Applications and challenges in designing VHH-based bispecific antibodies: leveraging machine learning solutions

ABSTRACT

The development of bispecific antibodies that bind at least two different targets relies on bringing together multiple binding domains with different binding properties and biophysical characteristics to produce a drug-like therapeutic. These building blocks play an important role in the overall quality of the molecule and can influence many important aspects from potency and specificity to stability and half-life. Single-domain antibodies, particularly camelid-derived variable heavy domain of heavy chain (VHH) antibodies, are becoming an increasingly popular choice for bispecific construction due to their single-domain modularity, favorable biophysical properties, and potential to work in multiple antibody formats. Here, we review the use of VHH domains as building blocks in the construction of multispecific antibodies and the challenges in creating optimized molecules. In addition to exploring traditional approaches to VHH development, we review the integration of machine learning techniques at various stages of the process. Specifically, the utilization of machine learning for structural prediction, lead identification, lead optimization, and humanization of VHH antibodies.

KEYWORDS:

• Bayesian optimization
• bispecific
• HCAb
• machine learning
• multispecific
• nanobody
• VHH

Introduction

Over the past 10 y bispecific antibodies (BsAbs) have emerged as a promising therapeutic for a wide variety of disease indications, most notably in oncology, autoimmune disease, and inflammation. Due to the unique ability to engage at least two separate binding sites within the same antibody format, BsAbs have been used to engage multiple soluble targets (e.g. cytokines), act as bispecific cell engagers (e.g. immune cell engagers) or bind to two different epitopes on a single molecule (e.g. biparatopic antibodies).¹ BsAbs can consist of different types of antibody-based fragments, such as single-chain variable fragments (scFv) and fragment antigen-binding regions (Fabs), within the same antibody scaffold. Single-domain antibodies are an increasingly popular choice of building blocks for bispecific antibody construction due to their small size, ease of production, and stability at a wide range of temperatures and chemical environments. In particular, the variable heavy domains of heavy-chain-only antibodies (VHH), derived from camelid heavy-chain-only antibodies (HCAbs) are emerging as integral building blocks, increasing the number of formats available to researchers for BsAb construction. A key feature of VHHs lies in the absence of a light chain, offering a streamlined approach to assembly and production. The absence of a light chain provides modularity, enhancing their suitability as building blocks for bispecifics and enabling the creation of versatile multispecific formats beyond traditional bispecific antibody structures. Although VHH therapeutics such as Sanofi’s caplacizumab (CABLIVI) and Ablynx-developed ozoralizumab (Nanozora) have been approved, the exploitation of VHHs as components of bispecific antibodies is still in its infancy. Here, we discuss recent applications and challenges of using VHH domains in bi- and multispecific antibody production. In addition, we explore how machine learning-based approaches can be used to help design optimal VHH-based bispecific and multispecific antibodies.

The structure and engineering of VHHs

Conventional antibodies are composed of two heavy chains and two light chains forming a tetrameric immunoglobulin G (IgG) molecule (Figure 1a). These chains have variable regions that bind the target antigen and constant regions to determine the mechanism of antigen elimination. The variable heavy chain and variable light chain domains are each composed of three complementarity-determining regions (CDRs) that together form the antigen binding site. The serum from camelids, including camels, llamas, and alpacas contains not only the conventional IgG isotype antibodies but also homodimeric HCAbs, which lack the light chain and the first constant Ig domain of the heavy chain (CH1) (Figure 1b).² These HCAbs possess a single variable binding domain responsible for antigen binding, termed VHH (Figure 1c). Similarly, sharks produce a unique type of single-domain antibody known as VNARs which are likewise composed exclusively of heavy chains, demonstrating convergent evolutionary adaptations in immune defense mechanisms across species.³ As with variable domains from conventional antibodies, VHHs are composed of four separate frameworks (FW1-4), which are separated by three hypervariable complementarity-determining loops (CDR1-3) and adopt the typical Ig fold structure consisting of nine β-strands in two closely packed layers of anti-parallel β-pleated sheets, resulting in a β-barrel structure (Figure 1c).^4–6 In VHH domains, CDR3 loops are generally longer compared to those found in human VH domains.^7,8 A comparison of CDR3 lengths in human VH sequences and VHH sequences within the Observed Antibody Space (OAS)⁹ revealed that, on average, VHH domains have CDR3 loops that are 1.4 residues longer.¹⁰ Due to the long CDR3s, VHH domains can exist with structural conformations beyond those seen in conventional antibodies.¹¹ Indeed, camel and llama VHHs may possess an additional disulfide bond between CDR1 and CDR3 or CDR2 and CDR3, respectively, thereby increasing conformational stability of these long CDR3 loop structures.^12,13 The longer extension of the CDR3 loop is thought to allow the VHH to bind concave epitopes such as active sites of enzymes and pore regions of membrane receptors.^14–16

Figure 1. Schematic representation of a conventional antibody vs a camelid-derived heavy chain-only antibody. (A) Conventional antibodies are composed of two heterodimeric chains, two identical heavy chain fragments consisting of the constant heavy domains CH1-CH3 and the antigen binding variable heavy domain (highlighted in pink) and two identical light chains composed of a constant light-chain domain and a variable light chain domain (highlighted in blue). (B) HCAb antibodies are composed of two identical heavy chain-only fragments that lack a CH1 domain and only have a single variable heavy domain (VHH). (C) Ribbon representation of the crystal structure of a VHH domain (PDB: 1I3V) showing the complementary-determining loops, CDR1 in green, CDR2 in magenta and CDR3 in orange. Framework regions are shown in gray. 3D structure created using Pymol.

Representation of human and llama antibodies using oval shapes for domains connected together by lines into a Y-shape. One of the llama oval shapes has a cartoon associated showing the structure of the domain in three dimensions.

Why are VHHs good building blocks?

