Structural basis of GM-CSF and IL-2 sequestration by the viral decoy receptor GIF

Subversion of the host immune system by viruses is often mediated by molecular decoys that sequester host proteins pivotal to mounting effective immune responses. The widespread mammalian pathogen parapox Orf virus deploys GIF, a member of the poxvirus immune evasion superfamily, to antagonize GM-CSF (granulocyte macrophage colony-stimulating factor) and IL-2 (interleukin-2), two pleiotropic cytokines of the mammalian immune system. However, structural and mechanistic insights into the unprecedented functional duality of GIF have remained elusive. Here we reveal that GIF employs a dimeric binding platform that sequesters two copies of its target cytokines with high affinity and slow dissociation kinetics to yield distinct complexes featuring mutually exclusive interaction footprints. We illustrate how GIF serves as a competitive decoy receptor by leveraging binding hotspots underlying the cognate receptor interactions of GM-CSF and IL-2, without sharing any structural similarity with the cytokine receptors. Our findings contribute to the tracing of novel molecular mimicry mechanisms employed by pathogenic viruses.

T hroughout evolution, mammalian viruses co-evolved with their hosts and developed numerous immunomodulatory strategies aimed at countering and evading antiviral responses by the host immune system. Among the many different immune evasion tactics deployed by viruses, are the use of antigenic variation, latency, interference with antigen presentation, inhibition of apoptosis of viral infected cells and viral mimicry 1,2 . The latter is mainly deployed by large DNA viruses of the pox and herpes family, which encode viral homologues of host cytokines, chemokines and their cognate receptors as primary molecular detractors of the host immune response [3][4][5] .
The orf virus, a 140 kb double-stranded DNA (dsDNA) virus from the parapoxfamily infects ruminants and humans worldwide causing contagious ecthyma, also known as orf, a highly spreadable skin condition characterized by severe lesions and ulcerations [6][7][8][9] . As such, the socioeconomic footprint of orf is estimated to be very large placing orf in the top 20 of the most impactful diseases in developing countries or poor communities with a heavy economic dependence on animal farming and agrarian activities 10,11 . Infection by the Orf virus through damaged skin elicits a severe and acute immune response characterized by local accumulation of T cells, B cells, dermal dendritic cells and neutrophils 12,13 . Although the Orf virus does not inflict systemic infections and orf is generally not fatal, it can lead to high mortality rates in young animals and children when ulcerations prohibit food intake or can drastically affect ovine flock sizes when lesions would hinder reproduction 8,14 .
The latter, GIF, is arguably one of the most intriguing viral proteins known because it targets two distinct host proteins pivotal to the host's ability to mount an effective immune response. For instance, GM-CSF is involved in the survival and activation of macrophages, neutrophils and eosinophils, the maturation of dendritic cells and differentiation of invariant natural killer T cells 21 , while IL-2 is crucial for the maintenance of regulatory T cells, and the proliferation of B cells, memory T cells and NK cells 22,23 . GIF is distantly related to type-II CKBP (CKBP-II), also known as the 'vCCI' or 'orthopox 35kDa major secreted virus' family [24][25][26] , and displays restrictive cross-species specificity since it is able to bind ovine GM-CSF and IL-2, but not their human or murine orthologues 20 . However, the structural and mechanistic basis for the remarkable functional duality displayed by GIF has remained uncharacterized.
In this study, we employ an integrative structural biology approach in an effort to elucidate the structural and mechanistic principles of how a single viral protein can sequester two different host proteins. Via a series of cross-validated structural snapshots including the crystal structure of GIF in complex with GM-CSF, two-dimensional (2D) classifications of GIF:GM-CSF and GIF:IL-2 complexes by negative-stain electron microscopy, the ensuing three-dimensional (3D) reconstruction of the GIF:IL-2 complex, and modeling of small-angle X-ray scattering (SAXS) data in conjunction with multi-angle laser light scattering (MALLS) measurements, we show that GIF adopts an obligate dimeric structure that can bind two molecules of its target cytokines. We further show via structure-based mutagenesis in combination with interaction studies that GIF does so with high affinity using a mutually exclusive interaction site for both cytokines and plays into binding principles utilized by the cognate receptors of GM-CSF and IL-2. Together, our findings help to unmask the molecular virtuosity gained during viral protein evolution towards new protein ligand specificities and the molecular mechanisms developed by viruses to subvert the host immune response.

