|Home | About | Journals | Submit | Contact Us | Français|
Protein–protein interactions play a key part in most biological processes and understanding their mechanism is a fundamental problem leading to numerous practical applications. The prediction of protein binding sites in particular is of paramount importance since proteins now represent a major class of therapeutic targets. Amongst others methods, docking simulations between two proteins known to interact can be a useful tool for the prediction of likely binding patches on a protein surface. From the analysis of the protein interfaces generated by a massive cross‐docking experiment using the 168 proteins of the Docking Benchmark 2.0, where all possible protein pairs, and not only experimental ones, have been docked together, we show that it is also possible to predict a protein's binding residues without having any prior knowledge regarding its potential interaction partners. Evaluating the performance of cross‐docking predictions using the area under the specificity‐sensitivity ROC curve (AUC) leads to an AUC value of 0.77 for the complete benchmark (compared to the 0.5 AUC value obtained for random predictions). Furthermore, a new clustering analysis performed on the binding patches that are scattered on the protein surface show that their distribution and growth will depend on the protein's functional group. Finally, in several cases, the binding‐site predictions resulting from the cross‐docking simulations will lead to the identification of an alternate interface, which corresponds to the interaction with a biomolecular partner that is not included in the original benchmark. Proteins 2016; 84:1408–1421. © 2016 The Authors Proteins: Structure, Function, and Bioinformatics Published by Wiley Periodicals, Inc.
Because of their essential role in performing, coordinating and regulating the majority of cell activities, proteins are undeniably amongst the most fascinating and complex macromolecules in living systems. Over recent decades, numerous studies have been devoted to the molecular properties and functions of individual proteins. However, proteins often fulfil their roles through interactions, and are capable of forming large edifices that can act as complex molecular machines.1 Transient interactions also form a complex protein network that controls these machines as well as a host of other cellular processes. As a consequence, protein–protein interactions (PPI) play a central role in biological systems,2, 3, 4 defining the interactome of an organism, or, as elegantly expressed by Robinson et al. the molecular sociology of the cell.5
Many experimental approaches are used to investigate PPI, including yeast two‐hybrid,6, 7 tandem affinity purification8, 9 and mass spectroscopy10 (see ref. 11 for a general review). These approaches have enabled PPI maps to be established for various organisms, including yeast,12 Escherichia coli,13 Drosophila melanogaster, 14 Caenorhabditis elegans 15 and humans.16 Experimental mapping of interactomes however suffers from several drawbacks. It involves expensive experiments and, despite continued progress, it still suffers from inaccuracies and generates significant numbers of false positives and negatives.17, 18, 19 As a complementary approach to in vitro methods, several in silico methods have also been developed for predicting binary protein interactions. Many of them are based on protein sequence information, using gene clustering or phylogenetic profiling.20, 21, 22, 23, 24, 25, 26 However, although these methods can predict interactions they do not provide any atomic‐level information on the conformation of the complex or on the origins of its formation and stability. Another approach has been to develop so‐called template methods, which predict interactions between pairs of proteins that are homologous (either globally, or at the interface region) to pairs of proteins within known binary complexes. These methods have achieved good results, but are naturally limited by the quality and coverage of the available template database.27, 28, 29 Still, important aspects of PPI, such as the influence of the crowded cellular environment,30, 31, 32 or the time‐dependence of these networks,33, 34, 35 cannot be addressed via any of these approaches.