In the traditional antibody format, the pairing of the VH and VL chains typically depends on hydrophobic interactions. These interactions often limit the solubility of human VHs when they are not paired with a light chain. A distinctive feature of VHH domains is the presence of hallmark residues in FW2 at positions 37 (F/Y), 44 (E), 45 (R) and 47 (G) that help form a hydrophilic surface patch, providing VHH domains with increased solubility and stability in the absence of a variable light domain.^4,17,18 Conventional VH domains pair these positions with the light chain variable domain, but VHH domains have evolved to function in the absence of this pairing and therefore have escaped those evolutionary constraints (Figure 2). In addition, the long CDR3 in VHHs is also known to fold back over FW2, providing additional stability in the absence of a light chain.^19,20 The presence of these hallmark residues enables improved expression of VHH domains compared to a human VH domain, which may display both yield and aggregation issues upon expression. Extensive engineering work is often required to improve human VH domain expression and solubility.^21,22 “Camelization” can involve mutating the corresponding framework 2 positions in a human VHs to those observed in a VHH.^21,23,24

Figure 2. Structural representation of framework 2 residues in human VH and camelid VHHs. Ribbon representation of the crystal structure of (A) human VH (PDB: 7OBF) and (B) camelid-derived VHH (PDB: 1JTT) highlighting the framework 2 positions for each domain. Position 37: Human Valine, Camelid Phenylalanine (red), Position 44: Human Glycine, Camelid Glutamic acid (blue), position 45: Human Leucine, Camelid Arginine (yellow) and position 47: Human Tryptophan, Camelid Glycine (cyan); Kabat numbering. The interaction position of the human variable light chain with the human variable heavy chain is represented by the dotted line. The complementary-determining loops, CDR1 in green, CDR2 in magenta and CDR3 in orange. Framework regions are shown in gray. 3D structure created using Pymol.

A cartoon representation of human and llama heavy chain domains with arrows and lines representing the structure of each domain. Spheres coloured in blue, red, yellow, and cyan represent amino acids in three-dimensional space.

In addition to the increased solubility, VHHs can also display enhanced thermal stability, with some VHHs reaching melting temperatures above 60°C.^25,26 More importantly, VHHs can also retain antigen binding after heat denaturation due to their ability to refold into a functionally relevant conformation.^27–29 The ability to refold is an attractive property, particularly if refolding is independent of formulation conditions. Despite high levels of thermal stability, there have been many attempts to improve VHH stability even further by mutagenesis of critical residues. Detailed analysis of CDR loop formation during library construction has shown that the CDR loops of VHH domains may adopt a particular structure depending on several factors.^30–32 Murakami et al. postulated that VHHs may prefer antigenic epitopes that form clefts or deep pockets that the longer CDR3 loops may be able to penetrate. To this end, they analyzed the crystal structures of VHHs in complex with antigens and proposed a classification for paratope formation. Depending on CDR3 loop length and structure, they define the loop formations as either upright, half-roll or roll. Interestingly, for upright loops, they showed that FW2 residues are involved in paratope formation, and therefore these residues may need to be considered when designing VHH-based phage libraries. Moreover, Kinoshita et al. have shown that variation in FW2 residues, along with CDR1–3 loop length, may also influence CDR conformation.³³

Compared to immunoglobulins from other species such as mouse and rat, camelid VHH domains share a higher sequence similarity with human VH domains, for example, > 80% for VH3–23, therefore reducing the potential for immunogenicity.^12,34 Although several studies have shown a lack of immune response to VHH-based antibodies in both mice and humans,^35,36 a successful humanization strategy has been adopted to generate a universal VHH-based scaffold.³⁷ Despite numerous studies showing minimal immune responses to VHHs, individuals have been identified with preexisting anti-drug reactivity (PEAR) against the C-termini of some VHHs, which complicates the immunogenicity assessment of these biotherapeutics.^38,39 The detection of PEAR during clinical immunogenicity assessment is significant because it can obscure the true immunogenic response to the drug and potentially affect both the safety and efficacy profiles of the therapeutic agents. This phenomenon is also observed in fully human heavy chains, Fabs and scFvs.^40,41 A comprehensive understanding of the biological mechanisms underlying this occurrence is still required.⁴² In one study, the addition of two proline residues at the VHH C-terminus was shown to substantially reduce PEAR.³⁸ In another study, the introduction of a C-terminal alanine proved effective in diminishing the binding of preexisting human anti-VH antibodies during in vitro screening.⁴³ These findings suggest potential pathways for mitigating the immunogenicity challenges posed by these antibodies in therapeutic applications.

In terms of production, VHH domains can be made in most heterologous expression systems, including E. coli, S. cerevisiae, P. pastoris and mammalian cells. E. coli systems are often the preferred choice due to the ease of use, cost of production and yield. However, the lack of relevant post-translational modifications in E. coli may affect physicochemical properties and function of VHHs, particularly when compared to production in mammalian cells.^44–46

One potential issue of using VHHs for therapeutic applications is their small size (~15 kDa). This may lead to fast renal clearance, resulting in poor pharmacokinetics and limited biodistribution. For example, the 55 kDa bispecific T-cell engager blinatumomab has a serum half-life of several hours, thereby requiring continuous intravenous infusion.⁴⁷ However, the rapid clearance of VHHs has a distinct advantage for some diagnostic applications, such as bio-imaging or for therapeutic applications where a short plasma half-life is sufficient or preferred.^48,49 To extend the half-life of VHH, binding domains against albumin can be fused into small antibody fragments such as scFv and single-domain antibody (sdAb) or an IgG Fc domain can be incorporated in the molecule. The inclusion of albumin-binding sdAb is an effective strategy to improve the serum half-life of small nanobody therapeutics, as the extended half-life of albumin is approximately 3 weeks mainly due to neonatal Fc receptor (FcRn)-mediated recycling, a property it has in common with the Fc domain of canonical IgG antibodies.⁵⁰ The genetic fusion to an IgG Fc domain also results in a longer plasma half-life and offers the added advantage of being able to use Fc-mediated purification and analytical strategies.

Generation of VHH antibody domains

VHH antibodies are typically sourced from the blood of naive or immunized llamas or alpacas and used to generate a VHH antibody phage library or from synthetic libraries. Muyldermans recently published a comprehensive overview covering different methods for nanobody generation.⁵¹ One of the main advantages of an immunized VHH library is that the VHH domains have been affinity matured in vivo by somatic hypermutation and expressed on the B cells, which may translate into a more favorable developability profile in vitro. As described above, camelid serum contains both conventional IgG antibodies and HCAbs. After serum collection from immunized animals, HCAb-specific IgGs are preferentially amplified by PCR and cloned into a phage display vector for in vitro selection purposes.

Recently, advances in computational approaches and data-mining techniques have allowed the development of synthetic VHH antibody library construction. Open source next-generation sequencing databases such as the Observed Antibody Space (OAS) and the Integrated Nanobody Database for Immunoinformatics (INDI) combined with structural information from the Protein Data Bank (PDB) has enabled the ability to construct VHH domain-based antibody libraries and allow complete control of structure-related sequences and removal of sequence-based manufacturing liabilities.^52–54 The application of large sets of antibody sequence analysis, combined with structural information, has led to the production of several sophisticated VHH-based libraries that focus on the unique properties of VHH CDR loop structure.^30,32,55,56

Using either an immunized or synthetic VHH-based phage display library, the combination of sequence and functional characterization and novel machine learning approaches is a powerful method for identifying therapeutic antibodies (Figure 3 and ML sections below).

