Abstract

Antibodies are a class of proteins indispensable for the vertebrate immune system. The general architecture of all antibodies is very similar, but they contain a hypervariable region which allows millions of antibody variants to exist, each of which can bind to different molecules. This binding malleability means that antibodies are an increasingly important category of biopharmaceuticals and biomarkers. We present Antibody i-Patch, a method that annotates the most likely antibody residues to be in contact with the antigen. We show that our predictions correlate with energetic importance and thus we argue that they may be useful in guiding mutations in the artificial affinity maturation process. Using our predictions as constraints for a rigid-body docking algorithm, we are able to obtain high-quality results in minutes. Our annotation method and re-scoring system for docking achieve their predictive power by using antibody-specific statistics. Antibody i-Patch is available from http://www.stats.ox.ac.uk/research/proteins/resources.

Introduction

Antibodies (immunoglobulins) are a class of proteins indispensable in mediating the acquired immune responses in vertebrates (Fleischman et al., 1963). The rich repertoire of these molecules allows them to form complexes with virtually any antigen. Due to their malleable binding properties, there is increasing industry interest in immunoglobulins, which has already resulted in several antibody-based drugs (Murad et al., 2012). Design of such therapeutic antibodies is generally carried out through experimental approaches such as phage display (Murad et al., 2012). There is, however, a surprising paucity of computational techniques aimed at antibody design. Computational antibody design so far has focused on affinity maturation (e.g. Lippow et al., 2007), antibody homology modeling (e.g. Sivasubramanian et al., 2009), binding residue prediction (e.g. Kunik et al., 2012) and antibody–antigen docking (e.g. Sircar and Gray, 2010), with the two last categories being the subject of this manuscript.

The general position of the antibody-binding site is known a priori in the form of its six complementarity determining regions (CDRs). Collectively, the CDRs cover on average 80% of all the antibody-binding residues (Kunik et al., 2012). The remaining binding sites are usually found in the immediate vicinity of the CDRs.

A lot of research in the last 30 years has focused on developing formal frameworks describing the antibody-binding site through the medium of CDRs. The first attempt was that undertaken by Wu and Kabat, which provided the initial definition of the CDRs through analysis of multiple sequence alignments (MSAs) of antibodies (Wu and Kabat, 1972). In contrast to this sequence-based approach, Chothia et al. developed a structural CDR loop definition (Chothia and Lesk, 1987; Al-Lazikani et al., 1997). MacCallum et al. provided a contact CDR definition, informed by the analysis of antibody–antigen binding sites (MacCallum et al., 1996). Lefranc later built the International Immunogenetics Information System (IMGT)-CDR definition, which is based on the IMGT unique numbering (Lefranc et al., 2003; Lefranc, 2011). More recently, Kunik et al. have defined antibody binding regions, where they attempt to obtain sequence segments containing the majority of antigen-binding residues (Kunik et al., 2012). Even though all of these characterizations revolve around the concept of CDRs, they are not refinements of one another but rather complementary concepts, focusing on different aspects of the antibody-binding site.

To the best of our knowledge, the first attempt at characterizing CDRs based on the antigen contacting residues was that undertaken by MacCallum et al. They analyzed a set of 26 crystal structures of antibody–antigen complexes and classified different types of paratope topographies depending on the antigen type and size. Specifically, they noted that the residues contacting the antigen fall in the center of the antibody combining site.

MacCallum et al. also provided an early CDR contact prediction method, which used the mean burial of each residue combined with the curvature of the binding site, to assess the likelihood of it being a contact site. A more recent antibody contact prediction method, which specifically addresses the problem of contact residues in the framework region, is Paratome. Given a sequence or structure of an antibody, Paratome annotates the region where the binding site residues are, by comparing the input sequence or structure to experimentally determined antibody structures with annotated contact sites. Using the definition of binding site residues as all antibody residues within 4.5 Å of the bound antigen, they achieve 0.31 precision at 0.96 recall. In comparison, precision of prediction of paratopes using CDR definitions is ∼0.3 with recall reaching around 0.8 for IMGT and Chothia and close to 0.9 for Kabat. Given that minor mutations to the binding site might lead to significant changes to the specificity and affinity profile of an antibody, knowing fewer binding residues but with a higher precision might be beneficial for guiding mutations in antibody engineering.

In the first part of this paper we describe our method, Antibody i-Patch, which rather than providing an entire binding region like Paratome, or only the CDRs like the existing definitions of Kabat, Chothia, AbM and IMGT, instead annotates each residue with a binding likelihood score. By doing so, one can differentiate between higher and lower confidence predictions, which might provide a better guide for introducing mutations to the CDR region.

In the second part of this manuscript we consider the problem of local docking (where the antigen-binding site is vaguely known). It is often the case that the general position of the epitope is known a priori and only the precise profile of the residues involved in an interaction with the antibody are unknown (Smith and Sternberg, 2003; McKinney et al., 2007; Covaceuszach et al., 2008). Furthermore, antibodies are sometimes designed for a specific site on the antigen, and once again local docking would be appropriate (McKinney et al., 2007; Covaceuszach et al., 2008). Even though there has been progress in improving the docking of general proteins, as demonstrated by successive rounds of the CAPRI experiment (Mendez et al., 2005), there are currently only two methods which specifically address antibodies—SnugDock and ADARS (Sircar and Gray, 2010; Brenke et al., 2012). This may be due to the specific biology of the antibody-binding site, meaning that general docking tools are not ideal for this problem (Sircar and Gray, 2010; de Vries and Bonvin, 2011; Brenke et al., 2012; Li and Kihara, 2012). Other methods, like ZDOCK, attempt the problem by constraining the binding region to the CDRs (Chen et al., 2003).

