Crystal structure of a diabody, a bivalent antibody fragment

Crystal structure of a diabody, a bivalent antibody fragment Olga Perisic1, Philip A Webb1', Philipp Holliger', Greg Winter1 , 2 and Roger LWilliamsl* Centre for Protein Engineering and 2 Laboratory of Molecular Biology, MRC Centre, Hills Road, Cambridge CB2 2QH, UK

1

Background: Diabodies are dimeric antibody fragments. In each polypeptide, a heavy-chain variable domain (VH) is linked to a light-chain variable domain (VL) but unlike single-chain Fv fragments, each antigenbinding site is formed by pairing of one VH and one VL domain from the two different polypeptides. Diabodies thus have two antigen-binding sites, and can be bispecific. Direct structural evidence is lacking for the connections and dimeric interactions between the two polypeptides of the diabody. Results: The 2.6 A resolution structure has been determined for a bivalent diabody with a flexible five-residue polypeptide linker between the (amino-terminal) VH and (carboxy-terminal) VL domains. The asymmetric unit of

the crystal consists of four polypeptides comprising two diabodies; for one of these polypeptides the linker can be traced between the VH and VL domains. Within each diabody the two associated VH and VL domains make backto-back interactions through the VH domains, and there is an extensive VL-VL interface between the two diabodies in the asymmetric unit. Conclusions: The structure of the diabody is very similar to that which had been predicted by molecular modelling. Diabodies directed against cell-surface antigens should be capable of bringing together two cells, such as in cell-targeted therapy, because the two antigen-binding sites of the diabody are at opposite ends of the molecule and separated by -65 A.

Structure 15 December 1994, 2:1217-1226 Key words: bivalent and bispecific antibody fragments, diabody, protein crystallography, single-chain Fv

Introduction The antigen-binding site of an antibody is formed by its heavy-chain and light-chain variable domains, VH and VL respectively. These variable domains can be expressed in bacteria [1] and they associate to form an Fv fragment. Such a fragment may retain the affinity and specificity of its parent antibody, but often dissociates [2]. The domain association can be stabilized by a flexible peptide linker of 15 amino acids or longer linking the VH and VL domains [3-4] to form a single-chain Fv (scFv) fragment. However, some scFvs were observed to form dimers in solution [5], and it was shown that the VH and VL domains forming each of the antigen-binding sites within the dimer were derived from different polypeptides [6]. Thus co-expression of chains derived from two different antibodies A and B (viz. VHA-VLB and

VHB-VLA chains) yielded dimeric Fv fragments that bound to both antigens. Furthermore, dimerization could be promoted by shortening the linker between the VH and VL domains to five amino acids or less [6], as this would preclude intramolecular pairing of VH and VL domains on the same chain. Such dimeric Fv fragments

were termed diabodies and a structural model was proposed for them [6]. In this model the two chains associate to form two 'Fv heads', with the antigen-binding sites located at opposite ends of the molecule and pointing in opposite directions. For V chains with an amino-terminal VH linked to a carboxy-terminal VL, it was predicted that the two Fv heads would pack together back-to-back through VH-VH interactions.

The crystallization [7] and structures [8,9] of two scFv fragments have recently been reported. In the crystal structure reported for the anti-sialidase scFv [8], two scFvs are oriented back-to-back across a crystallographic two-fold axis so that the two antigen-binding sites are pointing in opposite directions. The authors concluded that it was difficult to determine whether the associated VH and VL domains were on the same polypeptide chain (scFv fragment) or on two chains (diabody), as the course of the 15-residue linker could not be traced, and the distances between domains were consistent with either possibility. Likewise, most of the residues of a 19 amino acid linker of the other reported scFv could not be traced, preventing the unambiguous assignment of connections between the VH and VL domains [9]. We present here the structure of a diabody, L5MK16, that binds specifically to phosphatidylinositol (PI)-specific phospholipase C81 (PLC, 1). As predicted by the model of Holliger et al. [6], the VH and VL domains from different polypeptides are paired to form two antigen-binding sites arranged back-to-back. The linker between the VH and VL domains is evident and allows unequivocal definition of the dimeric interactions in the diabody. Results and discussion The sequence of the L5MK16 diabody is shown in Fig. 1. The diabody includes a carboxy-terminal tag of six consecutive histidine residues, (His) 6, for affinity

*Corresponding author.

© Current Biology Ltd ISSN 0969-2126

1217

1218

Structure 1994, Vol 2 No 12 purification by metal-chelate chromatography [10] and a peptide tag (c-myc) for detection by enzyme-linked immunosorbent assay (ELISA). SDS-PAGE of a crystal (data not shown) indicated a molecular mass consistent with the sequence presented in Fig. 1. ELISA was used to verify that the diabody specifically bound the antigen (PI-PLC8 1).

other antibody structures is less than 1.1 A when comparing all framework residues. The deviation of the diabody from the mean antibody structure is shown as a function of residue number in Fig. 3. The deviation between the diabody and the mean structure is great only in those regions where the standard deviation among all antibody structures is also great.