Molecular modeling potentially offers an alternative route for identifying protein interactions, while at the same time providing structural models for the corresponding complexes and insight into the physical principles behind complex formation. In particular, the characterization on the molecular level of protein interfaces represents a key issue from a therapeutic point of view, since these PPI sites are potential targets for drugs designed to modulate or mimic their effects.36 As a consequence, numerous interface prediction methods have been developed over the last years, which combine evolutionary and structural, and sometimes experimental, information.37, 38, 39, 40 In this perspective, docking methods, which were originally developed to predict the structure of a complex starting from the structures of two proteins that are known to interact,41 are of specific interest. The collection of docking poses between two protein partners can be used to derive a consensus of predicted interface residues. Following the NIP (Normalized Interface Propensity) approach that was originally developed by Fernandez‐Recio et al.,42 docking calculations have been used for binding sites predictions both in simple docking studies,43, 44, 45, 46, 47 i. e. studies involving only protein partners that are already known to interact, and complete cross‐docking (CC‐D) studies, which involve performing docking calculations on all possible protein pairs within a given dataset.48, 49
After an early CC‐D study on a reduced test set,48 the computational power of the public World Community Grid, has allowed us to carry out CC‐D calculations on the complete Mintseris Docking Benchmark 2.0,50 which comprises 168 proteins forming 84 known binary complexes. A first analysis of the PPI energies and interfaces resulting from this large scale study (14,196 potential binary interactions) showed how the combination of docking and evolutionary information could improve partner identification within the benchmark.51 This work focuses more specifically on the information that can be obtained using the PIP (Protein Interface Propensity) value regarding the protein binding sites, which does not necessitate any prior knowledge regarding a protein's interaction partner. In addition, we developed a clustering algorithm for residues with a high interface propensity, that is, residues that are the most commonly found in protein interfaces resulting from docking calculations. This new approach shows how the distribution of potential binding patches on a protein surface will depend on the functional category this protein belongs to. Finally, we highlight several cases where cross‐docking calculations will lead to the identification of alternate protein binding sites corresponding to interaction partners that are not included in the original dataset.
In this section, we describe the MAXDo (Molecular Association via Cross Docking) algorithm that was developed for CC‐D studies.48 Since CC‐D involves a much larger number of calculations than simple docking, we chose a rigid‐body docking approach using a reduced protein model in order to make rapid conformational searches.
All simulations were performed using the unbound conformations of the proteins from the Docking Benchmark 2.0 of Mintseris et al.50 with the exception of 12 antibodies for which the unbound structure is unavailable and the bound structure was used instead. Any further reference to these proteins uses their name, or the Protein Data Bank (PDB) code52 of the experimental complex they belong to with the r or l extension denoting a receptor or a ligand protein respectively. For example, 1AY7_r and 1AY7_l refer to barnase (receptor) and barstar (ligand) in the barnase‐barstar complex 1AY7. The coordinates for the bound and unbound structures of both receptor and ligand proteins are available in the PDB and can also be found at the following address: http://zlab.bu.edu/zdock/benchmark.shtml. The 84 binary complexes listed in the Docking Benchmark 2.0 cover three broad biochemical categories and three difficulty categories related to the degree of conformational change in the protein‐protein interface upon complex formation. They are classified as Enzyme‐Inhibitors (E, 23 complexes, enzymes with a r extension and inhibitors with a l extension after the PDB code), Antibody‐Antigen (A, 10 complexes, antibodies with a r extension and antigens with a l extension after the PDB code), Bound Antibody‐Antigen (AB, 12 complexes, bound‐antibodies with a r extension and antigens with a l extension after the PDB code) and Others (O, 36 complexes). Note that for three cases in the AB category, namely 1IR9, 1KXQ and 2HMI, there was an inversion in the pdb files names, the antigen protein has a r extension and the bound antibody has a l extension after the PDB code.
We use a coarse‐grain protein model developed by Zacharias,53 where each amino acid is represented by one pseudoatom located at the Cα position, and either one or two pseudoatoms representing the side‐chain (with the exception of Gly). Ala, Ser, Thr, Val, Leu, Ile, Asn, Asp, and Cys have a single pseudoatom located at the geometrical center of the side‐chain heavy atoms. For the remaining amino acids, a first pseudoatom is located midway between the Cβ and C atoms, while the second is placed at the geometrical center of the remaining side‐chain heavy atoms. This description, which allows different amino acids to be distinguished from one another, has already proved useful both in protein–protein docking53, 54, 55 and protein mechanics studies.56, 57, 58
Interactions between the pseudoatoms of the Zacharias representation are treated using a soft LJ‐type potential with appropriately adjusted parameters for each type of side‐chain, see Table 1 in Ref. 53. In the case of charged side‐chains, electrostatic interactions between net point charges located on the second side‐chain pseudoatom were calculated by using a distance‐dependent dielectric constant ε=15r, leading to the following equation for the interaction energy of the pseudoatom pair i,j at distance rij:
where Bij and Cij are the repulsive and attractive LJ‐type parameters respectively, and qi and qj are the charges of the pseudoatoms i and j.
