The Structure of the Macrophage Signal Regulatory Protein α (SIRPα) Inhibitory Receptor Reveals a Binding Face Reminiscent of That Used by T Cell Receptors*

Signal regulatory protein (SIRP) α is a membrane receptor that sends inhibitory signals to myeloid cells by engagement of CD47. The high resolution x-ray structure of the N-terminal ligand binding domain shows it to have a distinctive immunoglobulin superfamily V-like fold. Site-directed mutagenesis suggests that CD47 is bound at a surface involving the BC, FG, and DE loops, which distinguishes it from other immunoglobulin superfamily surface proteins that use the faces of the fold, but resembles antigen receptors. The SIRP interaction is confined to a single domain, and its use of an extended DE loop strengthens the similarity with T cell receptor binding and the suggestion that they are closely related in evolution. The employment of loops to form the CD47-binding surface provides a mechanism for small sequence changes to modulate binding specificity, explaining the different binding properties of SIRP family members.

Signal regulatory protein (SIRP) 4 ␣ (also known as CD172a or Shps-1) is a member of a family of surface receptors that are expressed mainly on myeloid cells and are involved in their regulation (1,2). The SIRPs belong to the class of membrane protein families called "paired receptors" that comprise several genes coding for proteins with similar extracellular regions but radically different transmembrane/cytoplasmic regions with different (activating or inhibitory) signaling potentials. There are many examples of paired receptors on NK cells where several are involved in recognizing MHC or MHC-related antigens (reviewed in Ref. 3) with fewer examples on macrophages and myeloid cells. The latter include the SIRPs where in humans there are three closely related genes coding for transmembrane proteins with three extracellular IgSF domains (reviewed in Refs. 1 and 2). SIRP␣ (CD172a or SHPS-1 (4)) interacts with a ligand CD47 expressed on many cells (5)(6)(7) and gives an inhibitory signal through immunoreceptor tyrosine-based inhibition motifs in the cytoplasmic region that interact with phosphatases SHP-1 and SHP-2 (8) (Fig. 1). In contrast, SIRP␤ has a short cytoplasmic region and associates with a transmembrane adapter protein DAP12 containing immunoreceptor tyrosinebased activation motifs to give an activating signal (9 -12). SIRP␥ contains a very short cytoplasmic region lacking obvious signaling motifs but also binds CD47, albeit 10 times weaker than SIRP␣ (13,14).
A number of features distinguishes the SIRPs from other cell surface protein interacting pairs. First, although the sequences of SIRP␣ and SIRP␤ are very similar (ϳ90% identity (Fig. 2)), they show very different binding to CD47. Second, SIRP␣, SIRP␤, and SIRP␥ all have three IgSF domains, and SIRP␣ and SIRP␥ interact with CD47 that contains only one IgSF domain. This three-plus-one domain topology is unique in interactions between leukocyte membrane proteins where the most common arrangement of interactions is end-on between proteins each with two IgSF domains; however, it could maintain the 14 nm distance between membranes that is a very common spacing in cell contacts involving immune cells, the immunological synapse (15)(16)(17) (Fig. 1). Third, CD47 has a single IgSF domain linked to a region that spans the membrane five times, a very unusual topology (18). Fourth, the topology of the protein at the membrane surface may be important. There is a disulfide link between the IgSF domain of CD47 and one of the loops between the transmembrane regions, which is required for optimal binding (19). Fifth, SIRP␤ but not SIRP␣ is present on the cell surface as a dimer because of a Cys bridge in IgSF domain 3 (20) suggesting that cis interactions may be important in this family. Sixth, the SIRPs are unusual in that they have immunoglobulin C1-set domains that are normally only present in proteins involved in antigen recognition, MHC antigens, TCR, and antibody (21). It has been suggested that SIRPs are closely related in evolution to TCR and primitive antigen recognition proteins (22).
