The potential of computational antibody design has been proposed for years but it finally appears to be a reality based on the recent advancement of the world’s first computationally designed antibody drug into a clinical trial. AU-007, the first of its kind in the Aulos Biosciences pipeline, is a computationally designed human antibody to the CD25-interacting domain of IL-2, that in preclinical studies displays strong anti-cancer activity. Such approaches to antibody design represent a real game-changer for antibody-based therapeutics, offering a path forward to targeting previously un-druggable epitopes. In this blog, we discuss the path to this achievement and its potential to change biologic drug development, diagnostics, and potentially the way we utilize antibodies in research and medicine.

So How Does Antibody-Epitope Binding Work?

The binding of the antibody to its epitope is often compared to that of a lock and key, and although this is a useful description, it doesn’t reflect the complexity and dynamics of this binding process

Several factors are involved in antibody-epitope binding; initially, it is the chemical nature and orientation of the amino acid residues of the antibody variable domains and of the epitope that enable the interaction of positively and negatively charged residues to interact with each other. Hydrogen bonds within these amino acids then strengthen the interaction between the antibody and the epitope, which is reinforced by Van der Waals contact and coupled with localized hydrophobic interactions between the two proteins, resulting in the expulsion of water molecules as the proteins come together. Since all of these forces are noncovalent, they are fully reversible and follow the basic thermodynamic principles for reversible biomolecular binding:

where KA is the affinity constant, [Ab-Ag] represents the molar concentration of the antibody-antigen complex, and [Ab] and [Ag] represent the molar concentrations of unoccupied binding sites on the antibody (Ab) or antigen (Ag), respectively.

KA values for antibodies can range from <10⁵/mol to well above 10¹²/mol depending on the antibody and are typically influenced by temperature, pH, and the ionic strength of the solvent. The strength of binding of an antibody to its target is measured by both affinity and avidity. Whereas affinity describes the binding strength at a single binding site, avidity is a measure of the total binding strength and is dependent on the affinity of the antibody for the epitope, the valency of both the antibody and the antigen, coupled with the structural arrangement of the interacting regions of both molecules. All naturally occurring antibodies are multivalent, (IgG are bivalent and IgMs are decavalent), whereas engineered nanobodies are monovalent.

How the Immune System Makes Effective Antibodies

B cells have unique B cell receptors (BCRs) expressed on their surface that are composed of the classic Y-shaped antibody molecule, coupled with a transmembrane domain that links this structure with non-covalently associated Igα/Igβ (CD79a/CD79b) heterodimer. When a B cell interacts with an antigen with which it has high affinity, the antigen-BCR complex is internalized and degraded, and the proteolytic products are presented on the surface of the B cell. The activation of this system results in the stimulation of the B-cell to survive and proliferate and engages other cells including macrophages, dendritic-, and T-cells to mount an effective immune response. As the B cells mature with the help of T-cells, they release the BCR in a soluble format in the form of an antibody.

Antigens have two key elements: a B cell epitope that is recognized by a BCR, which can be a protein, nucleic acid, carbohydrate, or lipid, as well as a T-cell epitope that represents a linear peptide that is displayed on the surface of the B-cell. It is this that elicits help from the T-cell to respond. Strong antigens therefore must possess both B-cell and T-cell epitopes, and typically these are distinct but physically connected regions of the antigen. This is critical because it means that there are real challenges to raising antibodies to many targets that display poor immunogenicity within their antigen structures. Conjugating these antigens to immunogenic carrier proteins and employing adjuvants to improve the immune system’s response to the antigen are often used to help generate antibodies to small molecules and highly conserved proteins. In other instances, novel approaches to antibody generation have been employed, including molecular-based antibody display systems that help to reduce the limitations of immune biology.

