US 5,639,641 AGrant
Resurfacing of Rodent Antibodies
Issue Date:1997-06-17
•16 Claims
•15 Drawing Sheets
Abstract
The invention relates to a method for providing, via resurfacing, humanized rodent antibodies that have improved therapeutic efficacy due to the presentation of a human surface in the variable region. In the method, (1) position alignments of a pool of antibody heavy and light chain variable regions is generated to give a set of heavy and light chain variable region framework surface exposed positions wherein the alignment positions for all variable regions are at least about 98% identical; (2) a set of heavy and light chain variable region framework surface exposed amino acid residues is defined for a rodent antibody (or fragment thereof); (3) a set of heavy and light chain variable region framework surface exposed amino acid residues that is most closely identical to the set of rodent surface exposed amino acid residues is identified; (4) the set of heavy and light chain variable region framework surface exposed amino acid residues defined in step (2) is substituted with the set of heavy and light chain variable region framework surface exposed amino acid residues identified in step (3), except for those amino acid residues that are within 5 .ANG.ngstroms of any atom of any residue of the complementarity determining regions of the rodent antibody; and (5) the humanized rodent antibody having binding specificity is produced.
Metadata
Assignee
- Immunogen Inc.
Inventors
- Jan T. Pedersen
- Stephen M. J. Searle
- Anthony R. Rees
- Michael A. Roguska
- Braydon C. Guild
Application Information
Application Number:US 9422455
Filing Date:1992-09-09
Priority Date:1992-09-09
Art Unit:186
Classifications
IPC:
C12N 1500C07K 1600A61K 39395
Field of Search:
530435387.3;387.7;388.869.6;172.1
Patent Drawings (15 sheets)
Description
FIELD OF THE INVENTION
The present invention relates to the development of prediction rules that can be used to accurately model the variable regions (V-regions) of antibodies. The development of these rules and their application in the predictive molecular restructuring of the surfaces of variable domains of non-human monoclonal antibodies enables changing of the surface, i.e., resurfacing, of these monoclonal antibody V-regions to replicate the surface characteristics found on human antibody V-regions. This method of resurfacing non-human monoclonal antibody V-regions to resemble human antibody V-regions is expected to permit the production of functional altered antibodies, which retain the binding parameters of the original non-human monoclonal antibody, with improved therapeutic efficacy in patients due to the presentation of a human surface on the V-region.
BACKGROUND OF THE INVENTION
General Background of Antibodies
Murine monoclonal antibodies are widely used as diagnostic and therapeutic agents in the treatment of human disease. Mice can be readily immunized with foreign antigens to produce a broad spectrum of high affinity antibodies. Invariably, the introduction of murine or other rodent antibodies into humans results in the production of a human anti-mouse antibody (HAMA) response due to the presentation of a foreign protein in the body. The production of HAMA in patients can result from the introduction of foreign antibody in a single dose or from extended use in therapy, for example, for the treatment of cancer. Extended use of murine antibody is generally limited to a term of days or weeks in patients before concerns of anaphylaxis arise. Moreover, once HAMA has developed in a patient, future use of murine antibodies for diagnostic or therapeutic purposes is often precluded for the same reasons.
Beyond ethical considerations, attempts to produce human monoclonal antibodies have not been highly successful for a number of reasons. The production in vitro of human monoclonals rarely results in high affinity antibodies. In vitro cultures of human lymphocytes yield a restricted range of antibody responses relative to the broad spectrum of reactive antibodies produced in vivo through direct immunization of mice. Additionally, in humans, immune tolerance prevents the successful generation of antibodies to self-antigens. All of these factors have contributed to the search for ways to modify the structures of murine monoclonal antibodies to improve their use in patients. Many investigators have attempted to alter, reshape or humanize murine monoclonal antibodies in an effort to improve the therapeutic application of these molecules in patients.
Strategies of Antibody Humanization
The earliest reports of the controlled rearrangement of antibody domains to create novel proteins was demonstrated using rabbit and human antibodies as described by Bobrzecka, K. et al. (Bobrzecka, K., Konieczny, L., Laidler, P. and Rybarska, J. (1980), Immunology Letters 2, pp. 151-155) and by Konieczny et al. (Konieczny, L., Bobrzecka, K., Laidler, P. and Rybarska, J. (1981), Haematologia 14 (I), pp. 95-99). In those reports, the protein subunits of antibodies, rabbit Fab fragments and human Fc fragments, were joined through protein disulfide bonds to form new, artificial protein molecules or chimetic antibodies.
Recombinant DNA technology was used to construct gene fusions between DNA sequences encoding mouse antibody variable light and heavy chain domains and human antibody light chain (LC) and heavy chain (HC) constant domains to permit expression of the first recombinant "near-human" antibody (chimeric antibody) product (Morrison, S. L., Johnson, M. J., Herzenberg, L. A. and Oi, V. T. (1984), Proc. Natl. Acad. Sci. U.S.A. 81, pp. 6851-6855).
The kinetics and immune response in man to chimeric antibodies has been examined (LoBuglio, A. F., Wheeler, R. H., Trang, J., Haynes, A., Rogers, K., Harvey, E. B., Sun, L., Ghrayeb, J. and Khazaeli, M. B. (1989), Proc. Natl. Acad. Sci. 86, pp. 4220-4224).
Chimeric antibodies contain a large number of non-human amino acid sequences and are immunogenic in man. The result is the production of human anti-chimera antibodies (HACA) in patients. HACA is directed against the murine V-region and can also be directed against the novel V-region/C-region (constant region) junctions present in recombinant chimeric antibodies.
To overcome some of the limitations presented by the immunogenicity of chimeric antibodies, the DNA sequences encoding the antigen binding portions or complementarity determining regions (CDR's) of murine monoclonal antibodies have been grafted by molecular means in the DNA sequences encoding the frameworks of human antibody heavy and light chains (Jones, P. T., Dear, P. H., Foote, J., Neuberger, M. S. and Winter, G. (1986), Nature 321, pp. 522-525; Riechmann, L., Clark, M., Waldmann, H. and Winter, G. (1988), Nature 332, pp. 323-327). The expressed recombinant products called reshaped or humanized antibodies are comprised of the framework of a human antibody light or heavy chain and the antigen recognition portions, CDR's, of a murine monoclonal antibody. Several patent applications have been filed in this area including, for example, European Patent Application, Publication No. 0239400; European Patent Application, Publication Nos. 0438310A1 and 0438310A2; International Patent Publication No. WO 91/09967; and International Patent Publication No. WO 90/07861.
However, it is questionable whether European Patent Application (EP), Publication No. 0239400 is truly enabling. It is not assured in this patent that the best fit is made to assure proper presentation of the CDR loops at the antibody combining site.
EP Publication Nos. 0438310A1 and 0438310A2 go a step beyond EP Publication No. 0239400 by protecting the importance of uniquely selected human frameworks for the human light chain (LC) and heavy chain (HC) V-regions. These V-region frameworks should show a high degree of sequence similarity with the frameworks of the murine monoclonal antibody and present the CDR's in the appropriate configuration. However, the criteria for sequence matching are no more sophisticated than simple homology searching of the antibody protein or DNA databases.
International Patent Publication No. WO 91/09967 attempts a further variation of the method disclosed in EP Publication No. 0239400. In International Patent Publication No. WO 91/09967, homology of the donor sequences and the acceptor framework is not important, rather it discloses that a selected set of residues in the LC and HC are critically important to humanization. The ability to make changes at these positions is the basis of International Patent Publication No. WO 91/09967.
International Patent Publication No. WO 90/07861 proposes four important criteria for designing humanized antibodies. 1) Homology between human acceptor and non-human donor sequences. 2) Use donor rather than acceptor amino acids where the acceptor amino acid is unusual at that position. 3) Use donor framework amino acids at positions adjacent to the CDR. 4) Use donor amino acids at framework positions where the sidechain atom is within 3 Angstroms of the CDR in a 3-D model. The first antibody humanized by this method retained less than 1/3 the affinity of the original monoclonal antibody.
None of the above methods for designing a humanized antibody is predictable due to the questions that surround CDR framework interactions. By replacement of murine framework with human framework, there is no guarantee of identical conformations for CDR's because i) the V.sub.L -V.sub.H interaction is not identical in all V-regions and ii) accurate prediction of the CDR-framework interactions are key to faithful reproduction of the antigen binding contacts.
The above methods do not offer a general solution to solving the issues surrounding antibody humanization, rather the methods as outlined in each reference above involve a substantial amount of trial and error searching to obtain the desired affinity in the final humanized product. More importantly, there is no guarantee that corrective changes in framework amino acids will leave the reshaped V-regions resembling the surface character of a truly human antibody. Therefore, it can be argued that antibodies humanized by the above methods may be immunogenic in man.
Antigenicity of Antibodies
The antigenicity/immunogenicity of an antibody, including recombinant reshaped antibody products, introduced into humans can be viewed as a surface phenomenon. In general one can view the immune system as scanning the surface of a protein introduced to the body. If the F.sub.V portion of a humanized antibody `opens-up` in the circulation then internal residues can be presented to the immune system. On the other hand, if the F.sub.V portion is stable and tightly packed then only the surface residues presented by the V-regions and the interface between the V.sub.L and V.sub.H regions will be `scanned`.
Surface Reshaping or Resurfacing of Antibodies
The notion of surface presentation of proteins to the immune system raises the prospect of redesigning murine monoclonal antibodies to resemble human antibodies by humanizing only those amino acids that are accessible at the surface of the V-regions of the recombinant F.sub.V. The resurfacing of murine monoclonal antibodies to reduce their immunogenicity could be beneficial in maintaining the avidity of the original monoclonal antibody in the reshaped version, because the natural framework-CDR interactions are retained. The value of maintaining the integrity of the framework-CDR interactions has been illustrated as summarized below.
In a recent research report, two different reshaped versions of the rat monoclonal antibody, Campath-9 (anti-human CD4), were generated (Gorman, S. D., Clark, M. R., Routledge, E. G., Cobbold, S. P. and Waldmann, H. (1991), Proc. Natl. Acad. Sci. U.S.A. 88, pp. 4181-4185). In one version, pV.sub.H NEW/C.sub.G1, the acceptor V.sub.H framework was from the human NEW-based heavy chain, which has 47% identical residues to the Campath-9 V.sub.H. While in the second version, pV.sub.H KOL/C.sub.G1, the acceptor V.sub.H framework was from the human KOL antibody, which has 72% identical residues to Campath-9 V.sub.H. Each reshaped antibody contained the identical V.sub.L domain from the human REI antibody sequence. However, the recombinant product of pV.sub.H KOL/C.sub.G1 had an avidity for CD4 that was substantially greater than the product of pV.sub.H NEW/C.sub.G1. The authors proposed a reshaping strategy where human sequences, that are highly homologous to the rodent antibody of interest, are transferred, by in vitro mutagenesis, into the rodent V-region to create a "bestfit" reshaped antibody. This strategy uses the term "bestfit" to describe the modeling process, however, there is no quantitative formula employed to assess "bestfit", and so in effect, the process is subjective. Additionally, there is no resurfacing concept presented in that paper.
