“Superhumanized” Antibodies: Reduction of Immunogenic Potential by Complementarity-Determining Region Grafting with Human Germline Sequences: Application to an Anti-CD28¹

Allogeneic bone marrow transplantation is a frequently successful therapy for hematologic malignancies. A serious limitation to the procedure is the fact that the majority of patients experience graft-versus-host disease (GVHD),⁴ resulting in high morbidity and mortality (1, 2, 3). T cells transferred with donor bone marrow have not been exposed to the tolerance mechanisms operative within a recipient’s thymus, hence they can react through their T cell Ag receptors with recipient self-peptide-MHC complexes. Therapeutic Abs directed to T cells hold great promise as a general approach to attacking the GVHD problem. One example is the mAb 9.3, which was raised to human T cells and subsequently used to identify the T cell surface Ag CD28 (4, 5). CD28 functions by receiving a costimulation signal for T cell activation (6) through interaction with the receptors CD80 and CD86 on APC (7). However, CD80 and CD86 can also interact with CTLA4, a molecule structurally related to CD28 that is also found on T cells (8). CTLA4 ligation is thought to produce inhibitory signals (9, 10, 11). Whether T cell Ag recognition leads to activation of the T cell, contributing to GVHD, depends on the balance between the stimulatory process mediated by CD28 and the inhibitory process mediated by CTLA4. Blocking of CD80/86 signals with a soluble CTLA4-Ig fusion protein can produce Ag-specific T cell inactivation (12), which may suppress GVHD and autoimmunity (13, 14). However, it was recently shown that blocking of the costimulus by an anti-CD28-specific Ab in mice prevented GVHD and was more selective and immunosuppressive than CTLA4-Ig (15). The 9.3 Ab may thus prove superior to CTLA4-Ig for treating GVHD. An anti-CD28 Ab may also find use in other immunosuppressive interventions, such as against autoimmune disease or graft rejection. Since this Ab might be used repetitively in the same patient, we sought to reduce its immunogenic potential by constructing a humanized version of 9.3.

Ab humanization was described in 1986, in which complementarity-determining regions (CDRs) were shown to transfer antigenic specificity when grafted from one Ab to another (16). The first clinical use of a humanized Ab revealed a striking absence of adverse reactions (17, 18). The subsequent safety record of humanized Abs in >10⁵ patients makes it clear that humanization has reduced the immunogenicity problem in immunotherapy.

Clark (19) has pointed out that the superiority of humanized Abs over chimeric Abs, in which human Ig constant domains are combined with the unmodified variable domains of a mouse Ab (20, 21, 22), has not formally been proven. Chimeric Abs are also strikingly less immunogenic than fully murine Abs, hence a significant share of the immunogenicity reduction of humanized Abs may be attributed to the replacement of murine constant regions. Both humanized and chimeric Abs can induce human anti-humanized (HAHA) or human anti-chimeric (HACA) Ab responses directed to variable regions. Distinguishing a general superiority through quantitative comparisons of the percentage of patients exhibiting HAHA or HACA can be equivocal, however, as the frequency of these responses depends critically on the treatment context. The variability of the human immune response to therapeutic Abs was shown vividly in a safety trial in which a humanized anti-TNFα was given to healthy volunteers. Cohorts receiving 2 mg/kg and above showed no HAHA, whereas all subjects in the 0.1-mg/kg cohort showed HAHA (23). Were one to compare the immunogenicity of a chimeric anti-TNF Ab using these data from the humanized one, the chimeric, whatever its value, could not possibly score lower than 0% HAHA in one set of subjects and could not possibly be higher than 100% HAHA in the other case.

An ideal comparison of the relative immunogenicity of humanized and chimeric Abs would take the form of quantitative measurement of immune responses to chimeric and humanized versions of the same, preferably nonimmunomodulatory, Ab used identically in therapy. To our knowledge, an experiment of this design has only been done in a murine system. Mice exposed to Igs with humanized variable domains and mouse constant regions developed an anti-variable region response that was absent in controls exposed to fully murine Igs (24). The closest human counterpart to this experiment is the use of humanized and chimeric anti-TNFα Abs, in separate studies, to treat Crohn’s disease. In this study, 7% of patients receiving the humanized anti-TNFα showed HAHA (25), whereas 10% of those receiving the chimeric showed HACA (Remicade prescribing information), despite simultaneous immunosuppression of the latter group with methotrexate. Available evidence, although imperfect, thus supports the original expectation that extension of human motifs into conserved parts of the variable region would reduce the immunogenicity of therapeutic Abs.