Often docking is further hampered by the lack of solved crystal structures (in our case antibodies and antigens) to serve as input (Sircar and Gray, 2010). One solution to this problem is to create a model structure of the molecule in question. In the case of antibody–antigen complexes the problem is somewhat easier than the general protein case as antibodies have a very well conserved overall structure (Sivasubramanian et al., 2009) and the CDRs (with the notable exception of CDR-H3) often adopt similar structures (Chothia and Lesk, 1987; Choi and Deane, 2011). This facilitates modeling, as demonstrated by RosettaAntibody, WAM and PIGS (Whitelegg and Rees, 2000; Marcatili et al., 2008; Sivasubramanian et al., 2009). Nonetheless, as currently no protein model is perfect, docking methods often struggle when they are used as input (Sircar and Gray, 2010). This issue was tackled by SnugDock which produces acceptable quality complexes through flexible docking of homology models, thus mitigating modeling errors and induced fit issues (Sircar and Gray, 2010). Coupling SnugDock with EnsembleDock (Chaudhury and Gray, 2008), an ensemble method, achieves even better results as it does not have to rely on a single model.

Here, we demonstrate Antibody i-Patch, a method which predicts antibody contact residues. In contrast to Paratome and the CDR definition methods, which indicate the extent of the general binding region, Antibody i-Patch assigns a contact likelihood score to each residue, allowing the user to choose a cut-off so as to achieve higher precision or better coverage. Using an in silico alanine scanning protocol, we show that residues with a higher score are more important energetically. We also show the applicability of Antibody i-Patch by using the predicted contact residues as constraints for local antibody–antigen docking. Using the fast docking algorithm, ZDOCK (Chen et al., 2003) and using our constraints we can achieve results similar to those of SnugDock on its homology model test set, but in minutes rather than hundreds of hours per target.

Materials and methods

Dataset

The datasets central to the analysis in this manuscript are described in Table I.

Table I.

Datasets used in this manuscript

Dataset name Dataset description 
NR-full Non-redundant dataset of 148 antibody–protein complexes. This dataset was used for the Antibody i-Patch analysis 
NR-subset Dataset of 30 structures chosen at random from NR-full, used for testing our docking procedure. The remaining 118 structures in NR-full were used for training Antibody i-Patch and for training the re-scoring algorithm docking 
SnugDock-H The 15 homology model docking targets used for the SnugDock analysis (Sircar and Gray, 2010
Dataset name Dataset description 
NR-full Non-redundant dataset of 148 antibody–protein complexes. This dataset was used for the Antibody i-Patch analysis 
NR-subset Dataset of 30 structures chosen at random from NR-full, used for testing our docking procedure. The remaining 118 structures in NR-full were used for training Antibody i-Patch and for training the re-scoring algorithm docking 
SnugDock-H The 15 homology model docking targets used for the SnugDock analysis (Sircar and Gray, 2010

Full list of PDB codes used in this analysis can be found in the Supplementary Information.

NR-full is a non-redundant set of antibody–protein complexes extracted from our antibody database ABSDB available at http://www.stats.ox.ac.uk/research/proteins/resources. The database extracts the antibody structures from the Protein Data Bank (Berman et al., 2000) as described by Dunbar et al. (2013). Only antibodies with both VL and VH present and of resolutions 3 Å or better were selected. Structures are reduced to a non-redundant set using CDHIT (Li and Godzik, 2006). In NR-full, no AB sequence was >99% identical and no antigen >90% identical. NR-full contains 148 non-redundant antibody–antigen complex structures.

The NR-subset consists of 30 structures chosen at random from NR-full. The 30 structures in the NR-subset served as the test set for the docking procedure, while the remaining 118 structures in NR-full were used to train the Antibody i-Patch method and the decoy re-scoring method for application to NR-subset.

SnugDock-H dataset consists of the 15 docking targets as described in Sircar and Gray (2010). The structures in SnugDock-H were the top-scoring homology models obtained from RosettaAntibody (Sivasubramanian et al., 2009). Since the models obtained did not contain H3 coordinates, these were modeled using FREAD (Choi and Deane, 2010).

Comparisons with other methods (e.g. Paratome, Kunik et al., 2012) were carried out on their datasets and the results of these analyses are shown in the Supplementary Material.

Antibody i-Patch

Antibody i-Patch is derived from the original i-Patch (Hamer et al., 2010). The binding site prediction algorithm, i-Patch requires the structures of two proteins (or models) that are assumed to interact together with the MSA of their homologs. Two modifications are needed to generate antibody i-Patch. The first is the removal of the need for an MSA. As antibody–antigen complexes are formed under very different processes than standard protein–protein interactions, creating corresponding MSAs is impossible. In order to establish whether i-Patch can operate without an MSA, we tested i-Patch on its original protein–protein binding site prediction dataset (Hamer et al., 2010) with and without MSA. In order to evaluate the statistical significance of the difference between the results of the two types of input an area under receiver operating characteristics (ROC) curve method was employed (Chen et al., 2008). We found the results from this test are not statistically significantly different from those using full MSAs.

The second is an asymmetry in residue interaction potential. As shown in Fig. 1, antibodies use different residues in their binding sites from proteins in general. Thus, the i-Patch scoring system needs to become asymmetric between the antibody and antigen.

Fig. 1.

Binding propensity differences between antibody, antigen and general proteins. Binding propensities of each residue type were calculated for antibodies and antigens. Values above ‘1’ indicate a preference to be in contact while those below ‘1’ correspond to a preference not to be in contact (red line in the figure above; for full details of how propensity was calculated please see the Supplementary Information). These propensities are contrasted to those of other proteins reported by the authors of i-Patch (Hamer et al., 2010). Antibodies appear to have radically different binding preferences from antigens and general proteins.

Fig. 1.

Binding propensity differences between antibody, antigen and general proteins. Binding propensities of each residue type were calculated for antibodies and antigens. Values above ‘1’ indicate a preference to be in contact while those below ‘1’ correspond to a preference not to be in contact (red line in the figure above; for full details of how propensity was calculated please see the Supplementary Information). These propensities are contrasted to those of other proteins reported by the authors of i-Patch (Hamer et al., 2010). Antibodies appear to have radically different binding preferences from antigens and general proteins.

