key: cord-0931818-1drybykg authors: Mahita, Jarjapu; Kim, Dong-Gun; Son, Sumin; Choi, Yoonjoo; Kim, Hak-Sung; Bailey-Kellogg, Chris title: Computational epitope binning reveals functional equivalence of sequence-divergent paratopes date: 2022-04-30 journal: Comput Struct Biotechnol J DOI: 10.1016/j.csbj.2022.04.036 sha: 2307e7c44478ad4a3aab785a319e19eb1c0cf5e2 doc_id: 931818 cord_uid: 1drybykg The therapeutic efficacy of a protein binder largely depends on two factors: its binding site and its binding affinity. Advances in in vitro library display screening and next-generation sequencing have enabled accelerated development of strong binders, yet identifying their binding sites still remains a major challenge. The differentiation, or “binning”, of binders into different groups that recognize distinct binding sites on their target is a promising approach that facilitates high-throughput screening of binders that may show different biological activity. Here we study the extent to which the information contained in the amino acid sequences comprising a set of target-specific binders can be leveraged to bin them, inferring functional equivalence of their binding regions, or paratopes, based directly on comparison of the sequences, their modeled structures, or their modeled interactions. Using a leucine-rich repeat binding scaffold known as a “repebody” as the source of diversity in recognition against interleukin-6 (IL-6), we show that the “Epibin” approach introduced here effectively utilized structural modelling and docking to extract specificity information encoded in the repebody amino acid sequences and thereby successfully recapitulate IL-6 binding competition observed in immunoassays. Furthermore, our computational binning provided a basis for designing in vitro mutagenesis experiments to pinpoint specificity-determining residues. Finally, we demonstrate that the Epibin approach can extend to antibodies, retrospectively comparing its predictions to results from antigen-specific antibody competition studies. The study thus demonstrates the utility of modeling structure and binding from the amino acid sequences of different binders against the same target, and paves the way for larger-scale binning and analysis of entire repertoires. Protein binders (e.g., antibodies, nanobodies, repebodies, affibodies, DARPins, galectins, and monobodies [1] [2] [3] ) are capable of specifically recognizing different target proteins with high binding affinity, making them attractive candidates for therapeutic purposes. In recent years, considerable advances have been made in the techniques and tools involved in selection and development of protein binders. In particular, adaptive immune receptor repertoire sequencing leverages next-generation sequencing technologies to characterize the sequences comprising a repertoire of Bcell receptors or antibodies [4] [5] [6] [7] [8] [9] [10] [11] [12] , recently even including paired antibody heavy and light chains [13] . Despite such advances in rapidly discovering high-affinity binders, the process of character- izing their biological functions still remains low throughput. The binding specificity of a protein binder is governed by its binding site on the target protein, also known as the 'epitope' of an antibody, a word we adopt here generically to include sites recognized by other classes of protein binder. Thus epitope mapping, the process of identifying the epitopes of high-affinity binders, is considered an essential component of understanding a binder's mechanism and function [14] . While experimental structural determination remains the gold standard for identifying epitopes and permitting deeper insights into the factors governing specific recognition of an antigen, their labor-intensive nature makes them infeasible for scaling up to large sets of interactions, as immune repertoires may yield [15] . Other methods for epitope mapping that do not involve structural determination include site-directed mutagenesis or alanine scanning mutagenesis combined with binding assays [14, 16] . An alternative to epitope mapping is epitope binning, which uses competitive binding assays to sort antibodies into bins [17] [18] [19] [20] [21] . Unlike epitope mapping, epitope binning does not provide information about the location of the epitope on the antigen, but rather just that two binders compete with each other and thus are likely to target the same epitope (though perhaps compete due to steric hindrance distal to the epitope, conformational change by the antigen upon binding of one, etc. [22] ). Subsequent characterization of a representative from each bin assists in elucidating their functions and epitopes of the whole bin. Nevertheless, experimental epitope binning, albeit higher-throughput than epitope mapping, is still limited by the size of the repertoire. Since the sequences of protein binders encode their structures and consequently the determinants of their binding site specificity, in theory computational methods should be able to decode this information and predict where and how a set of binders with different sequences will bind an antigen. A number of different computational methods have been developed for computationally predicting binding regions (epitopes on antigens and paratopes on antibodies), applying machine learning methods based on sequence alone (e.g., [23] [24] [25] ) or based on antibody structures or homology models (e.g., [26] [27] [28] [29] [30] [31] ). Unfortunately, the utility of purely computational prediction of epitopes and paratopes remains limited in general due to the relatively low accuracy of their predictions. To overcome the limitations of methods that are purely computational or purely experimental, and better identify epitopes, researchers have pursued methods that integrate computational predictions with experimental data. Recently, a novel approach called ''dock binning" successfully localized epitopes for a panel of antibodies against glycoprotein D of herpes simplex virus, using docking models to identify putative antigenic ''hot spots" on the antigen across the antibody panel, and then leveraging experimental data regarding these hot spots in order to constrain docking models and reveal each antibody's specific epitope regions [32] . This study provided a limited demonstration of the ability of dock binning to discriminate binders with different epitopes, considering only one representative antibody from each distinct ''community" of