Laboratory Engineering of Antibodies

It was the developments in molecular techniques in the early ’90s that enabled the genes for IgG molecules to be cloned into eukaryotic expression vectors (Winter and Milstein, 1991), followed by the application of random or directed mutagenesis that made it possible to optimize recombinant antibodies in the lab (Hoogenboom and Chames, 2000). But the selection of targeted antibodies was accelerated by the application of antibody display techniques that physically link these proteins to the DNA that encodes them. Antibody display systems (phage, bacteria, yeast, ribosomal, and more recently mammalian) have all been successfully used to screen antibody libraries for highly specific antibodies (Valldorf et. al. 2020). These approaches typically involve rounds of antigen-capture screening to enrich the library with clones encoding antibodies with the desired binding characteristics. Indeed, there are now several approved therapeutics that have been developed using display technologies (see Table 1 for a list of examples). Phage display typically employs a solid support matrix (plate of bead column) to capture the phage to the target antigen; with repeated rounds of washing and recapture, the highest affinity clones can be selected. Yeast surface display is often performed using FACS cell sorting. The yeast cells are incubated with fluorescently tagged antigen, and any yeast expressing antibodies that bind this target antigen are sorted out based on this fluorescence signal. This elegant approach enables near real-time assessment of library candidates and control over the binding characteristics, such as affinity and avidity, and the kinetics of binding (Doerner et. al. 2014).

Each display technology comes with its pros and cons, but all are essentially limited by the size of the antibody libraries. Maturation of the clones can be achieved using mutational clone diversification, which can improve binding characteristics. However, enhanced computational algorithms and machine learning approaches to protein structural modeling overcomes many of these restrictions, making it now possible to custom-build antibody paratopes to previously un-targetable protein epitopes, and with unparalleled precision.

Table 1. examples of biologic drugs that have been developed using phage or yeast display, that have received regulatory approval.

Computational Design of Human Antibodies

In silico methods for protein engineering have been under development for several decades and involve both deterministic and stochastic design methods and algorithms. Antibody design specifically focuses on the modeling of the six hypervariable loops in the Complementary Determining Region (CDR) within the antigen-binding domain, to form a custom-built paratope for any given epitope. This Artificial Intelligence (AI) driven approach can also be used to predict the ability of antibodies to penetrate cells, as well as optimize several critical features such as their half-life in vivo, solubility, immunogenicity, and toxicity (Roy et. al. 2017). All of this works to shorten the time and costs of drugs to progress from development into clinical trials and ultimately bring safer drugs into the market.

The latest milestone for this approach comes with the first clinical trial for a computationally designed antibody AU-007 in Australia. AU-007 is a computationally evolved, anti-IL-2 human monoclonal antibody that has been specifically engineered to activate IL-2 against tumors by eliminating the IL-2-driven induction of T_reg expansion. In addition, this antibody was engineered to prevent the binding of IL-2 to vascular endothelium which can lead to vascular leak syndrome and pulmonary edema (see figure 1). These features make this novel antibody very different from other anti-IL-2 biotherapeutics in the clinic that can be toxic, lead to cytokine storm or result in T_reg expansion that suppresses the immune response to cancer. Aulos were able to achieve this by engineering the antibody to bind only to the CD25-binding portion of IL-2, therefore eliminating the binding of IL-2 to trimeric receptors on T_regs, vascular, and pulmonary endothelial cells, while still allowing IL-2 to bind to and expand T_effector cells.

Figure 1. AU-007 is a computationally designed antibody that binds human IL-2 with pMolar affinity and completely inhibits the binding of IL-2 to CD25 without hindering its binding to CD122/CD132. This ultimately prevents IL-2 from binding to the CD25/CD122/CD132 trimeric receptor on T_regs, but not to the CD122/CD132 dimeric receptor expressed on T_effector and NK cells. In fact, AU-007 binding to IL-2 promotes T_effector, NKs, and NKTs activation, while preventing Treg expansion and the resulting negative feedback loop, and the binding of IL-2 to vascular endothelium that can lead to toxicity.

Aulos leveraged AI in a way that mimics the natural selection of antibodies. When facing a new epitope, our immune system looks for an existing antibody (germline of memory) to serve as a template for antibody maturation through somatic hypermutations. The computational platform behind this strategy, developed by Biolojic Design, uses machine learning algorithms that have been trained on published antibody-antigen structures, that include billions of human somatic hypermutations derived from antibody repertoires, and uses this information to predict the impact of specific amino acid residues within the CDRs. From this data, the complete paratope can be fine-tunned for precise binding characteristics. Biolojic Design is now working with several drug companies, including Eli Lilly and Nektar to develop novel biotherapeutics with specific functions including agonism, antagonism, and selective binding that have the potential to serve as molecular switches for clinical applications.