In the Supplementary Information, we give a brief outline of the i-Patch procedure (the full algorithm is described in (Hamer et al., 2010)) and show how it is converted to Antibody i-Patch.

Antibody i-Patch scores for the antibody were calculated for each CDR region plus two residues on either side according to the IMGT definition. This was done as the majority of the binding site of an antibody is known to be composed of the CDRs and a few residues on either side (for the details on the decision to take two residues please see Supplementary Information). The entire antigen surface was used.

Thus, given an input consisting of an antibody and an antigen structure, each residue within the IMGT CDR region augmented with two anchor residues on each side is assigned a non-negative Antibody i-Patch score. Scores are not normalized and they range from 0 to 300, with higher values corresponding to greater binding likelihood.

Evaluating the performance of antibody i-Patch

The evaluation procedure consisted of 2000 jackknife tests on NR-full.

One iteration of our evaluation procedure consisted of randomly splitting NR-full into training and test datasets, consisting of 118 and 30 structures, respectively. Antibody i-Patch was trained using the 118 structures in the training set and then applied to each of the 30 entries in the test set. Thus each antibody in the training set would have its residues in the IMGT CDR region together with two residues on each side annotated with an Antibody i-Patch score. The performance of the method was then assessed by predicting all residues with an Antibody i-Patch score above a cut-off as contacts. A subset of these will be the true positives (TP), and the rest false positives (FP). Those residues below the cut-off will be either false negatives (FN) or true negatives. Thus for each score cut-off we calculate its corresponding precision (TP/(TP + FP)) and recall (TP/(TP + FN)). All cut-offs between 0.1 and 300 at intervals of 0.1 were tested.

The evaluation procedure described above was repeated 2000 times on distinct partitions of NR-full into 118 training structures and 30 test structures. The precision and recall scores corresponding to a particular cut-off from each run were averaged over the 2000 iterations.

The predictive power of the static CDR loop and region definitions was carried out by annotating the antibody sequences with the corresponding Kabat, Chothia and Contact regions as defined using the Chothia numbering scheme. Antibody sequences were numbered using Abnum (Abhinandan and Martin, 2008). The IMGT definitions were copied directly from the IMGT website or annotated manually for those structures not yet in the database.

Computational alanine scanning

Following the protocol outlined by the authors of Paratome (Kunik et al., 2012) for alanine scanning (in silico), we have introduced alanine mutations to each binding site residue (of the IMGT CDRs together with two anchor residues on each side) which was in contact with the antigen in the native structure (closest heavy atoms within 4.5 Å from each other). FoldX (Schymkowitz et al., 2005) was used to calculate the interaction energy between the wild-type antibody and antigen and each alanine mutated antibody and antigen, allowing us to calculate the interaction energy change (ΔΔG) between the wild-type and antigen and the mutant and antigen. Following Paratome (Kunik et al., 2012), an interaction energy change ΔΔG < −0.25 kcal/mol was defined as stabilizing whereas if it was the case that ΔΔG > 0.25 kcal/mol, it was destabilizing.

In order to assess the energetic importance of the Antibody i-Patch score, we took all residues with an Antibody i-Patch score greater than a cut-off and calculated the percentage of these residues that are destabilizing (ΔΔG > 0.25 kcal/mol by computational alanine scanning). As we have 2000 independent runs, we calculate the average percentage of destabilizing positions above a given score. Thus, if the Antibody i-Patch score is related to energetic importance, a larger percentage of residues will be destabilizing at higher scores.

Docking methods

NR-subset and SnugDock-H were used for this analysis and two docking methods were tested: ZDOCK (Chen et al., 2003) and PatchDock (Duhovny et al., 2002; Schneidman-Duhovny et al., 2005). ZDOCK was used in its default mode to produce 2000 decoys. PatchDock was also run in its default mode, producing an indeterminate number of decoys. The antibody was treated as the receptor in both pieces of software. The top 200 scoring decoys for each target, according to each method were collected for further analysis. ZDOCK and PatchDock both allow for docking constraints to be specified in the input. We used this facility to enter constraints based on the Antibody i-Patch scores. In the manuscript, the results are shown for ZDOCK (PatchDock results which are similar are given in the Supplementary Information).

Antigen constraint

An extended epitope was provided as input to the docking algorithms, which consisted of the actual binding residues, together with all other residues within 4, 5 or 6 Å. Using any of these constraints, included far more residues than there are in the actual antigen epitope (see Supplementary Information). The use of an augmented epitope aims to simulate the case when the epitope position on the antigen is already known, but the precise residues involved in the interaction are not. For comparative purposes, the constraint approximates the initial position of the antibody and antigen to a similar extent to that employed in SnugDock (Sircar and Gray, 2010) where the antibody is pointed towards the antigen and then rotated to add randomness. In the manuscript, results are shown for the actual epitope plus all residues within 5 Å (the results for other sets are shown in Supplementary Information). The set of residues given as the antigen constraint is referred to as Cag.

Antibody constraint

The binding constraint for the antibody is the set of residues that are assumed to be involved in binding. There were three versions of the antibody constraint given to the docking algorithms: the IMGT CDRs (referred to as C-CDR), the residues that were predicted by Antibody i-Patch to be binding sites (referred to as C-Antibody i-Patch) and the actual residues constituting the paratope (referred to as C-Native). The residues for C-Antibody i-Patch were generated by applying Antibody i-Patch to antibody–antigen pairs in datasets NR-subset or SnugDock-H and taking everything above the cut-off of 40.0 so as to achieve good balance between precision and recall. The value of 40 was motivated by the average results of Antibody i-Patch on the training set portions of the Antibody i-Patch runs. The set of residues provided as the antibody constraint is referred to as Cab.