epitope binders. Thus, there is still a need for a purely computational epitope binning method that not only discriminates binders having different epitopes but also clusters together sequentially-divergent binders sharing the same epitope. In this work, we investigate the extent to which competition for binding to an epitope can be predicted directly from the amino acid sequences of a set of different but related binders, based on the sequences alone or using the sequences as the basis for models of structure and docking models which represent predicted interactions between the binders and the target. In particular, we seek to computationally analyze the functional equivalence of the binding regions, or loosely ''paratopes" (as with ''epitope", again adopting for general use with any protein binder the corresponding word used with antibodies), of a set of distinct binders of a common target. We take as a case study a set of interleukin-6 (IL-6) binding repebodies generated through phage display of a diversified repebody library [33] . A repebody is a binding scaffold comprised of leucine-rich repeat (LRR) modules whose b-strands mediate antigen binding [3, 34] ; it serves as an excellent system for the study at hand due to the ease of generating antigenspecific libraries with different binding modes driven by diversity in a relatively small set of residue positions in well-defined binding modules [35, 36] . Binding competition assay experiments on IL-6 binding repebodies indicate that there are largely two different sets (bins) of binders, with one bin composed of sequence dissimilar (yet competing) binders. Computationally predicted competition revealed that although sequence similarity was in general an indicative metric for binding competition, an alternative approach was required for the set of sequence-dissimilar binders that were grouped into the same bin. Consequently, we developed Epibin, inferring functional equivalence of paratopes by leveraging the docking hypothesis that key structural features of binders mediate binding and molecular docking largely captures binding similarity. Epibin successfully classified binned variants sharing low paratope sequence identity and identified specificity-determining residues that confer functional equivalence. In previous work [33] , a library of repebody clones was generated by diversification of six positions in two of the leucine-rich repeat modules (hereafter called hypervariable sites 1, 2, and 3 of modules 3 and 4) and screened by phage display for IL-6 binding. Seven clones ( Fig. 1) were selected for further analysis on the basis of expression, mutation variability, and binding. While the structure of one of these repebodies in complex with IL-6 (clone ''D3") has been solved (PDB ID: 4J4L [33] ), the details of the interactions of the other six repebodies with IL-6 remain unknown. The sequences show diversity throughout the two variable modules (Fig. 1b) . From previous structural studies [33, 37] , it is notable that the presence or absence of Pro in hypervariable sites (e.g. 3 of module 3, Fig. S1 ) induced a significant conformational change. Thus, structural models of the six repebodies were constructed accordingly, based either on the above-mentioned crystal structure for those with Pro or on a different repebody (PDB ID: 3RFS [3] ) for those without Pro (Fig. S1 ). Structural superposition of these two classes of repebodies consequently highlights distinct conformational differences in module 3 (Fig. 1c) . ClusPro-based docking [38] of these models to the crystal structure of unbound IL-6 (PDB ID: 1ALU [39] ) yielded approximately 30 models per repebody (Fig. 1d) , which were further subjected to computational binning analysis (''Epibin", Fig. 1e ) as we introduce below. Competition between repebodies was predicted based on sequence similarity (quantified as BLOSUM62 substitution scores summed across the six hypervariable positions of the repebody, Fig. 1b) , binder structure-based dissimilarity (root-mean-squared deviation of the C b atoms over the whole structure of the undocked repebody homology model, Fig. 1c ), and interaction similarity (the Epibin score we introduce here, Fig. 1e ) based on docking models. It was also experimentally assessed with competition binding assays. The Epibin score predicts competition for epitope binding based on pairwise similarities of complex models produced by docking. While many of the docking models are likely to be incorrect (with perhaps one near-native, some relatively close, and more well off-target [40] [41] [42] ), intuitively they rely on and thus reveal complementarity between a putative paratope and putative epitope. Therefore, if the docking models of two protein binders involve many of the same regions on the target, then the binders are likely to have similar binding preferences and compete for the (unknown) true epitope. To leverage this intuition in predicting the extent of competition between binders, we devised a scoring function (Epibin score) separately considering each pair of docking models, one from each binder, and evaluating the fraction of docking models from one binder that have a sufficiently close competitor from the other, and vice versa (Eqs. 1 and 2). The fraction of models with competitors is subtracted from 1 to be comparable to experimental measurements, with Epibin score of 0 for complete competition to 1 for no competition. Fig. 2 summarizes both the predicted competition, as heatmaps (Fig. 2a, Table S1 ), and the resulting binning, as dendrograms (Fig. 2b) . Sequence-based analysis predicts strong competition among repebodies B10, D3, and B3 (herby referred to as the ''D3 cluster") due to highly similar residues within modules 3 and 4 of these repebodies. Binning based on modelled binder structure-based dissimilarity, as estimated by RMSD, further tightly clusters A4, A10, and G3, due to the presence of Pro in the hypervariable site 2 of module 3. In contrast, Epibin predicts A4 to strongly compete with B10 and C6, to moderately compete with B3 and D3, and to weakly compete with G3. In addition, Epibin also predicts weak competition of G3 with B10 and C6. The competitive immunoassays (here symmetrized by averaging, in order to support comparison with the computational methods) showed that, consistent with Epibin but counter to sequence-and structure-based predictions, A4 indeed competes with B10 and to a lesser degree, with B3 and D3 (Fig. 2a, Fig S2) . On the other