We describe here the x-ray crystallographic analysis of the ligand binding domain of SIRP␣, and we show by site-directed mutagenesis that it interacts with its ligand CD47 in a novel way by involving the loops at the end of the domain. This may explain the unusual sensitivity of the protein specificity to amino acid changes. * The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked "advertisement" in accordance with 18 U.S.C. Section 1734 solely to indicate this fact. The atomic coordinates and structure factors (code 2uv3) have been deposited in

EXPERIMENTAL PROCEDURES
Production of the SIRP␣ N-terminal V Domain (SIRP␣ d1)-A construct for the production of SIRP␣ N-terminal V domain (SIRP␣ d1) consisted of the SIRP␣ signal sequence and the N-terminal IgSF domain (residues 1-148; GenBank TM accession number CAA71403) followed by the sequence TRHHH-HHH. The amino acid terminus is predicted to be residue 30 of the precursor, giving the sequence EEEL. The numbering in this paper is that of this mature form. This construct was expressed using the pEE14 vector in the Lec3.2.8.1 variant of CHO cells as described (23). The recombinant protein was purified by nickel affinity chromatography, eluted with imidazole. N-Linked sugars were removed by incubation with Endo Hf (New England Biolabs) at 1 unit/g for 90 min at 37°C. The protein was purified by gel filtration in 10 mM Hepes, pH 7.4, 140 mM NaCl, 0.02% NaN 3 (HBS) and concentrated for crystallization trials. Selenium-labeled SIRP␣ was similarly prepared, but once the transfected CHO cells were confluent, the media were removed and the cells washed with PBS and incubated with methionine-free Dulbecco's modification of Eagle's media containing 30 mg/liter L-selenomethionine, 50 mg/liter L-cystine, and 2 mM sodium butyrate. It was not possible to estimate the level of selenomethionine incorporation by mass spectrometry; however, subsequent crystallographic analysis indicated it to be ϳ60%, sufficient for structure determination.
Recombinant extracellular SIRP␣ consisting of all three domains was produced in a similar manner (residues 1-349; accession number CAA71403) followed by the sequence TRHHHHHH). Recombinant extracellular region of CD47 with a biotinylation site and a polyhistidine tag (CD47) was prepared by transient expression in 293T cells using the pEFBOS vector (24) and biotinylated using the Bir enzyme as described (25). The CD47 fragment was subcloned into pEE14 and expressed by CHO K1 cells. Recombinant protein was purified by nickel affinity chromatography and gel filtration in HBS.
Surface Plasmon Resonance Analysis of the Interaction of SIRP␣ d1 and d1-3 with CD47-The interactions were analyzed using a BIAcore TM 2000 at 37°C as described (26). Briefly, ϳ8000 RUs of streptavidin were coupled to a CM5 research grade chip using amine coupling, and biotinylated CD47 and CD4d3 ϩ 4 were bound (659.7 and 724.6 RU, respectively). For kinetic analysis serially diluted monomeric SIRP␣ purified proteins were injected at the indicated active concentrations over the flow cells. K D values were obtained by both nonlinear curve fitting and Scatchard transformations to the binding data. Extinction coefficients of 9770, 36,850, and 22,550 M Ϫ1 cm Ϫ1 for SIRP␣ d1, SIRP␣ d1-3, and CD47, respectively, were calculated by Vector NTI (Invitrogen).  Residues identical in all three sequences are boxed. The positions of the ␤-strands determined from the structure (see below) are indicated by bars and letters and the short ␣-helical region by a zigzag (residues 83-85). The mutants made are indicated by the residue above the sequence; red denotes residues giving Ͻ35% wild type binding, orange 35-65%, and green no effect (see Fig. 3). Numbering is for the mature protein used for structural analysis. GenBank TM accession numbers for the sequences are CAA71403, NP_006056, and NP_061026 for SIRP␣, SIRP␤, and SIRP␥, respectively.
Crystallization, Data Collection, Structure Determination, and Refinement-Recombinant SIRP␣ d1 protein was concentrated to 45 mg/ml. Crystallization screening experiments were set up at 21°C in the Oxford Protein Production Facility crystallization facility as vapor diffusion experiments using 200-nl plus 100-nl droplets of protein and precipitant. These experiments yielded crystals under several conditions, the best being with ammonium sulfate, sodium/potassium phosphate, and lithium-sulfate. At these conditions three-row optimization experiments were set up with protein/precipitant volumes of 100:100, 200:100, and 300:100 nl (27). All crystals had a plate-like morphology. Some crystals grew to their full size within 24 h, whereas others took well over 1 month to appear. The crystal used for native data collection was grown at protein/precipitant volumes of 200:100 nl from 2.4 M ammonium sulfate, 0.1 M MES, pH 6.0 (Hampton Grid Screen, C3); the SeMet crystal was grown from protein at 24 mg/ml from 1.6 M ammonium sulfate, 0.1 M Hepes, pH 7.0 (Hampton Grid Screen, B4), at a protein/precipitant ratio of 300:100 nl.