Results
Crystal structure of the GIF:GM-CSF complex. To obtain structural insights at high resolution into the functional duality of GIF as a cytokine-binding platform we sought to determine the structure of GIF in complex with its target cytokines. Large-scale production of recombinant GIF by transient expression in HEK293T cells and subsequent purification by immobilized-metal affinity chromatography (IMAC) and size-exclusion chromatography (SEC) initially led to prohibitory low amounts of purified GIF protein and uncovered the tendency for recombinant GIF to aggregate in standard purification buffers. However, addition of a molar excess of either recombinant ovine GM-CSF or IL-2 to culture medium containing secreted recombinant GIF enabled in situ reconstitution of GIF:GM-CSF and GIF:IL-2 complexes that could be purified to homogeneity for further biophysical and structural studies. Indeed, crystallization trials and subsequent crystal optimization efforts yielded crystals of the GIF:GM-CSF complex that enabled determination of the crystal structure to 2.8 Å resolution (Table 1) allowing us to obtain for the first time structural snapshots of GIF and its complex with GM-CSF.
Our structural analyses reveal that GIF is an obligatory dimer, consisting of two GIF subunits with an 11-strand b-sandwich topology stabilized by three intramolecular disulfide bonds. Two GIF protomers interact via b-strand complementation mediated by strand bB and the C-terminal strand bK, resulting in an extended and concave b-sandwich platform (Fig. 1a). The ensuing dimeric GIF assembly is decorated by long loops and two short a-helices at the ends of each GIF protomer, with a dense cluster of N-linked glycans on the bottom side (Fig. 1a). The GIF dimer interface is additionally stabilized by two sets of interactions, one cluster centered around the intersheet disulfide bridge between Cys59 on strand bB and Cys59 on bK on either side of the twofold axis, and a second set mediated by the tips of opposing G-H loops (Fig. 1b; Supplementary Table 1). Interestingly, predictions of structural disorder identify the C-terminal 30 residues of GIF as intrinsically disordered. It is now clear that the second half of that sequence is essential for constructing the dimeric core structure of GIF, while the first half in the JK loop could not be modeled and might indeed be intrinsically disordered. The b-sandwich core of GIF is very similar to the fold representing poxviral CKBPs of the poxviral immune evasion (PIE) superfamily (Fig. 2) consistent with prior predictions and structural annotations [27][28][29][30][31] .
The GIF:GM-CSF complex features two molecules of GM-CSF at equivalent interaction epitopes at the poles of the dimeric b-sandwich platform (Fig. 1a). GM-CSF employs the helixbundle face defined by helices aA and aD to interact with a rather broad surface on GIF contributed by strands bF, bA, bI and bJ burying B1,750 Å 2 of solvent-accessible surface area ( Fig. 1c; Supplementary Table 2). The interaction interface displays striking electrostatic complementarity with GIF contributing a highly basic surface flanking an island of aromatic residues, while GM-CSF projects a conspicuously acidic face of equal extent ( Fig. 1g; Supplementary Table 2). In contrast, the side of GIF not involved in binding to GM-CSF displays opposite electrostatic properties in that is highly acidic (Fig. 1g) and heavily glycosylated (Fig. 1a).
Given the extensive interactions between GM-CSF and GIF, we wondered whether GM-CSF undergoes any conformational changes upon binding to GIF. To enable such comparisons, we determined the crystal structure of ovine GM-CSF (Table 1). Structural superposition of the bound and unbound forms of GM-CSF reveals that the helical bundle core of GM-CSF only undergoes rather subtle main-chain and side-chain adjustments, most notably Lys20 and Asn27 on helix aA, which tilt downwards to interact with Gln41 and Trp43 on bA of GIF, respectively ( Supplementary Fig. 1). Furthermore, Glu108 on helix aD reaches out to Arg106 on the E-F loop of GIF, whereas Phe115 bends inwards to make place for His209 on bI of GIF. We further note that in one of the two-bound GM-CSF molecules, the N-terminal loop preceding helix aA at Gln14 is disordered, whereas in the second copy of GM-CSF it swings inwards by almost 180°to align roughly parallel to helix aA and interact with Phe189 and Tyr228 on GIF (Supplementary Fig. 1; Supplementary Table 2).
GIF-cytokine complexes are structurally distinct. Next we sought to obtain comparative structural information on GIF:GM-CSF and GIF:IL-2 complexes by negative-stain electron microscopy (EM) using purified complexes following co-expression of GIF with each of its target cytokines in HEK293T cells (Fig. 3a,b; Supplementary Fig. 2). The ensuing 2D class averages for the two complexes revealed particles with apparent twofold symmetry, albeit with striking differences. On one hand, the 2D class averages for the GIF:GM-CSF complex ( Fig. 3b) agreed very well with the crystal structure for the GIF:GM-CSF complex (Fig. 1a). On the other hand, however, the particle classification for the GIF:IL-2 complex (Fig. 3a) showed a drastically different assembly resembling a horseshoe-shaped structure. This prompted us to proceed to a 3D reconstruction of the particle representing the GIF:IL-2 assembly (Fig. 3c), which confirmed the shape incompatibilities with the crystal structure of the GIF:GM-CSF complex and revealed structural features to support a possible model representing the GIF:IL-2 complex (Fig. 3d). The 3D EM envelope representing the GIF:IL-2 complex can be interpreted in terms of a GIF dimer occupying the toe of the horseshoe mounted by two copies of IL-2 at the poles of the GIF dimer (Fig. 3d). The resolution of the 3D EM model did not allow us to discriminate the absolute orientation of IL-2, although the features of the 3D EM model strongly suggest that IL-2 is orientated with the longitude of its helical bundle nearly perpendicular to the plane of the GIF dimer (Fig. 3d). Indeed, similarly prepared and purified GIF:GM-CSF and GIF:IL-2 complexes characterized by SAXS yielded scattering data that despite the heterogeneous glycosylation levels of the samples, support this interpretation and further show that the two GIF-cytokine complexes have similar molecular weights (Supplementary Fig. 3a-d; Supplementary Table 3).