hand, immunoassays revealed little competition between A4 and C6. As expected from all analyses, the repebodies within the D3 cluster also experimentally manifested high levels of competition. Finally, the experiments revealed competition between A10 and G3 that was only predicted by binder structure model-based dissimilarity analysis. In inferring clustering and bins from the patterns of competition (Fig. 2b) , the experiments and all methods grouped the D3 cluster together, though Epibin brings A4 into the core of the group whereas the experiments place it just outside, and both sequence-and structure-based methods associate it most tightly with A10. Overall, the Pearson correlation coefficient (PCC) value The amino acid differences in the repebodies manifested as structural differences in homology models. (d) The structural differences in the repebody models led to predicted complex structure differences according to docking models (for clarity, only a few models out of $30 per repebody are shown). (e) The docking models were analyzed by Epibin in order to predict competition for binding between pairs of repebodies. First, Epibin computes the ''epitope overlap score (EOS)", the extent of common epitope residues (0: none to 1: all). In the example, B10_1 and B3_1 represent docking models of B10-IL-6 and B3-IL-6, respectively. Comparing the epitopes of these two docking models indicates that their epitopes overlap to some extent, quantified by an EOS of 0.6. Similarly, B10_2 (third row) and B3_4 (fifth column) docking models share common epitopes, yielding an EOS of 0.7. On the other hand, docking models B10_2 and B3_1 do not have any common epitope residues and hence the EOS between them is 0. Similarly, B10_4 does not share any common epitope residues with any of the B3 docking models, yielding EOS = 0.0 for the whole fourth row. Based on the set of EOSs, the overall ''Epibin score" capturing competition between two repebodies is computed as the average fraction of each repebody's docking models that have a sufficiently strong competitor for the same epitope among the other repebody's docking models (i.e., EOS exceeds a threshold). In the example, the cells containing the 'D' symbol indicate those pairs of models for which the EOS is above the threshold of 0.5 and hence included in the fraction. Among the illustrated pairs of docking models, we see that three of the four B10 models (rows) have a strong competitor among the B3 models (columns), and all four of the B3 models have a strong competitor among the B10 models. For consistency with experimental data, the average fraction of sufficiently overlapped models is subtracted from 1 to yield the Epibin competition score, here giving a score of 1 -0.5*(3/4 for B10 + 4/4 for B3) = 0.125. of 0.9 (Fig. 2c) indicates a strong positive correlation between Epibin-predicted competition and experimentally-observed competition, emphasizing the predictive utility of computational epitope binning. The Epibin results presented in Fig. 2 are based on a single EOS threshold (0.65) and the complete set of ClusPro-generated docking models (a total of 30 for each repebody -IL-6 pair except for A4 -IL-6 for which ClusPro returned 24 models). To assess the impact of varying the EOS stringency and docking model completeness, we further evaluated Epibin at different EOS thresholds ranging from 0.5 to 0.75 and with sets of the top 20, top 15, top 10, and top 5 docking models each (Table S2 ). The PCC between Epibinpredicted competition and the observed competition when using the set of 30 models ranges from 0.69 at an EOS threshold of 0.5 to 0.94 at 0.7, indicating that this control over stringency does impact the quality of the results. For the purpose of illustrating the Epibin results in Fig. 2 , we chose an EOS threshold of 0.65 (PCC of 0.9, marginally less than for the higher threshold of 0.7) since it balances accuracy over a range of test cases. Moreover, the range of PCCs over subsets of docking models varies only between 0.90 and 0.92 at this EOS of 0.65, suggesting that Epibin is robust to taking smaller subsets of the top docking models. We also evaluated the impact of the docking method, by applying Epibin to the models returned by the HADDOCK webserver [43] (a total of 400 unranked models for each repebody -IL-6 pair). However, Epibin did not fare quite as well as it did on the ClusPro docking models, with PCC ranging from 0.65 at an EOS threshold of 0.5 to a maximum of 0.78 at an EOS threshold of 0.65 (Table S2 ). While A4 and B10 have distinct differences in sequence and modelled structures, analysis of putative interactions via Epibin was able to successfully predict their experimentally observed competition. Therefore, we sought to delineate the factors contributing to this competition. For the sake of comparison across the repebodies we label paratope equivalent residues that may make similar contributions to recognition of the antigen, regardless of where they are in the sequence (Fig. 3a) . We base the labeling on the D3 cluster of repebodies (D3, B3, B10, and C6), which present the same paratope, and we label as 'a' the structural position contributed by their hypervariable site 1 on module 3, 'b' module 3 site 2, 'c' module 4 site 1, and 'd' module 4 site 2. Module 3 0 s hypervariable site 3 is labeled 'e', but the Pro occupying this position is pointed towards the interior of the repebody so does not participate in binding interactions for this cluster. By comparing the structure of A4 to that of B10 ( Fig. 3 .b.i), we see that in A4 structural positions 'a' and 'c' still derive from the same sequence positions as in B10, but in contrast 'b' is presented by module 3 site 3 instead of module 3 site 2. Similarly, 'd' is presented by module 4 site 3 in A4 as compared to module 4 site 2 in B10. Strikingly, these structurally equivalent positions are occupied by physicochemically similar amino acids (Fig. 3c) , with both repebodies using an Asp in position 'b' in A4 and B10, and the other competing repebodies B3 and D3 using Glu there (Figs. S3(a, b) ). In the solved D3 -IL-6 complex structure, this Glu (Glu 128) makes electrostatic interactions with Lys 27 and Arg 30 of IL-6, and it is likely that the Asp residues are also involved in similar interactions. Likewise, the 'd' position is filled by Phe in A4 and Trp in B10 and D3 ( Fig. 3b.i; Figs. S3(a, b) ), and in the solved D3 crystal structure this Trp (Trp 152) forms a hydrophobic pocket with two other