The precision score

Let Pr(Tab, Tag) denote the precision of a contacting pair of residues with types Tab on the antibody and Tag on the antigen to be correct if observed in a decoy from ZDOCK. The precision Pr(Tab, Tag) was estimated by executing ZDOCK on each antibody–antigen complex in dataset NR-full, which was not in NR-subset. As the constraint for the local docking we gave the paratope and epitope residues on each molecule, together with those within 5 Å away from them. Decision to take this cut-off was an arbitrary choice acting as a middle ground with respect to the epitope cut-offs we use in this manuscript.

For each of the 118 targets in NR-full that were not in NR-subset, we collected the top 200 decoys as ranked by ZDOCK. Over all, the decoys collected in this manner we counted the number of TPs and FPs for amino acid types, Tab and Tag, being observed within 4.5 Å. We denote the number of true positives and false positives collected in this manner (TP (Tab, Tag)) and (FP (Tab, Tag)), respectively. This leads to the estimate of ZDOCK pairing up Tab and Tag given below: 

formula

For example, if we consider Tab as cysteine residues on the antibody and Tag as tyrosine residues on the antigen, we count in the 23 600 decoys (118 targets times 200 decoys) the number of times cysteine residues on the antibody are in contact with tyrosines on the antigen and then partition these into true positives (TP(Tab, Tag)) and false positives (FP(Tab, Tag)) in order to calculate the precision (Pr(Tab, Tag)). An analogous procedure was carried out for PatchDock.

We have also applied this procedure for use on the SnugDock-H dataset. Here, we used CDHIT to remove those complexes from NR-full that had >90% antigen sequence identity and 99% antibody sequence identity between NR-full and SnugDock. The resulting set of 138 antibody–antigen complexes was used in an analogous fashion to create the precision score for re-scoring the decoys in SnugDock-H.

Reordering decoys

For the top 200 decoys generated by ZDOCK for a given target, they are rescored using the following procedure. For each decoy, all pairs of residues within 4.5 Å where one is on the antibody (rab) and the other on the antigen (rag) are selected. Those interacting residue pairs (rab, rag) that belong to the initial constraints (Cab, Cag) are then used to re-score the decoy. This is achieved by summing the precision values given in Equation (1) for every interacting residue pair (rab, rag) in the initial constraints set (Cab, Cag). The 200 decoys are then reordered using this score.

Evaluating docking performance

Decoy quality was evaluated according to the CAPRI criteria (Mendez et al., 2005) (see Supplementary Information for more details). Since it was computationally feasible to perform multiple runs of ZDOCK on the target set, we have sampled the results for each combination of inputs over 20 runs. The results for each run were in the form of vectors with the CAPRI ratings e.g (5***, 1**, 3*)—five very good results, one good result and three satisfying results. A result for each input combination consists, therefore, of the average of the result vectors over 20 runs. Since the multiple runs do not affect the results of PatchDock, only a single run was performed for each of the targets.

Results

The work presented in this manuscript is divided into two parts: antibody contact prediction and antibody–antigen docking. First, we discuss a method to predict binding residues on the antibody. Secondly, we show how our antibody-binding site predictions can be used to constrain and therefore improve local antibody–antigen docking algorithms.

Antibody i-Patch

Antibody i-Patch is an adaptation of the protein–protein contact predictor i-Patch (Hamer et al., 2010). The original i-Patch algorithm receives on input the structures of two proteins, which are assumed to interact, together with the MSA of interacting homologs (Hamer et al., 2010). The algorithm uses the contact propensity data contained in its training dataset of domain–domain interactions and protein complexes and the correlated mutations in the input MSA to generate a statistical score for each surface residue of the input proteins. The score for each residue is calculated using a patch of residues on the proteins surface centered on the residue of interest. The standard i-Patch score is relatively accurate at annotating protein contact residues (0.59 precision at 0.2 recall).

The i-Patch methodology is not immediately applicable to the antibody–antigen complexes for two reasons. First, because it is impossible to create the necessary MSAs, as antibody–antigen complexes do not undergo correlated mutations, but rather the antibody becomes fine tuned to bind a particular antigen, there is no evolutionary information (MSA) available for either the antibody or antigen. Adaptation of the algorithm to the special case of antibodies involves removing the MSA from the input.

Secondly, standard i-Patch uses contact propensities from protein–protein binding sites, which may not be applicable to antibody–antigen complexes. We therefore compared the binding propensity of amino acids (Hamer et al., 2010) in antibodies to those in general proteins—Fig. 1. In accord with previous analysis (Raghunathan et al., 2012), we find that antibodies appear to have a very different binding profile from general proteins. The antigen-binding profile on the other hand is not profoundly different from that of general proteins.

Figure 1 shows the propensities of the different residue types to be involved in binding in antibodies, in antigens and in proteins in general. A propensity value >‘1’ indicates that if this residue is seen on the surface of, say antibodies, it is likely to be involved in binding to an antigen and conversely for scores <‘1’ (for full details see Supplementary Information). We observed a very strong preference for tyrosine and tryptophan to be in the binding sites of antibodies, not seen for proteins in general. These high contact propensities of tyrosine and tryptophan in antibodies are in agreement with earlier theoretical (Mian et al., 1991) and experimental (Fellouse et al., 2005; Birtalan et al., 2008) findings. Moreover, the contribution of tryptophan can be explained by a conserved residue of this type (Collis et al., 2003). The presence of tryptophan in CDRs is relatively high as compared with other protein loops, where it has very low abundance (Collis et al., 2003). Serine, another residue known to be prevalent in the antibody-binding site (Golub et al., 1997; Zemlin et al., 2003), shows a different behavior. Its binding propensity profile in Fig. 1 does not differ significantly from that of general proteins.

Given that antibodies appear to have binding preferences distinct from those of general proteins, we recalculated the contact scores used in i-Patch so as to accommodate those differences. The propensity information used in Antibody i-Patch was calculated using a jackknife procedure on the NR-full dataset while keeping the original protein propensities for the antigen. Details of the procedure can be found in the Materials and methods section.