What the Future Holds

With the combination of bioinformatics, X-ray crystallography protein structure analysis, and AI, there is now a real path forward to the engineering of high-performance antibodies targeting previously undruggable targets. Biolojic Design is not the only company taking this approach; Schrodinger has a similar computational platform that is being implemented to advance the preclinical pipelines of several biotech and pharma companies including Bristol Myers, Takeda, and ZaiLabs, with applications in oncology, infectious disease (Natali et. al. 2021) and autoimmunity (Stafford et. al .2020). The SARS-CoV-2 pandemic has highlighted the need for expedited approaches to therapeutic antibody development, and the promise that AI could bring to rapidly addressing emerging variants of this and other viruses (Shan et. al. 2022). Monoclonal antibodies and their derivatives represent one of the most rapidly growing segments of the life science industry. The FDA approved its first therapeutic murine antibody, Orthoclone Okt3 in 1986, and by 2021 more than 100 monoclonal antibody therapeutics have been approved, with more than 15 under review by the FDA or EMA this year.

It has been stated that hybridoma-based antibody development is more art than science. These novel computational approaches are changing this paradigm and unlocking the opportunity to develop antibodies with unique binding and physical characteristics. For biological drugs, this translates into engineering favorable pharmacokinetic properties and long serum half-lives for improved dosing, along with bispecific antibodies with desired activation and T-cell engager properties. But beyond this, the ability to custom-build antibodies has the potential to change the way that we design high-performance, low-cost diagnostics and clinical reagents for specialized platforms such as flow cytometry and tumor imaging, as well as to vindicate expectations that AI-driven antibody engineering will deliver considerable medical benefits.

References

• Winter G, Milstein C. Man-made antibodies. Nature. (1991); 349:293–299.
• Hoogenboom HR, Chames P. Natural and designer binding sites made by phage display technology. Immunol Today. (2000); 21:371–378.
• Valldorf, Bernhard, Hinz, Steffen C., Russo, Giulio, Pekar, Lukas, Mohr, Laura, Klemm, Janina, Doerner, Achim, Krah, Simon, Hust, Michael and Zielonka, Stefan. "Antibody display technologies: selecting the cream of the crop" Biological Chemistry, vol. 403, no. 5-6, (2022), pp. 455-477. https://doi.org/10.1515/hsz-2020-0377.
• Doerner, A., Rhiel, L., Zielonka, S., and Kolmar, H. (2014). Therapeutic antibody engineering by high efficiency cell screening. FEBS (Fed. Eur. Biochem. Soc.) Lett. 588: 278–287, https://doi.org/10.1016/j.febslet.2013.11.025.
• Ankit Roy, Sanjana Nair, Neeladri Sen, Neelesh Soni, M.S. Madhusudhan, In silico methods for design of biological therapeutics. Methods, Vol. 131 (2017) Pp. 33-65. ISSN 1046-2023. https://doi.org/10.1016/j.ymeth.2017.09.008. (https://www.sciencedirect.com/science/article/pii/S1046202317300671)
• Stafford, I.S., Kellermann, M., Mossoto, E. et al. A systematic review of the applications of artificial intelligence and machine learning in autoimmune diseases. npj Digit. Med. 3, 30 (2020). https://doi.org/10.1038/s41746-020-0229-3
• Natali Eriberto N., Babrak Lmar M., Miho Enkelejd. Prospective Artificial Intelligence to Dissect the Dengue Immune Response and Discover Therapeutics. Frontiers in Immunology Vol.12 (2021) DOI=10.3389/fimmu.2021.574411; ISSN=1664-3224
• Shan, S. et. al. Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization. PNAS Vol. 119, No. 11. (2022) 119. https://doi.org/10.1073/pnas.2122954119