aromatic residues, Phe 177 and Tyr 201. We note again that only the interaction-based prediction of competition via Epibin was able to pick up on this paratope equivalence, which uses different residue positions from A4 and B10 to fill a similar structural/functional role. The repebodies A10 and G3 that do not compete with A4 and B10 place physicochemically different amino acids in these positions: 'b' is Ala in A10 and Trp in G3 (vs. Asp in A4 and B10), while 'd' is Arg in A10 and Glu in G3 (vs. Phe in A4 and Trp in B10) (Fig. 3b.ii; Fig. 3c; Fig. S3 .c). Thus, Epibin is able to pick up on the resulting differences in epitope preferences in predicting that they do not compete. In order to test the hypothesis that positions 'b' and 'd' are important for competition between A4 and the D3 cluster, we designed single-point and double-point mutant variants of A4, B10, and C6 (Table 1) . Based on our hypothesis that the negatively charged amino acid in position 'b' and the aromatic residue in position 'd' act as specificity determining positions, residues in these positions were mutated to either resemble or differ from the original physicochemical properties. In detail, the Asp residue in the 'b' position of A4 and B10 were mutated to Glu, Arg, or Lys; in the position 'd', Phe of A4 was mutated to either Trp or Val, and the Trp of B10 was mutated to either Phe or Val. Furthermore, since B10 and C6 have different residues in the 'b' and 'd' positions ( Fig. S3.d) , we reasoned that mutating the residues in these positions of C6 to mimic the interactions of B10 would increase the competition between A4 and C6 mutants (Fig. S3.e) . In particular, C6 variants were constructed with the Trp residue in position 'b' mutated to Asp and/or the Val residue in position 'd' mutated to Trp ( Table 1 ). The competition of the variants against wild-type forms of A4, B10 and C6 were computationally predicted with Epibin as well as experimentally evaluated. The prediction trends well (though without statistically significant correlation) with the experimental observations (Fig. 4a, Fig. S4 .a). Firstly, for A4 and B10 variants, in accordance with our hypothesis, we assumed that mutating Asp in position 'b' to either Lys or Arg would greatly abrogate competition, and even binding to IL-6. As per our expectations, all of the B10 mutations to Lys or Arg in 'b' eliminated IL-6 binding. The outcomes were similar for A4 variants, where all except A4KV lost their binding to IL-6 ( Fig. 4b, Fig. S4.b) . In the case of A4KV, the binding affinity was maintained to a similar extent to wild-type A4 but was not able to compete with any of the wild-type binders (Fig. 4a) . Mutations of Phe or Trp in 'd' to Val removes the aromatic functional group, but since it is not a replacement by an opposing functional group (as in Asp to Lys/Arg), we expected that these mutations would likely retain the ability of these variants to bind IL-6, though with weakened strength and competition. While A4DV and A4EV maintained weak binding, complete loss of IL-6 binding was observed in B10DV and B10EV. This may originate from the fact that the wild-type sequence in 'd' for A4 is Phe and B10 is Trp. In this sense, we assume that B10 undergoes greater reduction in the size of the functional group in 'd' when mutated to Val and therefore manifests a larger decrement in binding compared to A4. In line with our hypothesis, experimental evaluation showed that A4EV competed to a lesser extent with wild-type binders for IL-6 binding. Other than A4EW, the physicochemically-similar mutations (i.e., A4DW, B10DF, and B10EF), exhibited similar trends of predicted and observed competition. Next, we set out to investigate whether the introduction of Asp in the 'b' position of C6 and Trp in the 'd' position would lead to increased competition with A4. All of the C6 mutants were predicted to compete with the wild-type C6 to a similar extent. However, we found that the mutation in position 'b' to Asp induced a formation of oligomeric species for both C6DV and C6DW (Fig. S4c.(iii, iv) ), which is likely the main reason for these two variants displaying significantly weaker binding to IL-6. Nevertheless, increased competition for C6WW against both A4 and B10 wild-types indicates that an aromatic functional group in 'd' is indeed a key specificity-determining residue (Fig. 4a) . Overall, these results support Epibin's ability to uncover structural/functional equivalence in the paratope, thus outperforming sequence-based or binder structure-based predictions in predicting competition. While investigation of competition gives an idea about which repebodies are likely to bind to the same epitope, along with their paratope-equivalent residues, it does not provide any information about the location of the epitopes. The observed competition between A4 and B10 suggests that their epitopes are common or located in close proximity to one another. In addition, based on the high sequence identity between the members of the D3 cluster of repebodies (B3, B10, C6, and D3), it is conceivable that the D3 epitope is shared by the other members of the D3 cluster. Also, while immunoassays showed competition between A10 and G3, Epibin did not predict any competition between these two repebodies. To better understand the mechanisms at work here, we sought to localize the epitope on IL-6 for each of the six repebodies. We made use of EpiScope [44] , a method that views antigen-antibody docked models as hypothesized binding modes and designs sets of antigen variants to test these hypotheses. If a docking model is correct, then unfavorable mutations in the associated interface (e.g., switching amino acid charge, size, hydrophobicity, etc.) should disrupt binding. EpiScope optimizes small panels of such mutational variants to efficiently test the various docking models, using computational protein design to identify combinations of mutations likely to disrupt putative binding while maintaining antigen stability. EpiScope has been successfully applied to localize the binding site of many protein binders with diverse scaffolds [35, [44] [45] [46] . Here, EpiScope was applied separately to each repebody, designing each repebody-specific IL-6 triple-mutants of which their binding disruption was examined experimentally, and thereby localizing the binding region. Table 2 shows the list of mutated sequence for each repebody. EpiScope designed a minimal set of triple-mutants predicted to effectively disrupt