The clinical impact of the immunogenicity of humanized Abs runs the gamut from negligible to intolerable. At one end of the immunogenicity range, the humanized anti-HER2/neu Ab Herceptin gave just a 0.5% rate of HAHA in breast cancer patients (26); at the opposite end, the 65% rate of HAHA among colon cancer patients treated with a humanized A33 Ab caused development of this particular humanized Ab to be halted (27).

What causes humanized Abs to be immunogenic, other than the obvious potential for anti-idiotype formation? One possibility, articulated by Clark (19), is that current CDR graft strategies preserve nonhuman motifs. These strategies, depicted in Fig. 1, upper panel, are framework-centered. Early findings evinced a view that the conformation of the CDRs was exquisitely sensitive to the chemical environment of the surrounding framework. A single untoward mutation in the framework could abolish Ag binding or so weaken affinity as to render clinical use of the humanized Ab impractical (17, 28, 29). Thus, avoidance of any perturbation of the CDRs has been an overriding principle in the design of humanized Abs. Accordingly, human framework sequences are usually chosen because they are the database entries most homologous to the frameworks of the murine Ab to be humanized (28). This practice preserves unique somatic mutations in the human framework sequence. Initial CDR grafts often result in loss of binding affinity, which is remedied by further changes of amino acid residues in the framework region from human to murine (17, 28). The end result of retaining homologous somatic mutations and structurally important mouse residues is that the framework of a humanized Ab can become quite mouse-like. For example, a humanized anti-IFNγ Ab and its mouse antecedent are identical in the V_H sequence for 48 contiguous residues (30).

Human Ab genes are formed in vivo by rearrangement of germline gene segments, and unmodified germline sequences are the norm among Abs of the IgM class. Only later in B cell ontogeny does a hypermutation process uniquely tailor the initial sequences to improve recognition of specific Ags, and only later in life do somatically mutated serum Igs become predominant (31). The body thus may be highly tolerant to germline-encoded Abs. An uncomplicated extrapolation is that clinical problems would be reduced if germline sequences were used for constructing humanized Abs (32). However, current CDR graft practices do not favor germline sequences, which only by rare coincidence have a good match to framework regions of a mouse Ab to be humanized. Human germline genes are thus seldom used to make humanized Abs (32, 33, 34).

The “canonical structure” of a CDR is a distinct fold of the polypeptide backbone that is repeated in many Abs (35). Only one or a few canonical structures are possible for each CDR in each chain. In the vast majority of Abs, each CDR will adopt one of these canonical structures and the particular one used can be assigned from sequence alone, without need for computer modeling. The canonical structure class of an Ab is a list of the canonical structures at H chain CDRs 1 and 2 and L chain CDRs 1, 2, and 3. To a first approximation, the canonical structure class determines the gross structure of an Ag combining site. The side chains of CDR residues determine antigenic specificity, but do not alter the geometry imposed by the canonical structure class. For example, canonical structure 4 of H chain CDR 2 and canonical structures 3 and 4 of L chain CDR 1 have extended loops that together create a cleft into which a small molecule can bind. Accordingly, just two canonical structure classes that use these combinations account for the vast majority of anti-hapten Abs (36).

The kernel of our findings in this report is a way that germline sequences can easily be used to make humanized Abs. Our strategy (Fig. 1, lower panel) is based on structural homologies between mouse and human CDRs and essentially ignores the frameworks. The first step of our CDR grafting method is to identify human germline V genes that in combination have the same canonical structure class as the mouse Ab to be humanized. Within that matching subset, typically a half dozen genes of 44 functional V_H or 41 functional V_L genes in the human genome, we pick the H and L chain gene segments whose CDRs have the best residue-to-residue homology to the mouse Ab. In the selected sequences, we simply convert the remaining nonhomologous CDR residues to the mouse Ab sequence. Abs constructed by this design retain the ability to bind Ag. Because they are CDR grafted in a way that minimizes deviation from human sequences, we call such Abs “superhumanized”. In this article, we describe application of this method to humanizing the mouse anti-CD28 Ab 9.3.