The native Patterson map of SeMet peak data showed a strong peak (0.5, 0.2, 0.5) indicating noncrystallographic, translational symmetry. Anomalous Patterson maps of the peak data set showed good peaks on all of the Harker planes of space group P2 1 2 1 2 in addition to this translation peak. All Patterson peaks could be explained by two selenium sites, one per molecule as expected from the sequence for two molecules per asymmetric unit. The selenium positions were confirmed using the hkl2map (29), and phases were calculated with Solve and Resolve (30). Applying noncrystallographic symmetry information led to a clear electron density map. The initial model was built using Arp/Warp (31) followed by manual model building and several refinement cycles with refmac5 (32), leading to a model with R-factor of 28.7% (R-free ϭ 33.1%) in the resolution range 50 to 2.25 Å. A region of close contact between molecule A and B was apparent. This dimer was used as a starting model for molecular replacement to solve the structure of the native crystal because the self-rotation function of the native data had indicated the presence of a 2-fold noncrystallographic symmetry. The selfrotation function calculated with CNS (33) using data from 15 to 4 Å had a peak at ϭ 90.0°, ϭ 64.3°, and ϭ 180°with a peak height of 4.47 . Starting with the SeMet dimer model, a clear molecular replacement solution was found by the program Phaser (34) for the native data. The same dimer was found in the SeMet and native crystal. The initial model of the native structure was again built with Arp/Warp followed by several cycles of refinement and model building using REFMAC, including TLS modeling (32) and Coot (35). This led to the refined structure of SIRP␣ d1 (Table 1).
Some native as well as SeMet crystals were found to possess considerably larger unit cell dimensions (space group C222, unit cell dimensions a ϭ 115.4, b ϭ 139.4, and c ϭ 80.9 Å). These crystals were highly twinned and unsuitable for structural analysis. Twinned and untwinned crystals grew under identical conditions and could not be distinguished by their morphology.
Mutagenesis of SIRP␣-Mutants were introduced by PCR into the pEFBOS vector (24) containing SIRP␣ (d1-3) linked to rat CD4d3 ϩ 4 as an antigenic label (25). The proteins were expressed by transient transfection in 293T cells, concentrated, and immobilized on a BIAcore chip to which OX68 anti-rat CD4d3 ϩ 4 mAb had been coupled in BIAcore TM 2000 at 25°C (36). In each experiment using four flow cells, ϳ1200 response units (RU) of wild type, negative control (CD4), and two mutants were immobilized. Recombinant CD47 extracellular domain (0.2 M) (purified and expressed as for SIRP␣) was passed over the mutants to test for loss of ligand binding. The mutants were also tested with the SIRP␣ mAb Se5A5 that recognizes domain 1 and OX117 that recognizes domains 2 or 3 (13). 5 The use of 0.2 M CD47 (just below the K D value) provides a sensitive assay to detect changes in affinity. For each mutant the specific binding in response units was determined and compared with the binding obtained with wild type SIRP␣ (Fig. 3B). The binding values were adjusted for any differences in amounts of the SIRP proteins immobilized (36). Mutants were categorized as loss of binding site (less than 35% of wild type binding), intermediate (35-65%), and no effect (Ͼ65% of wild type binding).

RESULTS
Expression of SIRP␣ d1-The ligand binding, N-terminal domain of SIRP␣ (SIRP␣ d1) was expressed at high levels in the Lec3.2.8.1 variant of CHO cells and purified by nickel affinity chromatography. This cell line has defective glycosylation apparatus that renders glycoproteins sensitive to Endo Hf treatment (23). SIRP␣ d1 migrated as a major band of ϳ16 kDa compatible with the single IgSF domain and a minor band of higher molecular weight that was removed by Endo Hf (Fig. 3E). This is in line with expectations, because SIRP␣ d1 possesses a single potential N-linked glycosylation site, with sequence NITP, and it is established that a proline at that position can reduce the efficiency of glycosylation (37).