binding regardless of which docking model is most accurate. In the case of the D3 cluster of repebodies (i.e., B3 and B10), which have high sequence similarity, the available IL-6-bound crystal structure was directly used in place of docking models to guide the design. The residues that are mutated in these IL-6 designs are highlighted in Fig. 5a . Overall, three distinct candidates for the epitope regions can be discerned: one containing residues Arg 24, Lys 27, Arg 30, and Tyr 31 (colored green in Fig. 5a) , one Ala 38, Arg 40, Lys 41, Asn 45, Met 49, and Ser 52 (colored brown in Fig. 5a) , and one Ala 68, Phe 74, and Glu 80 (colored blue in Fig. 5a ). The first epitope region overlaps with the crystal structure-based design that abrogates D3 binding; thus, the bind- Table 1 List of designed variants of A4, B10 and C6. Mutations made in repebody modules 3 and 4 (position 'b' and/or position 'd') are highlighted in bold. ing of the D3 cluster and A4 to designs spanning this region were expected to be disrupted. The third region is unique for A10, which implies that A10 produces docking models different from those produced by the other repebodies. Binding disruption was evaluated via ELISA testing repebody binding against the respective designed IL-6 triple mutants vs. that against wild-type IL-6 ( Fig. 5b, Fig. S5 ). As expected, B3, B10, and C6 were all revealed to interact with IL-6 in a binding mode similar to that of D3, as inferred by binding disruption of the triple-mutant in the D3 binding interface. Surprisingly, A4 was shown to interact predominantly with the second epitope region, whose mutations largely disrupted its binding, while mutations in the first epitope region resulted in intermediate binding disruption. This implies that although A4 shares similar functional paratope equivalence with the D3 cluster of repebodies (position 'b' and 'd' in Fig. 3) , they do not employ identical binding modes as their backbone alignments may differ. This could also be one of the reasons A4 was more resistant than B10 to the mutation of the 'd' position Phe to Val. For A10, binding was disrupted only for the third epitope region, located opposite to the D3 epitope, and for G3 none of the variants resulted in any reduction in binding. Since in the competitive ELISA, A10 and G3 were shown to compete, we also tested binding of G3 to the IL-6 triple-mutant variants designed for A10. G3 was found to retain binding to these variants, which suggests that the G3 epitope is likely located in the vicinity of the A10 epitope and the observed competition between them could be due to steric clashes while attempting to bind their respective epitopes. This further illustrates that experimental competition assays cannot determine whether the competition is due to an overlap of the epitope or, as in this case, steric hindrance. Therefore, for the purpose of selecting and developing therapeutics, it may be very beneficial to leverage the information provided by Epibin regarding whether the binders actually share an epitope or not. The ability to group entire repertoires of antibody sequences by predicted epitope specificity could provide large-scale functional insights into the effects of infection and vaccination, as well as the results of antigen-specific library screening. While our main focus here is on repebodies against IL-6 due to the availability of high-quality experimental data and the ease of conducting follow-up experiments, we were also interested in assessing how well Epibin is able to bin groups of antigen-specific antibodies. In order to gain preliminary insights, we investigated three different sets of antibodies for which experimental cross-blocking data is available: twelve anti-D8 antibodies [47] , ten anti-Pfs25 antibodies [48] , and nine anti-SARS-CoV-2 antibodies [49] . Following an analogous approach to that implemented for repebodies, competition Correlation between Epibin-predicted competition (x-axis) and corresponding observed competition (y-axis) of all designed variants competing against the wild-type forms A4, B10, and C6 (separate panels by wild-type). Pearson's correlation coefficient is denoted as r. (b) Binding affinity measurement of the wild-type and variant forms of A4, B10, and C6 against IL-6 as measured by ITC. Bars show the means and SDs over three replicates. EpiScope-generated triple-mutant variants of anti-IL-6 repebodies to localize epitope regions on IL-6. Triple -mutant variants A10 A50T|F56L|E62K, K9E|Y13S|A20E, N27D| M31K|S34N A4 Y13S|R22S|K23E, R6K|K9E|Y13S, R22S|K23D|N27E B10 R6Q|R12K|Y13S, N27D|M31K|S34D, Y13S|R22K|K23E B3 R6Q|R12K|Y13S, N27D|N30D|M31K, R12K|Y13S|K23E C6 N27E|M31V|S34D, R6Q|R12K|Y13S, R12K|R22K|K23E G3 R22K|N27E|M31K, R6K|K9E|Y13S, R12K|Y13S|R22S 4J4L (D3 Clone) R6K|K9E|R12K, R6Q|K9E|R12K, L1G|R6K|K9E, L1G|R6Q|R12K, L1G|R6L|R12K between the antibodies was predicted based on sequence identity among the CDR loops, which mainly drive antibody recognition of the antigen, structural similarity of the variable regions, and Epibin-based comparison of docking models. Furthermore, for Epibin evaluation, both ClusPro and HADDOCK models were tested, using either the full ClusPro set of $25-30 models or just the top 5, 10, 15, or 20, and the EOS threshold was varied from 0.45 to 0.7. To evaluate the ability of the different computational binning methods to correctly classify pairs of antibodies as either competing or non-competing, we calculated the area under the precisionrecall curve (AU-PRC) evaluating agreement between predicted and observed competition (as classified in the original publications) over the range of prediction probability thresholds. We note that AU-PRC is preferred to area under receiver operating characteristic curve, AU-ROC, for imbalanced datasets like these. We observed that overall, the best results were obtained with the top 20 ClusPro models at an EOS of 0.45 (Table S3 ). In the case of the anti-Pfs25 antibodies, CDR sequence identity yielded 0.80 AU-PRC and Epibin 0.78, while structural similarity of the antibody variable region only obtained 0.56 (Table S3) . For the anti-SARS-CoV-2 dataset, Epibin substantially outperformed the others, with 0.60 compared to 0.32 for sequence and 0.25 for antibody variable region structure (Table S3 ). However, for the anti-D8 dataset, structural similarity of the antibodies as gauged from RMSD measurements was by far the best score, yielding 0.91 AU-PRC, compared to 0.40 for sequence similarity and 