The asymmetric unit of the crystal The overall structure of the diabody and the asymmetric unit of the crystal is illustrated in Fig. 2. The asymmetric unit of the crystal consists of two diabodies (made up of four polypeptides that will be referred to as A, C, E and G) as illustrated in Fig. 2b. The two polypeptides in a diabody are related to each other by a two-fold axis approximately parallel to the crystallographic c axis. The two diabodies in the asymmetric unit are related to each other by two-fold axes approximately parallel to the b and a axes. The principal intermolecular interactions within the asymmetric unit involve VH-VL pairing within each Fv head, VH-VH interactions between the two heads in each diabody and VL-VL interactions between diabodies. The myc and (His) 6 tags at the carboxyl termini of the light chains are not visible in the electron-density maps. In one of the four polypeptide chains, the interdomain linker is visible throughout its length; in the three other chains only the ends of the linkers are visible. No other breaks are apparent (when contoured at 1.0cr) in the four-fold averaged electrondensity maps (see the Materials and methods section).

The hypervariable regions are located at opposite ends of the diabody (Fig. 2). It has been shown that hypervariable regions L1, L2, L3, H1 and H2 of antibodies adopt a small repertoire of main-chain conformations known as canonical structures [12,13]. On the basis of its sequence, the canonical structures that would be expected for the L5MK16 diabody hypervariable regions are Ll:4, L2:1, L3: 1, H1:1 and H2:2. The conformations of these loops in the diabody structure are consistent with this classification. Only the H2 loop in the diabody is somewhat different from many members of the same class (Table 1). The H2 loop in the diabody is glycine-rich (containing three consecutive glycines) and has a well-defined conformation (Fig. 4a). The main-chain conformation of the H2 loop in L5MK16 is most similar to those seen in antibody-fragment structures ljhl [14] and lbbj [15] (Fig. 4b). The fourth residue of the H2 loop of L5MK16, Gly56, has positive 4 and qf angles, as do all of the members of this canonical-structure class. The most notable difference between the H2 main-chain conformation in L5MK16 and most other members of the same class is that the main-chain dihedral angles of the second residue of the loop, Gly54, could not be adopted by a nonglycine residue (=99 and 4t=-71). Two other structures have a glycine at this position with similar dihedral angles, lcgs [16] and lbbj [15]. The H3 loops of antibodies are structurally very diverse. The VH complementarity-determining region (CDR)3 in the diabody is 13

Structure of the VL and VH domains The framework structure [11] of the VH and VL domains in the L5MK16 diabody is very similar to that of other antibody structures. The root mean square (rms) deviation of Ca atoms between the L5MK16 diabody and

*VH domain 1 10 20 30 40 52A 60 ,1 10 20 30 CDR1 40 CDR2 QVQLQQSGTELMKPGRSLKISCKTTGYIFS NYWIE WVKQRPGHGLEWIC; KILPGGGSNTYN H2 H1 110 70 80 90 100ABCDEFGH 120 CDR2 70 80 90 CDR3 DKFKG KATFTADTSSNIAYMQLSSLTSEDSAVYYCAR GEDYYAYWYVLEDY WGQGTTVTVSS H3

Linker GGGGS

VL domain 50 1 10 20 27ABCDE 30 40 50 CDR2 1 10 20 CDR1 40 5 DIELTQSPLSLPVSLGDQASISC RSSOSLVHSNGNTYLH WYLKKPGQSF'KLLIY KVSTRFS L2 L1 60 70 80 90 100 70 80 90 CDR3 110 GVPDRFSGSGSGTDFTLKISRVEAEDLGVYFC SQSTHVPFT FGSGTKLEELK L3

Myc tag

His tag

RAAA EQKLISEEDL NGAA HHHHHH

Fig. 1. The sequence of the L5MK16 diabody with functional units numbered separately. Complementarity-determining regions (CDRs), as defined by Kabat et al. [11], are indicated by bars over the sequence and hypervariable (H and L) regions, as defined by Chothia et al. [13], are indicated by bars below the sequence. The sequential numbering used throughout the text is on the line closest to the sequence, and the numbering of Kabat et al. [11] is above the sequential numbering.

Crystal structure of a diabody Perisic et al.

Fig. 2. The overall structure of the L5MK16 diabody. (a) A diagram of one of the diabodies in the asymmetric unit (drawn with MOLSCRIPT [361). One polypeptide chain (C) is drawn in yellow and the other (A) in green. The fiveresidue linker of chain C that connects the amino-terminal VH domain to the carboxy-terminal VL domain is shown in magenta. This linker region is visible in the electron-density map. The other linkers in panels (a)-(d) are modelled according to this conformation and shown for clarity but are not entirely visible in the electron-derlsity map. (b) An overall diagram of the asymmetric unit of the crystals (drawn with GRASP [17]). A molecular surface is shown superimposed on the VL domain of chain C and the VH domain of chain A. The surface of the antigen-binding site is coloured purple. (c) Stereoview showing the C backbone trace of one polypeptide (chain A). Every tenth residue is labelled with a dot. H and L refer to the residues of the VH and VL domains respectively. (d) A stereoview showing the C. backbone trace of one diabody. Chain C is drawn in thick lines and chain A in thin lines. residues long with a conformation that differs from other H3 loops of similar size (Fig. 5).

composed of domains from different polypeptides and interact with each other through the V H domains.

