(Received for publication, December 20, 1996, and in revised form, February 7, 1997)
From the Departments of Protein Engineering and
§ Immunology, Genentech, Inc., South San Francisco,
California 94080
Antibody humanization often requires the replacement of key residues in the framework regions with corresponding residues from the parent non-human antibody. These changes are in addition to grafting of the antigen-binding loops. Although guided by molecular modeling, assessment of which framework changes are beneficial to antigen binding usually requires the analysis of many different antibody mutants. Here we describe a phage display method for optimizing the framework of humanized antibodies by random mutagenesis of important framework residues. We have applied this method to humanization of the anti-vascular endothelial growth factor murine monoclonal antibody A4.6.1. Affinity panning of a library of humanized A4.6.1 antibody mutants led to the selection of one variant with greater than 125-fold enhanced affinity for antigen relative to the initial humanized antibody with no framework changes. A single additional mutation gave a further 6-fold improvement in binding. The affinity of this variant, 9.3 nM, was only 6-fold weaker than that of a murine/human chimera of A4.6.1. This method provides a general means of rapidly selecting framework mutations that improve the binding of humanized antibodies to their cognate antigens and may prove an attractive alternative to current methods of framework optimization based on cycles of site-directed mutagenesis.
Monoclonal antibodies (mAbs)1 have enormous potential as therapeutic agents; however, most mAbs are derived from murine or other non-human sources, which severely limits their clinical efficacy. In addition to the immunogenicity of rodent mAbs when administered to humans (1-3), further limitations arise from weak recruitment of effector function (4, 5) and rapid clearance from serum (6, 7). As a means of circumventing these deficiences, the antigen-binding properties of murine mAbs can be conferred to human antibodies through a process known as antibody "humanization" (4, 8). A humanized antibody contains the amino acid sequences from the six complementarity-determining regions (CDRs) of the parent murine mAb, which are grafted onto a human antibody framework. For this reason, humanization of non-human antibodies is commonly referred to as CDR grafting. The low content of non-human sequence in humanized antibodies (~5%) has proven effective in both reducing the immunogenicity and prolonging the serum half-life in humans (7, 9).
Unfortunately, simple grafting of CDR sequences often yields humanized antibodies that bind antigen much more weakly than the parent murine mAb (4, 10-16), and decreases in affinity of up to several hundredfold have been reported (13-15). To restore high affinity, the antibody must be further engineered to fine tune the structure of the antigen-binding loops. This is achieved by replacing key residues in the framework regions of the antibody variable domains with the matching sequence from the parent murine antibody. These framework residues are usually involved in supporting the conformation of the CDR loops (17), although some framework residues may themselves directly contact the antigen (18). Chothia and Lesk (19) first noted the importance of certain framework residues to CDR conformation, and a comprehensive examination of all framework residues that can affect antigen binding was conducted by Foote and Winter (20). These authors tabulated a list of some 30 "Vernier" residues that can potentially contribute to CDR structure. Although higher antigen affinity would likely result from editing the entire set of Vernier residues within a humanized antibody to match the corresponding parent murine sequence, this is not generally desirable given the increased risk of immunogenicity imposed by adding further elements of murine sequence. Thus, from a therapeutic standpoint it is preferable to confine framework changes to the minimum set that affords a high affinity humanized antibody.
As shown in the literature (4, 10-16), the task of identifying framework residues that affect antigen binding is the most challenging and time-consuming aspect of antibody humanization. Although these critical framework residues are generally derived from the large set proposed by Foote and Winter (20), a small set of changes will usually suffice to optimize binding, although these often differ from one humanized antibody to the next. Whereas a molecular model of the humanized antibody can provide valuable guidance, the proper combination of mutations must be determined by experiment. Typically, this is achieved by preparing a panel of mutants with "suspect" framework residues replaced by their murine counterparts. These variants are each tested for antigen binding, and favorable mutations that enhance affinity are combined into a final variant.