Figure 3. Overview of VHH antibody discovery methods. VHH antibodies are typically sourced from either alpacas or llamas after immunization with an antigen of choice. An alternative source can be a naive library from llamas or alpacas, or a synthetic library design using a combination of sequence and structural information. Constructed libraries can be interrogated by phage display and combined with a machine learning approach to identify key attributes for affinity, developability and function.

Read the detailed description of this figure

A four section picture with cartoon llama shapes. One llama has a syringe above it to represent an immunized llama library. One llama has nothing above it to represent a naive llama antibody library, and a computer screen icon has protein sequences next to it, representing a semi- or synthetic antibody library. To the right of the cartoon llama and computer is a cartoon representation of a phage in rectangular shapes with small, colored balls on curving lines to represent llama antibodies. To the right of this image is a computer icon representing machine learning and to the right of the computer icon is llama antibodies represented by using oval shapes for domains connected together by lines into a Y-shape. Below the antibodies are two straight lines representing a graph with waving lines in the center of the graph representing data.

VHH architecture in bispecific antibodies

VHH topology can play a key role in the potency and biophysical properties of VHH-based bispecific antibodies, with the orientation of specific VHH domains known to affect binding and potency toward some targets as demonstrated by De Bruin et al. They generated two multispecific constructs consisting of an anti-Vγ9 Vδ2-TCR VHH domain, 6H4, and an anti-EGFR VHH domain, 7D12 fused via a flexible Gly4Ser linker. Both orientations were generated, with one construct having the anti-Vγ9 Vδ2-TCR VHH domain at the N-terminus, and the other having the anti-EGFR VHH domain at the N-terminus. De Bruin et al. reported that the orientation for an anti-Vγ9 Vδ2-TCR VHH 6H4 domain was not important to retain binding to the T-cell receptor (TCR) because the anti-Vγ9 Vδ2-TCR VHH 6H4 domain could bind at either the N- or C-terminal position. In contrast, the anti-epidermal growth factor receptor (EGFR) VHH position was critical with an N-terminal position favored. Moreover, the length of the Gly4Ser linker between the VHH domains had no influence on activity.⁵⁷

Positional effects were also documented for a tandem VHH bispecific molecule targeting CD16 (a natural killer (NK) cell marker) and EGFR.⁵⁸ The authors noted an N-terminal preference for each targeting VHH domain, with an N-terminal anti-CD16 VHH displaying a higher binding capacity for NK cells compared to an N-terminal anti-EGFR VHH domain and the reverse orientation displaying opposite effects. Interestingly, both molecules were effective at inducing NK cell-specific degranulation of EGFR-positive cells, but the bispecific with the N-terminal CD16-binding VHH domain was able to induce a stronger NK cell degranulation at lower concentrations.

The positional flexibility of VHH domains may depend on the targeted antigen and accessibility of the antigen-binding loops when the VHH is placed C-terminal to another protein domain. Alvarez-Cienfuegos et al. recently described an elegant example of producing trimeric molecules called trimerbodies. By fusing an anti-carcinoembryonic antigen (CEA) VHH N-terminally of a collagen trimerization domain (TIE^XVIII) they were able to express and purify monospecific trimeric VHH molecules that bound to CEA.⁵⁹ The addition of a Gly-Ser linker between each VHH to make a tandem trimerbody ensured stoichiometric presentation. The authors went on to show that the presentation of three different VHHs could produce a functional trispecific molecule, termed ATTACK (Asymmetric Tandem Trimerbody for T cell Activation and Cancer Killing).⁶⁰

While multispecific antibody constructs allow binding or neutralization of two or more targets with one molecule, they can also be engineered to target cells selectively. Using avidity as a guiding principle, a bispecific antibody could be engineered to bind only when two different targets are co-expressed on a cell, while ignoring cells that express only one of the two targets. Such an AND gate can be used to design, for instance, immune cell engagers with superior on-target off-tumor selectivity, resulting in better therapeutic windows while killing tumor cells expressing both targets and leaving normal cells expressing just one or the other target unharmed.^61,62 Here too, VHHs are ideally suited as building blocks for such multispecific immune cell engagers.

Importance of geometry/affinity/valency in VHH-based bispecifics

Engineering in multivalency either to multiple different targets or to the same target is an approach that could dramatically improve the selectivity and therefore the safety profile of, for example, an immuno-oncology antibody.⁶³

Increasing the valency by introducing multiple VHHs, which, for example, bind a target such as a tumor-associated antigen (TAA) at a lower affinity, can deliver avidity-driven selectivity. Such a design provides an inherent receptor density sensor enabling engagement only when the target TAA is present at a higher level than that found in healthy tissue, a property typically observed for many tumor targets. Since many of these targets retain low expression in healthy tissue, nonselective tumor-targeting antibodies (e.g. antibody–drug conjugate, immune cell engager molecules) would still engage normal cells and generate unwanted adverse events during treatment (Figure 4). A similar avidity-driven approach can be used to target two receptors co-expressed on a tumor cell but not on normal cells. An avidity-driven selective antibody should provide a much safer drug profile and reduce therapeutic adverse events.

Figure 4. Targeting tumor cells co-expressing two different receptors at the cell surface (left) not co-expressed on normal cells with a multispecific VHH, thereby minimizing antibody engagement against healthy cells.

A two section image showing how an antibody can bind to a tumor cell and T-cell at the same time. Gray shapes with two small lines projected outwards represent the cell membrane. Embedded in this membrane are colored rectangular boxes connected by thick lines. These structures represent the cell surface receptors responsible for tumor growth. On the left is an antibody represented by oval shapes connected by a wavy line, and the antibody is touching the t-cell and tumor cell at the same time. On the right, there is only one tumor cell with fewer receptors, and the antibody does not bind to the cell due to the low level of receptors.

Development of VHHs as bi-and multispecific therapeutics

Over the past 10 y many different formats for bispecific antibodies have been introduced, with many displaying more than two different antigen-binding sites on each arm (Figure 6). As complexity increased, the ability to purify homogeneous samples has also become more challenging. In particular, the standard monoclonal antibody format requirement to pair different light chains with their cognate heavy chains has resulted in unequal mixtures, often leading to extensive engineering solutions to address the problems. Several strategies have successfully overcome these challenges, such as the use of a common light chain or electrostatic engineering between heavy and light chains.^64–66 The use of single-domain VHHs overcomes these shortfalls due to the removal of the light chain pairing heterodimerization, dramatically improving the homogeneity of an overexpressed recombinant product. Along with greater product homogeneity, VHH biophysical properties, such as increased solubility and higher thermal stability compared to VH or VL alone, often also translate to a bispecific molecule with enhanced biophysical characteristics.