0.60 for Epibin (Table S3) . Epibin results using HADDOCK models were not quite as good as those using ClusPro models (Table S3 ; note that the server failed to return results for the anti-D8 Abs). Although Epibin demonstrated promising performance on antibodies, the modest results when compared to its performance on repebodies could be attributed to other factors such as the inherent nature of the datasets where the antibodies themselves could possibly have widely different epitope specificities with little room for competition, or have been chosen based on a sequence identity cutoff. Thus, despite having scope for improvement, these initial results also demonstrate the general utility of Epibin for binning large sets of antibodies based on available information about the antibody sequences and antigen. Epitope binning facilitates the selection of diverse antigenbinding candidates from a large pool by grouping into bins those with shared features. While experimental binning methods have scaled to assess competition between numerous clones, nextgeneration sequencing methods go much further, identifying thousands or even millions of different clones in repertoires [12, 50] . This gap begs for computational methods that can characterize the different sets of binding functions within a repertoire, based on sequence information alone, without needing to isolate and experimentally evaluate the different clones. The study presented here evaluates the extent to which comparison of sequences, modeled structures, or modeled interactions, leads to binning of a set of seven anti-IL-6 repebodies in a manner consistent with experimentally-observed competition. We show that the information provided by our Epibin computational epitope binning method was beneficial not only in distinguishing important com- petition, but also in pinpointing the paratope-equivalent residues contributing to the binding specificity of each bin. For both repebodies and antibodies, the accuracy of a computational predictor naturally depends on the quality of the models. Here, binning predictions are based on docking models which in turn are influenced by the nature of the interacting proteins. For example, protein interactions involving flexible binding sites or significant conformational changes are particularly challenging. The docking models further depend on the initial models of the individual proteins. In the case of repebodies, the presence of a Pro in one of the binding loops may affect the quality of the model, while in the case of antibodies, CDR-H3 may be difficult to model correctly. Furthermore, the antigen structure used in docking may not capture the same state as that probed experimentally, e.g., if there are post-translational modifications such as glycosylation which is common in viral antigens. Despite these limitations, with the methods for modeling proteins and interactions continuing to improve and the databases upon which they build continuing to expand, computational epitope binning promises to enable large-scale investigation of functional diversity across repertoires of antigen-specific binders. The generation and selection of a panel of diverse anti-IL-6 repebodies was previously described [33] . All 42 clones were compared against each other and those with more than 4 common sequences in the hypervariable region were discarded from the set. Among the 11 clones with sequence variance, only those with sufficient and comparable binding intensity were chosen through ELISA (data not shown). Due to the structural importance of Pro, repebody homology models were based on the D3 crystal structure (PDB ID: 4J4L, [33] ) for those containing Pro in the hypervariable site 3 of module 3 (B3, B10, and C6) and on a different repebody crystal structure (PDB ID: 3RFS [3] ) for those without Pro (A4, A10, and G3). Twenty models per repebody were generated using MODELLER v9.12 [51] and the best model was selected based on the DOPE score [52] . Models were further energy-minimized by the Tinker (ver. 6) molecular dynamics software [53] using the GB/SA implicit solvent model [54] with the AMBER99sb force field parameter [55] . The structure of unbound IL-6 (PDB ID: 1ALU) [39] contains a few missing residues which were filled using MODELLER and energyminimized as described above. The modeled repebodies were docked on the unbound structure of IL-6 using two docking webservers -ClusPro [38] and HADDOCK [43] . When utilizing the Clu-sPro webserver, the Antibody Mode scoring function [56] was employed, using residues making up the concave portion of the repebody to guide the docking. The ClusPro webserver generated 24-30 models for each repebody -IL-6 pair. For the HADDOCK webserver, residues comprising the concave portion of the repebody and all surface residues of IL-6 were used to guide the docking. The HADDOCK webserver generated 400 unranked models for each repebody -IL-6 pair. Repebodies were evaluated for competition based on comparisons of their sequences, their modeled structures, and their modeled interactions with the antigen. Repebody sequences vary only in six positions that are restricted to Modules 3 and 4. All-against-all pairwise BLOSUM62 substitution scores, used as a measure of similarity, was calculated on the basis of these six hypervariable positions. All-against-all pairwise root mean square deviations (RMSD) of C b atoms were calculated for the modelled structures of the six repebodies along with the D3 crystal structure (Fig. 1c) . For the antibodies, C b -RMSD was calculated on the variable regions of heavy and light chains. The Epibin score for a pair of protein binders integrates information about their docking models against a common target to assess how likely it is that the binder will compete for the same epitope on that target. We define an antigen residue to be in the epitope region according to a particular docking model if any of its non-hydrogen heavy atoms are within 6 Å of any of the binder's heavy atoms in that docking model. We compute the ''epitope overlap score" (EOS) (Fig. 1e) for a pair of docking models based on the similarity of their two sets of ''epitope" residues defined in the models. Let A and B be two docking models and eðÁÞ be a function that gives the epitope residues in a docking model as described above. Then the EOS is defined as the Jaccard index: For the Epibin score, we compute the fraction of one binder's docking models that have competitors from the other binder. Let R and S be