The concept of reducing rodent-derived antibody immunogenicity through the replacement of exposed residues in the antibody framework regions which differ from those of human origin is discussed in a recent paper (Padlan, E. A. (1991), Molecular Immunology 28, pp. 489-498). In that paper, the variable domains of two antibody structures, KOL (human) and J539 (mouse), are examined. The crystal structures of the Fab fragments of these two antibodies have been elucidated to high resolution. The solvent accessibility of the exposed framework residues in the variable domains of these two antibodies were compared to a sequence database of human and murine antibody V-region subgroups. On the basis of his findings, Padlan proposed to reduce the antigenicity of allogeneic variable domains [murine V-regions], through replacement of the exposed residues in the framework regions with residues usually found in human antibodies. In murine sequences with the highest similarity to a given human sequence, the number of changes necessary to "humanize" a murine V-region surface would range from 6-15 amino acid changes per V-region. This reference suggests how to convert one antibody surface into another but no general method is developed. Application of the procedure is provided by two examples, a worst-case and a best-case.
Worst Case:
Among the representative murine kappa V.sub.L sequences examined for which its autologous V.sub.H has been sequenced, S107V.sub.L has the most residues that need to be replaced to humanize it. S107V.sub.L is most similar to the members of the human subgroup VKIV and JK2. The exposed or partially exposed residues that need to be replaced are those at positions 9, 10, 14, 15, 16, 17, 18, 22, 41, 63, 80, 85, 85, 100 and 106. Murine V-region S107V.sub.H is most similar in its framework to the members of the human subgroup VHIII and JH6. The exposed or partially exposed residues in S107V.sub.H that need to be replaced are those at positions 3, 40, 68, 73, 75, 76, 82b and 89. A total of 23 residues need to be replaced to humanize the variable domains of S107.
Best Case:
Among the murine V.sub.H sequences examined for which the autologous V.sub.L has also been sequenced, MOPC21V.sub.H has the least number of residues that need to be replaced to humanize it. MOPC21V.sub.H is most similar in its framework to the members of the human subgroup HIII and JH6. The exposed or partially exposed residues that need to be replaced are those at positions 1, 42, 74, 82a, 84, 89 and 108. MOPC21V.sub.L is most similar in its framework to human subgroup VKIV and JK4. The exposed or partially exposed residues that need to be replaced are those at positions 1, 9, 12, 15, 22, 41, 63, 68, 83 and 85. A total of 17 amino acids need to be replaced to humanize the variable domains of MOPC21.
Of the light chains in the Best- and Worst-Case examples cited above, S107V.sub.L required changes at 15 positions and MOPC21V.sub.L required changes at 10 positions. Only seven of the changes are common to both of these light chain sequences (see underlined residues). Moreover, of the heavy chain residues that need to be replaced to humanize the respective V-regions, S107V.sub.H required changes at 8 positions and MOPC21V.sub.H required changes at 7 positions. In this instance, only one position is common to both of these heavy chain sequences (see residues in boldface).
An analysis of S107 V-regions alone would not have led to the prediction of which residues to change in MOPC21. The reason for this is that the surface residues in Padlan's analysis are only determined by reference to the crystal structure analysis of one antibody. In addition, the basis for defining the surface exposure of an amino acid at a particular position on that crystal structure is a continuous gradient of change, e.g., the fractional solvent accessibility values (Padlan, E. A. (1990), Molecular Immunology 28, pp. 489-498) were computed, where: 0 to 0.2=completely buried, 0.2 to 0.4=mostly buried, 0.4 to 0.6=partly buried/partly exposed, 0.6 to 0.8=mostly exposed, and 0.8 or above=completely exposed. By limiting the analysis of exposed surface residues to a single crystal structure and by superimposing a broad range of solvent accessibility ratios on exposed residues, such a modeling strategy could be expected to have a wide margin of error in its calculations. This model fails to take into account the great majority of structural information available in the database for other antibody crystal structures.
SUMMARY OF THE INVENTION
Accordingly, it is an object of this invention to provide humanized rodent antibodies or fragments thereof, and in particular, humanized rodent monoclonal antibodies that have impinged therapeutic efficacy in patients due to the presentation of a human surface on the V-region. This and other objects have been attained by providing a method for determining how to humanize a rodent antibody or fragment thereof by resurfacing the method comprising:
(a) determining the conformational structure of the variable region of the rodent antibody or fragment thereof by constructing a three-dimensional model of the rodent antibody variable region;
(b) generating sequence alignments from relative accessibility distributions from x-ray crystallographic structures of a sufficient number of rodent and human antibody variable region heavy and light chains to give a set of heavy and light chain framework positions wherein the alignment positions are identical in 98% of the sufficient number of rodent antibody heavy and light chains;
(c) defining for the rodent antibody or fragment thereof to be humanized a set of heavy and light chain surface exposed amino acid residues using the set of framework positions generated in step (b);
(d) identifying from human antibody amino acid sequences a set of heavy and light chain surface exposed amino acid residues that is most closely identical to the set of surface exposed amino acid residues defined in step (c), wherein the heavy and light chain from the human antibody are or are not naturally paired;
(e) substituting, in the amino acid sequence of the rodent antibody or fragment thereof to be humanized the set of heavy and light chain surface exposed amino acid residues defined in step (c) with the set of heavy and light chain surface exposed amino acid residues identified in step (d);
(f) constructing a three-dimensional model of the variable region of the rodent antibody or fragment thereof resulting from the substituting specified in step (e);
(g) identifying, by comparing the three-dimensional models constructed in steps (a) and (f), any amino acid residues from the sets identified in steps (c) or (d), that are within 5 Angstroms of any atom of any residue of the complementarity determining regions of the rodent antibody or fragment thereof to be humanized; and
(h) changing any residues identified in step (g) from the human to the original rodent amino acid residue to thereby define a rodent antibody humanizing set of surface exposed amino acid residues; with the proviso that step (a) need not be conducted first, but must be conducted prior to step (g).
Also provided is a method for producing a humanized rodent antibody or fragment thereof from a rodent antibody or fragment thereof, the method comprising:
(I) carrying out the above-described method for determining how to humanize a rodent antibody or fragment thereof by resurfacing; and
(II) modifying the rodent antibody or fragment thereof by replacing the set of rodent antibody surface exposed amino acid residues with the rodent antibody humanizing set of surface exposed amino acid residues defined in step (h) of the above-described method.
In a preferred embodiment, the rodent antibody or fragment thereof is a murine antibody, and most preferably murine antibody N901.
BRIEF DESCRIPTION OF THE FIGURES
FIG. 1 shows an algorithm that can be used for constructing a three-dimensional model of the rodent antibody variable region.
FIG. 2 is a diagram showing the approach to determine how to humanize a rodent antibody or fragment thereof according to the present invention.
FIGS. 3A SEQ ID NOS: 1-12 and 3B SEQ ID NOS: 13-24 are plots of relative accessibility of amino acid residues for twelve antibody F.sub.V structures, mapped onto the sequence alignment of these structures. Structures Glb2 (Jeffrey, P. D., Doctor of Philosophy Thesis, University of Oxford, United Kingdom, 1991), D1.3 (Amit, A. G., Mariuzza, R. A., Phillips, S. E. V. and Poljak, R. J. (1986), Science 233, pp. 747-753), 3D6 (Grunow, R., Jahn, S., Porstman, T., Kiessig, T., Steinkeller, H., Steindl, F., Mattanovich, D., Gurtler, L., Deinhardt, F., Katinger, H. and von R., B. (1988), J. Immunol. Meth. 106, pp. 257-265) and 36-71 (5fab) (Rose, D. R., Strong, R. K., Margolis, M. N., Gefter, M. L. and Petsko, G. A. (1990), Proc. Natl. Acad. Sci. U.S.A. 87, pp. 338-342) are not yet present in the Brookhaven database. The other structures used were: 2hfl (Sheriff, S., Silverton, E. W., Padlan, E. A., Cohen, G. H., Smith-Gill, S. J., Finzel, B. C. and Davies, D. R. (1987), Proc. Natl. Acad. Sci. U.S.A. 84, pp. 8075-8079), 3hfm (Padlan, E., Silverton, E., Sheriff, S., Cohen, G., Smith-Gill, S. and Davies, D. (1989), Proc. Natl. Acad. Sci. U.S.A. 86, pp. 5938-5942), 2fbj (Mainhart, C. R., Potter, M. and Feldmann, R. J. (1984), Mol. Immunol. 21, pp. 469-478), 3fab (Saul, F. A., Amzel, L. M. and Poljak, R. J. (1978), J. Biol. Chem. 253, pp. 585-597), 4fab (Herron, J., He, X., Mason, M., Voss, E. and Edmunson, A. (1989), Proteins: Struct., Funct., Genet. 5, pp. 271-280), 2mcp (Segal, D., Padlan, E., Cohen, G., Rudikoff, S., Potter, M. and Davies, D. (1974), Proc. Natl. Acad. Sci. U.S.A. 71, pp. 4298), 2fb4 (Marquart, M. Deisenhofer, J. and Huber, R. (1980), J. Mol. Biol. 141, pp. 369-391), and 1f19 (Lascombe, M. Alzari, P., Boulot, G., Salujian, P., Tougard, P., Berek, C., Haba, S., Rosen, E., Nisonof, A. and Poljak, R. (1989), Proc. Natl. Acad. Sci. U.S.A. 86, p. 607). These structures are designated by their Brookhaven entry code. The sequence numbering used here is described in FIGS. 4A and 4B. FIG. 3A graphically shows the relative accessibility for the heavy chain and FIG. 3B graphically shows the relative accessibility for the light chain.
FIGS. 4A SEQ ID NOS: 25-31 and 4B SEQ ID NOS: 32-38 show alignments of sequences generated using the three methods of humanization. Sequences are: 1) Original rodent N901. 2+3) KOL (Marquart, M. Deisenhofer, J. and Huber, R. (1980), J. Mol. Biol. 141, pp. 369-391) and reshaped N901 using KOL surface. 4+5) Most homologous sequences, L(KV2F) (Klobeck, H., Meindl, A., Combriato, G., Solomon, A. and Zachau, H. (1985), Nucleic Acids Res. pp. 6499-6513) and H(G36005) (Schroeder, H. and Wang, J. (1990), Proc. Natl. Acad. Sci. U.S.A. 87), and reshaped N901 using these sequences. 6+7) Most homologous with respect to surface residues, L(KV4B) (Klobeck, H., Bronkamp, G., Combriato, G., Mocikat, R., Pohelnz, H. and Zachau, H. (1985), Nucleic Acids Res. 3, pp. 6515-6529) and H(PLO123) (Bird, J., Galili, N., Link, M., Sites, D. and Sklar, J. (1988), J. Exp. Med. 168, pp. 229-245), and reshaped N901 using these sequences. The numbering is the same as used in the antibody modelling program ABM (trademark for commercial software, Oxford Molecular Ltd., Oxford, U.K.), which is based on structural conservation and not sequence homology as used by Padlan et al. (Kabat, E. A., Wu, T. T., Reid-Miller, M., Perry, H. M. and Gottesman, K. S. (1987), Sequences of Proteins of Immunological Interest. U.S. Department of Health and Human Services, Fourth Edition). The sequence changes which have to be introduced in order to resurface N901 with a given sequence are marked with bars, back-mutations as determined from F.sub.V models are marked with stars. The sequence homology of given sequences to N901 are shown in brackets after each sequence.