The Ligand Binding Potential Is Present in the N-terminal Domain of SIRP␣-Previous studies had shown that the N-terminal domain of SIRP␣ could bind the ligand CD47 (5), but a 5 D. Hatherley, unpublished data.
is the intensity of an individual measurement of a reflection, and ͗I(hkl)͘ is the average intensity of that reflection. c R-factor ϭ 100 ϫ (⌺ hkl ʈF obs ͉ Ϫ͉F calc ʈ/⌺ hkl ͉F obs ͉), where ͉F obs ͉ and ͉F calc ͉ are the observed and calculated structure facture amplitudes. d R-free equals the R-factor of test set (5% of the data removed prior to refinement). e Other atoms include 299 water molecules, 6 sulfate ions, and 2 MES molecules. f r.m.s.d. indicates root mean square deviation from ideal geometry. contribution of the other domains could not be excluded. This was tested by preparing recombinant SIRP␣ containing all three IgSF domains (d1-d3). A recombinant protein consisting of the single IgSF domain of CD47 together with a tag that could be biotinylated was prepared to enable immobilization of CD47 to streptavidin previously coupled to the BIAcore chip. Fig. 3 shows that SIRP␣ d1 and d1-3 bound CD47 with comparable affinities, K D of 0.3 and 1.3 M at 37°C and 0.2 and 0.6 M at 25°C (data not shown), respectively. The experiment was repeated in the opposite orientation (passing recombinant soluble CD47 over the chip). Despite using freshly prepared CD47, some aggregation occurred, but an estimate for the K D of 0.8 M at 25°C was obtained (Fig. 3C). The values are comparable with 2 M at 37°C from previous data (13). Thus the best estimate for the SIRP␣/CD47 interaction is K D ϳ 1.5 M with no significant contribution from domains 2 and 3.
Crystallization and Structure Determination-SIRP␣ d1 formed crystals rapidly that diffracted to ϳ1.8 Å (see "Experimental Procedures"). Molecular replacement phasing was unsuccessful, and the structure was solved by multiwavelength anomalous diffraction (MAD) analysis of a crystal (belonging to a different space group, see Table 1) of selenomethionated SIRP␣ d1 prepared by growing the CHO clone in the presence of selenomethionine. The analysis was complicated by the presence of pseudo symmetry (see "Experimental Procedures"); however, the structure was eventually refined satisfactorily at 1.8 Å resolution (R-free ϭ 24.9%; for other details see Table 1) to yield a model for the protein consisting of 299 water molecules, 6 sulfates, and 2 molecules of MES buffer. The crystallographic unit cell contains two SIRP␣ d1s related by a noncrystallographic 2-fold axis, with structures that are similar (overall fold shown in Fig.  4A, root mean square deviation on 117 C-␣s of 0.55 Å) apart from a few differences, which are minor with the exception of a significant difference in the DE loop that makes up the core of the major crystal contact (ϳ3.5 Å deviation at C-␣ of residue Thr-67). This contact is between the two noncrystallographically related molecules of SIRP␣ d1, which are in a trans-orientation, inconsistent with the formation of cis dimers at the cell surface (see Fig. 4 and "Discussion").
Structural Characteristics of SIRP␣ d1-The structure consists of a typical V-like Ig domain with two ␤-sheets linked by a conventional Cys bridge between ␤-strands B and F (Fig. 4). The A strand switches from the GFC to BED faces as found in many Ig V domains. There is also a short 3 10 helix (residues 83-85). The closest structures to SIRP␣ d1 were all IgSF domains, as judged by systematic comparison using the Dali and SSM servers (38,39); however, none were especially close and indeed the ranking of closest structures differed between the two servers. A composite, representative list of the 12 most similar domains was therefore compiled, and by using this information and a general knowledge of the field, definitive superpositions were obtained using the program SHP (39). The result is presented in Fig. 5, with an overlay of the C-␣ backbones of these structures shown in Fig. 5B, whereas Fig. 5A presents the global fold comparison data in a systematic way, by using pairwise analysis of all structures to produce a phylogenetic tree (40). This confirms that SIRP␣ d1 is substantially different from all other known structures, although it is more similar to the V domains of immunoglobulins and TCRs than it is to the CD2 adhesion molecules. The closest similarities are to TCR-like molecules and CTLA4. The distinctive features of SIRP␣ d1 include, most notably, an unusually large DE loop but also protruding FG and CCЈ loops and the absence of a CЉ strand (Figs. 4 and 5).