Using the fact that the general position of the antibody-binding site is known, we constrained the region which we consider for the analysis to the IMGT CDR definition augmented by two further residues on each side. This augmentation of the IMGT definition was motivated by an analysis of which framework regions were most likely to be in contact with the antigen (see the Supplementary Information).

Antibody i-Patch predicts antigen binding residues

The performance of Antibody i-Patch was evaluated using a P-ROC graph (Shiu and Gatsonis, 2008). Here the precision (or accuracy) is plotted against the recall (or coverage). A perfect predictor would give precision of 1.0 for all values of recall. In our case, if we impose a high score cut-off, we achieve precision values up to 0.77 but at a recall of only 0.1 (Fig. 2). At a low score cut-off Antibody i-Patch precision falls to 0.42 with a coverage of 0.93. The results presented here are the averages of P-ROC plots of 2000 jackknife tests on NR-full set.

Fig. 2.

P-ROC plot of Antibody i-Patch results averaged over 2000 runs. Comparison of the performance of Antibody i-Patch with static antibody-binding site annotation methods for the contact distance of 4.5 Å. Standard error is shown for the values of precision, (recall errors can be found in the Supplementary Information). In contrast to the static antibody-binding site annotation methods of Kabat, Chothia, Contact and the International Immunogenetics Information System (IMGT), Antibody i-Patch produces results for a wide spectrum of precision and recall values. As all the residues outside the window of IMGT definition augmented with two framework residues on either side are considered to be non-binding, the recall starts at 0.93. If the original IMGT definition had been used the same graph would be truncated to the point corresponding to that of IMGT, i.e. recall of 0.83.

Fig. 2.

P-ROC plot of Antibody i-Patch results averaged over 2000 runs. Comparison of the performance of Antibody i-Patch with static antibody-binding site annotation methods for the contact distance of 4.5 Å. Standard error is shown for the values of precision, (recall errors can be found in the Supplementary Information). In contrast to the static antibody-binding site annotation methods of Kabat, Chothia, Contact and the International Immunogenetics Information System (IMGT), Antibody i-Patch produces results for a wide spectrum of precision and recall values. As all the residues outside the window of IMGT definition augmented with two framework residues on either side are considered to be non-binding, the recall starts at 0.93. If the original IMGT definition had been used the same graph would be truncated to the point corresponding to that of IMGT, i.e. recall of 0.83.

The performance of Antibody i-Patch on homology models is very similar to that on crystal structures (see Supplementary Information). We also tested the use of an entirely sequence-based Antibody i- Patch where the patch was defined by the sequence neighbors rather than structural neighbors, which considerably weakened the performance of the algorithm, thus the construction of homology models is an essential step in any prediction pipeline.

The fact that Antibody i-Patch performs as well on homology models as on crystal structures probably arises from the fact that it requires structural information only for the purpose of determining patches on the surface not exact neighbors. Therefore, even fairly low-resolution homology models could provide an acceptable approximation of the native structure. The above results indicate that Antibody i-Patch needs only the sequence of an antibody as input.

Comparing antibody i-Patch to static CDR annotation methods

The average precision and recall scores over the 2000 jackknife tests on dataset NR-full are presented in Fig. 2. The results indicate that Antibody i-Patch achieves both higher precision and recall as compared with the static CDR definitions. The recall of the different methods is limited by the residues included in the predictions of each of the CDRs. This can be seen in a comparison between Antibody i-Patch and IMGT. Antibody i-Patch reaches a highest recall of 0.93, whereas IMGTs recall is only around 0.83. This increase is entirely due to the inclusion of the two extra framework residues on either side of each of the CDRs in the Antibody i-Patch predictions.

Out of the static CDR region and loop annotation methods, Kabat regions achieve highest recall with a value of 0.9. The Contact definition, which was designed based on antibody–antigen contacts underperforms in this experiment, achieving the lowest recall of all—0.77. This is probably due to the small size of its original training set of 26 antibody–antigen complexes (not all of which were protein antigens). All the static methods achieve precision in the region of 0.3.

We have also tested the Antibody i-Patch algorithm on the Paratome test set (results in the Supplementary Information). On this dataset Antibody i-Patch achieves highest recall of 0.94 at 0.4 precision compared with 0.96 recall and 0.31 precision by Paratome. This indicates that i-Patch does not achieve recall quite as high as Paratome but is more precise.

Manual inspection of the Antibody i-Patch annotation scores has not revealed any definite topological patterns. The method does not annotate a circular patch within the binding interface as discovered by MacCallum et al. (1996). The simplified model for antibody contact residue prediction proposed by MacCallum et al. in the form of the mean burial of each residue combined with the curvature of the binding site might also prove to be a useful antibody design tool. Nevertheless, the Contact CDR definition appears to underperform with respect to other methods due to the fact that the 26 complexes used for the study by MacCallum et al., not all of which are protein complexes, was a very small dataset at a time. The CDR prediction method proposed by MacCallum et al., as well as the CDR Contact definition could be re-evaluated using the more abundant data available today.

Unlike the other methods including Paratome, Antibody i-Patch does not just give a yes–no result for each residue, instead each residue is given a score related to its likelihood to be in the binding site. Thus, if an user was not interested in identifying the entire binding region (which all methods achieve with relatively low precision) but instead wished to know with great certainty a small number of residues involved in the binding site (low recall, high precision), they could pick a high Antibody i-Patch score cut-off, say 160, which corresponds to recall of 0.4 and a precision of 0.63 (Fig. 2).

Residues with higher antibody i-Patch scores are more energetically Significant

We have investigated the energetic significance of the results returned by Antibody i-Patch. As a background distribution, we use the residues from across the entire IMGT-CDR regions augmented by two residues on either side of each CDR. We contrast the effect of mutating residues in this background set with mutating residues which achieve high Antibody i-Patch scores.

