Rapid Structural Characterization of Human Antibody–Antigen Complexes through Experimentally...

doi:10.1016/j.jmb.2009.12.053 J. Mol. Biol. (2010) 396, 1491–1507

Available online at www.sciencedirect.com

Rapid Structural Characterization of HumanAntibody–Antigen Complexes through ExperimentallyValidated Computational Docking

Luca Simonelli1, Martina Beltramello1, Zinaida Yudina1,Annalisa Macagno1, Luigi Calzolai2 and Luca Varani1⁎
1Institute for Research inBiomedicine, Via Vela 6, 6500Bellinzona, Switzerland2Institute for Health andConsumer Protection, JointResearch Centre, TP 203,I-21020 Ispra (VA), Italy
Received 28 July 2009;received in revised form25 November 2009;accepted 28 December 2009Available online4 January 2010

*Corresponding author. E-mail [email protected] used: Ab, antibody

monoclonal antibody; ADE, antibodenhancement; PDB, Protein Data Bacomplementarity-determining regioheteronuclear single quantum coheraccessible surface area.

0022-2836/$ - see front matter © 2009 E

If we understand the structural rules governing antibody (Ab)–antigen (Ag)interactions in a given virus, then we have the molecular basis to attempt todesign and synthesize new epitopes to be used as vaccines or optimize theantibodies themselves for passive immunization. Comparing the binding ofseveral different antibodies to related Ags should also further ourunderstanding of general principles of recognition.To obtain and compare the three-dimensional structure of a large number

of different complexes, however, we need a faster method than traditionalexperimental techniques. While biocomputational docking is fast, its resultsmight not be accurate. Combining experimental validation with computa-tional prediction may be a solution.As a proof of concept, here we isolated amonoclonal Ab from the blood of

a human donor recovered from dengue virus infection, characterized itsimmunological properties, and identified its epitope on domain III ofdengue virus E protein through simple and rapid NMR chemical shiftmapping experiments. We then obtained the three-dimensional structure ofthe Ab/Ag complex by computational docking, using the NMR data todrive and validate the results. In an attempt to represent the multipleconformations available to flexible Ab loops, we docked several differentstarting models and present the result as an ensemble of models equallyagreeing with the experimental data. The Ab was shown to bind a regionaccessible only in part on the viral surface, explaining why it cannoteffectively neutralize the virus.

© 2009 Elsevier Ltd. All rights reserved.

Keywords: docking; antibody–antigen complex; denguevirus;NMRmapping;epitope mapping
Edited by M. Levitt
Introduction

Individuals that survive a viral infection haveantibodies (Abs) capable of detecting and neutraliz-ing subsequent attacks by the same virus. These Absbind antigens (Ags), often viral proteins, through

ess:

; Ag, antigen; mAb,y-dependentnk; CDR,n; HSQC,ence; SASA, solvent-

lsevier Ltd. All rights reserve

specific atomic interactions between the Ab and theregion of the Ag that it recognizes (epitope). A betterunderstanding of these interactions is expected toaccelerate vaccine development, since most currentvaccines are based on the generation of neutralizingAb responses. Recently developed technologyallows us to interrogate the immune response ofhuman donors recovered from a given infection,identify all the Abs against a given Ag, and isolate,produce, and purify milligram quantities of suchAbs.1 We are thus offered the chance to characterizea panel of human Abs and try to understand whichare more effective and why, as we illustrate here as aproof of concept.If we understand the biochemical rules governing

Ab/Ag interactions in a given virus, then we havethe molecular basis to attempt to design and

d.

http://dx.doi.org/10.1016/j.jmb.2009.12.053

1492 Ab/Ag Complex Structure via Computational Docking

synthesize new epitopes to be used as vaccines,optimize the Abs themselves for passive immuni-zation, or design new drugs mimicking the Abs ortheir effect.The best way to study atomic interactions is to

obtain the three-dimensional structure of Ab/Agcomplexes by x-ray crystallography: an often longprocess with high failure rate. However, to under-stand general principles of Ab/Ag recognition, itwould be best to compare the binding of severaldifferent Abs to the same Ag and, if possible,binding of the same Ab to slightly different Ags(mutants or pathogen variants). Indeed, studying asingle complex has further limitations, since it maynot thoroughly represent viruses with significantvariations over time or geographical distance suchas influenza, dengue, or many others.In order to obtain and compare the three-

dimensional structure of a large number of differentcomplexes, it is important to have faster methodsthan traditional experimental techniques. Advancesin algorithms and computing power allow us toapply computational docking2 (the process ofobtaining the structure of a complex between twomolecular components) to the study of Ab/Agcomplexes. While computational docking is fast,the structures it provides might not be as accurate orprecise as experimental ones; however, whensearching for general rules, it is probably better tostudy trends in a large panel of slightly inaccuratestructures rather than a single, highly precisecomplex. This kind of analysis could then suggesthypotheses and direct more focused experimentalefforts on specific Abs of the panel.Several recent studies and the Critical Assessment

of Protein Interactions experiment3 have shown thatcomputational docking can provide relatively accu-rate solutions,4–14 but it often lacks the ability todiscriminate inaccurate results; it is therefore im-portant to experimentally validate the models.We here propose a novel, rapid, experimentally

validated computational approach that can be used

for the structural characterization of a large panel ofdifferent Abs bound to the same Ag. We use anexisting program for docking, RosettaDock,15 butimprove its accuracy with rapidly obtained exper-imental data used both to drive and to validate thecomputational results. In particular, we identifyinterface residues with NMR chemical shift map-ping, which is very well suited to the characteriza-tion of intermolecular interfaces.16 Althoughconditions and exact requirements vary with differ-ent samples, NMR data of sufficient quality canusually be obtained in less than 24 h with 300 μl of a100-μM sample; isotopically labeled material isrequired for NMR, but we show in this article thatinexpensive nitrogen labeling is enough for satisfac-tory results. Interpretation of the NMR mappingresults is usually achieved in less than a couple ofdays, provided that NMR assignments are available.Obtaining NMR assignments requires anythingbetween 2 weeks and 2 years according to themolecule of interest, size, sample behavior, and soforth. Once assignments of a given Ag are available,though, the binding footprint of any Ab can berapidly obtained.As a proof of concept, we apply the approach to

the complex between a humanmonoclonal antibody(mAb) and a fragment of the surface protein ofdengue virus (DENV), as schematically illustrated inFig. 1. DENV is a flavivirus17,18 responsible for ∼100million annual human cases,19 including 500,000hospitalizations and 20,000 deaths with an economicburden rivaling that of malaria. Although DENVhas been mainly restricted to the tropical region,both its epidemic activity and its geographicexpansion are increasing as travel, urbanization,and climate changes create favorable conditions forvector and virus dissemination,20 with an estimated2.5 billion people at risk of infection.No cure or vaccine for DENV is currently

available; the effort to find one has been hamperedby the presence of four different dengue serotypes(DENV1–4) and by a poorly understood process

Fig. 1. Schematic workflow forAb characterization. Characterizinga large panel ofAb/Ag complexes isexpected to further our understand-ing of their interaction and acceler-ate vaccine development. As shownhere as a proof of concept, humanmAbs are isolated from donorsinfected with a given virus, theirneutralization efficacy is assessed,and they are structurally and bio-chemically characterized. The infor-mation is used to drive and validatecomputational docking. In an itera-tive process, computational modelscan suggest hypothesis to be testedby new experiments.