Quaternary structure The structure of the diabody confirms the nature of the VH-linker-VL dimer that was originally proposed on the basis of molecular modelling [6]. The Fv heads are

Packing interactions within the antigen-binding domain (VHV L pairing) The VH-VL pairing involves the most extensive interface 2 of all of the diabody interactions: 1602 A of solvent-

1219

1220

Structure 1994, Vol 2 No 12 accessible surface area (1.7 A solvent probe radius, default radii of GRASP [17]) is buried by this interface. Although the area buried in the VH-VL interface of the bivalent diabody is extensive, there is an unusually large rotation of the VH domain relative to the VL domain about an axis perpendicular to the interface. The rotation is -12 ° relative to an arbitrary reference structure (see the Materials and methods section); all other antibody-fragment structures available in the Protein Data Base have smaller rotations with respect to the same reference. The orientation of VH relative to VL is most similar to that found in lbbj [15]. This may explain why lbbj gave the strongest signal among the rotation functions calculated using a series of VH-VL pairs as the search models (see the Materials and methods section). However, it is unlikely that the similarity between the diabody and lbbj is a consequence of bias introduced in the molecular-replacement solution because the rotation of VH with respect to VL in the diabody is -4 ° different from the rotation found for lbbj. In addition, the hypervariable regions with conformations distinct from lbbj showed up clearly in the electron-density maps.

(a) 3.0

2.5 2.0 C

1.5

0 4-

o

._ 0D

1.0

I

I

I~~~~~~~I

0.5 0.0 L 0

20

40 60 80 Residue number

100

120

(b) 3.0 2.5 2.0

I 1.5 a 1.0 0.5 0.0 0

20

40 60 80 Residue number

100

120

Fig. 3. The deviation of the diabody VL domain (a) and VH domain (b) from the mean antibody structure as a function of residue number. The deviations of the diabody Cs from the mean position of a group of 29 superimposed antibody fragments is indicated by the solid line. The standard deviation from the mean position for the same group of antibody fragments is indicated by the dotted line. The antibody fragments were aligned and superimposed using the programs AMPS and STAMP [37]. The thick horizontal bars indicate the hypervariable regions and the thin bars denote the framework regions, as defined by Chothia and Lesk [12].

Packing interactionsbetween antigen-bindingdomains (VH-VH interactions) As illustrated in Fig. 6a, the molecular modelling [6] had correctly identified the essential nature of the interactions between domains in a diabody. As predicted, the Fv heads were packed back-to-back through VH-VH interactions, with the antigen-binding sites located at opposite ends of the molecule. The solvent-accessible surface area buried in the VH-VH interaction is 1052 A2 (1.7 A probe radius, default atomic radii of GRASP). The residues involved in this VH-VH interface (excluding linker residues) are 14-15, 62-63, 65-67, 84-89, and 122; only one of these residues (residue 14) is typically involved in the interaction between a VH and a CH domain in a Fab fragment. The diabody VH-VH interface does not include the residues that form the 'molecular ball-and-socket' [18] in the VH-CH interface of Fabs (residues Thrll19 and Serl21 using the L5MK16 numbering scheme). In general, a different region of the surface is utilized for the back-to-back diabody VH packing than for the VH-CH interface of

Table 1. Rms deviations (in A) for H2 loops of several antibody fragments having canonical classes H1:1 and H2:2.

2hfl 1nca 1bbj 1fgv ljhl 1bbd 1igj ldba Itet L5MK16

2hfl

1nca

lbbj

1fgv

ljhl

lbbd

ligj

Idba

Itet

L5MK16

0.41 0.45 0.38 0.51 0.37 0.50 0.43 0.37 0.64

0.89 0.45 0.32 0.56 0.27 0.46 0.27 0.24 0.72

0.84 1.02 0.46 0.58 0.50 0.66 0.54 0.41 0.52

0.65 0.44 0.95 0.34 0.27 0.37 0.38 0.36 0.65

0.71 0.65 1.03 0.37 0.52 0.47 0.57 0.54 0.48

0.69 0.42 0.95 0.28 0.52 0.49 0.26 0.29 0.75

0.83 0.71 1.16 0.51 0.58 0.59 0.54 0.54 0.72

0.86 0.38 1.02 0.50 0.68 0.40 0.74

0.74 0.31 0.94 0.40 0.58 0.39 0.73 0.34 0.68

1.02 0.90 0.79 0.84 - 0.80 0.89 1.05 0.94 0.90

0.23 0.75

The loops were superimposed using either main-chain atoms (above the diagonal) or Cs (below the diagonal). Residues 51-58 (L5MK16 numbering) were used for the superimposition. With the exception of the L5MK16 diabody, all other antibody fragments are referred to by their PDB entry code [40].

Crystal structure of a diabody Perisic et al.