two sets of docking models and IfÁg the indicator function yielding 1 if a predicate is true and 0 if it is false, based on the overlap threshold h. Then the Epibin score is defined as follows: A score of ''0" indicates the two sets of docking models are most likely overlapped, whereas ''1" means no docking models share a sufficient number of epitope overlaps according to the epitope overlap threshold (h). The results presented here use h ¼ 0:65 except when examining the effect of different values of h on the Epibin scores for the repebodies and antibodies as indicated in the Results section. Repebodies were hierarchically clustered (calculated by average linkage) based on one of these similarity/dissimilarity scores using the clustermap function of the Python data visualization software Seaborn [57] . Human interleukin-6 (IL-6; amino acids 19-184) was cloned into a pET21a vector with 6xHis-tag and a thrombin cleavage sequence inserted at the N-terminus of the IL-6 gene. Repebody genes were selected from the previous report [33] and synthesized by Integrated DNA Technologies. Repebody genes were cloned into a pET21a expression vector with either a myc tag or a flag tag attached to the C-terminus. All cloned genes were transformed to Origami B(DE3) cells for expression. Cells were cultured at 37°C, 200 rpm using LB media to OD 0.6 and induced with 0.5 mM IPTG. Induced cells were harvested by centrifuge at 4,000 g, after 20 h of culture at 18°C, 200 rpm. Harvested cells were lysed through sonication and centrifuged at 13,000 rpm for 1 h. The supernatant was filtered, and affinity purified with Ni-NTA (Qiagen). Proteins were further purified with 20 mM Tris-HCl (pH 8.0), 50 mM NaCl buffer through size exclusion chromatography (AKTA FPLC, GE Healthcare). Competitive ELISAs were conducted with the seven repebodies. Human IL-6 (10 lg/mL) was immobilized to 96-well maxisorp plate (SPL Life Sciences) overnight at 4°C. Plates were blocked with 2% PBST-BSA for 2 h at 4°C. Purified repebody A (myc tag) was mixed with another repebody B (flag tag) in molar ratios of 1:0, 1:1, 1:2 and 1:4 to a final concentration of repebody A of 100 lg/ mL. After washing the 96-well plate with PBST, the repebody mixture was treated and incubated for 1 h at room temperature. Plates were washed three times before addition of anti-myc tag antibody-HRP conjugate in 1:400 dilution (sc-40 HRP; Santa Cruz Biotechnology). After 1 h of incubation with antibody at room temperature, tetramethylbenzidine solution (Sigma) was added to each well, and the reaction was stopped immediately by adding 1 M sulfuric acid. The results were obtained by scanning absorbance at 450 nm using Infinite M200 plate reader (Tecan). The experimental competition score was calculated as the ratio of the absorbance taken at 1:0 (myc:flag) to the absorbance observed at 1:4. Competition between two repebodies was considered to be present if this ratio was 0.5. In order to render the results directly comparable with the computational prediction, which are necessarily symmetric, the A:B score and B:A score were averaged to yield a symmetric competition score (Table S1 ). Repebody epitopes were localized by application of the computational-experimental EpiScope method [44] : design IL-6 variants mutated so as to disrupt docking model interface regions, and experimentally evaluate the effects on binding of the mutations in order to test the hypothesized binding modes. The docking models described above were provided as input to EpiScope, which designed for each repebody a set of repebodyspecific triple-mutant IL-6 variants predicted to maintain IL-6 stability while disrupting each of the binding interfaces present in the models. There were a total of 12 triple-mutant variants, three each for A4, A10, C6, and G3. Due to their high sequence similarity with D3, clones B3 and B10 were expected to resemble the binding mode of D3. Thus, triple-mutant IL-6 variants based on the crystal structure of D3 in complex with IL-6 (PDB code 4J4L) were designed and tested prior to EpiScope. EpiScope was not applied to B3 and B10 as the crystal structure-based triple-mutant successfully disrupted binding of the two clones. The designed variants of IL-6 with triple point mutations were cloned, expressed, and purified in the same methods as wild-type human IL-6. For the binding assays, 10 lg/mL of each IL-6 variant was immobilized onto 96-well plates by incubation overnight at 4°C. The plates were blocked with 2% PBST-BSA solution for 2 h at 4°C. Following washing with PBST, each repebody and corresponding IL-6 variants were treated in a range of concentrations and incubated for 1 h at room temperature. The subsequent steps from antibody treatment to absorbance scanning were identical to those described above for competitive ELISA. The binding affinity was determined through ITC at 25°C (iTC200 system; Microcal). Repebody was prepared in 0.2 mM, and IL-6 in 0.02 mM. 2 lL of repebody was injected twenty times to IL-6 in an isothermal chamber. Integrated peaks were fitted to a one-site binding model, and the dissociation constant was calculated using the Origin program. Anti-D8, anti-Pfs25 and anti-SARS-CoV-2 antibodies whose structures were unavailable were modelled using ABodybuilder [58] and the top-ranked models were chosen for docking to the crystal structure of the vaccinia virus D8 antigen (PDB ID: 4E9O), Pfs25 antigen (PDB ID: 6AZZ) and the RBD domain of SARS-CoV-2 (PDB ID: 6XDG), respectively using either the ClusPro docking server with the Antibody Mode scoring function or the HADDOCK webserver [43] . Residues making up the CDRs were used to guide the docking in the case of ClusPro, while for HADDOCK, the antibody CDR residues and all surface-exposed residues of the respective antigen structures were used to guide the docking. The HADDOCK webserver failed to return results for the anti-D8 antibodies, so only results for the other two are provided here. Epibin was then implemented on the antibody docking models in the same manner as that done for the repebody-IL-6 docked models. The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Glossary Epitope: the region on an antigen recognized by an antibody, here generalized to include the region on any target protein recognized by any target-specific protein binder. Epitope binning: grouping antigen variants into sets sharing common epitopes by which antibodies recognize them, here generalized to the grouping of a set of any protein target