Derivation of the Mu9.3 hybridoma has been described previously (12). Hybridoma cells used in the current study were obtained from frozen stocks of Dr. P. Martin (Fred Hutchinson Cancer Research Center, Seattle, WA). The H chain variable region gene of mAb 9.3 was cloned by reverse transcription and PCR three times, and the L chain gene twice. In each case, total cell RNA was extracted by a guanidinium isothiocyanate procedure (37) using Tri-reagent (Sigma-Aldrich, St. Louis, MO). mRNA was then isolated using Oligo-Tex columns (Qiagen, Hilden, Germany). In the case of the L chain, reverse transcription was primed with the constant region-specific oligonucleotide 5′-TAGAAGCTTCTCATTCCTGTTGAAGCTCTTGAC. PCR was then performed with the same primer and a reverse primer corresponding to the N-terminal framework sequence 5′-GACAGAATTCGACATTGAGCTCACCCAGTCTCCA. In two independent clonings of the H chain, reverse transcription was primed with the constant region-specific oligonucleotide 5′-CCCAAGCTTAATTTTCTTGTCCAC. PCR was then performed with the same primer and with a degenerate oligonucleotide corresponding to the N-terminal framework sequence 5′-GACAGAATTCSAGGTSMARCTGCAGSAGTCWGG. In a third experiment, the constant region-specific oligonucleotide 5′-GACAACGCGTCTCAATTTTCTTGTCCACCTTGGTGC was used to prime H chain mRNA for reverse transcription. This oligonucleotide and the signal peptide-specific oligonucleotide 5′-GACAGTGCACATGAAGTTGTGGTTAAACTGGGTTTT were then used for PCR amplification of the H chain variable region. Amplification products were cloned in pUC119 or derivatives with modified multiple cloning sites and sequenced. The coding sequences of Mu9.3 V_H and V_L were reconfigured as a single chain Fv construct. The corresponding protein was shown to bind to the CD28-Ig fusion protein, confirming that the two fragments were correctly cloned (data not shown). Humanized 9.3 V_L (Hu9.3 V_L) and V_H (Hu9.3 V_H) sequences were each generated according to the principles outlined in Results.

The DNA fragments for Hu9.3 V_L and Hu9.3 V_H were generated by overlap extension of a set of six oligonucleotides as described by Ye et al. (38) followed by amplification using two short terminal oligonucleotide primers. The constructs were directly cloned into the vector pT7Blue using the manufacturer’s Perfectly Blunt Cloning protocol (Novagen, Madison, WI). Clones bearing the correct sequence were confirmed by sequencing using dye terminator chemistry.

Different 9.3 Fabs were constructed by assembling Hu9.3 V_H, Hu9.3 V_L, Mu9.3 V_H, Mu9.3 V_L, human C_κ, and human C_H1 sequences in different combinations using PCR. See Table I for details. A hexahistidine sequence was added to the C-terminal end of the H chain for the purpose of purification (39). The constructs were then subcloned into the vector pAK19 (40). Expression in this vector is regulated by the alkaline phosphatase promoter and each chain has its own translational start site and leader sequence. When derepressed under low phosphate conditions, recombinant proteins are directed into the periplasmic space. E. coli K12 strain 39C1 (W3110 fhuA Δ(argF-lac) deoC phoS*), which derepresses at higher phosphate levels, was used as a host strain.

Cells were grown in 400 ml of 2×YT medium overnight, then this culture was used to inoculate 15 L minimal medium (41) in a New Brunswick MPP-30 fermentor. Cells were allowed to grow for 24–30 h with ample aeration and agitation (250 rpm). Cells were harvested, resuspended in binding buffer (50 mM NaH₂PO₄, 500 mM NaCl, and 20 mM imidazole, pH 8.0) and sonicated. Following sonication, the mixture was spun at 16,000 × g for 30 min. The supernatant fraction was saved and spun again to ensure that all insoluble materials had been removed. The lysate was adjusted to pH 8.0 and passed over a 25-ml Ni-NTA column (Qiagen) that had been pre-equilibrated with binding buffer. The column was washed extensively with the binding buffer and eluted with the same buffer supplemented with 250 mM imidazole. The eluted protein was dialyzed overnight against PBS and passed over a streptococcal protein G column that had been pre-equilibrated with PBS (42). After extensive washing with PBS, the protein was eluted with 0.2 M glycine (pH 3.0) and dialyzed overnight against PBS. Each protein preparation was analyzed by SDS-PAGE to check its purity. Preparations purified by this method showed >98% electrophoretic purity, but still contained high levels of endotoxin. (Endotoxin assays were performed by the Biologics Production Shared Resource at our center using a BioWhittaker QCL-1000 Chromogenic Limulus Amebocyte Lysate test kit (BioWhittaker, Walkersville, MD).) Samples to be used for MLR were therefore passed over Detoxi-gel minicolumns (Pierce, Rockford, IL), which reduced this contaminant to <1 endotoxin unit/μg of Fab. Concentrations of each Fab were determined by the absorbance at 280 nm using the extinction coefficient and molecular mass calculated from their amino acid sequences (43).