The presence of two molecules in the crystallographic asymmetric unit allows us to provide at least a partial assessment of the flexibility of the molecule. The linker region between d1 and d2 (residues 114 -120) is notable for the absence of flexibility (despite being involved in different interactions in the two molecules). This suggests that these two domains are likely to be locked together as a rather rigid unit, despite this sequence only containing a single proline residue. In fact there are only two regions of significant flexibility (excluding the C-terminal His 6 tag); the CЈ D loop (which is significantly displaced from the position characteristic of molecules containing a CЉ strand, see Fig. 5B) shows modest differences, but the greatest variation is seen in the extended DE loop (up to ϳ4 Å). This loop is crucial for the striking and extensive contact in the crystal between the two SIRP␣ d1 molecules (Fig. 5B), and despite the deviations between the two molecules their mutual interaction conspires to produce an almost exact 2-fold relationship (179.6°rotation and 0.3 Å translation). This interface is novel for interactions of Ig-like domains and is substantial, occluding 750 Å 2 of surface area, compared with 650 -690 Å 2 for the CD2/CD2 interaction observed in the crystal (41,42) and about 600 Å for the actual ligand interaction between CD2 and CD58 (43). Furthermore the anti-parallel orientation of these domains (see Fig. 4B) is such that the contact could potentially mimic interactions with membrane proteins on other cells, either through ligand binding (CD47) or homophilic interactions. This surface also shows some variation between SIRP family members (Fig. 2) making it an attractive candidate for the CD47-binding site, but mutagenesis experiments (see below) allow us to discount this possibility. It also seems unlikely that it represents a biologically relevant homophilic interaction, as no interaction of SIRP␣ was observed during biochemical analysis or by surface plasmon resonance (data not shown).

Identification of the Binding Region on SIRP␣ for CD47 by
Site-directed Mutagenesis-On the basis of the SIRP␣ d1 structure, a panel of 19 single residue mutants was designed to map the CD47-binding surface (shown in Figs. 2 and 6). There was a particular focus on three regions as follows: (i) the GFCCЈ face first identified for CD2 (42) that is frequently involved in leukocyte protein interactions involving IgSF domains (six mutations); (ii) the interface observed in the crystal contact involving the DE loop (four mutations); and (iii) the loops corresponding to the complementarity determining regions found in antibodies and TCRs (the BC, CЈCЉ, and FG loops, seven mutations). The strategy was to select outpointing side chains and to introduce major changes, the aim being to disrupt the interface, enabling its extent to be mapped, rather than finding the energetically most important residues in the interaction, a strategy we developed with the CD2/CD48 and CD200/CD200R systems (36,44). Mutants were introduced into a chimeric protein that contained CD4 domains 3 ϩ 4 as an antigenic tag, produced in a transient mammalian expression system, immobilized in the BIAcore TM using a mAb against the CD4d3 ϩ 4 tag, and tested for binding to CD47 and mAb recognizing SIRP␣ (Fig. 3). The concentration of CD47 used was close to the K D value; this allows for sensitive detection of mutants. All the mutants were expressed at satisfactory levels, could be immobilized through the CD4 mAb, and reacted with the OX117 mAb that recognizes a determinant in domains 2 or 3 of SIRP␣, indicating that the proteins were well folded. The mutants were also tested with a mAb against the N-terminal domain SE5A5. This mAb blocks ligand binding (6), and three mutants that destroyed ligand binding activity also lost SE5A5 activity (V33E, K96E, and I31K); three mutants (F94R, D100K, and The structures were superimposed pairwise, and a distance matrix was constructed with SHP (39,40). The tree was prepared with PHYLIP (65). B, superposition of these domains in a wire-type representation with SIRP␣ d1 in boldface green shown in the same view as in Fig. 4A. The inset shows a view onto the top of the molecule. The differences to other structures are particularly pronounced in the loop regions. Color coding for the molecules is according to the labels in A. S66D) were fully active for mAb binding but lost ligand binding activity, and two (M72R, R69E) had partial SE5A5 activity. This indicates that there is overlap between the antibody and ligandbinding sites. CD47 binding data for the mutants are summarized in Figs. 2 and 6.