To provide additional evidence for the molecular basis and stoichiometry of the two structurally distinct GIF-cytokine complexes we carried out a comparative analysis by SEC and MALLS (Fig. 4a,   Our analyses show that both the GIF:GM-CSF and the GIF:IL-2 complexes are chromatographically relatively monodisperse as evidenced by the molecular weight profiles across each elution peak, and that both complexes obey a 2:2 stoichiometry assuming fully N-glycosylated proteins (Fig. 4a,b; Supplementary Table 4). Addition of a large molar excess of purified cytokine to its respective complex with GIF did not change the chromatographic behaviour of the complex indicating that the purified GIF-cytokine complexes were stable and fully occupied by cytokines. Along these lines and to cross-validate the consistency of the deduced 2:2 stoichiometry of the GIF:cytokine complexes, we treated a given GIF-cytokine complex with a molar excess of the second cytokine partner and analysed the resultant protein mixtures by SEC-MALLS (Fig. 4c,d). Indeed, the SEC-MALLS profiles showed that the 2:2 assemblies were conserved and that the chromatographic profiles and fits were consistent with the exchange of one cytokine for the other (Fig. 4c,d). Therefore, we conclude that GIF is able to adopt well defined 2:2 stoichiometric assemblies in the presence of GM-CSF and IL-2, despite its aggregative propensity. In this regard, we note that we were only able to observe chromatographically well-behaved GIF in the presence of detergent (0.15% (w/v) CHAPS), albeit as an unresolved mixture of dimers and tetramers ( Supplementary  Fig. 4a) 20 . Finally, we note that glycan analyses of the two complexes unmasked very complex N-glycan profiles suggestive of high levels of sialylation, and by extrapolation expected structural heterogeneity with respect to the glycan structures.
Thus, our structural analyses of GIF:GM-CSF and GIF:IL-2 complexes by several orthogonal approaches including X-ray crystallography, EM, SAXS and SEC-MALLS illustrate that the two GIF-cytokine complexes are structurally distinct based on the ability of dimeric GIF to sequester two copies of each cytokine to yield stable assemblies obeying 2:2 stoichiometry.

GIF binds GM-CSF and IL-2 via mutually exclusive epitopes.
Previous studies demonstrated that GIF inhibits the biological activities of both ovine GM-CSF and ovine IL-2 in soft bone marrow cell colony and T-cell proliferation assays, respectively 20 . To characterize the interaction between GIF and its two     target cytokines kinetically and in terms of binding affinity we employed bio-layer interferometry (BLI) using biotinylated GIF immobilized onto streptavidin-coated BLI sensors. Our measurements reveal that GIF binds GM-CSF and IL-2 with high affinities yet with quite distinct kinetic binding profiles (Fig. 5a,b). On one hand, binding of GM-CSF to GIF is characterized by rather slow association kinetics and very slow dissociation kinetics leading to a K D ¼ 27 nM, whereas IL-2 binds to GIF with B60-fold higher affinity (K D ¼ 0.47 nM) that can be recapitulated by fast association kinetics and moderately slow dissociation kinetics (Fig. 5a,b). The unique capacity of GIF to bind tightly to two different cytokines resulting in structurally distinct complexes prompted us to explore the possible overlap between the GM-CSF and IL-2 binding epitopes. We thus carried out BLI experiments meant to measure the possible binding of GM-CSF and IL-2 to preassembled GIF:IL-2 and GIF:GM-CSF complexes, respectively, at concentrations that would warrant the stability of the preassembled complexes (Fig. 5c,d). We found that neither GM-CSF nor IL-2 were able to bind additively to GIF already occupied by the other cytokine, revealing that the cytokine binding sites on GIF are mutually exclusive. In light of our structural analyses, it would appear that the binding footprints of GM-CSF and IL-2 on GIF might be overlapping to a certain degree (Figs 1 and 3).