variants based on common binding regions by which target-specific protein binders recognize them. Paratope: the region on an antibody recognizing an antigen, here generalized to include the region on any target-specific protein binder recognizing its target Writing -original draft. Dong-Gun Kim: Investigation, Methodology, Validation, Formal analysis, Visualization, Writing -review & editing. Sumin Son: Validation. Yoonjoo Choi: Conceptualization, Supervision, Funding acquisition, Writing -review & editing. Hak-Sung Kim: Conceptualization, Supervision, Funding acquisition, Writing -review & editing. Chris Bailey-Kellogg: Conceptualization, Supervision, Funding acquisition Protein binders and their applications in developmental biology Reflections on the use of protein binders to study protein function in developmental biology Design of a binding scaffold based on variable lymphocyte receptors of jawless vertebrates by module engineering How repertoire data are changing antibody science Next-generation sequencing of antibody display repertoires Sequencing the functional antibody repertoire Deep mining of human antibody repertoires: concepts, methodologies, and applications Advanced methodologies in highthroughput sequencing of immune repertoires Novel approaches to analyze immunoglobulin repertoires Deep sequencing of phage display libraries to support antibody discovery Commonality despite exceptional diversity in the baseline human antibody repertoire Functional interrogation and mining of natively paired human v H: V L antibody repertoires An introduction to epitope mapping Epitope mapping of antibody-antigen interactions with X-ray crystallography Epitope mapping by phage display High-throughput epitope binning of therapeutic monoclonal antibodies: Why you need to bin the fridge Broad epitope coverage of a human in vitro antibody library Highthroughput epitope binning assays on label-free array-based biosensors can yield exquisite epitope discrimination that facilitates the selection of monoclonal antibodies with functional activity Flow cytometry-based epitope binning using competitive binding profiles for the characterization of monoclonal antibodies against cellular and soluble protein targets A novel method of Multiplexed Competitive Antibody Binning for the characterization of monoclonal antibodies Antibodies targeting closely adjacent or minimally overlapping epitopes can displace one another Machine learning-based methods for prediction of linear Bcell epitopes PEASE: Predicting B-cell epitopes utilizing antibody sequence Deep learning methods improve linear B-cell epitope prediction Improving B-cell epitope prediction and its application to global antibody-antigen docking Antibody specific Bcell epitope predictions: leveraging information from antibody-antigen protein complexes Learning context-aware structural representations to predict antigen and antibody binding interfaces Protein interaction interface region prediction by geometric deep learning Epitopes of protein binders are related to the structural flexibility of a target protein surface Ab-Ligity: identifying sequence-dissimilar antibodies that bind to the same epitope Characterizing epitope binding regions of entire antibody panels by combining experimental and computational analysis of antibody: antigen binding competition A high-affinity protein binder that blocks the IL-6/STAT3 signaling pathway effectively suppresses non-small cell lung cancer Protein binder for affinity purification of human immunoglobulin antibodies Computer-guided binding mode identification and affinity improvement of an LRR protein binder without structure determination Tracking protein-protein interaction and localization in living cells using a high-affinity molecular binder Effective suppression of C5a-induced proinflammatory response using anti-human C5a repebody The ClusPro web server for protein-protein docking 1.9 Å crystal structure of interleukin 6: implications for a novel mode of receptor dimerization and signaling Score_set: A CAPRI benchmark for scoring protein complexes Assessment of ab initio models of protein complexes by molecular dynamics Native or non-native protein-protein docking models? Molecular dynamics to the rescue The HADDOCK web server for data-driven biomolecular docking Computationally-driven identification of antibody epitopes Effective inhibition of C3a-mediated pro-inflammatory response by a human C3a-specific protein binder A Chimeric antigen receptor that binds to a conserved site on MICA Using a combined computational-experimental approach to predict antibody-specific B cell epitopes Molecular definition of multiple sites of antibody inhibition of malaria transmissionblocking vaccine antigen Pfs25 Studies in humanized mice and convalescent humans yield a SARS-CoV-2 antibody cocktail Influenza virus-specific human antibody repertoire studies MODELLER A program for protein structure modeling Statistical potential for assessment and prediction of protein structures Force fields for protein simulations Semianalytical treatment of solvation for molecular mechanics and dynamics Comparison of multiple amber force fields and development of improved protein backbone parameters Application of asymmetric statistical potentials to antibody-protein docking seaborn: statistical data visualization ABodyBuilder: Automated antibody structure prediction with data-driven accuracy estimation This work was supported in part by the Global Research Laboratory Program (NRF-2015K1A1A2033346) to Hak-Sung Kim, the Combinatorial Tumor Immunotherapy MRC grant of the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (NRF-2020R1A5A2031185) to Y.C., and NIH R01 2R01GM098977 to CBK.The authors also thank Arij Elfaki (Dartmouth) for her earlier methodology development on which Epibin is based, as well as the Research Computing (Information Technology and Consulting) group at Dartmouth for access to and maintenance of the Discovery cluster. The authors declare no competing interests. All docking models, intermediate files, processed experimental data, and analysis scripts developed for the studies described in this paper have been deposited in a GitHub repository which can be accessed at https://github.com/jmahita/Epibin. The experimental data for the antibody studies was obtained from the supplementary material of the respectively cited publications. Supplementary data to this article can be found online at https://doi.org/10.1016/j.csbj.2022.04.036.