Our systematic docking algorithm (see Supporting Information Fig. S1) is derived from the ATTRACT protocol53 and uses a multiple energy minimization scheme. For each pair of proteins, the first molecule (called the receptor) is fixed in space, while the second (termed the ligand) is used as a probe and placed at multiple positions on the surface of the receptor. The initial distance of the probe from the receptor is chosen so that no pair of probe‐receptor pseudoatoms comes closer than 6 Å. Starting probe positions are randomly created around the receptor surface with a density of one position per 10 Å2, and for each starting position, 210 different ligand orientations are generated, resulting in a total number of start configurations ranging from roughly 100,000 to 450,000 depending on the size of the receptor.
During each energy minimization, the ligand protein is kept at a given location over the surface of the receptor protein, using a harmonic restraint to maintain its center of mass on a vector passing through the center of mass of the receptor protein. The direction of this vector is defined by two Euler angles θ and , (where θ==0° was chosen to pass through the center of the binding interface of the receptor protein) as shown in Supporting Information Figure S1. By using a Korobov grid59 and varying the Euler angles from 0°→360° and 0°→180° respectively, it is possible to uniformly sample interactions over the complete surface of the receptor and to represent its binding potential using 2D energy maps (each point corresponding to the best ligand orientation for the chosen θ/ pair). These maps where developed during the first phase of this project for validating the docking algorithm.48
Each energy minimization for a pair of interacting proteins typically takes 5 s on a single 2 GHz processor. As noted above, approximately 1,00,000 to 4,50,000 minimizations are needed to probe all possible interaction conformations, as a function of the size of the interacting proteins. Therefore, a CC‐D search on the benchmark, namely 168 × 168=28,224 receptor/ligand pairs, would require several thousand years of computation on a single processor. However, since each minimization is independent of the others, this problem belongs to the “embarrassingly parallel” category and is well adapted to multiprocessor machines, and particularly to grid‐computing systems. In the present case, our calculations have been carried out using the public World Community Grid (WCG, www.worldcommunitygrid.org) during the first phase of the Help Cure Muscular Dystrophy (HCMD) project. It took approximately six months to perform CC‐D calculations on the complete dataset of 168 proteins. More technical details regarding the execution of the program on WCG can be found in Ref. 60. The resulting CC‐D data is available for download at the following address: http://www.lcqb.upmc.fr/CCDMintseris /
Surface residues have a relative solvent accessible surface area larger than 5%. The accessibility is calculated with the NACCESS program,61 using a 1.4 Å probe. Interface residues present at least a 10% decrease of their accessible surface area in the protein bound structure compared to the unbound form.
In order to see whether cross‐docking simulations can give us information regarding protein interaction sites, we use the interaction propensity approach initially developed by Fernandez‐Recio et al.42 That is, we count the number of docking hits for each surface residue ri in protein P1, that is, the number of times each surface residue belongs to a docked interface between P1 and all its interaction partners in the benchmark. In earlier works,48, 51 we used a Boltzmann weighting factor which would favor docked interfaces with low energies. As a consequence, for a given protein pair P1P2, all interfaces with a 2.7 kcal.mol−1 or more energy difference from the lowest energy docked interface would have a Boltzmann weight lower than 1% (see ref. 51 for more details). Here, in order to limit the number of docked interfaces that would have to be reconstructed for determining the interface residues, which is the time consuming part of the analysis process, we chose to calculate the residues PIP (Protein Interface Propensity) values using only the lowest energy docking poses within this 2.7 kcal.mol−1 energy criterion, therefore we have
where Npos,P1P2 is the number of retained docking poses of P1 and P2 (which will vary with protein P2) satisfying the energy criterion, and Nint,P1P2(i) is the number of these conformations where residue i belongs to the binding interface. Finally, the PIP value for a given residue i belonging to protein P1 taking into account the CC‐D calculations within the whole benchmark will simply be the average PIP of this residue over all the possible partner proteins P2, that is
PIP values are comprised between 0 (the residue does not appear in any docked interface) and 1 (the residue is present in every single docked interface involving protein P1) and will be used for the prediction of binding sites. For each protein pair in the benchmark, between 1 and 215 docking poses were kept using the 2.7 kcal.mol−1 energy criterion, with an average of 11 docking poses (see Supporting Information Fig. S2a in the supplementary material for the distribution of the number of conserved poses for each protein pair), and for 60% of the protein pairs, five docking poses or less are kept after filtering on the interaction energy. These low statistics on each individual protein pair are compensated by the fact that every protein was docked with 168 different partners. Eventually, for each protein in the benchmark, between 900 and 4000 docking poses were used for to calculate the residues PIP values (see Supporting Information Fig. S2b).