FIG. 5 is a stereo plot of mean antibody .beta.-barrel, coordinates determined by iterative multiple fitting of eight antibody structures. Strands 7 and 8 comprise the `take off` positions for CDR H3 and are not included in the fitting of V.sub.L and V.sub.H regions.
FIG. 6 is a plot of RMS deviation (RMSD) from the mean of the eight .beta.-sheet strands comprising the framework. The RMSD was calculated from structures F19.9, 4-4-20, NEW, FBJ, KOL, HyHEL-5, HyHEL-10 and McPC603. N,C.alpha.,C atoms are included in the plot. The residues used are shown in the alignment (Table 2). The most disordered residues are all the residues of strand HFR4, the last residue of LFR1, and the first and last residue of HFR2. The nomenclature of the strands is explained in the alignment in Table 2. LFR1-#1, LFR2-#2, LFR3-#3, LFR4-#4, HFR1-#5, HFR2-#6, HFR3-#7, HFRS4-#8.
FIG. 7 is a flowchart of the overall modelling protocol known as CAMAL.
FIG. 8 is a plot of superimposed loop backbones for models and x-ray structures discussed in Example 2. The loops are positioned after global framework fit. This does not represent the best local least squares fit, but shows how the loops are positioned globally onto the framework.
FIGS. 9A to 9D are stereo (N,C-.alpha.,C,O) representations of crystal structures and models of D1.3, 3671 and Gloop-2 variable domain and .beta.-barrel strands described in Example 2. Crystal structures are shown with open bonds, model with solid bonds. The difference between the 3D6-H3 in the model and the crystal structure is due to a 5.degree.-7.degree. twist in the extended .beta.-sheet conformation of this loop, FIG. 9A: D1.3, FIG. 9B: 36-71, FIG. 9C: Gloop-2, FIG. 9D: 3D6.
FIG. 10 is a histogram showing the distribution of loop length for CDR H3 loops, data from Kabat et al. (Kabat, E. A., Wu, T. T., Reid-Miller, M., Perry, H. M. and Gottesman, K. S. (1987), Sequences of Proteins of Immunological Interest. U.S. Department of Health and Human Services, Fourth Edition).
DETAILED DESCRIPTION OF THE INVENTION
The existence of specific, yet different, surface patches in murine and human antibodies may be the origin of the inherited immunogenicity of murine antibodies in humans. Statistical analysis of a database of unique human and murine antibody F.sub.V fragments has revealed that certain combinations of residues in exposed surface positions are specific for human and murine sequences. The combinations are not the same in human and murine F.sub.V domains. However, it is possible to define families of surface residues for the two species of antibodies. These families reveal a novel method for the "humanization" or reshaping of murine antibodies. Humanization is the modification of the solvent accessible surface of a non-human antibody or fragment thereof to resemble the surface of a chosen human antibody or fragment thereof such that the modified non-human antibody or fragment thereof exhibits lower immunogenicity when administered to humans. Such a process applies in the present application to antibody variable regions but could equally well apply to any other antibody fragment. The method is considered to be generally applicable to humanization of rodent antibodies.
According to the present invention, a statistical analysis is presented which is based on accessibility calculated for a range of antibody crystal structures. When this information is applied to an antibody sequence database, it is possible to discriminate between human and murine antibodies at the sequence level purely on the basis of their surface residue profiles.
Rational Resurfacing Approach
There are several key features of the resurfacing approach of the present invention.
1) This method uses as a starting point, construction of a three-dimensional model of a rodent variable region by known methods;
2) A large number (e.g., twelve) of antibody F.sub.V or Fab fragment x-ray crystallographic structures are analyzed to produce an unambiguous set of surface exposed amino acid residues that will be positionally identical (about 98%) for a majority of antibodies. The set is produced by identifying all those residues whose solvent accessibility is above a given cut-off (typically 30%), calculated using a modification of the method of Kabsch and Sander (Kabsch, W. and Sander, C. (1983), Biopolymers 22, pp. 2257-2637) in which explicit atomic radii are used for each atom type to predict sidechain positions as is described below in more detail;
3) Using a complete human antibody database, the best set of human heavy and light chain surface exposed amino acid residues is selected on the basis of their closest identity to the set of surface amino acid residues of the murine antibody;
4) In order to retain the conformational structure of the CDRs of the rodent antibody, replacement of any human surface exposed amino acid with the original rodent surface exposed amino acid residue is carried out whenever a surface residue in either the rodent or resurfaced antibody is calculated from the three-dimensional model to be within 5 Angstroms of a CDR residue.
The general resurfacing approach of the present invention is illustrated in FIG. 2. The approach can be divided into two stages. In the first, the rodent framework (white) is retained and only the surface residues changed from rodent (dark grey circles) to the closest human pattern (light grey circles). This should remove the antigenicity of the rodent antibody. In the second stage, surface residues within 5 Angstroms of the CDRs are replaced with the rodent equivalents in an attempt to retain antigen binding and CDR conformation.
The method of the present invention is applicable to whole antibodies as well as antibody fragments. Suitable antibody fragments that can be used can readily be determined by the skilled artisan. Examples of some suitable fragments include a single chain antibody (SCA), an antibody F.sub.V fragment, Fab fragment, F(ab').sub.2 fragment, Fab' fragment, or other portion of an antibody comprising the binding site thereof.
According to the present invention, an important step in the method for determining how to modify a rodent antibody or fragment thereof by resurfacing is to determine the conformational structure of the variable region of the rodent antibody or fragment thereof to be humanized by constructing a three-dimensional model of the rodent antibody variable region. This can be done by known methods such as those described, for example, in Martin et al. (Martin, A. C. R., Cheetham, J. C. and Rees, A. R. (1989), Proc. Natl. Acad. Sci. U.S.A. 86, pp. 9268-9272; Methods in Enzymology (1991), 203, pp. 121-152) and as described in detail in Example 2.
Martin et al. describe an algorithm which is depicted in FIG. 1. The algorithm applies to murine and human antibodies equally well. The present inventors therefore expect that, based on sequence similarity between antibodies of different species (Kabat, E. A. Segments of Proteins of Immunological Interest, National Institutes of Health, U.S.A. 1991), the algorithm will work equally well for rat and other rodent antibodies.
Briefly, the algorithm depicted in FIG. 1 can be summarized as follows. The framework region of an antibody to be modelled is selected on the basis of sequence homology and constructed by a least squares fit onto the six conserved strands of the variable region .beta.-barrel. Light and heavy chain complementarity determining regions are constructed using a combination of canonical structures (Chothia, C. and Lesk, A. M. (1987), J. Molec. Bio. 196, pp. 901-917), database searching and conformational searching. Detailed descriptions of these methods are described in Example 2 herein and in the above two references (Martin et al. 1989 and 1991).
According to the present invention, another three-dimensional model is also constructed. The other three-dimensional model is of the rodent antibody variable region having human antibody surface amino acid residues substituted therein at particular rodent antibody surface residue positions.
This other three-dimensional model is constructed by carrying out the series of steps described next.
The first of the steps is to generate sequence alignments from relative accessibility distributions from x-ray crystallographic structures of a sufficient number of rodent antibody variable region heavy and light chains to give a set of framework positions of surface exposed amino acid residues the positions of which are 98% identical in a majority of the variable regions.
As used herein, the term "framework" means the antibody variable region from which the complementarity determining regions have been excluded.
"Complementarity determining regions" means those amino acid sequences corresponding to the following numbering system as defined by Kabat, E. A. (In Sequences of Immunological Interest, N.I.H., U.S.A., 1991).
A sufficient number of rodent antibody fragments that need to be analyzed in order to produce the set of framework positions of surface exposed amino acid residues can readily be determined by the skilled artisan through routine experimentation using a database of antibody sequences. Thus, this step can be conducted using suitable databases now in existence or later compiled.
The x-ray crystallographic structures are used to determine relative accessibility distributions of surface exposed amino acid residues. The relative accessibility distributions identify all those residues whose solvent accessibility is above a given cut-off (typically 30%), calculated using a modification of the method of Kabsch and Sander (Kabsch, W. and Sander C. (1983), Biopolymers 22, pp. 2257-2637) in which explicit atomic radii are used for each atom type.
The relative accessibility distributions determined from the x-ray crystallographic structures can then be used to generate sequence alignments which give a set of framework positions of surface exposed amino acid residues the positions of which are 98% identical in a majority of the variable regions.
The set of framework positions of surface exposed amino acid residues for the variable regions of murine antibodies is shown in Table 1, set forth in Example 1, and was produced using the sequence alignments and accessibility distributions shown in FIGS. 3A and 3B.
Once a set of framework positions of surface exposed amino acid residues for the variable regions of the rodent antibodies have been generated, the surface exposed residues of the heavy and light chain pair of the rodent antibody, or fragment thereof, to be humanized can be identified using an alignment procedure such as that described in Example 1 and shown in FIGS. 3A and 3B. This defines a set of surface exposed amino acid residues of a heavy and light chain pair of a rodent antibody or antibody fragment to be humanized.
Next, a complete human antibody sequence database is used to identify a set of surface exposed amino acid residues from a human antibody variable region that have the closest positional identity to the set of surface exposed amino acid residues of the variable region of the rodent antibody that is to be humanized. The set of surface exposed amino acid residues from the human antibodies can be separately identified for a heavy chain and for a light chain that are not naturally paired and/or a set can be identified from a natural human heavy and light chain pair, that is, a pair originating from a single B cell or hybridoma clone. Preferably, the set is one from a natural human heavy and light chain pair.
A humanized rodent antibody that gives the appearance of a human antibody is then predicted by substituting the set of surface exposed amino acid residues from the rodent antibody or fragment thereof to be humanized with the set of surface exposed amino acid residues from the human antibody.
A three-dimensional model can then be constructed from the resulting, fully substituted variable region of the rodent antibody or fragment thereof. The three-dimensional model is constructed using the same known methods mentioned above for constructing a 3-D model of the original rodent antibody or fragment thereof.
While the antigenicity of this fully "resurfaced" or humanized antibody should be removed, an additional factor to be addressed is the binding affinity or the binding strength of the resurfaced antibody. Changes in the framework of the variable domain introduced through resurfacing can influence the conformation of the CDR loops and therefore antigen binding of the antibody. According to the present invention, this problem is removed by the next step which is to identify, by means of a comparison of both of the above-described three-dimensional models of the rodent antibody variable region, any residues from the set of surface exposed amino acid residues of the variable region heavy and light chain pair of the rodent or human antibody identified that are within 5 Angstroms of any atom of any residue of the rodent antibody or antibody fragment complementarity determining regions (CDRs).
Any residue(s) so identified is then changed back from the human to the original rodent amino acid residue(s).
The results of this method can then be applied to a particular rodent antibody by well known methods. Briefly, genes for the humanized variable heavy and light chain regions are constructed using standard recombinant DNA methods (Sambrook, J., Fritsch, E. F. and Maniatis, T. (1989), Molecular Cloning, Second Edition). For example, a PCR method can be used (Daugherty et al. (1991), Nucleic Acids Research 19, pp. 2471-2476).