As an initial step, VHHs have been genetically fused together into a tandem VHH format using a flexible peptide linker such as a glycine-serine linker, yielding a bispecific molecule. The relatively small size of a tandem VHH molecule may provide improved tissue penetration compared to full IgG-based therapies. Caplacizumab, the first such tandem bivalent, but monospecific, VHH-based molecule to be approved for marketing, consists of two VHH domains separated by a short peptide linker (three alanine residues) illustrated a prototype architecture for similar bispecific formats.⁶⁷ Caplacizumab binds the von Willebrand factor protein and prevents platelet aggregation, the main contributing factor in acquired thrombotic thrombocytopenic purpura. Caplacizumab has no built-in mechanism to extend its circulation in plasma and, as such, has a short plasma half-life (~16–27 h), therefore requiring daily injections. Ozoralizumab, a bispecific and trivalent tandem VHH molecule originally developed by Ablynx and licensed to Taisho Pharmaceutical, has been approved for rheumatoid arthritis and is marketed in Japan. It consists of two anti-TNFα VHHs separated by an anti-human serum albumin (HSA) VHH.⁶⁸ These VHHs are fused via a simple G₄Sx2 peptide linker to improve flexibility, while the fusion to an anti-HSA nanobody offers an increase in plasma half-life to 18 d. Gefurulimab (ALXN-1720, Alexion) is another example of a bispecific VHH where the half-life of an anti-complement 5 VHH is improved by genetically fusing it to an anti-HSA VHH linked by a G₄Sx3 flexible peptide.⁶⁹ Gefurulimab is currently undergoing evaluation in a Phase 3 clinical trial as a treatment for myasthenia gravis (NCT05556096).

Further examples of tandem multispecific VHHs include M1095 and BI836880. M1095/ALX0761 (sonelokimab) is a trispecific tandem VHH targeting the cytokines IL17A and IL17F as well as HSA. An N-terminal anti-IL17F VHH is linked via a glycine-serine linker to an anti-HSA VHH followed by a C-terminally positioned anti-IL17A VHH. Originally discovered at Ablynx in collaboration with Merck Serono, it has been investigated as a treatment for patients with moderate-to-severe psoriasis (Phase 2b, NCT03384745), moderate-to-severe hidradenitis suppurativa (Phase 2, NCT05322473) and for patients with active psoriatic arthritis (Phase 2, NCT05640245).⁷⁰

BI836880 is a humanized trispecific nanobody that targets vascular endothelial growth factor (VEGF), angiopoietin 2 and HSA. Targeting the latter supports a longer plasma half-life of the molecule. Similar to M1095, the overall architecture of the therapeutic consists of two VHHs each binding to their respective ligands, separated by an anti-HSA VHH. In blocking two ligands and two separate downstream pathways, the bispecific aims to block both angiogenesis and tumor cell proliferation. It thus has potential for the treatment of solid tumors.^71,72

Unless a short plasma half-life is preferred (e.g., for imaging purposes or agonistic antibodies), a moiety that provides a half-life extension is usually added to the antibody construct. This can be a genetic fusion to a VHH that binds to HSA, as described above for M1095 and BI836880, or a genetic fusion to the Fc portion of IgG. In both instances, adding an Fc not only makes the molecules larger and above the typical cutoff size for renal clearance, but importantly, it also facilitates recycling via the FcRn receptor. An example of the latter is the anti-CXCR2 COH105.⁷³ In this construct, two VHHs each binding a different epitope on the CXCR2 chemokine receptor and linked together via a glycine-serine linker are fused to an Fc. As such, this forms a symmetric dimer of a biparatopic receptor antagonist with a half-life in cynomolgus monkeys and humans similar to an IgG. Here too, the final topology of the molecule was arrived at after experimental evaluation of constructs with different glycine-serine linker lengths and different sequential order of the individual anti-CXCR2 VHH elements. Other bispecifics built with VHHs have used a similar architecture. For instance, Alphamed Oncology has evaluated Erfonrilimab (KN046), a VHH-based bispecific targeting PD-L1 and CTLA4, as a treatment for cancer (NCT05149326, see Table 1 and Figure 5 for a topological representation of some key molecules).⁷⁴

Figure 5. Topological representation of examples of clinically relevant VHH-based therapeutic bispecifics.

A two section picture representing different types of antibody formats. On the left-hand side are five images one above the other. These images contain colored oval shapes representing llama or human antibody domains connected by wavy lines. On the right hand side of the image are three more colored oval shapes representing llama or human antibody domains connected by wavy lines. Below are two images that represent antibodies using oval shapes for domains connected together by lines into a Y-shape.

Table 1. Preclinical and clinical stage VHH-based bispecific molecules. All data correct as of 1st September 2023.