We follow the procedure in Kunik et al. (2012) and excluded the non-binding residues from this analysis because our aim was to explore the link between the Antibody i-Patch score—which as demonstrated correlates with higher likelihood to be a contact site—and energetic importance. Even though there are non-binding residues that are crucial for the stability of the antibody–antigen complex, our aim was to establish if the higher-likelihood binding residues are energetically important, thus inclusion of non-contacts would be meaningless and could potentially skew the energetic difference distribution.

Using FoldX (Schymkowitz et al., 2005), we have performed in silico alanine scanning of the antibody contact residues from the region of IMGT-CDRs augmented by two framework residues on either side for each structure in NR-full. In each of the 2000 runs of Antibody i-Patch on NR-full, we have noted the fraction of destabilizing residues above each score cut-off. We have averaged the proportions over those 2000 runs for each score cut-off. As in the previous analysis (Kunik et al., 2012), the majority of the mutations lie in the neutral region of −0.25 ≤ ΔΔG ≤ 0.25 kcal/mol. For increasing values of the Antibody i-Patch score, the fraction of the contact residues with destabilizing mutations (ΔΔG ≥ 0.25 kcal/mol) rises steadily, as shown in Fig. 3.

Fig. 3.

The energetic importance of antibody–antigen contact residues is correlated with the Antibody i-Patch score. The number of contact residues with an Antibody i-Patch score greater than a cut-off which lead to a ΔΔG > 0.25 kcal/mol when mutated to alanine compared with the number of contact residues in general that lead to a ΔΔG > 0.25 kcal/mol when mutated to alanine. As the Antibody i-Patch score cut-off is increased, the ratio of residues which cause an energetic change upon alanine mutation increases. In other words, residues with a high Antibody i-Patch score tend to be energetically more important.

Fig. 3.

The energetic importance of antibody–antigen contact residues is correlated with the Antibody i-Patch score. The number of contact residues with an Antibody i-Patch score greater than a cut-off which lead to a ΔΔG > 0.25 kcal/mol when mutated to alanine compared with the number of contact residues in general that lead to a ΔΔG > 0.25 kcal/mol when mutated to alanine. As the Antibody i-Patch score cut-off is increased, the ratio of residues which cause an energetic change upon alanine mutation increases. In other words, residues with a high Antibody i-Patch score tend to be energetically more important.

In silico alanine scanning is not a substitute for the actual experimental data, but it acts as an indicator of the general energetic behavior of the mutations. Based on this FoldX analysis, we suggest that residues receiving higher Antibody i-Patch scores not only have a higher likelihood of being in contact but also tend to be more energetically important.

Antibody–antigen docking

Prediction of antibody-binding residues has a potential application to antibody docking. As demonstrated by CPORT and PI-LZerD, prediction of contact residues can reduce the global docking problem to a local one, by greatly constraining the search space (de Vries and Bonvin, 2011; Li and Kihara, 2012).

In many cases, docking of antibody–antigen complexes can be reduced to a local docking problem. The binding site of the antibody is known a priori, in the form of the CDRs and its surrounding regions. The epitope location on the antigen is also often known either through experimental data or a desire to target an antibody to a particular location on an antigen (Smith and Sternberg, 2003; McKinney et al., 2007; Covaceuszach et al., 2008). Here, we investigate the extent to which our accurate identification of the antibody-binding residues facilitates the task of local docking.

We chose two fast rigid-body docking algorithms for our analysis: ZDOCK and PatchDock. The aim here is to generate a very rapid, relatively accurate local docking pipeline for antibody–antigen complexes. We have tested the performance of our method on the NR-subset (Fig. 4) and SnugDock-H—the 15 homology model targets used in the SnugDock analysis (Fig. 5). For each target in the datasets NR-subset and SnugDock-H, we used two docking constraints, one for the antibody and one for the antigen. Each constraint consisted of a set of residues, which describe the binding site to a greater or lesser degree.

Fig. 4.

Results from running our docking pipeline on the test dataset of crystal structures docking targets NR-subset. The results are the averages of the three element vectors given by the CAPRI classification into three quality groups: satisfying (*), medium (**) or good (***) quality. The individual three element vectors were collected over 20 runs of the pipeline on the dataset NR-subset. The standard deviations are given in the Supplementary Information.

Fig. 4.

Results from running our docking pipeline on the test dataset of crystal structures docking targets NR-subset. The results are the averages of the three element vectors given by the CAPRI classification into three quality groups: satisfying (*), medium (**) or good (***) quality. The individual three element vectors were collected over 20 runs of the pipeline on the dataset NR-subset. The standard deviations are given in the Supplementary Information.

Fig. 5.

Results from running our docking pipeline on the test dataset of homology models docking targets SnugDock-H. The results are the averages of the three element vectors given by the CAPRI classification into three quality groups: satisfying (*), medium (**) or good (***) quality. The individual three element vectors were collected over 20 runs of the pipeline on the dataset SnugDock-H. The standard deviations are given in the Supplementary Information.

Fig. 5.

Results from running our docking pipeline on the test dataset of homology models docking targets SnugDock-H. The results are the averages of the three element vectors given by the CAPRI classification into three quality groups: satisfying (*), medium (**) or good (***) quality. The individual three element vectors were collected over 20 runs of the pipeline on the dataset SnugDock-H. The standard deviations are given in the Supplementary Information.

The constraint for the antigen consisted of the actual binding residues together with residues on the antigen within 5 Å. In the case of the antibody, we tested three different constraints: Native, CDR and Antibody i-Patch. The first, Native, consisted of the correct binding site. The second, CDR consisted of the IMGT-CDR residues. The last set, Antibody i-Patch consisted of the binding residues predicted by Antibody i-Patch with a score cut-off >‘40’. This Antibody i-Patch score cut-off was chosen as it gave an acceptable trade off between the values of recall and precision. On the dataset NR-full it gave on average 0.45 precision and 0.9 recall (see Fig. 2 for the P-ROC plot and Supplementary Information for an example constrained input). In the case of the 15 targets in the docking test set (SnugDock-H), this cut-off of ‘40’ gives precision of 0.47 and recall of 0.88.