T T T T ORF_GIF
To delineate the possible boundaries of the two distinct binding epitopes utilized by GIF to sequester GM-CSF and IL-2, we resorted to our high resolution snapshots of the GIF:GM-CSF complex to design GIF variants carrying alanine mutations at amino acids involved in the GIF:GM-CSF binding interface (Figs 1c-e and 6a). Thus, given the overwhelmingly positive electrostatic potential of the GIF-binding epitope (Fig. 1g), we hypothesized that mutation of arginine, lysine and histidine residues involved in more than one interaction with amino acids on GM-CSF along the helix-A/D face of the helical bundle might be appropriate candidate amino acids to mutate in GIF.
Our BLI data revealed that all mutations caused a B10 to B120-fold decrease in the affinity of GIF for GM-CSF, with Lys211 on GIF constituting a true hotspot at the epicenter of the binding interface ( Fig. 6a; Supplementary Fig. 5e; Supplementary  Table 5). Interestingly, all mutations negatively affected the off-rate of the interaction, which is quite remarkable given that all mutations were alanine substitutions in the context of an interaction footprint described by very pronounced electrostatics (Fig. 1g). Such kinetic profiles lend support to the specificity of the interactions concerned. Importantly, while all but one mutation had no marked effect on the GIF:IL-2 interaction, Arg188Ala at the far end of the GIF:GM-CSF interaction epitope bordering the putative IL-2 binding site (Fig. 1c) elicited a 10-fold reduction in the affinity of the GIF:IL-2 interaction and a near 20-fold decrease in the affinity of the GIF:GM-CSF interaction by drastically impacting the interaction off-rate ( Fig. 6a; Supplementary Fig. 5i,f; Supplementary Table 5). In light of our entire alanine-scanning mutagenesis data set, this very fortuitous observation lends strong support to the localization of the IL-2-binding footprint as modeled immediately adjacent to, and quite possibly partly overlapping with, the GM-CSF binding site (Fig. 6a). Indeed, such mode of binding is corroborated well by the mutual exclusivity of the two cytokine binding sites revealed by our binding studies using wild-type GIF (Fig. 5c,d), because the presence of one cytokine on GIF would sterically prevent binding of the other. neutralizes ovine GM-CSF and IL-2 (Fig. 5c,d) but not the human counterparts (Fig. 5e,f) and in light of the GIF-mediated abrogation of the bioactivity of the two cytokines in cellular assays 20 , we sought to trace the structural basis for the antagonistic properties of GIF. In the first instance we compared the binding mode of ovine GM-CSF to GIF with the structure of human GM-CSF in complex with its cognate receptors GMRa and bc (refs 32,33). The two GM-CSF orthologues share 80% sequence identity and can be superposed with an r.m.s.d. of 0.7 Å (backbone atoms). Our analyses show that binding of GIF to GM-CSF drastically overlaps with both cognate binding epitopes of GM-CSF to GMRa and b c (Fig. 6b). Consequently, GIF can be classified at the outset as a competitive decoy receptor for GM-CSF.
A closer look at the structural superimpositions reveals that competitive antagonism by GIF goes well beyond a steric exclusion mechanism. Remarkably, GIF engages several residues in helices aA and aD of GM-CSF that would be involved in binding to GMRa (Fig. 6b, sites 1 and 2). In particular we highlight the deployment of Lys 211 by GIF to cap helix aD in GM-CSF (Fig. 6b, site 2), which is exactly what the functionally important Arg302 of GMRa does 34,35 in the GMRa:GM-CSF complex 32 . Furthermore, GIF employs a triad of residues (Gln 108, Tyr 110 and Arg 45) to engage residue Asp112 in GM-CSF, which constitutes a structural and functional hotspot for interacting with GMRa (refs 34-36). Together, these analyses show that GIF has adopted a mechanism of interface mimicry 37 and illustrate how a viral protein bearing no sequence or structural similarity to the cognate receptors of the target cytokine has evolved convergently to sequester the target cytokine by partially hijacking key principles and interactions underlying the cognate cytokine-receptor complex. Indeed, such intricate mimicry might present a daunting challenge to the host in terms of developing an effective counter.
Regrettably, the lack of structural information at high resolution for the binding of IL-2 to GIF limits us from making a robust proposal for the possible structural mechanism underpinning the role of GIF as a viral decoy receptor to antagonize IL-2-mediated signalling in the host. However, we note that the likely near vertical orientation of the long axis of the helical bundle of IL-2 with respect to the GIF dimer (Fig. 3d) in the context of IL-2 in complex with IL-2Ra, IL-2Rb and g c (ref. 38), would create severe steric hindrance for binding of either the IL-2Rb or the g c receptor chains. Interestingly, it is well established that binding of IL-2 to IL-2Ra primes IL-2 for binding with a higher affinity to IL-2Rb (refs 38,39). It is thus tempting to consider whether GIF might play into the mechanism of IL-2-mediated receptor activation by binding the activated form of IL-2 with high affinity to compete more effectively against the cognate cytokine receptors.