BsAb (Sponser)	Targets	Half life extension format	Indication	Status	Sequence (when available from REF^83)
Gefurulimab (ALXN-1720) Alexion/AZ	C5	HSA	Dermatomyositis Proteinuria Myasthenia gravis	Phase 3 NCT05556096	EVQLVESGGGLVKPGGSLRLSCAASGRPVSNYAAAWFRQAPGKEREFVSAINWQKTATYADSVKGRFTISRDNAKNSLYLQMNSLRAEDTA VYYCAAVFRVVAPKTQYDYDYWGQGTLVTVSSGGGGAGGGGAGGGGSEVQLVESGGGLVQPGGSLRLSCAASGRAHSDYAMAWFRQA PGQEREFVAGIGWSGGDTLYADSVRGRFTNSRDNSKNTLYLQMNSLRAEDTAVYYCAARQGQYIYSSMRSDSYDYWGQGTLVTVSS
KN046 Alphamab Oncology	PD-L1 × CTLA4	wt IgG1-Fc	Pancreatic cancer	Phase 3 NCT05149326NCT03733951	QVQLVESGGGLVQPGGSLRLSCAASGKMSSRRCMAWFRQAPGKERERVAKLLTTSGSTYLADSVKGRFTISRDNSKNTVYLQMNSLRAEDT AVYYCAADSFEDPTCTLVTSSGAFQYWGQGTLVTVSSGAPQVQLVESGGGLVQPGGSLRLSCAASGYIYSAYCMGWFRQAPGKGLEGV AAIYIGGGSTYYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYCAADVIPTETCLGGSWSGPFGYWGQGTLVTVSSGSEPKSSDK THTCPPCPAPELLGGPSVFLFPPKPKDTLMISRTPEVTCVVVDVSHEDPEVKFNWYVDGVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWL NGKEYKCKVSNKALPAPIEKTISKAKGQPREPQVYTLPPSRDELTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDG SFFLYSKLTVDKSRWQQGNVFSCSVMHEALHNHYTQKSLSLSPGK
Solenokimab (M1095/ALX0761) Avillion	IL17-A x IL17F	HSA	Moderate-to-severe psoriasis	Phase 2 NCT03384745	DVQLVESGGGLVQPGGSLRLSCAASGRTFSSYVVGWFRQAPGKEREFIGAISGSGESIYYAVSEKGRFTISRDNSKNTLYLQMNSLRPEDTAVYYCTA DQEFGYLRFGRSEYWGQGTLVTVSSGGGGSGGGSEVQLVESGGGLVQPGNSLRLSCAASGFTFSSFGMSWVRQAPGKGLEWVSSISGSGSDTL YADSVKGRFTISRDNAKTTLYLQMNSLRPEDTAVYYCTIGGSLSRSSQGTLVTVSSGGGGSGGGSEVQLVESGGGLVQPGGSLRLSCAASGRTYD AMGWLRQAPGKEREFVAAISGSGDDTYYADSVKGRFTISRDNSKNTLYLQMNSLRPEDTAVYYCATRRGLYYVWDANDYENWGQGTLVTVSS
Enristomig (INBRX-105) Inhibrix	PD-L1 × 4-1BB	IgG1-Fc ΔGlu255, ΔLeu 256, ΔLeu257	Solid tumors (HNC)	Phase 2 NCT03809624	EVQLLESGGGEVQPGGSLRLSCAASGGIFAIKPISWYRQAPGKQREWVSTTTSSGATNYAESVKGRFTISRDNAKNTLYLQMSSLRAEDTAVYYCNV FEYWGQGTLVTVKPGGSGGSEVQLLESGGGEVQPGGSLRLSCAASGFSFSINAMGWYRQAPGKRREFVAAIESGRNTVYAESVKGRFTISRDNA KNTVYLQMSSLRAEDTAVYYCGLLKGNRVVSPSVAYWGQGTLVTVKPGGGGDKTHTCPPCPAPGGPSVFLFPPKPKDTLMISRTPEVTCVVVDV SHEDPEVKFNWYVDGVEVHNAKTKPREEQYNSTYRVVSVLTVLHQDWLNGKEYKCKVSNKALPAPIEKTISKAKGQPREPQVYTLPPSR DELTKNQVSLTCLVKGFYPSDIAVEWESNGQPENNYKTTPPVLDSDGSFFLYSKLTVDKSRWQQGNVFSCSVMHEALHNHYTQKSLSLSPGK
LAVA-1207 LAVA Therapeutics	Vδ2-TCR x PMSA	Human IgG1-Fc, modified hinge	Metastatic castration-resistant prostate cancer	Phase 2 NCT05369000
Tarperprumig (ALXN-1820) Alexion/AZ	Properdin	HSA	Sickle cell disease	Phase 2 NCT04631562	QVQLVESGGGLVKPGGSLRLSCAASGRPVSNYAAAWFRQAPGKEREFVSAINWQKTATYADSVKGRFTISRDNAKNSLYLQMNSLRAEDTAVYY CAAVFRVVAPKTQYDYDYWGQGTLVTVSSGGGGEGGGGEGGGGEVQLLESGGGLVQPGGSLRLSCAASGRISSIIHMAWFRQAPGKEREL VSEISRVGTTVYADSVKGRFTISRDNSKNTLYLQMNSLKPEDTAVYYCNALQYEKHGGADYWGQGTLVTVSS
SAR442970 Sanofi	TNFa x OX40L	HSA	Hidradenitis suppurativa	Phase 2 ACT16852
PM1032 Biotheus	CLDN18.2 × 4-1BB	Fc(LALA)	Solid tumours	Phase 2a NCT05839106
HPN536 Harpoon Therapeutics	CD3 × MSLN	HSA	Advanced cancers associated with mesothelin expression.	Phase 1/2a NCT03872206
HPN328 Harpoon Therapeutics	CD3 × DLL3	HSA	Metastatic small cell lung and other neuroendocrine cancers	Phase 1/2 NCT04471727
SAR444200	GPC3 × TCRab	none	Solid tumours	Phase 1/2 NCT05450562
SAR443765 Sanofi	TSLP x IL-13	HSA	Asthma	Phase 1b NCT05366764
BI836880 Boehringer Ingelheim	VEGF x Ang2	HSA	Solid tumors	Phase 1 NCT02674152 NCT02689505
HPN217 Harpoon Therapeutics	CD3 × BCMA	HSA	Hematological R/R MM	Phase 1 NCT04184050
Cancer Hospital Affiliate to Zhengzhou University and Henan Cancer Hospital, Henan, China	CD19 × CD20	CAR-T	Hematological R/R B-cell lymphoma	Phase 1 NCT03881761
IBI333 Innovent Biologics	VEGF-A x VEGF-C	wt IgG1-Fc	Neovascular age-related macular degeneration	Phase 1 CTR20222674
LAVA-1223 LAVA Therapeutics	EGFR x Vγ9 Vδ2	Human IgG1-Fc, modified hinge	EGFR+ solid tumours	Phase 1
PM1003 Biotheus	PD-L1 × 4-1BB	Fc(LALA)	Solid tumours	Phase 1 ChiCTR2100052887
PM1022 Biotheus	PD-L1 ×TIGIT	Fc(LALA)	Solid tumours	Phase 1 NCT05867771
RC1416 Regenecore	IL-4Ra x IL-5	IgG	Asthma	Phase 1 NCT06067490
GTB-3650 Second generation camelid GT Biopharma	CD16 × IL15 × CD33	N/A	Hematological Leukemia	IND-Enabling GMP manufacturing

Inhibrx’s bispecific tetravalent antibody construct (INBRX105) is another example of sdAb domains fused to an Fc domain for half-life extension. The sdAb modules of INBRX105 target PD-L1 and CD137 (4-1BB). The molecule is currently in a Phase 2 clinical trial in combination with pembrolizumab (NCT03809624).

Trispecific T cell activating constructs (TriTACs) offer yet a different architecture incorporating single-domain antibodies. Here, two sdAbs, one targeting a TAA and one binding to HSA, are genetically fused to an anti-CD3 scFv. With a molecular weight of only about 50KDa, these molecules offer the potential for better tumor penetration yet should have a longer plasma half-life mediated by binding to HSA. Such constructs targeting DLL3 (HPN328,) BCMA (HPN217) and MSLN (HPN536) are currently in early-phase clinical trials (NCT04471727, NCT04184050, NCT03872206, respectively; see Table 1 and Figure 5 for a topological representation of some key molecules).^75,76

VHH-based bispecific antibodies can also be part of the targeting component of CAR-T cells, such as in the CD19/CD20-bispecific CAR-T for the treatment of B-cell lymphoma, currently in Phase 1 clinical trials (NCT03881761).⁷⁷ For each of these molecules, different individual VHH building blocks will be tested with different linkers of varying length to arrive at an optimally efficacious, potent and manufacturable bispecific molecule.