Considering the PIP values results for all the residues, we define as predicted interface residues, residues whose PIP value lies above a chosen cutoff, and we can use the classical notions of sensitivity, specificity and the error function to evaluate their efficiency for the identification of protein interaction sites. Sensitivity (Sen.) is defined as the number of surface residues that are correctly predicted as interface residues (true positives, TP) divided by the total number of experimentally identified interface residues in the set (T). Specificity (Spe.) is defined as the fraction of surface residues that do not belong to an experimental protein interface and that are predicted as such (true negatives, TN). Additional useful notions that are commonly used include the positive predicted value (PPV, also called precision, Prec.), which is the fraction of predicted interface residues that are indeed experimental interface residues (TP/P), and the negative predicted value (NPV), which is the fraction of residues that are not predicted to be in the interface and which do not belong to an experimental interface (TN/N).
An optimal prediction tool would have all notions (Sen., Spe., Prec. and NPV) equal to unity. If this cannot be achieved, a compromise can obtained by minimizing a normalized error function based on the sensitivity and specificity values, which is comprised between 0 and 1 and defined as:
On a classic receiver operating characteristics (ROC) curve (with the sensitivity plotted as a function of 1‐specificity) the minimum error corresponds to the point on the curve that is the farthest away from the diagonal (which corresponds to random prediction).
In order to visualize how binding patches composed of residues with the highest PIP values can form on a protein surface, we use the following clustering algorithm: for a given protein, its surface residues are ordered following their PIP value. Starting with the residue with the largest PIP, each surface residue leads to the creation of a new isolated cluster, or, if any of its heavy atoms is <5 Å away from the heavy atoms of a residue already included in a cluster, is added to an already existing cluster. This process is implemented as long as the average PIP value of every cluster is larger than a given threshold named PIP clust. PIP clust is a relative value, and is expressed as a percentage of the maximum PIP value found on a single protein surface. For PIPclust=100% only the residue with the largest PIP value on this specific surface is selected, while for PIPclust=0% all surface residues are selected and form a single cluster covering the whole protein's surface. We define a protein's clustering profile as the curve showing the number of clusters on its surface as a function of the PIPclust criterion.
We must recall that, since the point of this work is to investigate the general binding behavior of protein surfaces with no prior knowledge of the binding partners, and not the correct docking of experimentally known partners, which can be achieved via other more effective but much more computationally demanding methods,62 we did not evaluate the quality of the best structural predictions for the 84 docked complexes. However, in an earlier work,48 where we performed cross‐docking simulations on a limited test‐set involving 12 proteins (using their bound structures), our method was able to predict correctly the position of the ligand protein with respect to its receptor with an rsmd of the Cα pseudoatoms below 3 Å; thus validating the quality of the force‐field used in the MAXDo program. Furthermore, this force‐field, which was originally developed by Zacharias for protein‐protein docking,53 has been successfully used on numerous occasions for the prediction of protein complex structures, especially during the CAPRI contest where the unbound structures of the protein partners are used.54, 63, 64
Figure Figure1(a)1(a) gives us a quantitative view of the results using the sensitivity and specificity notions defined in the Methods section. The ROC curve for the complete dataset (which contains >48,000 protein residues) with the variation of the sensitivity and specificity of the predictions is plotted in Figure Figure1(a),1(a), while Figure Figure1(b)1(b) shows the selection of residues potentially belonging to a protein interface as a function of a PIP cutoff comprised between 0 and 1. In Figure Figure1(a),1(a), the dotted diagonal corresponds to a random sampling of surface residues and divides the ROC space into areas of correct (above the diagonal) and incorrect (below the diagonal) classification. The greatest distance between the ROC curve and the diagonal yields the lowest error estimate, which is indicated by a dashed arrow in Figure Figure1(a).1(a). With a minimum error of 0.29 [see Supporting Information Fig. S3(a)], the optimal PIP cutoff [0.09, vertical dotted line in Fig. Fig.1(b)]1(b)] enables us to select 32% of the residues with a sensitivity of 71% and a specificity of 71% for the interface residues. These results are comparable to those that were obtained in our previous work cross‐docking the unbound structures of 12 proteins from a reduced dataset and which led to 30% coverage and 70% and 75% sensitivity and specificity, respectively. To measure the efficiency of our predictions independently of the cutoff value, we can also use the area under the curve (AUC) value, which is 0.77 for the black line in Figure Figure1(a)1(a) (while random predictions would yield an AUC value of 0.50). Finally, Figure Figure1(c)1(c) shows the variation of the precision with the PIP cutoff. Experimental interface residues represent only 11% of all the surface residues from our data set, which corresponds to the random precision value that would be obtained with a PIP cutoff set to 0. Increasing the PIP cutoff will increase the precision and the specificity of the prediction while decreasing the sensitivity. For example, setting the PIP cutoff to 0.4 would lead to a 40% precision, 98% specificity, but only 17% sensitivity.