Mutants in that region of the DE loop (Fig. 2) involved in the extensive crystal contact (Fig. 4B) had no effect on CD47 binding, indicating that this interface is not mimicking the CD47 interaction. Residues in the FG and BC loops caused dramatic loss of activity, whereas mutants in the main faces of the ␤-sheets had no effect. Some mutants had partial effects, and most of these were in the putative binding surface involving the BC, CЈD, and FG loops together with the top of the DE loop. One mutant that disrupts CD47 binding (S66D) is somewhat distant (almost 10 Å) from the putative binding face, but this side chain forms a hydrogen bond with the main chain so it may have an indirect effect by stabilizing the conformation of the DE loop. Overall, the mutants affecting ligand binding clearly suggest that CD47 binds through the ends of the domain rather than the face. This is illustrated in Fig. 6B, top view, that shows that the binding face involves almost all the top of the SIRP␣ d1 molecule, covering, at a minimum, a triangular area of ϳ17 Å in width and ϳ10 Å in height. Fig. 6C shows that the binding face is strongly polarized electrostatically, comprising an extensive basic patch.

DISCUSSION
The CD47 binding domain of SIRP␣ is a classic Ig V-type domain. IgSF domains are particularly common in surface membrane proteins of leukocytes, and in many cases N-terminal V-type domains mediate interactions with membrane proteins on opposing cells (18). In this respect SIRP␣ d1 is fairly typical; however, the structure shows a number of distinctive features, and the mutagenesis analysis shows that the binding face is unusual in that it involves the loops that are the equivalent of the hypervariable loops of antibodies and TCRs. This is in line with a very recent study (45) using mutational analysis of SIRP␤ and SIRP␣ to find residues crucial in binding CD47. Residues Val-27 and Gln-37 had effects on binding in agreement with the involvement of the BC loop.
The involvement of the loops is unusual in that IgSF domains involved in cell adhesion tend to associate via the faces of the ␤-sheets. This face to face interaction was first proposed on the basis of pseudo-homophilic interactions seen in crystals of CD2 that involved the GFCCЈ face (41,42) and subsequently observed in, for instance, B7/CTLA4 (46), the Coxsackie and adenovirus receptor (47), and CD2/CD58 (43). This mode of association is probably common to many interactions in this family of proteins, as illustrated by the recent structure of the homophilic interacting NTB-A (48). The homodimerization in Necl-1 involves the CCЈCЉD strands and intervening loops (49), FIGURE 6. Mapping the CD47 binding region by mutagenesis. A, position of SIRP␣ d1 residues mutated and their influence on CD47 binding as follows: residues colored red inhibit binding; residues colored yellow have an intermediate effect; and residues colored green have no effect on binding. Residues that have a strong influence on binding form a cluster toward the top of the molecule and define our proposed CD47-binding site. View is as in Fig. 4A with a slight, 2°, turn so that some residues can be seen more clearly. B and C, surface of SIRP␣ d1 when viewed from above. B, the mutagenesis coloring scheme of A is used. C, the surface is colored according to the electrostatic potential (blue for positive potential and red for negative; the potential was calculated by Pymol and should only be regarded as a qualitative guide). The proposed CD47-binding site as characterized by the red residues in A and extends along a pronounced basic patch on the surface as shown in C. D shows SIRP␣ represented as in A with CTLA4 superimposed (shown in semi-transparent orange). The CTLA4 molecule is complexed with B7-2 (58), and B7-2 has been moved to maintain the correct position with respect to CTLA4 and is shown in cyan. Note that although there is a striking correlation between the critical binding residues of SIRP␣ and the position of B7-2, there are serious steric clashes between the two molecules, so that although there may be similarities in the mode of engagement between SIRP␣ and CTLA4, there are also likely to be major differences. and the interaction with CD4 and MHC class II involves mainly the face of the fold (50). CD8 does interact with MHC class I through its loops, but it interacts as a dimer and thus resembles TCR and Ig antigen recognition (51,52). The recognition of MHC class I by the leukocyte LIR and KIR receptor family also involve the loops of two domains, but the interaction is centered around the junction of the two domains (53)(54)(55). In contrast in SIRP␣ the binding involves the complementarity determining region-like loops of only one domain and in addition the top of the unusually large DE loop (Arg-69). The use of loops presented at the ends of IgSF domains for ligand binding is, however, a recurrent theme in the rapid generation of different fine binding specificities; it is key to antigen recognition in the context of both antibodies and TCRs and is used by the LIR and KIR receptors. There is, however, also a parallel among the adhesive interactions, the interaction of CTLA4 with B7-1 and B7-2 (56,57). In that system the B7s contribute a classical "side" face of the ␤-sandwich; however, CTLA4 approaches this face in such a way that the corner of the molecule forms the point of contact, i.e. in a more "end-on" fashion than in the classical adhesive interactions, but not so squarely onto the loops as is seen in the immune receptors and, apparently, in SIRP␣. Nevertheless, the residues involved have some similarity with those implicated in SIRP␣ (58). This is illustrated in Fig. 6D where the CTLA4 of the CTLA4/B7-2 complex (56) has been superposed on SIRP␣ d1 to show how the region occluded by B7-2 overlaps the critical binding residues of position of SIRP␣ d1. Note that there are major differences in the binding loops between CTLA4 and SIRP␣, which is reflected in the steric clashes between SIRP␣ and B7-2 , largely because of the prominent DE loop of SIRP␣ and the shortening of the FG loop in CTLA4, so the B7-2 slips down and is disconnected from the DE loop of CTLA4 (this is even more pronounced in the complex of CTLA4 with B7-1 (57)). In this respect SIRP␣ instead resembles TCRs; it has an extended DE loop, which appears to be directly involved in ligand binding. Some TCRs use the DE loop to recognize MHC peptide, and this region is termed the HV4 or 4th hypervariable region (59,60). In contrast the DE loop is not used for antigen recognition by antibodies.