As a means of simplifying antibody humanization, a number of different approaches have been developed. A common method is to select the human framework most homologous in sequence to that of the murine antibody of interest (10-12). In this way, the number of mismatches between the humanized and parent murine framework is minimized, and it becomes less likely that a key framework residue will need to be replaced. However, it should be noted that even a single incorrect framework residue can have a dramatic effect on antigen binding (11). An alternative humanization strategy was originally proposed by Padlan (21): here, rather than grafting the CDRs onto a human framework, the entire murine variable domains are retained and select framework residues are replaced with human sequence. Only those residues on the surface of the antibody framework, as defined by exposure to solvent, are targeted for replacement. These residues are considered unimportant for antigen binding but most likely to contribute to any potential immunogenicity. On the other hand, buried residues are retained as the murine sequence since these are most likely to affect the structure (and thus antigen binding) of the antibody but are of low immunogenicity since they are in effect "hidden" from the immune system. Although this approach has been successfully used to humanize two different murine mAbs (22, 23), it remains to be seen whether these antibodies are indeed nonimmunogenic, since even buried residues can invoke a T-cell immune response if presented as a denatured peptide by a class II major histocompatability complex molecule (24).
In contrast to these other techniques, an alternative approach is to
humanize antibodies using only a single human framework, regardless of
the sequence of the parent murine antibody. This method has been
successfully used to humanize a number of murine antibodies (13-15,
25) using a framework derived from consensus sequences of the most
abundant human subclasses, namely VL subgroup I
(VL
I) and VH subgroup III
(VHIII) (26). Use of the most common human VL
and VH frameworks minimizes any potential immunogenicity of
the humanized antibody and also eliminates possible idiosyncracies associated with any one particular framework. In addition, this framework has been demonstrated to give good yields of antibody when
expressed recombinantly in either Escherichia coli or
eukaryotic expression systems, an important consideration for
antibodies that are to go into large scale development for clinical
applications. Based on the collective data from these and other
unpublished antibody humanizations, we have observed that the framework
residues that most often influence antigen binding are consistently
derived from a set of only 11 residues. We have incorporated this
information into the design of a random mutagenesis approach to
antibody humanization. By randomizing this small set of critical
framework residues and by monovalent display of the resultant library
of antibody molecules on the surface of filamentous phage (27, 28),
optimal framework sequences can be identified via affinity-based
selection. Here we describe the application of this approach to the
humanization of the murine antibody A4.6.1 (29, 30), an antibody that
binds to vascular endothelial growth factor (VEGF).
The
murine anti-VEGF mAb A4.6.1 blocks VEGF receptor binding and has been
described previously (29, 30). The first Fab variant of humanized
A4.6.1, hu2.0, was constructed by site-directed mutagenesis using a
deoxyuridine-containing template of plasmid pAK2 (13) that codes for a
human VLI-C
1 light chain and human VHIII-CH1
1 heavy chain Fd
fragment. The transplanted A4.6.1 CDR sequences were chosen according
to the sequence definition of Kabat et al. (26), except for
CDR-H1, which we extended to encompass both sequence and structural
(19) definitions, viz. VH residues 26-35. The
Fab encoding sequence was subcloned into the phagemid vector phGHamg3
(27, 28). This final construct, pMB4-19, encodes the initial humanized
A4.6.1 Fab, hu2.0, with the carboxyl terminus of the heavy chain fused
precisely to the carboxyl portion of the M13 gene III coat protein.
pMB4-19 is similar in construction to pDH188, a previously described
plasmid for monovalent display of Fab fragments (31). Notable
differences between pMB4-19 and pDH188 include a shorter M13 gene III
segment (codons 249-406) and use of an amber stop codon immediately
following the antibody heavy chain Fd fragment. This permits expression
of both secreted heavy chain or heavy chain-gene III fusions in
supE suppressor strains of E. coli.
