to falsify it.

Highlights.

• The majority of proteins encoded in a genome are accessible by structure modelling

• Modelling can provide accurate target predictions for structure-based drug design

• Models can capture crucial receptor flexibility in a protein family beyond apo structures

• Ligand steered modelling methods often outperform docking into static models

• Automated modelling techniques have matured to provide accurate models