1493Ab/Ag Complex Structure via Computational Docking

unique in human medicine, antibody-dependentenhancement (ADE). Abs raised against a previousdengue infection facilitate subsequent infection by adifferent serotype and lead to dengue hemorrhagicfever,21–23 a more severe form of the disease withfatality rates that can surpass 20% but fall below 1%with appropriate supportive care. This featurecomplicates the task of finding a vaccine, since avaccine that would not protect equally against allfour serotypes would actually contribute to theemergence of dengue hemorrhagic fever.E protein is the principal component of the

external surface of the DENV virion and is adominant target of the response consisting of Absagainst DENV. Crystal structures of E protein24–26

have shown it to be a homodimer, with eachmonomer composed of three domains: DI, DII, andDIII, to which several neutralizing Abs bind.25,27

Large domain movements of E protein arenecessary for successful fusion of the viral mem-brane to the host cellular membrane, leading toinfection.28–30 In West Nile virus, which is closelyrelated to DENV, this movement can be blocked byan Ab binding to DIII.30

A structural comparison of several different Ab/DIII complexes could provide insightful clues on theAb features needed for DENV neutralization andenhancement. Ideally, we would like to observe theinteraction of several Abs with each of the fourserotypes and would have to determine severalstructures. The task, long and difficult with tradi-tionally experimental techniques, is ideally suited tocomputational studies.

Results and Discussion

As proof of concept, we apply a rapid, experi-mentally validated computational approach toobtain the structure of the complex between DIIIof DENV4 and the human mAb DV32.6, isolatedfrom a donor recovered from dengue infection.In extreme synthesis, computational docking starts

with the structures of the individual components andmust face two problems: (1) finding the correctsolution, achieved by repeating the simulationthousands of times (technically, it is a matter offinding a global energy minimum), and (2) discrim-inating the correct solution from the thousands ofwrong ones by use of a so-called “scoring function”.The structure ofDENV4DIII is available31 and theAbcan be reliably predicted by homologymodeling.32,33

All that is needed is the amino acid sequence of theAg-binding region, which we can easily obtain fromthe isolated mAb. Great progress has been made infinding the correct solution whereas there is still a lotof room for improving the scoring functions.

Computational docking can achieve precise andaccurate results

Although several papers have proven its validity,computational docking is a relatively new and

constantly evolving technique; hence, it is worthtesting its accuracy on our system by reproducing arelated, known experimental structure34 [complexbetween 1A1D-2 Ab and DENV2 DIII; Protein DataBank (PDB) code: 2R29].Three different starting structures can be consid-

ered for the Ag: the bound conformation, that is, thestructure of DIII in the crystallographic complex; theunbound one, that is, DIII from the x-ray structure ofthe free protein (PDB code: 1TG826); and a homol-ogy model of DIII (based on DENV4 DIII), which weobtained from the I-TASSERweb server.35 Similarly,the starting Ab conformation can be taken directlyfrom the x-ray structure of the complex (bound) orfrom homology models.Good agreement between the docking and crystal

structure, in terms of both spatial similarity andintermolecular contacts, is obtained when using abound Ab conformation with any Ag conformation(Fig. 2). RMSD of the six complementarity-deter-mining region (CDR) loops to the crystal structureis 1.4 Å when using a bound DIII, 2.6 Å forunbound, and 1.8 Å for the homology model, arather impressive result (details in the Supplemen-tary Material).The above scenario does not reflect a blind

prediction, where the bound structure of the Ab isnot available. Thus, we predicted the structure of thevariable domain of 1A1D with the web serversPIGS36 and RosettaAntibody33 and docked eachmodel with DENV2 DIII. Ab loops can be reliablypredicted but the main weakness lies in long CDRloops, particularly the H3 loop, which is poorlyconserved and does not follow canonical structuralrules.32,37–39 Models with different structures can begenerated and appear equally reliable when evalu-ated in the context of the free Ab. In order to test theeffect of different starting structures on docking, wethus generated 11models of the Ab, mainly differingin the position of the H3 and L3 loops. We thendocked each of them independently with the boundand unbound Ag structure. The model generatedwith PIGS yields docking results spatially close tothe x-ray structure of the complex; RMSD of the sixCDR loops is 2.1 Å when using bound DIII, 2.7 Å forunbound, and 2.8 Å when docking the homologymodel of DIII. All structures, however, havecontacts between the Ag and Ab residues outsidethe canonical loops. The results of the 10 modelsgenerated with RosettaAntibody vary greatly: whendocking a bound DIII, the CDR loop RMSD to the x-ray complex is between 2.8 and 7.1 Å. It variesbetween 2.4 and 6.3 Å when docking unbound DIII,instead. Curiously, although the Rosetta5 model isthe most similar to the x-ray in both the bound andthe unbound docking test, there is no correlationbetween the other models in the two cases. Themodel Rosetta2, for instance, has the second bestRMSD when docked to unbound DIII (3.1 Å) but arelatively poor 6.3-Å RMSD when docked to boundDIII. Users might be tempted to consider only thefirst model (Rosetta1) offered by the RosettaAnti-body server, yet in this case, with RMSD of 5.0 and

Fig. 2. Docking a known experimental structure. (a) The surface of DIII is shown in blue while the Ag-binding loops of the Ab are colored as follows (see the text for details):green, x-ray structure; yellow, docking result, replacing the structure of the Ab with its homology model; purple, decoy with good score but low accuracy. (b) Docking result of arandomized global search. Each dot in the plot represents a docking decoy. The y-axis shows the Rosetta score so that better decoys are at the bottom; the x-axis shows the RMSD(in angstroms) to the x-ray structure so that more accurate results are to the left. R1 is the most accurate decoy, yellow in (a); R2 is a decoy with similar score but low accuracy,purple in (a). The scoring function alone cannot differentiate between the two. (c) Interface residues, colored as above and defined as Ag residues within 5 Å of the Ab, are mappedon the DIII sequence. There is good agreement in the contacts between the x-ray structure (green) and the accurate model, whereas the wrongmodel (purple) can be discriminated.

1494Ab/A

gCom

plexStructure

viaCom

putationalDocking


5.7 Å, it provides much poorer results than othermodels.Although it is difficult to draw general conclu-

sions from a single test case, precise and accurateresults can, in principle, be obtained when dockingAbs to DIII of dengue E protein. Equally goodresults are achieved when using bound, unbound,or homology-modeled Ag structure. The choice ofstarting Ab model, instead, may have a critical effecton docking.Generally speaking, Abs with non-standard CDR

loops would be more difficult to model and mightyield worst docking results; however, it must bepointed out that more and more experimental Abstructures will be available to use as reference andthat there are constant improvements in the algo-rithms used to model protein loops.

Docking is right but we cannot know it

In the previous test case, the docking decoys werechosen according to their similarity with the knowncrystal structure. In a blind prediction, however, thetarget complex is not known and the result must bechosen according to its computational score.In such general cases, the orientation of the Ab

can be fixed, since we know that it interactsthrough the CDR loops, but the starting positionof the Ag must be randomized so that every part ofits surface has the possibility to interact with the Abduring the docking process. The result of such arandomized search between the 1A1D modelgenerated with PIGS and the bound DENV2 DIIIis summarized by the plot in Fig. 2b. One decoy(R1) has a good score and is similar to the crystalstructure (RMSD of the CDR loops, 2.1 Å); howev-er, other decoys have just as good scores but are notaccurate, as shown by the high RMSD to theexperimental structure. Clearly, the scoring func-tion by itself cannot discriminate the right solution(Model R1) from the wrong ones with the samescore (R2 and others).When comparing the Ag residues in contact with

the Ab (Fig. 2c), we can find good agreementbetween the correct model and x-ray structures,while it is not so for the inaccurate model, suggest-ing that experimental information capable of iden-tifying the contact residues should help todiscriminate accurate solutions. Finally, all thecomputational models seem to have slightly morecontacts than the x-ray structure, suggesting that thescoring functions favor interfaces with larger areas.Even so, it is promising for our intended use(studying Ab/Ag interactions) that even modelswith relatively high RMSD to the correct structurecan identify the binding region, if not specificcontacts, rather accurately.Of course, this is just an example, and several

studies, as well as the Critical Assessment of ProteinInteractions experiment, have shown how scoringfunctions can, at times, identify accurate resultsamong different models.40,41 Nonetheless, the inad-equacy of the scoring function is a big limitation that

we overcome using rapid experimental results todrive and validate the docking.