E. coli strain 34B8, a nonsuppressor, was
transformed with phagemid pMB4-19 or variants thereof. Single colonies
were grown overnight at 37 °C in 5 mL of 2YT medium containing 50 µg/mL carbenicillin. These cultures were diluted into 200 mL of AP5
medium (32) containing 20 µg/mL carbenicillin and incubated for
26 h at 30 °C. The cells were pelleted at 4000 × g and frozen at 20 °C for at least 2 h. Cell
pellets were then resuspended in 5 mL of 10 mM Tris-HCl (pH
7.6) containing 1 mM EDTA, shaken at 4 °C for 90 min,
and centrifuged at 10,000 × g for 15 min. The
supernatant was applied to a 1-mL streptococcal protein G-Sepharose
column (Pharmacia Biotech Inc.) and washed with 10 mL of 10 mM MES (pH 5.5). The bound Fab fragment was eluted with 2.5 mL of 100 mM acetic acid and immediately neutralized with
0.75 mL of 1 M Tris-HCl, pH 8.0. Fab preparations were
buffer-exchanged into PBS and concentrated using Centricon-30
concentrators (Amicon). Typical yields of Fab were ~1 mg/L culture
after protein G purification. Purified Fab samples were characterized
by electrospray mass spectrometry, and concentrations were determined
by amino acid analysis.
The humanized A4.6.1 phagemid library was constructed by site-directed mutagenesis according to the method of Kunkel et al. (33). A derivative of pMB4-19 containing TAA stop triplets at VH codons 24, 37, 67, and 93 was prepared for use as the mutagenesis template (all sequence numbering is according to Kabat et al. (26)). This modification was to prevent subsequent background contamination by wild type sequences. The codons targeted for randomization were 4 and 71 (light chain) and 24, 37, 67, 69, 71, 73, 75, 76, 78, 93, and 94 (heavy chain).
To randomize heavy chain codons 67, 69, 71, 73, 75, 76, 78, 93, and 94 with a single mutagenic oligonucleotide, two 126-mer oligonucleotides
were first preassembled from 60-mer and 66-mer fragments by
template-assisted enzymatic ligation. Briefly, 1.5 nmol of 5
phosphorylated oligonucleotide 503-1 (GAT TTC AAA CGT CGT
ACT TCT AGA GAC AAC TCC AAA AAC ACA
TAC CTG CAG ATG AAC) or 503-2 (GAT TTC AAA CGT CGT
ACT TCT GAC TCC ACA TAC CTG CAG
ATG AAC) were combined with 1.5 nmol of 503-3 (AGC CTG CGC GCT GAG GAC
ACT GCC GTC TAT TAC TGT TAC CCC CAC TAT TAT GGG)
(randomized codons underlined; N = A/G/T/C; W = A/T; B = G/T/C; D = G/A/T; R = A/G; Y = C/T). 1.5 nmol of
template oligonucleotide (CTC AGC GCG CAG GCT GTT CAT CTG CAG GTA),
with complementary sequence to the 5
ends of 503-1/503-2 and the 3
end of 503-3, was added to hybridize each end of the ligation junction.
Taq ligase and ligase buffer (New England Biolabs) were added, and the
reaction mixture was subjected to 40 rounds of thermal cycling,
(95 °C for 1.25 min, 50 °C for 5 min) to cycle the template
oligonucleotide between ligated and unligated junctions. The product
126-mer oligonucleotides were purified on a 6% urea, TBE
polyacrylamide gel and extracted from the polyacrylamide in buffer. The
two 126-mer products were combined in equal ratio, ethanol
precipitated, and finally solubilized in 10 mM Tris-HCl, 1 mM EDTA. The mixed 126-mer oligonucleotide product was
designated 504-01.
Randomization of select framework codons (VL, 4 and 71; VH, 24, 37, 67, 69, 71, 73, 75, 76, 93, and 94) was effected in two steps. First, VL randomization was achieved by preparing three additional derivatives of the modified pMB4-19 template. Framework codons 4 and 71 in the light chain were replaced individually or pairwise using the two mutagenic oligonucleotides GCT GAT ATC CAG ACC CAG TCC CCG and TCT GGG ACG GAT ACT CTG ACC ATC. Deoxyuridine-containing template was prepared from each of these new derivatives. Together with the original template, these four constructs coded for each of the four possible light chain framework sequence combinations (see Table I).
|
Oligonucleotides 504-1, a mixture of two 126-mer oligonucleotides (see above), and CGT TTG TCC TGT GCA TCT GGC TAT ACC TTC ACC AAC TAT GGT ATG AAC TGG CGT CAG GCC CCG GGT AAG were used to randomize heavy chain framework codons using each of the four templates just described. The four libraries were electroporated into E. coli XL-1 Blue cells (Stratagene) and combined. The total number of independent transformants was estimated at >1.2 × 108, approximately 1500-fold greater than the maximum number of DNA sequences in the library.