Variable heavy chain or variable light chain gene constructs are subcloned into appropriate expression vectors. Suitable expression vectors contain either a human gamma or human kappa constant region gene, a suitable promoter, a sequence coding for a human immunoglobulin leader peptide (for example: met-gly-trp-ser-cys-ile-ile-leu-phe-leu-val-ala-thr-ala-thr (SEQ ID NO:39), Olandi et al. (1989), PNAS 86, pp. 3833-3837), and a drug selectable marker.
Heavy and light chain expression plasmids can be co-transfected, for example, by electroporation into suitable cells, for example, SP2/0 cells, and selected with an appropriate drug, G418, for example. Screening for intact antibody can be accomplished by ELISA assay. 96-well plates are coated with, for example, goat anti-human kappa chain antibody, and light chains are detected with, for example, goat anti-human antibody conjugated to alkaline phosphatase.
As another approach, light chain constructs are transfected, for example, by electroporation into suitable cells, for example, SP2/0 cells and selected, for example, in hygromycin. Screening for light chain expression can be accomplished by ELISA assay. 96-well plates are coated with, for example, goat anti-human kappa chain antibody, and light chains are detected with, for example, goat anti-human antibody conjugated to alkaline phosphatase.
A light chain producing line is then used as a host to electroporate in the heavy chain construct. The heavy chain plasmid is co-transfected with a plasmid containing the gene coding for another drug marker, for example, neomycin resistance and selected in the presence of the drug G418. Screening for intact antibody is accomplished by ELISA assay. 96-well plates are coated with, for example, goat anti-human Fc and detected with, for example, goat anti-human light chain conjugated to alkaline phosphatase.
EXAMPLE 1 AND COMPARATIVE EXAMPLES
The superiority of the presently claimed method for determining how to modify a rodent antibody or fragment thereof by resurfacing in order to produce a humanized rodent antibody will now be described by reference to the following example and comparative examples which are illustrative and are not meant to limit the present invention.
A) Analysis for Murine Antibodies
In order to determine the positions which are usually accessible on the surface of the F.sub.V domain of murine antibodies, the accessibility was calculated for twelve Fab x-ray crystallographic structures obtained from the Brookhaven database (Bernstein, F., Koetzle, T., Williams, G., Meyer, E., Brice, M., Rodgers, J., Kennard, O., Shimanouchi, T. and Tasumi, M. (1977), J. Mol. Biol. 112, pp. 535-542). The relative accessibility was calculated using the program MC (Pedersen, J. (1991)), which implements a modified version of the DSSP (Kabsch, W. and Sander, C. (1983), Biopolymers 22, pp. 2257-2637) accessibility calculation routine in which explicit atomic radii are specified for every atom. A residue was defined as being surface accessible when the relative accessibility was greater than 30%. The alignment positions of these residues were conserved in all twelve structures (98% identity). Surface accessible framework positions constitute 40% of the F.sub.V surface area. The remaining surface accessible residues are in the CDRs and in the interdomain C-terminal region. FIGS. 3A and 3B show a sequence alignment of the twelve crystal structures, the average relative accessibility, and the 30% accessibility cutoff. FIG. 3A shows the alignments and relative accessibility for the twelve murine antibody light chains and FIG. 3B shows the alignments and relative accessibility for the murine antibody heavy chains.
The surface accessible framework positions were mapped onto a database of unique human and mouse F.sub.V sequences (see lists at the end of this Example). The frequency of particular residues in each of these positions is shown in Table 1. Only residue frequencies higher than 5% are listed.
None of the entire combinations of surface residues in the human sequences are found in the murine sequences and vice versa (see lists at the end of this Example). However, the residues in individual positions appear to be conserved (see Table 1). There are few residues which differ significantly between the species; these are at positions 54 and 91 of the L chain and 168 and 216 of the H chain. Of these positions only position 216 is a non conservative (V to Y) mutation. Differences between human and murine antigenicities are therefore believed to arise from the combinations of residues in these positions.
In order to determine whether the mouse sequences are more distantly related to human F.sub.V sequences than to other mouse F.sub.V sequences, the homology was calculated using a Dayhoff mutation matrix (Dayhoff, M., Barker, W. and Hunt, L. (1983), Meth. Enz. 91, pp. 524-545). The homology was calculated between all the sequences in a pool of both human and mouse sequence patches made up of the surface accessible residues. The data was then represented as a density map (not shown) in which the sequences are plotted against each other. The density map can be used to discriminate "murine surfaces" from "human surfaces".
B) Reshaping of Antibody N901
In order to test the resurfacing approach suggested by the above analysis, three humanization experiments were set up. 1) Traditional loop grafting (Verhoeyen, M. E., Saunders, J. A., Broderick, E. L., Eida, S. J. and Badley, R. A. (1991), Disease markers 9, pp. 3-4) onto a human F.sub.V framework of known structure (KOL). 2) Resurfacing approach using most similar chain. 3) Resurfacing approach using human sequences with most similar surface residues.
The antibody used was the murine anti-N901 antibody (Griffin et al. (1983), J. Imm. 130, pp. 2947-2951). The anti-N901 antibody (also referred to herein as the "N901 antibody") is available commercially from Coulter Corporation under the name NKH-1.
The alignment of the light chain sequences and heavy chain sequences in FIGS. 4A and 4B, respectively, show the original N901 antibody and the sequences used in each of the three approaches outlined here.
FIGS. 4A and 4B show alignments of sequences generated using the three methods of humanization. Sequences are: 1) Original rodent N901. 2+3) KOL (Marquart, M. Deisenhofer, J. and Huber, R. (1980), J. Mol. Biol. 141, pp. 369-391) and reshaped N901 using KOL surface. 4+5) Most homologous sequences, L(KV2F) (Klobeck, H., Meindl, A., Combriato, G., Solomon, A. and Zachau, H. (1985), Nucleic Acids Res., pp. 6499-6513) and H(G36005) (Schroeder, H. and Wang, J. (1990), Proc. Natl. Acad. Sci. U.S.A. 87) and reshaped N901 using these sequences. 6+7) Most homologous with respect to surface residues, L(KV4B) (Klobeck, H., Bronkamp, G., Combriato, G., Mocikat, R., Pohelnz, H. and Zachau, H. (1985), Nucleic Acids Res. 3, pp. 6515-6529) and H(PLO123) (Bird, J., Galili, N., Link, M., Sites, D. and Sklar, J. (1988), J. Exp. Med. 168, pp. 229-245), and reshaped N901 using these sequences. The numbering is the same as used in the antibody modelling program ABM (ABM is a trademark for commercial software, Oxford Molecular Ltd., Oxford, U.K.), which is based on structural conservation and not sequence homology as used by Padlan et al. (Kabat, E. A., Wu, T. T., Reid-Miller, M., Perry, H. M. and Gottesman, K. S. (1987), Sequences of Proteins of Immunological Interest. U.S. Department of Health and Human Services, Fourth Edition). The sequence changes which have to be introduced in order to reshape N901 with a given sequence are marked with bars, and back-mutations as determined from F.sub.V models are marked with stars. The sequence homology of a given sequence to N901 is shown in brackets after each sequence.
(1) Classical Humanization
In classical humanization the rationale is to graft the rodent CDR's onto a framework of known structure, such that CDR-framework interactions can be accurately monitored by homology modelling. The model of the humanized antibody is compared to that of the original rodent antibody, and possible CDR interacting framework residues are back mutated (marked with `*` in alignment) in order to retain the three-dimensional shape of the CDR's. In this example the antibody KOL was used, giving a low homology score of only 77 and 46 in the heavy and light chains respectively.
(2) Most Similar Chain Resurfacing
A database of nonredundant human antibody sequences was compiled from available protein and nucleotide sequences. A total of 164 H and 129 L chains were sampled.
Each of the rodent chains, L and H, were then matched and the most similar human sequence found independently (G36005/KV2F) (Schroeder, H. and Wang, J. (1990), Proc. Natl. Acad. Sci. U.S.A. 87); Klobeck, H., Meindl, A., Combriato, G., Solomon, A. and Zachau, H. (1985), Nucleic Acids Res., pp. 6499-6513). Surface residues, as outlined in Table 1, were then changed in the rodent sequences to match those of the human sequences. Subsequently a model was built of the resurfaced antibody and compared to the model of the original rodent antibody and back mutation of any CDR interacting residues was performed.
(3) Most Similar Surface Replacement According to the Present Invention
This method is identical to the above method, except that the similarity is calculated only over the surface residues outlined in Table 1 above.
The same procedure of surface mutation and subsequent back mutation was performed as in the previous methods. In this case the chosen sequences were PLO123/KV4B (Bird, J., Galili, N., Link, M., Sites, D. and Sklar, J. (1988), J. Exp. Med. 168, pp. 229-245); Klobeck, H., Bronkamp, G., Combriato, G., Mocikat, R., Pohelnz, M. and Zachau, H. (1985), Nucleic Acids Res. 3, pp. 6515-6529).
The following lists show the surface residue patterns in mouse and human light and heavy chain antibody variable regions. The sequences are ordered on similarity to one another. There are no pattern matches between mouse and human sequences although there are matches within a species.
EXAMPLE 2
Detailed Description of Method for Constructing Three-Dimensional Model of Antibody Variable Region
The references cited in the text below are listed at the end of this Example.
The first antibody Fab structure was determined in 1972. Since then, no more than about twelve Fab structures have been published, a number that represents a very small fraction of the total antibody repertoire (>10.sup.8 antibodies). To understand the molecular basis of this antibody diversity will require knowledge of either a large number of x-ray structures, or the rules by which combining site topography is governed. The development of such prediction rules has now reached the point where variable regions of antibodies can be modelled to an accuracy approaching that of the medium resolution x-ray structure.
The interaction of an antibody with its cognate antigen is one of the most widely accepted paradigms of molecular recognition. To understand the antibody-antigen interaction in atomic detail requires knowledge of the three-dimensional structure of antibodies and of their antigen complexes. Traditionally such information has come from x-ray crystallographic studies (see Davies et al. for review (Davies et al., 1988)).
The modelling of antibody combining sites was first attempted by Padlan & Davies (Padlan et al., 1976) at a time when very few antibody structures were known. Nonetheless, Padlan and colleagues recognized that the key lay in high structural homology that existed within the .beta.-sheet framework regions of different antibody variable domains. The antigen combining site is formed by the juxtaposition of six interstrand loops, or CDRs (Complementarity Determining Regions) (Kabat et al., 1987), on this framework. If the framework could be modelled by homology then it might be possible to model the CDRs in the same way. Padlan and Davies (Padlan et al., 1976) reasoned that CDR length was the important determinant of backbone conformation though the number of antibody structures was insufficient to thoroughly test this maximum overlap procedure (MOP). This notion was not picked up again until the early 1980's when Pedersen and Rees proposed a similar approach to modelling antibody combining sites based on a more extensive analysis of antibody structures (de la Pas et al., 1986).