A CD28-Ig fusion protein has been described previously (44). We made a similar construct that differed from the earlier, primarily in the sequences at the junctions between domains. Our expression construct consisted of the leader sequence and first two mature residues of human CD5 (45); a linker with a KpnI restriction site, encoding the amino acids RVP; the CD28 extracellular domain from residues 1 through 121 (46); and an additional Glu residue introduced as part of a splice site. The CD28 portion of the construct was made by polymerization of synthetic oligonucleotides (38), then subcloned with human IgG1 hinge, CH2, and CH3 exons (47) in the expression vector pcDNA3.1neo (Invitrogen, Carlsbad, CA). The construct was stably transfected into Chinese hamster ovary (CHO)-K1 cells (48), and drug-resistant colonies were screened for recombinant protein production by ELISA using 9.3-coated plates. A high-producing cloned cell line was propagated and grown for protein production in a Cell Factory (Nunc, Rochester, NY). CD28-Ig fusion protein was isolated from spent culture medium by chromatography on protein A-Sepharose.

A 96-well plate was coated with 2 μg/ml human CD28-Ig in the presence of 50 mM carbonate buffer (pH 9.6) overnight at 4°C. The wells were washed three times with 0.05% Tween 20 in PBS and the remaining binding sites in the wells were blocked by incubating with 200 μl of 1% BSA for 1 h. One hundred microliters of different Fabs of varying concentrations in PBS in triplicate was added to each well, followed by an incubation time of 1 h. The wells were washed three times with 0.05% Tween 20 in PBS. One hundred microliters of goat anti-human κ-chain IgG-HRP conjugate (Sigma-Aldrich) was added to each well. After 1 h of incubation, the wells were washed six times and 100 μl of 1 mM 2,2′-azino-bis-3-ethylbenzthiazoline sulfonic acid/4 mM H₂O₂ substrate solution in citrate buffer (0.1 M, pH 4.5) was added for color development. Absorbance at 405 nm was taken after 15 min using a plate reader (Molecular Devices, Menlo Park, CA). The data were processed with the program Igor (WaveMetrics, Lake Oswego, OR) to determine the equivalent concentration for half-maximal binding of the Abs to CD28-Ig (EC₅₀). An empirical four-parameter fitting function was used (49): y = a + (d − a)/{1 + exp (b(c − x))}, where a, b, c, and d are adjustable fitting parameters, x is the logarithm of Ab concentration, and y is the response; in this case absorbance of peroxidase product. The logarithm of EC₅₀ is given by the value of the parameter c for the best fit of the fitting function for the binding isotherm.

CHO cells expressing recombinant human CD80 on their surface were suspended at a density of 20 × 10⁶/ml. Fifty-microliter aliquots of the cell suspension were placed in 3-ml tubes. Twenty-five microliters of recombinant Fab (Chi9.3, Hu9.3, or hybrid Fab 9.3 (Hy9.3)) was added to each tube at concentrations ranging from 0.1 nM to 12 μM, followed by 25 μl of 0.8 mg/ml CD28-Ig-FITC, and allowed to stand on ice for 1 h. Each tube was washed twice with HBSS containing 2% heat-inactivated horse serum and 5 mM sodium azide. The cells in each tube were then resuspended in 0.3 ml of 1% paraformaldehyde and analyzed on the FACScan flow cytometer (BD Biosciences, Mountain View, CA). The mean fluorescence for each sample was determined and plotted to determine the blocking isotherms for each Ab using the flow cytometry analysis program CellQuest (BD Biosciences). EC₅₀ values for the blocking of CD28-Ig binding was determined as described above.