Figure 4 shows the results of running ZDOCK with different antibody constraints on the crystal structure test dataset NR-subset. The docking results are evaluated using the CAPRI criteria (Mendez et al., 2005). Each decoy is placed in one of the four categories from 0 stars up to 3 stars, with 3 stars denoting higher quality results and the best result among the top 10 decoys is reported for each target. As ZDOCK is non deterministic, but very rapid, we decided to check whether our results were consistent across multiple runs. For each test case in the set we sampled the results from 20 runs, thus the results shown in Figs 4 and 5 and those in the Supplementary Information are the averages of these runs.

The Antibody i-Patch predictions and our antibody–antigen specific scoring system compared with the general protein score of the docking software enriches the top 10 predictions with useful decoys (Fig. 4). Specifically, using our docking pipeline we achieve better results than using native contacts as constraints for the docking algorithm. We argue that this is due to the fact that the docking algorithms like ZDOCK and PatchDock are trained on general proteins datasets, which have distinctly different binding profiles from those of antibodies (Fig. 1). These results demonstrate that our antibody-specific docking pipeline consistently selects better decoys, approaching the maximum possible score given in the top 200 results of ZDOCK. Standard deviations are given in the Supplementary Information.

On the 30 crystal structures in the NR-subset even given native contacts, ZDOCK only achieves a CAPRI vector of (6.5***, 10.2**, 5.2*), whereas using the Antibody i-Patch constraints and scoring system a vector of (8***, 14.8**, 5.1*) is given (Fig. 4).

Thus using the antibody-specific scoring improves our ability to identify correct docking poses. In fact, most of the good quality docks in the top 200 results are sorted into the top 10 by the antibody-specific scoring system (Fig. 4).

Figure 5 shows the docking results on antibody homology models (dataset SnugDock-H). Once again reasonable results are found in the top 200 for any of the constraints. However, it is only in the antibody-specific procedure that these decoys are placed in the top 10.

Comparing the rigid-body docking pipeline to other methods

The results of our pipeline are not as good as those obtained by more complex docking procedures such as SnugDock and EnsembleDock, but are similar to those of standard RosettaDock. For example, as shown in Fig. 5, the top 10 results for the Antibody i-Patch procedure on the 15 homology model test cases lead to a CAPRI vector of (0.7***,3.1**, 7.0*), while Ensemble Dock with SnugDock on the same set has a CAPRI vector of (0***, 9**, 5*) and Standard RosettaDock achieves (0***, 5**, 6*) (Sircar and Gray, 2010). In general the flexibility in the more complex methods allows them to generate higher quality results (e.g. move decoys from being of * quality up to **) something which cannot be achieved in a rigid docking scenario.

Another comparator between the methods, however, is the time taken to achieve these results. The Antibody i-Patch ZDOCK pipeline can produce results for our targets in minutes rather than the hundreds of CPU hours needed for the flexible methods (Sircar and Gray, 2010; Pierce et al., 2011). Therefore, our pipeline could provide good initial poses at little computational cost for ensemble or flexible docking methods like SnugDock for further refinement. Thus, the method could be used where speed is critical to tackling the problem, for example, all against all antibody/antigen screening.

Supplementary data

Supplementary data are available at PEDS online.

Funding

This work was supported by the Engineering and Physical Sciences Research Council, and UCB Pharma.

Acknowledgements

We thank James Dunbar for help in creating the non-redundant dataset as well as valuable comments and discussions.

References

Abhinandan
K.R.
Martin
A.C.R.
Mol. Immunol.
 , 
2008
, vol. 
14
 (pg. 
3832
-
3839
)
Al-Lazikani
B.
Lesk
A.M.
Chothia
C.
J. Mol. Biol.
 , 
1997
, vol. 
4
 (pg. 
927
-
948
)
Birtalan
S.
Zhang
Y.
Fellouse
F.
Shao
L.
Schaefer
G.
Sidhu
S.
J. Mol. Biol.
 , 
2008
, vol. 
377
 
5
(pg. 
1518
-
1528
)
Berman
H.M.
Westbrook
J.
Feng
Z.
Gilliland
G.
Bhat
T.N.
Weissig
H.
Shindyalov
I.N.
Bourne
P.E.
Nucleic Acids Res.
 , 
2000
, vol. 
28
 
1
(pg. 
235
-
242
)
Brenke
R.
Hall
D.
Chuang
G.
Comeau
S.
Bohnuud
T.
Beglov
D.
Schueler-Furman
O.
Vajda
S.
Kozakov
D.
Bioinformatics
 , 
2012
, vol. 
28
 
20
(pg. 
2608
-
2614
)
Chaudhury
S.
Gray
J.J.
J. Mol. Biol.
 , 
2008
, vol. 
381
 pg. 
10681087
 
Chen
P.Y.
Deane
C.M.
Reinert
G.
PLoS Comput. Biol.
 , 
2008
, vol. 
4
 pg. 
e1000118
 
Chen
R.
Li
L.
Weng
Z.
Proteins
 , 
2003
, vol. 
1
 (pg. 
80
-
87
)
Choi
Y.
Deane
C.
Mol. BioSyst.
 , 
2011
, vol. 
7
 