NMR epitope mapping improves the accuracy ofcomputational docking

The previous example demonstrates how compu-tational docking by itself may not yet provide anaccurate result and highlights the need for improve-ment. One obvious solution is to provide betterscoring functions, which remains one of the long-term goals of computational docking. A betterapproach, at least for now, might be to rule outinaccurate models due to disagreement with bio-chemical information, which should ideally becollected rapidly so as to maintain one of the mainadvantages of computational biology: speed.Solution NMR chemical shift mapping is rapid

and extremely powerful at identifying interfaceresidues,42 and here we show that it is a valid toolto improve the accuracy of computational docking.It was used twice to study Ab/Ag interactions,43,44but here, for the first time, we apply it to thecharacterization of a full-sized humanmAb, DV32.6,in complex with its Ag, DIII from DENV4 E protein.Briefly, the NMR signal is exquisitely sensitive to

the local chemical environment; when an Ab bindsto a protein, the chemical environment of residues atthe binding interface changes, as does signal arisingfrom those residues. By comparing the NMRspectrum of DIII either free or in complex withDV32.6 (Fig. 3a), we are able to identify whichresidues are affected by Ab binding. One limitationis that chemical shift mapping cannot distinguishwhether a residue is at the interface or if itsenvironment changes due to allosteric or indirecteffects; nonetheless, analysis of the mapping resultin the context of the three-dimensional structure isusually enough to differentiate the two situations.In a typical experiment, the DIII Ag is isotopically

labeled while the full Ab remains unlabeled. Ifneeded, the isolated human mAb can be enzymat-ically cut to yield a smaller (50 kDa) Ag-bindingfragment, but we have satisfactory results even withthe full, uncut Ab despite its large molecular mass(150 kDa). Once the NMR assignments (knowingwhich protein residue produces which NMR signal)are available,31 chemical shift mapping results forany Ab complex can be obtained in a matter of hoursthrough simple and sensitive heteronuclear singlequantum coherence (HSQC) experiments. Moresophisticated techniques, such as cross-relaxationor deuterium exchange experiments,44–47 are avail-able to reach similar results, but direct comparisonof 15N-HSQCs is the simplest and most economicalway to obtain information on intermolecular inter-faces at the residue level.Twenty-six Ag residues show significant chemical

shift changes upon formation of the complexbetween DIII of DENV4 and DV32.6; structuralconsiderations suggest that another residue (Q316)should be part of the epitope but no clear NMR datacan be included due to spectral overlap. All the

Fig. 3. NMR epitope mapping of DV32.6 on dengue DIII. (a) Superposition of 15N-HSQC of DENV4 DIII either free (blue) or in complex with Ab DV32.6 (red). The Ab isunlabeled and not visible in the spectrum. Some of the residues showing chemical shift changes are indicated. (b) Schematic representation of the NMR sample: the unlabeled full-length Ab was mixed in a 1:2 ratio with 15N-labeled DIII; each Ab molecule binds two DIII molecules (175 kDa). Residues showing significant (red) and minor (purple) chemicalshift changes define the epitope on the blue surface of DIII. (c) The epitope residues are mapped on the DIII sequence of the four dengue serotypes and colored according to aminoacid properties. Some of them are conserved, explaining why the Ab binds to all serotypes, and some are not, accounting for the differences in neutralization.

1496Ab/A

gCom

plexStructure

viaCom

putationalDocking


residues whose resonances are perturbed by com-plex formation are indicated in Figs. 4c and 5c. OurNMR results are sufficient to identify the epitope butthey do not give any information on the Ab residuesrequired to achieve binding: an answer that com-putational docking can provide.NMRmapping can have three roles in driving and

validating the docking process.

(1) Once the region of interaction has beendefined, the two molecules can be manuallyput in the correct orientation before startingthe docking, thus limiting the necessity toexplore a large space in a randomized globalsearch, saving computational time andresources. Since the docking procedureexplores a relative large area around thestarting position even in a “local search”, verycareful initial positioning of the dockingpartners is not required, nor would it beeasily achievable.

(2) Interface residues can be constrained withinthe docking protocol so that a bonus score isassigned when they end up at the interfaceand a penalty when they do not. This isalready done for Abs,48 where such penaltiesare assigned outside Ag-binding loops. Theapproach has pitfalls and must be used withcaution, however. Imposing tight constraintsor giving too much weight to selectedresidues might yield many similar models,increasing the apparent precision but actuallydecreasing their accuracy. In the currentwork, we preferred to use penalties for theAb, but no bonus or penalties for the Agresidues predicted to be at the interface.

(3) NMR mapping results can be used to dis-criminate correct docking models from inac-curate ones through visual inspection of thesequence and three-dimensional structure.Typically, the best scoring decoys after alocal search are analyzed and all those withgood agreement with the experimental dataare subsequently refined. In our experience,almost all decoys selected after the initialsearch can generate a scoring funnel afterrefinement. Furthermore, the ranking of thedecoys can be changed by refinement, in thesense that refining a decoy might provide avery good fit with the NMR data even if itwas not the case before. At the same time,decoys that are clearly far away from theNMR data will not provide accurate solutionsafter refinement. It is thus often important torefine more than one decoy, but not necessaryto do so for every one with a good score. Atthe end of the process, the decoy that agreesthe most with the experimental data is chosenas final model. In the absence of experimentalinformation, the choice is dictated by thescore itself or by other considerations such asthe number of decoys populating clusters of

similar structures. As an example, Model S1in Fig. 4 (yellow) has the best score and wouldhave to be preferred, yet it clearly does notagree with the NMR data as well as Model S2(green), which is instead chosen for refine-ment. The differences between the decoyswould be even more significant if we hadconducted a randomized search of the wholeAg instead of a local search around the NMR-derived epitope. The example clearly showshow NMR epitope mapping increases theaccuracy of computational docking by over-coming the limitations of scoring functions.

Docking validation

It is easy to decide if a decoy fits the experimentaldata in a case like the one illustrated in Fig. 4, wherevisual inspection (Fig. 4b) or analysis of the interfaceresidues (Fig. 4c) is obvious. But is there anyindicator allowing discrimination of inaccuratemodels at a finer level? We propose two possibili-ties, although any criterion for accepting or rejectinga model is somehow subjective, and in our view, itwould be a mistake to apply it too rigorously; visualconsiderations and analysis of the resulting struc-ture should always have a part.If a model is correct, all interface residues should

show NMR chemical shift changes upon complexformation; the opposite is not true: residues withNMR changes might not be at the interface becausethe shift can arise from indirect effects. Takentogether, we may reject models with interfaceresidues that do not show chemical shift changesbut we cannot impose that all the NMR-mappedresidues are in contact with the Ab. The definition of“interface” is somehow subjective, but we consideran Ag residue in contact if its NH proton is within7 Å of any Ab atom. The value is somehowsubjective but we chose it taking into considerationthe following: the resolution of the x-ray or dockedstructure; consistency with our test cases, the notionthat NMR dipolar effects tend to disappear aftersuch distance and that we can only observe NHmoieties in 15N-HSQCs: it is possible that a residuewith side-chain atoms at the interface does notexperience any change at its backbone NH uponcomplex formation.We tested this approach on a TCR/pMHC

complex in which NMR mapping was later provenaccurate by an x-ray structure.16 Similarly to Abs,TCRs recognize pMHCs through six variable loops,and the quality of the NMR data is comparable tothat of our Ab/Ag experiments. After separating thex-ray components by 25 Å, we docked the TCR/pMHC complex and evaluated the available x-raystructure as well as the resulting decoys. In the x-raystructure, there are 17 NMR contacts, that is, 17 NHgroups in assignedMHC surface residues within 7 Åof the TCR: they are all defined as valid contactsbecause they all show chemical shift changes in theNMR mapping experiment, a part from one residuethat could not be properly assessed due to spectral

Fig. 4. NMR validation of computational docking. (a) Docking result of a local search for DV32.6 on DENV4 DIII. The x-axis shows the RMSD to an arbitrary structure; the restis as described for Fig. 1b. Model S1 (yellow) has the best score but does not agree with the experimental data as well as Model S2 (green). (b) DIII is represented as in Fig. 3; the Abis shown as cartoon. (c) The top line shows the NMR epitope on the sequence of DIII; residues showing significant chemical shift changes are in red; residues with small changesare in purple. Interface residues are mapped on the two bottom lines, colored in green and yellow as above. The accepted model (green) has good agreement between the dockinginterface and NMR data while the rejected model does not.