In the case of the SIRPs, the use of loops for binding may provide a molecular explanation for the different recognition properties of the three SIRPs despite close sequence similarity (Fig. 2). Examination of the sequence differences between the SIRPs, which can bind CD47 (SIRP␣ and SIRP␥), and SIRP␤, which does not, shows few candidates for residues essential for CD47 binding especially when one looks at other SIRP␣ sequences (61) that presumably correspond to polymorphisms as only one gene has been found in the human genome (2). The use of loops for binding provides a mechanism whereby small chemical differences in side chains can have large effects on the shape of the loops without disrupting the core ␤ fold of the molecule, as exemplified by the variation observed in the complementarity determining regions of antibodies and TCRs.
The SIRPs show additional similarities to TCRs and antibodies, in that their other two Ig-like domains have sequence similarity to the class of IgSF C domains found only in proteins associated with antigen recognition such as TCR, Ig itself, MHC antigens, and ␤ 2 -microglobulin. One theory proposed to explain this was that they represent the precursors of the antigen receptors (22); however, because no SIRP homologues have been recognized in species such as amphibians and fish, it may be more likely that SIRP evolved from an antigen receptor (1,2). The similarities in the binding face of SIRP␣ and the antigen receptors, together with the overall fold similarity, provides further support for a close evolutionary link between these molecules.
The data presented here indicate that the top of the Ig-like domain of SIRP␣ interacts with the single Ig-like domain of CD47. Assuming that the three domains of SIRP␣ are in a linear array (and the rigidity of the d1-d2 linker is consistent with this), then this interaction would span four Ig domains, as in the immunological synapse as illustrated in the scheme in Fig. 1. The orientation of CD47 may be important because it contains an unusual disulfide bond linking the extracellular domain of CD47 to one of the small extracellular loops of the 5-transmembrane segment (19).
Another aspect of proteins mediating cell-cell interactions is the use of the multimeric state (dimeric or higher order) of the proteins on the cell surface to modulate signaling outcome. This has been shown to be important in the B7 CD28/CTLA-4 system (46). We found no evidence of dimerization of the recombinant SIRP␣ N-terminal domain, and the crystal contacts observed are not compatible with cis interactions at the cell surface. However, it is still possible that the full extracellular domain can associate.
The finding that SIRP␣ recognition uses the loops at the top of the domain rather than the faces of the fold provides a mechanism to manipulate ligand binding specificity by modest chemical changes in amino acid side chains. In addition to antigen receptors, this mode of recognition is used by NK paired receptors LIR and KIRs, where there is evidence for rapid changes in specificity driven by pathogens. There are similarities between the NK self-recognition system and the highly polymorphic SIRPs in that CD47 is a marker of self, because cells deficient in CD47 are more susceptible to lysis by phagocytes (62). The main difference between the two systems is that, unlike the MHC antigen targets of the NK receptors, CD47 is not polymorphic. Nevertheless, the SIRP CD47 system of paired receptors recognition has evolved, presumably in response to pathogen load, to provide a highly sensitive recognition system regulating myeloid cell activity.