Phagemid particles displaying the humanized A4.6.1 Fab fragments were propagated in E. coli XL-1 Blue cells. Briefly, cells harboring the randomized pMB4-19 construct were grown overnight at 37 °C in 25 mL of 2YT medium containing 50 µg/mL carbenicillin and approximately 1010 M13KO7 helper phage (34). Phagemid stocks were purified from culture supernatants by precipitation with a saline polyethylene glycol solution and resuspended in 100 µL of PBS (~1014 phagemid/mL)
Selection of Humanized A4.6.1 Fab VariantsPurified VEGF1-121 (100 µL at 10 µg/mL in PBS) was coated onto a microtiter plate well overnight at 4 °C. The coating solution was discarded and this well, in addition to an uncoated well, was blocked with 6% skim milk for 1 h and washed with PBS containing 0.05% Tween 20. 10 µL of phagemid stock, diluted to 100 µL with 20 mM Tris (pH 7.5) containing 0.1% bovine serum albumin and 0.05% Tween 20, was added to each well. After 2 h the wells were washed and the bound phage eluted with 100 µL of 0.1 M glycine (pH 2.0) and neutralized with 25 µL of 1 M Tris, pH 8.0. An aliquot of this was used to titer the number of phage eluted. The remaining phage eluted from the VEGF-coated well were propagated for use in the next selection cycle. A total of eight rounds of selection was performed, after which time 20 individual clones were selected and sequenced (35).
Determination of VEGF Binding AffinitiesAssociation and dissociation rate constants for binding of humanized A4.6.1 Fab variants to VEGF1-121 were measured by surface plasmon resonance (36) on a Pharmacia BIAcore instrument. VEGF1-121 was covalently immobilized on the biosensor chip via primary amino groups. Binding of humanized A4.6.1 Fab variants was measured by flowing solutions of Fab in PBS, 0.05% Tween 20 over the chip at a flow rate of 20 µL/min. Following each binding measurement, residual Fab was stripped from the immobilized ligand by washing with 5 µL of 50 mM aqueous HCl at 3 µL/min. Binding profiles were analyzed by nonlinear regression using a simple monovalent binding model (BIAevaluation software version 2.0, Pharmacia).
An initial humanized A4.6.1
Fab fragment was constructed (hu2.0) (Fig. 1) in which
the CDRs from A4.6.1 were grafted onto a human
VLI-VHIII framework. All other residues in
hu2.0 were maintained as the human sequence. Binding of this variant to
VEGF was so weak that it was undetectable. Based on the relative
affinity of other weakly binding humanized A4.6.1 variants (data not
shown), the Kd for binding of hu2.0 was estimated at
>7 µM. This contrasts with an affinity of 1.6 nM for a chimeric Fab construct consisting of the intact
VL and VH domains from murine A4.6.1 and human
constant domains. Thus, binding of hu2.0 to VEGF was reduced at least
4000-fold relative to the chimera. The absence of detectable VEGF
binding by variant hu2.0 highlights the important role of the framework
in properly structuring the CDR loops for antigen binding.
Design of Antibody Library
Based on the cumulative results
from humanizing a number of murine antibodies onto a human
VLI-VHIII framework (13-15, 25), we have
observed that most framework changes, as required to optimize antigen
binding, are limited to some subset of the residues shown in Table
I and Fig. 2. We reasoned that this
information could serve as the basis for a combinatorial approach to
selecting required framework mutations. Accordingly, we designed a
combinatorial approach to the humanization of mAb A4.6.1 by randomizing
these key framework residues and displaying the resultant library of Fab variants on the surface of filamentous phage.