The application of VHHs in immune-cell recruitment has been elegantly demonstrated by LAVA Therapeutics. LAVA-051 is a humanized VHH-based bispecific antibody that binds specifically to the Vδ2-TCR chain of gamma/delta T-cells (Vγ9 Vδ2) and the MHC class-I like molecule CD1d that is expressed on tumor cells. Interestingly, LAVA-051 is able to recruit not only Vγ9 Vδ2 T-cells but also type-I NK cells (iNKT) cells, resulting in a novel dual immune cell-mediated killing response on CLL and MM cells.^78,79 LAVA has extended their GammaBody^TM technology with LAVA-1207, a humanized bispecific antibody containing a modified human IgG1-Fc domain for half-life extension that engages Vγ9 Vδ2 and the prostate-specific antigen (PMSA). LAVA-1207 has shown potent and selective tumor cell killing with low treatment emergent adverse events (NCT05369000).⁸⁰

The superior utilization of VHHs in a trispecific antibody format has been shown in the application to trispecific killer cell engagers (TriKES^TM). Initial studies of scFv16-m15–33, a TriKE consisting of anti-CD16 scFv linked to a mutated IL15 followed by an anti-CD33 scFv displayed potent NK recruitment and cytotoxicity, albeit with increased T-cell proliferation.⁸¹ However, previous formats of this TriKE using a wild-type IL15 molecule were nonfunctional, leading the authors to conclude anti-CD16 scFv folding may affect the function of wild-type IL15. Importantly, the direct replacement of the anti-CD16 scFv with a humanized llama anti-CD16 VHH generated a TriKE, cam16-wtIL15–33, that not only allowed the utilization of wild-type IL15 but displayed superior NK cell proliferation, activation and cell killing.⁸² scFv16-m15–33 (GTB-3550) was evaluated in clinical studies as a treatment for CD33-positive myelodysplastic syndromes but then terminated due to the development of the superior nanobody containing TriKE product (GTB-3650, Clinical Trials.gov NCT03214666).

Although camelid-derived VHHs offer a distinct advantage over human single-domain antibodies in terms of favorable biophysical characteristics, recent advances in the production of human VH domain-only platforms have been documented by Harbour Biomed, Nona Biomed and Crescendo Biologics. However, the production of transgenic mice platforms expressing human VH antibody fragments remains a challenging and costly exercise compared to the relatively simple approach of llama immunizations combined with phage display.

VHH design using machine learning

The application of machine learning to VHH design is a rapidly emerging area of research. Given the recency of machine learning methods being applied to VHH design, the relative scarcity of literature in this field highlights an opportunity for substantial breakthroughs. In this section, we highlight both methods for generic antibody design that may be applicable to VHH and those tailored specifically toward VHHs. We highlight methods that include structural prediction, lead identification, synthetic library design, sequence optimization methods, and humanization.^83–85

Computational structural modeling of VHHs

The three-dimensional structure of both the protein target and the VHH is key to developing fully de novo antibodies against a given target. While sequence information provides considerable insight into how the molecule behaves, structural information can reveal important insights, such as which residues are on the surface of the protein, what are the side-chain orientations, and which residues are involved in target binding. In the absence of an experimentally determined structure generated using X-ray crystallography or cryo-EM, computational methods can be used to predict the structure of the protein.^86–89 Traditionally, structural prediction has been performed using homology modeling and an energy minimization refinement step to predict the structures of proteins based on a template structure. The introduction of AlphaFold and subsequently AlphaFold2 (AF2) marked a major breakthrough in the protein structural prediction space.^87,90 AlphaFold uses a deep learning model that efficiently uses evolutionary and structural information to accurately predict the three-dimensional structure of a given protein. Building on this breakthrough, antibody and VHH-specific structural prediction models that show notable progress in predicting these immune molecules have been developed. These include IgFold, ABLooper, NanoNet, ABodyBuilder2, and also its VHH-specific counterpart NanoBodyBuilder2 (Figure 6).^84,86,88,89 It is important to highlight the disparity in the availability of structural data between VHHs and human antibodies that are available for model training and evaluation. As of December 2023, there are 7,977 antibody structures on the Structural Antibody Database (SAbDab),⁹¹ but only 1,186 of these are VHH structures, potentially making VHH structural prediction more challenging. However, given the absence of a light chain or VH/VL interface to predict, VHH structures could be considered less complex to predict structurally compared to traditional antibodies. The advances in antibody and VHH structure prediction have enabled more accurate prediction of other important protein properties, such as solubility, thermostability, aggregation, and binding affinity.^92,93 Despite advances in structural prediction, CDR loop prediction remains a challenge, in particular for CDR3, the most diverse loop in sequence and length.⁹⁴ Additionally, predicting the structure of antibodies and nanobodies using these tools can yield similar yet different predicted structures, introducing uncertainty when using these predicted models to characterize other properties of the molecule.

Figure 6. Overview of machine learning tools for antibodies and VHHs. (A) Tools for structural prediction, encompassing the ImmuneBuilder suite, IgFold, and Nanonet. (B) Language model-based sequence optimization with AbLang, AntiBERTy, IgLM, Sapiens and NanoBERT, alongside Bayesian Optimization techniques like AntBO and LaMBO2. Additionally, humanization tools include Hu-Mab, PLAN and CUMAb.

A four section picture with curved arrows and lines curving into a barrel-like structure representing the three-dimensional structure of a llama heavy chain antibody domain. To the right of this image is a gray rectangular box with letters connected to the gray box by straight lines representing amino acids. To the right of this image is a three-dimensional graph represented by three gray surfaces at right-angles to each other. In the center of this graph is a blue mountain-like structure with yellow peaks. This image represents how Bayesian optimization works. To the right of this graph is a cartoon image of a llama, with an arrow pointing to a cartoon image of a male human figure. This image represents the humanization process.

Predicting the structure of multivalent molecules creates an additional challenge given the conformational flexibility provided by the adjoining linkers used between the individual units of the monovalent molecule. Huang et al. used molecular dynamics simulations in conjunction with AF2 structural predictions to rationalize why multivalent anti-VEGF designed VHHs exhibited more than 100-fold improvement in affinity and ~six-fold increase in VEGF-VEGFR pathway inhibition compared with the monovalent molecule.⁹⁵ This demonstrates the potential of coupling physics-based approaches such as molecular dynamics simulations with machine learning-based structural predictions for the development of superior multivalent VHHs.