1498Ab/A

gCom

plexStructure

viaCom

putationalDocking

Fig.

5(legend

onnext

page)



overlap. As expected, some residues have NMRchanges but appear relatively far away from theinterface in the x-ray structure. In the most accuratedecoy (RMSD of 3.1 Å to the x-ray structure for theCDR loops), there are 16 valid NMR contacts and 2“NMR violations”, that is, residues within 7 Å in thedocking decoy but far away in the x-ray. As anexample, one inaccurate decoy (RMSD of 8.8 Å forthe CDR loops) has 18 valid NMR contacts but also 7NMR violations; another one (RMSD 22.1 Å) hasonly 12 valid contacts and 6 violations. In summary,the “NMR violation” criterion seems capable ofdiscriminating inaccurate TCR/pMHC models.In a complementary approach, we calculate the

solvent-accessible surface area (SASA) of eachresidue both in the structure of the isolated pMHCand in the complex with the TCR, defining a “SASAcontact” if there is a difference of at least 10%between the two cases. The 10% threshold isarbitrary, but it provides the best agreementbetween the SASA contacts in the TCR/pMHC x-ray structure and the NMR mapping. A fewdifferences between the SASA and NMR mappingare expected, since we calculate the NMR criteriononly on the backbone NH while the SASA criterionaccounts for the side chains, which might be closerto the interface than the NH is. Nonetheless, it isintriguing that the few x-ray residues with SASAcontacts but no NMR changes are marred byspectral overlap, suggesting that they might actuallybe shifting but are not clearly detected in themapping experiments. Similarly as before, if wedefine a “SASA violation” as a residue with a SASAcontact in a docked decoy but not in the x-raystructure, we have a reasonable ranking of thedecoys: the docked decoy most similar to the x-rayhas 18 SASA contacts and only 1 SASA violation,while the decoy with 8.8 Å RMSD has 14 valid SASAcontacts but also 4 violations. There are even moreviolations in less accurate decoys. Just like the NMRcriterion, the SASA criterion is capable of discrim-inating inaccurate models and we suggest that acombination of the two, in addition to visualconsiderations, might be used to provide experi-mental validation of docking results.Evaluating the DENV2 DIII/1A1D-2 models with

the SASA criterion is straightforward, since we candefine the correct SASA contacts with the availablex-ray structure. Not having NMR data, however, itis difficult to test the NMR criterion: the position of

Fig. 5. Docking multiple models of the Ab. (a) Superpositioepitope are colored as in previous figures; the Abs are shown agreen, respectively. (b) Ribbon representation of the Ab CDshowing the variability in the different models. Not drawn tviolations are in green, while the others are in gray. (c) Top lResidues with significant chemical shift changes are in red, whwhich no NMR mapping information is available, either becBottom lines: interface residues are highlighted for each modeSASA criteria (see the text) are in green. If they are at the interfare shown in dark green. If they are at the interface according to1, 2, 5, 7, and 8 have no violations.

the protein residues is known only within the limitsof the x-ray resolution (3.1 Å for this test case) andwe do not know exactly at which distance from theAb we would expect Ag signals to change uponcomplex formation. Nonetheless, we tried to deducea virtual NMR chemical shift footprint for Ab 1A1D-2, defined as any Ag residue whose NH is within 7 Åof the Ab in the x-ray. For each docked model, wethen consider an NMR contact any DENV2 residuewith NHwithin 7 Å of the Ab: if the residue is part ofthe virtual footprint, we count it as a valid NMRcontact; otherwise, we count it as an NMR violation.The DENV2 DIII/1A1D-2 models whose CDR

loops have RMSD to the x-ray structure of less than2 Å have a large number of valid SASA contacts (19)and no violations (0); they also have a large numberof valid NMR contacts (19 and 21) with only 1 NMRviolation. All docked complexes starting with an Abmodel generated with PIGS have low RMSD (lessthan 3 Å), a large number of valid SASA and NMRcontacts (between 17 and 20), but also several SASA(4) and NMR violations (1 to 3). Whereas theposition of the loops is similar to the x-ray structure,visual analysis reveals that other Ab residues getclose to the Ag: a configuration not likely to becorrect. The docked complex between unbound DIIIand the Ab model Rosetta8 is not particularlyaccurate (CDR loop RMSD of 5.2 Å to the x-ray)but has no SASA violations and only two NMRviolations. The number of valid contacts is ratherlow, however (13 SASA and 11 NMR), and visualinspection easily reveals that only part of the epitopeis covered by the Ab in the docked complex.Generally speaking, the more accurate docked

complexes have a large number of valid contactsand few, if any, violations. Decoys not following thistrend usually have problems readily spotted byvisual inspection of the three-dimensional structureand biological considerations. Although any decisionwill have to rely on subjective judgment, assessmentof the NMR and SASA criteria as well as structuralconsiderations appears to be a valid approach for thevalidation of docking results.

Docking multiple models to simulate loopflexibility

Proteins often undergo significant conformationalchanges upon complex formation,49 particularly ininherently flexible loop regions. Docking programs

n of all DV32.6 models showing no violation; DIII and thes cartoon. The heavy and light chains are in dark and lightR loops, indicated according to standard nomenclature,o scale with (a). All 11 models are shown; those withoutine: the NMR epitope is mapped on the sequence of DIII.ile residues with small changes are in purple. Residues forause of overlap or because of lack of signal, are in gray.l. Residues at the interface according to both the NMR andace according to the NMR but not the SASA criterion, theythe SASA but not NMR criterion, they are in cyan. Models


usually take side-chain flexibility into considerationbut modeling backbone conformational changes isdifficult, due to the complexity and numbers ofdegrees of freedom of the backbone conformationalspace.47 Different strategies have been proposedbut the problem remains one of the biggestlimitations of computational docking (for reviews,see Refs. 47 and 50).The issue is particularly relevant for Abs, since

their whole interface is composed by Ag-bindingloops, which may experience different conforma-tions. Furthermore, while it is straightforward topredict the conserved framework region by homol-ogy modeling, assembling the heavy and light chainin the correct orientation and predicting the struc-ture of the Ag-binding loops are less obvious.36,51