Fig. 4. Conformation of the H2 loop. (a)The H2 loop of the L5MK16 diabody superimposed on the four-fold averaged 2Fo-F c electrondensity map contoured at 1.0a (drawn with the program 0 [381). (b) The main-chain conformation of the H2 loop of L5MK16 as compared with the canonical class 2 H2 loops from antibody fragments bbj and ljhl. Residues 53 and 54 of the L5MK16 loop have a main-chain conformation very similar to that of 1bbj whereas the main-chain conformation of residues 55 and 56 is more similar to that of 1jhl (L5MK16 residue numbering). Some differences are observed between the L5MK16 structure and the diabody model of Holliger et al. [6] (see Fig. 6b). In the model, the VH domains were brought as close to each other as possible, roughly equivalent to a one to two residue linker, because it had been shown that some diabodies could be assembled without extra residues of linker, by direct coupling of the VH and VL domains [6]. Indeed, diabodies have been assembled even after deletion of a residue from the carboxyl terminus of the VH (or amino terminus of VL) (A Pope, PH Holliger and G Winter, unpublished data). In contrast, the L5MK16 diabody has a five-residue linker. Relative to the model, this allows rotation of the VH domains mainly around an axis passing through the centroids of the interacting VH domains. A tilt of the axis of one of the interacting VH domains with respect to the other is also observed (Fig. 6b). The structure of diabodies with very short linkers may have domain orientations more closely resembling those of the initial model, and work is in progress to determine the structures of such diabodies.

Fig. 5. The H3 loop (residues 98-111) of the L5MK16 diabody (red) superimposed on several H3 loops from other antibody fragments (PDB entry codes: 1hil, magenta; 1igm, yellow; 2f19, cyan; and ldfb, green). The antibody fragments were structurally aligned using framework regions. Fabs. Indeed, the packing between VH domains in a diabody is more extensive (involving approximately twice the surface area) than that between the VH domain and the CH1 domain of an Fab fragment. These VH-VH interactions may help to stabilize the diabody.

Packing interactions between diabodies within the asymmetric unit (VL-VL interactions) The surface area buried by the VL-VL interactions between the two diabodies is quite extensive. The interactions between the VL domain -sheets are such that the orientation of the strands in the interacting sheets ° with respect to each other, whereas make an angle of-O0 within protein domains in general 3-sheets usually pack with an angle of between -20 ° and -50 ° with respect to each other [19]. This interaction buries a total of 2281 A2, or -570 A2 per VL domain. Several other diabodies have been shown to form multimers with molecular masses greater than that of a dimer. However, the results of gel filtration of L5MK16 at 1 mg ml-', a protein concentration lower than that used for crystallization, are consistent with the formation of a dimer rather than the tetramer found in the crystal (data not shown). The linker region The linker region is completely visible in one of the four chains (chain C) in the asymmetric unit (Fig. 7). For the

1221

1222

Structure 1994, Vol 2 No 12 other three chains (A, E and G), the ends of the linker are visible, but the intervening three residues are not. One of the reasons for the differences in linker order could be crystal packing. The linker region is located in two distinct crystal packing environments within the crystal. In one of these environments, the carboxyl terminus of the light chain is near the first residue of the linker. Two of the linkers that are not visible (E and G) are in this environment. The other type of environment has a much larger solvent-occupied cavity in the region of the linker. The linker that is visible (chain C) is in this type of environment as is the linker for chain A. Inspection of the packing environment for linker A reveals no apparent reason for its disorder relative to the linker in polypeptide C. For each of the linkers, the distance between C atoms of the first and last residues (residues that are visible in each of the polypeptides) is -10 A, and the only connectivity possible is that illustrated in Fig. 2.

Fig. 7. The C trace of the polypeptide containing the linker that is visible in the electron-density map (linker shown in pink). The Ca trace of the VL domain is shown in green, and that of the VH domain is shown in yellow. Diabodies as cell cross-linking agents Diabodies have great potential for practical applications [20]. Bispecific diabodies might be used to link two different types of surface antigens, for example, a T-cell receptor-CD3 complex could be cross-linked with a tumour cell antigen to trigger lysis of the tumour cell (see Fig. 8).

Fig. 6. (a) A comparison of the L5MK1 6 diabody conformation (top) with the model for the D1.3 diabody (bottom) that was predicted by Holliger et al. [61. For both structures the view is approximately perpendicular to the two-fold axis relating one polypeptide of the diabody to the other. (b) A schematic illustration of the differences in packing between the predicted structure and the L5MK16 diabody. The domains are represented as spheres with centres located at the centroids of the domains. The solid bars connecting the spheres represent packing interactions within the diabody. The upper pair of diagrams represents the L5MK16 interactions, and the lower pair represents the interactions for the predicted structure. Each of the pairs of diagrams represents two orthogonal views of the interactions.

The distance between antigen-binding sites in the L5MK16 diabody is -65 A (as measured from the middle of the H3 loops). This is less than the distance between antigen-binding sites in an intact antibody (up to 150 A). Nevertheless, this may be sufficient for cell-targeted therapy (for a review, see [20]). In order to link two cells, a cross-linking agent must be able to overcome the repulsive forces between cells. To address the question of whether the distance spanned by a diabody is sufficient to overcome these repulsive forces and successfully cross-link two cells, it is instructive to compare a diabody with the major histocompatibility complex (MHC) that is readily able to link two types of cell. We note that the distance between the

Crystal structure of a diabody Perisic et al. affect the interaction with the antigen as they come close to CDR3 of the VL domain (Fig. 9). As noted above, an additional interaction occurs between VL domains of two diabodies within the L5MK16 crystal that gives rise to a tetramer. Because all four antigen-binding sites in the tetramer are accessible, this may be a feature that could be exploited in some applications.