Each of residues 4 and 71 in the light chain and 24, 37, 67, 78, and 93 from the heavy chain were partially randomized to allow the selection
of either the murine A4.6.1, human VLI-VHIII sequence, or sequences commonly found in other human and murine frameworks (Table I). Note that randomization of these residues was not
confined to a choice between the human
VL
I-VHIII consensus or A4.6.1 framework
sequences. Rather, inclusion of additional amino acids commonly found
in other human and murine framework sequences allows for the
possibility that additional diversity may lead to the selection of
tighter binding variants.
Additional heavy chain framework residues were randomized in a binary fashion according to the human VHIII and murine A4.6.1 framework sequences. Residues VH 71, 73, 75, and 76 are positioned in a hairpin loop adjacent to the antigen-binding site. The side chains of VH 71 and 73 are largely buried in canonical antibody structures, and their potential role in shaping the conformation of CDR-H2 and CDR-H3 is well known (11, 13, 25, 37). On the other hand, although the side chains of VH 75 and 76 are exposed to solvent (Fig. 2), it has nevertheless been observed that these two residues can also influence antigen binding (15), presumably due to direct antigen contact in some antibody-antigen complexes. Because of their proximity in sequence and possible interdependence, we randomized VH 71, 73, 75, and 76 en bloc so that only two possible combinations of this tetrad could be selected: either all human VHIII or all murine A4.6.1 sequence. Finally, VH residues 69 and 94 were randomized, but only to represent the VHIII and A4.6.1 sequences. We have not replaced VH 69 and 94 in previous antibody humanizations, but because they differ between the VHIII consensus and A4.6.1 sequences (Fig. 1) and have been noted as potentially important for proper CDR conformation (20), we opted to include them in the randomization strategy.
Humanized A4.6.1 Fab Library Displayed on the Surface of PhagemidA variety of systems have been developed for the
functional display of antibody fragments on the surface of filamentous
phage (38). These include the display of Fab fragments or single chain variable fragments as fusions to either the gene III or gene VIII coat
proteins of M13 bacteriophage. We opted to use a system similar to that
described by Garrard et al. (31) in which a Fab fragment is
monovalently displayed as a gene III fusion (Fig. 3).
This system has two notable features: unlike single chain variable fragments, Fab fragments have no tendency to form dimeric species, the
presence of which can prevent selection of the tightest binders due to
avidity effects. Second, the monovalency of the displayed protein
eliminates a second potential source of avidity effects that would
otherwise result from the presence of multiple copies of a protein on
each phagemid particle (27, 28).
A concern in designing the humanized A4.6.1 phagemid library was that residues targeted for randomization were widely distributed across the VL and VH sequences. Limitations in the length of synthetic oligonucleotides require that simultaneous randomization of each of these framework positions can be achieved only through the use of multiple oligonucleotides. However, as the total number of oligonucleotides increases, the efficiency of mutagenesis decreases (i.e. the proportion of mutants obtained that incorporate sequence derived from all of the mutagenic oligonucleotides). To circumvent this problem, we incorporated two features into our library construction. The first was to prepare four different mutagenesis templates coding for each of the possible VL framework combinations. This was simple to do given the limited diversity of the light chain framework (only four different sequences) but was beneficial in that it eliminated the need for two oligonucleotides from the mutagenesis strategy. Second, we preassembled two 126-base oligonucleotides from smaller synthetic fragments. This made possible randomization of VH codons 67, 69, 71, 73, 75, 76, 93, and 94 with a single long oligonucleotide rather than two smaller ones. The final randomization mutagenesis strategy therefore employed only two oligonucleotides simultaneously onto four different templates.