Two factors help us in this regard: (1) most CDRloops have access to only a limited number of clearlydefined canonical structural classes;52 (2) we oftenknow the structure of Abs in their bound confor-mation; hence, we can use that as a reference forbound homology models. Nonetheless, the mainweakness in the modeling lies in long CDR loops,particularly the H3 loop that does not appear to fitinto canonical structural classes.Here, we suggest a slightly different approach to

the docking of potentially flexible or inaccuratelymodeled Ab loops: after having obtained thesequence of the variable region for both the heavyand light chain, we used both the PIGS server36 andRosettaAntibody Modeling server33,51,53 to generate11 different models of the human mAb DV32.6. Themain differences involve the position of H3 and L3,whereas the other loops are almost identical. Wethen docked and evaluated each model indepen-dently (Fig. 5). This allows exploring the effect ofdifferent loop conformations and, on a morepractical note, limits the chances of starting with awrong model. If only one Ab model is chosen forsubsequent docking, in fact, it has to be decidedwhich one is best, but we have no experimentalinformation to guide us. Conversely, dockingmultiple conformations, either as an ensemble orserially, represents a particular challenge for thescoring function, because backbone changes affectnot only inter- but also intra-molecular energies andmake a direct score comparison problematic. Ourapproach of repeating the calculations with 11different models has the obvious disadvantage ofincreasing the workload but has the strong advan-tage of allowing experimental validation of eachdifferent Ab homology model in the context of thecomplex conformation.Not all the different homology models of DV32.6

yield a final docking result, after refinement, equallycompatible with the experimental evidence (Fig. 5c).According to the NMR and SASA criteria describedearlier, Models 1, 2, 5, 7, and 8 show no violation andtheir ensemble is accepted as final result. Model8 has more contacts in agreement with the NMR andSASA criteria than other models. Although it isusual to indicate a single structure as final dockingresult, we believe that it is more accurate, albeit less

precise, to include all the models with no violation,since they equally satisfy the experimental evidence.Model 6 has no SASA violation and 1 NMR

violation, but simply changing the distance cutofffrom 6 to 5.5 Å, a small difference given theuncertainty of docking, would result in no violation.However, it can be discarded because it has contactsbetween the Ag and Ab residues well outside theAg-binding loops. Finally, Models 3, 4, 9, and 10 andthe one generated with PIGS have 2 to 11 violationsand may be rejected. We were not able to discrim-inate accurate from inaccurate models by scoringfunction, by the presence or absence of a funnel inthe plot of the docking result,40 or by using theFunHunt server.41

In summary, different Ab loop conformationsapparently yield equally accurate results; it remainsto be seen whether this is a relevant biologicalfeature or a simple shortcoming of the computa-tional docking that cannot discriminate wrongconformers. The idea that Abs might adoptdifferent conformations to interact with the sameAg or to adapt to slightly different Ags is, however,intriguing.

Amix of flexibility and specificity at the interface?

The Ab models mainly differ in the 13-residue-long H3 loop (average RMSD to Model 8 is 2.6 Å); afeature shared by the models with no violations isthat H3 extends out towards the Ag (SupplementaryFig. 4), although this is less true for Model 8.Conversely, the loop is bent in Model 2, presenting arather flat surface to the Ag. All mentioned modelshave H3 residues in proximity of K310 and A313 onDIII, in addition to other residues different for eachmodel, but no common specific interaction can bedetected (the contacted H3 residues vary betweenW101, S103, S105, and T106).By contrast, Models 1, 5, 7, and 8 show specific

interactions and possible hydrogen bonding be-tween DIII K310 and K323 and D50 and D53 on theAb L2 loop (Fig. 6): the side-chain oxygens of D50appear to bridge the NH3 groups of the two Aglysines, right at the core of the epitope. Models 1, 7,and 8 but not Model 5 show further H-bondpossibilities between K310 and S32 on the L1loop, while Models 5, 7, 8, and 2 show additionalH-bond between D53 side chain and K323. Curi-ously, D50 interacts with the two lysines even inModel 2, but using the backbone carbonyl withK323 instead of the side chain. In support of the keyrole played by the two lysines, mutating either ofthem abolishes Ab binding (M. Diamond, personalcommunication).Care must be taken when discussing specific

atomic interactions with the low resolution ofcomputational docking, but it is tantalizing tothink that sterically constrained parts of the interfacemight provide specific and conserved contacts at theprice of significant entropic costs. At the same time,more flexible regions such as H3 might use multipleconformations, providing less specific interactions,

Fig. 6. Hydrogen bonds network at the Ab/Ag interface. Hydrogen-bond interactions at the core of the Ab/Aginterface. Mutating any of the two indicated DIII lysines abolishes binding. Nitrogen and oxygen atoms are shown in blueand red, respectively, whereas hydrogens are in white. Possible hydrogen bonds are indicated by broken lines. DIII isshown in light blue, while the Ab light chain is in green. Model 8 is shown.


contributing less to the overall binding energy butoffering the advantage of lower entropic costs.Indeed, similar observations have been reportedfor RNA–protein and protein–protein interfaces.54

The issue is worth investigating by other meanssuch as molecular dynamics or direct NMR dynamicstudies.

Binding of DV32.6 to DENV4 E protein domain III

DV32.6 is a human mAb isolated from a donorrecovered from a dengue infection. It binds thewhole virus, the full E protein dimer and DIII alonein all four dengue serotypes; it does not effectivelyneutralize DENV4 and weakly neutralizes the otherserotypes. It also shows significant ADE (Fig. 7a).Here, we analyzed its complex with DIII because, inaddition to being biologically relevant, its limitedsize (112 residues) makes it particularly suitable toNMR structural studies and computer simulations.The Ab/Ag interface has a surface of approxi-

mately 650 Å2, with roughly equal amounts of polar,nonpolar, and charged amino acids. Electrostaticinteractions play a major role, with the acceptedmodels showing between 22 and 37 possiblehydrogen bonds and salt bridges. The epitopecomprises residues absolutely conserved, explainingwhy DV32.6 is capable of binding to all serotypes,and residues that are not, thus accounting for thedifferences in neutralization among serotypes. Wecannot understand which residues cause the differ-ences, however, simply by looking at the primarysequence of the epitope.

Whenanalyzing the epitope in the context of the fullE protein dimer, it appears evident that it is partlyinaccessible on the dimer and, therefore, on theknown structure of the virus (Fig. 7b). We tried torepeat the dockingusing the full E protein but failed toobtain any satisfactory solution, either by score or byagreementwith theNMRepitope still accessible in thefull protein; this, although not conclusive, suggeststhat DV32.6 does not have an epitope touching bothDIII and the rest of E protein. Furthermore, if DV32.6had been raised from injection of DIII into a mouse,we could think of this as an artifact, since DIII by itselfexposes regions not available on the full virus, but ourAb has been isolated from a human donor exposedto the virus and does, indeed, bind to the full virus inELISA experiments. Epitopes, instead, may becomeavailable or unavailable on the viral surface due toconformational dynamics of the virion as was shownfor another dengue Ab.34

It has been suggested that ADE is a result of loweroccupancy55 but that Abs binding to partiallyoccluded epitopes might promote surface rearrange-ment capable of blocking the virus.34 The issue is notstraightforward, however: DV32.6 is a very poorneutralizer of serotype 4, yet, since it also binds to thefull E protein dimer, it should elicit some conforma-tional rearrangements to open up the inaccessibleregion of DIII. It is possible that binding to DENV4might not free enough energy to drive a conforma-tional change on the full viral surface, while bindingto other serotypes could.The number of Abs bound to the full virion at any

given concentration depends on the binding strength

Fig. 7. Binding to a partially inaccessible epitope. DV32.6, which does not effectively neutralize DENV4, binds to an epitope partially inaccessible on the viral surface;furthermore, not all the binding sites are equally available on the full viral surface due to steric hindrance. (a) Neutralization assay: empty circles indicate the amount of infectedVERO cells at different Ab concentration. Cells are infected even at high Ab concentration. Enhancing assay: filled circles indicate the amount of infected K562 cells. At low Abconcentration, no cells are infected; at increasing Ab concentrations, cells are infected through ADE. (b) Surface representation of the full E protein dimer, in gray. DIII is in blue,with the NMR-derived epitope in red and purple. Only part of the epitope is accessible on the surface (compare to previous figures). Part of the E dimer must be displaced to allowDV32.6 (green cartoon) binding. (c) DV32.6 variable domain (green) is overlaid on the viral surface, highlighting possible steric clashes among individual units.