Biological implications

Fig. 8. Schematic representations of cross-linking a T cell and a target cell (a)through a diabody and (b) through an MHC complex. The size of the MHC complex [39] issimilar to that of the diabody. TCR, T-cell receptor. antigen-binding sites of the diabody is similar to that between the base of the MHC and the bound peptide antigen [21] (Fig. 8). As the MHC of antigen-presenting cells is readily capable of engaging T cells, we predict that diabodies binding to the cell-surface antigen of a target cell should likewise be capable of interacting with T cells. It has been reported that bispecific antibodies can mimic the function of superantigen [22] by binding to both the T-cell receptor and the MHC complex without disturbing the complex. The L5MK16 structure indicates that a suitable diabody may also be able to function as a superantigen mimic because the distance spanned by the diabody is similar to the distance spanned by superantigen [23]. The diabody has antigen-binding domains with conformations that are similar to those of Fabs whose structures have been determined. This means that it is unlikely that a diabody with a five-residue linker would bind antigen differently from its parent antibody. Both the structure analysis of this diabody and experimental data [6] indicate that considerably shorter linkers can be accommodated in a diabody. However, very short linkers may

Fig. 9. The hypervariable region closest to the linker is L3 of the VL domain. The L3 loop is coloured green, the linker is shown in magenta, and the other hypervariable regions are coloured yellow.

Diabodies are small, bivalent or bispecific antibody fragments that can readily be expressed in bacteria. They are composed of two polypeptides, each having a heavy-chain variable domain (VH) linked to a light-chain variable domain (VL). Until now, unequivocal definition of the dimeric interactions within the diabody was not possible. The X-ray structure of the diabody presented here confirms the nature of the dimer that was originally proposed on the basis of molecular modelling. VH and VL domains from different polypeptides are paired to form two antigen-binding sites arranged back-to-back. The two antigenbinding sites of the diabody are at opposite ends of the molecule, allowing them to bind simultaneously even to large protein antigens. Bivalent and bispecific diabodies offer enormous potential for immunodiagnostic and therapeutic applications [20]. Diabodies could play a key role in cell-targeted therapy of cancer and other diseases - either by delivering enzymes, toxins, or drugs to target cells or by bringing an effector cell to a target cell (e.g. diabodies may be used to recruit cytotoxic T cells to tumour cells to trigger tumour cell lysis). Bivalency can also allow diabodies to bind to multimeric antigens with high avidity. The diabody described here has a five amino acid linker between the VH and VL domains. Both the structure analysis of this diabody and experimental

1223

1224

Structure 1994, Vol 2 No 12

data indicate that considerably shorter linkers can be accommodated without loosing antigen binding. By shortening the linker peptide, it may be possible to alter the relationship of the antigenbinding sites to each other. This may influence the stability of the diabodies or bring about cooperativity in binding antigens.

Materials and methods Gene construction A mouse hybridoma cell line MK2-16b that secretes an antibody against a PLC8 1 from rat was generously provided by Chris Dean and Maurizio Valeri (Institute of Cancer Research, Sutton, UK). Total RNA was isolated from 4x10 7 cells [24]. The first strand cDNA was synthesized and the VH and VL domains were then amplified by polymerase chain reaction (PCR) using V-domain specific primers [25,26]. PCR products were purified and reamplified with primers that introduced restriction sites and appended the linker (encoding GGGGS) at the 5'-end of the VL chain. The VH domain was amplified with the VHlBACKSFI15 [27] and VH1FOR-2 [28] primers which contain the SfiI and BstEII sites, respectively. The VL domain was amplified with primer 4 [6], which contains a BstEII site and encodes the five amino acid linker, and a VK4FOR-Not primer mix [26]. After enzyme digestions, the PCR products were cloned by three-way ligation into the vector pUC119SNmyc(His) 6 which had been cut with SfiI and NotI; this vector is a derivative ofpUC119, and contains a pelB leader, a polylinker (with SfiI and NotI sites), a myc peptide tag as in pHEN1 [27] and a (His) 6 tag for affinity purification. The ligation was transformed into E. coli strain HB2151. Both periplasmic lysates and culture supernatants from individual colonies were screened by ELISA for the expression of antibody fragments that bind to PLC81 . Plates were coated with 10 pLg m-1 PLC, 1 in phosphate-buffered saline (PBS). The binding of the fragments to the antigen was detected using monoclonal antibody 9E10, which recognizes the myc peptide tag [28]. The construct was sequenced by the dideoxy nucleotide method using double-stranded DNA and a sequenase II kit (United States Biochemical).

Protein expression and purification Protein was purified from 10 1 of cell culture grown in 2xTY medium (1.6% w/v tryptone, 1% w/v yeast extract, 0.5% w/v NaCI) with 0.1% (w/v) glucose and 0.1 mg ml- l ampicillin at 37°C. At an A600 of 0.8, cells were induced with 1 mM isopropyl -D-thiogalactopyranoside (IPTG) and grown for an additional 20 h at 28°C. Cells were resuspended in buffer containing 50mM sodium phosphate, pH 7.2, 0.1 M NaCI and 1 mM EDTA, centrifuged at 10000 g, and the periplasmic fraction was dialyzed against PBS. Protein was purified by batch absorption to Ni 2 +-NTA (nitrilo-tri-acetic acid) Sepharose CL-6B resin (Qiagen) [10] equilibrated in 50 mM sodium phosphate, pH 7.2, 1 M NaCI and 20 mM imidazole. The resin was washed with the same buffer containing 60 mM imidazole, and the His-tagged diabody was eluted with a buffer containing 200 mM imidazole. After buffer exchange to 20 mM Tris-HC, pH 8.0, the protein was further purified on a MonoQ column (Pharmacia). The procedure yielded 3-5 mg of a pure protein as judged by electrophoresis on 20% polyacrylamide SDS gels and isoelectric focusing on PhastGel IEF 3-9 (Pharmacia). The protein was concentrated with a Centricon 10 (Amicon) to 10 mg ml-1 (determined by BioRad protein assay, with bovine serum albumin as a standard) and stored at 4°C.