Discussion
The Orf virus GIF protein provides a fascinating example of viral protein evolution to enable evasion of the host immune response. A decade and a half after the discovery of GIF as a viral protein able to antagonize the bioactivity of both GM-CSF and IL-2 (ref. 20), two distinct cytokines pivotal to the host immune response, we have presented here the structural and mechanistic determinants underlying cytokine sequestration by GIF. GIF neutralizes GM-CSF and IL-2 with high affinity by engaging the two cytokines on mutually exclusive binding epitopes on the concave surface of its dimeric scaffold yielding two structurally distinct complexes (Fig. 7). The structural and evolutionary origins of GIF can be traced to a conserved beta-sandwich fold that serves as the hallmark of the PIE domain superfamily 40 . Although GIF does not bear any resemblance to the cognate receptors of GM-CSF and IL-2 whatsoever, it targets hotspots and mechanistic binding principles that are pivotal to the assembly of productive complexes between GM-CSF and IL-2 and their cognate receptors. In this regard, GIF joins a very sparsely populated class of viral decoy cytokine receptors that are structurally distinct from the cognate receptors, as was also recently shown for BARF1 from the Epstein-Barr virus 41 , and the Yaba-like disease virus 2L protein 42 .
The herein structural characterization of GIF now provides the structural prototype for PIE viral proteins that target cytokines, defining a distinct functional branch in the phylogenetic tree of the PIE superfamily (Fig. 8) With such a wealth of structural data at hand we deemed it opportune to trace the common and differentiating structural features of the PIE superfamily (Fig. 9) 11 12 13 Elution volume (ml) 11 12 13 Elution volume (ml) 11 12 13 Elution volume (ml)  of the common b-sandwich core becomes apparent when comparing the targeting (CC and CXC chemokines, cytokines, MHC class I molecules) as well as the binding epitope of ligands binding the aforementioned proteins. Whereas Orf CKBP, A41 and vCCI all bind chemokines on the acidic surface formed by b-sheet II, Orf GIF binds both GM-CSF and IL-2 through a basic patch present on b-sheet I (Fig. 9a). Moreover, although the surface charge distribution of b-sheets I and II is generally conserved among PIE superfamily members, the ectromelia virus SECRET domain displays reversed surface charge properties of b-sheets I and II, and differs from other viral CKBPs since it binds chemokines on the negatively charged surface of b-sheet I. Conversely, b-sheet I is shielded away by two loops in A41 and vCCI: one connecting b-strands bG and bH, and the second present at the C-terminal end of both proteins. The bG-bH loop serves a different role in Orf GIF, where it is involved in GM-CSF binding through the functionally important Arg188 (Fig. 1c) as well as stabilizing the GIF dimer through binding to the bG 0 -bH 0 loop of an opposing GIF monomer. In Orf CKBP, b-sheet I is occupied by the extra a-helix aC, which connects b-strands bJ and bK. Cowpox virus CPXV203 on the other hand uses the extra C-terminal helices aB and aC covering b-sheet I to bind into a groove formed by subunits a2 and a3 of MHC1 (Fig. 9a). Helix aC is not present in Orf GIF, instead it is replaced by an arginine-and proline-rich stretch of 15 amino acids connecting bJ and bK. This loop region is predicted to be unstructured and in fact we were not able to model it in the electron density maps for the GIF:GM-CSF complex. Phylogenetic analysis suggests that the binding specificity of GIF towards ovine GM-CSF and IL-2 most likely appeared after speciation of the Orf virus 16 . It thus appears that the J-K loop diverged to an ahelix in Orf CKBP and an unstructured loop in Orf GIF, which allows the surface formed by b-sheet I to bind GM-CSF and IL-2 (Fig. 9b).
An intriguing aspect of the PIE superfamily b-sandwich fold concerns its oligomerization capacity. A41, vCCI, CPXV203 and the ectromelia virus SECRET domain form monomers in solution, while Orf CKBP is an obligate dimer (Fig. 9a). Interestingly, the SECRET domain present in the CrmB gene from variola virus was also found to be a dimer 47 . Orf CKBP dimerizes via b-sheet complementation involving its C-terminal b-strand bK. A similar interface involving b-strand bK is found in GIF dimers, where the binding between GIF monomers is further strengthened by additional interactions mediated by the G-H loop. By combining an oligomeric state with binding multivalence on a single molecular platform that can engage the target molecules with high affinity into long-lived complexes, GIF appears to recapitulate well the general strategy employed by viral effector proteins and decoy receptors for inflicting a potent and efficient modulation of the host immune system within the relatively short timeframes available to them 41,48 . Arguably, what appears to set GIF apart from all other viral decoy receptors characterized to date is not only its ability to target two different cytokines but also the fact that it does so via structurally distinct assemblies (Fig. 7). In the case of GM-CSF, the mechanistic consequences of its sequestration by GIF can be traced to interference with the cognate cytokine receptor binding interfaces (Fig. 6b), which is likely the mechanism deployed for the antagonism of IL-2 activity as well. Together, our work contributes a novel example of the virtuosity and potency of viral effector proteins and molecular decoys to achieve immunomodulation of the host.