Conformational changes upon binding usually define the difficulty level for docking experimentally identified complexes. The Docking Benchmark 2.0 comprises three groups labelled as rigid (63 complexes), medium (13 complexes) and difficult (8 complexes), which present average RMSDs of the Cα atoms from the interface residues of 0.82 Å, 1.63 Å and 3.67 Å, respectively. Interestingly, the prediction of binding interfaces using the PIP values performs slightly better for the medium, than for the rigid and difficult groups with average AUCs of 0.78 for the medium group, and 0.77 for the rigid and difficult groups respectively (see Supporting Information Table S1).
Using the biochemical categories discussed above (Enzymes, Inhibitors, Antigens, AntiBodies, Bound AntiBodies and Others) (see Table 1), we obtained the ROC curves of the PIP predictions for each of these sub‐groups [see Fig. Fig.2(a)].2(a)]. We can observe noticeable variations in the quality of the predictions depending on the protein type. While the curves for proteins from the Others group (purple line) and inhibitors (green line) stay close to the overall curve (black line), the predictions for enzymes and antigens interfaces appear to be much less effective, with ROC curves (in red and yellow respectively) below the overall curve, an increased minimum error function (0.33 for both groups, compared to 0.29) and decreased AUCs of 0.73 and 0.72 respectively. In contrast, interfaces from antibodies (both in the unbound, AB, and bound, BAB, groups), are better predicted than average, with ROC curves (dark and light blue lines) above the overall one, reduced minimum errors of 0.16 and 0.20 respectively, and increased AUCs of 0.90 and 0.85 respectively.
Instead of using the classic Sen./(1‐Spe.) ROC curve, one can also evaluate the performance of the PIP predictions with the Precision/Sensitivity (Recall) curve like what has been done by Hwang et al. for the prediction of binding interfaces from simple docking calculations.47 In that case the random AUC is no longer a constant and will depend on the proportion of experimental interface residues in each dataset, which corresponds to the Random Precision column on the far right of Table 1. Figure Figure2(b)2(b) shows the Precision/Sensitivity curves for the complete benchmark (in black) and all the functional subgroups. Once again we can observe a striking difference between the antigens (yellow line), with predictions that are barely above the random line, and the antibodies (dark blue), which present a curve that is markedly above all the others.
Using the clustering algorithm presented in the Material an Methods section, we can plot an average clustering profile for the complete benchmark, see Figure Figure3(a).3(a). As expected, for a maximum threshold PIPclust=1, we find a single cluster formed by the residue with the largest PIP value on the protein surface. As the threshold decreases, the number of clusters will either increase, when high‐PIP residues are scattered on different parts of the surface, or decrease, since adding a residue that belongs to two disjoint clusters will lead to their merging and the formation of a single cluster. The average number of clusters on the ensemble of protein surfaces will pass through a maximum value Nclustmax=2 for PIPclust=60%, before decreasing and reaching a final value Nclust=1 for PIPclust=0%, where all surface residues are selected. Figure Figure44 presents the progressive binding clusters growth on the protein surface for representative proteins from the enzyme, inhibitor, antibody and antigen categories. For example, reading the first line on Figure Figure44 from right to left shows how binding clusters (in blue), will increase in size and number on the surface of the α‐amylase enzyme (1BVN_r) upon decreasing the PIPclust threshold value. As could be expected, there are important variations in the proteins individual cluster profiles, in particular regarding the maximum number of clusters that can be found on the protein surface, which is comprised between 1 and 21 and appears to be strongly correlated with the proteins size (see Supporting Information Fig. S4). The position of the PIPclust threshold corresponding to this maximum will also vary widely from one protein to the other.