1503Ab/A

gCom

plexStructure

viaCom

putationalDocking


and on/off rates. These cannot be inferred fromdocking, but superimposing the bound DV32.6 onthe known structure of the virus (Fig. 7c) gives anidea of the steric hindrance that it might be subjectedto. It appears that the Abmight not be able to bind toall the available epitopes without clashes and thatmaximum occupancy might be limited to 120 sites ofthe 180 surface sites.17 This number is ratherspeculative, given the limited resolution of bothour model and the viral structure, but it might proveuseful if taken not as an absolute but in relation tosimilar analysis for other Abs.

Perspective

Abs play a pivotal role in the immune system andare very attractive targets for both pharmaceuticalcompanies and basic research, yet we have ratherlimited structural knowledge of the way theyrecognize Ags. Given a specific Ab, we cannotpredict what it recognizes or its binding strength orefficiency. We do not yet have the ability to designex novo an Ab with desired therapeutic propertiesor an Ag capable of eliciting a specific Ab response(vaccine).Comparative studies of different Ab/Ag com-

plexes might eventually teach us what the maindeterminants of structural recognition are. Search-ing for trends and general consensus in thehundreds of currently available structures is notsimple, however, partly because the Ags are mostlyunrelated. It would probably be easier to concen-trate on a set of different Abs bound to the same Ag,include in the analysis both Abs that are effectiveneutralizers and Abs that are not, and try tounderstand what are the differences among them.While non-neutralizing Abs are usually not thefocus of long and expensive structural studies, theycould actually provide a valid comparison in asystematic analysis.If we want to characterize a large panel of Ab/Ag

complexes, it is important to have access to fasterand cheaper techniques than those traditionallyavailable to structural biology. Computational dock-ing has the potentiality to become such a technique:whereas major technical breakthroughs in estab-lished techniques such as x-ray or NMR are not verylikely, it is undeniable that computational methodswill become more accurate, rapid, and economicalover the years, improving the relevance of acomputational approach to structural interactions.It is debatable whether computational docking willever reach the accuracy and precision of experimen-tal techniques, but when looking for general trends,we might learn more by comparing a set of slightlyinaccurate structures, sacrificing precision for speedand quantity. Representative and promising com-plexes identified by this first analysis could then bethe object of more focused and expensive experi-mental studies.One can envision an integrated approach exploit-

ing the advantages of each structural technique.X-ray crystallography is best at determining the

structure of individual components but sufferswhen dealing with flexible elements or weakaffinity, which might be expected when an Abbinds to “wrong” serotypes and/or is not a goodneutralizer. Obtaining the structure of a largecomplex by solution NMR spectroscopy is long,difficult, and expensive, but NMR can rapidly andreliably identify interface residues through simplechemical shift mapping experiments. Finally, com-puters can take the structures of the individualcomponents and dock them together, using theNMR-derived information to drive and validate theprocess.

Conclusions

As a proof of concept, we isolated a mAb fromthe blood of a human donor recovered fromDENV infection, characterized its immunologicalproperties, and produced and purified it inmilligram quantities. We identified its epitope ondomain III of DENV E protein through simple andrapid NMR chemical shift mapping experiments.Finally, we obtained the three-dimensional struc-ture of the Ab/Ag complex by computationaldocking, using the NMR data to drive and validatethe results. In an attempt to represent the multipleconformations available to flexible Ab loops, wedocked several different starting models andpresent the result as an ensemble of models equallyagreeing with the experimental data. The Ab wasshown to bind a region accessible only in part onthe viral surface.Today, computational docking is capable of

finding accurate solutions among thousands ofgenerated models, but it often lacks the ability todiscern them from inaccurate ones due to deficien-cies in the scoring functions. Experimental valida-tion is, in our opinion, an important step of anycomputational approach, and NMR mapping isideally posed to provide it for docking.

Materials and Methods

Samples production and purification

DIII was expressed in Escherichia coli BL21(DE3) cellswith a pET21a vector (Novagen), induced with IPTG andharvested after 3 h. Uniformly 15N-labeled NMR sampleswere grown in M9 minimal media using 15NH4Cl as solesource of nitrogen. The bacterial pellets were harvested,sonicated, and centrifuged; the unsoluble fraction contain-ing DIII was washed using 1M of urea and 1% of Triton X-100 and then dissolved in 8 M urea and refolded inphosphate buffer by rapid dilution. DIII was purified byFPLC over a cation and size exchange column. Abs fromB-cell culture supernatant were purified by protein-Aaffinity column followed by size-exclusion columnaccording to standard protocols. DIII and DV32.6 werefinally eluted in 10 mMNa–phosphate buffer and 150 mMNaCl, concentrated independently and mixed immediate-ly before the required experiment.


NMR experiments

15N-HSQC experiments were collected on a Bruker 700-MHz spectrometer with cryoprobe. Transverse relaxationoptimized spectroscopy experiments proved to be oflower quality due to lack of sample deuteration andwere not used for data analysis. Acquisition time variedbetween 15 min (free DIII) and several hours (complex). Ina typical experiment, 8 mg of unlabeled full-length Abwasmixed with a 0.25-mM 15N-labeled DIII sample; each Abmolecule binds two DIII molecules (total molecular mass,175 kDa). Assignments were obtained by comparison ofHSQC spectra to previously published data (BiologicalMagnetic Resonance Bank entry 708731).

Chemical shift mapping

NMR epitope mapping starts by identifying the Agresidues showing clear chemical shift differences upon Abbinding. A quantitative interpretation is formally impos-sible without having assignments of the complex, butobtaining them is long and difficult and defeats the ideabehind a rapid approach. However, if a peak of the freeprotein shifts or disappears, then the residue is affected bycomplex formation even if we do not know exactly wherethe peak ends up in the complex. The residues with cleardifferences are thus mapped on the protein structure,defining the epitope; at this stage, it might be possible toadd residues whose change is suspected but vitiated bypoor spectral quality, if they clearly fit within theboundaries of the structural epitope. This process mightintroduce errors, however, and must be exercised withcaution, if at all. The defined epitope is used to guide andprovide a first validation of the docking results. At thevery last stage, if Ab/Ag contacts are detected betweenresidues without chemical shift changes, then the NMRspectrum is checked again to see if those residues showchemical shift changes not previously recognized. In ourexperience, it is very rare to add residues in this final stage.It is worth pointing out that the procedure is similar to theone used in NMR structure determination, where dubiousspectral information is added if warranted by initialstructural calculations.

Ab modeling

Abs were modeled with the programs PIGS36 andRosettaAntibody,33 using as templates Abs not related todengue. With PIGS, we opted to use the same Ab astemplate for both the heavy chain and the light chain, thuslimiting the problem of correctly assembling the twochains together. No such option is available withRosettaAntibody. The PDB codes of the templates for1A1D, used in the test cases, are as follows (sequenceidentity is indicated in parentheses). PIGS: 1J05 for bothheavy (86% identity) and light chain (83%). RosettaAnti-body: 1TY7 for the heavy-chain framework (94%) and1EJO for the light chain (97%); 1EJO for L1 (87%), 1J05 forL2 (86%), and 1EGJ for L3 (75%); 35C8 for H1 (100%), 1TY7for H2 (94%), and 1AXT for H3 (same length, no identity).The PDB codes of the templates for Ab DV32.6 are asfollows. PIGS: 1W72 for both chains (sequence identity90% for the light chain and 76% for the heavy chain).RosettaAntibody: 1W72 for the heavy-chain frameworkregion (94%), 1ADQ for the light chain (98%); 1ADQ for L1(100%); 1W72 for L2 (100%); 1ADQ for L3 (100%); 3CK0for H1 (82%); 1U0Q for H2 (100%).