Crystallization Crystals were grown in hanging drops at 17°C using the vapor diffusion method [29]. The protein stock was mixed with an equal volume of a precipitant solution containing 20% (v/v) glycerol, 15% (v/v) polyethylene glycol 3000, 0.1 M Tris-HCI, pH 8.8, and 0.2 M sodium acetate, and equilibrated against a reservoir containing the precipitant. Typical crystals had the dimensions 0.3 mmx0.3 mmxO.1 mm. The crystals had P2 1 symmetry with unit cell dimensions a=72.1 A, b=80.7 A, c=88.0 A, 13=99.8 °.

Diffraction data collection All data were obtained from a single crystal. For data collection, the crystal was flash frozen in a nitrogen cold stream at 95 K. Freezing and data collection were carried out with the crystals suspended in a film of freezing solution in a rayon loop. The freezing solution was the same as the precipitant solution used for crystallization. Data were collected with an Enraf-Nonius GX-13 rotating anode CuK a X-ray source operating at 40 kV and 60 mA with a Supper double nickel-coated mirror focusing system and a Mar Research 300 mm image plate scanner. Images were collected with an oscillation angle of 2.0 ° for 60 min. Diffraction data were processed with the program MOS-FLM [30] and the CCP4 package [SERC (UK) Collaborative Computer Project No 4, Daresbury Laboratory, Warrington, UK, 1979]. Scaling the images revealed no discernible decay throughout the period of data collection (90 h). A total of 124443 observations were made for 31023 unique reflections giving an average redundancy of 4. The data were 98% complete to 2.6 A. The overall Rsy m was 0.125.

Structure solution The self-rotation function indicated the presence of two-fold non-crystallographic symmetry. An initial rotation function was calculated using search models derived from a library of 34 antibody fragment structures. The structure yielding the rotation function solution with greatest ratio of peak to rms deviation was chosen for further analysis. This was the structure of the Fab' monoclonal antibody fragment lbbj [15]. The molecular replacement solution was carried out with the program AMORE [31]. The highest eight peaks in the rotation function solution (all >3o) were used for translation function searches. For the translation function, all data from 8.0 A to 3.2 A were used. The third highest peak in the rotation map (8.7tr) resulted in the translation function with the greatest correlation coefficient and highest ratio of peak to rms deviation (11.3(r). This rotation/translation solution was refined as a rigid body and fixed for a second set of translation function searches. The same eight rotation function solutions were employed for the next set of translation functions. The rotation function peak that yielded the highest correlation in the second set of translation function searches (14.7) was the highest peak in the rotation function map (9.2ur). This rotation/translation solution was refined as a rigid body and then fixed along with the first rotation/translation solution for a third set of translation function searches. The translation function with the greatest correlation coefficient and greatest ratio of peak to rms deviation (18.6or) in the third set of translation functions was the fifth highest peak in the rotation function map (8.6u) and differed in rotation by only ~4° from the first rotation function solution employed. All three rotation/translation function solutions were refined simultaneously as rigid bodies and fixed for a fourth set of translation functions. The fourth set of translation function searches yielded a 21.3r translation function solution. Further translation function searches yielded no new solutions.