Molecular docking can be used to model the structure of an antibody or VHH in complex with its target, but its application is complicated by the flexibility of CDR loops and limited experimental data on epitope–paratope interactions. General multimeric protein structure prediction tools do not yet generalize well to antibody-antigen binding prediction, though recent versions of AF2 demonstrate consistent improvement.^96,97 Notably, AF2 exhibits greater success modeling nanobody-antigen complexes than larger antibody-antigen complexes.⁹⁷ Larger antibody sequence datasets have enabled improved paratope prediction. For instance, Parapred uses deep learning to predict the paratope residues that may be involved in target binding from the CDR sequence alone.⁹⁸ Epitope prediction remains a significant challenge in the field; nonetheless, the application of deep learning techniques has demonstrated improvements in antibody-antigen docking.^99–101 DLAB-Re is one example of a machine learning method designed for the accurate identification and ranking of docked poses.¹⁰² NbX has further demonstrated an ability to improve the ranking of VHH-antigen binding poses.¹⁰³ One challenge with docking predictions is understanding the energy of the systems. Hummer et al. observed that there is currently insufficient experimental data to predict ∆∆G accurately and robustly, with orders of magnitude more likely needed.¹⁰⁴ In contrast, Gaudreault et al. derived a novel scoring method using native output values from the AF2 model that is better calibrated for scoring binding energy than physics-based approaches.⁹⁶

It is clear that limitations still exist with antibody and VHH structural prediction, particularly around CDR prediction, B-cell epitope prediction, and generating accurate antibody-antigen poses. However, structure-based design is accelerating due to the increased availability of accurate antibody and antigen structure models. As more antibody and VHH structures become available and as foundational machine learning models improve, we expect to see substantial progress in the computational prediction of antibody and VHH structure. Hummer et al. notes that although full in silico antibody design is not yet viable, the foundations have been laid.¹⁰⁵ The progress in computational structural prediction not only promises a more comprehensive understanding of antibodies and VHHs but also paves the way for the design of bispecific VHHs by providing crucial insights into the biophysical properties and binding characteristics of individual domains.

Generative machine learning-based approaches to VHH lead identification

Generative machine learning is a class of machine learning that goes beyond predicting values given some input and is capable of generating new data, e.g. sequences or structures, that are distinct from the training data.¹⁰⁶ Generative ML approaches for antibodies typically involve the use of deep learning models, such as graph neural networks, diffusion networks, and large language models (LLMs).¹⁰⁷ These models are trained on a large dataset of antibody or VHH sequences and/or structural data with the goal of learning the underlying patterns inherent to the molecules. These patterns can then be used to inform the design of novel molecules, with some models capable of co-generating sequence and structure simultaneously.¹⁰⁸

While a large set of general protein LLMs (including ProtGPT2,¹⁰⁹ ProGen2,¹¹⁰ ESM-2,¹¹¹ and ProtTrans)¹¹² have been trained on the UniRef¹¹³ corpus of protein sequences, the nuances of antibody sequences have stimulated interest in developing antibody-specific LLMs. A burgeoning set of LLMs tuned to the antibody sequence space have been made available (including AntiBERTY,¹¹⁴ AbLang,¹¹⁵ and IgLM,¹⁰⁷ [Table 2]), all drawing from the OAS database that collates antibody sequences and contains nearly 2 billion sequences spanning 68 different studies⁹ [as of September 2023]. The models are distinct in the subsets of OAS they trained on and the model architecture (BERT vs. GPT2). Importantly, the models differ in the way they handle different subsets of the antibody space: IgLM is conditional on species (human, camelid, etc.); AbLang has separate models for the heavy and light chain; and AntiBERTy uses a single, unconditional model. nanoBERT was recently developed¹¹⁶ and is even more specialized as it is an LLM trained exclusively on VHH sequences from the Integrated Nanobody Database for Immunoinformatics.⁵³

Table 2. Summary of language models applicable to antibodies and VHHs.

Name	Domain	Data	Base Model Type	Reference	Notes
IgLM	Antibody	OAS (558 M subset)	GPT2 (transformer; decoder only)	Generative language modeling for antibody design	Conditional predictions on species
AbLang	Antibody	OAS (14 M heavy chains and 187K light chains)	BERT (transformer; encoder-only)	AbLang: an antibody language model for completing antibody sequences	Separate models for heavy and light chain
AntiBERTy	Antibody	OAS (558 M subset)	BERT (transformer; encoder-only)	Deciphering antibody affinity maturation with language models and weakly supervised learning
nanoBERT	VHH	INDI (10 M VHH chains)	BERT (transformer; encoder-only)	nanoBERT: a deep learning model for gene agnostic navigation of the nanobody mutational space
Sapiens (BioPhi)	Antibody	OAS	Deep neural network based on the transformer encoder architecture	BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning
ProtGPT2	Protein	UniRef50 2021_04	GPT2 (transformer; decoder only)	ProtGPT2 is a deep unsupervised language model for protein design
ProGen2	Protein	UniRef90 & BFD30	BERT (transformer; encoder-only)	ProGen2: Exploring the Boundaries of Protein Language Models
ESM-2	Protein	UniRef50 2021_04	RoBERTa (transformer; encoder-only)	Language models generalize beyond natural proteins
ProtTrans/Prot-T5	Protein	UniRef50 2021_04 (best)	T5 (transformer; encoder-decoder) (best)	ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning	A number of models are trained on several datasets

In practice, generative approaches for antibodies have gained the most traction with diversifying sequences to modulate functional properties. To give the reader a sample of interesting applications, we highlight some case studies below. Shin et al. used a deep generative model to develop a 10⁵ VHH library that illustrated better expression than a 1000-fold larger synthetic library.⁵⁶ Generative deep learning models applied to the CDRs of preexisting binders were able to produce some antibodies that improved binding over wildtype.¹¹⁴ DiffAb uses antibody-antigen bound structure as input into diffusion probabilistic frameworks and equivariant neural networks to generate novel CDRs that are predicted to preserve target binding (though the manuscript lacks experimental validation).¹¹⁷ Hie et al. apply LLMs to a two-round affinity maturation campaign and see variants with improvement in 7- to 170-fold improvements in affinity.¹¹⁸ Remarkably, Hie et al. noted that general protein LLMs exhibited superior performance compared to antibody-specific LLMs in their case studies. We highlight the need for rigorous evaluation and comparison of the LLMs in the context of the target application. For instance, the application of LLMs to the design of multispecific antibodies is likely to require substantial design constraints to overcome the lack of similar training data (linker regions, multivalent designs, etc.).

