Demystifying humanization

 Antibodies are composed of two identical heavy chains and two light chains, with the N-terminal domain of each chain, known as the variable heavy (VH) and variable light (VL) domains being responsible for antigen binding. Variable domains contain two b-sheets with loops connecting the b-strands (1). Three of the loops in each variable domain are highly variable in both length and amino acid composition and are thus referred to as hypervariable (HV) loops. The regions between the HV loops are known as framework regions (FR). HV loops are denoted as H1, H2, H3 and L1, L2 and L3 for the VH and VL respectively. It is the interplay of these 6 loops that determines antibody specificity.

Prior to the first antibody structure Wu and Kabat (2) aligned primary amino acid sequences of available antibodies and identified the regions involved in antigen binding as the three segments on each variable domain with highest variability. These are known as complementarity determining regions (CDRs). Therefore, CDR is a sequence based definition and HV is a structural based one. Despite this the terms are often used interchangeably. The CDRs and HV loops mostly overlap, as shown in the figure below, but this depends on the CDR scheme that is utilised (covered in more detail below). Further analysis of antibody sequences (3) and early antibody-antigen complexes (4) demonstrated that come CDR positions show less variability, with such residues playing a structural role, while those sites with highest variability dominate antibody-antigen interactions. This smaller subset of CDRs has been termed abbreviated CDRs (aCDRs) or, more commonly, specificity determining regions (SDRs).

Chothia and Lesk (5) analysed the first antibody structures and discovered that all HV-loops except H3 form a small number of main-chain conformations or canonical structures. Canonical structures are determined by the HV length and conserved amino acids in the HV and FR regions known as canonical structure residues (CSRs) (5-8). Foote and Winter (9) then described that a number of residues underlying the CDRs (16 in the VH and 14 in the VL) are responsible for stabilising HV loop structure and position. These amino acids fine-tune antibody specificity and are collectively called Vernier Zone (VZ) positions.

In addition to the non-CDR amino acids that support the CDRs and determine their canonical structure the amino acids at the interface of the VH and VL domain can also play a role as they determine the mutual orientation of the two domains. Although there is much plasticity in VH-VL pairing, humanization at these sites can occasionally have a significant impact on antigen binding affinity and so these sites should also be taken into consideration.

In addition to these interconnected and often overlapping terms to define amino acid positions that play a role, either directly or indirectly, in antigen binding there are also numerous antibody numbering schemes with differing CDR definitions. Antibody numbering is not as straightforward as it may seem. Antibodies are of uneven length with insertions and deletions. This is especially true in the CDR regions but also in the FRs and in constant domains with differences across subtypes within and across species. The solution is to devise a numbering scheme that creates a consistent scaffold based on a large number of antibodies that allows for certain conserved amino acids or motifs to be fixed. 

The first numbering scheme was described by Wu and Kabat, and is now commonly referred to simply as the Kabat scheme (2, 10). Based on amino acid sequences only they assigned variability scores at each position and identified three hypervariable regions in each variable domain, which they referred to as CDRs. In 1987, Chothia and Lesk introduced a structure based numbering scheme, which is now commonly known as the Chothia scheme (5). They aligned crystal structures and defined the loops that form the CDRs. More recently this scheme has been revised to correct for insertion/deletion positions in the Martin scheme (11). In 1997 Lefranc and colleagues introduced a new standardized system for all protein sequences of the immunoglobulin superfamily, including antibody variable and constant domains (12-14). This scheme, known as IMGT, is based on amino acid sequence alignment of the germline V-genes and, along with a variety of associated tools, it has become adopted by the World Health Organization. In my experience these four schemes are the most commonly used ones, with Kabat and IMGT being the most popular. It should be noted though that there are a number of other schemes such as AHo (15) and Gelfand (16).

The early success of mouse monoclonal antibodies led to the U.S. Food and Drug Administration (FDA) approval of the first therapeutic antibody OKT3 (muromonab) in 1986, as a treatment of kidney transplant rejection. However, most mouse antibodies were shown to have limited use as therapeutic agents because of a short serum half-life, an inability to trigger human effector functions and in particular they were recognized by the patients’ immune systems as foreign proteins resulting in a human anti-mouse antibody (HAMA) response. In an attempt to reduce the immunogenicity of the mouse antibodies, genetic engineering was used to generate chimeric antibodies containing human constant domains and the mouse variable domains to retain the specificity (17, 18). The process of humanizing an antibody, which takes this a step further, was first described by Sir Greg Winter and colleagues in a publication in 1986 (19) and the accompanying patent EP0239400A2. Winter describes a method for grafting the CDR regions from an original mouse antibody (the donor) onto a human framework (the acceptor). Winter also states that framework regions may also play a role in binding, as had been previously described in a crystal structure (Amit et al, 1986). However, the original work did not specifically define which FR amino acids may be critical. This more nuanced humanization was defined by 3 main families of patents/publications from Protein Design Labs (Queen), Celltech (Adair) and Genentech (Carter) (20-22). The associated patents are as follows:

• Queen: US7,022,500; US5,693,762; US5,585,089; US5,530,101; US6,180,370
• Adair: US7,566,771; US7,262,050; US7,244,832; US7,244,615; US7,241,877; US6,632,927; US5,859,205; US5,929,212
• Carter: US6,407,213; US6,054,297; US6,639,055; US6,719,971; US6,800,738

Queen is a structure based approach and describes four criteria to humanize an antibody:

• The use of a human acceptor framework from a particular human immunoglobulin that is homologous to the non-human donor or a consensus framework from many human antibodies.
• Use the donor amino acid rather than the human acceptor residue if the human residue is unusual and the donor residue is typical for human sequences at a specific residue in the framework.
• Use the donor framework amino acid residue rather than the acceptor at positions immediately adjacent to the CDRs
• Use the donor amino acid residue at framework positions at which the amino acid is predicted to have a side chain atom within about 3 Å of the CDRs in a three-dimensional immunoglobulin model and to be capable of interacting with the antigen or with the CDRs of the humanized immunoglobulin.

Adair describes a hierarchy of positions within the framework of the variable regions at which the amino acid identities of the residues are important for obtaining CDR-grafted products with satisfactory binding affinity. The hierarchy of framework residues is described as follows:

For heavy chains:

• Choose donor residues at positions 2, 4, 6, 23-25, 36, 37, 39, 47-49, 71, 73, 78, 93, 94, 103, 104, 106 and 107.
• To further optimise affinity consider choosing donor residues at positions 1, 3, 9, 11, 18, 20, 38, 41, 46, 67, 69, 72, 76, 80, 82, 87, 88, 91, 108, 110 and 112.

For light chains:

• Chose donor residues at positions 2, 4, 6, 35, 38, 44, 46-49, 58, 62, 64-69, 71, 85, 87, 98, 99, 101, 102.
• To further optimise affinity consider choosing donor residues at positions 1, 3, 10, 12, 21, 24, 37, 40, 45, 60, 63, 70, 73, 80, 103, 105.

Carter describes a process for CDR grafting onto a consensus sequence generated through computer modelling. A number of framework amino acids are described as potential sites for incorporation of the donor amino acid:

• For heavy chains: 2, 4, 24, 36, 37, 39, 43, 45, 49, 58, 60, 67-69, 70, 73-76, 78, 91-93, 103
• For light chains: 4, 35, 36, 38, 43, 44, 46, 58, 62, 64, 65-71, 73, 85, 87, 98

It is clear that there is no single approach to humanization, and many variations on these have been described over the years. Despite this, they all tend to follow the same basic principles, which I describe in more detail below.

Antibody humanization processes can be grouped into two distinct categories: rational and empirical. Rational approaches involve the systematic design, production and assessment of a small number of variants (typically 5-50) with iterative design-build-test cycles performed if the initial humanized antibodies do not meet the desired criteria. In contrast, empirical approaches rely on the generation and screening of large libraries, typically via phage, ribosome, yeast or mammalian display. This section will cover rational methods, which are the most common approach to antibody humanization. All rational methods follow the same basic design cycle consisting of three inter-connected steps, as shown in the image below. 

1. Selection of amino acids to graft

The idea is simple, select the amino acids in the parental antibody (also known as the donor) that are responsible for antigen binding and then graft these onto a human framework. It is unusual at the early stages of an antibody development project to have access to an antibody-antigen structure and so there is often no antibody specific structural information to guide this process. Instead it relies on the knowledge that antibody structure at a top level is very consistent as described earlier. Typically amino acids from the HV loops, CDRs or SDRs are selected for grafting. By far the most common approach is to use CDRs (by one of any number of definitions). In the early years of antibody humanization this process was known as antibody reshaping but now tends to simply be referred to as CDR grafting, even though it could be HV loops or SDRs that are selected for grafting. 

2. Selection of human framework

The second step in the humanization process is the selection of a human framework, often referred to as the acceptor. In early years when very few human antibody sequences were available donor amino acids were grafted onto a fixed human framework irrespective of homology. These days a ‘best-fit’ approach is typically employed where human sequences with highest homology to the non-human antibody are selected as acceptors to increase chances of successfully retention of affinity (23). A number of alternative but less frequently utilised methods of framework selection have also been described. As first described by Carter, a consensus human sequence can be used as the template. Alternatively, Foote reported on the process of superhumanization (24). In this approach framework homology is irrelevant and instead frameworks are selected based on those that contain the same or closely related canonical structures. Within the genes sharing canonical structures those with the most homologous CDRs are then chosen as framework donors. 