Selection of Tight Binding Humanized A4.6.1 Fab VariantsVariants from the humanized A4.6.1 Fab phagemid library were selected based on binding to VEGF. Enrichment of functional phagemid, as measured by comparing titers for phage eluted from a VEGF-coated versus uncoated microtiter plate well, increased up to the seventh round of affinity panning. After one additional round of sorting, 20 clones were sequenced to identify preferred framework residues selected at each position randomized. These results, summarized in Table II, revealed strong consensus among the clones selected. Ten out of the 20 clones had the identical DNA sequence, designated hu2.10. Of the 13 framework positions randomized, eight substitutions were selected in hu2.10 (VL 71 and VH 37, 71, 73, 75, 76, 78, and 94). Interestingly, residues VH 37 (Ile) and 78 (Val) were selected neither as the human VHIII or murine A4.6.1 sequence. This result suggests that some framework positions may benefit from extending the diversity beyond the target human and parent murine framework sequences, although we have not tested this by making the appropriate mutants.
|
There were four other unique amino acid sequences among the remaining 10 clones analyzed: hu2.1, hu2.2, hu2.6, and hu2.7. All of these clones, in addition to hu2.10, contained identical framework substitutions at positions VH 37 (Ile), 78 (Val), and 94 (Lys) but retained the human VHIII consensus sequence at positions 24 and 93. Four clones had lost the light chain coding sequence and did not bind VEGF when tested in a phage enzyme-linked immunosorbent assay (39). We have occasionally noted the loss of heavy or light chain sequence with other Fab phagemid libraries,2 and these clones are presumably selected for on the basis of enhanced expression. Such artifacts can often be minimized by reducing the number of sorting cycles or by propagating libraries on solid media.
Expression and Binding Affinity of Humanized A4.6.1 VariantsPhage-selected variants hu2.1, hu2.2, hu2.6, hu2.7, and hu2.10 were expressed in E. coli using shake flasks, and Fab fragments were purified from periplasmic extracts by protein G affinity chromatography. Recovered yields of Fab for these five clones ranged from 0.2 (hu2.6) to 1.7 mg/L (hu2.1). The affinity of each of these variants for antigen (VEGF) was measured by surface plasmon resonance on a BIAcore instrument (Table III). Analysis of these binding data revealed that the consensus clone hu2.10 possessed the highest affinity for VEGF out of the five variants tested. Thus, our Fab phagemid library was selectively enriched for the tightest binding clone. The calculated Kd for hu2.10 was 55 nM, at least 125-fold tighter than for hu2.0, which contains no framework changes (Kd > 7 µM). The other four selected variants all exhibited weaker binding to VEGF, ranging down to a Kd of 360 nM for the weakest (hu2.7). Interestingly, the Kd for hu2.6, 67 nM, was only marginally weaker than that of hu2.10, and yet only one copy of this clone was found among 20 clones sequenced. This may have been due to a lower level of expression and display, as was the case when the soluble Fab of this variant was expressed.
|
Despite
the large improvement in antigen affinity over the initial humanized
variant, binding of hu2.10 to VEGF was still 35-fold weaker than that
of a chimeric Fab fragment containing the murine A4.6.1 VL
and VH domains. This considerable difference suggested that
further optimization of the humanized framework might be possible
through additional mutations. Of the Vernier residues identified by
Foote and Winter (20), only residues VL 46, VH
2, and VH 48 differed in the A4.6.1 versus human
VLI-VHIII framework (Fig. 1), but these were
not randomized in our phagemid library. A molecular model of the
humanized A4.6.1 Fv fragment showed that VL 46 sits at the
VL-VH interface and could influence the
conformation of CDR-H3. Furthermore, this amino acid is almost always
leucine in most VL
frameworks (26) but is valine in A4.6.1. Accordingly, a Leu
Val substitution was made at this position in the background of hu2.10. Analysis of binding kinetics for
this new variant, hu2.10V, indicated a further 6-fold improvement in
the Kd for VEGF binding, demonstrating the
importance of valine at position VL 46 in antibody A4.6.1.
The Kd for hu2.10V (9.3 nM) was thus
within 6-fold that of the chimera. In contrast to VL 46, no
improvement in the binding affinity of hu2.10 was observed for
replacement of either VH 2 or VH 48 with the
corresponding residue from murine A4.6.1.