Methods
Protein production and purification. Sequence optimized constructs containing residues 18-144 of ovine GM-CSF (Uniprot code: P28773) and residues 21-155 of ovine IL-2 (Uniprot code: P19114) were cloned in the pET-15b vector (Novagen) and pHLsec vector 49 encoding an N and C-terminal His 6 -tag respectively. Ovine GM-CSF and IL-2 cloned in the pET-15b vector were expressed in BL21(DE3) (Novagen) Escherichia coli cells and produced from inclusion bodies by rapid dilution refolding (described in reference Verstraete et al. 50 ), while the same constructs cloned in the pHLsec vector were transiently expressed in mammalian HEK293T cells 49 . Refolded oGM-CSF and oIL-2 were purified by IMAC using prepacked 5 ml Ni-NTA Superflow cartridges (QIAGEN). After binding, oGM-CSF or oIL-2 were washed with 10 column volumes washing buffer (50 mM NaH 2 PO 4 , 300 mM NaCl, 15 mM Imidazole pH 7.0) and eluted with 10 column volumes elution buffer (50 mM NaH 2 PO 4 , 300 mM NaCl, 500 mM Imidazole, pH 7.0). After elution protein fractions were pooled, concentrated to a volume of 1 ml and injected onto a Prep-Grade Hiload 16/600 SD 75 size exclusion column (GE Healthcare) equilibrated with HEPES-buffered saline buffer (HBS, 20 mM HEPES, pH 7.5, 150 mM NaCl). oGM-CSF and oIL-2 expressed in HEK293T cells were purified using the same purification protocol as described above, but instead of Ni-NTA cartridges, a Co 2 þ -packed TALON FF column (Clontech) was used for IMAC.
A construct containing residues 1-265 of the Orf virus GM-CSF/IL-2 inhibition factor (GIF, Uniprot code: Q9J5U5) was cloned into the pHLsec vector encoding a C-terminal His 6 -tag 49 . His-tagged GIF was transiently expressed in HEK293T cells. During transfection, sodium valproic acid (3.6 mM) was added to the transfection medium to boost expression levels. After 5 days of expression, medium was harvested, supplemented with 0.15% (weight per volume) CHAPS and loaded onto a Co 2 þ -packed TALON FF column (Clontech). During IMAC, the same buffers as described above were used, supplemented with 0.15% (weight per volume) CHAPS. Eluted GIF was pooled, concentrated and injected onto a Prep-Grade Hiload 16/600 SD 200 size exclusion column (GE healthcare) equilibrated with HBS supplemented with 0.15% (weight per volume) CHAPS.
To generate GIF:oGM-CSF and GIF:oIL-2 complexes, GIF was either co-expressed with oGM-CSF or oIL-2 in HEK293T cells by transient transfection 49 , or purified after addition of an excess or refolded oGM-CSF or oIL-2 to the HEK293T expression medium containing GIF. In each case, expression medium was harvested after 5 days of co-expression and loaded onto a Co 2 þ -packed TALON FF column (Clontech) or Ni 2 þ -packed complete column (Roche) for oIL-2 and oGM-CSF, respectively. Subsequent IMAC and SEC purification steps were performed using the protocol as described for unliganded GIF, but without adding CHAPS to purification buffers.
Crystallization and data collection. All crystallization trials were set up using a Mosquito robot (TTP LabTech) in a sitting drop vapour diffusion set-up with drop volumes of 200 nl (100 nl protein solution þ 100 nl reservoir solution). Refolded oGM-CSF was concentrated to 5 mg ml À 1 before initiating crystallization trials. Large plate-like crystals were obtained at 293 K in a crystallization condition containing 0.2 M sodium acetate trihydrate, 0.1 M TRIS hydrochloride pH 8.5 and 30% (weight per volume) PEG 4,000 (Hampton Crystal Screen 1, condition B10). Crystals were scooped, transferred to mother liquor supplemented with 25% ethylene glycol, and subsequently flash-cooled in liquid nitrogen. Diffraction data The obtained data set was processed using XDS 51 .
Purified complex between GIF produced in HEK293T cells and in vitro refolded oGM-CSF produced in E. coli was concentrated to 6 mg ml À 1 . Small and highly mosaic cube-like crystals were obtained at 293 K in a crystallization condition containing 0.15 M ammonium sulfate, 0.1 M MES pH 5.5 and 25% (weight per volume) PEG4000 (Molecular Dimensions ProPlex screen, condition C10). After optimization of this initial hit, new crystals with dimensions of 40 Â 40 Â 250 mm slowly grew over a period of 3 months in a condition containing 0.15 M ammonium sulfate, 0.1 M MES pH 6 and 24% (weight per volume) PEG 4,000. Before X-ray data collection, crystals were scooped, transferred to mother liquor supplemented with 25% ethylene glycol, and flash cooled in liquid nitrogen. Diffraction data were collected to 2.84 Å at the P14 microfocus beam line at PETRA III, Hamburg, and subsequently processed using XDS 51 . An overview of all data collection and refinement statistics can be found in Table 1.