12
(pg. 
3327
-
3334
)
Choi
Y.
Deane
C.M.
Proteins
 , 
2010
, vol. 
78
 (pg. 
1431
-
1440
)
Chothia
C.
Lesk
A.M.
J. Mol. Biol.
 , 
1987
, vol. 
4
 (pg. 
901
-
917
)
Collis
A.V.J.
Brouwer
A.P.
Martin
A.C.R.
J. Mol. Biol.
 , 
2003
, vol. 
2
 (pg. 
337
-
354
)
Covaceuszach
S.
Cassetta
A.
Konarev
P.V.
Gonfloni
S.
Rudolph
R.
Svergun
D.I.
Lamba
D.
Cattaneo
A.
J. Mol. Biol.
 , 
2008
, vol. 
4
 (pg. 
881
-
896
)
de Vries
S.J.
Bonvin
A.M.J.J.
PLoS ONE
 , 
2011
, vol. 
3
 pg. 
e17695
 
Duhovny
D.
Nussinov
R.
Wolfson
H.J.
Lecture Notes in Computer Science
 , 
2002
, vol. 
2452
 (pg. 
185
-
200
Springer Verlag, WABI
Dunbar
J.
Fuchs
A.
Shi
J.
Deane
C.M.
2013
 
ABangle: Characterising the VH-VL orientation in antibodies, PEDS Advance Access Article
Fellouse
F.
Li
B.
Compaan
D.
Peden
A.
Hymowitz
S.
Sidhu
S.
J. Mol. Biol.
 , 
2005
, vol. 
348
 
5
(pg. 
1153
-
1162
)
Fleischman
J.B.
Porter
R.R.
Press
E.M.
Biochem. J.
 , 
1963
, vol. 
2
 (pg. 
220
-
228
)
Golub
R.
Fellah
J.
Charlemagne
J.
Immunogenetics
 , 
1997
, vol. 
46
 
5
(pg. 
402
-
409
)
Hamer
R.
Luo
Q.
Armitage
J.P.
Reinert
G.
Deane
C.M.
Proteins
 , 
2010
, vol. 
78
 (pg. 
2781
-
2797
)
Kunik
V.
Peters
B.
Ofran
Y.
PLoS Comput. Biol.
 , 
2012
, vol. 
8
 pg. 
e1002388
 
Lefranc
M.P.
Cold Spring Harb. Protoc.
 , 
2011
, vol. 
6
 (pg. 
633
-
642
)
Lefranc
M.P.
Pommié
C.
Ruiz
M.
Giudicelli
V.
Foulquier
E.
Truong
L.
Thouvenin-Content
V.
Lefranc
G.
Dev. Comp. Immunol.
 , 
2003
, vol. 
27
 
1
(pg. 
55
-
77
)
Li
B.
Kihara
D.
BMC Bioinformatics
 , 
2012
, vol. 
13
 pg. 
7
 
Li
W.
Godzik
A.
Bioinformatics
 , 
2006
, vol. 
22
 (pg. 
1658
-
1659
)
Lippow
S.M.
Wittrup
K.D.
Tidor
B.
Nat. Biotechnol.
 , 
2007
, vol. 
10
 (pg. 
1171
-
1176
)
MacCallum
R.M.
Martin
A.C.R.
Thornton
J.M.
JMB
 , 
1996
, vol. 
262
 (pg. 
732
-
745
)
Marcatili
P.
Rosi
A.
Tramontano
A.
Bioinformatics
 , 
2008
, vol. 
24
 
17
(pg. 
1953
-
1954
)
McKinney
B.A.
Kallewaard
N.L.
Crowe
J.E.
Jr
Meiler
J.
Immunome Res.
 , 
2007
, vol. 
3
 pg. 
8
 
Mendez
R.
Leplae
R.
Lensink
M.F.
Wodak
S.J.
BMC Bioinformatics
 , 
2005
, vol. 
60
 (pg. 
150
-
169
)
Mian
I.
Bradwell
A.
Olson
A.
J. Mol. Biol.
 , 
1991
, vol. 
217
 
1
(pg. 
133
-
151
)
Murad
J.
Lin
O.
Paez
E.
Khasawneh
F.
Curr. Mol. Med.
 , 
2012
, vol. 
2
 (pg. 
165
-
178
)
Pierce
B.
Hourai
Y.
Weng
Z.
PloS ONE
 , 
2011
, vol. 
6
 
9
pg. 
e24657
 
Raghunathan
G.
Smart
J.
Williams
J.
Almagro
J.
J. Mol. Recognit.
 , 
2012
, vol. 
25
 
3
(pg. 
103
-
113
)
Schneidman-Duhovny
D.
Inbar
Y.
Nussinov
R.
Wolfson
H.J.
Nucleic Acids Res.
 , 
2005
, vol. 
33
 (pg. 
W363
-
W367
)
Schymkowitz
J.
Borg
J.
Stricher
F.
Nys
R.
Rousseau
F.
Serrano
L.
Nucleic Acids Res.
 , 
2005
, vol. 
33
 
Suppl. 2
(pg. 
W382
-
W388
)
Shiu
S.Y.
Gatsonis
C.
Philos Trans. A, Math. Phys. Eng. Sci.
 , 
2008
, vol. 
366
 (pg. 
2313
-
2333
)
Sircar
A.
Gray
J.J.
PLoS Comput. Biol.
 , 
2010
, vol. 
6
 pg. 
e1000644
 
Sivasubramanian
A.
Sircar
A.
Chaudhury
S.
Gray
J.J.
Proteins
 , 
2009
, vol. 
74
 (pg. 
497
-
514
)
Smith
G.R.
Sternberg
M.J.
Proteins
 , 
2003
, vol. 
52
 pg. 
7479
 
Whitelegg
N.R.
Rees
A.R.
Protein Eng.
 , 
2000
, vol. 
12
 (pg. 
819
-
824
)
Wu
T.T.
Kabat
E.A.
J. Exp. Med.
 , 
1972
, vol. 
2
 (pg. 
211
-
250
)
Zemlin
M.
Klinger
M.
Link
J.
Zemlin
C.
Bauer
K.
Engler
J.
Schroeder
H.
Kirkham
P.
J. Mol. Biol.
 , 
2003
, vol. 
334
 
4
(pg. 
733
-
749
)

Author notes

Edited by Anthony Rees

Supplementary data