Figure Figure3(b–g)3(b–g) show the average clustering profiles for the six biochemical categories that are represented in the benchmark. As seen in Figure Figure3(b,c),3(b,c), proteins from the others and enzymes categories present an average profile that is similar to the general cluster profile from Figure Figure3(a),3(a), while inhibitors [Fig. [Fig.3(d)]3(d)] have a somewhat flatter profile, with a maximum cluster number under 2. This is probably due to the fact that, since inhibitors represent the smallest proteins on the benchmark, their surface is only large enough for two clusters at most, which will rapidly merge into a single one as the PIPclust criterion decreases. On the second line of Figure Figure44 we can see for example how the tendamistat inhibitor (1BVN_l) only presents one or two binding clusters, which will cover most of its surface at an early stage (PIPclust around 60% and under).
On the other hand, the average profiles for proteins from the antibody [bound, Figure Figure3(e),3(e), or unbound, Fig. Fig.3(f)]3(f)] and antigen [Fig. [Fig.3(g)]3(g)] categories are very specific. Antibodies usually present a single dominant cluster that will grow on the protein surface, and alternative binding patches will only appear at a later stage for low values of the PIPclust criterion (under 40%). As an illustration, the third line of Figure Figure44 shows the binding clusters growth in the case of the FAB Hyhel63 antibody (1DQJ_r), with a very late coverage of the protein surface. On the opposite, antigens will often present various disjoint binding clusters on their surface for high values of the PIPclust criterion (around 70%), which will afterwards merge in a single binding patch as the PIPclust criterion keeps on decreasing, see for example the fourth line in Figure Figure44 with the case of the pancreatic α‐amylase antigen (1KXQ_l). The individual clustering profiles for all the proteins in the dataset are available in the Supporting Information Figure S5.
Promiscuity, or multispecificity, is a frequent feature in the protein world65, 66 and a key issue when trying to predict proteic interfaces, since many proteins can actually present multiple binding sites, even when interacting with a single partner.67 This is also the case for proteins from the docking benchmark 2.0, and the apparent failure of cross‐docking for predicting experimental binding sites can result from the detection of alternative interfaces formed by the protein with other biomolecular partners. For example, Figure Figure5(a)5(a) presents the case of HIV1 reverse transcriptase (2HMI_r). The prediction of the binding site between this antigen and the FAB 28 antibody (2HMI_l) leads to a AUC value of 0.71, which is slightly smaller than the average AUC obtained for the prediction of interface residues in antigens (see Table 1). Coloring the antigen's surface using the PIP values resulting from cross‐docking, we can see that we only have a weak binding signal [Fig. [Fig.5(a),5(a), left panel, in white] for the antigen/antibody binding site. However, the 2HMI complex also comprises a DNA double strand that binds HIV1 reverse transcriptase on the opposite side of the antibody binding site [Fig. [Fig.5(a),5(a), center panel]. As we can see in the right panel from Figure Figure5(a),5(a), the antigen surface residues surrounding the DNA double strand present high PIP values (in blue), which means that our procedure can lead to the prediction of binding sites with a molecular partner (in that case DNA) that do not belong to the benchmark used for the cross‐docking calculations. We used the list of alternate interfaces and partners that has been established by Martin and Lavery49 for proteins from the Docking Benchmark 4.068 (which comprises many complexes from the 2.0 version) and found out several other cases of proteins with a low AUC value where the binding patches formed by surface residues with high PIP‐values correspond to experimental interfaces with alternate biomolecular partners. Figure Figure5(b,c)5(b,c) show how cross‐docking lead to the prediction of an alternate protein binding site for then enzyme colicin E7 nuclease (7CEI_r) and a DNA binding‐site for its inhibitor, the Im7 immunity protein (7CEI_l). More alternate interfaces are shown in Supporting Information Figure S6 with the references partner and the alternate partner systematically shown in black and green, respectively. Using the PiQSi webserver,69 we also searched for alternate experimental interfaces for the two proteins in the benchmark which presented the poorest binding site predictions (as expressed by their AUC value, see Supporting Information Fig. S7), namely the CMTI‐1 squash inhibitor (1PPE_l, AUC=0.21, green dot surrounded by a red circle in Supporting Information Fig. S7) and the T‐cell receptor β (1SBB_r, AUC=0.33, purple dot surrounded by a red circle in Supporting Information Fig. S7). Interestingly, in both cases the alternate binding sites predicted via the PIP‐values matches with the interfaces formed in homodimeric structures as can be seen on Figure Figure55(d,e).