Those essentially knowledge-based procedures are best exemplified for antibodies by the work of Chothia & Lesk (Chothia et al., 1986) who, in 1986, extended and modified the MOP procedure by introducing the concept of "key" residues. These residues allow the further subdivision of CDRs of the same length into "canonical" structures which differ in having residues at specified positions that, through packing, hydrogen bonding or the ability to assume unusual values of the torsion angels .phi., .psi. and .omega., determine the precise CDR conformation (Chothia et al., 1989). Similar knowledge-based methods have been proposed for predicting loop conformations in general (Thornton et al., 1988; Tramontano et al., 1989). These methods rely on the crystallographic database of protein structures. However, none of the above knowledge-based methods has been totally successful. In particular, the MOP or canonical structure approaches have succeeded in modelling only five of the six CDRs. This stems from the fact that the third CDR of the heavy chain, H3, is more variable in sequence, length and structure than any of the other CDRs.
To deal with this problem several groups have attempted to use ab initio methods to model the combining site (Bruccoleri and Karplus, 1987). The requirement with such methods is that the total allowable conformational space accessible to a particular CDR is sampled. Typical of purely geometric approaches is that of Go & Sheraga (Go and Sheraga, 1970) and more recently Palmer & Sheraga (Palmer and Sheraga, 1991), where the problem is reduced to one in which the central region of the polypeptide backbone, having characteristic bond length and bond angles, is constructed between the end points of the loop (CDR if an antibody loop) by a "chain closure" algorithm. In a modification of this algorithm, Bruccoleri & Karplus (Bruccoleri and Karplus, 1987) introduced an energy minimization procedure which greatly expanded the domain of conformational space searched during the chain closure procedure. This modification is incorporated into the conformational search program CONGEN (Bruccoleri and Karplus, 1987), which also allows the user to choose any set of standard bond length and bond angels such as the CHARMM (Brooks et al., 1983) standard geometry parameter sets. Other approaches such as minimization (Moult and James, 1986), or molecular dynamics (Fine et al., 1986) either fail to saturate conformational space or are unable to deal with the problem of long CDRs. Whichever of the ab initio methods is employed however, the problem is one of defining the selection criteria in such a way as to allow the unambiguous identification of the correct structure (in this context correct is defined by reference to an appropriate X-ray structure) within the ensemble of candidates, for every CDR. To date this has not been possible.
Recently a more holistic approach has been taken to the modelling of CDRs which combines the advantages of knowledge-based and ab initio methods in a single algorithm known as CAMAL (Combined Algorithm for Modelling Antibody Loops) (Martin et al., 1989; Martin et al., 1991). Previously this algorithm has been used to model individual CDRs in the presence of the crystal structure conformations of the other five. As is demonstrated below, CAMAL is able to predict the backbone conformations of all six CDRs of the antibody combining site to an accuracy approaching that of medium resolution x-ray structures. In addition the algorithm includes a procedure for selecting and fitting together the light and heavy chain framework regions prior to generation of CDR conformations, thus making possible the prediction of the entire variable region. Furthermore a new Monte Carlo (MC) simulated annealing method has been developed for the determination of sidechain conformations.
the Framework Region
Antibody framework regions consist of conserved .beta.-strands that form the .beta.-barrel structure characteristic of immunoglobulin V-type regions. In the procedure described here each V-region is built from a database of known antibody structures, using sequence homology for selection of the light (L) and heavy (H) chain V-domains. The two domains are then paired by least squares fitting on the most conserved strands of the antibody .beta.-barrel (Table 2 and FIGS. 5 & 6). The strand orientations were determined by analyzing the barrels of known antibody crystal structures. Eight antibodies were analyzed using a multiple structure fitting program as follows. Seven structures were fitted onto one of the set selected at random and mean coordinates were calculated. All eight structures were then fitted onto these mean coordinates and new mean coordinates determined. This procedure was iterated until the mean coordinate set converged (5-10 cycles). The variance for the mean coordinates at each barrel point (N,C.alpha.,C) was calculated. In FIG. 5 this variance is plotted against the projected positions of these points onto the conjugate axis of the barrel.
Strand 8 and all but two residues of strand 7 in both light and heavy chains were eliminated as they showed deviations greater than 3.sigma. (standard deviation units) from the mean coordinates. These two strands comprised the takeoff points of CDR H3, and suggests that any knowledge-based prediction of CDR H3 would have to account not only for sequence and length variation in the CDR itself, but also for the position of the participating strands. The remaining mean coordinates were used as a scaffold onto which the L and H chains were fitted. Strands 7 and 8 in the final framework were obtained from the database structure used in the construction. The framework strands are marked + in the multialignment in Table 2.
The sidechains were then replaced using a . `maximum overlap` method, in which sidechain templates were fitted on backbone atoms with the sidechain torsion angles being adjusted to match those of equivalent torsions in the parent sidechain.
The Combining Site
The procedure for predicting the structure of combining sites combines a database search with a conformational search procedure. The architecture of the program suite to perform this task is outlined in FIG. 7.
The database search utilizes distance constraints for each of the six CDR loops determined from known antibody structures. These constraints were determined by calculating C.alpha.--C.alpha. distances within known loops and using a search range of x+3.5.sigma. (the mean .+-.3.5 standard deviation units). A database containing all the proteins in the Brookhaven Protein Databank (Bernstein et al., 1977) is then searched for fragments which satisfy the constraints for a loop of the required length. The middle section of the loop is then deleted and reconstructed using the conformational search program CONGEN (Bruccoleri and Karplus, 1987). For loops of six or seven residues, the structure database appears to saturate the conformational space available to the backbone adequately and only sidechains are built by conformational search. Loops shorter than six residues are built by conformational search alone since this is computationally feasible and the number of loops selected from the database becomes unacceptably large as loop length decreases.
When modelling a complete combining site, loops of 6 or more residues are modelled individually with the other loops absent. If the loops are built consecutively, small errors can accumulate leading to a poor result (Martin, 1990). All the loop conformations are then evaluated using a solvent modified potential, which excludes the attractive van der Waals and electrostatic terms of the non-bonded energy function contained within the GROMOS (.ANG.qvist et al., 1985) potential. The lowest five energy conformations are selected and filtered using a "structurally determining residue" algorithm (FILTER), based on backbone torsion angles observed in the original database loops. Since the database search is not used for the shortest loops of 5 residues or fewer, the FILTER algorithm cannot be used. Energy is thus the only available selection criterion and the short loops are built last, in the presence of the longer loops.
Side Chains
The determination of sidechain positions was previously done using the iterative sidechain determination algorithm described by Bruccoleri et al. (Bruccoleri and Karplus, 1987). Unfortunately the CHARMM (Brooks et al., 1983) force field fails to select the correct conformations of exposed hydrophobic sidechains. There is no penalty for having an exposed uncharged atom, without solvent present. CONGEN is also unable to saturate the conformational space for a large number of sidechains (more than 6 residues).
Recently Lee et al. (Lee and Levitt, 1991; Lee and Subbiah, 1991) has proposed a method for searching conformational space for a large number of sidechains using MC simulated annealing. A simple energy function is used for the evaluation of conformations generated by a biased random walk: ##EQU1## Where the first term is a simple Lennard-Jones potential which evaluates the non-bonded contacts between the atoms in a given molecule, the second term is a simple torsional term which only applies to C--C bonds. The torsional term biases the function towards 60.degree. rotamers. .epsilon..sub.o and .kappa..sub.o are constants. The metropolis function: ##EQU2## is used to evaluate the energy function. Any move which results in a decrease in energy is accepted, and any move which results in a positive .delta.E is only accepted with the probability P. This simple method can be used to search the large conformational space defined by a set of torsion angles in amino-acid sidechains, and find or define the global minimum which exist for a set of sidechains. T is the simulation temperature.
When searching sidechain conformations using this method the simulation system usually gets trapped in an energetic minima well before the global minimum is encountered, at a high temperature, without the solution space having been searched sufficiently. This problem can be solved by truncating the Lennard-Jones potential, thus allowing atoms to pass through each other. In reality this function would converge towards infinity when the distance r between the atoms approaches zero.
The evaluation of sidechain conformations generated is done solely on the basis of energy, for internal (core) residues, since good van der Waal's interactions are considered to be equal to a good packing of the sidechains. The situation becomes more complicated when trying to predict the conformation of surface residues. The lowest van der Waal's interaction is obtained by a combination of sidechain conformations which minimize the overlap of atoms, this means that the lowest energy is obtained with extended conformations of sidechains, without considering good packing of sidechains.
Using the fact that hydrophobic, bulky residues will be shielded by the hydrophilic sidechains, and will be buried in the surface, it is possible to generate a simple function which will evaluate these macroscopic observations. These functions can either be implemented in the objective evaluation function of the Monte Carlo simulation, or as is done here, added as a post processing step. Including an accessibility/hydrophobicity term in the evaluation function would slow down the calculation considerably, hence the term has been added as a post processing function. The function used is a sum of the product of relative exposed surface area multiplied by the residual hydrophobicities. The hydrophobicities used are taken from Cornette et al. (Cornette et al., 1987). ##EQU3## n is the number of sidechains reconstructed. The surface area is calculated using the tesselated icosahedron approach (Chau and Dean, 1987), which is not very precise (0.1 percent), but is able to evaluate a large number of conformations. The function is evaluated for the final 2,000 conformations and the lowest value conformation selected as the best.
Using this simple approach it is possible to integrate over a large phase space with many degrees of freedom, and get a complete sampling of the space.
Predicted Structures of an Anti-hapten, Anti-peptide and Two Anti-protein Antibodies
In the following section the predicted structures of four different antibody F.sub.V regions are presented and analyzed. The antibodies are:
Gloop-2 (Darsley and Rees, 1985), an anti-lysozyme antibody whose Fab structure was previously determined and which was used as a learning exercise during the development of CAMAL.
D1.3 (Amir et al., 1986), an anti-lysozyme antibody whose uncomplexed F.sub.V coordinates were supplied by R. Poljak et al. after the model coordinates had been deposited.
36-71 (Rose et al., 1990), an anti-phenylarsonate antibody whose Fab structure was carried out by D. R. Rose, et al., and whose coordinates were obtained after the model coordinates had been deposited.
3D6 (Grunow et al., 1988), an anti-protein (GP41 of HIV) antibody whose Fab structure was carried out by D. Carter et al. (Carter, 1991) and whose coordinates were obtained after the model coordinates had been deposited. For this antibody, the model was generated using the canonical loop method of Chothia & Lesk (Chothia et al., 1989; Chothia et al., 1986) for CDRs L1, L2, H1 and H2, while L3 and H3, which cannot be modelled using canonical structures, were constructed using CAMAL.
All four models were subjected to both restrained and unrestrained energy minimization using the DISCOVER (TM Biosym Technology) potential with 300 cycles of steepest descents, followed by conjugate gradient minimization until convergence to within 0.01 Kcal occurred.
The resolution and R-factors of the x-ray structures are given in Table 3 together with the parent frameworks selected in building the models. The structures and models were compared by global fits of the loops. The .beta.-barrel strands 1 to 6, as described above, were least squares fitted and the RMS deviation was then calculated over the loops. The backbone (N,C.alpha.,C) RMS values for fitting model and crystal structure frameworks were between 0.4 and 0.9 .ANG., illustrating the conservation of the core .beta.-barrel. Using all eight strands RMS deviations between 0.6 and 1.2 .ANG. were observed.
Global fits (Table 4) give a more realistic measure of the accuracy of the model than a local least-squares fit over the loops since they account for the overall positioning of the loops in the context of the F.sub.V structure. Local fits, which give lower RMS deviations, are also shown in Table 4. Differences between local and global RMS deviations arise from differences in V.sub.H /V.sub.L domain packing and differences in loop `take off` angles and positions.