Human PBMC were prepared by density gradient centrifugation on Ficoll-Hypaque. The cells were resuspended in RPMI 1640 medium containing 1 U/ml penicillin, 1 μg/ml streptomycin, and 10% heat-inactivated FBS. A total of 5 × 10⁴ responder cells from one individual was mixed with 5 × 10⁴ irradiated stimulator cells (3000 rad) from an unrelated individual in each well of round bottom 96-well plates. Fabs were added and the plates were incubated at 37°C in a 5% CO₂ atmosphere for 6 days. Cultures were then pulse labeled with 1 μCi of [³H]thymidine and harvested 18 h later. Assays were performed in triplicate and data are reported as mean cpm of three replicates.

Multiple independent isolation of Mu9.3 V_H and V_L genes by RT-PCR is described above. The two L chain and three H chain clonings all gave consistent sequences. Our sequence determined for the Mu9.3 L chain agrees with a published sequence (50) and is clearly different from both the productively and unproductively rearranged MOPC21 L chain genes found in NS1 cells, the fusion partner used for generating the 9.3 hybridoma (51, 52). Differences in our H chain sequence from the published version of 9.3 (50) affect three coding positions. V_H residues 47, 76, and 82a are W, S, and N in our sequence, and C, G, and K, respectively, in the published version. We have no explanation for this discrepancy. Cells used to derive the respective sequences can be traced to the same hybridoma stock and recombinant molecules with either sequence are apparently functional. Positions 76 and 82a are quite variable in known mouse sequences, and there is no a priori reason to favor one sequence over the other. A Cys at position 47 is more remarkable. Cysteine residues are rare in Igs outside the standard positions for disulfide formation. In contrast, tryptophan at position 47 has an essential role in forming the V_H–V_L interface (53), and this residue is >98% conserved in most mouse V_H subgroups (54).

We first sought to identify human V genes whose hypervariable loops have the same canonical structures as 9.3. Six canonical structures have been described for the hypervariable loop in CDR1 of mouse κ-chains (55). The loop portion of Mu9.3 Vκ CDR1 has five residues, therefore is likely to adopt canonical structure 5. Sequences with this canonical structure are not represented in the human germline, but six human Vκ genes have canonical structure 4, with six residue loops, and two have canonical structure 3, with seven residue loops. Both canonical structures 3 and 4 resemble canonical structure 5 (56) and were considered further. Only a single canonical structure is known for the second hypervariable loop in human and mouse κ-chains. Six canonical structures have been described for the hypervariable region of CDR3 in κ-chains. Mu9.3 Vκ CDR3 has six residues in its hypervariable loop and residue 95 is a Pro, and therefore is likely to adopt canonical structure 1. All eight of the human Vκ genes previously considered also have canonical structure 1 in CDR2 and canonical structure 1 in CDR3.

We next sought to rank the eight human germline Vκ genes according to residue-to-residue homology within CDR2 and CDR3. CDR1 was not considered because the differences in length make residue-to-residue comparisons ambiguous. Of the eight Vκ genes, the best matches in CDR3 were A2 and B3 (57), which had three of seven residues identical to Mu9.3. B3 had superior homology in CDR2 (four of seven identical), hence it was chosen as a starting point for the L chain design. The 5′-encoded Tyr residue of human Jκ2 (58) matched the corresponding position of Mu9.3 exactly, hence this germline fragment was used. Residues in Kabat CDRs of B3 that did not already match Mu9.3 were changed to match the Mu9.3 CDRs, with one exception. Residue 34, which probably is not critical to Ag recognition, was left identical to B3. In addition, a glycosylation motif that appears in Mu9.3 at positions 70–72 was retained, but no other changes were made to the human germline sequences of B3 and Jκ2. The final sequence of the Hu9.3 L chain, as shown in Fig. 2, has 21 mutational differences from the human germline, all but 2 of which are in the CDRs.