Docking

Ab/Ag structures were predicted with Rosetta Docking2.3; only the variable domain of the Ab was used fordocking. The starting structures were visually orientedwith the Ab CDR loops facing the NMR-derived epitopeand then separated by 25 Å. Since the initial dockingprocedure explores a relative large area around thestarting position, very careful initial positioning of thedocking partners is not required. Docking runs wereconducted as described previously.7 The Ab position waskept fixed in all cases. In “randomized global searches”,the position of the Ag was randomized and broughttowards the Ab with 5–15–25 perturbations [perturbationalong the line of centers (in angstroms)–perturbation in theplane perpendicular to the line of center (in angstroms)–rotational perturbation (in degrees)]; 3–8–8 perturbationswere used both for local searches and for refinements. Atotal of 10,000–15,000 decoys were generated for eachdocking run. After a local search, the lowest scoringmodels were visually analyzed and the best ones wereselected in accordance to their agreement with the NMRepitope mapping data. Each of these decoys was refinedwith 3–8–8 perturbations, but the Ab and Ag were notseparated as in local searches. For each homology modelof the Ab, the refined decoy in best agreement with theNMR data was chosen as final result.

RMSD calculations

Cα RMSD values, calculated with the programProFit,56 are shown throughout the manuscript. Whencomparing complexes, the Ags were superimposed andRMSD values were obtained for the subset of residuesindicated in the various cases (e.g., H3 loop). The heavyand light chains were superimposed when comparingfree Abs.

Memory B-cell immortalization and cloning

Peripheral blood samples were obtained from a healthyadult donor who had been diagnosed with dengue.Memory B cells were isolated, purified, and immortalizedas described previously.57 B-cell culture supernatantswere screened by ELISA using DIII or full E protein ascoating Ag according to standard techniques.

Neutralization and enhancement assays

DENV–Ab complexes were formed by pre-incubationof DV32.6 with DENV attenuated virus and used to infectVERO (for neutralization) and K562 cells (for enhance-ment). After 3 days of infection, the percentage of infectedcells was determined by fluorescence-activated cell sortingusing standard procedure and staining cells with mousemAb 4G2 (ATCC, D1-4G2-4-15).58

Acknowledgements

This work was supported by generous grants fromSVRI, BSI, and Canton Ticino. The work would nothave been possible without the CSCS Swiss NationalSupercomputer Center providing computationalresources and support, in particular Maria Grazia


Giuffreda and Angelo Mangili, as well as the NMRfacility of NIMR-MRC London, where experimentswere collected. We are indebted to Paolo Marcatili,Anna Tramontano, Aroop Sircar, and Jeff Gray forprecious help with Ab homology modeling. Wewish to thank Gabriele Varani, Silvia Monticelli,Federica Sallusto, Antonio Lanzavecchia, andLuisa Granziero for discussion and critical readingof the manuscript.

Supplementary Data

Supplementary data associated with this articlecan be found, in the online version, at doi:10.1016/j.jmb.2009.12.053

References

1. Lanzavecchia, A., Bernasconi, N., Traggiai, E.,Ruprecht, C. R., Corti, D. & Sallusto, F. (2006).Understanding and making use of human memoryB cells. Immunol. Rev. 211, 303–309.

2. Gray, J. J. (2006). High-resolution protein–proteindocking. Curr. Opin. Struct. Biol. 16, 183–193.

3. Janin, J., Henrick, K., Moult, J., Eyck, L. T., Sternberg,M. J., Vajda, S. et al. (2003). CAPRI: a Critical Assess-ment of PRedicted Interactions. Proteins, 52, 2–9.

4. Carter, P., Lesk, V. I., Islam, S. A. & Sternberg, M. J.(2005). Protein–protein docking using 3D-Dock inrounds 3, 4, and 5 of CAPRI. Proteins, 60, 281–288.

5. Fernandez-Recio, J., Abagyan, R. & Totrov, M. (2005).Improving CAPRI predictions: optimized desolvationfor rigid-body docking. Proteins, 60, 308–313.

6. Wiehe, K., Pierce, B., Mintseris, J., Tong, W. W.,Anderson, R., Chen, R. & Weng, Z. (2005). ZDOCKand RDOCK performance in CAPRI rounds 3, 4, and5. Proteins, 60, 207–213.

7. Daily, M. D., Masica, D., Sivasubramanian, A.,Somarouthu, S. & Gray, J. J. (2005). CAPRI rounds3–5 reveal promising successes and future challengesfor RosettaDock. Proteins, 60, 181–186.

8. Mendez, R., Leplae, R., Lensink, M. F. & Wodak, S. J.(2005). Assessment of CAPRI predictions in rounds3–5 shows progress in docking procedures. Proteins,60, 150–169.

9. Smith, G. R., Fitzjohn, P. W., Page, C. S. & Bates, P. A.(2005). Incorporation of flexibility into rigid-bodydocking: applications in rounds 3–5 of CAPRI.Proteins, 60, 263–268.

10. Camacho, C. J. (2005). Modeling side-chains usingmolecular dynamics improve recognition of bindingregion in CAPRI targets. Proteins, 60, 245–251.

11. van Dijk, A. D., de Vries, S. J., Dominguez, C., Chen,H., Zhou, H. X. & Bonvin, A. M. (2005). Data-drivendocking: HADDOCK's adventures in CAPRI. Proteins,60, 232–238.

12. Inbar, Y., Schneidman-Duhovny, D., Halperin, I.,Oron, A., Nussinov, R. & Wolfson, H. J. (2005).Approaching the CAPRI challenge with an efficientgeometry-based docking. Proteins, 60, 217–223.

13. Schueler-Furman, O., Wang, C. & Baker, D. (2005).Progress in protein–protein docking: atomic resolu-tion predictions in the CAPRI experiment usingRosettaDock with an improved treatment of side-chain flexibility. Proteins, 60, 187–194.

14. Janin, J. (2005). Assessing predictions of protein–protein interaction: the CAPRI experiment. Protein Sci.14, 278–283.

15. Gray, J. J., Moughon, S., Wang, C., Schueler-Furman,O., Kuhlman, B., Rohl, C. A. & Baker, D. (2003).Protein–protein docking with simultaneous optimiza-tion of rigid-body displacement and side-chain con-formations. J. Mol. Biol. 331, 281–299.

16. Varani, L., Bankovich, A. J., Liu, C. W., Colf, L. A.,Jones, L. L., Kranz, D. M. et al. (2007). Solutionmapping of T cell receptor docking footprints onpeptide-MHC. Proc. Natl Acad. Sci. USA, 104,13080–13085.

17. Mukhopadhyay, S., Kuhn, R. J. & Rossmann, M. G.(2005). A structural perspective of the flavivirus lifecycle. Nat. Rev., Microbiol. 3, 13–22.

18. Lindenbach, B. D. (2001). Flaviviridae: the viruses andtheir replication. In Fields Virology (Knipe, D. M. &Howley, P.M., eds), Fields Virology, vol II, pp. 991–1042,Lippincott Williams & Wilkins, Hagerstown, MD.

19. Halstead, S. B. (1988). Pathogenesis of dengue:challenges to molecular biology. Science, 239, 476–481.

20. WHO. (1997). Dengue Haemorrhagic Fever: Diagno-sis, Treatment, Prevention andControlWHO,Geneva,Switzerland.

21. Shu, P. Y. & Huang, J. H. (2004). Current advancesin dengue diagnosis. Clin. Diagn. Lab. Immunol. 11,642–650.

22. Rothman, A. L. (2004). Dengue: defining protectiveversus pathologic immunity. J. Clin. Invest. 113, 946–951.

23. Sabin, A. B. (1952). Research on dengue during WorldWar II. Am. J. Trop. Med. Hyg. 1, 30–50.

24. Modis, Y., Ogata, S., Clements, D. & Harrison, S. C.(2003). A ligand-binding pocket in the dengue virusenvelope glycoprotein. Proc. Natl Acad. Sci. USA, 100,6986–6991.