We have shown that the PIP index provides useful information on residues more likely to be involved in protein‐protein interfaces, although its performance depends on the protein family (see Fig. Fig.2).2). Antigen surfaces in particular, seem to be more difficult to locate than the other interaction sites, including those of antibodies. Interestingly, this result agrees with the observations of Kowalsman and Eisenstein.70 In their work analysing protein‐protein docking models, they showed that interaction interfaces are enriched with high interface propensity residues, with the exception of antigenic sites. Both enzymes and inhibitors yield similar results regarding the interface propensity of their interface core residues (that is, residues that are totally buried in the experimental protein interface). In contrast, while antibody binding sites seem very well defined, including an important proportion of core residues with high interface propensity, antigen interfaces on the contrary are enriched with residues that have low interface propensities. Distinctions between enzyme/inhibitor and antigen/antibody interfaces have also been noted in other studies,71, 72, 73, 74, 75 reflecting the fact that, while both surfaces evolve to promote binding in the former case, only the antibody surface evolves in the latter.
Our cross‐docking results also show how the different evolutionary histories of antigens and antibody proteins impact their surface binding properties. In Figure Figure2(a,b),2(a,b), we can see that enzymes and inhibitors have comparable ROC curves, while the PIP index performs much better for the determination of antibody interfaces than for antigens. In Figure Figure2(b),2(b), it is quite striking that unbound antibodies are the only functional group for which it is possible to make precisions that combine a high precision (above 50%) and high sensitivity (80%). Another interesting feature of Figure Figure22 is that unbound antibodies (dark blue curve) perform better that bound antibodies (light blue curve) for the prediction of interface residues. This somewhat puzzling result may explained by the fact that, in the case of bound antibodies, the protein interface underwent specific conformational changes so as to specifically fit a given antigen, leading to a decrease of the quality of binding with every other potential partner. In contrast, for unbound antibodies, the protein interface may be less adapted to the specific antigen, but performs better with all the other proteins. This effect also appears when looking at the average AUC for the three classes denoting conformational changes upon binding in the benchmark, with cross‐docking on proteins from the medium category leading to slightly better binding site predictions than for protein from the rigid and difficult groups. Again, the unbound structure of proteins in the rigid group might be already adapted to their specific partner and less suitable for arbitrary docking, while a random partner will dock more easily on the unbound, and therefore unadapted, structure of a protein from the medium group. On the other hand, the conformational change undergone upon binding for proteins from the difficult group might this time be too important, thus making the binding site predictions more complicated for these systems. Interestingly, we would observe the opposite effect when looking at the prediction of interaction partners within the benchmark,51 that is, proteins from the rigid group would naturally perform better than those from the medium and difficult groups, since their unbound structure is already adapted to the specific partner we were looking for, thus highlighting the fact that partner and binding site prediction for protein are two different issues that each have to be dealt with via specific tools. In particular, while efficient binding site prediction tools, using for example evolutionary data,25 are nowadays available, binding partner prediction seems to be a much more complex problem which is still difficult to address.48, 51
Looking at the average clustering profile for proteins in different biochemical groups, we can see a specific binding behavior emerge for the surfaces of antibodies and antigens. Antibodies usually present a single binding cluster that will be located on a very specific part of the protein surface, while antigens have several binding clusters that are scattered all over the protein's surface. These trends are in agreement with the different performances that are observed for the prediction of the experimental interfaces in these two groups using the PIP values, and again reflect the different evolutionary pressures undergone by the antigen and the antibody functional groups.