In designing a superhumanized 9.3 V_H, we again began by matching the canonical structures of CDR1 and CDR2 of Mu9.3 with the corresponding canonical structures in human germline V_H genes. The canonical structures of H chain CDR3 are notoriously difficult to predict (59) and some segments of CDR3 are not encoded in the germline. We therefore did not consider CDR3 in our choice of germline V_H genes for constructing Hu9.3. Three canonical structures have been described for the first hypervariable loop in Ig H chains (55). This loop in Mu9.3 (V_H positions 26–32) has six residues, with a hydrophobic side chain at position 29, suggesting that it adopts canonical structure 1. Four canonical structures have been described for the hypervariable loop in H chain CDR2. This region in Mu9.3 (positions 52–56) has five residues, including Gly at position 55, suggesting that it adopts canonical structure 1 (55). Eleven human V_H genes follow this particular combination of canonical structures (60). Of these, gene DP-45 has two of five and three of five residues in hypervariable loops 1 and 2 identical to Mu9.3, greater than any of the other genes, hence our humanized H chain design started from DP-45. We made no effort to match human D segment sequences. Of the human J_H segments, J_H4 had the closest homology to the C-terminal end of CDR3 in Mu9.3, hence it was used in the construction (61). The final sequence of the Hu9.3 H chain (Fig. 2) has 18 mutational differences from the human germline, all of which are in the CDRs.

The Hu9.3 V_H and V_L were configured as a Fab by fusion with the human IgG1 CH1 and human Cκ constant domain genes, respectively. As a control, Mu9.3 V_H and V_L were fused to the same human constant domains to yield a chimeric molecule (see Table I). Initial binding studies revealed a moderate reduction of avidity of the humanized Fab in comparison to Chi9.3. A hybrid Fab construct was therefore created (Hy9.3), consisting of Hu9.3 V_L and Mu9.3 V_H fused to the same human constant domain genes. Each of the three Fabs was expressed in E. coli. These were isolated without need for in vitro refolding and purified by affinity chromatography on Ni-NTA and protein G-Sepharose resins. All three purified proteins were found to be homogenous by SDS-PAGE.

The ability of the three Fabs to bind to CD28 was examined by ELISA. CD28-Ig-coated plates were incubated with Fab solutions at concentrations ranging from 1 pM to 10 μM. Binding was then assayed with an anti-human κ immunoconjugate. The binding isotherms generated were fit to an empirical equation to determine the respective Fab concentrations at half-maximal binding (49). This analysis, shown in Fig. 3, indicated that the Chi9.3 Fab had the best binding activity, with an EC₅₀ of 20 nM. The EC₅₀ of Hu9.3 was 630 nM, showing a significant reduction of binding activity. The EC₅₀ of Hy9.3 was 30 nM, close to that of Chi9.3, showing that most of the reduction in binding by Hu9.3 could be attributed to weakened interactions involving the H chain.

CHO cells stably transfected with the human CD80 gene and expressing this protein on their surface were obtained from J. A. Ledbetter (Pacific Northwest Research Institute, Seattle, WA). Binding of FITC-labeled CD28-Ig to CD80-CHO cells was quantified by flow cytometry, as shown in Fig. 4. Varying concentrations of the Fabs were mixed with CD28-Ig-FITC before incubation with CD80-CHO cells and were found to inhibit binding of the fluorescently tagged ligand. We generated blocking isotherms of the three Abs over a concentration range of 0.1 nM to 20 μM and extracted the EC₅₀ values from the analysis of these isotherms. The Chi9.3 Fab exhibited the best blocking activity followed by Hy9.3 and Hu9.3, with EC₅₀ values of 0.2, 0.4, and 1.2 μM, respectively. The order and relative ratio of EC₅₀ values show that this assay is consistent with the data from the previous section showing that Chi9.3 Fab exhibited the most binding activity followed by Hy9.3 Fab and Hu9.3 Fab. This inhibition confirms that Hu9.3 prevents CD80-CD28 ligation. A control humanized Ab Fab recognizing an irrelevant Ag, lysozyme (29), did not inhibit, showing that blocking of the CD80-CD28-Ig interaction is restricted to the anti-CD28 activity of the Fabs.

This bioassay tested whether Hu9.3 has biological activity comparable to that of the two other constructs. At 10 μg/ml, Hu9.3, Chi9.3, and Hy9.3 Fabs inhibit MLR by 30, 37, and 35%, respectively (Table II). The negative control, humanized antilysozyme (HuLys) Fab, had essentially no effect. We tested the dose dependence of MLR inhibition by these Fabs in the experiment shown in Fig. 5. The MLR was conducted in the presence of concentrations of Fabs ranging from 0.01 to 5 μg/ml, and in all three cases inhibition was seen above 0.1 μg/ml. Again, the negative control HuLys Fab had no effect.