Table 5 shows the canonical loops selected from modelling 3D6. Backbone structures of the modelled CDRs, superimposed on the x-ray structures after global fitting are shown in FIG. 8. General features and points of interest for each of the six CDRs are discussed below.
Analysis of the CDR Regions
During the comparison of CDR conformations in the V-region models and the x-ray Fab structures it was observed that at certain positions in a CDR, the peptide backbone may adopt either of two conformations by undergoing a "peptide flip" (1,4 shift). This phenomenon is also seen in type 2 .beta.-turns (Paul et al., 1990). Dynamics simulations of .beta.-turns show that the transformation energy between .phi.1=-00, .psi.1=-30, .phi.2=-90, .psi.2=0 and .phi.1=-00, .psi.1=120, .phi.2=90, .psi.2=0 has a maximum value of 5 kcal (Paul et al., 1990). This is low enough to allow selection of either conformation. The peptide flip is observed within several canonical classes (as described by Chothia et al. (Chothia et al., 1989)) and the hydrogen bonding pattern used to determine the conformation of a canonical class does not disallow the peptide flip. Any modelling procedure should therefore take these, or any other multiple conformations, into consideration where the transformation energies are sufficiently low to permit population of the different conformational forms. Table 6 shows an example of the "peptide-flip" phenomenon from the crystallographic database of antibody structures. It should be noted that a single crystal structure will not show multiple conformations since the crystallization will `freeze out` one of the conformations. During the modelling procedure the two populations of conformers are easily extracted from a set of ab initio generated loops, by using a torsional clustering algorithm.
CDR-L1
In Gloop-2 and D1.3, all five low energy conformations were very similar with RMS deviations differing by less than 0.25 .ANG. (backbone) and 0.35 .ANG. (all atoms). The FILTER algorithm was unable to distinguish between the conformations and the lowest energy structure was selected.
Although CDR-L1 of 3D6 was originally built using the canonical loop from HyHEL-10, the mid-section was rebuilt by conformational search, for the following reason. HyHEL-10 and REI CDR-L1 loops are placed in the same canonical ensemble (Chothia et al., 1989) although they contain a 1-4 shift (peptide flip) relative to one another between the fifth and eighth residues of the loop (residues 28-31) (see Table 6).
36-71 shows the same 1-4 shift between the model and crystal structure CDRs. Both crystal structure and model were compared with other loops of the same canonical class as defined by Chothia et al. (Chothia et al., 1989). It was found that the hydrogen bonding pattern which determines the conformation was conserved.
CDR-L2
CDR-L2 of D1.3 has two adjacent threonines (49, 50) which in the x-ray structure are packed against the tyrosine at the fourth position of CDR-H3, thus minimizing the exposed hydrophobic sidechains. In the unminimized model the threonine sidechains are exposed to the solvent, but after energy minimization, this packing is observed.
CDR-L3
In Gloop-2, D1.3 and 36-71 the proline at the seventh position in the loop is correctly predicted in the cis conformation. It has previously been suggested that the conformation of CDR-L3 is dictated by the presence of a proline in position 8 or 9 (Chothia et al., 1989) within the loop. 3D6 does not have a proline in either position. Only 7 out of 290 CDR-L3 sequences (Kabat et al., 1987) lack a proline at both positions and in all of the published x-ray structures this proline is present. This is an example of a situation where either a new canonical class may need to be defined or where the canonical rule breaks down altogether, and an alternative method must be employed.
The 3D6 L3 loop is 7 residues in length and was built using database loops alone where conformational space is saturated by means of fragments selected from the crystallographic database (Global RMS: 2.01 .ANG., N,C.alpha.,C), and by using CAMAL (Construction: Q[Q(YNS)Y]S, Global RMS: 1.97 .ANG., N,C.alpha.,C). The similarity of the structures generated by the two procedures illustrates the utility of the database search and suggests that, for shorter loops it is capable of saturating the available conformational space.
CDR-H1
Using the Kabat and Wu definition of CDR-H1 places this loop as an extension of the .beta.-sheet. The extended nature of this stretch of peptide limits its conformational flexibility and CDR-H1 is generally modelled accurately (Martin et al., 1989; Chothia et al., 1989).
In Gloop-2 and D1.3, the Phe or Tyr sidechain at the second position in the loop is poorly placed and packs against Leu at the penultimate position in HFR1 (see Table 2). 36-71 has a well-placed Asn at this position, rather than the more common bulky hydrophobic sidechain.
CDR-H2
CDR-H2 of 36-71 is similar in sequence to F19.9 (Strong et al., 1991), (36-71: YNNPGNGYIA (SEQ ID NO:492); F19.9: YINPGKGYLS (SEQ ID NO:493)). While the structurally determining residues specified by Chothia and Lesk (Chothia et al., 1989) are conserved, the backbone conformations are different: F19.9 has a bulge at the --PGN-- Gly, compared with 36-71, giving the loop a `kink` in the middle. The model of 36-71 shows a 1-4 shift, though the sidechains are still well placed.
In Gloop-2, the all atom RMS deviation is poor (3.00 .ANG.) when compared with the P2.sub.1 crystal structure, owing to rotations of the Phe at position 3 in the loop and Tyr at position 10 by approximately 120.degree. about the .chi..sub.2 torsion angle. Gloop-2 has been solved in two different crystal forms, P2.sub.1 and P1 Jeffrey, 1989). When compared with the P1 structure, the sidechains are placed almost perfectly and the all atom RMS (global fit) drops to 2.23 .ANG..
This concerted sidechain motion between crystal forms illustrates the effects of crystallization conditions on surface sidechain placement. Even though surface sidechains may show low temperature factors indicating low mobility in the crystal, their mobility in solution may be high. In the Gloop-2 P1 structure, the mean sidechain temperature factor for the F.sub.V domain is 13.46 (.sigma.=8.20) while the sidechains of these two residues of H2 show mean temperature factors of 5.56 (.sigma.=0.68) for the Phe at position 3 and 7.10 (.sigma.=1.73) for the Tyr at position 10.
CDR-H3
CDR-H3 is the most variable of the six CDR's with all lengths up to 21 residues being represented in Kabat et al., (Kabat et al., 1987). This extreme variability results from V-D-J splicing (Schilling et al., 1980) and has always been a problem when attempting to model antibodies. Such loops may be divided into short (up to 7 residues), medium (up to 14 residues) and long (15 or more residues). Using the CAMAL procedure, short and medium CDR-H3's can be modelled as accurately as other CDR's of similar lengths. Although long CDR-H3's are more difficult and cannot, at present, be built to the same accuracy, the chain trace is still correct.
It is unlikely that the longer loops consist of `pure` loops (i.e., all random coil or turn). In crystal structures of antibodies with medium to long CDR-H3 loops (McPC603 (Rudikoff et al., 1981): 11 amino acids (aa); KOL (Marquart et al., 1980): 17 aa; F19.9 (Lascombe et al., 1989): 15 aa) the loops consist of a disordered .beta.-sheet extension from the .beta.-barrel core and a 5-8 residue random coil/turn connecting these two strands.
To determine the nature of medium to long loops (>8 residues) which satisfy the CDR-H3 constraints, a complete search of the Protein Databank for loops of length 8-20 residues, was performed using the inter-C.alpha. distance constraints determined from known antibody crystal structures for CDR-H3. The resulting loops were then analyzed using the DSSP (Kabsch and Sander, 1983) program, which is able to assign secondary structure to polypeptide structures. The amount of secondary structure for each length of loop was calculated, and it was observed that for loops longer than 12 residues the amount of secondary structure within each of the classes described in DSSP was constant. The number of loops selected is also constant (approximately 150 loops) for loops longer than 12 residues. A closer inspection of each of the length ensembles shows indeed that the loops are the same between the groups.
This analysis shows that, like the long CDR-H3 crystal structures, the selected fragments consist of .beta.-strands connected by 5-8 residue loops. For loops above 12-13 residues in length, the same loops are selected, but with extensions to the .beta.-strands. This is called the "sliding-ladder" effect. In addition, the maximum size of a random coil or turn fragment in any of the structures contained in the Protein Databank tends not to exceed 8 residues, as determined by DSSP. This implies that the conformational space of longer loops is not saturated by the database and, although it is unlikely that long loops in antibodies will differ significantly from long loops in other structures, confidence in the prediction must be correspondingly reduced.
By how much is the usefulness of the CAMAL algorithm reduced by this observation?
The frequency of occurrence of different CDR-H3 lengths in antibody sequences described by Kabat et al. (Kabat et al., 1987) was analyzed. FIG. 10 shows that more than 85% of H3 loops have lengths between 4 and 14 residues which can be modelled accurately by the CAMAL algorithm.
CDR-H3 of D1.3 is of average length (8 residues), though no loops of this length are seen in the available antibody structures. The crystal structure coordinate set showed an RMS of 1.9 .ANG. compared with the model.
The 36-71 loop is 12 residues long. The conformation is correctly predicted as a short loop connecting an extension of the .beta.-sheet.
The 3D6 H3 loop is 17 residues long. While KOL (Marquart et al., 1980) has the same length it has only one residue in common with 3D6 and only one conservative mutation. There is thus no reason to believe that the conformations would be similar. The final predicted conformation of 3D6 is an extended .beta.-sheet, as in the crystal structure. The difference between the predicted and the crystal structure of 3D6-H3 is due to a twist of 5.degree.-7.degree. in the extended .beta.-sheet loop (see FIGS. 9A-9D). Such a twist has also been observed for complexed and uncomplexed antibodies by Wilson et al. (Wilson and others). This suggests that long CDR-H3 loops may be flexible and actively involved in antigen binding.
The Complete Variable Region
Prediction of the strand positions and V.sub.L -V.sub.H orientation in the framework .beta.-barrel was exact for all of the four antibodies. The backbone (N,C.alpha.,C) RMS deviations from the crystal structures were between 0.56 and 0.86 .ANG., despite the fact that, in all cases the V.sub.L and V.sub.H regions of a particular model were derived from different antibody structures. This suggests that this method will do well in procedures such as humanization (Gorman et al., 1991), where correct framework positioning is important. The backbones of all six CDRs in all four antibodies are essentially correctly predicted, as shown in FIG. 8. There are two important points to make about these predictions. First, the position of each CDR on its framework barrel is correct. Thus, CDR-framework interactions can be confidently monitored. The only deviation from the x-ray structure is CDR-H3 of antibody 3D6 which has been previously discussed. Second, the all atom RMS deviation between models and x-ray structures is dominated by sidechain positions. In most instances this deviation is due to a small number of incorrectly positioned, exposed sidechains (for example, in D1.3 the only sidechains which are incorrectly predicted are Tyr 9 of L1, Trp 4 of L3, Tyr 2 of H1 and Tyr 4 of H3). Since each CDR is constructed in the absence of other CDRs, the force field may choose a rotamer which is 120.degree. away from that found in the crystal structure. This effect has also been observed by Lee et al. (Lee and Levitt, 1991).