A combinatorial approach to antibody humanization has a distinct advantage over the conventional site-directed mutagenesis-based method in that a much larger pool of mutants can be explored simultaneously, with the best binders being rapidly selected from a phage display library based on relative antigen affinity. We have applied such an approach to humanization of the murine mAb A4.6.1, a neutralizing anti-VEGF antibody known to inhibit mitogenic signaling and potentially useful as an anti-tumor therapeutic (30). A library of framework-randomized variants was sorted, and from this a number of variants were identified that bound VEGF considerably tighter than the straight CDR graft variant hu2.0 (Tables II and III). The tightest binding variant in the final pool was the consensus clone hu2.10, and incorporation of one additional mutation at VL 46 yielded a humanized A4.6.1 variant with an affinity within 6-fold that of the chimeric Fab.
Although hu2.10V contained nine framework changes, these did not
contribute equally to the improvements in antigen binding. The data in
Table III provide some insight into the relative contributions that
individual framework changes made to antigen binding. As noted
previously, a 6-fold improvement in antigen affinity was associated
with replacing residue VL 46 (leucine) with the
corresponding amino acid found in A4.6.1 (valine). Interestingly, part
of the improvement in affinity was due to an increase in the
association rate constant, suggesting that VL 46 may play a
role in preorganizing the antibody structure into a conformation more
suitable for antigen binding. Other mutations that affected antigen
affinity were primarily due to changes in the dissociation rate
constant for binding. Comparison of hu2.1 and hu2.10 reveals a 5-fold
improvement in affinity for substitution of VH residues 71, 73, 75, and 76 with the A4.6.1 sequence. Conversion of VL
71 to the A4.6.1 sequence (Phe Tyr) had negligible effect on
binding (hu2.2 versus hu2.7), whereas variants with leucine
at VL 4 bound marginally worse (<2-fold) than those with
methionine, the naturally occurring residue in both the A4.6.1 and
human VL
I frameworks (hu2.2 versus hu2.1). Comparison of other humanized A4.6.1 variants not shown here revealed that the VH 94 Arg
Lys change resulted in a 5-fold
improvement in Kd, either due to direct antigen
contact by this residue or to a structural role in maintaining the
proper conformation of CDR-H3. Curiously, variant hu2.6 has three
sequence differences relative to the consensus clone hu2.10, but
nevertheless has a similar Kd. This suggests that
these three substitutions have little effect on antigen binding.
Although the negligible effect of conservative changes at
VL 4 and 71 concurs with binding data for other variants,
it is somewhat surprising that the more drastic change at
VH 67 (Phe
Thr) has little effect on binding.
Recently, Rosok et al. (40) reported a combinatorial
approach to the humanization of murine mAb BR96. However, the method used in that work differs from that presented here in several key
features. For reasons enunciated earlier, we used the same human
VLI-VHIII framework that has been used in
previous antibody humanizations (13-15, 25), whereas Rosok et
al. selected a novel human framework based on homology to the
sequence of the parent murine mAb. Furthermore, we humanized the entire
framework region (except for CDRs), then randomized only those
framework residues that have been empirically found to be important to
antigen binding. In contrast, Rosok et al. humanized only
surface-exposed residues, then randomized the remaining buried residues
that differed in sequence between the murine mAb and human template.
Thus, the respective sets of residues randomized were almost entirely
different, reflecting the use of dissimilar humanization
strategies.
We have demonstrated that phage display methods can be readily applied to the humanization of a murine monoclonal antibody. Application to the humanization of mAb A4.6.1 led to selection of a humanized variant that bound VEGF greater than 125-fold tighter than the variant with no framework changes. One additional mutation at VL 46 within the background of the best phage selected clone was required to improve the affinity to within 6-fold that of murine A4.6.1. It should be noted, however, that VL 46 does not usually require replacement in antibody humanizations given the overwhelming predominance of leucine at this position in most antibody frameworks (26). Given this consideration, we feel that the general phage method described will usually suffice for the selection of tight binding humanized antibodies. Finally, the success of this humanization, in combination with previous results (13-15, 25), again illustrates the feasibility of using a single human framework as a generic scaffold for humanized antibodies.
We thank Marcus Ballinger for assistance with the oligonucleotide preassembly, James Bourell for mass spectrometric analyses, Allan Padua for amino acid analyses, and the oligonucleotide synthesis group.