Irrespective of the method there are two sources of human sequence: mature and germline gene sequences. Antibody protein molecules are encoded by several recombined germline gene segments prior to antigen exposure and maturation. These are known as germline sequences. Mature sequences are products of the immune response and as such have undergone hypersomatic mutation.

Although both approaches have proved successful, in my experience the use of germline sequences as acceptor frameworks is more common. This may be due to a number of potential advantages. Due to the random process of hypersomatic mutation, mature sequences may contain immunogenic amino acids. Additionally, comparison of germline and mature antibody structures in both free and antigen-bound state have shown that germline sequences are more flexible (25, 26). This plasticity in theory might enable the accommodation of a greater range of CDRs.

3. Retention of donor content

Straight-forward grafting of CDRs onto an appropriate human framework often leads to a decrease in antigen binding affinity due to incompatibilities between the human framework and non-human CDRs. Therefore, a third critical step is required to prevent the loss of binding affinity. This is known as the introduction of back-mutations, or rather the retention of a certain amount of donor (non-human) amino acids within the framework regions. The number and location of back mutations is highly dependent on the antibody. This makes humanization somewhat unpredictable and a bit more of an art than a science. Typically the amino acids of most interest for back mutation are those belonging to the Vernier Zone (VZ) and canonical structure residues (CSRs), although amino acids outside of these regions can also play a role, such as those at the VH-VL interface. Humanization is ultimately a balance between increasing humanness and maintaining binding activity. For this reason it is rare to generate just one humanized variant. Instead a small panel of humanized variants are typically generated with differing numbers and positions of back-mutations, as well as possibly using different human frameworks.

Alternatively, the process of veneering or resurfacing (not to be confused with reshaping) has also been described (27-29). This method shares the same grafting and donor selection steps but all non-surface exposed amino acids are retained as the non-human donor amino acids. This is to try and maintain the core of the variable domain and thus minimize perturbation of the amino acids determining the specificity of the antibody while eliminating potentially immunogenic B-cell epitopes on the surface. 

In contrast to the rational humanization design process described above, empirical approaches rely on the generation and screening of large libraries, typically via phage, ribosome, yeast or mammalian display. Due to the cost and time required in such library-based approaches these are often better suited to affinity maturation rather than the more straight forward process of humanization. Despite this, these empirical approaches have been utilised for humanization and so will be covered briefly here.

Framework libraries (30) and framework shuffling (31) are two similar methods. In both a library of human frameworks is built to determine the framework(s) that best supports the CDRs. In the former method the frameworks contain variants, typically at the sites of buried amino acids critical for maintenance of canonical structures. In the latter combinatorial libraries of variants are not generated.

In contrast to the FR library or shuffling approaches, in which CDRs remain non-human, guided selection allows for the isolation of fully human antibodies (32). The first fully human antibody approved by the FDA was isolated in this manner (33). The strategy consists of combining one of the non-human variable domains with a library of human variable domains. The resulting antibodies are screened to select a human variable domain that maintains binding activity. This selected variable domain is then fixed and combined with a human library of the alternative variable domain, resulting in selection of a fully human VH/VL pair. A potential disadvantage of shuffling one antibody chain while maintaining the other is potential epitope drift (34). To try to prevent this one or more CDRs are commonly retained (35, 36).

Humaneering is based on the experimental identification of minimum specificity determinants by sequential replacement of non-human fragments into human frameworks and assessment of binding. It starts with regions of CDR-H3 and L3 add progressively replaces other regions of the non-human antibody into human FRs. 

Although affinity maturation is not my area of expertise, nor a service I offer, it would be wrong to discuss humanization without mentioning affinity maturation. The aim of humanization is typically to maintain the affinity of the original antibody, whereas with maturation the aim is to increase the binding affinity, often by an order of magnitude or more. Antibodies coming out of a discovery campaign often don’t have the affinity required for therapeutic use, potentially due to an affinity ceiling in a mammalian immune response, limitation of the discovery platform being used or the fact that it is a challenging target/epitope. In simple terms at least, affinity maturation therefore offers an opportunity to take a pre-existing antibody with an unacceptable affinity and generate one that meets the desired criteria. As with humanization, affinity maturation can be broken down into rational and empirical approaches.

Rational approaches typically utilise structure guided computational design to generate a small panel of variants for testing. This relies on existing crystal structure information or highly accurate modelling as well as expert structural biology skills to predict the impact of one or multiple amino acid substitutions at the antibody-antigen interface. In years to come with AI/ML aided design I expect this kind of approach to become more routine but for now the majority of affinity maturation is performed by empirical library based approaches. Libraries can be designed in a wide variety of different ways with the two principal tasks being the selection of positions for diversification (e.g. one or multiple CDRs, vernier zone positions etc) and the choice of amino acids to substitute in. These libraries are then screened and antibodies with new, hopefully improved, sequences are isolated. It should also be noted that although humanization and affinity maturation are often considered as separate steps they can be achieved simultaneously in library based approaches.

Section coming soon.