Conclusion
For antibodies having CDR H3 regions of 14 residues or less the complete variable domain can be modelled to an accuracy approaching that of medium resolution x-ray structures. For antibodies with longer H3 loops the CAMAL algorithm is likely to need an additional procedure in which molecular dynamics simulations are also incorporated.
The canonical approach of Chothia et al. appears to work well (at least in modelling backbones) where it may be applied and may be used successfully in combination with the CAMAL procedure.
One important observation that has emerged from these studies is that a given loop can exist in several conformations. In particular, this seems likely for CDR-L1 and, to a lesser extent, CDR-L3 and longer CDR-H3's. A simple combinatorial calculation shows that, if each of these three loops can exist in three separate conformations, a given combining site can have 3.sup.3 =27 different topographies. Clearly, this would explain the origins of cross reactivity and would allow for induced fit of antigens. - Alignment of antibody sequences used in the modeling. `*` indicates CDR regions; `+` indicates .beta.-strand regions used in the fitting for modelling frameworks. Nomenclature for .beta.-barrel strands is (H or L - Chain) - FR(Framewor k region)-(Strand number), thus for example strand one of the heavy chain becomes HFR1. SEQ. Antibody ID. No. sequence L-chain ##STR1## ##STR2## ##STR3## ##STR4## ##STR5## ##STR6## ##STR7## ##STR8## ##STR9## ##STR10## ##STR11## ##STR12## ##STR13## ##STR14## ##STR15## Seq. Antibody ID. No. sequence L-chain ##STR16## ##STR17## ##STR18## ##STR19## ##STR20## ##STR21## ##STR22## ##STR23## ##STR24## ##STR25## ##STR26## ##STR27## ##STR28## ##STR29## ##STR30## ##STR31## ##STR32##
M. J. Darsley, P de al Paz, D. C. Phillips and A. R. Rees in Methodological Surveys in Biochemistry and Analysis, pages 63-68, Volume 15, 1985, Plenum Press (Eds. E. Reid, G. M. W. Cook and D. J. Morre), Presented at the Ninth International Subcellular Methodology Forum, Sep. 3-6, 1984, Guildford, UK.
Amit, A. G., Mariuzza, R. A., Phillips, S. E. V. and Poljak, R. J. (1986). The Three-dimensional Structures of an Antigen-antibody Complex at 2.8 .ANG. Resolution. Science 233, pp. 747-753.
.ANG.qvist, J., van Gunsieren, W., Leifonmark, M. and Tapia, O. (1985), J. Mol. Biol. 183, pp. 461-477.
Bernstein, F., Koetzle, T., Williams, G., Meyer, E., Brice, M., Rodgers, J., Kennard, O., Shimanouchi, T. and Tasumi, M. (1977), J. Mol. Biol. 112, pp. 535-542.
Brooks, B., Bruccoleri, R., Olaison, B., Statcs, D., Swaminathan, S. and Karplus, M. (1983), J. Comp. Chem. 4, pp. 187-217.
Bruccoleri, R. E. and Karplus, M. (1987), Prediction of the Folding of Short Polypeptide Segments by Uniform Conformational Sampling. Biopolymers 26, pp. 137-168.
Carter, D. et al. (1991). Protein Engineering, p. 9999.
Chau, P. and Dean P. (1987). Molecular Recognition: 3d Surface Structure Comparison by Gnomonic Projection. J. Mol. Graph. 5, pp. 97-100.
Chothia, C., Lesk, A., Levitt, M., Amit, A., Mariuzza, R., Phillips, S. and Poljak, R. (1986). The Predicted Structure of Immunoglobulin D1.3 and its Comparison with the Crystal Structure. Science 233, pp. 755-758.
Chothia, C., Lesk, A. M., Tramontano, A., Levitt, M., Smith-Gill, S. J., Air, G., Sheriff, S., Padlan, E. A., Davies, D. R., Tulip, W. R., Colman, P. M., Alzri, P. M. and Poljak, R. J. (1989). Conformations of Immunoglobulin Hypervariable Regions. Nature (London) 342, pp. 877-883.
Cornette, J. L., Cease, K. B., Margalit, H., Spouge, J. L., Berzofsky, J. A. and Delisi, C. (1987). Hydrophobicity Scales and Computational Techniques for Detecting Amphipatic Structures in Proteins. Journal of Molecular Biology 195.3, pp. 659-685.
Darsley, M. and Rees, A. (1985), EMBO J. 4, pp. 383-392.
Davies, D., Sheriff, S. and Padlan, E. (1988). Antibody Antigen Complexes. J. Biol. Chem. 263, pp. 10541-10544.
de la Paz, P., Sutton, B., Darsly, M. and Rees, A. (1986). Modelling of the Combining Sites of Three Anti-lysozyme Monoclonal Antibodies and of the Complex Between One of the Antibodies and Its Epitope. EMBO J. 5, pp. 415-425.
Fine, R., Wang, H., Shenkin, P., Yarmush, D. and Levinthar, C. (1986). Predicting Antibody Hypervariable Loop Conformations ii: Minimization and Molecular Dynamics Studies of McPC603 from Many Randomly Generated Loop Conformations. Proteins: Struct., Funct., Genet. 1, pp. 342-362.
Go, N. and Sheraga, H. (1970). Ring Closure and Local Conformational Deformations of Chain Molecules. Macromolecules 3, pp. 178-187.
Gorman, S., Clark, M., Rutledge, E., Cobbold, S. and Waldman, H. (1991). Reshaping a Therapeutic CD4 Antibody. Proc. Natl. Acad. Sci. U.S.A. 88, pp. 4181-4185.
Grunow, R., Jahn, S., Porstman, T., Kiessig, T., Steinkeller, H., Steindl, F., Mattanovich, D., Gurtler, L., Deinhardt, F., Katinger, H. and von R., B. (1988). The High Efficiency, Human B Cell Immortalizing Heteromyeloma cb-f7. J. Immunol. Meth. 106, pp. 257-265.
Jeffrey, P. (1989). The Structure and Specificity of Immunoglobulins. D. Phil. Thesis, University of Oxford.
Lascombe, M., Alzari, P., Boulot, G., Salujian, P., Tougard, P., Berek, C., Haba, S., Rosen, E., Nisonof, A. and Poljak, R. (1989). Three-dimensional Structure of Fab r19.9, A Monoclonal Murine Antibody Specific for the pazobenzenearsonate Group. Proc. Natl. Acad. Sci. U.S.A. 86, p. 607.
Lee, C. and Levitt, M. (1991). Accurate Prediction of the Stability and Activity Effects of Site-directed Mutagenesis on a Protein Core. Nature 352.6334, pp. 448-451.
Lee, C. and Subbiah, S. (1991). Prediction of Protein Side-chain Conformation by Packing Optimization. Journal of Molecular Biology 217.2, pp. 373-388.
Marquart, M., Deisenhofer, J. and Huber, R. (1980), Crystallographic Refinement and Atomic Models of the Intact Immunoglobulin Molecule KOL and Its Antigen-binding Fragment at 3.0 .ANG. and 1.9 .ANG. Resolution. J. Mol. Biol. 141, pp. 369-391.
Martin, A. C. R. (1990). Molecular Modelling of Antibody Combining Sites. D. Phil. Thesis, University of Oxford.
Martin, A. C. R., Cheetham, J. C. and Rees, A. R. (1989). Modelling Antibody Hypervariable Loops: A Combined Algorithm. Proc. Natl. Acad. Sci. U.S.A. 86, pp. 9268-9272.
Martin, A. C. R., Cheetham, J. C. and Rees, A. R. (1991). Modelling Antibody Hypervariable Loops using a `Combined Algorithm`. Meth. Enz. In press.
Moult, J. and James, N. (1986). Proteins: Struct., Funct., Genet. 1, p. 146.
Padlan, E., Davies, D., Pecht, I., Givol, D. and Wright, C. (1976). Model Building Studies of Antigen-binding Sites: The Hapten-Binding Site of MOPC-315. Cold Spring Harbor Quant. Symp. Biochem. 41, pp. 627-637.
Palmer, K. and Sheraga, J. (1991). Standard-geometry Chains Fitted to X-ray Deviated Structures: Validation of the Rigid-geometry Approximation. I. Chain Closure through a Limited Search of Loop Conformations. J. Comp. Chem. 12, pp. 505-526.
Paul, P., Burney, P., Campbell, M. and Odguthorpe, D. (1990). The Conformational Preferences of .gamma.-lactam and Its Role in Constraining Peptide Structure. J. Comp.-aided. Mol. Des. 4, pp. 239-253.
Rose, D. R., Strong, R. K., Margolis, M. N., Gefter, M. L. and Petsko, G. A. (1990). Crystal Structure of the Antigen-binding Fragment of the Murine Anti-arsonate Monoclonal Antibody 36-71 at 2.9 .ANG. Resolution. Proc. Natl. Acad. Sci. U.S.A. 87, pp. 338-342.
Rudikoff, S., Satow, Y., Padlan, E. A., Davies, D. R. and Potter, M. (1981). Kappa Chain Structure from a Crystallized Murine Fab': The Role of the Joining Segment in Hapten Binding. Mol. Immunol. 18, pp. 705-711.
Schilling, J., Clevinger, B., Davie, J. M. and Hood, L. (1980). Amino Acid Sequence of Homogeneous Antibodies to Dextran and DNA Rearrangements in Heavy Chain V-region Gene Segments. Nature (London) 283, pp. 35-40.
Strong, R., Campbell, R. , Rose, D., Petsko, G., Sharon, J. and Margolies, M. (1991). Three-dimensional Structure of Murine Anti-pazophenylarsonate Fab 36-71.1, X-ray Crystallography, Site-directed Mutagenesis, and Modeling of the Complex with Hapten. Biochemistry 30, pp. 3739-3748.
Thornton, J., Sibanda, B., Edwards, M. and Barlow, D. (1988). Analysis, Design and Modification of Loop Regions in Proteins. BioEssays 8, pp. 63-69.
Tramontano, A. Chothia, C. and Lesk, A. (1989). Structural Determinants of the Conformations of Medium-sized Loops in Proteins. Proteins: Struct., Funct., Genet. 6, pp. 382-394.
Wilson, I. et al., Presented at Structure and Function Meeting in Honour of Sir David Phillips, 1-3 Jul., 1991, Oxford, UK.
Claims
Certificate of Correction present
The claims shown below may not include correction changes. Use the corrected PDF for authoritative text.
What is claimed is:
1. A method for producing a humanized rodent antibody or fragment thereof by resurfacing, said method consisting essentially of: (a) generating position alignments from relative accessibility distributions from x-ray crystallographic structures of a pool of rodent and human antibody heavy and light chain variable regions to give a set of heavy and light chain variable region framework surface exposed positions wherein the alignment positions for all rodent and human variable regions are at least about 98% identical; (b) defining for a rodent antibody or fragment thereof a set of heavy and light chain variable region framework surface exposed amino acid residues using said set of heavy and light chain variable region framework surface exposed positions generated in said step (a); (c) identifying from human antibody amino acid sequences a set of heavy and light chain variable region framework surface exposed amino acid residues that is most closely identical to said set of rodent surface exposed amino acid residues defined in said step (b); (d) substituting, in the variable region framework amino acid sequence of said rodent antibody or fragment thereof, said set of heavy and light chain variable region framework surface exposed amino acid residues defined in said step (b) with said set of heavy and light chain variable region framework surface exposed amino acid residues identified in said step (c); (e) constructing three-dimensional models of said variable region of said rodent antibody or fragment thereof and of said variable region of said rodent antibody or fragment thereof resulting from the substituting specified in said step (d); (f) comparing said three-dimensional models constructed in said step (e), and identifying any amino acid residues from said sets identified in said steps (b) or (c), that are within 5 .ANG.ngstroms of any atom of any residue of the complementarity determining regions of said rodent antibody or fragment thereof; (g) changing any residues identified in said step (f) from the human to the original rodent amino acid residue to thereby define a humanizing set of surface exposed amino acid residues; (h) replacing the set of rodent antibody variable region framework surface exposed amino acid residues defined in said step (b) with the humanizing set of variable region framework surface exposed amino acid residues defined in said step (g); and (i) producing said humanized rodent antibody or fragment thereof having binding specificity.