An interesting feature of binding sites prediction via cross‐docking simulations, is that for some cases it will lead to the prediction of an alternate interface that corresponds to the interaction of the target protein with biomolecular partners that are not present in the original protein benchmark. This is the case for several proteins in the enzyme, inhibitor, antigen and others categories, for which, at first sight, the binding site predictions via cross‐docking seemed to be working badly, as shown from their weak AUC values compared to the average obtained for their functional category. In particular, it is quite striking that for two cases (the antigen HIV1 reverse transcriptase and the inhibitor Im7 immunity protein), this alternate binding site is actually a DNA binding site while the cross‐docking experiment was exclusively performed on PPI. Alternate predicted interfaces can also correspond to the interfaces found in homodimeric structures of the protein under study. In their work on arbitrary protein‐protein docking,49 Martin and Lavery could list 59 cases of protein with multiple interfaces. For over one half (31 proteins) of these cases, arbitrary docking would target the alternate experimental interface instead of the reference experimental interface corresponding to the binding partner present in the docking benchmark. In our case, for those nine proteins (shown in Fig. Fig.55 and Supporting Information Fig. S6) where alternate experimental interfaces were found, taking into account all the experimental interfaces residues (both from the reference and the alternate interface) for the evaluation of the binding site predictions via cross‐docking leads to a slight increase of the AUC (from 0.74 to 0.75) and a noticeable increase of the method's precision as can be seen on Supporting Information Figure S8.
In this work, we use high‐throughput cross‐docking calculations on a dataset of 168 protein unbound structures to develop a PIP index for the prediction of interface residues on the protein surface. The PIP index does not require any prior knowledge of a protein's potential interaction partners, and its predictive power depends on the protein functional group and is poorer for antigens. We also develop a clustering algorithm which permits to group together surface residues with high PIP‐values. The resulting clustering profiles for the various biochemical categories show remarkable differences, especially between the antibody and antigen groups, thus suggesting that the binding ability of the surface residues actually holds some information regarding a protein's potential function. To the best of our knowledge, this is the first time such a clustering analysis has been performed on proteins surfaces, and that a close connection between a protein's function and its binding behaviour has been established. Finally, the PIP index can also lead to the prediction of alternate interface that are not present in our original benchmark, but are still biologically significant. This means that the evaluation of its predictive power, that is based only on the prediction of a restricted number of proteic interfaces already present in the docking benchmark, is probably well underestimated. The regular finding of alternate experimental interfaces which concur with our cross‐docking predictions also highlight the fact that defining «true negatives» when dealing with PPI is actually a difficult matter.
In future work, we plan to apply the analysis tools that have been developed within the frame of this first experiment to the results from the cross‐docking calculations subsequently performed on a much‐larger protein database (made of 2246 proteins) that has been used in the second stage of the HCMD project. This database comprises a subset of >200 proteins that are known to be mutated and over‐expressed in neuromuscular disorder. Combining data resulting from cross‐docking simulations and evolutionary approaches, we aim to retrieve key information regarding the function, binding sites and potential binding partners of these target proteins.
Globally it is quite remarkable how a simple coarse‐grain rigid approach can still bring us many information on protein sociology. Just like we cannot expect every single individual in a group to behave the same way, a protein will not always behave and bind the way we expect it to, no matter how exact or detailed our protein model might be. Because proteins, like people, are complex systems, driven by a profusion of parameters that we cannot hope to identify in their entirety. However, if we set aside individual cases, that cannot, and should not, be trusted to tell us exactly how a protein behaves, and now turn our attention to the statistical data emerging from high‐throughput calculations, we can still retrieve valuable information regarding proteins general binding behaviour and function.
This work was originally carried out in the framework of the DECRYPTHON Project, set up by the CNRS (Centre National de la Recherche Scientifique), the AFM (French Muscular Distrophy Association) and IBM. The cross‐docking calculations were carried out on the World Community Grid (www.worldcommunitygrid.org) in phase 1 of the Help Cure Muscular Distrophy project. The authors wish to thank all the members of the WCG team for adapting our docking program for use on this grid, and also WCG volunteers for donating the computing time that made this work possible. S. S.‐M., A. C. and L. V. thank the Mapping Project (Investissement d'Avenir) ANR‐11‐BINF‐003 for postdoctoral funding.