Crystal structure of a diabody Perisic et al. The four molecules that had been positioned with the rotation/translation functions were refined as eight rigid bodies (VH and VL domains refined independently) with the program X-PLOR yielding a crystallographic R-factor of 0.51 with all data between 30.0 A and 3.0 A. A 3Fo-2Fc electron-density map calculated from the rigid-body refined model indicated clear density for much of the polypeptide but with several breaks in the main-chain density and many poorly defined side chains. The atomic model was used to define one mask for a VH domain and one for its paired VL domain with the program MAMA [32,33]. These masks and non-crystallographic operators derived from least-squares superposition of the rigidly refined models were used for density averaging with the program RAVE [32,33]. Ten cycles of four-fold averaging of the VH and VL domains, and combination of observed amplitudes with phases from the averaged electron-density map were carried out to yield a map that was easily interpretable. The averaging included all data to a resolution limit of 2.6 A. A model for the diabodies was built into the 2Fo-F c map calculated with averaged phases and observed amplitudes. This model was refined with the program X-PLOR [34] (least-squares atomic refinement) and then used to calculate a new mask and new non-crystallographic operators. The mask and operators were then used for another 10 cycles of averaging with RAVE. The R-factor relating the structure-factor amplitudes calculated from the averaged map on the last cycle and the observed amplitudes was 0.19. The model was rebuilt using the 2Fo-F c map resulting from the second round of density averaging. The rebuilt model was further refined with X-PLOR (least-squares minimization) allowing atomic coordinates and temperature factors to vary. The model was then subjected to simulated annealing refinement with X-PLOR [35] followed by least-squares minimization. Difference electron-density maps were examined to locate ordered solvent molecules. Waters were added at positions with density >1.2a in the 3Fo-2F c map and >3cr in the Fo-F c map provided that plausible hydrogen-bond donors or acceptors were found in the vicinity of the position. The final R-factor was 0.20 using all data between 6.0 A and 2.6 A for 7613 non-hydrogen atoms (including 261 solvent atoms). The rms deviations from ideal geometry were 0.18 A, 2.40 and 26.70 for bond distances, bond angles, and dihedral angles respectively. Residue Val56 of the VL domain is in a disallowed region of the Ramachandran map (Fig. 10). This residue is located in the L2 loop and has well-defined electron density. The main-chain dihedral angles ° (0=79, ,=-39 ) are very similar to those that have been observed for the residue at this position in other antibody structures. The average B-factor for all atoms is 15.8 A2 . The average B-factors for main-chain and side-chain atoms are respectively 15.9 A2 and 15.4 A2 . The average solvent B-factor is 21.5 A2. Analysis of the orientation of VH relative to VL To compare the orientation of VH to VL in the diabody with the orientations observed in other antibody fragments, the transformation that optimally superimposes the VH domain of the diabody on to the VH domain of a reference antibody fragment was determined. This transformation was applied to the VH-VL pair in the diabody. The transformation necessary to then superimpose the VL domain of the transformed diabody on to the VL domain of the reference antibody fragment was then calculated. This transformation was decomposed into rotations around three mutually perpendicular vectors. One of these vectors was the vector connecting the centroids of the VH and VL domains that is roughly perpendicular to the interface between the two domains. The arbitrary reference employed in the analysis was molecule 1 of the PDB entry 8FAB. The rotation around this vector was calculated for a set of 26 antibody fragments.

Fig. 10. Ramachandran plot for the four chains in the asymmetric unit of the crystal. The disallowed regions are shaded with the lightest grey. Glycine residues are indicated as triangles, and non-glycine residues as squares. The only residue in the disallowed region is Val56 of the VL domain in each of the chains. Coordinates and structure factors have been deposited with the Protein Data Bank, Brookhaven National Laboratory, (accession number ILMK). Acknouledgements: We thank Matilda Katan for providing us with PLC. 1 used for the ELISAs. We are grateful to Alexey Murzin for his many helpful suggestions and for the analysis of the relative orientations of the VH and VL domains in the diabody. We thank Cyrus Chothia for valuable discussions about antibody structure and Tim Green for comments on the manuscript. The work was supported by a grant from the ZENECA/MRC/DTI LINK programme and British Heart Foundation grant PG/93154 to RLW.

References

1. Skerra, A. & Pluckthun, A. (1988). Assembly of a functional immunoglobulin Fv fragment in Escherichia coil. Science 240,

1038-1041. 2. Glockshuber, R., Malia, M., Pfitzinger, . & Pluckthun, A. (1990). A comparison of strategies to stabilize immunoglobulin Fv-fragments. Biochemistry 29, 1362-1367. 3. Huston, J.S., et al., & Opperman, H. (1988). Protein engineering of antibody-binding sites: recovery of specific activity in an antidigoxin single-chain Fv analogue produced in Escherichia coli. Proc. Natl. Acad. Sci. USA 85, 5879-5883.

4. Bird, R.E., et al., & Whitlow, M. (1988). Single-chain antigen-binding proteins. Science 242, 423-426. 5. Griffiths, A.D., et al., & Winter, G. (1993). Human anti-self antibodies with high specificity from phage display libraries. EMBO J. 12, 725-734. 6. Holliger, P., Prospero, T. & Winter, G. (1993). 'Diabodies': small bivalent and bispecific antibody fragments. Proc. Natl. Acad. Sci. USA

90, 6444-6448. 7. Essig, N.Z., Wood, J.F., Howard, A.J., Raag, R. & Whitlow, M. (1993). Crystallization of single-chain Fv proteins. J. Mol. Biol. 234, 897-901. 8. Kortt, A.A., etal., & Colman, P.M. (1994). Recombinant anti-sialidase single-chain variable fragment antibody. Eur. J. Biochem. 221, 151-157.