Computational approaches to VHH lead optimization

Due to the vast sequence space largely resulting from the sequence and length diversity of CDR3 in the heavy chain, performing an efficient search for the optimal sequence across therapeutically relevant parameters (e.g. affinity, stability, purity) using computational approaches is highly challenging.

Several promising machine learning approaches for antibody sequence optimization include Bayesian optimization (BO), genetic algorithms, reinforcement learning, and active learning.^119,120 BO is particularly well suited for navigating the complex design space of antibody and VHH molecules due to its efficient exploration, ability to optimize multiple properties, and flexibility to incorporate diverse input feature spaces.¹¹⁹ By using probabilistic models and making informed decisions about where to sample next, BO effectively balances the trade-off between exploration and exploitation of the design space to converge toward optimal sequences with desirable therapeutic properties. As a first example, AntBO combinatorial BO framework used for the in silico design of CDRH3 that restricts the search to sequences with a favorable developability profile.¹²¹ Similarly, Li et al. coupled LLMs for feature representation with BO to navigate a machine learning-derived fitness landscape, leading to the design of high-affinity scFv libraries.¹²² Finally, LaMBO2 integrates two approaches (NOS for the generation of protein sequences and LaMBO for multiple objective BO with edit-based constraints) to improve antibody expression and binding affinity while maintaining developability.^123,124 Collectively, these methods demonstrate that BO is a powerful tool for designing optimized building blocks of single-domain antibodies, which can be subsequently employed in the construction of multi-specific VHHs with enhanced therapeutic properties.

BO offers a promising approach for optimizing multispecific VHHs and antibodies, presenting opportunities not only at the sequence level but also for optimizing additional features of multispecific antibodies and VHHs. These additional features include linker length, linker composition, the number of VHH or antibody units (valency), and the molecule’s topology. By framing the optimization problem in terms of abstract geometric, topological and biological features, rather than as sequence optimization, the search space can be dramatically reduced whilst explicitly focusing design on geometric and biological features directly relevant to identifying multispecific VHHs with the potential for optimal selectivity, efficacy, and developability. Additionally, optimization of the individual parts of the multispecific molecules helps expedite the design, build, test and learn cycle, crucial in enabling swift iterations and the generation of ample data for refining the models and ultimately reducing the time required to identify lead molecules.

Once a lead VHH has been chosen, humanization is often performed to reduce the risk of immunogenicity. While general strategies for VHH humanization have been established,¹²⁵ several machine learning-based tools, including Hu-Mab,¹²⁶ Sapiens,¹²⁷ PLAN⁸⁵ and CUMab,¹²⁸ have been developed for antibody humanization (Figure 6). The applicability of these machine learning tools for VHH humanization is in doubt as humanizing hallmark residues in framework region 2 unique to VHHs are generally, but not always, found to have a detrimental effect on properties such as binding affinity, solubility, and purity.^37,129,130 Consequently, the machine learning models trained on traditional antibody sequences may not be applicable for VHH humanization. Recently, AbNatiV was developed to assess the degree of humanness and VHH-nativeness of a given sequence. This enables users to fine-tune VHH humanization by ensuring the retention of crucial camelid residues while avoiding especially immunogenic non-human residues.¹³¹

Conclusion

VHHs are a promising class of therapeutic molecules due to their size, modularity, and stability. The modular nature of VHHs allows for the creation of bi- and multispecific antibodies with high specificity and affinity. VHHs are also capable of being grafted together in a highly flexible fashion, allowing the ability to explore a vast combinatorial space in a range of highly adaptable formats, a distinct advantage in the multispecific space. Machine learning is already enhancing various stages of the development of therapeutic VHHs by aiding in processes such as lead discovery, lead optimization, and structural prediction. However, it is evident that the application of machine learning in nanobody development lags behind its use in antibody development. Although there is a critical need for the generation of high-quality, domain-specific data to support the analysis of fragment interactions, with this data there exists a notable opportunity for researchers to propel future breakthroughs in VHH-enabled drug discovery by exploring and enhancing machine learning approaches specifically tailored for VHHs.

Abbreviations

Ab	=	Antibody
ADA	=	Anti-drug antibody
AML	=	Acute myeloid leukemia
APC	=	Antigen-presenting cell
ATUE	=	Antibody understanding evaluation language model
BiTE	=	Bispecific T-cell engager
BO	=	Bayesian optimization
BsAbs	=	Bispecific antibodies
CAR-T	=	chimeric antigen receptor T-cell
CDR	=	Complementarity-determining region (1–3)
CEA	=	Carcinoembryonic antigen
CH1	=	Constant heavy domain 1
CODV	=	Cross-over dual variable
CRS	=	Cytokine release syndrome
Cryo-EM	=	Cryogenic electron microscopy
EGFR	=	Epidermal growth factor receptor
EpCAM	=	Epithelial cell adhesion molecule
Fabs	=	Fragment antigen-binding regions
Fc domain	=	Fragment crystallizable domain
FcRn	=	Neonatal Fc receptor
FW	=	Framework region of an antibody (1–4)
GANS	=	Generative adversarial networks
GEPIA2	=	Gene expression profiling interactive analysis
HCAbs	=	Heavy chain-only antibodies
IgG	=	Immunoglobulin G
IgLM	=	Immunoglobulin language model
kDa	=	Kilodalton
LLMs	=	Large language models
Mab	=	Monoclonal antibody
MD	=	Molecular dynamics
ML	=	Machine learning
NK	=	Natural killer
NLP	=	Natural language processing
NOS	=	Diffusion optimized sampling
OAS	=	Observed antibody space
PEAR	=	Pre-existing anti-drug reactivity
PCR	=	Polymerase chain reaction
PDB	=	Protein data bank
PD-L1	=	Programmed cell death ligand 1
Pmhc	=	Peptide loaded major histocompatibility complex
PLM	=	Protein language model
ROR1	=	Receptor tyrosine kinase like orphan receptor 1
scFv	=	Single-chain variable fragments
sdAb	=	Single-domain antibody
VAEs	=	Variational autoencoders
VHH	=	Variable heavy domain of heavy chain
VH	=	Variable heavy (domain of an antibody)
VL	=	Variable light (domain of an antibody)
TCE	=	T-cell engager
TCR	=	T-cell receptor
TriTE	=	Trispecific T-cell engager
TsAb	=	Tribody format of antibody

Acknowledgments

We would like to thank members of LabGenius for the scientific discussion of the review. In particular, we would like to thank Lucy Shaw, Tonya Frolov and Leonard Wossnig for detailed comments and careful review of the manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The author(s) reported that there is no funding associated with the work featured in this article.