2. The method of claim 1, wherein said rodent antibody is an antibody fragment.
3. The method of claim 2, wherein said rodent antibody fragment is a single chain antibody, a F.sub.V fragment, a Fab fragment, a F(ab').sub.2 fragment or a Fab' fragment.
4. The method of claim 1 or 2, wherein said step (c) identifies a set of naturally paired heavy and light chain surface exposed amino acid residues that is most closely identical to said set of surface exposed amino acid residues defined in said step (b).
5. The method of claim 1 or 2, wherein the set of heavy and light chain surface exposed positions in step (a) is identical in at least 98% of said sequence alignment positions.
6. The method of claim 1 or 2, wherein said surface exposed amino acid residues are those residues whose solvent accessibility is above 30%.
7. The method of claim 1 or 2, wherein the rodent antibody or fragment thereof is a murine antibody.
8. The method of claim 1 or 2, wherein the rodent antibody or fragment thereof is murine antibody anti-N901.
9. The method of claim 1 or 2, wherein said set of framework positions of surface exposed amino acid residues is defined by the set shown in Table 1 and the alignments set forth in FIGS. 3A and 3B: ______________________________________ Posi- tion Human Mouse ______________________________________ Light chain 1 D 51 E 34 A 5 S 5 D 76 Q 9 E 6 3 V 38 Q 24 S 24 Y 6 V 63 Q 22 L 5 5 T 61 L 37 T 87 9 9 26 S 26 G 17 A 14 L 7 S 36 A 29 L 17 P 5 15 P 62 V 25 L 12 L 47 P 30 V 8 A 7 18 R 57 S 18 T 13 P 6 R 38 K 22 S 13 Q 12 T 9 46 P 94 P 82 S 9 47 G 89 G 71 D 18 51 K 43 R 31 K 70 Q 13 R 8 T 5 63 G 91 G 98 66 D 43 S 25 A 9 D 38 A 26 S 26 73 S 96 S 90 I 5 86 P 44 A 27 S 17 T 8 A 50 P 11 T 8 E 7 Q 6 87 E 71 D 11 G 7 E 91 D 6 111 K 74 R 12 N 6 K 93 115 K 54 L 40 K 87 L 5 116 R 60 G 33 S 5 R 89 G 9 117 Q 50 T 37 E 6 P 6 A 74 Q 14 P 5 R 5 Heavy chain 118 E 47 Q 46 E 59 Q 29 D 10 120 Q 83 T 7 Q 68 K 26 122 V 59 L 15 Q 13 Q 57 V 27 L 5 K 5 126 G 54 A 23 P 18 G 36 9 30 A 29 127 G 53 E 22 A 14 D 7 E 45 G 43 S 6 128 L 61 V 31 F 7 L 96 130 K 46 Q 41 E 5 K 52 Q 27 R 17 131 P 95 P 91 A 5 132 G 74 S 16 T 7 G 82 S 17 136 R 53 K 23 S 17 T 7 K 66 S 17 R 13 143 G 96 G 98 145 T 46 S 32 N 9 I 7 T 63 S 19 N 7 A 5 D 5 160 P 84 S 10 P 89 H 7 161 G 93 G 71 E 24 162 K 76 Q 10 R 8 K 50 Q 30 N 10 H 5 183 D 26 P 25 A 17 Q 10 T 7 E 31 P 22 D 17 A 12 Q 11 184 S 70 K 9 P 8 K 42 S 37 T 6 186 K 53 Q 22 R 7 N 5 K 83 Q 7 187 G 66 S 21 T 5 G 62 S 18 D 10 195 T 30 D 26 N 19 K 7 T 36 K 30 N 26 D 6 196 S 91 S 76 A 16 197 K 65 I 8 T 8 R 5 S 46 K 34 Q 11 208 R 46 T 18 K 17 D 6 T 55 R 26 K 8 209 A 50 P 21 S 13 T 8 S 67 A 14 T 11 210 E 46 A 18 D 13 S 9 Z 8 V 5 E 88 D 7 212 T 91 T 53 S 43 222 G 17 D 11 P 10 Y 9 V N 8 D 67 A 18. ______________________________________
10. A resurfaced rodent antibody or fragment thereof made by the method of claim 1.
11. The resurfaced antibody or fragment thereof of claim 10, wherein said rodent antibody is an antibody fragment.
12. The resurfaced antibody or fragment thereof of claim 10, wherein said rodent antibody fragment is a single chain antibody, a F.sub.V fragment, a Fab fragment, a F(ab').sub.2 fragment or a Fab' fragment.
13. The resurfaced antibody or fragment thereof of claim 10 or 11, wherein said surface exposed amino acid residues are those residues whose solvent accessibility is above 30%.
14. The resurfaced antibody or fragment thereof of claim 10, or 11, wherein the rodent antibody or fragment thereof is a murine antibody.
15. The resurfaced antibody or fragment thereof of claim 10 or 11, wherein the rodent antibody or fragment thereof to be humanized is murine antibody anti-N901.
16. The resurfaced antibody or fragment thereof of claim 10 or 11, wherein said set of framework positions of surface exposed amino acid residues is defined by the set shown in Table 1 and the alignments set forth in FIGS. 3A and 3B: ______________________________________ Posi- tion Human Mouse ______________________________________ Light chain 1 D 51 E 34 A 5 S 5 D 76 Q 9 E 6 3 V 38 Q 24 S 24 Y 6 V 63 Q 22 L 5 5 T 61 L 37 T 87 9 9 26 S 26 G 17 A 14 L 7 S 36 A 29 L 17 P 5 15 P 62 V 25 L 12 L 47 P 30 V 8 A 7 18 R 57 S 18 T 13 P 6 R 38 K 22 S 13 Q 12 T 9 46 P 94 P 82 S 9 47 G 89 G 71 D 18 51 K 43 R 31 K 70 Q 13 R 8 T 5 63 G 91 G 98 66 D 43 S 25 A 9 D 38 A 26 S 26 73 S 96 S 90 I 5 86 P 44 A 27 S 17 T 8 A 50 P 11 T 8 E 7 Q 6 87 E 71 D 11 G 7 E 91 D 6 111 K 74 R 12 N 6 K 93 115 K 54 L 40 K 87 L 5 116 R 60 G 33 S 5 R 89 G 9 117 Q 50 T 37 E 6 P 6 A 74 Q 14 P 5 R 5 Heavy chain 118 E 47 Q 46 E 59 Q 29 D 10 120 Q 83 T 7 Q 68 K 26 122 V 59 L 15 Q 13 Q 57 V 27 L 5 K 5 126 G 54 A 23 P 18 G 36 9 30 A 29 127 G 53 E 22 A 14 D 7 E 45 G 43 S 6 128 L 61 V 31 F 7 L 96 130 K 46 Q 41 E 5 K 52 Q 27 R 17 131 P 95 P 91 A 5 132 G 74 S 16 T 7 G 82 S 17 136 R 53 K 23 S 17 T 7 K 66 S 17 R 13 143 G 96 G 98 145 T 46 S 32 N 9 I 7 T 63 S 19 N 7 A 5 D 5 160 P 84 S 10 P 89 H 7 161 G 93 G 71 E 24 162 K 76 Q 10 R 8 K 50 Q 30 N 10 H 5 183 D 26 P 25 A 17 Q 10 T 7 E 31 P 22 D 17 A 12 Q 11 184 S 70 K 9 P 8 K 42 S 37 T 6 186 K 53 Q 22 R 7 N 5 K 83 Q 7 187 G 66 S 21 T 5 G 62 S 18 D 10 195 T 30 D 26 N 19 K 7 T 36 K 30 N 26 D 6 196 S 91 S 76 A 16 197 K 65 I 8 T 8 R 5 S 46 K 34 Q 11 208 R 46 T 18 K 17 D 6 T 55 R 26 K 8 209 A 50 P 21 S 13 T 8 S 67 A 14 T 11 210 E 46 A 18 D 13 S 9 Z 8 V 5 E 88 D 7 212 T 91 T 53 S 43 222 G 17 D 11 P 10 Y 9 V N 8 D 67 A 18. ______________________________________
Patent Citations (11)
Non-Patent Literature (12)
- Gorman et al., "Reshaping a Therapeutic CD4 Antibody", Proc. Natl. Acad. Sci., U.S.A., vol. 88, (1991) pp. 4181-4185.
- Padlan, "A Possible Procedure for Reducing the Immunogenicity of Antibody Variable Domains While Preserving Their Ligand-Binding Properties", Molecular Immunology, vol. 28, No. 4/5, (1991) pp. 489-498.
- Bobrzecka et al., "The Method of Controlled Rearrangement of Protein Disulphides and Its Use for Synthesis of Chimeric Immunoglobulin G", Immunology Letters, vol. 2, (1980) pp. 151-155.
- Konieczny et al., "The Combination of IgM Subunits and Proteolytic IgG Fragments by Controlled Formation of Interchain Disulphides", Haematologia, vol. 14, No. (1), (1981) pp. 95-99.
- Morrison et al., "Chimeric Human Antibody Molecules: Mouse Antigen-Binding Domains with Human Constant Region Domains", Proc. Natl. Acad. Sci., U.S.A., vol. 81, (1984) pp. 6851-6855.
- LoBuglio et al., "Mouse/Human Chimeric Monoclonal Antibody in Man: Kinetics and Immune Response", Proc. Natl. Acad. Sci., U.S.A., vol. 86, (1989) pp. 4220-4224.
- Martin et al., "Molecular Modeling of Antibody Combining Sites", Methods In Enzymology, vol. 203, (1991) pp. 121-153.
- Martin et al., "Modeling Antibody Hypervariable Loops: A Combined Algorithm", Proc. Natl. Acad. Sci., U.S.A., vol. 86, (1989) pp. 9268-9272.
- Morrison, "In Vitro Antibodies: Strategies for Production and Application", Annu. Rev. Immunol., vol. 10, (1992) pp. 239-265.
- Waldmann, "Monoclonal Antibodies in Diagnosis and Therapy", Science, vol. 252, (1991) pp. 1657-1662 pp. 121-153.
- Eduardo A. Padlan, A Possible Procedure for Reducing the Immunogenicity of Antibody Variable Domains While Preserving Their Ligand-Binding Properties, Molecular Immunology, vol. 28, pp. 489-498, 1991.
- Stanfield et al. TibTelh vol. 12 1994 p. 275.