25. Modis, Y., Ogata, S., Clements, D. & Harrison, S. C.(2005). Variable surface epitopes in the crystal structureof dengue virus type 3 envelope glycoprotein. J. Virol.79, 1223–1231.

26. Zhang, Y., Zhang, W., Ogata, S., Clements, D.,Strauss, J. H., Baker, T. S. et al. (2004). Conformationalchanges of the flavivirus E glycoprotein. Structure, 12,1607–1618.

27. Oliphant, T., Engle, M., Nybakken, G. E., Doane, C.,Johnson, S., Huang, L. et al. (2005). Development of ahumanized monoclonal antibody with therapeuticpotential against West Nile virus. Nat. Med. 11,522–530.

28. Bressanelli, S., Stiasny, K., Allison, S. L., Stura, E. A.,Duquerroy, S., Lescar, J. et al. (2004). Structure of aflavivirus envelope glycoprotein in its low-pH-in-duced membrane fusion conformation. EMBO J. 23,728–738.

29. Modis, Y., Ogata, S., Clements, D. & Harrison, S. C.(2004). Structure of the dengue virus envelope proteinafter membrane fusion. Nature, 427, 313–319.

30. Nybakken, G. E., Oliphant, T., Johnson, S., Burke, S.,Diamond, M. S. & Fremont, D. H. (2005). Structuralbasis ofWestNile virus neutralization by a therapeuticantibody. Nature, 437, 764–769.

31. Volk, D. E., Lee, Y. C., Li, X., Thiviyanathan, V.,Gromowski, G. D., Li, L. et al. (2007). Solutionstructure of the envelope protein domain III ofdengue-4 virus. Virology, 364, 147–154.

32. Morea, V., Lesk, A. M. & Tramontano, A. (2000).Antibody modeling: implications for engineering anddesign. Methods, 20, 267–279.

33. Sircar, A., Kim, E. T. & Gray, J. J. (2009). RosettaAnti-




body: antibody variable region homology modelingserver. Nucleic Acids Res. 37, W474–479.

34. Lok, S. M., Kostyuchenko, V., Nybakken, G. E.,Holdaway, H. A., Battisti, A. J., Sukupolvi-Petty, S.et al. (2008). Binding of a neutralizing antibody todengue virus alters the arrangement of surfaceglycoproteins. Nat. Struct. Mol. Biol. 15, 312–317.

35. Zhang, Y. (2008). I-TASSER server for protein 3Dstructure prediction. BMC Bioinformatics, 9, 40.

36. Marcatili, P., Rosi, A. & Tramontano, A. (2008). PIGS:automatic prediction of antibody structures. Bioinfor-matics, 24, 1953–1954.

37. Decanniere, K., Muyldermans, S. & Wyns, L. (2000).Canonical antigen-binding loop structures in immu-noglobulins: more structures, more canonical classes?J. Mol. Biol. 300, 83–91.

38. Martin, A. C. & Thornton, J. M. (1996). Structuralfamilies in loops of homologous proteins: automaticclassification, modelling and application to antibo-dies. J. Mol. Biol. 263, 800–815.

39. Abhinandan, K. R. &Martin, A. C. (2008). Analysis andimprovements to Kabat and structurally correct num-bering of antibody variable domains.Mol. Immunol. 45,3832–3839.

40. London, N. & Schueler-Furman, O. (2007). Assessingthe energy landscape of CAPRI targets by FunHunt.Proteins, 69, 809–815.

41. London, N. & Schueler-Furman, O. (2008). FunHunt:model selection based on energy landscape character-istics. Biochem. Soc. Trans. 36, 1418–1421.

42. Gao, G., Williams, J. G. & Campbell, S. L. (2004).Protein–protein interaction analysis by nuclear magnet-ic resonance spectroscopy.MethodsMol. Biol. 261, 79–92.

43. Sasakawa, H., Sakata, E., Yamaguchi, Y., Masuda, M.,Mori, T., Kurimoto, E. et al. (2007). Ultra-high fieldNMR studies of antibody binding and site-specificphosphorylation of alpha-synuclein. Biochem. Biophys.Res. Commun. 363, 795–799.

44. Morgan, W. D., Frenkiel, T. A., Lock, M. J., Grainger,M. & Holder, A. A. (2005). Precise epitope mapping ofmalaria parasite inhibitory antibodies by TROSYNMR cross-saturation. Biochemistry, 44, 518–523.

45. Takahashi, H., Nakanishi, T., Kami, K., Arata, Y. &Shimada, I. (2000). A novel NMR method for deter-mining the interfaces of large protein–protein com-plexes. Nat. Struct. Biol. 7, 220–223.

46. Pellecchia, M., Sem, D. S. & Wuthrich, K. (2002). NMRin drug discovery. Nat. Rev., Drug. Discov. 1, 211–219.

47. Bonvin, A. M. (2006). Flexible protein–protein dock-ing. Curr. Opin. Struct. Biol. 16, 194–200.

48. Gray, J. J., Moughon, S. E., Kortemme, T., Schueler-Furman, O., Misura, K. M., Morozov, A. V. & Baker,D. (2003). Protein–protein docking predictions for theCAPRI experiment. Proteins, 52, 118–122.

49. Lo Conte, L., Chothia, C. & Janin, J. (1999). The atomicstructure of protein–protein recognition sites. J. Mol.Biol. 285, 2177–2198.

50. Chaudhury, S. & Gray, J. J. (2008). Conformerselection and induced fit in flexible backbone pro-tein–protein docking using computational and NMRensembles. J. Mol. Biol. 381, 1068–1087.

51. Sivasubramanian, A., Sircar, A., Chaudhury, S. &Gray, J. J. (2009). Toward high-resolution homologymodeling of antibody Fv regions and application toantibody–antigen docking. Proteins, 74, 497–514.

52. Vargas-Madrazo, E., Lara-Ochoa, F. & Almagro, J. C.(1995). Canonical structure repertoire of the antigen-binding site of immunoglobulins suggests stronggeometrical restrictions associated to the mechanismof immune recognition. J. Mol. Biol. 254, 497–504.

53. Narayanan, A., Sellers, B. D. & Jacobson, M. P. (2009).Energy-based analysis and prediction of the orienta-tion between light- and heavy-chain antibody variabledomains. J. Mol. Biol. 388, 941–953.

54. Mittermaier, A., Varani, L., Muhandiram, D. R., Kay,L. E. & Varani, G. (1999). Changes in side-chain andbackbone dynamics identify determinants of specific-ity in RNA recognition by human U1A protein. J. Mol.Biol. 294, 967–979.

55. Pierson, T. C., Xu, Q., Nelson, S., Oliphant, T.,Nybakken, G. E., Fremont, D. H. & Diamond, M. S.(2007). The stoichiometry of antibody-mediated neu-tralization and enhancement of West Nile virusinfection. Cell Host Microbe, 1, 135–145.

56. Martin, A. C. http://www.bioinf.org.uk/software/profit/.

57. Traggiai, E., Becker, S., Subbarao, K., Kolesnikova, L.,Uematsu, Y., Gismondo, M. R. et al. (2004). Anefficient method to make human monoclonal anti-bodies from memory B cells: potent neutralization ofSARS coronavirus. Nat. Med. 10, 871–875.

58. Henchal, E. A., Gentry, M. K., McCown, J. M. &Brandt, W. E. (1982). Dengue virus-specific andflavivirus group determinants identified with mono-clonal antibodies by indirect immunofluorescence.Am. J. Trop. Med. Hyg. 31, 830–836.

Rapid Structural Characterization of Human Antibody–Antigen Complexes through Experimentally...

Documents

Transcript of Rapid Structural Characterization of Human Antibody–Antigen Complexes through Experimentally...