1225

1226

Structure 1994, Vol 2 No 12 9. Zdanov, A., et al., & Cygler, M. (1994). Structure of a single-chain antibody variable domain (Fv) fragment complexed with a carbohydrate antigen at 1.7 A resolution. Proc. Natl. Acad. Sci. USA 91, 6423-6427. 10. Hochuli, E., Bannwarth, W., Dobeli, H., Gentz, R. & StiJber, D. (1988). Genetic approach to facilitate purification of recombinant proteins with a novel metal chelate adsorbent. Biotechnology 6, 1321-1325. 11. Kabat, E.A., Wu, T.T., Perry, H.M., Gottesman, K.S. & Foeller, C. (1991). Sequences of Proteins of Immunological Interest. 5th edn. U.S. Department of Health and Human Services, Public Health Services of National Institute of Health. 12. Chothia, C. & Lesk, A.M. (1987). Canonical structures for the hypervariable regions of immunoglobulins. J. Mol. Biol. 196, 901-917. 13. Chothia, C., et al., & Poljak, R.J. (1989). Conformations of immunoglobulin hypervariable regions. Nature 342, 877-883. 14. Chitarra, V., et al,, & Polijak, R.J. (1993). Three-dimensional structure of a heteroclitic antigen-antibody cross-reaction complex. Proc. Natl. Acad. Sci. USA 90, 7711-7715. 15. Brady, R.L., et al., & Low, D.C. (1992). Crystal structure of a chimeric Fab' fragment of an antibody binding tumor cells. J. Mol. Biol. 227, 253-264. 16. Guddat, L.W., Shan, L., Anchin, J.M., Linthicum, D.S. & Edmundson, A.B. (1994). Local and transmitted conformational changes on complexation of an anti-sweetener Fab. J. Mol. Biol. 236, 247-274. 17. Nicholls, A. (1992). GRASP: Graphical Representation and Analysis of Surface Properties. Columbia University, New York. 18. Chothia, C. & Lesk, A. (1988). Elbow motion in the immunoglobulins involves a molecular ball-and-socket joint. Nature 335, 188-190. 19. Chothia, C., Novotny, J., Bruccoleri, R.& Karplus, M. (1985). Domain association in immunoglobulin molecules: the packing of variable domains. J. Mol. Biol. 186, 651-663. 20. Holliger, P. & Winter, G. (1993). Engineering bispecific antibodies. Curr. Opin. Biotech. 4, 446-449. 21. Bjorkman, P.J., Saper, M.A., Samraoui, B., Bennett, W.S., Strominger, J.L. & Wiley, D.C. (1987). Structure of the human class I histocompatibility antigen. Nature 329, 506-512. 22. Yagi, J., et al., & Yammamoto, K. (1994). Superantigen-like properties of an antibody bispecific for MHC class II molecules and the V domain of the T cell antigen receptor. J. Immunol. 152, 3833-3841. 23. Jardetzky, T.S., et al., & Wiley, D.C. (1994). The three-dimensional structure of a human class-lI histocompatibility molecule complexed with superantigen. Nature 368, 711-718. 24. Clackson, T., Gussow, D. & Jones, P.T. (1991). General applications of PCR to gene cloning and manipulation. In PCR: A Practical Approach. (McPherson, M.J., Quirke, P. and Taylor, G.R., eds), pp. 187-214, IRL Press, Oxford.

25. Orlandi, R., Gussow, D.H., Jones, P.T. & Winter, G. (1989). Cloning immunoglobulin variable domains for expression by the polymerase chain reaction. Proc. Natl. Acad. Sci. USA 86, 3833-3837. 26. Clackson, T., Hoogenboom, H.R., Griffiths, A.D., & Winter, G. (1991). Making antibody fragments using phage display libraries. Nature 352, 624-628. 27. Hoogenboom, H.R., Griffiths, A.D., Johnson, K.S., Chiswell, D.J., Hudson, P. & Winter, GC.(1991). Multi-subunit proteins on the surface of filamentous phage: methodologies for displaying antibody (Fab) heavy and light chains. Nucleic Acids Res. 19, 4133-4137. 28. Ward, E.S., Gussow, D., Griffiths, A.D., Jones, P.T. & Winter, G. (1989). Binding activities of a repertoire of single immunoglobulin variable domains secreted from Escherichia coli. Nature 341, 544-546. 29. McPherson, A. (1982). Preparation and Analysis of Protein Crystals. John Wiley & Sons, New York. 30. Leslie, A.G.W., Brick, P. & Wonacott, A.T. (1986). MOSFLM. Daresbury Lab. Inf. Quart. Protein Crystallogr. 18, 33-39. 31. Navaza, J. (1994). AMORE - an automated procedure for molecular replacement. Acta Crystallogr. A 50, 157-163. 32. Kleywegt, G.J. & Jones, T.A. (1994). Halloween - Masks and Bones. In From First Map to Final Model, pp. 59-66, SERC Daresbury Laboratory, Warrington, UK. 33. Kleywegt, G.J. & Jones, T.A. (1993). RAVE Uppsala, Sweden. 34. Bronger, A.T. (1992). X-PLOR Version 3.1. Yale University, New Haven. 35. Br nger, A.T. (1988). Crystallographic refinement by simulated annealing: application to a 2.8 A resolution structure of aspartate aminotransferase. J. Mol. Biol. 203, 803-816. 36. Kraulis, P.J. (1991). MOLSCRIPT: a program to produce both detailed and schematic plots of protein structures. J. Appl. Crystallogr. 24, 946-950. 37. Russell, R.B. & Barton, G.J. (1992). Multiple protein sequence alignment from tertiary structure comparison: assignment of global and residue confidence levels. Proteins 14, 309-323. 38. Jones, T.A., Zou, J.-Y., Cowan, S.W., & Kjeldgaard, M. (1991). Improved methods for building protein models in electron density maps and the location of errors in these models. Acta Crystallogr. A 47, 110-119. 39. Brown, J.H., et al., & Wiley, D.C. (1993). Three-dimensional structure of the human class II histocompatibility antigen HLA-DR1. Nature 364, 33-39. 40. Bernstein, F.C., et al., & Tasumi, M. (1977). The protein data bank: a computer-based archival file for macromolecular structures. J. Mol. Biol. 112, 535-542. Received: 28 Sep 1994; revisions requested: 12 Oct 1994; revisions received: 21 Oct 1994. Accepted: 24 Oct 1994.