Our two goals in this study were to provide a reagent for intervention in GVHD and to explore a general approach to Ab humanization that potentially could lower the in vivo immunogenicity of humanized Abs. The preceding sections detail the initial step of the first goal, creation of a humanized anti-CD28 Ab in the form of a Fab. Our previous work indicates that an Ab with this specificity is effective in preventing GVHD in a mouse model (15). We show above that Hu9.3 binds CD28, blocks ligation of CD28 to CD80, and acts as an immunosuppressant in vitro, in that it inhibits replication of responder T cells in the MLR. Hu9.3 therefore possesses the qualitative biological properties of the parent mouse Ab and shows potential as a clinical reagent for GVHD. Hu9.3 differs from murine 9.3 in showing a 30-fold lower Ag avidity, an observation whose structural significance we discuss below. Whether the shift to lower avidity compromises in vivo use may depend on the particular clinical application. A higher avidity version of Hu9.3 derived by a program of mutagenesis and selection (62) would likely have greater potency. Reconstruction of Hu9.3 Fab as an IgG would likely improve pharmacokinetics.

Our second goal amounts to a “reinvention” of humanized Abs, this time assigning the reduction of sequence differences between the humanized molecule and human germline Ig genes a higher priority than maintenance of strong Ag binding. Here, we consider the extent to which balancing these two structural goals affected overall success in creating a clinical reagent.

Because no effort was made to fabricate humanized frameworks homologous to mouse frameworks, nor were “critical” mouse framework residues retained, the reduction in nonhuman motifs in the superhumanized molecule is substantial. As shown in Table III, Hu9.3 has fewer mutational differences from germline genes than conventionally CDR-grafted Abs, a total of 31 over both chains. This number would have been lower still, 29, had the glycosylation motif been eliminated. Glycosylation and the glycosylation motif peptide sequence do not affect Ag binding (P. Tan, unpublished data). By comparison, the conventionally humanized Abs in Table III have 35–37 mutational differences from the closest human germline sequences.

Superhumanizing the L chain caused an insignificant avidity loss, the EC₅₀ value of 30 nM in the half-humanized Hy9.3 being close to EC₅₀ value of 20 nM in the fully murine Chi9.3. The remaining factor of 20 loss of avidity attributable to superhumanizing the 9.3 H chain would appear to compare poorly with humanized Abs in clinical use, which generally have avidity within a factor of 4 of the parental murine Ab. However, making a humanized Ab is usually a two-step process, the first of which is design and creation of the initial CDR-grafted construct, as we have presented here. The initial construct almost always incurs a loss of avidity, so a second step of in vitro affinity maturation follows. Comparing final, affinity-matured humanized Abs is therefore not an appropriate way to assess how different approaches to the initial CDR graft affect avidity. Many first constructs did not detectably bind Ag at all (58, 59, 60). Others bound weakly, even when designed with homology matching (61). Of the market-approved Abs in Table III, Campath-1H had a 39-fold avidity loss in its initial form and Herceptin had 83-fold. The initial form of the humanized anti-lysozyme Ab HuLys showed a 70-fold affinity loss (29), despite the availability at the time of its construction of a crystal structure of the parent Ab in complex with Ag. Thus, when Hu9.3 is compared with other initial constructs, its 30-fold avidity loss is moderate.

The ease with which human germline genes chosen without regard for homology to framework sequences were used for CDR grafting of mouse 9.3 has a structural implication. The accepted model of plastic CDRs embedded in a rock-solid framework is not necessarily invalid in light of our findings, but is an unsafe axiom. Indeed, crystallographic studies of HuLys showed that self-correcting mechanisms made the conformation of CDRs resistant to changes induced by the shift to a human framework (63, 64). Furthermore, mutations at supposedly critical framework positions failed to elicit crystallographically detectable changes in the CDRs (65). However, CDRs undeformable regardless of framework is not an apt model either. Harris and coworkers (32) showed this most generally in their “fixed framework” method of CDR grafting; NEWM and REI frameworks were used to humanize at least eight Abs, for which each new set of CDRs required unique mutational remodeling of the framework to retain Ag binding at high avidity. Superhumanizing may “work” because its axiom lies in a middle ground: CDRs are deformable and are not compatible with every framework set, but CDRs are compatible with those frameworks that have evolved to support other CDRs with the same canonical backbone structure.

We thank Jeffrey A. Ledbetter for providing the CD80-CHO cell line and Nancy McFarland and James R. Swartz for providing E. coli strain 39C1.