Structure determination and refinement. Diffraction data for ovine GM-CSF was processed in spacegroup P2 1 (a ¼ 41.1 Å, b ¼ 77.02 Å, c ¼ 47.45 Å, b ¼ 111.5°). Structure determination was performed by maximum-likelihood molecular replacement (MR) in Phaser 52 using the crystal structure of human GM-CSF (PDB code 2GMF 53 ) as a search model. The resulting single MR solution was refined in Phenix 54 using positional and real-space refinement with non-crystallographic symmetry (NCS) restraints, individual anisotropic atomic displacement parameter (ADP) refinement and optimized X-ray/stereochemistry and X-ray/ADP weights (Table 1; Supplementary Fig. 6).
The obtained data set for the GIF:oGM-CSF complex was processed in space group C2 (a ¼ 145 After manual inspection of the collected data, severe anisotropy was observed. The data were ellipsoidally truncated and anisotropically scaled on-line using the anisotropy server 55 , which recommended resolution limits along the a*, b* and c* axes of 3.3 Å, 2.8 Å and 2.7 Å, respectively. As no homologues of the Orf GIF protein are available, a trimmed ensemble model was made with the program Ensembler of the Phenix package by using two proteins distantly related to GIF, vaccinia virus A41L (ref. 27) (PDB code ¼ 2VGA, 19% sequence identity) and cowpox virus vCCI (ref. 28) (PDB code ¼ 1CQ3, 11% sequence identity). This trimmed ensemble and the refined structure of ovine GM-CSF, were used as search models for maximum-likelihood MR in Phaser 52 . This led to an MR solution containing two trimmed ensembles and two oGM-CSF molecules, whereby a GIF dimer is formed decorated by two oGM-CSF molecules at equivalent positions on GIF. Initial refinement cycles were performed in Buster 56 using NCS refinement and anisotropic ADP refinement with TLS. After tracing of the main chain and most of the missing side-chains of GIF, further refinement cycles were carried out in Phenix using positional and real-space refinement with NCS restraints, individual anisotropic ADP refinement combined with TLS and optimized X-ray/stereochemistry weights (Table 1; Supplementary  Fig. 6).
Production of biotinylated proteins for kinetic studies. Residues 1-265 of Orf GIF as well as six interface residue mutant GIF species (R45A, R45A/R106A, R188A, R106A/R188A, K211A and H232A) were cloned in the pHLsec vector containing a C-terminal AviTag 49 . The pHLsec-GIF-AviTag constructs were transiently co-transfected in HEK293T cells with E. coli biotin ligase (BirA) cloned in the pDisplay vector (pDisplayBirA-ER 57 ). Before co-transfection, cell medium was replaced by serum-free DMEM medium containing 100 mM of D-biotin. After 5 days of co-expression, expression medium was harvested and desalted using a HiPrep 26/10 Desalting column (GE Healthcare) equilibrated with HBS to remove excess D-biotin.
Bio-layer interferometry. All experiments were performed in 1 Â PBS (137 mM NaCl, 2.7 mM KCl, 10 mM Na 2 HPO 4 , 1.8 mM KH 2 PO 4 , pH 7.4) supplemented with 0.01% (w/v) bovine serum albumin and 0.002% Tween-20, using an Octet RED96 instrument (ForteBio), operating at 298 K. Streptavidin-coated biosensors were functionalized with biotinylated GIF, quenched with a 10 mg ml À 1 biocytin solution and then exposed to different concentrations of ligand (GM-CSF: 1-8 mM, IL-2: 3.75-30 nM). To verify that no nonspecific binding was present during the interaction assay, non-functionalized biosensors were used as a control by measuring in parallel all ligand concentrations as well as running buffer. The measurements presented in Fig. 5c were carried out based on sensor tips bound with immobilized GIF saturated with 250 nM of oIL-2 and were used for measurements against 20 mM oGMCSF. The measurements presented in Fig. 5d were carried out based on sensor tips bound with immobilized GIF saturated with 20 mM oGMCSF and were used for measurements against 250 nM of oIL-2. All data were fitted with the FortéBio Data Analysis 7.1 software using a 1:1 interaction model. All binding experiments were performed in triplicate. Displayed K D , k d and k a values represent the averages of these triplicates.
Negative-stain electron microscopy and image analysis. GIF:IL-2 was applied to the clean side of carbon on a carbon-mica interface and stained with 2% (weight per volume) uranyl acetate. Images were recorded under low-dose conditions with a JEOL 1,200 EX II microscope at 100 kV and at a nominal magnification of 40,000x. Negatives were digitized on a Zeiss scanner (Photoscan TD) to a pixel size of 3.5 Å at the object scale ( Supplementary Fig. 2a). A semi-automatic particle selection with BOXER 58 led to an extraction of a total of 26,910 GIF:IL-2 subframes of 64 Â 64 pixels. CTF correction and 2D reference-free classification was performed in RELION 59 (Supplementary Fig. 2b). 2D class averages suggested the presence of a two-fold symmetry axis. Three different sets of 30 to 40 best class averages were extracted and used to calculate 12 ab initio 3D reconstructions of GIF:IL-2 with VIPER applying a C2 symmetry [60][61][62] . All plausible solutions had a similar horseshoe appearance with a well-defined global core. Therefore, they were aligned in 3D, averaged and filtered to 50 Å to create a low resolution GIF:IL-2 3D model ( Supplementary Fig. 2c). This model was further refined in RELION using 3D auto-refinement and 3D classification procedures (Supplementary Fig. 2d). In an attempt to better take the SAXS data into account, a SAXS model of GIF:IL2 was filtered to 50 Å resolution and used as an initial model for another cycle of RELION refinement of the previously selected 24,324 GIF:IL-2 particles. This final refinement resulted in 3D maps with a resolution of 17 Å according to the FSC ¼ 0.5 criterion (Supplementary Fig. 2e,f) GIF:GM-CSF was applied to the clear side of carbon on a carbon-mica interface and stained with 2% (w/v) sodium silicotungstate, pH 7.5. Images were recorded under low-dose conditions with a FEI Tecnai12 microscope operating at 120 kV and at a nominal magnification of 49,000 Â using a Gatan Orius SC1,000 CCD camera. The CTF for each micrograph was determined with CTFFIND3 (ref. 63). Using projections of the crystal structure of GIF-GM-CSF filtered to 18 Å as templates, particles were automatically selected from all the micrographs using the Fast Projection Matching (FPM) algorithm 64 . The resulting data set was cleaned for bad images. The final 73,153 particles were filtered to 25 Å and two-dimensional (2D) reference-free classification was performed in RELION 63 with no CTF correction.
Modeling of GIF and oIL-2 in the EM map. The crystal structure of dimeric GIF and two copies of an oIL-2 homology model generated by MODELLER 65 were fitted in the EM map of the GIF:oIL-2 complex using the autofit procedure in UCSF Chimera 66 .
Small-angle X-ray scattering. SEC-SAXS data were collected at the SWING beamline at SOLEIL (France) using an integrated online HPLC set-up. Purified samples of GIF:oGM-CSF and GIF:oIL-2 co-expressed in HEK293T cells were first concentrated to 10 and 8.8 mg ml À 1 respectively before injection on a Superdex Increase 10/300 GL size exclusion column (GE Healthcare) equilibrated with HBS (pH 7.5) using an injection volume of 100 ml and a flow rate of 0.5 ml min À 1 . Measurements were performed at 293 K, within a momentum transfer range of 0.01 Å À 1 oso0.550 Å À 1 (GIF:GM-CSF) and 0.01 Å À 1 oso0.622 Å À 1 (GIF:IL-2), where s ¼ 4psin(y)/l and 2y is the scattering angle. Eighty nine buffer frames with an exposure time of 1 s were averaged and subtracted from 10 Â 1 s frames (GIF:GM-CSF) or 16 Â 1 s frames (GIF:IL-2) taken from the eluted protein peaks. Scattering data were collected on an AVIEX PCCD170170 detector. Initial data integration, averaging and subtraction were performed using the Foxtrot software at SWING. The forward scattering (I 0 ) and radius of gyration (R g ) were determined by PRIMUS-QT 67 using Guinier approximation 68 , and compared with values for I 0 and R g calculated by SCÅTTER 69  was evaluated using Autoporod 70 , the maximum particle dimension D max and corresponding distance distribution function P(r) were evaluated using SCÅTTER 69 . Molecular masses of GIF:oGM-CSF and GIF:oIL-2 complexes were calculated using methods independent of calibration standards provided by the SAXSMoW 71 and SCÅTTER 69 software. An overview of SAXS data collection, structural parameters and MW determination can be found in Supplementary  Table 3.
Modeling of SAXS data. The crystal structure of the GIF:oGM-CSF complex and the model of GIF:IL-2 based on the fit in the EM map were used as an input for the Allosmod-FoXS web server 72 to model missing loops, N and C termini and N-linked glycosylation. During each Allosmod-FoXS run, data up to a scattering angle of 0.5 Å À 1 was used. Fits of the resulting SAXS models of GIF:GM-CSF and GIF:IL-2 complexes to the experimental SAXS data were calculated using the FoXS software 73 .